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 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 330 {
 331
 332         struct btrfs_fs_info *fs_info = NULL;
 333         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 334                                               rb_node);
 335         if (bnode->root)
 336                 fs_info = bnode->root->fs_info;
 337         btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
 338                     "found at offset %llu\n", (unsigned long long)bytenr);
 339 }
 340
 341 /*
 342  * walk up backref nodes until reach node presents tree root
 343  */
 344 static struct backref_node *walk_up_backref(struct backref_node *node,
 345                                             struct backref_edge *edges[],
 346                                             int *index)
 347 {
 348         struct backref_edge *edge;
 349         int idx = *index;
 350
 351         while (!list_empty(&node->upper)) {
 352                 edge = list_entry(node->upper.next,
 353                                   struct backref_edge, list[LOWER]);
 354                 edges[idx++] = edge;
 355                 node = edge->node[UPPER];
 356         }
 357         BUG_ON(node->detached);
 358         *index = idx;
 359         return node;
 360 }
 361
 362 /*
 363  * walk down backref nodes to find start of next reference path
 364  */
 365 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 366                                               int *index)
 367 {
 368         struct backref_edge *edge;
 369         struct backref_node *lower;
 370         int idx = *index;
 371
 372         while (idx > 0) {
 373                 edge = edges[idx - 1];
 374                 lower = edge->node[LOWER];
 375                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 376                         idx--;
 377                         continue;
 378                 }
 379                 edge = list_entry(edge->list[LOWER].next,
 380                                   struct backref_edge, list[LOWER]);
 381                 edges[idx - 1] = edge;
 382                 *index = idx;
 383                 return edge->node[UPPER];
 384         }
 385         *index = 0;
 386         return NULL;
 387 }
 388
 389 static void unlock_node_buffer(struct backref_node *node)
 390 {
 391         if (node->locked) {
 392                 btrfs_tree_unlock(node->eb);
 393                 node->locked = 0;
 394         }
 395 }
 396
 397 static void drop_node_buffer(struct backref_node *node)
 398 {
 399         if (node->eb) {
 400                 unlock_node_buffer(node);
 401                 free_extent_buffer(node->eb);
 402                 node->eb = NULL;
 403         }
 404 }
 405
 406 static void drop_backref_node(struct backref_cache *tree,
 407                               struct backref_node *node)
 408 {
 409         BUG_ON(!list_empty(&node->upper));
 410
 411         drop_node_buffer(node);
 412         list_del(&node->list);
 413         list_del(&node->lower);
 414         if (!RB_EMPTY_NODE(&node->rb_node))
 415                 rb_erase(&node->rb_node, &tree->rb_root);
 416         free_backref_node(tree, node);
 417 }
 418
 419 /*
 420  * remove a backref node from the backref cache
 421  */
 422 static void remove_backref_node(struct backref_cache *cache,
 423                                 struct backref_node *node)
 424 {
 425         struct backref_node *upper;
 426         struct backref_edge *edge;
 427
 428         if (!node)
 429                 return;
 430
 431         BUG_ON(!node->lowest && !node->detached);
 432         while (!list_empty(&node->upper)) {
 433                 edge = list_entry(node->upper.next, struct backref_edge,
 434                                   list[LOWER]);
 435                 upper = edge->node[UPPER];
 436                 list_del(&edge->list[LOWER]);
 437                 list_del(&edge->list[UPPER]);
 438                 free_backref_edge(cache, edge);
 439
 440                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 441                         BUG_ON(!list_empty(&node->upper));
 442                         drop_backref_node(cache, node);
 443                         node = upper;
 444                         node->lowest = 1;
 445                         continue;
 446                 }
 447                 /*
 448                  * add the node to leaf node list if no other
 449                  * child block cached.
 450                  */
 451                 if (list_empty(&upper->lower)) {
 452                         list_add_tail(&upper->lower, &cache->leaves);
 453                         upper->lowest = 1;
 454                 }
 455         }
 456
 457         drop_backref_node(cache, node);
 458 }
 459
 460 static void update_backref_node(struct backref_cache *cache,
 461                                 struct backref_node *node, u64 bytenr)
 462 {
 463         struct rb_node *rb_node;
 464         rb_erase(&node->rb_node, &cache->rb_root);
 465         node->bytenr = bytenr;
 466         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 467         if (rb_node)
 468                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 469 }
 470
 471 /*
 472  * update backref cache after a transaction commit
 473  */
 474 static int update_backref_cache(struct btrfs_trans_handle *trans,
 475                                 struct backref_cache *cache)
 476 {
 477         struct backref_node *node;
 478         int level = 0;
 479
 480         if (cache->last_trans == 0) {
 481                 cache->last_trans = trans->transid;
 482                 return 0;
 483         }
 484
 485         if (cache->last_trans == trans->transid)
 486                 return 0;
 487
 488         /*
 489          * detached nodes are used to avoid unnecessary backref
 490          * lookup. transaction commit changes the extent tree.
 491          * so the detached nodes are no longer useful.
 492          */
 493         while (!list_empty(&cache->detached)) {
 494                 node = list_entry(cache->detached.next,
 495                                   struct backref_node, list);
 496                 remove_backref_node(cache, node);
 497         }
 498
 499         while (!list_empty(&cache->changed)) {
 500                 node = list_entry(cache->changed.next,
 501                                   struct backref_node, list);
 502                 list_del_init(&node->list);
 503                 BUG_ON(node->pending);
 504                 update_backref_node(cache, node, node->new_bytenr);
 505         }
 506
 507         /*
 508          * some nodes can be left in the pending list if there were
 509          * errors during processing the pending nodes.
 510          */
 511         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 512                 list_for_each_entry(node, &cache->pending[level], list) {
 513                         BUG_ON(!node->pending);
 514                         if (node->bytenr == node->new_bytenr)
 515                                 continue;
 516                         update_backref_node(cache, node, node->new_bytenr);
 517                 }
 518         }
 519
 520         cache->last_trans = 0;
 521         return 1;
 522 }
 523
 524
 525 static int should_ignore_root(struct btrfs_root *root)
 526 {
 527         struct btrfs_root *reloc_root;
 528
 529         if (!root->ref_cows)
 530                 return 0;
 531
 532         reloc_root = root->reloc_root;
 533         if (!reloc_root)
 534                 return 0;
 535
 536         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 537             root->fs_info->running_transaction->transid - 1)
 538                 return 0;
 539         /*
 540          * if there is reloc tree and it was created in previous
 541          * transaction backref lookup can find the reloc tree,
 542          * so backref node for the fs tree root is useless for
 543          * relocation.
 544          */
 545         return 1;
 546 }
 547 /*
 548  * find reloc tree by address of tree root
 549  */
 550 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 551                                           u64 bytenr)
 552 {
 553         struct rb_node *rb_node;
 554         struct mapping_node *node;
 555         struct btrfs_root *root = NULL;
 556
 557         spin_lock(&rc->reloc_root_tree.lock);
 558         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 559         if (rb_node) {
 560                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 561                 root = (struct btrfs_root *)node->data;
 562         }
 563         spin_unlock(&rc->reloc_root_tree.lock);
 564         return root;
 565 }
 566
 567 static int is_cowonly_root(u64 root_objectid)
 568 {
 569         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 570             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 571             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 572             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 573             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 574             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
 575                 return 1;
 576         return 0;
 577 }
 578
 579 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 580                                         u64 root_objectid)
 581 {
 582         struct btrfs_key key;
 583
 584         key.objectid = root_objectid;
 585         key.type = BTRFS_ROOT_ITEM_KEY;
 586         if (is_cowonly_root(root_objectid))
 587                 key.offset = 0;
 588         else
 589                 key.offset = (u64)-1;
 590
 591         return btrfs_read_fs_root_no_name(fs_info, &key);
 592 }
 593
 594 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 595 static noinline_for_stack
 596 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 597                                   struct extent_buffer *leaf,
 598                                   struct btrfs_extent_ref_v0 *ref0)
 599 {
 600         struct btrfs_root *root;
 601         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 602         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 603
 604         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 605
 606         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 607         BUG_ON(IS_ERR(root));
 608
 609         if (root->ref_cows &&
 610             generation != btrfs_root_generation(&root->root_item))
 611                 return NULL;
 612
 613         return root;
 614 }
 615 #endif
 616
 617 static noinline_for_stack
 618 int find_inline_backref(struct extent_buffer *leaf, int slot,
 619                         unsigned long *ptr, unsigned long *end)
 620 {
 621         struct btrfs_key key;
 622         struct btrfs_extent_item *ei;
 623         struct btrfs_tree_block_info *bi;
 624         u32 item_size;
 625
 626         btrfs_item_key_to_cpu(leaf, &key, slot);
 627
 628         item_size = btrfs_item_size_nr(leaf, slot);
 629 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 630         if (item_size < sizeof(*ei)) {
 631                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 632                 return 1;
 633         }
 634 #endif
 635         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 636         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 637                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 638
 639         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 640             item_size <= sizeof(*ei) + sizeof(*bi)) {
 641                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 642                 return 1;
 643         }
 644
 645         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 646                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 647                 *ptr = (unsigned long)(bi + 1);
 648         } else {
 649                 *ptr = (unsigned long)(ei + 1);
 650         }
 651         *end = (unsigned long)ei + item_size;
 652         return 0;
 653 }
 654
 655 /*
 656  * build backref tree for a given tree block. root of the backref tree
 657  * corresponds the tree block, leaves of the backref tree correspond
 658  * roots of b-trees that reference the tree block.
 659  *
 660  * the basic idea of this function is check backrefs of a given block
 661  * to find upper level blocks that refernece the block, and then check
 662  * bakcrefs of these upper level blocks recursively. the recursion stop
 663  * when tree root is reached or backrefs for the block is cached.
 664  *
 665  * NOTE: if we find backrefs for a block are cached, we know backrefs
 666  * for all upper level blocks that directly/indirectly reference the
 667  * block are also cached.
 668  */
 669 static noinline_for_stack
 670 struct backref_node *build_backref_tree(struct reloc_control *rc,
 671                                         struct btrfs_key *node_key,
 672                                         int level, u64 bytenr)
 673 {
 674         struct backref_cache *cache = &rc->backref_cache;
 675         struct btrfs_path *path1;
 676         struct btrfs_path *path2;
 677         struct extent_buffer *eb;
 678         struct btrfs_root *root;
 679         struct backref_node *cur;
 680         struct backref_node *upper;
 681         struct backref_node *lower;
 682         struct backref_node *node = NULL;
 683         struct backref_node *exist = NULL;
 684         struct backref_edge *edge;
 685         struct rb_node *rb_node;
 686         struct btrfs_key key;
 687         unsigned long end;
 688         unsigned long ptr;
 689         LIST_HEAD(list);
 690         LIST_HEAD(useless);
 691         int cowonly;
 692         int ret;
 693         int err = 0;
 694
 695         path1 = btrfs_alloc_path();
 696         path2 = btrfs_alloc_path();
 697         if (!path1 || !path2) {
 698                 err = -ENOMEM;
 699                 goto out;
 700         }
 701         path1->reada = 1;
 702         path2->reada = 2;
 703
 704         node = alloc_backref_node(cache);
 705         if (!node) {
 706                 err = -ENOMEM;
 707                 goto out;
 708         }
 709
 710         node->bytenr = bytenr;
 711         node->level = level;
 712         node->lowest = 1;
 713         cur = node;
 714 again:
 715         end = 0;
 716         ptr = 0;
 717         key.objectid = cur->bytenr;
 718         key.type = BTRFS_METADATA_ITEM_KEY;
 719         key.offset = (u64)-1;
 720
 721         path1->search_commit_root = 1;
 722         path1->skip_locking = 1;
 723         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 724                                 0, 0);
 725         if (ret < 0) {
 726                 err = ret;
 727                 goto out;
 728         }
 729         BUG_ON(!ret || !path1->slots[0]);
 730
 731         path1->slots[0]--;
 732
 733         WARN_ON(cur->checked);
 734         if (!list_empty(&cur->upper)) {
 735                 /*
 736                  * the backref was added previously when processing
 737                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 738                  */
 739                 BUG_ON(!list_is_singular(&cur->upper));
 740                 edge = list_entry(cur->upper.next, struct backref_edge,
 741                                   list[LOWER]);
 742                 BUG_ON(!list_empty(&edge->list[UPPER]));
 743                 exist = edge->node[UPPER];
 744                 /*
 745                  * add the upper level block to pending list if we need
 746                  * check its backrefs
 747                  */
 748                 if (!exist->checked)
 749                         list_add_tail(&edge->list[UPPER], &list);
 750         } else {
 751                 exist = NULL;
 752         }
 753
 754         while (1) {
 755                 cond_resched();
 756                 eb = path1->nodes[0];
 757
 758                 if (ptr >= end) {
 759                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 760                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 761                                 if (ret < 0) {
 762                                         err = ret;
 763                                         goto out;
 764                                 }
 765                                 if (ret > 0)
 766                                         break;
 767                                 eb = path1->nodes[0];
 768                         }
 769
 770                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 771                         if (key.objectid != cur->bytenr) {
 772                                 WARN_ON(exist);
 773                                 break;
 774                         }
 775
 776                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 777                             key.type == BTRFS_METADATA_ITEM_KEY) {
 778                                 ret = find_inline_backref(eb, path1->slots[0],
 779                                                           &ptr, &end);
 780                                 if (ret)
 781                                         goto next;
 782                         }
 783                 }
 784
 785                 if (ptr < end) {
 786                         /* update key for inline back ref */
 787                         struct btrfs_extent_inline_ref *iref;
 788                         iref = (struct btrfs_extent_inline_ref *)ptr;
 789                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 790                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 791                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 792                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 793                 }
 794
 795                 if (exist &&
 796                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 797                       exist->owner == key.offset) ||
 798                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 799                       exist->bytenr == key.offset))) {
 800                         exist = NULL;
 801                         goto next;
 802                 }
 803
 804 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 805                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 806                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 807                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 808                                 struct btrfs_extent_ref_v0 *ref0;
 809                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 810                                                 struct btrfs_extent_ref_v0);
 811                                 if (key.objectid == key.offset) {
 812                                         root = find_tree_root(rc, eb, ref0);
 813                                         if (root && !should_ignore_root(root))
 814                                                 cur->root = root;
 815                                         else
 816                                                 list_add(&cur->list, &useless);
 817                                         break;
 818                                 }
 819                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 820                                                                       ref0)))
 821                                         cur->cowonly = 1;
 822                         }
 823 #else
 824                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 825                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 826 #endif
 827                         if (key.objectid == key.offset) {
 828                                 /*
 829                                  * only root blocks of reloc trees use
 830                                  * backref of this type.
 831                                  */
 832                                 root = find_reloc_root(rc, cur->bytenr);
 833                                 BUG_ON(!root);
 834                                 cur->root = root;
 835                                 break;
 836                         }
 837
 838                         edge = alloc_backref_edge(cache);
 839                         if (!edge) {
 840                                 err = -ENOMEM;
 841                                 goto out;
 842                         }
 843                         rb_node = tree_search(&cache->rb_root, key.offset);
 844                         if (!rb_node) {
 845                                 upper = alloc_backref_node(cache);
 846                                 if (!upper) {
 847                                         free_backref_edge(cache, edge);
 848                                         err = -ENOMEM;
 849                                         goto out;
 850                                 }
 851                                 upper->bytenr = key.offset;
 852                                 upper->level = cur->level + 1;
 853                                 /*
 854                                  *  backrefs for the upper level block isn't
 855                                  *  cached, add the block to pending list
 856                                  */
 857                                 list_add_tail(&edge->list[UPPER], &list);
 858                         } else {
 859                                 upper = rb_entry(rb_node, struct backref_node,
 860                                                  rb_node);
 861                                 BUG_ON(!upper->checked);
 862                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 863                         }
 864                         list_add_tail(&edge->list[LOWER], &cur->upper);
 865                         edge->node[LOWER] = cur;
 866                         edge->node[UPPER] = upper;
 867
 868                         goto next;
 869                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 870                         goto next;
 871                 }
 872
 873                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 874                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 875                 if (IS_ERR(root)) {
 876                         err = PTR_ERR(root);
 877                         goto out;
 878                 }
 879
 880                 if (!root->ref_cows)
 881                         cur->cowonly = 1;
 882
 883                 if (btrfs_root_level(&root->root_item) == cur->level) {
 884                         /* tree root */
 885                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 886                                cur->bytenr);
 887                         if (should_ignore_root(root))
 888                                 list_add(&cur->list, &useless);
 889                         else
 890                                 cur->root = root;
 891                         break;
 892                 }
 893
 894                 level = cur->level + 1;
 895
 896                 /*
 897                  * searching the tree to find upper level blocks
 898                  * reference the block.
 899                  */
 900                 path2->search_commit_root = 1;
 901                 path2->skip_locking = 1;
 902                 path2->lowest_level = level;
 903                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 904                 path2->lowest_level = 0;
 905                 if (ret < 0) {
 906                         err = ret;
 907                         goto out;
 908                 }
 909                 if (ret > 0 && path2->slots[level] > 0)
 910                         path2->slots[level]--;
 911
 912                 eb = path2->nodes[level];
 913                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 914                         cur->bytenr);
 915
 916                 lower = cur;
 917                 for (; level < BTRFS_MAX_LEVEL; level++) {
 918                         if (!path2->nodes[level]) {
 919                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 920                                        lower->bytenr);
 921                                 if (should_ignore_root(root))
 922                                         list_add(&lower->list, &useless);
 923                                 else
 924                                         lower->root = root;
 925                                 break;
 926                         }
 927
 928                         edge = alloc_backref_edge(cache);
 929                         if (!edge) {
 930                                 err = -ENOMEM;
 931                                 goto out;
 932                         }
 933
 934                         eb = path2->nodes[level];
 935                         rb_node = tree_search(&cache->rb_root, eb->start);
 936                         if (!rb_node) {
 937                                 upper = alloc_backref_node(cache);
 938                                 if (!upper) {
 939                                         free_backref_edge(cache, edge);
 940                                         err = -ENOMEM;
 941                                         goto out;
 942                                 }
 943                                 upper->bytenr = eb->start;
 944                                 upper->owner = btrfs_header_owner(eb);
 945                                 upper->level = lower->level + 1;
 946                                 if (!root->ref_cows)
 947                                         upper->cowonly = 1;
 948
 949                                 /*
 950                                  * if we know the block isn't shared
 951                                  * we can void checking its backrefs.
 952                                  */
 953                                 if (btrfs_block_can_be_shared(root, eb))
 954                                         upper->checked = 0;
 955                                 else
 956                                         upper->checked = 1;
 957
 958                                 /*
 959                                  * add the block to pending list if we
 960                                  * need check its backrefs. only block
 961                                  * at 'cur->level + 1' is added to the
 962                                  * tail of pending list. this guarantees
 963                                  * we check backrefs from lower level
 964                                  * blocks to upper level blocks.
 965                                  */
 966                                 if (!upper->checked &&
 967                                     level == cur->level + 1) {
 968                                         list_add_tail(&edge->list[UPPER],
 969                                                       &list);
 970                                 } else
 971                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 972                         } else {
 973                                 upper = rb_entry(rb_node, struct backref_node,
 974                                                  rb_node);
 975                                 BUG_ON(!upper->checked);
 976                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 977                                 if (!upper->owner)
 978                                         upper->owner = btrfs_header_owner(eb);
 979                         }
 980                         list_add_tail(&edge->list[LOWER], &lower->upper);
 981                         edge->node[LOWER] = lower;
 982                         edge->node[UPPER] = upper;
 983
 984                         if (rb_node)
 985                                 break;
 986                         lower = upper;
 987                         upper = NULL;
 988                 }
 989                 btrfs_release_path(path2);
 990 next:
 991                 if (ptr < end) {
 992                         ptr += btrfs_extent_inline_ref_size(key.type);
 993                         if (ptr >= end) {
 994                                 WARN_ON(ptr > end);
 995                                 ptr = 0;
 996                                 end = 0;
 997                         }
 998                 }
 999                 if (ptr >= end)
1000                         path1->slots[0]++;
1001         }
1002         btrfs_release_path(path1);
1003
1004         cur->checked = 1;
1005         WARN_ON(exist);
1006
1007         /* the pending list isn't empty, take the first block to process */
1008         if (!list_empty(&list)) {
1009                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1010                 list_del_init(&edge->list[UPPER]);
1011                 cur = edge->node[UPPER];
1012                 goto again;
1013         }
1014
1015         /*
1016          * everything goes well, connect backref nodes and insert backref nodes
1017          * into the cache.
1018          */
1019         BUG_ON(!node->checked);
1020         cowonly = node->cowonly;
1021         if (!cowonly) {
1022                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1023                                       &node->rb_node);
1024                 if (rb_node)
1025                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1026                 list_add_tail(&node->lower, &cache->leaves);
1027         }
1028
1029         list_for_each_entry(edge, &node->upper, list[LOWER])
1030                 list_add_tail(&edge->list[UPPER], &list);
1031
1032         while (!list_empty(&list)) {
1033                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1034                 list_del_init(&edge->list[UPPER]);
1035                 upper = edge->node[UPPER];
1036                 if (upper->detached) {
1037                         list_del(&edge->list[LOWER]);
1038                         lower = edge->node[LOWER];
1039                         free_backref_edge(cache, edge);
1040                         if (list_empty(&lower->upper))
1041                                 list_add(&lower->list, &useless);
1042                         continue;
1043                 }
1044
1045                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1046                         if (upper->lowest) {
1047                                 list_del_init(&upper->lower);
1048                                 upper->lowest = 0;
1049                         }
1050
1051                         list_add_tail(&edge->list[UPPER], &upper->lower);
1052                         continue;
1053                 }
1054
1055                 BUG_ON(!upper->checked);
1056                 BUG_ON(cowonly != upper->cowonly);
1057                 if (!cowonly) {
1058                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1059                                               &upper->rb_node);
1060                         if (rb_node)
1061                                 backref_tree_panic(rb_node, -EEXIST,
1062                                                    upper->bytenr);
1063                 }
1064
1065                 list_add_tail(&edge->list[UPPER], &upper->lower);
1066
1067                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1068                         list_add_tail(&edge->list[UPPER], &list);
1069         }
1070         /*
1071          * process useless backref nodes. backref nodes for tree leaves
1072          * are deleted from the cache. backref nodes for upper level
1073          * tree blocks are left in the cache to avoid unnecessary backref
1074          * lookup.
1075          */
1076         while (!list_empty(&useless)) {
1077                 upper = list_entry(useless.next, struct backref_node, list);
1078                 list_del_init(&upper->list);
1079                 BUG_ON(!list_empty(&upper->upper));
1080                 if (upper == node)
1081                         node = NULL;
1082                 if (upper->lowest) {
1083                         list_del_init(&upper->lower);
1084                         upper->lowest = 0;
1085                 }
1086                 while (!list_empty(&upper->lower)) {
1087                         edge = list_entry(upper->lower.next,
1088                                           struct backref_edge, list[UPPER]);
1089                         list_del(&edge->list[UPPER]);
1090                         list_del(&edge->list[LOWER]);
1091                         lower = edge->node[LOWER];
1092                         free_backref_edge(cache, edge);
1093
1094                         if (list_empty(&lower->upper))
1095                                 list_add(&lower->list, &useless);
1096                 }
1097                 __mark_block_processed(rc, upper);
1098                 if (upper->level > 0) {
1099                         list_add(&upper->list, &cache->detached);
1100                         upper->detached = 1;
1101                 } else {
1102                         rb_erase(&upper->rb_node, &cache->rb_root);
1103                         free_backref_node(cache, upper);
1104                 }
1105         }
1106 out:
1107         btrfs_free_path(path1);
1108         btrfs_free_path(path2);
1109         if (err) {
1110                 while (!list_empty(&useless)) {
1111                         lower = list_entry(useless.next,
1112                                            struct backref_node, upper);
1113                         list_del_init(&lower->upper);
1114                 }
1115                 upper = node;
1116                 INIT_LIST_HEAD(&list);
1117                 while (upper) {
1118                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1119                                 list_splice_tail(&upper->upper, &list);
1120                                 free_backref_node(cache, upper);
1121                         }
1122
1123                         if (list_empty(&list))
1124                                 break;
1125
1126                         edge = list_entry(list.next, struct backref_edge,
1127                                           list[LOWER]);
1128                         list_del(&edge->list[LOWER]);
1129                         upper = edge->node[UPPER];
1130                         free_backref_edge(cache, edge);
1131                 }
1132                 return ERR_PTR(err);
1133         }
1134         BUG_ON(node && node->detached);
1135         return node;
1136 }
1137
1138 /*
1139  * helper to add backref node for the newly created snapshot.
1140  * the backref node is created by cloning backref node that
1141  * corresponds to root of source tree
1142  */
1143 static int clone_backref_node(struct btrfs_trans_handle *trans,
1144                               struct reloc_control *rc,
1145                               struct btrfs_root *src,
1146                               struct btrfs_root *dest)
1147 {
1148         struct btrfs_root *reloc_root = src->reloc_root;
1149         struct backref_cache *cache = &rc->backref_cache;
1150         struct backref_node *node = NULL;
1151         struct backref_node *new_node;
1152         struct backref_edge *edge;
1153         struct backref_edge *new_edge;
1154         struct rb_node *rb_node;
1155
1156         if (cache->last_trans > 0)
1157                 update_backref_cache(trans, cache);
1158
1159         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1160         if (rb_node) {
1161                 node = rb_entry(rb_node, struct backref_node, rb_node);
1162                 if (node->detached)
1163                         node = NULL;
1164                 else
1165                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1166         }
1167
1168         if (!node) {
1169                 rb_node = tree_search(&cache->rb_root,
1170                                       reloc_root->commit_root->start);
1171                 if (rb_node) {
1172                         node = rb_entry(rb_node, struct backref_node,
1173                                         rb_node);
1174                         BUG_ON(node->detached);
1175                 }
1176         }
1177
1178         if (!node)
1179                 return 0;
1180
1181         new_node = alloc_backref_node(cache);
1182         if (!new_node)
1183                 return -ENOMEM;
1184
1185         new_node->bytenr = dest->node->start;
1186         new_node->level = node->level;
1187         new_node->lowest = node->lowest;
1188         new_node->checked = 1;
1189         new_node->root = dest;
1190
1191         if (!node->lowest) {
1192                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1193                         new_edge = alloc_backref_edge(cache);
1194                         if (!new_edge)
1195                                 goto fail;
1196
1197                         new_edge->node[UPPER] = new_node;
1198                         new_edge->node[LOWER] = edge->node[LOWER];
1199                         list_add_tail(&new_edge->list[UPPER],
1200                                       &new_node->lower);
1201                 }
1202         } else {
1203                 list_add_tail(&new_node->lower, &cache->leaves);
1204         }
1205
1206         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1207                               &new_node->rb_node);
1208         if (rb_node)
1209                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1210
1211         if (!new_node->lowest) {
1212                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1213                         list_add_tail(&new_edge->list[LOWER],
1214                                       &new_edge->node[LOWER]->upper);
1215                 }
1216         }
1217         return 0;
1218 fail:
1219         while (!list_empty(&new_node->lower)) {
1220                 new_edge = list_entry(new_node->lower.next,
1221                                       struct backref_edge, list[UPPER]);
1222                 list_del(&new_edge->list[UPPER]);
1223                 free_backref_edge(cache, new_edge);
1224         }
1225         free_backref_node(cache, new_node);
1226         return -ENOMEM;
1227 }
1228
1229 /*
1230  * helper to add 'address of tree root -> reloc tree' mapping
1231  */
1232 static int __must_check __add_reloc_root(struct btrfs_root *root)
1233 {
1234         struct rb_node *rb_node;
1235         struct mapping_node *node;
1236         struct reloc_control *rc = root->fs_info->reloc_ctl;
1237
1238         node = kmalloc(sizeof(*node), GFP_NOFS);
1239         if (!node)
1240                 return -ENOMEM;
1241
1242         node->bytenr = root->node->start;
1243         node->data = root;
1244
1245         spin_lock(&rc->reloc_root_tree.lock);
1246         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1247                               node->bytenr, &node->rb_node);
1248         spin_unlock(&rc->reloc_root_tree.lock);
1249         if (rb_node) {
1250                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1251                             "for start=%llu while inserting into relocation "
1252                             "tree\n", node->bytenr);
1253                 kfree(node);
1254                 return -EEXIST;
1255         }
1256
1257         list_add_tail(&root->root_list, &rc->reloc_roots);
1258         return 0;
1259 }
1260
1261 /*
1262  * helper to update/delete the 'address of tree root -> reloc tree'
1263  * mapping
1264  */
1265 static int __update_reloc_root(struct btrfs_root *root, int del)
1266 {
1267         struct rb_node *rb_node;
1268         struct mapping_node *node = NULL;
1269         struct reloc_control *rc = root->fs_info->reloc_ctl;
1270
1271         spin_lock(&rc->reloc_root_tree.lock);
1272         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1273                               root->commit_root->start);
1274         if (rb_node) {
1275                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1276                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1277         }
1278         spin_unlock(&rc->reloc_root_tree.lock);
1279
1280         if (!node)
1281                 return 0;
1282         BUG_ON((struct btrfs_root *)node->data != root);
1283
1284         if (!del) {
1285                 spin_lock(&rc->reloc_root_tree.lock);
1286                 node->bytenr = root->node->start;
1287                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1288                                       node->bytenr, &node->rb_node);
1289                 spin_unlock(&rc->reloc_root_tree.lock);
1290                 if (rb_node)
1291                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1292         } else {
1293                 spin_lock(&root->fs_info->trans_lock);
1294                 list_del_init(&root->root_list);
1295                 spin_unlock(&root->fs_info->trans_lock);
1296                 kfree(node);
1297         }
1298         return 0;
1299 }
1300
1301 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1302                                         struct btrfs_root *root, u64 objectid)
1303 {
1304         struct btrfs_root *reloc_root;
1305         struct extent_buffer *eb;
1306         struct btrfs_root_item *root_item;
1307         struct btrfs_key root_key;
1308         int ret;
1309
1310         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1311         BUG_ON(!root_item);
1312
1313         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1314         root_key.type = BTRFS_ROOT_ITEM_KEY;
1315         root_key.offset = objectid;
1316
1317         if (root->root_key.objectid == objectid) {
1318                 /* called by btrfs_init_reloc_root */
1319                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1320                                       BTRFS_TREE_RELOC_OBJECTID);
1321                 BUG_ON(ret);
1322
1323                 btrfs_set_root_last_snapshot(&root->root_item,
1324                                              trans->transid - 1);
1325         } else {
1326                 /*
1327                  * called by btrfs_reloc_post_snapshot_hook.
1328                  * the source tree is a reloc tree, all tree blocks
1329                  * modified after it was created have RELOC flag
1330                  * set in their headers. so it's OK to not update
1331                  * the 'last_snapshot'.
1332                  */
1333                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1334                                       BTRFS_TREE_RELOC_OBJECTID);
1335                 BUG_ON(ret);
1336         }
1337
1338         memcpy(root_item, &root->root_item, sizeof(*root_item));
1339         btrfs_set_root_bytenr(root_item, eb->start);
1340         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1341         btrfs_set_root_generation(root_item, trans->transid);
1342
1343         if (root->root_key.objectid == objectid) {
1344                 btrfs_set_root_refs(root_item, 0);
1345                 memset(&root_item->drop_progress, 0,
1346                        sizeof(struct btrfs_disk_key));
1347                 root_item->drop_level = 0;
1348         }
1349
1350         btrfs_tree_unlock(eb);
1351         free_extent_buffer(eb);
1352
1353         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1354                                 &root_key, root_item);
1355         BUG_ON(ret);
1356         kfree(root_item);
1357
1358         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1359         BUG_ON(IS_ERR(reloc_root));
1360         reloc_root->last_trans = trans->transid;
1361         return reloc_root;
1362 }
1363
1364 /*
1365  * create reloc tree for a given fs tree. reloc tree is just a
1366  * snapshot of the fs tree with special root objectid.
1367  */
1368 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1369                           struct btrfs_root *root)
1370 {
1371         struct btrfs_root *reloc_root;
1372         struct reloc_control *rc = root->fs_info->reloc_ctl;
1373         int clear_rsv = 0;
1374         int ret;
1375
1376         if (root->reloc_root) {
1377                 reloc_root = root->reloc_root;
1378                 reloc_root->last_trans = trans->transid;
1379                 return 0;
1380         }
1381
1382         if (!rc || !rc->create_reloc_tree ||
1383             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1384                 return 0;
1385
1386         if (!trans->block_rsv) {
1387                 trans->block_rsv = rc->block_rsv;
1388                 clear_rsv = 1;
1389         }
1390         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1391         if (clear_rsv)
1392                 trans->block_rsv = NULL;
1393
1394         ret = __add_reloc_root(reloc_root);
1395         BUG_ON(ret < 0);
1396         root->reloc_root = reloc_root;
1397         return 0;
1398 }
1399
1400 /*
1401  * update root item of reloc tree
1402  */
1403 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1404                             struct btrfs_root *root)
1405 {
1406         struct btrfs_root *reloc_root;
1407         struct btrfs_root_item *root_item;
1408         int del = 0;
1409         int ret;
1410
1411         if (!root->reloc_root)
1412                 goto out;
1413
1414         reloc_root = root->reloc_root;
1415         root_item = &reloc_root->root_item;
1416
1417         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1418             btrfs_root_refs(root_item) == 0) {
1419                 root->reloc_root = NULL;
1420                 del = 1;
1421         }
1422
1423         __update_reloc_root(reloc_root, del);
1424
1425         if (reloc_root->commit_root != reloc_root->node) {
1426                 btrfs_set_root_node(root_item, reloc_root->node);
1427                 free_extent_buffer(reloc_root->commit_root);
1428                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1429         }
1430
1431         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1432                                 &reloc_root->root_key, root_item);
1433         BUG_ON(ret);
1434
1435 out:
1436         return 0;
1437 }
1438
1439 /*
1440  * helper to find first cached inode with inode number >= objectid
1441  * in a subvolume
1442  */
1443 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1444 {
1445         struct rb_node *node;
1446         struct rb_node *prev;
1447         struct btrfs_inode *entry;
1448         struct inode *inode;
1449
1450         spin_lock(&root->inode_lock);
1451 again:
1452         node = root->inode_tree.rb_node;
1453         prev = NULL;
1454         while (node) {
1455                 prev = node;
1456                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1457
1458                 if (objectid < btrfs_ino(&entry->vfs_inode))
1459                         node = node->rb_left;
1460                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1461                         node = node->rb_right;
1462                 else
1463                         break;
1464         }
1465         if (!node) {
1466                 while (prev) {
1467                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1468                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1469                                 node = prev;
1470                                 break;
1471                         }
1472                         prev = rb_next(prev);
1473                 }
1474         }
1475         while (node) {
1476                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1477                 inode = igrab(&entry->vfs_inode);
1478                 if (inode) {
1479                         spin_unlock(&root->inode_lock);
1480                         return inode;
1481                 }
1482
1483                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1484                 if (cond_resched_lock(&root->inode_lock))
1485                         goto again;
1486
1487                 node = rb_next(node);
1488         }
1489         spin_unlock(&root->inode_lock);
1490         return NULL;
1491 }
1492
1493 static int in_block_group(u64 bytenr,
1494                           struct btrfs_block_group_cache *block_group)
1495 {
1496         if (bytenr >= block_group->key.objectid &&
1497             bytenr < block_group->key.objectid + block_group->key.offset)
1498                 return 1;
1499         return 0;
1500 }
1501
1502 /*
1503  * get new location of data
1504  */
1505 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1506                             u64 bytenr, u64 num_bytes)
1507 {
1508         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1509         struct btrfs_path *path;
1510         struct btrfs_file_extent_item *fi;
1511         struct extent_buffer *leaf;
1512         int ret;
1513
1514         path = btrfs_alloc_path();
1515         if (!path)
1516                 return -ENOMEM;
1517
1518         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1519         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1520                                        bytenr, 0);
1521         if (ret < 0)
1522                 goto out;
1523         if (ret > 0) {
1524                 ret = -ENOENT;
1525                 goto out;
1526         }
1527
1528         leaf = path->nodes[0];
1529         fi = btrfs_item_ptr(leaf, path->slots[0],
1530                             struct btrfs_file_extent_item);
1531
1532         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1533                btrfs_file_extent_compression(leaf, fi) ||
1534                btrfs_file_extent_encryption(leaf, fi) ||
1535                btrfs_file_extent_other_encoding(leaf, fi));
1536
1537         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1538                 ret = 1;
1539                 goto out;
1540         }
1541
1542         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1543         ret = 0;
1544 out:
1545         btrfs_free_path(path);
1546         return ret;
1547 }
1548
1549 /*
1550  * update file extent items in the tree leaf to point to
1551  * the new locations.
1552  */
1553 static noinline_for_stack
1554 int replace_file_extents(struct btrfs_trans_handle *trans,
1555                          struct reloc_control *rc,
1556                          struct btrfs_root *root,
1557                          struct extent_buffer *leaf)
1558 {
1559         struct btrfs_key key;
1560         struct btrfs_file_extent_item *fi;
1561         struct inode *inode = NULL;
1562         u64 parent;
1563         u64 bytenr;
1564         u64 new_bytenr = 0;
1565         u64 num_bytes;
1566         u64 end;
1567         u32 nritems;
1568         u32 i;
1569         int ret;
1570         int first = 1;
1571         int dirty = 0;
1572
1573         if (rc->stage != UPDATE_DATA_PTRS)
1574                 return 0;
1575
1576         /* reloc trees always use full backref */
1577         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1578                 parent = leaf->start;
1579         else
1580                 parent = 0;
1581
1582         nritems = btrfs_header_nritems(leaf);
1583         for (i = 0; i < nritems; i++) {
1584                 cond_resched();
1585                 btrfs_item_key_to_cpu(leaf, &key, i);
1586                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1587                         continue;
1588                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1589                 if (btrfs_file_extent_type(leaf, fi) ==
1590                     BTRFS_FILE_EXTENT_INLINE)
1591                         continue;
1592                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1593                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1594                 if (bytenr == 0)
1595                         continue;
1596                 if (!in_block_group(bytenr, rc->block_group))
1597                         continue;
1598
1599                 /*
1600                  * if we are modifying block in fs tree, wait for readpage
1601                  * to complete and drop the extent cache
1602                  */
1603                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1604                         if (first) {
1605                                 inode = find_next_inode(root, key.objectid);
1606                                 first = 0;
1607                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1608                                 btrfs_add_delayed_iput(inode);
1609                                 inode = find_next_inode(root, key.objectid);
1610                         }
1611                         if (inode && btrfs_ino(inode) == key.objectid) {
1612                                 end = key.offset +
1613                                       btrfs_file_extent_num_bytes(leaf, fi);
1614                                 WARN_ON(!IS_ALIGNED(key.offset,
1615                                                     root->sectorsize));
1616                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1617                                 end--;
1618                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1619                                                       key.offset, end);
1620                                 if (!ret)
1621                                         continue;
1622
1623                                 btrfs_drop_extent_cache(inode, key.offset, end,
1624                                                         1);
1625                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1626                                               key.offset, end);
1627                         }
1628                 }
1629
1630                 ret = get_new_location(rc->data_inode, &new_bytenr,
1631                                        bytenr, num_bytes);
1632                 if (ret > 0) {
1633                         WARN_ON(1);
1634                         continue;
1635                 }
1636                 BUG_ON(ret < 0);
1637
1638                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1639                 dirty = 1;
1640
1641                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1642                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1643                                            num_bytes, parent,
1644                                            btrfs_header_owner(leaf),
1645                                            key.objectid, key.offset, 1);
1646                 BUG_ON(ret);
1647
1648                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1649                                         parent, btrfs_header_owner(leaf),
1650                                         key.objectid, key.offset, 1);
1651                 BUG_ON(ret);
1652         }
1653         if (dirty)
1654                 btrfs_mark_buffer_dirty(leaf);
1655         if (inode)
1656                 btrfs_add_delayed_iput(inode);
1657         return 0;
1658 }
1659
1660 static noinline_for_stack
1661 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1662                      struct btrfs_path *path, int level)
1663 {
1664         struct btrfs_disk_key key1;
1665         struct btrfs_disk_key key2;
1666         btrfs_node_key(eb, &key1, slot);
1667         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1668         return memcmp(&key1, &key2, sizeof(key1));
1669 }
1670
1671 /*
1672  * try to replace tree blocks in fs tree with the new blocks
1673  * in reloc tree. tree blocks haven't been modified since the
1674  * reloc tree was create can be replaced.
1675  *
1676  * if a block was replaced, level of the block + 1 is returned.
1677  * if no block got replaced, 0 is returned. if there are other
1678  * errors, a negative error number is returned.
1679  */
1680 static noinline_for_stack
1681 int replace_path(struct btrfs_trans_handle *trans,
1682                  struct btrfs_root *dest, struct btrfs_root *src,
1683                  struct btrfs_path *path, struct btrfs_key *next_key,
1684                  int lowest_level, int max_level)
1685 {
1686         struct extent_buffer *eb;
1687         struct extent_buffer *parent;
1688         struct btrfs_key key;
1689         u64 old_bytenr;
1690         u64 new_bytenr;
1691         u64 old_ptr_gen;
1692         u64 new_ptr_gen;
1693         u64 last_snapshot;
1694         u32 blocksize;
1695         int cow = 0;
1696         int level;
1697         int ret;
1698         int slot;
1699
1700         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1701         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1702
1703         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1704 again:
1705         slot = path->slots[lowest_level];
1706         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1707
1708         eb = btrfs_lock_root_node(dest);
1709         btrfs_set_lock_blocking(eb);
1710         level = btrfs_header_level(eb);
1711
1712         if (level < lowest_level) {
1713                 btrfs_tree_unlock(eb);
1714                 free_extent_buffer(eb);
1715                 return 0;
1716         }
1717
1718         if (cow) {
1719                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1720                 BUG_ON(ret);
1721         }
1722         btrfs_set_lock_blocking(eb);
1723
1724         if (next_key) {
1725                 next_key->objectid = (u64)-1;
1726                 next_key->type = (u8)-1;
1727                 next_key->offset = (u64)-1;
1728         }
1729
1730         parent = eb;
1731         while (1) {
1732                 level = btrfs_header_level(parent);
1733                 BUG_ON(level < lowest_level);
1734
1735                 ret = btrfs_bin_search(parent, &key, level, &slot);
1736                 if (ret && slot > 0)
1737                         slot--;
1738
1739                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1740                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1741
1742                 old_bytenr = btrfs_node_blockptr(parent, slot);
1743                 blocksize = btrfs_level_size(dest, level - 1);
1744                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1745
1746                 if (level <= max_level) {
1747                         eb = path->nodes[level];
1748                         new_bytenr = btrfs_node_blockptr(eb,
1749                                                         path->slots[level]);
1750                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1751                                                         path->slots[level]);
1752                 } else {
1753                         new_bytenr = 0;
1754                         new_ptr_gen = 0;
1755                 }
1756
1757                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1758                         WARN_ON(1);
1759                         ret = level;
1760                         break;
1761                 }
1762
1763                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1764                     memcmp_node_keys(parent, slot, path, level)) {
1765                         if (level <= lowest_level) {
1766                                 ret = 0;
1767                                 break;
1768                         }
1769
1770                         eb = read_tree_block(dest, old_bytenr, blocksize,
1771                                              old_ptr_gen);
1772                         if (!eb || !extent_buffer_uptodate(eb)) {
1773                                 ret = (!eb) ? -ENOMEM : -EIO;
1774                                 free_extent_buffer(eb);
1775                                 break;
1776                         }
1777                         btrfs_tree_lock(eb);
1778                         if (cow) {
1779                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1780                                                       slot, &eb);
1781                                 BUG_ON(ret);
1782                         }
1783                         btrfs_set_lock_blocking(eb);
1784
1785                         btrfs_tree_unlock(parent);
1786                         free_extent_buffer(parent);
1787
1788                         parent = eb;
1789                         continue;
1790                 }
1791
1792                 if (!cow) {
1793                         btrfs_tree_unlock(parent);
1794                         free_extent_buffer(parent);
1795                         cow = 1;
1796                         goto again;
1797                 }
1798
1799                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1800                                       path->slots[level]);
1801                 btrfs_release_path(path);
1802
1803                 path->lowest_level = level;
1804                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1805                 path->lowest_level = 0;
1806                 BUG_ON(ret);
1807
1808                 /*
1809                  * swap blocks in fs tree and reloc tree.
1810                  */
1811                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1812                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1813                 btrfs_mark_buffer_dirty(parent);
1814
1815                 btrfs_set_node_blockptr(path->nodes[level],
1816                                         path->slots[level], old_bytenr);
1817                 btrfs_set_node_ptr_generation(path->nodes[level],
1818                                               path->slots[level], old_ptr_gen);
1819                 btrfs_mark_buffer_dirty(path->nodes[level]);
1820
1821                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1822                                         path->nodes[level]->start,
1823                                         src->root_key.objectid, level - 1, 0,
1824                                         1);
1825                 BUG_ON(ret);
1826                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1827                                         0, dest->root_key.objectid, level - 1,
1828                                         0, 1);
1829                 BUG_ON(ret);
1830
1831                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1832                                         path->nodes[level]->start,
1833                                         src->root_key.objectid, level - 1, 0,
1834                                         1);
1835                 BUG_ON(ret);
1836
1837                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1838                                         0, dest->root_key.objectid, level - 1,
1839                                         0, 1);
1840                 BUG_ON(ret);
1841
1842                 btrfs_unlock_up_safe(path, 0);
1843
1844                 ret = level;
1845                 break;
1846         }
1847         btrfs_tree_unlock(parent);
1848         free_extent_buffer(parent);
1849         return ret;
1850 }
1851
1852 /*
1853  * helper to find next relocated block in reloc tree
1854  */
1855 static noinline_for_stack
1856 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1857                        int *level)
1858 {
1859         struct extent_buffer *eb;
1860         int i;
1861         u64 last_snapshot;
1862         u32 nritems;
1863
1864         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1865
1866         for (i = 0; i < *level; i++) {
1867                 free_extent_buffer(path->nodes[i]);
1868                 path->nodes[i] = NULL;
1869         }
1870
1871         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1872                 eb = path->nodes[i];
1873                 nritems = btrfs_header_nritems(eb);
1874                 while (path->slots[i] + 1 < nritems) {
1875                         path->slots[i]++;
1876                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1877                             last_snapshot)
1878                                 continue;
1879
1880                         *level = i;
1881                         return 0;
1882                 }
1883                 free_extent_buffer(path->nodes[i]);
1884                 path->nodes[i] = NULL;
1885         }
1886         return 1;
1887 }
1888
1889 /*
1890  * walk down reloc tree to find relocated block of lowest level
1891  */
1892 static noinline_for_stack
1893 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1894                          int *level)
1895 {
1896         struct extent_buffer *eb = NULL;
1897         int i;
1898         u64 bytenr;
1899         u64 ptr_gen = 0;
1900         u64 last_snapshot;
1901         u32 blocksize;
1902         u32 nritems;
1903
1904         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1905
1906         for (i = *level; i > 0; i--) {
1907                 eb = path->nodes[i];
1908                 nritems = btrfs_header_nritems(eb);
1909                 while (path->slots[i] < nritems) {
1910                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1911                         if (ptr_gen > last_snapshot)
1912                                 break;
1913                         path->slots[i]++;
1914                 }
1915                 if (path->slots[i] >= nritems) {
1916                         if (i == *level)
1917                                 break;
1918                         *level = i + 1;
1919                         return 0;
1920                 }
1921                 if (i == 1) {
1922                         *level = i;
1923                         return 0;
1924                 }
1925
1926                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1927                 blocksize = btrfs_level_size(root, i - 1);
1928                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1929                 if (!eb || !extent_buffer_uptodate(eb)) {
1930                         free_extent_buffer(eb);
1931                         return -EIO;
1932                 }
1933                 BUG_ON(btrfs_header_level(eb) != i - 1);
1934                 path->nodes[i - 1] = eb;
1935                 path->slots[i - 1] = 0;
1936         }
1937         return 1;
1938 }
1939
1940 /*
1941  * invalidate extent cache for file extents whose key in range of
1942  * [min_key, max_key)
1943  */
1944 static int invalidate_extent_cache(struct btrfs_root *root,
1945                                    struct btrfs_key *min_key,
1946                                    struct btrfs_key *max_key)
1947 {
1948         struct inode *inode = NULL;
1949         u64 objectid;
1950         u64 start, end;
1951         u64 ino;
1952
1953         objectid = min_key->objectid;
1954         while (1) {
1955                 cond_resched();
1956                 iput(inode);
1957
1958                 if (objectid > max_key->objectid)
1959                         break;
1960
1961                 inode = find_next_inode(root, objectid);
1962                 if (!inode)
1963                         break;
1964                 ino = btrfs_ino(inode);
1965
1966                 if (ino > max_key->objectid) {
1967                         iput(inode);
1968                         break;
1969                 }
1970
1971                 objectid = ino + 1;
1972                 if (!S_ISREG(inode->i_mode))
1973                         continue;
1974
1975                 if (unlikely(min_key->objectid == ino)) {
1976                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1977                                 continue;
1978                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1979                                 start = 0;
1980                         else {
1981                                 start = min_key->offset;
1982                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1983                         }
1984                 } else {
1985                         start = 0;
1986                 }
1987
1988                 if (unlikely(max_key->objectid == ino)) {
1989                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1990                                 continue;
1991                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1992                                 end = (u64)-1;
1993                         } else {
1994                                 if (max_key->offset == 0)
1995                                         continue;
1996                                 end = max_key->offset;
1997                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1998                                 end--;
1999                         }
2000                 } else {
2001                         end = (u64)-1;
2002                 }
2003
2004                 /* the lock_extent waits for readpage to complete */
2005                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2006                 btrfs_drop_extent_cache(inode, start, end, 1);
2007                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2008         }
2009         return 0;
2010 }
2011
2012 static int find_next_key(struct btrfs_path *path, int level,
2013                          struct btrfs_key *key)
2014
2015 {
2016         while (level < BTRFS_MAX_LEVEL) {
2017                 if (!path->nodes[level])
2018                         break;
2019                 if (path->slots[level] + 1 <
2020                     btrfs_header_nritems(path->nodes[level])) {
2021                         btrfs_node_key_to_cpu(path->nodes[level], key,
2022                                               path->slots[level] + 1);
2023                         return 0;
2024                 }
2025                 level++;
2026         }
2027         return 1;
2028 }
2029
2030 /*
2031  * merge the relocated tree blocks in reloc tree with corresponding
2032  * fs tree.
2033  */
2034 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2035                                                struct btrfs_root *root)
2036 {
2037         LIST_HEAD(inode_list);
2038         struct btrfs_key key;
2039         struct btrfs_key next_key;
2040         struct btrfs_trans_handle *trans;
2041         struct btrfs_root *reloc_root;
2042         struct btrfs_root_item *root_item;
2043         struct btrfs_path *path;
2044         struct extent_buffer *leaf;
2045         int level;
2046         int max_level;
2047         int replaced = 0;
2048         int ret;
2049         int err = 0;
2050         u32 min_reserved;
2051
2052         path = btrfs_alloc_path();
2053         if (!path)
2054                 return -ENOMEM;
2055         path->reada = 1;
2056
2057         reloc_root = root->reloc_root;
2058         root_item = &reloc_root->root_item;
2059
2060         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2061                 level = btrfs_root_level(root_item);
2062                 extent_buffer_get(reloc_root->node);
2063                 path->nodes[level] = reloc_root->node;
2064                 path->slots[level] = 0;
2065         } else {
2066                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2067
2068                 level = root_item->drop_level;
2069                 BUG_ON(level == 0);
2070                 path->lowest_level = level;
2071                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2072                 path->lowest_level = 0;
2073                 if (ret < 0) {
2074                         btrfs_free_path(path);
2075                         return ret;
2076                 }
2077
2078                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2079                                       path->slots[level]);
2080                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2081
2082                 btrfs_unlock_up_safe(path, 0);
2083         }
2084
2085         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2086         memset(&next_key, 0, sizeof(next_key));
2087
2088         while (1) {
2089                 trans = btrfs_start_transaction(root, 0);
2090                 BUG_ON(IS_ERR(trans));
2091                 trans->block_rsv = rc->block_rsv;
2092
2093                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2094                                              BTRFS_RESERVE_FLUSH_ALL);
2095                 if (ret) {
2096                         BUG_ON(ret != -EAGAIN);
2097                         ret = btrfs_commit_transaction(trans, root);
2098                         BUG_ON(ret);
2099                         continue;
2100                 }
2101
2102                 replaced = 0;
2103                 max_level = level;
2104
2105                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2106                 if (ret < 0) {
2107                         err = ret;
2108                         goto out;
2109                 }
2110                 if (ret > 0)
2111                         break;
2112
2113                 if (!find_next_key(path, level, &key) &&
2114                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2115                         ret = 0;
2116                 } else {
2117                         ret = replace_path(trans, root, reloc_root, path,
2118                                            &next_key, level, max_level);
2119                 }
2120                 if (ret < 0) {
2121                         err = ret;
2122                         goto out;
2123                 }
2124
2125                 if (ret > 0) {
2126                         level = ret;
2127                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2128                                               path->slots[level]);
2129                         replaced = 1;
2130                 }
2131
2132                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2133                 if (ret > 0)
2134                         break;
2135
2136                 BUG_ON(level == 0);
2137                 /*
2138                  * save the merging progress in the drop_progress.
2139                  * this is OK since root refs == 1 in this case.
2140                  */
2141                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2142                                path->slots[level]);
2143                 root_item->drop_level = level;
2144
2145                 btrfs_end_transaction_throttle(trans, root);
2146
2147                 btrfs_btree_balance_dirty(root);
2148
2149                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2150                         invalidate_extent_cache(root, &key, &next_key);
2151         }
2152
2153         /*
2154          * handle the case only one block in the fs tree need to be
2155          * relocated and the block is tree root.
2156          */
2157         leaf = btrfs_lock_root_node(root);
2158         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2159         btrfs_tree_unlock(leaf);
2160         free_extent_buffer(leaf);
2161         if (ret < 0)
2162                 err = ret;
2163 out:
2164         btrfs_free_path(path);
2165
2166         if (err == 0) {
2167                 memset(&root_item->drop_progress, 0,
2168                        sizeof(root_item->drop_progress));
2169                 root_item->drop_level = 0;
2170                 btrfs_set_root_refs(root_item, 0);
2171                 btrfs_update_reloc_root(trans, root);
2172         }
2173
2174         btrfs_end_transaction_throttle(trans, root);
2175
2176         btrfs_btree_balance_dirty(root);
2177
2178         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2179                 invalidate_extent_cache(root, &key, &next_key);
2180
2181         return err;
2182 }
2183
2184 static noinline_for_stack
2185 int prepare_to_merge(struct reloc_control *rc, int err)
2186 {
2187         struct btrfs_root *root = rc->extent_root;
2188         struct btrfs_root *reloc_root;
2189         struct btrfs_trans_handle *trans;
2190         LIST_HEAD(reloc_roots);
2191         u64 num_bytes = 0;
2192         int ret;
2193
2194         mutex_lock(&root->fs_info->reloc_mutex);
2195         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2196         rc->merging_rsv_size += rc->nodes_relocated * 2;
2197         mutex_unlock(&root->fs_info->reloc_mutex);
2198
2199 again:
2200         if (!err) {
2201                 num_bytes = rc->merging_rsv_size;
2202                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2203                                           BTRFS_RESERVE_FLUSH_ALL);
2204                 if (ret)
2205                         err = ret;
2206         }
2207
2208         trans = btrfs_join_transaction(rc->extent_root);
2209         if (IS_ERR(trans)) {
2210                 if (!err)
2211                         btrfs_block_rsv_release(rc->extent_root,
2212                                                 rc->block_rsv, num_bytes);
2213                 return PTR_ERR(trans);
2214         }
2215
2216         if (!err) {
2217                 if (num_bytes != rc->merging_rsv_size) {
2218                         btrfs_end_transaction(trans, rc->extent_root);
2219                         btrfs_block_rsv_release(rc->extent_root,
2220                                                 rc->block_rsv, num_bytes);
2221                         goto again;
2222                 }
2223         }
2224
2225         rc->merge_reloc_tree = 1;
2226
2227         while (!list_empty(&rc->reloc_roots)) {
2228                 reloc_root = list_entry(rc->reloc_roots.next,
2229                                         struct btrfs_root, root_list);
2230                 list_del_init(&reloc_root->root_list);
2231
2232                 root = read_fs_root(reloc_root->fs_info,
2233                                     reloc_root->root_key.offset);
2234                 BUG_ON(IS_ERR(root));
2235                 BUG_ON(root->reloc_root != reloc_root);
2236
2237                 /*
2238                  * set reference count to 1, so btrfs_recover_relocation
2239                  * knows it should resumes merging
2240                  */
2241                 if (!err)
2242                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2243                 btrfs_update_reloc_root(trans, root);
2244
2245                 list_add(&reloc_root->root_list, &reloc_roots);
2246         }
2247
2248         list_splice(&reloc_roots, &rc->reloc_roots);
2249
2250         if (!err)
2251                 btrfs_commit_transaction(trans, rc->extent_root);
2252         else
2253                 btrfs_end_transaction(trans, rc->extent_root);
2254         return err;
2255 }
2256
2257 static noinline_for_stack
2258 void free_reloc_roots(struct list_head *list)
2259 {
2260         struct btrfs_root *reloc_root;
2261
2262         while (!list_empty(list)) {
2263                 reloc_root = list_entry(list->next, struct btrfs_root,
2264                                         root_list);
2265                 __update_reloc_root(reloc_root, 1);
2266                 free_extent_buffer(reloc_root->node);
2267                 free_extent_buffer(reloc_root->commit_root);
2268                 kfree(reloc_root);
2269         }
2270 }
2271
2272 static noinline_for_stack
2273 int merge_reloc_roots(struct reloc_control *rc)
2274 {
2275         struct btrfs_root *root;
2276         struct btrfs_root *reloc_root;
2277         LIST_HEAD(reloc_roots);
2278         int found = 0;
2279         int ret = 0;
2280 again:
2281         root = rc->extent_root;
2282
2283         /*
2284          * this serializes us with btrfs_record_root_in_transaction,
2285          * we have to make sure nobody is in the middle of
2286          * adding their roots to the list while we are
2287          * doing this splice
2288          */
2289         mutex_lock(&root->fs_info->reloc_mutex);
2290         list_splice_init(&rc->reloc_roots, &reloc_roots);
2291         mutex_unlock(&root->fs_info->reloc_mutex);
2292
2293         while (!list_empty(&reloc_roots)) {
2294                 found = 1;
2295                 reloc_root = list_entry(reloc_roots.next,
2296                                         struct btrfs_root, root_list);
2297
2298                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2299                         root = read_fs_root(reloc_root->fs_info,
2300                                             reloc_root->root_key.offset);
2301                         BUG_ON(IS_ERR(root));
2302                         BUG_ON(root->reloc_root != reloc_root);
2303
2304                         ret = merge_reloc_root(rc, root);
2305                         if (ret)
2306                                 goto out;
2307                 } else {
2308                         list_del_init(&reloc_root->root_list);
2309                 }
2310                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2311                 if (ret < 0) {
2312                         if (list_empty(&reloc_root->root_list))
2313                                 list_add_tail(&reloc_root->root_list,
2314                                               &reloc_roots);
2315                         goto out;
2316                 }
2317         }
2318
2319         if (found) {
2320                 found = 0;
2321                 goto again;
2322         }
2323 out:
2324         if (ret) {
2325                 btrfs_std_error(root->fs_info, ret);
2326                 if (!list_empty(&reloc_roots))
2327                         free_reloc_roots(&reloc_roots);
2328         }
2329
2330         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2331         return ret;
2332 }
2333
2334 static void free_block_list(struct rb_root *blocks)
2335 {
2336         struct tree_block *block;
2337         struct rb_node *rb_node;
2338         while ((rb_node = rb_first(blocks))) {
2339                 block = rb_entry(rb_node, struct tree_block, rb_node);
2340                 rb_erase(rb_node, blocks);
2341                 kfree(block);
2342         }
2343 }
2344
2345 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2346                                       struct btrfs_root *reloc_root)
2347 {
2348         struct btrfs_root *root;
2349
2350         if (reloc_root->last_trans == trans->transid)
2351                 return 0;
2352
2353         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2354         BUG_ON(IS_ERR(root));
2355         BUG_ON(root->reloc_root != reloc_root);
2356
2357         return btrfs_record_root_in_trans(trans, root);
2358 }
2359
2360 static noinline_for_stack
2361 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2362                                      struct reloc_control *rc,
2363                                      struct backref_node *node,
2364                                      struct backref_edge *edges[], int *nr)
2365 {
2366         struct backref_node *next;
2367         struct btrfs_root *root;
2368         int index = 0;
2369
2370         next = node;
2371         while (1) {
2372                 cond_resched();
2373                 next = walk_up_backref(next, edges, &index);
2374                 root = next->root;
2375                 BUG_ON(!root);
2376                 BUG_ON(!root->ref_cows);
2377
2378                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2379                         record_reloc_root_in_trans(trans, root);
2380                         break;
2381                 }
2382
2383                 btrfs_record_root_in_trans(trans, root);
2384                 root = root->reloc_root;
2385
2386                 if (next->new_bytenr != root->node->start) {
2387                         BUG_ON(next->new_bytenr);
2388                         BUG_ON(!list_empty(&next->list));
2389                         next->new_bytenr = root->node->start;
2390                         next->root = root;
2391                         list_add_tail(&next->list,
2392                                       &rc->backref_cache.changed);
2393                         __mark_block_processed(rc, next);
2394                         break;
2395                 }
2396
2397                 WARN_ON(1);
2398                 root = NULL;
2399                 next = walk_down_backref(edges, &index);
2400                 if (!next || next->level <= node->level)
2401                         break;
2402         }
2403         if (!root)
2404                 return NULL;
2405
2406         *nr = index;
2407         next = node;
2408         /* setup backref node path for btrfs_reloc_cow_block */
2409         while (1) {
2410                 rc->backref_cache.path[next->level] = next;
2411                 if (--index < 0)
2412                         break;
2413                 next = edges[index]->node[UPPER];
2414         }
2415         return root;
2416 }
2417
2418 /*
2419  * select a tree root for relocation. return NULL if the block
2420  * is reference counted. we should use do_relocation() in this
2421  * case. return a tree root pointer if the block isn't reference
2422  * counted. return -ENOENT if the block is root of reloc tree.
2423  */
2424 static noinline_for_stack
2425 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2426                                    struct backref_node *node)
2427 {
2428         struct backref_node *next;
2429         struct btrfs_root *root;
2430         struct btrfs_root *fs_root = NULL;
2431         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2432         int index = 0;
2433
2434         next = node;
2435         while (1) {
2436                 cond_resched();
2437                 next = walk_up_backref(next, edges, &index);
2438                 root = next->root;
2439                 BUG_ON(!root);
2440
2441                 /* no other choice for non-references counted tree */
2442                 if (!root->ref_cows)
2443                         return root;
2444
2445                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2446                         fs_root = root;
2447
2448                 if (next != node)
2449                         return NULL;
2450
2451                 next = walk_down_backref(edges, &index);
2452                 if (!next || next->level <= node->level)
2453                         break;
2454         }
2455
2456         if (!fs_root)
2457                 return ERR_PTR(-ENOENT);
2458         return fs_root;
2459 }
2460
2461 static noinline_for_stack
2462 u64 calcu_metadata_size(struct reloc_control *rc,
2463                         struct backref_node *node, int reserve)
2464 {
2465         struct backref_node *next = node;
2466         struct backref_edge *edge;
2467         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2468         u64 num_bytes = 0;
2469         int index = 0;
2470
2471         BUG_ON(reserve && node->processed);
2472
2473         while (next) {
2474                 cond_resched();
2475                 while (1) {
2476                         if (next->processed && (reserve || next != node))
2477                                 break;
2478
2479                         num_bytes += btrfs_level_size(rc->extent_root,
2480                                                       next->level);
2481
2482                         if (list_empty(&next->upper))
2483                                 break;
2484
2485                         edge = list_entry(next->upper.next,
2486                                           struct backref_edge, list[LOWER]);
2487                         edges[index++] = edge;
2488                         next = edge->node[UPPER];
2489                 }
2490                 next = walk_down_backref(edges, &index);
2491         }
2492         return num_bytes;
2493 }
2494
2495 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2496                                   struct reloc_control *rc,
2497                                   struct backref_node *node)
2498 {
2499         struct btrfs_root *root = rc->extent_root;
2500         u64 num_bytes;
2501         int ret;
2502
2503         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2504
2505         trans->block_rsv = rc->block_rsv;
2506         ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2507                                   BTRFS_RESERVE_FLUSH_ALL);
2508         if (ret) {
2509                 if (ret == -EAGAIN)
2510                         rc->commit_transaction = 1;
2511                 return ret;
2512         }
2513
2514         return 0;
2515 }
2516
2517 static void release_metadata_space(struct reloc_control *rc,
2518                                    struct backref_node *node)
2519 {
2520         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2521         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2522 }
2523
2524 /*
2525  * relocate a block tree, and then update pointers in upper level
2526  * blocks that reference the block to point to the new location.
2527  *
2528  * if called by link_to_upper, the block has already been relocated.
2529  * in that case this function just updates pointers.
2530  */
2531 static int do_relocation(struct btrfs_trans_handle *trans,
2532                          struct reloc_control *rc,
2533                          struct backref_node *node,
2534                          struct btrfs_key *key,
2535                          struct btrfs_path *path, int lowest)
2536 {
2537         struct backref_node *upper;
2538         struct backref_edge *edge;
2539         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2540         struct btrfs_root *root;
2541         struct extent_buffer *eb;
2542         u32 blocksize;
2543         u64 bytenr;
2544         u64 generation;
2545         int nr;
2546         int slot;
2547         int ret;
2548         int err = 0;
2549
2550         BUG_ON(lowest && node->eb);
2551
2552         path->lowest_level = node->level + 1;
2553         rc->backref_cache.path[node->level] = node;
2554         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2555                 cond_resched();
2556
2557                 upper = edge->node[UPPER];
2558                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2559                 BUG_ON(!root);
2560
2561                 if (upper->eb && !upper->locked) {
2562                         if (!lowest) {
2563                                 ret = btrfs_bin_search(upper->eb, key,
2564                                                        upper->level, &slot);
2565                                 BUG_ON(ret);
2566                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2567                                 if (node->eb->start == bytenr)
2568                                         goto next;
2569                         }
2570                         drop_node_buffer(upper);
2571                 }
2572
2573                 if (!upper->eb) {
2574                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2575                         if (ret < 0) {
2576                                 err = ret;
2577                                 break;
2578                         }
2579                         BUG_ON(ret > 0);
2580
2581                         if (!upper->eb) {
2582                                 upper->eb = path->nodes[upper->level];
2583                                 path->nodes[upper->level] = NULL;
2584                         } else {
2585                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2586                         }
2587
2588                         upper->locked = 1;
2589                         path->locks[upper->level] = 0;
2590
2591                         slot = path->slots[upper->level];
2592                         btrfs_release_path(path);
2593                 } else {
2594                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2595                                                &slot);
2596                         BUG_ON(ret);
2597                 }
2598
2599                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2600                 if (lowest) {
2601                         BUG_ON(bytenr != node->bytenr);
2602                 } else {
2603                         if (node->eb->start == bytenr)
2604                                 goto next;
2605                 }
2606
2607                 blocksize = btrfs_level_size(root, node->level);
2608                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2609                 eb = read_tree_block(root, bytenr, blocksize, generation);
2610                 if (!eb || !extent_buffer_uptodate(eb)) {
2611                         free_extent_buffer(eb);
2612                         err = -EIO;
2613                         goto next;
2614                 }
2615                 btrfs_tree_lock(eb);
2616                 btrfs_set_lock_blocking(eb);
2617
2618                 if (!node->eb) {
2619                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2620                                               slot, &eb);
2621                         btrfs_tree_unlock(eb);
2622                         free_extent_buffer(eb);
2623                         if (ret < 0) {
2624                                 err = ret;
2625                                 goto next;
2626                         }
2627                         BUG_ON(node->eb != eb);
2628                 } else {
2629                         btrfs_set_node_blockptr(upper->eb, slot,
2630                                                 node->eb->start);
2631                         btrfs_set_node_ptr_generation(upper->eb, slot,
2632                                                       trans->transid);
2633                         btrfs_mark_buffer_dirty(upper->eb);
2634
2635                         ret = btrfs_inc_extent_ref(trans, root,
2636                                                 node->eb->start, blocksize,
2637                                                 upper->eb->start,
2638                                                 btrfs_header_owner(upper->eb),
2639                                                 node->level, 0, 1);
2640                         BUG_ON(ret);
2641
2642                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2643                         BUG_ON(ret);
2644                 }
2645 next:
2646                 if (!upper->pending)
2647                         drop_node_buffer(upper);
2648                 else
2649                         unlock_node_buffer(upper);
2650                 if (err)
2651                         break;
2652         }
2653
2654         if (!err && node->pending) {
2655                 drop_node_buffer(node);
2656                 list_move_tail(&node->list, &rc->backref_cache.changed);
2657                 node->pending = 0;
2658         }
2659
2660         path->lowest_level = 0;
2661         BUG_ON(err == -ENOSPC);
2662         return err;
2663 }
2664
2665 static int link_to_upper(struct btrfs_trans_handle *trans,
2666                          struct reloc_control *rc,
2667                          struct backref_node *node,
2668                          struct btrfs_path *path)
2669 {
2670         struct btrfs_key key;
2671
2672         btrfs_node_key_to_cpu(node->eb, &key, 0);
2673         return do_relocation(trans, rc, node, &key, path, 0);
2674 }
2675
2676 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2677                                 struct reloc_control *rc,
2678                                 struct btrfs_path *path, int err)
2679 {
2680         LIST_HEAD(list);
2681         struct backref_cache *cache = &rc->backref_cache;
2682         struct backref_node *node;
2683         int level;
2684         int ret;
2685
2686         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2687                 while (!list_empty(&cache->pending[level])) {
2688                         node = list_entry(cache->pending[level].next,
2689                                           struct backref_node, list);
2690                         list_move_tail(&node->list, &list);
2691                         BUG_ON(!node->pending);
2692
2693                         if (!err) {
2694                                 ret = link_to_upper(trans, rc, node, path);
2695                                 if (ret < 0)
2696                                         err = ret;
2697                         }
2698                 }
2699                 list_splice_init(&list, &cache->pending[level]);
2700         }
2701         return err;
2702 }
2703
2704 static void mark_block_processed(struct reloc_control *rc,
2705                                  u64 bytenr, u32 blocksize)
2706 {
2707         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2708                         EXTENT_DIRTY, GFP_NOFS);
2709 }
2710
2711 static void __mark_block_processed(struct reloc_control *rc,
2712                                    struct backref_node *node)
2713 {
2714         u32 blocksize;
2715         if (node->level == 0 ||
2716             in_block_group(node->bytenr, rc->block_group)) {
2717                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2718                 mark_block_processed(rc, node->bytenr, blocksize);
2719         }
2720         node->processed = 1;
2721 }
2722
2723 /*
2724  * mark a block and all blocks directly/indirectly reference the block
2725  * as processed.
2726  */
2727 static void update_processed_blocks(struct reloc_control *rc,
2728                                     struct backref_node *node)
2729 {
2730         struct backref_node *next = node;
2731         struct backref_edge *edge;
2732         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2733         int index = 0;
2734
2735         while (next) {
2736                 cond_resched();
2737                 while (1) {
2738                         if (next->processed)
2739                                 break;
2740
2741                         __mark_block_processed(rc, next);
2742
2743                         if (list_empty(&next->upper))
2744                                 break;
2745
2746                         edge = list_entry(next->upper.next,
2747                                           struct backref_edge, list[LOWER]);
2748                         edges[index++] = edge;
2749                         next = edge->node[UPPER];
2750                 }
2751                 next = walk_down_backref(edges, &index);
2752         }
2753 }
2754
2755 static int tree_block_processed(u64 bytenr, u32 blocksize,
2756                                 struct reloc_control *rc)
2757 {
2758         if (test_range_bit(&rc->processed_blocks, bytenr,
2759                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2760                 return 1;
2761         return 0;
2762 }
2763
2764 static int get_tree_block_key(struct reloc_control *rc,
2765                               struct tree_block *block)
2766 {
2767         struct extent_buffer *eb;
2768
2769         BUG_ON(block->key_ready);
2770         eb = read_tree_block(rc->extent_root, block->bytenr,
2771                              block->key.objectid, block->key.offset);
2772         if (!eb || !extent_buffer_uptodate(eb)) {
2773                 free_extent_buffer(eb);
2774                 return -EIO;
2775         }
2776         WARN_ON(btrfs_header_level(eb) != block->level);
2777         if (block->level == 0)
2778                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2779         else
2780                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2781         free_extent_buffer(eb);
2782         block->key_ready = 1;
2783         return 0;
2784 }
2785
2786 static int reada_tree_block(struct reloc_control *rc,
2787                             struct tree_block *block)
2788 {
2789         BUG_ON(block->key_ready);
2790         if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2791                 readahead_tree_block(rc->extent_root, block->bytenr,
2792                                      block->key.objectid,
2793                                      rc->extent_root->leafsize);
2794         else
2795                 readahead_tree_block(rc->extent_root, block->bytenr,
2796                                      block->key.objectid, block->key.offset);
2797         return 0;
2798 }
2799
2800 /*
2801  * helper function to relocate a tree block
2802  */
2803 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2804                                 struct reloc_control *rc,
2805                                 struct backref_node *node,
2806                                 struct btrfs_key *key,
2807                                 struct btrfs_path *path)
2808 {
2809         struct btrfs_root *root;
2810         int release = 0;
2811         int ret = 0;
2812
2813         if (!node)
2814                 return 0;
2815
2816         BUG_ON(node->processed);
2817         root = select_one_root(trans, node);
2818         if (root == ERR_PTR(-ENOENT)) {
2819                 update_processed_blocks(rc, node);
2820                 goto out;
2821         }
2822
2823         if (!root || root->ref_cows) {
2824                 ret = reserve_metadata_space(trans, rc, node);
2825                 if (ret)
2826                         goto out;
2827                 release = 1;
2828         }
2829
2830         if (root) {
2831                 if (root->ref_cows) {
2832                         BUG_ON(node->new_bytenr);
2833                         BUG_ON(!list_empty(&node->list));
2834                         btrfs_record_root_in_trans(trans, root);
2835                         root = root->reloc_root;
2836                         node->new_bytenr = root->node->start;
2837                         node->root = root;
2838                         list_add_tail(&node->list, &rc->backref_cache.changed);
2839                 } else {
2840                         path->lowest_level = node->level;
2841                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2842                         btrfs_release_path(path);
2843                         if (ret > 0)
2844                                 ret = 0;
2845                 }
2846                 if (!ret)
2847                         update_processed_blocks(rc, node);
2848         } else {
2849                 ret = do_relocation(trans, rc, node, key, path, 1);
2850         }
2851 out:
2852         if (ret || node->level == 0 || node->cowonly) {
2853                 if (release)
2854                         release_metadata_space(rc, node);
2855                 remove_backref_node(&rc->backref_cache, node);
2856         }
2857         return ret;
2858 }
2859
2860 /*
2861  * relocate a list of blocks
2862  */
2863 static noinline_for_stack
2864 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2865                          struct reloc_control *rc, struct rb_root *blocks)
2866 {
2867         struct backref_node *node;
2868         struct btrfs_path *path;
2869         struct tree_block *block;
2870         struct rb_node *rb_node;
2871         int ret;
2872         int err = 0;
2873
2874         path = btrfs_alloc_path();
2875         if (!path) {
2876                 err = -ENOMEM;
2877                 goto out_free_blocks;
2878         }
2879
2880         rb_node = rb_first(blocks);
2881         while (rb_node) {
2882                 block = rb_entry(rb_node, struct tree_block, rb_node);
2883                 if (!block->key_ready)
2884                         reada_tree_block(rc, block);
2885                 rb_node = rb_next(rb_node);
2886         }
2887
2888         rb_node = rb_first(blocks);
2889         while (rb_node) {
2890                 block = rb_entry(rb_node, struct tree_block, rb_node);
2891                 if (!block->key_ready) {
2892                         err = get_tree_block_key(rc, block);
2893                         if (err)
2894                                 goto out_free_path;
2895                 }
2896                 rb_node = rb_next(rb_node);
2897         }
2898
2899         rb_node = rb_first(blocks);
2900         while (rb_node) {
2901                 block = rb_entry(rb_node, struct tree_block, rb_node);
2902
2903                 node = build_backref_tree(rc, &block->key,
2904                                           block->level, block->bytenr);
2905                 if (IS_ERR(node)) {
2906                         err = PTR_ERR(node);
2907                         goto out;
2908                 }
2909
2910                 ret = relocate_tree_block(trans, rc, node, &block->key,
2911                                           path);
2912                 if (ret < 0) {
2913                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2914                                 err = ret;
2915                         goto out;
2916                 }
2917                 rb_node = rb_next(rb_node);
2918         }
2919 out:
2920         err = finish_pending_nodes(trans, rc, path, err);
2921
2922 out_free_path:
2923         btrfs_free_path(path);
2924 out_free_blocks:
2925         free_block_list(blocks);
2926         return err;
2927 }
2928
2929 static noinline_for_stack
2930 int prealloc_file_extent_cluster(struct inode *inode,
2931                                  struct file_extent_cluster *cluster)
2932 {
2933         u64 alloc_hint = 0;
2934         u64 start;
2935         u64 end;
2936         u64 offset = BTRFS_I(inode)->index_cnt;
2937         u64 num_bytes;
2938         int nr = 0;
2939         int ret = 0;
2940
2941         BUG_ON(cluster->start != cluster->boundary[0]);
2942         mutex_lock(&inode->i_mutex);
2943
2944         ret = btrfs_check_data_free_space(inode, cluster->end +
2945                                           1 - cluster->start);
2946         if (ret)
2947                 goto out;
2948
2949         while (nr < cluster->nr) {
2950                 start = cluster->boundary[nr] - offset;
2951                 if (nr + 1 < cluster->nr)
2952                         end = cluster->boundary[nr + 1] - 1 - offset;
2953                 else
2954                         end = cluster->end - offset;
2955
2956                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2957                 num_bytes = end + 1 - start;
2958                 ret = btrfs_prealloc_file_range(inode, 0, start,
2959                                                 num_bytes, num_bytes,
2960                                                 end + 1, &alloc_hint);
2961                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2962                 if (ret)
2963                         break;
2964                 nr++;
2965         }
2966         btrfs_free_reserved_data_space(inode, cluster->end +
2967                                        1 - cluster->start);
2968 out:
2969         mutex_unlock(&inode->i_mutex);
2970         return ret;
2971 }
2972
2973 static noinline_for_stack
2974 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2975                          u64 block_start)
2976 {
2977         struct btrfs_root *root = BTRFS_I(inode)->root;
2978         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2979         struct extent_map *em;
2980         int ret = 0;
2981
2982         em = alloc_extent_map();
2983         if (!em)
2984                 return -ENOMEM;
2985
2986         em->start = start;
2987         em->len = end + 1 - start;
2988         em->block_len = em->len;
2989         em->block_start = block_start;
2990         em->bdev = root->fs_info->fs_devices->latest_bdev;
2991         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2992
2993         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2994         while (1) {
2995                 write_lock(&em_tree->lock);
2996                 ret = add_extent_mapping(em_tree, em, 0);
2997                 write_unlock(&em_tree->lock);
2998                 if (ret != -EEXIST) {
2999                         free_extent_map(em);
3000                         break;
3001                 }
3002                 btrfs_drop_extent_cache(inode, start, end, 0);
3003         }
3004         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3005         return ret;
3006 }
3007
3008 static int relocate_file_extent_cluster(struct inode *inode,
3009                                         struct file_extent_cluster *cluster)
3010 {
3011         u64 page_start;
3012         u64 page_end;
3013         u64 offset = BTRFS_I(inode)->index_cnt;
3014         unsigned long index;
3015         unsigned long last_index;
3016         struct page *page;
3017         struct file_ra_state *ra;
3018         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3019         int nr = 0;
3020         int ret = 0;
3021
3022         if (!cluster->nr)
3023                 return 0;
3024
3025         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3026         if (!ra)
3027                 return -ENOMEM;
3028
3029         ret = prealloc_file_extent_cluster(inode, cluster);
3030         if (ret)
3031                 goto out;
3032
3033         file_ra_state_init(ra, inode->i_mapping);
3034
3035         ret = setup_extent_mapping(inode, cluster->start - offset,
3036                                    cluster->end - offset, cluster->start);
3037         if (ret)
3038                 goto out;
3039
3040         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3041         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3042         while (index <= last_index) {
3043                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3044                 if (ret)
3045                         goto out;
3046
3047                 page = find_lock_page(inode->i_mapping, index);
3048                 if (!page) {
3049                         page_cache_sync_readahead(inode->i_mapping,
3050                                                   ra, NULL, index,
3051                                                   last_index + 1 - index);
3052                         page = find_or_create_page(inode->i_mapping, index,
3053                                                    mask);
3054                         if (!page) {
3055                                 btrfs_delalloc_release_metadata(inode,
3056                                                         PAGE_CACHE_SIZE);
3057                                 ret = -ENOMEM;
3058                                 goto out;
3059                         }
3060                 }
3061
3062                 if (PageReadahead(page)) {
3063                         page_cache_async_readahead(inode->i_mapping,
3064                                                    ra, NULL, page, index,
3065                                                    last_index + 1 - index);
3066                 }
3067
3068                 if (!PageUptodate(page)) {
3069                         btrfs_readpage(NULL, page);
3070                         lock_page(page);
3071                         if (!PageUptodate(page)) {
3072                                 unlock_page(page);
3073                                 page_cache_release(page);
3074                                 btrfs_delalloc_release_metadata(inode,
3075                                                         PAGE_CACHE_SIZE);
3076                                 ret = -EIO;
3077                                 goto out;
3078                         }
3079                 }
3080
3081                 page_start = page_offset(page);
3082                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3083
3084                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3085
3086                 set_page_extent_mapped(page);
3087
3088                 if (nr < cluster->nr &&
3089                     page_start + offset == cluster->boundary[nr]) {
3090                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3091                                         page_start, page_end,
3092                                         EXTENT_BOUNDARY, GFP_NOFS);
3093                         nr++;
3094                 }
3095
3096                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3097                 set_page_dirty(page);
3098
3099                 unlock_extent(&BTRFS_I(inode)->io_tree,
3100                               page_start, page_end);
3101                 unlock_page(page);
3102                 page_cache_release(page);
3103
3104                 index++;
3105                 balance_dirty_pages_ratelimited(inode->i_mapping);
3106                 btrfs_throttle(BTRFS_I(inode)->root);
3107         }
3108         WARN_ON(nr != cluster->nr);
3109 out:
3110         kfree(ra);
3111         return ret;
3112 }
3113
3114 static noinline_for_stack
3115 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3116                          struct file_extent_cluster *cluster)
3117 {
3118         int ret;
3119
3120         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3121                 ret = relocate_file_extent_cluster(inode, cluster);
3122                 if (ret)
3123                         return ret;
3124                 cluster->nr = 0;
3125         }
3126
3127         if (!cluster->nr)
3128                 cluster->start = extent_key->objectid;
3129         else
3130                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3131         cluster->end = extent_key->objectid + extent_key->offset - 1;
3132         cluster->boundary[cluster->nr] = extent_key->objectid;
3133         cluster->nr++;
3134
3135         if (cluster->nr >= MAX_EXTENTS) {
3136                 ret = relocate_file_extent_cluster(inode, cluster);
3137                 if (ret)
3138                         return ret;
3139                 cluster->nr = 0;
3140         }
3141         return 0;
3142 }
3143
3144 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3145 static int get_ref_objectid_v0(struct reloc_control *rc,
3146                                struct btrfs_path *path,
3147                                struct btrfs_key *extent_key,
3148                                u64 *ref_objectid, int *path_change)
3149 {
3150         struct btrfs_key key;
3151         struct extent_buffer *leaf;
3152         struct btrfs_extent_ref_v0 *ref0;
3153         int ret;
3154         int slot;
3155
3156         leaf = path->nodes[0];
3157         slot = path->slots[0];
3158         while (1) {
3159                 if (slot >= btrfs_header_nritems(leaf)) {
3160                         ret = btrfs_next_leaf(rc->extent_root, path);
3161                         if (ret < 0)
3162                                 return ret;
3163                         BUG_ON(ret > 0);
3164                         leaf = path->nodes[0];
3165                         slot = path->slots[0];
3166                         if (path_change)
3167                                 *path_change = 1;
3168                 }
3169                 btrfs_item_key_to_cpu(leaf, &key, slot);
3170                 if (key.objectid != extent_key->objectid)
3171                         return -ENOENT;
3172
3173                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3174                         slot++;
3175                         continue;
3176                 }
3177                 ref0 = btrfs_item_ptr(leaf, slot,
3178                                 struct btrfs_extent_ref_v0);
3179                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3180                 break;
3181         }
3182         return 0;
3183 }
3184 #endif
3185
3186 /*
3187  * helper to add a tree block to the list.
3188  * the major work is getting the generation and level of the block
3189  */
3190 static int add_tree_block(struct reloc_control *rc,
3191                           struct btrfs_key *extent_key,
3192                           struct btrfs_path *path,
3193                           struct rb_root *blocks)
3194 {
3195         struct extent_buffer *eb;
3196         struct btrfs_extent_item *ei;
3197         struct btrfs_tree_block_info *bi;
3198         struct tree_block *block;
3199         struct rb_node *rb_node;
3200         u32 item_size;
3201         int level = -1;
3202         int generation;
3203
3204         eb =  path->nodes[0];
3205         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3206
3207         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3208             item_size >= sizeof(*ei) + sizeof(*bi)) {
3209                 ei = btrfs_item_ptr(eb, path->slots[0],
3210                                 struct btrfs_extent_item);
3211                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3212                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3213                         level = btrfs_tree_block_level(eb, bi);
3214                 } else {
3215                         level = (int)extent_key->offset;
3216                 }
3217                 generation = btrfs_extent_generation(eb, ei);
3218         } else {
3219 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3220                 u64 ref_owner;
3221                 int ret;
3222
3223                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3224                 ret = get_ref_objectid_v0(rc, path, extent_key,
3225                                           &ref_owner, NULL);
3226                 if (ret < 0)
3227                         return ret;
3228                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3229                 level = (int)ref_owner;
3230                 /* FIXME: get real generation */
3231                 generation = 0;
3232 #else
3233                 BUG();
3234 #endif
3235         }
3236
3237         btrfs_release_path(path);
3238
3239         BUG_ON(level == -1);
3240
3241         block = kmalloc(sizeof(*block), GFP_NOFS);
3242         if (!block)
3243                 return -ENOMEM;
3244
3245         block->bytenr = extent_key->objectid;
3246         block->key.objectid = rc->extent_root->leafsize;
3247         block->key.offset = generation;
3248         block->level = level;
3249         block->key_ready = 0;
3250
3251         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3252         if (rb_node)
3253                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3254
3255         return 0;
3256 }
3257
3258 /*
3259  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3260  */
3261 static int __add_tree_block(struct reloc_control *rc,
3262                             u64 bytenr, u32 blocksize,
3263                             struct rb_root *blocks)
3264 {
3265         struct btrfs_path *path;
3266         struct btrfs_key key;
3267         int ret;
3268
3269         if (tree_block_processed(bytenr, blocksize, rc))
3270                 return 0;
3271
3272         if (tree_search(blocks, bytenr))
3273                 return 0;
3274
3275         path = btrfs_alloc_path();
3276         if (!path)
3277                 return -ENOMEM;
3278
3279         key.objectid = bytenr;
3280         key.type = BTRFS_EXTENT_ITEM_KEY;
3281         key.offset = blocksize;
3282
3283         path->search_commit_root = 1;
3284         path->skip_locking = 1;
3285         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3286         if (ret < 0)
3287                 goto out;
3288
3289         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3290         if (ret > 0) {
3291                 if (key.objectid == bytenr &&
3292                     key.type == BTRFS_METADATA_ITEM_KEY)
3293                         ret = 0;
3294         }
3295         BUG_ON(ret);
3296
3297         ret = add_tree_block(rc, &key, path, blocks);
3298 out:
3299         btrfs_free_path(path);
3300         return ret;
3301 }
3302
3303 /*
3304  * helper to check if the block use full backrefs for pointers in it
3305  */
3306 static int block_use_full_backref(struct reloc_control *rc,
3307                                   struct extent_buffer *eb)
3308 {
3309         u64 flags;
3310         int ret;
3311
3312         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3313             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3314                 return 1;
3315
3316         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3317                                        eb->start, btrfs_header_level(eb), 1,
3318                                        NULL, &flags);
3319         BUG_ON(ret);
3320
3321         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3322                 ret = 1;
3323         else
3324                 ret = 0;
3325         return ret;
3326 }
3327
3328 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3329                                     struct inode *inode, u64 ino)
3330 {
3331         struct btrfs_key key;
3332         struct btrfs_path *path;
3333         struct btrfs_root *root = fs_info->tree_root;
3334         struct btrfs_trans_handle *trans;
3335         int ret = 0;
3336
3337         if (inode)
3338                 goto truncate;
3339
3340         key.objectid = ino;
3341         key.type = BTRFS_INODE_ITEM_KEY;
3342         key.offset = 0;
3343
3344         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3345         if (IS_ERR(inode) || is_bad_inode(inode)) {
3346                 if (!IS_ERR(inode))
3347                         iput(inode);
3348                 return -ENOENT;
3349         }
3350
3351 truncate:
3352         ret = btrfs_check_trunc_cache_free_space(root,
3353                                                  &fs_info->global_block_rsv);
3354         if (ret)
3355                 goto out;
3356
3357         path = btrfs_alloc_path();
3358         if (!path) {
3359                 ret = -ENOMEM;
3360                 goto out;
3361         }
3362
3363         trans = btrfs_join_transaction(root);
3364         if (IS_ERR(trans)) {
3365                 btrfs_free_path(path);
3366                 ret = PTR_ERR(trans);
3367                 goto out;
3368         }
3369
3370         ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3371
3372         btrfs_free_path(path);
3373         btrfs_end_transaction(trans, root);
3374         btrfs_btree_balance_dirty(root);
3375 out:
3376         iput(inode);
3377         return ret;
3378 }
3379
3380 /*
3381  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3382  * this function scans fs tree to find blocks reference the data extent
3383  */
3384 static int find_data_references(struct reloc_control *rc,
3385                                 struct btrfs_key *extent_key,
3386                                 struct extent_buffer *leaf,
3387                                 struct btrfs_extent_data_ref *ref,
3388                                 struct rb_root *blocks)
3389 {
3390         struct btrfs_path *path;
3391         struct tree_block *block;
3392         struct btrfs_root *root;
3393         struct btrfs_file_extent_item *fi;
3394         struct rb_node *rb_node;
3395         struct btrfs_key key;
3396         u64 ref_root;
3397         u64 ref_objectid;
3398         u64 ref_offset;
3399         u32 ref_count;
3400         u32 nritems;
3401         int err = 0;
3402         int added = 0;
3403         int counted;
3404         int ret;
3405
3406         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3407         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3408         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3409         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3410
3411         /*
3412          * This is an extent belonging to the free space cache, lets just delete
3413          * it and redo the search.
3414          */
3415         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3416                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3417                                                NULL, ref_objectid);
3418                 if (ret != -ENOENT)
3419                         return ret;
3420                 ret = 0;
3421         }
3422
3423         path = btrfs_alloc_path();
3424         if (!path)
3425                 return -ENOMEM;
3426         path->reada = 1;
3427
3428         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3429         if (IS_ERR(root)) {
3430                 err = PTR_ERR(root);
3431                 goto out;
3432         }
3433
3434         key.objectid = ref_objectid;
3435         key.type = BTRFS_EXTENT_DATA_KEY;
3436         if (ref_offset > ((u64)-1 << 32))
3437                 key.offset = 0;
3438         else
3439                 key.offset = ref_offset;
3440
3441         path->search_commit_root = 1;
3442         path->skip_locking = 1;
3443         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3444         if (ret < 0) {
3445                 err = ret;
3446                 goto out;
3447         }
3448
3449         leaf = path->nodes[0];
3450         nritems = btrfs_header_nritems(leaf);
3451         /*
3452          * the references in tree blocks that use full backrefs
3453          * are not counted in
3454          */
3455         if (block_use_full_backref(rc, leaf))
3456                 counted = 0;
3457         else
3458                 counted = 1;
3459         rb_node = tree_search(blocks, leaf->start);
3460         if (rb_node) {
3461                 if (counted)
3462                         added = 1;
3463                 else
3464                         path->slots[0] = nritems;
3465         }
3466
3467         while (ref_count > 0) {
3468                 while (path->slots[0] >= nritems) {
3469                         ret = btrfs_next_leaf(root, path);
3470                         if (ret < 0) {
3471                                 err = ret;
3472                                 goto out;
3473                         }
3474                         if (ret > 0) {
3475                                 WARN_ON(1);
3476                                 goto out;
3477                         }
3478
3479                         leaf = path->nodes[0];
3480                         nritems = btrfs_header_nritems(leaf);
3481                         added = 0;
3482
3483                         if (block_use_full_backref(rc, leaf))
3484                                 counted = 0;
3485                         else
3486                                 counted = 1;
3487                         rb_node = tree_search(blocks, leaf->start);
3488                         if (rb_node) {
3489                                 if (counted)
3490                                         added = 1;
3491                                 else
3492                                         path->slots[0] = nritems;
3493                         }
3494                 }
3495
3496                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3497                 if (key.objectid != ref_objectid ||
3498                     key.type != BTRFS_EXTENT_DATA_KEY) {
3499                         WARN_ON(1);
3500                         break;
3501                 }
3502
3503                 fi = btrfs_item_ptr(leaf, path->slots[0],
3504                                     struct btrfs_file_extent_item);
3505
3506                 if (btrfs_file_extent_type(leaf, fi) ==
3507                     BTRFS_FILE_EXTENT_INLINE)
3508                         goto next;
3509
3510                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3511                     extent_key->objectid)
3512                         goto next;
3513
3514                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3515                 if (key.offset != ref_offset)
3516                         goto next;
3517
3518                 if (counted)
3519                         ref_count--;
3520                 if (added)
3521                         goto next;
3522
3523                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3524                         block = kmalloc(sizeof(*block), GFP_NOFS);
3525                         if (!block) {
3526                                 err = -ENOMEM;
3527                                 break;
3528                         }
3529                         block->bytenr = leaf->start;
3530                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3531                         block->level = 0;
3532                         block->key_ready = 1;
3533                         rb_node = tree_insert(blocks, block->bytenr,
3534                                               &block->rb_node);
3535                         if (rb_node)
3536                                 backref_tree_panic(rb_node, -EEXIST,
3537                                                    block->bytenr);
3538                 }
3539                 if (counted)
3540                         added = 1;
3541                 else
3542                         path->slots[0] = nritems;
3543 next:
3544                 path->slots[0]++;
3545
3546         }
3547 out:
3548         btrfs_free_path(path);
3549         return err;
3550 }
3551
3552 /*
3553  * helper to find all tree blocks that reference a given data extent
3554  */
3555 static noinline_for_stack
3556 int add_data_references(struct reloc_control *rc,
3557                         struct btrfs_key *extent_key,
3558                         struct btrfs_path *path,
3559                         struct rb_root *blocks)
3560 {
3561         struct btrfs_key key;
3562         struct extent_buffer *eb;
3563         struct btrfs_extent_data_ref *dref;
3564         struct btrfs_extent_inline_ref *iref;
3565         unsigned long ptr;
3566         unsigned long end;
3567         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3568         int ret;
3569         int err = 0;
3570
3571         eb = path->nodes[0];
3572         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3573         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3574 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3575         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3576                 ptr = end;
3577         else
3578 #endif
3579                 ptr += sizeof(struct btrfs_extent_item);
3580
3581         while (ptr < end) {
3582                 iref = (struct btrfs_extent_inline_ref *)ptr;
3583                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3584                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3585                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3586                         ret = __add_tree_block(rc, key.offset, blocksize,
3587                                                blocks);
3588                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3589                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3590                         ret = find_data_references(rc, extent_key,
3591                                                    eb, dref, blocks);
3592                 } else {
3593                         BUG();
3594                 }
3595                 ptr += btrfs_extent_inline_ref_size(key.type);
3596         }
3597         WARN_ON(ptr > end);
3598
3599         while (1) {
3600                 cond_resched();
3601                 eb = path->nodes[0];
3602                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3603                         ret = btrfs_next_leaf(rc->extent_root, path);
3604                         if (ret < 0) {
3605                                 err = ret;
3606                                 break;
3607                         }
3608                         if (ret > 0)
3609                                 break;
3610                         eb = path->nodes[0];
3611                 }
3612
3613                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3614                 if (key.objectid != extent_key->objectid)
3615                         break;
3616
3617 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3618                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3619                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3620 #else
3621                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3622                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3623 #endif
3624                         ret = __add_tree_block(rc, key.offset, blocksize,
3625                                                blocks);
3626                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3627                         dref = btrfs_item_ptr(eb, path->slots[0],
3628                                               struct btrfs_extent_data_ref);
3629                         ret = find_data_references(rc, extent_key,
3630                                                    eb, dref, blocks);
3631                 } else {
3632                         ret = 0;
3633                 }
3634                 if (ret) {
3635                         err = ret;
3636                         break;
3637                 }
3638                 path->slots[0]++;
3639         }
3640         btrfs_release_path(path);
3641         if (err)
3642                 free_block_list(blocks);
3643         return err;
3644 }
3645
3646 /*
3647  * helper to find next unprocessed extent
3648  */
3649 static noinline_for_stack
3650 int find_next_extent(struct btrfs_trans_handle *trans,
3651                      struct reloc_control *rc, struct btrfs_path *path,
3652                      struct btrfs_key *extent_key)
3653 {
3654         struct btrfs_key key;
3655         struct extent_buffer *leaf;
3656         u64 start, end, last;
3657         int ret;
3658
3659         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3660         while (1) {
3661                 cond_resched();
3662                 if (rc->search_start >= last) {
3663                         ret = 1;
3664                         break;
3665                 }
3666
3667                 key.objectid = rc->search_start;
3668                 key.type = BTRFS_EXTENT_ITEM_KEY;
3669                 key.offset = 0;
3670
3671                 path->search_commit_root = 1;
3672                 path->skip_locking = 1;
3673                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3674                                         0, 0);
3675                 if (ret < 0)
3676                         break;
3677 next:
3678                 leaf = path->nodes[0];
3679                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3680                         ret = btrfs_next_leaf(rc->extent_root, path);
3681                         if (ret != 0)
3682                                 break;
3683                         leaf = path->nodes[0];
3684                 }
3685
3686                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3687                 if (key.objectid >= last) {
3688                         ret = 1;
3689                         break;
3690                 }
3691
3692                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3693                     key.type != BTRFS_METADATA_ITEM_KEY) {
3694                         path->slots[0]++;
3695                         goto next;
3696                 }
3697
3698                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3699                     key.objectid + key.offset <= rc->search_start) {
3700                         path->slots[0]++;
3701                         goto next;
3702                 }
3703
3704                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3705                     key.objectid + rc->extent_root->leafsize <=
3706                     rc->search_start) {
3707                         path->slots[0]++;
3708                         goto next;
3709                 }
3710
3711                 ret = find_first_extent_bit(&rc->processed_blocks,
3712                                             key.objectid, &start, &end,
3713                                             EXTENT_DIRTY, NULL);
3714
3715                 if (ret == 0 && start <= key.objectid) {
3716                         btrfs_release_path(path);
3717                         rc->search_start = end + 1;
3718                 } else {
3719                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3720                                 rc->search_start = key.objectid + key.offset;
3721                         else
3722                                 rc->search_start = key.objectid +
3723                                         rc->extent_root->leafsize;
3724                         memcpy(extent_key, &key, sizeof(key));
3725                         return 0;
3726                 }
3727         }
3728         btrfs_release_path(path);
3729         return ret;
3730 }
3731
3732 static void set_reloc_control(struct reloc_control *rc)
3733 {
3734         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3735
3736         mutex_lock(&fs_info->reloc_mutex);
3737         fs_info->reloc_ctl = rc;
3738         mutex_unlock(&fs_info->reloc_mutex);
3739 }
3740
3741 static void unset_reloc_control(struct reloc_control *rc)
3742 {
3743         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3744
3745         mutex_lock(&fs_info->reloc_mutex);
3746         fs_info->reloc_ctl = NULL;
3747         mutex_unlock(&fs_info->reloc_mutex);
3748 }
3749
3750 static int check_extent_flags(u64 flags)
3751 {
3752         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3753             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3754                 return 1;
3755         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3756             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3757                 return 1;
3758         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3759             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3760                 return 1;
3761         return 0;
3762 }
3763
3764 static noinline_for_stack
3765 int prepare_to_relocate(struct reloc_control *rc)
3766 {
3767         struct btrfs_trans_handle *trans;
3768         int ret;
3769
3770         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3771                                               BTRFS_BLOCK_RSV_TEMP);
3772         if (!rc->block_rsv)
3773                 return -ENOMEM;
3774
3775         /*
3776          * reserve some space for creating reloc trees.
3777          * btrfs_init_reloc_root will use them when there
3778          * is no reservation in transaction handle.
3779          */
3780         ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3781                                   rc->extent_root->nodesize * 256,
3782                                   BTRFS_RESERVE_FLUSH_ALL);
3783         if (ret)
3784                 return ret;
3785
3786         memset(&rc->cluster, 0, sizeof(rc->cluster));
3787         rc->search_start = rc->block_group->key.objectid;
3788         rc->extents_found = 0;
3789         rc->nodes_relocated = 0;
3790         rc->merging_rsv_size = 0;
3791
3792         rc->create_reloc_tree = 1;
3793         set_reloc_control(rc);
3794
3795         trans = btrfs_join_transaction(rc->extent_root);
3796         if (IS_ERR(trans)) {
3797                 unset_reloc_control(rc);
3798                 /*
3799                  * extent tree is not a ref_cow tree and has no reloc_root to
3800                  * cleanup.  And callers are responsible to free the above
3801                  * block rsv.
3802                  */
3803                 return PTR_ERR(trans);
3804         }
3805         btrfs_commit_transaction(trans, rc->extent_root);
3806         return 0;
3807 }
3808
3809 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3810 {
3811         struct rb_root blocks = RB_ROOT;
3812         struct btrfs_key key;
3813         struct btrfs_trans_handle *trans = NULL;
3814         struct btrfs_path *path;
3815         struct btrfs_extent_item *ei;
3816         u64 flags;
3817         u32 item_size;
3818         int ret;
3819         int err = 0;
3820         int progress = 0;
3821
3822         path = btrfs_alloc_path();
3823         if (!path)
3824                 return -ENOMEM;
3825         path->reada = 1;
3826
3827         ret = prepare_to_relocate(rc);
3828         if (ret) {
3829                 err = ret;
3830                 goto out_free;
3831         }
3832
3833         while (1) {
3834                 progress++;
3835                 trans = btrfs_start_transaction(rc->extent_root, 0);
3836                 if (IS_ERR(trans)) {
3837                         err = PTR_ERR(trans);
3838                         trans = NULL;
3839                         break;
3840                 }
3841 restart:
3842                 if (update_backref_cache(trans, &rc->backref_cache)) {
3843                         btrfs_end_transaction(trans, rc->extent_root);
3844                         continue;
3845                 }
3846
3847                 ret = find_next_extent(trans, rc, path, &key);
3848                 if (ret < 0)
3849                         err = ret;
3850                 if (ret != 0)
3851                         break;
3852
3853                 rc->extents_found++;
3854
3855                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3856                                     struct btrfs_extent_item);
3857                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3858                 if (item_size >= sizeof(*ei)) {
3859                         flags = btrfs_extent_flags(path->nodes[0], ei);
3860                         ret = check_extent_flags(flags);
3861                         BUG_ON(ret);
3862
3863                 } else {
3864 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3865                         u64 ref_owner;
3866                         int path_change = 0;
3867
3868                         BUG_ON(item_size !=
3869                                sizeof(struct btrfs_extent_item_v0));
3870                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3871                                                   &path_change);
3872                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3873                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3874                         else
3875                                 flags = BTRFS_EXTENT_FLAG_DATA;
3876
3877                         if (path_change) {
3878                                 btrfs_release_path(path);
3879
3880                                 path->search_commit_root = 1;
3881                                 path->skip_locking = 1;
3882                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3883                                                         &key, path, 0, 0);
3884                                 if (ret < 0) {
3885                                         err = ret;
3886                                         break;
3887                                 }
3888                                 BUG_ON(ret > 0);
3889                         }
3890 #else
3891                         BUG();
3892 #endif
3893                 }
3894
3895                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3896                         ret = add_tree_block(rc, &key, path, &blocks);
3897                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3898                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3899                         ret = add_data_references(rc, &key, path, &blocks);
3900                 } else {
3901                         btrfs_release_path(path);
3902                         ret = 0;
3903                 }
3904                 if (ret < 0) {
3905                         err = ret;
3906                         break;
3907                 }
3908
3909                 if (!RB_EMPTY_ROOT(&blocks)) {
3910                         ret = relocate_tree_blocks(trans, rc, &blocks);
3911                         if (ret < 0) {
3912                                 if (ret != -EAGAIN) {
3913                                         err = ret;
3914                                         break;
3915                                 }
3916                                 rc->extents_found--;
3917                                 rc->search_start = key.objectid;
3918                         }
3919                 }
3920
3921                 ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5);
3922                 if (ret < 0) {
3923                         if (ret != -ENOSPC) {
3924                                 err = ret;
3925                                 WARN_ON(1);
3926                                 break;
3927                         }
3928                         rc->commit_transaction = 1;
3929                 }
3930
3931                 if (rc->commit_transaction) {
3932                         rc->commit_transaction = 0;
3933                         ret = btrfs_commit_transaction(trans, rc->extent_root);
3934                         BUG_ON(ret);
3935                 } else {
3936                         btrfs_end_transaction_throttle(trans, rc->extent_root);
3937                         btrfs_btree_balance_dirty(rc->extent_root);
3938                 }
3939                 trans = NULL;
3940
3941                 if (rc->stage == MOVE_DATA_EXTENTS &&
3942                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3943                         rc->found_file_extent = 1;
3944                         ret = relocate_data_extent(rc->data_inode,
3945                                                    &key, &rc->cluster);
3946                         if (ret < 0) {
3947                                 err = ret;
3948                                 break;
3949                         }
3950                 }
3951         }
3952         if (trans && progress && err == -ENOSPC) {
3953                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
3954                                               rc->block_group->flags);
3955                 if (ret == 0) {
3956                         err = 0;
3957                         progress = 0;
3958                         goto restart;
3959                 }
3960         }
3961
3962         btrfs_release_path(path);
3963         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3964                           GFP_NOFS);
3965
3966         if (trans) {
3967                 btrfs_end_transaction_throttle(trans, rc->extent_root);
3968                 btrfs_btree_balance_dirty(rc->extent_root);
3969         }
3970
3971         if (!err) {
3972                 ret = relocate_file_extent_cluster(rc->data_inode,
3973                                                    &rc->cluster);
3974                 if (ret < 0)
3975                         err = ret;
3976         }
3977
3978         rc->create_reloc_tree = 0;
3979         set_reloc_control(rc);
3980
3981         backref_cache_cleanup(&rc->backref_cache);
3982         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3983
3984         err = prepare_to_merge(rc, err);
3985
3986         merge_reloc_roots(rc);
3987
3988         rc->merge_reloc_tree = 0;
3989         unset_reloc_control(rc);
3990         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3991
3992         /* get rid of pinned extents */
3993         trans = btrfs_join_transaction(rc->extent_root);
3994         if (IS_ERR(trans))
3995                 err = PTR_ERR(trans);
3996         else
3997                 btrfs_commit_transaction(trans, rc->extent_root);
3998 out_free:
3999         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4000         btrfs_free_path(path);
4001         return err;
4002 }
4003
4004 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4005                                  struct btrfs_root *root, u64 objectid)
4006 {
4007         struct btrfs_path *path;
4008         struct btrfs_inode_item *item;
4009         struct extent_buffer *leaf;
4010         int ret;
4011
4012         path = btrfs_alloc_path();
4013         if (!path)
4014                 return -ENOMEM;
4015
4016         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4017         if (ret)
4018                 goto out;
4019
4020         leaf = path->nodes[0];
4021         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4022         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4023         btrfs_set_inode_generation(leaf, item, 1);
4024         btrfs_set_inode_size(leaf, item, 0);
4025         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4026         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4027                                           BTRFS_INODE_PREALLOC);
4028         btrfs_mark_buffer_dirty(leaf);
4029         btrfs_release_path(path);
4030 out:
4031         btrfs_free_path(path);
4032         return ret;
4033 }
4034
4035 /*
4036  * helper to create inode for data relocation.
4037  * the inode is in data relocation tree and its link count is 0
4038  */
4039 static noinline_for_stack
4040 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4041                                  struct btrfs_block_group_cache *group)
4042 {
4043         struct inode *inode = NULL;
4044         struct btrfs_trans_handle *trans;
4045         struct btrfs_root *root;
4046         struct btrfs_key key;
4047         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4048         int err = 0;
4049
4050         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4051         if (IS_ERR(root))
4052                 return ERR_CAST(root);
4053
4054         trans = btrfs_start_transaction(root, 6);
4055         if (IS_ERR(trans))
4056                 return ERR_CAST(trans);
4057
4058         err = btrfs_find_free_objectid(root, &objectid);
4059         if (err)
4060                 goto out;
4061
4062         err = __insert_orphan_inode(trans, root, objectid);
4063         BUG_ON(err);
4064
4065         key.objectid = objectid;
4066         key.type = BTRFS_INODE_ITEM_KEY;
4067         key.offset = 0;
4068         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4069         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4070         BTRFS_I(inode)->index_cnt = group->key.objectid;
4071
4072         err = btrfs_orphan_add(trans, inode);
4073 out:
4074         btrfs_end_transaction(trans, root);
4075         btrfs_btree_balance_dirty(root);
4076         if (err) {
4077                 if (inode)
4078                         iput(inode);
4079                 inode = ERR_PTR(err);
4080         }
4081         return inode;
4082 }
4083
4084 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4085 {
4086         struct reloc_control *rc;
4087
4088         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4089         if (!rc)
4090                 return NULL;
4091
4092         INIT_LIST_HEAD(&rc->reloc_roots);
4093         backref_cache_init(&rc->backref_cache);
4094         mapping_tree_init(&rc->reloc_root_tree);
4095         extent_io_tree_init(&rc->processed_blocks,
4096                             fs_info->btree_inode->i_mapping);
4097         return rc;
4098 }
4099
4100 /*
4101  * function to relocate all extents in a block group.
4102  */
4103 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4104 {
4105         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4106         struct reloc_control *rc;
4107         struct inode *inode;
4108         struct btrfs_path *path;
4109         int ret;
4110         int rw = 0;
4111         int err = 0;
4112
4113         rc = alloc_reloc_control(fs_info);
4114         if (!rc)
4115                 return -ENOMEM;
4116
4117         rc->extent_root = extent_root;
4118
4119         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4120         BUG_ON(!rc->block_group);
4121
4122         if (!rc->block_group->ro) {
4123                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4124                 if (ret) {
4125                         err = ret;
4126                         goto out;
4127                 }
4128                 rw = 1;
4129         }
4130
4131         path = btrfs_alloc_path();
4132         if (!path) {
4133                 err = -ENOMEM;
4134                 goto out;
4135         }
4136
4137         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4138                                         path);
4139         btrfs_free_path(path);
4140
4141         if (!IS_ERR(inode))
4142                 ret = delete_block_group_cache(fs_info, inode, 0);
4143         else
4144                 ret = PTR_ERR(inode);
4145
4146         if (ret && ret != -ENOENT) {
4147                 err = ret;
4148                 goto out;
4149         }
4150
4151         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4152         if (IS_ERR(rc->data_inode)) {
4153                 err = PTR_ERR(rc->data_inode);
4154                 rc->data_inode = NULL;
4155                 goto out;
4156         }
4157
4158         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4159                (unsigned long long)rc->block_group->key.objectid,
4160                (unsigned long long)rc->block_group->flags);
4161
4162         ret = btrfs_start_all_delalloc_inodes(fs_info, 0);
4163         if (ret < 0) {
4164                 err = ret;
4165                 goto out;
4166         }
4167         btrfs_wait_ordered_extents(fs_info->tree_root, 0);
4168
4169         while (1) {
4170                 mutex_lock(&fs_info->cleaner_mutex);
4171                 ret = relocate_block_group(rc);
4172                 mutex_unlock(&fs_info->cleaner_mutex);
4173                 if (ret < 0) {
4174                         err = ret;
4175                         goto out;
4176                 }
4177
4178                 if (rc->extents_found == 0)
4179                         break;
4180
4181                 printk(KERN_INFO "btrfs: found %llu extents\n",
4182                         (unsigned long long)rc->extents_found);
4183
4184                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4185                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4186                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4187                                                  0, -1);
4188                         rc->stage = UPDATE_DATA_PTRS;
4189                 }
4190         }
4191
4192         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4193                                      rc->block_group->key.objectid,
4194                                      rc->block_group->key.objectid +
4195                                      rc->block_group->key.offset - 1);
4196
4197         WARN_ON(rc->block_group->pinned > 0);
4198         WARN_ON(rc->block_group->reserved > 0);
4199         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4200 out:
4201         if (err && rw)
4202                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4203         iput(rc->data_inode);
4204         btrfs_put_block_group(rc->block_group);
4205         kfree(rc);
4206         return err;
4207 }
4208
4209 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4210 {
4211         struct btrfs_trans_handle *trans;
4212         int ret, err;
4213
4214         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4215         if (IS_ERR(trans))
4216                 return PTR_ERR(trans);
4217
4218         memset(&root->root_item.drop_progress, 0,
4219                 sizeof(root->root_item.drop_progress));
4220         root->root_item.drop_level = 0;
4221         btrfs_set_root_refs(&root->root_item, 0);
4222         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4223                                 &root->root_key, &root->root_item);
4224
4225         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4226         if (err)
4227                 return err;
4228         return ret;
4229 }
4230
4231 /*
4232  * recover relocation interrupted by system crash.
4233  *
4234  * this function resumes merging reloc trees with corresponding fs trees.
4235  * this is important for keeping the sharing of tree blocks
4236  */
4237 int btrfs_recover_relocation(struct btrfs_root *root)
4238 {
4239         LIST_HEAD(reloc_roots);
4240         struct btrfs_key key;
4241         struct btrfs_root *fs_root;
4242         struct btrfs_root *reloc_root;
4243         struct btrfs_path *path;
4244         struct extent_buffer *leaf;
4245         struct reloc_control *rc = NULL;
4246         struct btrfs_trans_handle *trans;
4247         int ret;
4248         int err = 0;
4249
4250         path = btrfs_alloc_path();
4251         if (!path)
4252                 return -ENOMEM;
4253         path->reada = -1;
4254
4255         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4256         key.type = BTRFS_ROOT_ITEM_KEY;
4257         key.offset = (u64)-1;
4258
4259         while (1) {
4260                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4261                                         path, 0, 0);
4262                 if (ret < 0) {
4263                         err = ret;
4264                         goto out;
4265                 }
4266                 if (ret > 0) {
4267                         if (path->slots[0] == 0)
4268                                 break;
4269                         path->slots[0]--;
4270                 }
4271                 leaf = path->nodes[0];
4272                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4273                 btrfs_release_path(path);
4274
4275                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4276                     key.type != BTRFS_ROOT_ITEM_KEY)
4277                         break;
4278
4279                 reloc_root = btrfs_read_fs_root(root, &key);
4280                 if (IS_ERR(reloc_root)) {
4281                         err = PTR_ERR(reloc_root);
4282                         goto out;
4283                 }
4284
4285                 list_add(&reloc_root->root_list, &reloc_roots);
4286
4287                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4288                         fs_root = read_fs_root(root->fs_info,
4289                                                reloc_root->root_key.offset);
4290                         if (IS_ERR(fs_root)) {
4291                                 ret = PTR_ERR(fs_root);
4292                                 if (ret != -ENOENT) {
4293                                         err = ret;
4294                                         goto out;
4295                                 }
4296                                 ret = mark_garbage_root(reloc_root);
4297                                 if (ret < 0) {
4298                                         err = ret;
4299                                         goto out;
4300                                 }
4301                         }
4302                 }
4303
4304                 if (key.offset == 0)
4305                         break;
4306
4307                 key.offset--;
4308         }
4309         btrfs_release_path(path);
4310
4311         if (list_empty(&reloc_roots))
4312                 goto out;
4313
4314         rc = alloc_reloc_control(root->fs_info);
4315         if (!rc) {
4316                 err = -ENOMEM;
4317                 goto out;
4318         }
4319
4320         rc->extent_root = root->fs_info->extent_root;
4321
4322         set_reloc_control(rc);
4323
4324         trans = btrfs_join_transaction(rc->extent_root);
4325         if (IS_ERR(trans)) {
4326                 unset_reloc_control(rc);
4327                 err = PTR_ERR(trans);
4328                 goto out_free;
4329         }
4330
4331         rc->merge_reloc_tree = 1;
4332
4333         while (!list_empty(&reloc_roots)) {
4334                 reloc_root = list_entry(reloc_roots.next,
4335                                         struct btrfs_root, root_list);
4336                 list_del(&reloc_root->root_list);
4337
4338                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4339                         list_add_tail(&reloc_root->root_list,
4340                                       &rc->reloc_roots);
4341                         continue;
4342                 }
4343
4344                 fs_root = read_fs_root(root->fs_info,
4345                                        reloc_root->root_key.offset);
4346                 if (IS_ERR(fs_root)) {
4347                         err = PTR_ERR(fs_root);
4348                         goto out_free;
4349                 }
4350
4351                 err = __add_reloc_root(reloc_root);
4352                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4353                 fs_root->reloc_root = reloc_root;
4354         }
4355
4356         err = btrfs_commit_transaction(trans, rc->extent_root);
4357         if (err)
4358                 goto out_free;
4359
4360         merge_reloc_roots(rc);
4361
4362         unset_reloc_control(rc);
4363
4364         trans = btrfs_join_transaction(rc->extent_root);
4365         if (IS_ERR(trans))
4366                 err = PTR_ERR(trans);
4367         else
4368                 err = btrfs_commit_transaction(trans, rc->extent_root);
4369 out_free:
4370         kfree(rc);
4371 out:
4372         if (!list_empty(&reloc_roots))
4373                 free_reloc_roots(&reloc_roots);
4374
4375         btrfs_free_path(path);
4376
4377         if (err == 0) {
4378                 /* cleanup orphan inode in data relocation tree */
4379                 fs_root = read_fs_root(root->fs_info,
4380                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4381                 if (IS_ERR(fs_root))
4382                         err = PTR_ERR(fs_root);
4383                 else
4384                         err = btrfs_orphan_cleanup(fs_root);
4385         }
4386         return err;
4387 }
4388
4389 /*
4390  * helper to add ordered checksum for data relocation.
4391  *
4392  * cloning checksum properly handles the nodatasum extents.
4393  * it also saves CPU time to re-calculate the checksum.
4394  */
4395 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4396 {
4397         struct btrfs_ordered_sum *sums;
4398         struct btrfs_sector_sum *sector_sum;
4399         struct btrfs_ordered_extent *ordered;
4400         struct btrfs_root *root = BTRFS_I(inode)->root;
4401         size_t offset;
4402         int ret;
4403         u64 disk_bytenr;
4404         LIST_HEAD(list);
4405
4406         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4407         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4408
4409         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4410         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4411                                        disk_bytenr + len - 1, &list, 0);
4412         if (ret)
4413                 goto out;
4414
4415         while (!list_empty(&list)) {
4416                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4417                 list_del_init(&sums->list);
4418
4419                 sector_sum = sums->sums;
4420                 sums->bytenr = ordered->start;
4421
4422                 offset = 0;
4423                 while (offset < sums->len) {
4424                         sector_sum->bytenr += ordered->start - disk_bytenr;
4425                         sector_sum++;
4426                         offset += root->sectorsize;
4427                 }
4428
4429                 btrfs_add_ordered_sum(inode, ordered, sums);
4430         }
4431 out:
4432         btrfs_put_ordered_extent(ordered);
4433         return ret;
4434 }
4435
4436 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4437                            struct btrfs_root *root, struct extent_buffer *buf,
4438                            struct extent_buffer *cow)
4439 {
4440         struct reloc_control *rc;
4441         struct backref_node *node;
4442         int first_cow = 0;
4443         int level;
4444         int ret;
4445
4446         rc = root->fs_info->reloc_ctl;
4447         if (!rc)
4448                 return;
4449
4450         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4451                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4452
4453         level = btrfs_header_level(buf);
4454         if (btrfs_header_generation(buf) <=
4455             btrfs_root_last_snapshot(&root->root_item))
4456                 first_cow = 1;
4457
4458         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4459             rc->create_reloc_tree) {
4460                 WARN_ON(!first_cow && level == 0);
4461
4462                 node = rc->backref_cache.path[level];
4463                 BUG_ON(node->bytenr != buf->start &&
4464                        node->new_bytenr != buf->start);
4465
4466                 drop_node_buffer(node);
4467                 extent_buffer_get(cow);
4468                 node->eb = cow;
4469                 node->new_bytenr = cow->start;
4470
4471                 if (!node->pending) {
4472                         list_move_tail(&node->list,
4473                                        &rc->backref_cache.pending[level]);
4474                         node->pending = 1;
4475                 }
4476
4477                 if (first_cow)
4478                         __mark_block_processed(rc, node);
4479
4480                 if (first_cow && level > 0)
4481                         rc->nodes_relocated += buf->len;
4482         }
4483
4484         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4485                 ret = replace_file_extents(trans, rc, root, cow);
4486                 BUG_ON(ret);
4487         }
4488 }
4489
4490 /*
4491  * called before creating snapshot. it calculates metadata reservation
4492  * requried for relocating tree blocks in the snapshot
4493  */
4494 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4495                               struct btrfs_pending_snapshot *pending,
4496                               u64 *bytes_to_reserve)
4497 {
4498         struct btrfs_root *root;
4499         struct reloc_control *rc;
4500
4501         root = pending->root;
4502         if (!root->reloc_root)
4503                 return;
4504
4505         rc = root->fs_info->reloc_ctl;
4506         if (!rc->merge_reloc_tree)
4507                 return;
4508
4509         root = root->reloc_root;
4510         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4511         /*
4512          * relocation is in the stage of merging trees. the space
4513          * used by merging a reloc tree is twice the size of
4514          * relocated tree nodes in the worst case. half for cowing
4515          * the reloc tree, half for cowing the fs tree. the space
4516          * used by cowing the reloc tree will be freed after the
4517          * tree is dropped. if we create snapshot, cowing the fs
4518          * tree may use more space than it frees. so we need
4519          * reserve extra space.
4520          */
4521         *bytes_to_reserve += rc->nodes_relocated;
4522 }
4523
4524 /*
4525  * called after snapshot is created. migrate block reservation
4526  * and create reloc root for the newly created snapshot
4527  */
4528 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4529                                struct btrfs_pending_snapshot *pending)
4530 {
4531         struct btrfs_root *root = pending->root;
4532         struct btrfs_root *reloc_root;
4533         struct btrfs_root *new_root;
4534         struct reloc_control *rc;
4535         int ret;
4536
4537         if (!root->reloc_root)
4538                 return 0;
4539
4540         rc = root->fs_info->reloc_ctl;
4541         rc->merging_rsv_size += rc->nodes_relocated;
4542
4543         if (rc->merge_reloc_tree) {
4544                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4545                                               rc->block_rsv,
4546                                               rc->nodes_relocated);
4547                 if (ret)
4548                         return ret;
4549         }
4550
4551         new_root = pending->snap;
4552         reloc_root = create_reloc_root(trans, root->reloc_root,
4553                                        new_root->root_key.objectid);
4554         if (IS_ERR(reloc_root))
4555                 return PTR_ERR(reloc_root);
4556
4557         ret = __add_reloc_root(reloc_root);
4558         BUG_ON(ret < 0);
4559         new_root->reloc_root = reloc_root;
4560
4561         if (rc->create_reloc_tree)
4562                 ret = clone_backref_node(trans, rc, root, reloc_root);
4563         return ret;
4564 }