Pileus Git - ~andy/linux/blob - 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", 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_get_fs_root(fs_info, &key, false);
 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         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 645             item_size <= sizeof(*ei)) {
 646                 WARN_ON(item_size < sizeof(*ei));
 647                 return 1;
 648         }
 649
 650         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 651                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 652                 *ptr = (unsigned long)(bi + 1);
 653         } else {
 654                 *ptr = (unsigned long)(ei + 1);
 655         }
 656         *end = (unsigned long)ei + item_size;
 657         return 0;
 658 }
 659
 660 /*
 661  * build backref tree for a given tree block. root of the backref tree
 662  * corresponds the tree block, leaves of the backref tree correspond
 663  * roots of b-trees that reference the tree block.
 664  *
 665  * the basic idea of this function is check backrefs of a given block
 666  * to find upper level blocks that refernece the block, and then check
 667  * bakcrefs of these upper level blocks recursively. the recursion stop
 668  * when tree root is reached or backrefs for the block is cached.
 669  *
 670  * NOTE: if we find backrefs for a block are cached, we know backrefs
 671  * for all upper level blocks that directly/indirectly reference the
 672  * block are also cached.
 673  */
 674 static noinline_for_stack
 675 struct backref_node *build_backref_tree(struct reloc_control *rc,
 676                                         struct btrfs_key *node_key,
 677                                         int level, u64 bytenr)
 678 {
 679         struct backref_cache *cache = &rc->backref_cache;
 680         struct btrfs_path *path1;
 681         struct btrfs_path *path2;
 682         struct extent_buffer *eb;
 683         struct btrfs_root *root;
 684         struct backref_node *cur;
 685         struct backref_node *upper;
 686         struct backref_node *lower;
 687         struct backref_node *node = NULL;
 688         struct backref_node *exist = NULL;
 689         struct backref_edge *edge;
 690         struct rb_node *rb_node;
 691         struct btrfs_key key;
 692         unsigned long end;
 693         unsigned long ptr;
 694         LIST_HEAD(list);
 695         LIST_HEAD(useless);
 696         int cowonly;
 697         int ret;
 698         int err = 0;
 699         bool need_check = true;
 700
 701         path1 = btrfs_alloc_path();
 702         path2 = btrfs_alloc_path();
 703         if (!path1 || !path2) {
 704                 err = -ENOMEM;
 705                 goto out;
 706         }
 707         path1->reada = 1;
 708         path2->reada = 2;
 709
 710         node = alloc_backref_node(cache);
 711         if (!node) {
 712                 err = -ENOMEM;
 713                 goto out;
 714         }
 715
 716         node->bytenr = bytenr;
 717         node->level = level;
 718         node->lowest = 1;
 719         cur = node;
 720 again:
 721         end = 0;
 722         ptr = 0;
 723         key.objectid = cur->bytenr;
 724         key.type = BTRFS_METADATA_ITEM_KEY;
 725         key.offset = (u64)-1;
 726
 727         path1->search_commit_root = 1;
 728         path1->skip_locking = 1;
 729         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 730                                 0, 0);
 731         if (ret < 0) {
 732                 err = ret;
 733                 goto out;
 734         }
 735         BUG_ON(!ret || !path1->slots[0]);
 736
 737         path1->slots[0]--;
 738
 739         WARN_ON(cur->checked);
 740         if (!list_empty(&cur->upper)) {
 741                 /*
 742                  * the backref was added previously when processing
 743                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 744                  */
 745                 BUG_ON(!list_is_singular(&cur->upper));
 746                 edge = list_entry(cur->upper.next, struct backref_edge,
 747                                   list[LOWER]);
 748                 BUG_ON(!list_empty(&edge->list[UPPER]));
 749                 exist = edge->node[UPPER];
 750                 /*
 751                  * add the upper level block to pending list if we need
 752                  * check its backrefs
 753                  */
 754                 if (!exist->checked)
 755                         list_add_tail(&edge->list[UPPER], &list);
 756         } else {
 757                 exist = NULL;
 758         }
 759
 760         while (1) {
 761                 cond_resched();
 762                 eb = path1->nodes[0];
 763
 764                 if (ptr >= end) {
 765                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 766                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 767                                 if (ret < 0) {
 768                                         err = ret;
 769                                         goto out;
 770                                 }
 771                                 if (ret > 0)
 772                                         break;
 773                                 eb = path1->nodes[0];
 774                         }
 775
 776                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 777                         if (key.objectid != cur->bytenr) {
 778                                 WARN_ON(exist);
 779                                 break;
 780                         }
 781
 782                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 783                             key.type == BTRFS_METADATA_ITEM_KEY) {
 784                                 ret = find_inline_backref(eb, path1->slots[0],
 785                                                           &ptr, &end);
 786                                 if (ret)
 787                                         goto next;
 788                         }
 789                 }
 790
 791                 if (ptr < end) {
 792                         /* update key for inline back ref */
 793                         struct btrfs_extent_inline_ref *iref;
 794                         iref = (struct btrfs_extent_inline_ref *)ptr;
 795                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 796                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 797                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 798                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 799                 }
 800
 801                 if (exist &&
 802                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 803                       exist->owner == key.offset) ||
 804                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 805                       exist->bytenr == key.offset))) {
 806                         exist = NULL;
 807                         goto next;
 808                 }
 809
 810 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 811                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 812                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 813                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 814                                 struct btrfs_extent_ref_v0 *ref0;
 815                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 816                                                 struct btrfs_extent_ref_v0);
 817                                 if (key.objectid == key.offset) {
 818                                         root = find_tree_root(rc, eb, ref0);
 819                                         if (root && !should_ignore_root(root))
 820                                                 cur->root = root;
 821                                         else
 822                                                 list_add(&cur->list, &useless);
 823                                         break;
 824                                 }
 825                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 826                                                                       ref0)))
 827                                         cur->cowonly = 1;
 828                         }
 829 #else
 830                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 831                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 832 #endif
 833                         if (key.objectid == key.offset) {
 834                                 /*
 835                                  * only root blocks of reloc trees use
 836                                  * backref of this type.
 837                                  */
 838                                 root = find_reloc_root(rc, cur->bytenr);
 839                                 BUG_ON(!root);
 840                                 cur->root = root;
 841                                 break;
 842                         }
 843
 844                         edge = alloc_backref_edge(cache);
 845                         if (!edge) {
 846                                 err = -ENOMEM;
 847                                 goto out;
 848                         }
 849                         rb_node = tree_search(&cache->rb_root, key.offset);
 850                         if (!rb_node) {
 851                                 upper = alloc_backref_node(cache);
 852                                 if (!upper) {
 853                                         free_backref_edge(cache, edge);
 854                                         err = -ENOMEM;
 855                                         goto out;
 856                                 }
 857                                 upper->bytenr = key.offset;
 858                                 upper->level = cur->level + 1;
 859                                 /*
 860                                  *  backrefs for the upper level block isn't
 861                                  *  cached, add the block to pending list
 862                                  */
 863                                 list_add_tail(&edge->list[UPPER], &list);
 864                         } else {
 865                                 upper = rb_entry(rb_node, struct backref_node,
 866                                                  rb_node);
 867                                 BUG_ON(!upper->checked);
 868                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 869                         }
 870                         list_add_tail(&edge->list[LOWER], &cur->upper);
 871                         edge->node[LOWER] = cur;
 872                         edge->node[UPPER] = upper;
 873
 874                         goto next;
 875                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 876                         goto next;
 877                 }
 878
 879                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 880                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 881                 if (IS_ERR(root)) {
 882                         err = PTR_ERR(root);
 883                         goto out;
 884                 }
 885
 886                 if (!root->ref_cows)
 887                         cur->cowonly = 1;
 888
 889                 if (btrfs_root_level(&root->root_item) == cur->level) {
 890                         /* tree root */
 891                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 892                                cur->bytenr);
 893                         if (should_ignore_root(root))
 894                                 list_add(&cur->list, &useless);
 895                         else
 896                                 cur->root = root;
 897                         break;
 898                 }
 899
 900                 level = cur->level + 1;
 901
 902                 /*
 903                  * searching the tree to find upper level blocks
 904                  * reference the block.
 905                  */
 906                 path2->search_commit_root = 1;
 907                 path2->skip_locking = 1;
 908                 path2->lowest_level = level;
 909                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 910                 path2->lowest_level = 0;
 911                 if (ret < 0) {
 912                         err = ret;
 913                         goto out;
 914                 }
 915                 if (ret > 0 && path2->slots[level] > 0)
 916                         path2->slots[level]--;
 917
 918                 eb = path2->nodes[level];
 919                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 920                         cur->bytenr);
 921
 922                 lower = cur;
 923                 need_check = true;
 924                 for (; level < BTRFS_MAX_LEVEL; level++) {
 925                         if (!path2->nodes[level]) {
 926                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 927                                        lower->bytenr);
 928                                 if (should_ignore_root(root))
 929                                         list_add(&lower->list, &useless);
 930                                 else
 931                                         lower->root = root;
 932                                 break;
 933                         }
 934
 935                         edge = alloc_backref_edge(cache);
 936                         if (!edge) {
 937                                 err = -ENOMEM;
 938                                 goto out;
 939                         }
 940
 941                         eb = path2->nodes[level];
 942                         rb_node = tree_search(&cache->rb_root, eb->start);
 943                         if (!rb_node) {
 944                                 upper = alloc_backref_node(cache);
 945                                 if (!upper) {
 946                                         free_backref_edge(cache, edge);
 947                                         err = -ENOMEM;
 948                                         goto out;
 949                                 }
 950                                 upper->bytenr = eb->start;
 951                                 upper->owner = btrfs_header_owner(eb);
 952                                 upper->level = lower->level + 1;
 953                                 if (!root->ref_cows)
 954                                         upper->cowonly = 1;
 955
 956                                 /*
 957                                  * if we know the block isn't shared
 958                                  * we can void checking its backrefs.
 959                                  */
 960                                 if (btrfs_block_can_be_shared(root, eb))
 961                                         upper->checked = 0;
 962                                 else
 963                                         upper->checked = 1;
 964
 965                                 /*
 966                                  * add the block to pending list if we
 967                                  * need check its backrefs, we only do this once
 968                                  * while walking up a tree as we will catch
 969                                  * anything else later on.
 970                                  */
 971                                 if (!upper->checked && need_check) {
 972                                         need_check = false;
 973                                         list_add_tail(&edge->list[UPPER],
 974                                                       &list);
 975                                 } else
 976                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 977                         } else {
 978                                 upper = rb_entry(rb_node, struct backref_node,
 979                                                  rb_node);
 980                                 BUG_ON(!upper->checked);
 981                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 982                                 if (!upper->owner)
 983                                         upper->owner = btrfs_header_owner(eb);
 984                         }
 985                         list_add_tail(&edge->list[LOWER], &lower->upper);
 986                         edge->node[LOWER] = lower;
 987                         edge->node[UPPER] = upper;
 988
 989                         if (rb_node)
 990                                 break;
 991                         lower = upper;
 992                         upper = NULL;
 993                 }
 994                 btrfs_release_path(path2);
 995 next:
 996                 if (ptr < end) {
 997                         ptr += btrfs_extent_inline_ref_size(key.type);
 998                         if (ptr >= end) {
 999                                 WARN_ON(ptr > end);
1000                                 ptr = 0;
1001                                 end = 0;
1002                         }
1003                 }
1004                 if (ptr >= end)
1005                         path1->slots[0]++;
1006         }
1007         btrfs_release_path(path1);
1008
1009         cur->checked = 1;
1010         WARN_ON(exist);
1011
1012         /* the pending list isn't empty, take the first block to process */
1013         if (!list_empty(&list)) {
1014                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1015                 list_del_init(&edge->list[UPPER]);
1016                 cur = edge->node[UPPER];
1017                 goto again;
1018         }
1019
1020         /*
1021          * everything goes well, connect backref nodes and insert backref nodes
1022          * into the cache.
1023          */
1024         BUG_ON(!node->checked);
1025         cowonly = node->cowonly;
1026         if (!cowonly) {
1027                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1028                                       &node->rb_node);
1029                 if (rb_node)
1030                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1031                 list_add_tail(&node->lower, &cache->leaves);
1032         }
1033
1034         list_for_each_entry(edge, &node->upper, list[LOWER])
1035                 list_add_tail(&edge->list[UPPER], &list);
1036
1037         while (!list_empty(&list)) {
1038                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1039                 list_del_init(&edge->list[UPPER]);
1040                 upper = edge->node[UPPER];
1041                 if (upper->detached) {
1042                         list_del(&edge->list[LOWER]);
1043                         lower = edge->node[LOWER];
1044                         free_backref_edge(cache, edge);
1045                         if (list_empty(&lower->upper))
1046                                 list_add(&lower->list, &useless);
1047                         continue;
1048                 }
1049
1050                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1051                         if (upper->lowest) {
1052                                 list_del_init(&upper->lower);
1053                                 upper->lowest = 0;
1054                         }
1055
1056                         list_add_tail(&edge->list[UPPER], &upper->lower);
1057                         continue;
1058                 }
1059
1060                 BUG_ON(!upper->checked);
1061                 BUG_ON(cowonly != upper->cowonly);
1062                 if (!cowonly) {
1063                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1064                                               &upper->rb_node);
1065                         if (rb_node)
1066                                 backref_tree_panic(rb_node, -EEXIST,
1067                                                    upper->bytenr);
1068                 }
1069
1070                 list_add_tail(&edge->list[UPPER], &upper->lower);
1071
1072                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1073                         list_add_tail(&edge->list[UPPER], &list);
1074         }
1075         /*
1076          * process useless backref nodes. backref nodes for tree leaves
1077          * are deleted from the cache. backref nodes for upper level
1078          * tree blocks are left in the cache to avoid unnecessary backref
1079          * lookup.
1080          */
1081         while (!list_empty(&useless)) {
1082                 upper = list_entry(useless.next, struct backref_node, list);
1083                 list_del_init(&upper->list);
1084                 BUG_ON(!list_empty(&upper->upper));
1085                 if (upper == node)
1086                         node = NULL;
1087                 if (upper->lowest) {
1088                         list_del_init(&upper->lower);
1089                         upper->lowest = 0;
1090                 }
1091                 while (!list_empty(&upper->lower)) {
1092                         edge = list_entry(upper->lower.next,
1093                                           struct backref_edge, list[UPPER]);
1094                         list_del(&edge->list[UPPER]);
1095                         list_del(&edge->list[LOWER]);
1096                         lower = edge->node[LOWER];
1097                         free_backref_edge(cache, edge);
1098
1099                         if (list_empty(&lower->upper))
1100                                 list_add(&lower->list, &useless);
1101                 }
1102                 __mark_block_processed(rc, upper);
1103                 if (upper->level > 0) {
1104                         list_add(&upper->list, &cache->detached);
1105                         upper->detached = 1;
1106                 } else {
1107                         rb_erase(&upper->rb_node, &cache->rb_root);
1108                         free_backref_node(cache, upper);
1109                 }
1110         }
1111 out:
1112         btrfs_free_path(path1);
1113         btrfs_free_path(path2);
1114         if (err) {
1115                 while (!list_empty(&useless)) {
1116                         lower = list_entry(useless.next,
1117                                            struct backref_node, upper);
1118                         list_del_init(&lower->upper);
1119                 }
1120                 upper = node;
1121                 INIT_LIST_HEAD(&list);
1122                 while (upper) {
1123                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1124                                 list_splice_tail(&upper->upper, &list);
1125                                 free_backref_node(cache, upper);
1126                         }
1127
1128                         if (list_empty(&list))
1129                                 break;
1130
1131                         edge = list_entry(list.next, struct backref_edge,
1132                                           list[LOWER]);
1133                         list_del(&edge->list[LOWER]);
1134                         upper = edge->node[UPPER];
1135                         free_backref_edge(cache, edge);
1136                 }
1137                 return ERR_PTR(err);
1138         }
1139         BUG_ON(node && node->detached);
1140         return node;
1141 }
1142
1143 /*
1144  * helper to add backref node for the newly created snapshot.
1145  * the backref node is created by cloning backref node that
1146  * corresponds to root of source tree
1147  */
1148 static int clone_backref_node(struct btrfs_trans_handle *trans,
1149                               struct reloc_control *rc,
1150                               struct btrfs_root *src,
1151                               struct btrfs_root *dest)
1152 {
1153         struct btrfs_root *reloc_root = src->reloc_root;
1154         struct backref_cache *cache = &rc->backref_cache;
1155         struct backref_node *node = NULL;
1156         struct backref_node *new_node;
1157         struct backref_edge *edge;
1158         struct backref_edge *new_edge;
1159         struct rb_node *rb_node;
1160
1161         if (cache->last_trans > 0)
1162                 update_backref_cache(trans, cache);
1163
1164         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1165         if (rb_node) {
1166                 node = rb_entry(rb_node, struct backref_node, rb_node);
1167                 if (node->detached)
1168                         node = NULL;
1169                 else
1170                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1171         }
1172
1173         if (!node) {
1174                 rb_node = tree_search(&cache->rb_root,
1175                                       reloc_root->commit_root->start);
1176                 if (rb_node) {
1177                         node = rb_entry(rb_node, struct backref_node,
1178                                         rb_node);
1179                         BUG_ON(node->detached);
1180                 }
1181         }
1182
1183         if (!node)
1184                 return 0;
1185
1186         new_node = alloc_backref_node(cache);
1187         if (!new_node)
1188                 return -ENOMEM;
1189
1190         new_node->bytenr = dest->node->start;
1191         new_node->level = node->level;
1192         new_node->lowest = node->lowest;
1193         new_node->checked = 1;
1194         new_node->root = dest;
1195
1196         if (!node->lowest) {
1197                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1198                         new_edge = alloc_backref_edge(cache);
1199                         if (!new_edge)
1200                                 goto fail;
1201
1202                         new_edge->node[UPPER] = new_node;
1203                         new_edge->node[LOWER] = edge->node[LOWER];
1204                         list_add_tail(&new_edge->list[UPPER],
1205                                       &new_node->lower);
1206                 }
1207         } else {
1208                 list_add_tail(&new_node->lower, &cache->leaves);
1209         }
1210
1211         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1212                               &new_node->rb_node);
1213         if (rb_node)
1214                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1215
1216         if (!new_node->lowest) {
1217                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1218                         list_add_tail(&new_edge->list[LOWER],
1219                                       &new_edge->node[LOWER]->upper);
1220                 }
1221         }
1222         return 0;
1223 fail:
1224         while (!list_empty(&new_node->lower)) {
1225                 new_edge = list_entry(new_node->lower.next,
1226                                       struct backref_edge, list[UPPER]);
1227                 list_del(&new_edge->list[UPPER]);
1228                 free_backref_edge(cache, new_edge);
1229         }
1230         free_backref_node(cache, new_node);
1231         return -ENOMEM;
1232 }
1233
1234 /*
1235  * helper to add 'address of tree root -> reloc tree' mapping
1236  */
1237 static int __must_check __add_reloc_root(struct btrfs_root *root)
1238 {
1239         struct rb_node *rb_node;
1240         struct mapping_node *node;
1241         struct reloc_control *rc = root->fs_info->reloc_ctl;
1242
1243         node = kmalloc(sizeof(*node), GFP_NOFS);
1244         if (!node)
1245                 return -ENOMEM;
1246
1247         node->bytenr = root->node->start;
1248         node->data = root;
1249
1250         spin_lock(&rc->reloc_root_tree.lock);
1251         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1252                               node->bytenr, &node->rb_node);
1253         spin_unlock(&rc->reloc_root_tree.lock);
1254         if (rb_node) {
1255                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1256                             "for start=%llu while inserting into relocation "
1257                             "tree\n", node->bytenr);
1258                 kfree(node);
1259                 return -EEXIST;
1260         }
1261
1262         list_add_tail(&root->root_list, &rc->reloc_roots);
1263         return 0;
1264 }
1265
1266 /*
1267  * helper to delete the 'address of tree root -> reloc tree'
1268  * mapping
1269  */
1270 static void __del_reloc_root(struct btrfs_root *root)
1271 {
1272         struct rb_node *rb_node;
1273         struct mapping_node *node = NULL;
1274         struct reloc_control *rc = root->fs_info->reloc_ctl;
1275
1276         spin_lock(&rc->reloc_root_tree.lock);
1277         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1278                               root->node->start);
1279         if (rb_node) {
1280                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1281                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1282         }
1283         spin_unlock(&rc->reloc_root_tree.lock);
1284
1285         if (!node)
1286                 return;
1287         BUG_ON((struct btrfs_root *)node->data != root);
1288
1289         spin_lock(&root->fs_info->trans_lock);
1290         list_del_init(&root->root_list);
1291         spin_unlock(&root->fs_info->trans_lock);
1292         kfree(node);
1293 }
1294
1295 /*
1296  * helper to update the 'address of tree root -> reloc tree'
1297  * mapping
1298  */
1299 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1300 {
1301         struct rb_node *rb_node;
1302         struct mapping_node *node = NULL;
1303         struct reloc_control *rc = root->fs_info->reloc_ctl;
1304
1305         spin_lock(&rc->reloc_root_tree.lock);
1306         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1307                               root->node->start);
1308         if (rb_node) {
1309                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1310                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1311         }
1312         spin_unlock(&rc->reloc_root_tree.lock);
1313
1314         if (!node)
1315                 return 0;
1316         BUG_ON((struct btrfs_root *)node->data != root);
1317
1318         spin_lock(&rc->reloc_root_tree.lock);
1319         node->bytenr = new_bytenr;
1320         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1321                               node->bytenr, &node->rb_node);
1322         spin_unlock(&rc->reloc_root_tree.lock);
1323         if (rb_node)
1324                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1325         return 0;
1326 }
1327
1328 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1329                                         struct btrfs_root *root, u64 objectid)
1330 {
1331         struct btrfs_root *reloc_root;
1332         struct extent_buffer *eb;
1333         struct btrfs_root_item *root_item;
1334         struct btrfs_key root_key;
1335         u64 last_snap = 0;
1336         int ret;
1337
1338         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1339         BUG_ON(!root_item);
1340
1341         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1342         root_key.type = BTRFS_ROOT_ITEM_KEY;
1343         root_key.offset = objectid;
1344
1345         if (root->root_key.objectid == objectid) {
1346                 /* called by btrfs_init_reloc_root */
1347                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1348                                       BTRFS_TREE_RELOC_OBJECTID);
1349                 BUG_ON(ret);
1350
1351                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1352                 btrfs_set_root_last_snapshot(&root->root_item,
1353                                              trans->transid - 1);
1354         } else {
1355                 /*
1356                  * called by btrfs_reloc_post_snapshot_hook.
1357                  * the source tree is a reloc tree, all tree blocks
1358                  * modified after it was created have RELOC flag
1359                  * set in their headers. so it's OK to not update
1360                  * the 'last_snapshot'.
1361                  */
1362                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1363                                       BTRFS_TREE_RELOC_OBJECTID);
1364                 BUG_ON(ret);
1365         }
1366
1367         memcpy(root_item, &root->root_item, sizeof(*root_item));
1368         btrfs_set_root_bytenr(root_item, eb->start);
1369         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1370         btrfs_set_root_generation(root_item, trans->transid);
1371
1372         if (root->root_key.objectid == objectid) {
1373                 btrfs_set_root_refs(root_item, 0);
1374                 memset(&root_item->drop_progress, 0,
1375                        sizeof(struct btrfs_disk_key));
1376                 root_item->drop_level = 0;
1377                 /*
1378                  * abuse rtransid, it is safe because it is impossible to
1379                  * receive data into a relocation tree.
1380                  */
1381                 btrfs_set_root_rtransid(root_item, last_snap);
1382                 btrfs_set_root_otransid(root_item, trans->transid);
1383         }
1384
1385         btrfs_tree_unlock(eb);
1386         free_extent_buffer(eb);
1387
1388         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1389                                 &root_key, root_item);
1390         BUG_ON(ret);
1391         kfree(root_item);
1392
1393         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1394         BUG_ON(IS_ERR(reloc_root));
1395         reloc_root->last_trans = trans->transid;
1396         return reloc_root;
1397 }
1398
1399 /*
1400  * create reloc tree for a given fs tree. reloc tree is just a
1401  * snapshot of the fs tree with special root objectid.
1402  */
1403 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1404                           struct btrfs_root *root)
1405 {
1406         struct btrfs_root *reloc_root;
1407         struct reloc_control *rc = root->fs_info->reloc_ctl;
1408         struct btrfs_block_rsv *rsv;
1409         int clear_rsv = 0;
1410         int ret;
1411
1412         if (root->reloc_root) {
1413                 reloc_root = root->reloc_root;
1414                 reloc_root->last_trans = trans->transid;
1415                 return 0;
1416         }
1417
1418         if (!rc || !rc->create_reloc_tree ||
1419             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1420                 return 0;
1421
1422         if (!trans->reloc_reserved) {
1423                 rsv = trans->block_rsv;
1424                 trans->block_rsv = rc->block_rsv;
1425                 clear_rsv = 1;
1426         }
1427         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1428         if (clear_rsv)
1429                 trans->block_rsv = rsv;
1430
1431         ret = __add_reloc_root(reloc_root);
1432         BUG_ON(ret < 0);
1433         root->reloc_root = reloc_root;
1434         return 0;
1435 }
1436
1437 /*
1438  * update root item of reloc tree
1439  */
1440 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1441                             struct btrfs_root *root)
1442 {
1443         struct btrfs_root *reloc_root;
1444         struct btrfs_root_item *root_item;
1445         int ret;
1446
1447         if (!root->reloc_root)
1448                 goto out;
1449
1450         reloc_root = root->reloc_root;
1451         root_item = &reloc_root->root_item;
1452
1453         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1454             btrfs_root_refs(root_item) == 0) {
1455                 root->reloc_root = NULL;
1456                 __del_reloc_root(reloc_root);
1457         }
1458
1459         if (reloc_root->commit_root != reloc_root->node) {
1460                 btrfs_set_root_node(root_item, reloc_root->node);
1461                 free_extent_buffer(reloc_root->commit_root);
1462                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1463         }
1464
1465         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1466                                 &reloc_root->root_key, root_item);
1467         BUG_ON(ret);
1468
1469 out:
1470         return 0;
1471 }
1472
1473 /*
1474  * helper to find first cached inode with inode number >= objectid
1475  * in a subvolume
1476  */
1477 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1478 {
1479         struct rb_node *node;
1480         struct rb_node *prev;
1481         struct btrfs_inode *entry;
1482         struct inode *inode;
1483
1484         spin_lock(&root->inode_lock);
1485 again:
1486         node = root->inode_tree.rb_node;
1487         prev = NULL;
1488         while (node) {
1489                 prev = node;
1490                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1491
1492                 if (objectid < btrfs_ino(&entry->vfs_inode))
1493                         node = node->rb_left;
1494                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1495                         node = node->rb_right;
1496                 else
1497                         break;
1498         }
1499         if (!node) {
1500                 while (prev) {
1501                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1502                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1503                                 node = prev;
1504                                 break;
1505                         }
1506                         prev = rb_next(prev);
1507                 }
1508         }
1509         while (node) {
1510                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1511                 inode = igrab(&entry->vfs_inode);
1512                 if (inode) {
1513                         spin_unlock(&root->inode_lock);
1514                         return inode;
1515                 }
1516
1517                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1518                 if (cond_resched_lock(&root->inode_lock))
1519                         goto again;
1520
1521                 node = rb_next(node);
1522         }
1523         spin_unlock(&root->inode_lock);
1524         return NULL;
1525 }
1526
1527 static int in_block_group(u64 bytenr,
1528                           struct btrfs_block_group_cache *block_group)
1529 {
1530         if (bytenr >= block_group->key.objectid &&
1531             bytenr < block_group->key.objectid + block_group->key.offset)
1532                 return 1;
1533         return 0;
1534 }
1535
1536 /*
1537  * get new location of data
1538  */
1539 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1540                             u64 bytenr, u64 num_bytes)
1541 {
1542         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1543         struct btrfs_path *path;
1544         struct btrfs_file_extent_item *fi;
1545         struct extent_buffer *leaf;
1546         int ret;
1547
1548         path = btrfs_alloc_path();
1549         if (!path)
1550                 return -ENOMEM;
1551
1552         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1553         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1554                                        bytenr, 0);
1555         if (ret < 0)
1556                 goto out;
1557         if (ret > 0) {
1558                 ret = -ENOENT;
1559                 goto out;
1560         }
1561
1562         leaf = path->nodes[0];
1563         fi = btrfs_item_ptr(leaf, path->slots[0],
1564                             struct btrfs_file_extent_item);
1565
1566         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1567                btrfs_file_extent_compression(leaf, fi) ||
1568                btrfs_file_extent_encryption(leaf, fi) ||
1569                btrfs_file_extent_other_encoding(leaf, fi));
1570
1571         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1572                 ret = -EINVAL;
1573                 goto out;
1574         }
1575
1576         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1577         ret = 0;
1578 out:
1579         btrfs_free_path(path);
1580         return ret;
1581 }
1582
1583 /*
1584  * update file extent items in the tree leaf to point to
1585  * the new locations.
1586  */
1587 static noinline_for_stack
1588 int replace_file_extents(struct btrfs_trans_handle *trans,
1589                          struct reloc_control *rc,
1590                          struct btrfs_root *root,
1591                          struct extent_buffer *leaf)
1592 {
1593         struct btrfs_key key;
1594         struct btrfs_file_extent_item *fi;
1595         struct inode *inode = NULL;
1596         u64 parent;
1597         u64 bytenr;
1598         u64 new_bytenr = 0;
1599         u64 num_bytes;
1600         u64 end;
1601         u32 nritems;
1602         u32 i;
1603         int ret = 0;
1604         int first = 1;
1605         int dirty = 0;
1606
1607         if (rc->stage != UPDATE_DATA_PTRS)
1608                 return 0;
1609
1610         /* reloc trees always use full backref */
1611         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1612                 parent = leaf->start;
1613         else
1614                 parent = 0;
1615
1616         nritems = btrfs_header_nritems(leaf);
1617         for (i = 0; i < nritems; i++) {
1618                 cond_resched();
1619                 btrfs_item_key_to_cpu(leaf, &key, i);
1620                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1621                         continue;
1622                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1623                 if (btrfs_file_extent_type(leaf, fi) ==
1624                     BTRFS_FILE_EXTENT_INLINE)
1625                         continue;
1626                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1627                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1628                 if (bytenr == 0)
1629                         continue;
1630                 if (!in_block_group(bytenr, rc->block_group))
1631                         continue;
1632
1633                 /*
1634                  * if we are modifying block in fs tree, wait for readpage
1635                  * to complete and drop the extent cache
1636                  */
1637                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1638                         if (first) {
1639                                 inode = find_next_inode(root, key.objectid);
1640                                 first = 0;
1641                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1642                                 btrfs_add_delayed_iput(inode);
1643                                 inode = find_next_inode(root, key.objectid);
1644                         }
1645                         if (inode && btrfs_ino(inode) == key.objectid) {
1646                                 end = key.offset +
1647                                       btrfs_file_extent_num_bytes(leaf, fi);
1648                                 WARN_ON(!IS_ALIGNED(key.offset,
1649                                                     root->sectorsize));
1650                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1651                                 end--;
1652                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1653                                                       key.offset, end);
1654                                 if (!ret)
1655                                         continue;
1656
1657                                 btrfs_drop_extent_cache(inode, key.offset, end,
1658                                                         1);
1659                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1660                                               key.offset, end);
1661                         }
1662                 }
1663
1664                 ret = get_new_location(rc->data_inode, &new_bytenr,
1665                                        bytenr, num_bytes);
1666                 if (ret) {
1667                         /*
1668                          * Don't have to abort since we've not changed anything
1669                          * in the file extent yet.
1670                          */
1671                         break;
1672                 }
1673
1674                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1675                 dirty = 1;
1676
1677                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1678                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1679                                            num_bytes, parent,
1680                                            btrfs_header_owner(leaf),
1681                                            key.objectid, key.offset, 1);
1682                 if (ret) {
1683                         btrfs_abort_transaction(trans, root, ret);
1684                         break;
1685                 }
1686
1687                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1688                                         parent, btrfs_header_owner(leaf),
1689                                         key.objectid, key.offset, 1);
1690                 if (ret) {
1691                         btrfs_abort_transaction(trans, root, ret);
1692                         break;
1693                 }
1694         }
1695         if (dirty)
1696                 btrfs_mark_buffer_dirty(leaf);
1697         if (inode)
1698                 btrfs_add_delayed_iput(inode);
1699         return ret;
1700 }
1701
1702 static noinline_for_stack
1703 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1704                      struct btrfs_path *path, int level)
1705 {
1706         struct btrfs_disk_key key1;
1707         struct btrfs_disk_key key2;
1708         btrfs_node_key(eb, &key1, slot);
1709         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1710         return memcmp(&key1, &key2, sizeof(key1));
1711 }
1712
1713 /*
1714  * try to replace tree blocks in fs tree with the new blocks
1715  * in reloc tree. tree blocks haven't been modified since the
1716  * reloc tree was create can be replaced.
1717  *
1718  * if a block was replaced, level of the block + 1 is returned.
1719  * if no block got replaced, 0 is returned. if there are other
1720  * errors, a negative error number is returned.
1721  */
1722 static noinline_for_stack
1723 int replace_path(struct btrfs_trans_handle *trans,
1724                  struct btrfs_root *dest, struct btrfs_root *src,
1725                  struct btrfs_path *path, struct btrfs_key *next_key,
1726                  int lowest_level, int max_level)
1727 {
1728         struct extent_buffer *eb;
1729         struct extent_buffer *parent;
1730         struct btrfs_key key;
1731         u64 old_bytenr;
1732         u64 new_bytenr;
1733         u64 old_ptr_gen;
1734         u64 new_ptr_gen;
1735         u64 last_snapshot;
1736         u32 blocksize;
1737         int cow = 0;
1738         int level;
1739         int ret;
1740         int slot;
1741
1742         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1743         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1744
1745         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1746 again:
1747         slot = path->slots[lowest_level];
1748         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1749
1750         eb = btrfs_lock_root_node(dest);
1751         btrfs_set_lock_blocking(eb);
1752         level = btrfs_header_level(eb);
1753
1754         if (level < lowest_level) {
1755                 btrfs_tree_unlock(eb);
1756                 free_extent_buffer(eb);
1757                 return 0;
1758         }
1759
1760         if (cow) {
1761                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1762                 BUG_ON(ret);
1763         }
1764         btrfs_set_lock_blocking(eb);
1765
1766         if (next_key) {
1767                 next_key->objectid = (u64)-1;
1768                 next_key->type = (u8)-1;
1769                 next_key->offset = (u64)-1;
1770         }
1771
1772         parent = eb;
1773         while (1) {
1774                 level = btrfs_header_level(parent);
1775                 BUG_ON(level < lowest_level);
1776
1777                 ret = btrfs_bin_search(parent, &key, level, &slot);
1778                 if (ret && slot > 0)
1779                         slot--;
1780
1781                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1782                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1783
1784                 old_bytenr = btrfs_node_blockptr(parent, slot);
1785                 blocksize = btrfs_level_size(dest, level - 1);
1786                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1787
1788                 if (level <= max_level) {
1789                         eb = path->nodes[level];
1790                         new_bytenr = btrfs_node_blockptr(eb,
1791                                                         path->slots[level]);
1792                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1793                                                         path->slots[level]);
1794                 } else {
1795                         new_bytenr = 0;
1796                         new_ptr_gen = 0;
1797                 }
1798
1799                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1800                         ret = level;
1801                         break;
1802                 }
1803
1804                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1805                     memcmp_node_keys(parent, slot, path, level)) {
1806                         if (level <= lowest_level) {
1807                                 ret = 0;
1808                                 break;
1809                         }
1810
1811                         eb = read_tree_block(dest, old_bytenr, blocksize,
1812                                              old_ptr_gen);
1813                         if (!eb || !extent_buffer_uptodate(eb)) {
1814                                 ret = (!eb) ? -ENOMEM : -EIO;
1815                                 free_extent_buffer(eb);
1816                                 break;
1817                         }
1818                         btrfs_tree_lock(eb);
1819                         if (cow) {
1820                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1821                                                       slot, &eb);
1822                                 BUG_ON(ret);
1823                         }
1824                         btrfs_set_lock_blocking(eb);
1825
1826                         btrfs_tree_unlock(parent);
1827                         free_extent_buffer(parent);
1828
1829                         parent = eb;
1830                         continue;
1831                 }
1832
1833                 if (!cow) {
1834                         btrfs_tree_unlock(parent);
1835                         free_extent_buffer(parent);
1836                         cow = 1;
1837                         goto again;
1838                 }
1839
1840                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1841                                       path->slots[level]);
1842                 btrfs_release_path(path);
1843
1844                 path->lowest_level = level;
1845                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1846                 path->lowest_level = 0;
1847                 BUG_ON(ret);
1848
1849                 /*
1850                  * swap blocks in fs tree and reloc tree.
1851                  */
1852                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1853                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1854                 btrfs_mark_buffer_dirty(parent);
1855
1856                 btrfs_set_node_blockptr(path->nodes[level],
1857                                         path->slots[level], old_bytenr);
1858                 btrfs_set_node_ptr_generation(path->nodes[level],
1859                                               path->slots[level], old_ptr_gen);
1860                 btrfs_mark_buffer_dirty(path->nodes[level]);
1861
1862                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1863                                         path->nodes[level]->start,
1864                                         src->root_key.objectid, level - 1, 0,
1865                                         1);
1866                 BUG_ON(ret);
1867                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1868                                         0, dest->root_key.objectid, level - 1,
1869                                         0, 1);
1870                 BUG_ON(ret);
1871
1872                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1873                                         path->nodes[level]->start,
1874                                         src->root_key.objectid, level - 1, 0,
1875                                         1);
1876                 BUG_ON(ret);
1877
1878                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1879                                         0, dest->root_key.objectid, level - 1,
1880                                         0, 1);
1881                 BUG_ON(ret);
1882
1883                 btrfs_unlock_up_safe(path, 0);
1884
1885                 ret = level;
1886                 break;
1887         }
1888         btrfs_tree_unlock(parent);
1889         free_extent_buffer(parent);
1890         return ret;
1891 }
1892
1893 /*
1894  * helper to find next relocated block in reloc tree
1895  */
1896 static noinline_for_stack
1897 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1898                        int *level)
1899 {
1900         struct extent_buffer *eb;
1901         int i;
1902         u64 last_snapshot;
1903         u32 nritems;
1904
1905         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1906
1907         for (i = 0; i < *level; i++) {
1908                 free_extent_buffer(path->nodes[i]);
1909                 path->nodes[i] = NULL;
1910         }
1911
1912         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1913                 eb = path->nodes[i];
1914                 nritems = btrfs_header_nritems(eb);
1915                 while (path->slots[i] + 1 < nritems) {
1916                         path->slots[i]++;
1917                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1918                             last_snapshot)
1919                                 continue;
1920
1921                         *level = i;
1922                         return 0;
1923                 }
1924                 free_extent_buffer(path->nodes[i]);
1925                 path->nodes[i] = NULL;
1926         }
1927         return 1;
1928 }
1929
1930 /*
1931  * walk down reloc tree to find relocated block of lowest level
1932  */
1933 static noinline_for_stack
1934 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1935                          int *level)
1936 {
1937         struct extent_buffer *eb = NULL;
1938         int i;
1939         u64 bytenr;
1940         u64 ptr_gen = 0;
1941         u64 last_snapshot;
1942         u32 blocksize;
1943         u32 nritems;
1944
1945         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1946
1947         for (i = *level; i > 0; i--) {
1948                 eb = path->nodes[i];
1949                 nritems = btrfs_header_nritems(eb);
1950                 while (path->slots[i] < nritems) {
1951                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1952                         if (ptr_gen > last_snapshot)
1953                                 break;
1954                         path->slots[i]++;
1955                 }
1956                 if (path->slots[i] >= nritems) {
1957                         if (i == *level)
1958                                 break;
1959                         *level = i + 1;
1960                         return 0;
1961                 }
1962                 if (i == 1) {
1963                         *level = i;
1964                         return 0;
1965                 }
1966
1967                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1968                 blocksize = btrfs_level_size(root, i - 1);
1969                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1970                 if (!eb || !extent_buffer_uptodate(eb)) {
1971                         free_extent_buffer(eb);
1972                         return -EIO;
1973                 }
1974                 BUG_ON(btrfs_header_level(eb) != i - 1);
1975                 path->nodes[i - 1] = eb;
1976                 path->slots[i - 1] = 0;
1977         }
1978         return 1;
1979 }
1980
1981 /*
1982  * invalidate extent cache for file extents whose key in range of
1983  * [min_key, max_key)
1984  */
1985 static int invalidate_extent_cache(struct btrfs_root *root,
1986                                    struct btrfs_key *min_key,
1987                                    struct btrfs_key *max_key)
1988 {
1989         struct inode *inode = NULL;
1990         u64 objectid;
1991         u64 start, end;
1992         u64 ino;
1993
1994         objectid = min_key->objectid;
1995         while (1) {
1996                 cond_resched();
1997                 iput(inode);
1998
1999                 if (objectid > max_key->objectid)
2000                         break;
2001
2002                 inode = find_next_inode(root, objectid);
2003                 if (!inode)
2004                         break;
2005                 ino = btrfs_ino(inode);
2006
2007                 if (ino > max_key->objectid) {
2008                         iput(inode);
2009                         break;
2010                 }
2011
2012                 objectid = ino + 1;
2013                 if (!S_ISREG(inode->i_mode))
2014                         continue;
2015
2016                 if (unlikely(min_key->objectid == ino)) {
2017                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2018                                 continue;
2019                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2020                                 start = 0;
2021                         else {
2022                                 start = min_key->offset;
2023                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2024                         }
2025                 } else {
2026                         start = 0;
2027                 }
2028
2029                 if (unlikely(max_key->objectid == ino)) {
2030                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2031                                 continue;
2032                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2033                                 end = (u64)-1;
2034                         } else {
2035                                 if (max_key->offset == 0)
2036                                         continue;
2037                                 end = max_key->offset;
2038                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2039                                 end--;
2040                         }
2041                 } else {
2042                         end = (u64)-1;
2043                 }
2044
2045                 /* the lock_extent waits for readpage to complete */
2046                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2047                 btrfs_drop_extent_cache(inode, start, end, 1);
2048                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2049         }
2050         return 0;
2051 }
2052
2053 static int find_next_key(struct btrfs_path *path, int level,
2054                          struct btrfs_key *key)
2055
2056 {
2057         while (level < BTRFS_MAX_LEVEL) {
2058                 if (!path->nodes[level])
2059                         break;
2060                 if (path->slots[level] + 1 <
2061                     btrfs_header_nritems(path->nodes[level])) {
2062                         btrfs_node_key_to_cpu(path->nodes[level], key,
2063                                               path->slots[level] + 1);
2064                         return 0;
2065                 }
2066                 level++;
2067         }
2068         return 1;
2069 }
2070
2071 /*
2072  * merge the relocated tree blocks in reloc tree with corresponding
2073  * fs tree.
2074  */
2075 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2076                                                struct btrfs_root *root)
2077 {
2078         LIST_HEAD(inode_list);
2079         struct btrfs_key key;
2080         struct btrfs_key next_key;
2081         struct btrfs_trans_handle *trans = NULL;
2082         struct btrfs_root *reloc_root;
2083         struct btrfs_root_item *root_item;
2084         struct btrfs_path *path;
2085         struct extent_buffer *leaf;
2086         int level;
2087         int max_level;
2088         int replaced = 0;
2089         int ret;
2090         int err = 0;
2091         u32 min_reserved;
2092
2093         path = btrfs_alloc_path();
2094         if (!path)
2095                 return -ENOMEM;
2096         path->reada = 1;
2097
2098         reloc_root = root->reloc_root;
2099         root_item = &reloc_root->root_item;
2100
2101         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2102                 level = btrfs_root_level(root_item);
2103                 extent_buffer_get(reloc_root->node);
2104                 path->nodes[level] = reloc_root->node;
2105                 path->slots[level] = 0;
2106         } else {
2107                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2108
2109                 level = root_item->drop_level;
2110                 BUG_ON(level == 0);
2111                 path->lowest_level = level;
2112                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2113                 path->lowest_level = 0;
2114                 if (ret < 0) {
2115                         btrfs_free_path(path);
2116                         return ret;
2117                 }
2118
2119                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2120                                       path->slots[level]);
2121                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2122
2123                 btrfs_unlock_up_safe(path, 0);
2124         }
2125
2126         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2127         memset(&next_key, 0, sizeof(next_key));
2128
2129         while (1) {
2130                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2131                                              BTRFS_RESERVE_FLUSH_ALL);
2132                 if (ret) {
2133                         err = ret;
2134                         goto out;
2135                 }
2136                 trans = btrfs_start_transaction(root, 0);
2137                 if (IS_ERR(trans)) {
2138                         err = PTR_ERR(trans);
2139                         trans = NULL;
2140                         goto out;
2141                 }
2142                 trans->block_rsv = rc->block_rsv;
2143
2144                 replaced = 0;
2145                 max_level = level;
2146
2147                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2148                 if (ret < 0) {
2149                         err = ret;
2150                         goto out;
2151                 }
2152                 if (ret > 0)
2153                         break;
2154
2155                 if (!find_next_key(path, level, &key) &&
2156                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2157                         ret = 0;
2158                 } else {
2159                         ret = replace_path(trans, root, reloc_root, path,
2160                                            &next_key, level, max_level);
2161                 }
2162                 if (ret < 0) {
2163                         err = ret;
2164                         goto out;
2165                 }
2166
2167                 if (ret > 0) {
2168                         level = ret;
2169                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2170                                               path->slots[level]);
2171                         replaced = 1;
2172                 }
2173
2174                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2175                 if (ret > 0)
2176                         break;
2177
2178                 BUG_ON(level == 0);
2179                 /*
2180                  * save the merging progress in the drop_progress.
2181                  * this is OK since root refs == 1 in this case.
2182                  */
2183                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2184                                path->slots[level]);
2185                 root_item->drop_level = level;
2186
2187                 btrfs_end_transaction_throttle(trans, root);
2188                 trans = NULL;
2189
2190                 btrfs_btree_balance_dirty(root);
2191
2192                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2193                         invalidate_extent_cache(root, &key, &next_key);
2194         }
2195
2196         /*
2197          * handle the case only one block in the fs tree need to be
2198          * relocated and the block is tree root.
2199          */
2200         leaf = btrfs_lock_root_node(root);
2201         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2202         btrfs_tree_unlock(leaf);
2203         free_extent_buffer(leaf);
2204         if (ret < 0)
2205                 err = ret;
2206 out:
2207         btrfs_free_path(path);
2208
2209         if (err == 0) {
2210                 memset(&root_item->drop_progress, 0,
2211                        sizeof(root_item->drop_progress));
2212                 root_item->drop_level = 0;
2213                 btrfs_set_root_refs(root_item, 0);
2214                 btrfs_update_reloc_root(trans, root);
2215         }
2216
2217         if (trans)
2218                 btrfs_end_transaction_throttle(trans, root);
2219
2220         btrfs_btree_balance_dirty(root);
2221
2222         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2223                 invalidate_extent_cache(root, &key, &next_key);
2224
2225         return err;
2226 }
2227
2228 static noinline_for_stack
2229 int prepare_to_merge(struct reloc_control *rc, int err)
2230 {
2231         struct btrfs_root *root = rc->extent_root;
2232         struct btrfs_root *reloc_root;
2233         struct btrfs_trans_handle *trans;
2234         LIST_HEAD(reloc_roots);
2235         u64 num_bytes = 0;
2236         int ret;
2237
2238         mutex_lock(&root->fs_info->reloc_mutex);
2239         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2240         rc->merging_rsv_size += rc->nodes_relocated * 2;
2241         mutex_unlock(&root->fs_info->reloc_mutex);
2242
2243 again:
2244         if (!err) {
2245                 num_bytes = rc->merging_rsv_size;
2246                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2247                                           BTRFS_RESERVE_FLUSH_ALL);
2248                 if (ret)
2249                         err = ret;
2250         }
2251
2252         trans = btrfs_join_transaction(rc->extent_root);
2253         if (IS_ERR(trans)) {
2254                 if (!err)
2255                         btrfs_block_rsv_release(rc->extent_root,
2256                                                 rc->block_rsv, num_bytes);
2257                 return PTR_ERR(trans);
2258         }
2259
2260         if (!err) {
2261                 if (num_bytes != rc->merging_rsv_size) {
2262                         btrfs_end_transaction(trans, rc->extent_root);
2263                         btrfs_block_rsv_release(rc->extent_root,
2264                                                 rc->block_rsv, num_bytes);
2265                         goto again;
2266                 }
2267         }
2268
2269         rc->merge_reloc_tree = 1;
2270
2271         while (!list_empty(&rc->reloc_roots)) {
2272                 reloc_root = list_entry(rc->reloc_roots.next,
2273                                         struct btrfs_root, root_list);
2274                 list_del_init(&reloc_root->root_list);
2275
2276                 root = read_fs_root(reloc_root->fs_info,
2277                                     reloc_root->root_key.offset);
2278                 BUG_ON(IS_ERR(root));
2279                 BUG_ON(root->reloc_root != reloc_root);
2280
2281                 /*
2282                  * set reference count to 1, so btrfs_recover_relocation
2283                  * knows it should resumes merging
2284                  */
2285                 if (!err)
2286                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2287                 btrfs_update_reloc_root(trans, root);
2288
2289                 list_add(&reloc_root->root_list, &reloc_roots);
2290         }
2291
2292         list_splice(&reloc_roots, &rc->reloc_roots);
2293
2294         if (!err)
2295                 btrfs_commit_transaction(trans, rc->extent_root);
2296         else
2297                 btrfs_end_transaction(trans, rc->extent_root);
2298         return err;
2299 }
2300
2301 static noinline_for_stack
2302 void free_reloc_roots(struct list_head *list)
2303 {
2304         struct btrfs_root *reloc_root;
2305
2306         while (!list_empty(list)) {
2307                 reloc_root = list_entry(list->next, struct btrfs_root,
2308                                         root_list);
2309                 __del_reloc_root(reloc_root);
2310                 free_extent_buffer(reloc_root->node);
2311                 free_extent_buffer(reloc_root->commit_root);
2312                 kfree(reloc_root);
2313         }
2314 }
2315
2316 static noinline_for_stack
2317 int merge_reloc_roots(struct reloc_control *rc)
2318 {
2319         struct btrfs_trans_handle *trans;
2320         struct btrfs_root *root;
2321         struct btrfs_root *reloc_root;
2322         u64 last_snap;
2323         u64 otransid;
2324         u64 objectid;
2325         LIST_HEAD(reloc_roots);
2326         int found = 0;
2327         int ret = 0;
2328 again:
2329         root = rc->extent_root;
2330
2331         /*
2332          * this serializes us with btrfs_record_root_in_transaction,
2333          * we have to make sure nobody is in the middle of
2334          * adding their roots to the list while we are
2335          * doing this splice
2336          */
2337         mutex_lock(&root->fs_info->reloc_mutex);
2338         list_splice_init(&rc->reloc_roots, &reloc_roots);
2339         mutex_unlock(&root->fs_info->reloc_mutex);
2340
2341         while (!list_empty(&reloc_roots)) {
2342                 found = 1;
2343                 reloc_root = list_entry(reloc_roots.next,
2344                                         struct btrfs_root, root_list);
2345
2346                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2347                         root = read_fs_root(reloc_root->fs_info,
2348                                             reloc_root->root_key.offset);
2349                         BUG_ON(IS_ERR(root));
2350                         BUG_ON(root->reloc_root != reloc_root);
2351
2352                         ret = merge_reloc_root(rc, root);
2353                         if (ret) {
2354                                 __del_reloc_root(reloc_root);
2355                                 free_extent_buffer(reloc_root->node);
2356                                 free_extent_buffer(reloc_root->commit_root);
2357                                 kfree(reloc_root);
2358                                 goto out;
2359                         }
2360                 } else {
2361                         list_del_init(&reloc_root->root_list);
2362                 }
2363
2364                 /*
2365                  * we keep the old last snapshod transid in rtranid when we
2366                  * created the relocation tree.
2367                  */
2368                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2369                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2370                 objectid = reloc_root->root_key.offset;
2371
2372                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2373                 if (ret < 0) {
2374                         if (list_empty(&reloc_root->root_list))
2375                                 list_add_tail(&reloc_root->root_list,
2376                                               &reloc_roots);
2377                         goto out;
2378                 } else if (!ret) {
2379                         /*
2380                          * recover the last snapshot tranid to avoid
2381                          * the space balance break NOCOW.
2382                          */
2383                         root = read_fs_root(rc->extent_root->fs_info,
2384                                             objectid);
2385                         if (IS_ERR(root))
2386                                 continue;
2387
2388                         trans = btrfs_join_transaction(root);
2389                         BUG_ON(IS_ERR(trans));
2390
2391                         /* Check if the fs/file tree was snapshoted or not. */
2392                         if (btrfs_root_last_snapshot(&root->root_item) ==
2393                             otransid - 1)
2394                                 btrfs_set_root_last_snapshot(&root->root_item,
2395                                                              last_snap);
2396
2397                         btrfs_end_transaction(trans, root);
2398                 }
2399         }
2400
2401         if (found) {
2402                 found = 0;
2403                 goto again;
2404         }
2405 out:
2406         if (ret) {
2407                 btrfs_std_error(root->fs_info, ret);
2408                 if (!list_empty(&reloc_roots))
2409                         free_reloc_roots(&reloc_roots);
2410         }
2411
2412         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2413         return ret;
2414 }
2415
2416 static void free_block_list(struct rb_root *blocks)
2417 {
2418         struct tree_block *block;
2419         struct rb_node *rb_node;
2420         while ((rb_node = rb_first(blocks))) {
2421                 block = rb_entry(rb_node, struct tree_block, rb_node);
2422                 rb_erase(rb_node, blocks);
2423                 kfree(block);
2424         }
2425 }
2426
2427 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2428                                       struct btrfs_root *reloc_root)
2429 {
2430         struct btrfs_root *root;
2431
2432         if (reloc_root->last_trans == trans->transid)
2433                 return 0;
2434
2435         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2436         BUG_ON(IS_ERR(root));
2437         BUG_ON(root->reloc_root != reloc_root);
2438
2439         return btrfs_record_root_in_trans(trans, root);
2440 }
2441
2442 static noinline_for_stack
2443 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2444                                      struct reloc_control *rc,
2445                                      struct backref_node *node,
2446                                      struct backref_edge *edges[], int *nr)
2447 {
2448         struct backref_node *next;
2449         struct btrfs_root *root;
2450         int index = 0;
2451
2452         next = node;
2453         while (1) {
2454                 cond_resched();
2455                 next = walk_up_backref(next, edges, &index);
2456                 root = next->root;
2457                 BUG_ON(!root);
2458                 BUG_ON(!root->ref_cows);
2459
2460                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2461                         record_reloc_root_in_trans(trans, root);
2462                         break;
2463                 }
2464
2465                 btrfs_record_root_in_trans(trans, root);
2466                 root = root->reloc_root;
2467
2468                 if (next->new_bytenr != root->node->start) {
2469                         BUG_ON(next->new_bytenr);
2470                         BUG_ON(!list_empty(&next->list));
2471                         next->new_bytenr = root->node->start;
2472                         next->root = root;
2473                         list_add_tail(&next->list,
2474                                       &rc->backref_cache.changed);
2475                         __mark_block_processed(rc, next);
2476                         break;
2477                 }
2478
2479                 WARN_ON(1);
2480                 root = NULL;
2481                 next = walk_down_backref(edges, &index);
2482                 if (!next || next->level <= node->level)
2483                         break;
2484         }
2485         if (!root)
2486                 return NULL;
2487
2488         *nr = index;
2489         next = node;
2490         /* setup backref node path for btrfs_reloc_cow_block */
2491         while (1) {
2492                 rc->backref_cache.path[next->level] = next;
2493                 if (--index < 0)
2494                         break;
2495                 next = edges[index]->node[UPPER];
2496         }
2497         return root;
2498 }
2499
2500 /*
2501  * select a tree root for relocation. return NULL if the block
2502  * is reference counted. we should use do_relocation() in this
2503  * case. return a tree root pointer if the block isn't reference
2504  * counted. return -ENOENT if the block is root of reloc tree.
2505  */
2506 static noinline_for_stack
2507 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2508                                    struct backref_node *node)
2509 {
2510         struct backref_node *next;
2511         struct btrfs_root *root;
2512         struct btrfs_root *fs_root = NULL;
2513         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2514         int index = 0;
2515
2516         next = node;
2517         while (1) {
2518                 cond_resched();
2519                 next = walk_up_backref(next, edges, &index);
2520                 root = next->root;
2521                 BUG_ON(!root);
2522
2523                 /* no other choice for non-references counted tree */
2524                 if (!root->ref_cows)
2525                         return root;
2526
2527                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2528                         fs_root = root;
2529
2530                 if (next != node)
2531                         return NULL;
2532
2533                 next = walk_down_backref(edges, &index);
2534                 if (!next || next->level <= node->level)
2535                         break;
2536         }
2537
2538         if (!fs_root)
2539                 return ERR_PTR(-ENOENT);
2540         return fs_root;
2541 }
2542
2543 static noinline_for_stack
2544 u64 calcu_metadata_size(struct reloc_control *rc,
2545                         struct backref_node *node, int reserve)
2546 {
2547         struct backref_node *next = node;
2548         struct backref_edge *edge;
2549         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2550         u64 num_bytes = 0;
2551         int index = 0;
2552
2553         BUG_ON(reserve && node->processed);
2554
2555         while (next) {
2556                 cond_resched();
2557                 while (1) {
2558                         if (next->processed && (reserve || next != node))
2559                                 break;
2560
2561                         num_bytes += btrfs_level_size(rc->extent_root,
2562                                                       next->level);
2563
2564                         if (list_empty(&next->upper))
2565                                 break;
2566
2567                         edge = list_entry(next->upper.next,
2568                                           struct backref_edge, list[LOWER]);
2569                         edges[index++] = edge;
2570                         next = edge->node[UPPER];
2571                 }
2572                 next = walk_down_backref(edges, &index);
2573         }
2574         return num_bytes;
2575 }
2576
2577 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2578                                   struct reloc_control *rc,
2579                                   struct backref_node *node)
2580 {
2581         struct btrfs_root *root = rc->extent_root;
2582         u64 num_bytes;
2583         int ret;
2584
2585         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2586
2587         trans->block_rsv = rc->block_rsv;
2588         ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2589                                   BTRFS_RESERVE_FLUSH_ALL);
2590         if (ret) {
2591                 if (ret == -EAGAIN)
2592                         rc->commit_transaction = 1;
2593                 return ret;
2594         }
2595
2596         return 0;
2597 }
2598
2599 static void release_metadata_space(struct reloc_control *rc,
2600                                    struct backref_node *node)
2601 {
2602         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2603         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2604 }
2605
2606 /*
2607  * relocate a block tree, and then update pointers in upper level
2608  * blocks that reference the block to point to the new location.
2609  *
2610  * if called by link_to_upper, the block has already been relocated.
2611  * in that case this function just updates pointers.
2612  */
2613 static int do_relocation(struct btrfs_trans_handle *trans,
2614                          struct reloc_control *rc,
2615                          struct backref_node *node,
2616                          struct btrfs_key *key,
2617                          struct btrfs_path *path, int lowest)
2618 {
2619         struct backref_node *upper;
2620         struct backref_edge *edge;
2621         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2622         struct btrfs_root *root;
2623         struct extent_buffer *eb;
2624         u32 blocksize;
2625         u64 bytenr;
2626         u64 generation;
2627         int nr;
2628         int slot;
2629         int ret;
2630         int err = 0;
2631
2632         BUG_ON(lowest && node->eb);
2633
2634         path->lowest_level = node->level + 1;
2635         rc->backref_cache.path[node->level] = node;
2636         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2637                 cond_resched();
2638
2639                 upper = edge->node[UPPER];
2640                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2641                 BUG_ON(!root);
2642
2643                 if (upper->eb && !upper->locked) {
2644                         if (!lowest) {
2645                                 ret = btrfs_bin_search(upper->eb, key,
2646                                                        upper->level, &slot);
2647                                 BUG_ON(ret);
2648                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2649                                 if (node->eb->start == bytenr)
2650                                         goto next;
2651                         }
2652                         drop_node_buffer(upper);
2653                 }
2654
2655                 if (!upper->eb) {
2656                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2657                         if (ret < 0) {
2658                                 err = ret;
2659                                 break;
2660                         }
2661                         BUG_ON(ret > 0);
2662
2663                         if (!upper->eb) {
2664                                 upper->eb = path->nodes[upper->level];
2665                                 path->nodes[upper->level] = NULL;
2666                         } else {
2667                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2668                         }
2669
2670                         upper->locked = 1;
2671                         path->locks[upper->level] = 0;
2672
2673                         slot = path->slots[upper->level];
2674                         btrfs_release_path(path);
2675                 } else {
2676                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2677                                                &slot);
2678                         BUG_ON(ret);
2679                 }
2680
2681                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2682                 if (lowest) {
2683                         BUG_ON(bytenr != node->bytenr);
2684                 } else {
2685                         if (node->eb->start == bytenr)
2686                                 goto next;
2687                 }
2688
2689                 blocksize = btrfs_level_size(root, node->level);
2690                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2691                 eb = read_tree_block(root, bytenr, blocksize, generation);
2692                 if (!eb || !extent_buffer_uptodate(eb)) {
2693                         free_extent_buffer(eb);
2694                         err = -EIO;
2695                         goto next;
2696                 }
2697                 btrfs_tree_lock(eb);
2698                 btrfs_set_lock_blocking(eb);
2699
2700                 if (!node->eb) {
2701                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2702                                               slot, &eb);
2703                         btrfs_tree_unlock(eb);
2704                         free_extent_buffer(eb);
2705                         if (ret < 0) {
2706                                 err = ret;
2707                                 goto next;
2708                         }
2709                         BUG_ON(node->eb != eb);
2710                 } else {
2711                         btrfs_set_node_blockptr(upper->eb, slot,
2712                                                 node->eb->start);
2713                         btrfs_set_node_ptr_generation(upper->eb, slot,
2714                                                       trans->transid);
2715                         btrfs_mark_buffer_dirty(upper->eb);
2716
2717                         ret = btrfs_inc_extent_ref(trans, root,
2718                                                 node->eb->start, blocksize,
2719                                                 upper->eb->start,
2720                                                 btrfs_header_owner(upper->eb),
2721                                                 node->level, 0, 1);
2722                         BUG_ON(ret);
2723
2724                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2725                         BUG_ON(ret);
2726                 }
2727 next:
2728                 if (!upper->pending)
2729                         drop_node_buffer(upper);
2730                 else
2731                         unlock_node_buffer(upper);
2732                 if (err)
2733                         break;
2734         }
2735
2736         if (!err && node->pending) {
2737                 drop_node_buffer(node);
2738                 list_move_tail(&node->list, &rc->backref_cache.changed);
2739                 node->pending = 0;
2740         }
2741
2742         path->lowest_level = 0;
2743         BUG_ON(err == -ENOSPC);
2744         return err;
2745 }
2746
2747 static int link_to_upper(struct btrfs_trans_handle *trans,
2748                          struct reloc_control *rc,
2749                          struct backref_node *node,
2750                          struct btrfs_path *path)
2751 {
2752         struct btrfs_key key;
2753
2754         btrfs_node_key_to_cpu(node->eb, &key, 0);
2755         return do_relocation(trans, rc, node, &key, path, 0);
2756 }
2757
2758 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2759                                 struct reloc_control *rc,
2760                                 struct btrfs_path *path, int err)
2761 {
2762         LIST_HEAD(list);
2763         struct backref_cache *cache = &rc->backref_cache;
2764         struct backref_node *node;
2765         int level;
2766         int ret;
2767
2768         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2769                 while (!list_empty(&cache->pending[level])) {
2770                         node = list_entry(cache->pending[level].next,
2771                                           struct backref_node, list);
2772                         list_move_tail(&node->list, &list);
2773                         BUG_ON(!node->pending);
2774
2775                         if (!err) {
2776                                 ret = link_to_upper(trans, rc, node, path);
2777                                 if (ret < 0)
2778                                         err = ret;
2779                         }
2780                 }
2781                 list_splice_init(&list, &cache->pending[level]);
2782         }
2783         return err;
2784 }
2785
2786 static void mark_block_processed(struct reloc_control *rc,
2787                                  u64 bytenr, u32 blocksize)
2788 {
2789         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2790                         EXTENT_DIRTY, GFP_NOFS);
2791 }
2792
2793 static void __mark_block_processed(struct reloc_control *rc,
2794                                    struct backref_node *node)
2795 {
2796         u32 blocksize;
2797         if (node->level == 0 ||
2798             in_block_group(node->bytenr, rc->block_group)) {
2799                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2800                 mark_block_processed(rc, node->bytenr, blocksize);
2801         }
2802         node->processed = 1;
2803 }
2804
2805 /*
2806  * mark a block and all blocks directly/indirectly reference the block
2807  * as processed.
2808  */
2809 static void update_processed_blocks(struct reloc_control *rc,
2810                                     struct backref_node *node)
2811 {
2812         struct backref_node *next = node;
2813         struct backref_edge *edge;
2814         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2815         int index = 0;
2816
2817         while (next) {
2818                 cond_resched();
2819                 while (1) {
2820                         if (next->processed)
2821                                 break;
2822
2823                         __mark_block_processed(rc, next);
2824
2825                         if (list_empty(&next->upper))
2826                                 break;
2827
2828                         edge = list_entry(next->upper.next,
2829                                           struct backref_edge, list[LOWER]);
2830                         edges[index++] = edge;
2831                         next = edge->node[UPPER];
2832                 }
2833                 next = walk_down_backref(edges, &index);
2834         }
2835 }
2836
2837 static int tree_block_processed(u64 bytenr, u32 blocksize,
2838                                 struct reloc_control *rc)
2839 {
2840         if (test_range_bit(&rc->processed_blocks, bytenr,
2841                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2842                 return 1;
2843         return 0;
2844 }
2845
2846 static int get_tree_block_key(struct reloc_control *rc,
2847                               struct tree_block *block)
2848 {
2849         struct extent_buffer *eb;
2850
2851         BUG_ON(block->key_ready);
2852         eb = read_tree_block(rc->extent_root, block->bytenr,
2853                              block->key.objectid, block->key.offset);
2854         if (!eb || !extent_buffer_uptodate(eb)) {
2855                 free_extent_buffer(eb);
2856                 return -EIO;
2857         }
2858         WARN_ON(btrfs_header_level(eb) != block->level);
2859         if (block->level == 0)
2860                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2861         else
2862                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2863         free_extent_buffer(eb);
2864         block->key_ready = 1;
2865         return 0;
2866 }
2867
2868 static int reada_tree_block(struct reloc_control *rc,
2869                             struct tree_block *block)
2870 {
2871         BUG_ON(block->key_ready);
2872         if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2873                 readahead_tree_block(rc->extent_root, block->bytenr,
2874                                      block->key.objectid,
2875                                      rc->extent_root->leafsize);
2876         else
2877                 readahead_tree_block(rc->extent_root, block->bytenr,
2878                                      block->key.objectid, block->key.offset);
2879         return 0;
2880 }
2881
2882 /*
2883  * helper function to relocate a tree block
2884  */
2885 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2886                                 struct reloc_control *rc,
2887                                 struct backref_node *node,
2888                                 struct btrfs_key *key,
2889                                 struct btrfs_path *path)
2890 {
2891         struct btrfs_root *root;
2892         int release = 0;
2893         int ret = 0;
2894
2895         if (!node)
2896                 return 0;
2897
2898         BUG_ON(node->processed);
2899         root = select_one_root(trans, node);
2900         if (root == ERR_PTR(-ENOENT)) {
2901                 update_processed_blocks(rc, node);
2902                 goto out;
2903         }
2904
2905         if (!root || root->ref_cows) {
2906                 ret = reserve_metadata_space(trans, rc, node);
2907                 if (ret)
2908                         goto out;
2909                 release = 1;
2910         }
2911
2912         if (root) {
2913                 if (root->ref_cows) {
2914                         BUG_ON(node->new_bytenr);
2915                         BUG_ON(!list_empty(&node->list));
2916                         btrfs_record_root_in_trans(trans, root);
2917                         root = root->reloc_root;
2918                         node->new_bytenr = root->node->start;
2919                         node->root = root;
2920                         list_add_tail(&node->list, &rc->backref_cache.changed);
2921                 } else {
2922                         path->lowest_level = node->level;
2923                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2924                         btrfs_release_path(path);
2925                         if (ret > 0)
2926                                 ret = 0;
2927                 }
2928                 if (!ret)
2929                         update_processed_blocks(rc, node);
2930         } else {
2931                 ret = do_relocation(trans, rc, node, key, path, 1);
2932         }
2933 out:
2934         if (ret || node->level == 0 || node->cowonly) {
2935                 if (release)
2936                         release_metadata_space(rc, node);
2937                 remove_backref_node(&rc->backref_cache, node);
2938         }
2939         return ret;
2940 }
2941
2942 /*
2943  * relocate a list of blocks
2944  */
2945 static noinline_for_stack
2946 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2947                          struct reloc_control *rc, struct rb_root *blocks)
2948 {
2949         struct backref_node *node;
2950         struct btrfs_path *path;
2951         struct tree_block *block;
2952         struct rb_node *rb_node;
2953         int ret;
2954         int err = 0;
2955
2956         path = btrfs_alloc_path();
2957         if (!path) {
2958                 err = -ENOMEM;
2959                 goto out_free_blocks;
2960         }
2961
2962         rb_node = rb_first(blocks);
2963         while (rb_node) {
2964                 block = rb_entry(rb_node, struct tree_block, rb_node);
2965                 if (!block->key_ready)
2966                         reada_tree_block(rc, block);
2967                 rb_node = rb_next(rb_node);
2968         }
2969
2970         rb_node = rb_first(blocks);
2971         while (rb_node) {
2972                 block = rb_entry(rb_node, struct tree_block, rb_node);
2973                 if (!block->key_ready) {
2974                         err = get_tree_block_key(rc, block);
2975                         if (err)
2976                                 goto out_free_path;
2977                 }
2978                 rb_node = rb_next(rb_node);
2979         }
2980
2981         rb_node = rb_first(blocks);
2982         while (rb_node) {
2983                 block = rb_entry(rb_node, struct tree_block, rb_node);
2984
2985                 node = build_backref_tree(rc, &block->key,
2986                                           block->level, block->bytenr);
2987                 if (IS_ERR(node)) {
2988                         err = PTR_ERR(node);
2989                         goto out;
2990                 }
2991
2992                 ret = relocate_tree_block(trans, rc, node, &block->key,
2993                                           path);
2994                 if (ret < 0) {
2995                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2996                                 err = ret;
2997                         goto out;
2998                 }
2999                 rb_node = rb_next(rb_node);
3000         }
3001 out:
3002         err = finish_pending_nodes(trans, rc, path, err);
3003
3004 out_free_path:
3005         btrfs_free_path(path);
3006 out_free_blocks:
3007         free_block_list(blocks);
3008         return err;
3009 }
3010
3011 static noinline_for_stack
3012 int prealloc_file_extent_cluster(struct inode *inode,
3013                                  struct file_extent_cluster *cluster)
3014 {
3015         u64 alloc_hint = 0;
3016         u64 start;
3017         u64 end;
3018         u64 offset = BTRFS_I(inode)->index_cnt;
3019         u64 num_bytes;
3020         int nr = 0;
3021         int ret = 0;
3022
3023         BUG_ON(cluster->start != cluster->boundary[0]);
3024         mutex_lock(&inode->i_mutex);
3025
3026         ret = btrfs_check_data_free_space(inode, cluster->end +
3027                                           1 - cluster->start);
3028         if (ret)
3029                 goto out;
3030
3031         while (nr < cluster->nr) {
3032                 start = cluster->boundary[nr] - offset;
3033                 if (nr + 1 < cluster->nr)
3034                         end = cluster->boundary[nr + 1] - 1 - offset;
3035                 else
3036                         end = cluster->end - offset;
3037
3038                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3039                 num_bytes = end + 1 - start;
3040                 ret = btrfs_prealloc_file_range(inode, 0, start,
3041                                                 num_bytes, num_bytes,
3042                                                 end + 1, &alloc_hint);
3043                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3044                 if (ret)
3045                         break;
3046                 nr++;
3047         }
3048         btrfs_free_reserved_data_space(inode, cluster->end +
3049                                        1 - cluster->start);
3050 out:
3051         mutex_unlock(&inode->i_mutex);
3052         return ret;
3053 }
3054
3055 static noinline_for_stack
3056 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3057                          u64 block_start)
3058 {
3059         struct btrfs_root *root = BTRFS_I(inode)->root;
3060         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3061         struct extent_map *em;
3062         int ret = 0;
3063
3064         em = alloc_extent_map();
3065         if (!em)
3066                 return -ENOMEM;
3067
3068         em->start = start;
3069         em->len = end + 1 - start;
3070         em->block_len = em->len;
3071         em->block_start = block_start;
3072         em->bdev = root->fs_info->fs_devices->latest_bdev;
3073         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3074
3075         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3076         while (1) {
3077                 write_lock(&em_tree->lock);
3078                 ret = add_extent_mapping(em_tree, em, 0);
3079                 write_unlock(&em_tree->lock);
3080                 if (ret != -EEXIST) {
3081                         free_extent_map(em);
3082                         break;
3083                 }
3084                 btrfs_drop_extent_cache(inode, start, end, 0);
3085         }
3086         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3087         return ret;
3088 }
3089
3090 static int relocate_file_extent_cluster(struct inode *inode,
3091                                         struct file_extent_cluster *cluster)
3092 {
3093         u64 page_start;
3094         u64 page_end;
3095         u64 offset = BTRFS_I(inode)->index_cnt;
3096         unsigned long index;
3097         unsigned long last_index;
3098         struct page *page;
3099         struct file_ra_state *ra;
3100         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3101         int nr = 0;
3102         int ret = 0;
3103
3104         if (!cluster->nr)
3105                 return 0;
3106
3107         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3108         if (!ra)
3109                 return -ENOMEM;
3110
3111         ret = prealloc_file_extent_cluster(inode, cluster);
3112         if (ret)
3113                 goto out;
3114
3115         file_ra_state_init(ra, inode->i_mapping);
3116
3117         ret = setup_extent_mapping(inode, cluster->start - offset,
3118                                    cluster->end - offset, cluster->start);
3119         if (ret)
3120                 goto out;
3121
3122         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3123         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3124         while (index <= last_index) {
3125                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3126                 if (ret)
3127                         goto out;
3128
3129                 page = find_lock_page(inode->i_mapping, index);
3130                 if (!page) {
3131                         page_cache_sync_readahead(inode->i_mapping,
3132                                                   ra, NULL, index,
3133                                                   last_index + 1 - index);
3134                         page = find_or_create_page(inode->i_mapping, index,
3135                                                    mask);
3136                         if (!page) {
3137                                 btrfs_delalloc_release_metadata(inode,
3138                                                         PAGE_CACHE_SIZE);
3139                                 ret = -ENOMEM;
3140                                 goto out;
3141                         }
3142                 }
3143
3144                 if (PageReadahead(page)) {
3145                         page_cache_async_readahead(inode->i_mapping,
3146                                                    ra, NULL, page, index,
3147                                                    last_index + 1 - index);
3148                 }
3149
3150                 if (!PageUptodate(page)) {
3151                         btrfs_readpage(NULL, page);
3152                         lock_page(page);
3153                         if (!PageUptodate(page)) {
3154                                 unlock_page(page);
3155                                 page_cache_release(page);
3156                                 btrfs_delalloc_release_metadata(inode,
3157                                                         PAGE_CACHE_SIZE);
3158                                 ret = -EIO;
3159                                 goto out;
3160                         }
3161                 }
3162
3163                 page_start = page_offset(page);
3164                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3165
3166                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3167
3168                 set_page_extent_mapped(page);
3169
3170                 if (nr < cluster->nr &&
3171                     page_start + offset == cluster->boundary[nr]) {
3172                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3173                                         page_start, page_end,
3174                                         EXTENT_BOUNDARY, GFP_NOFS);
3175                         nr++;
3176                 }
3177
3178                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3179                 set_page_dirty(page);
3180
3181                 unlock_extent(&BTRFS_I(inode)->io_tree,
3182                               page_start, page_end);
3183                 unlock_page(page);
3184                 page_cache_release(page);
3185
3186                 index++;
3187                 balance_dirty_pages_ratelimited(inode->i_mapping);
3188                 btrfs_throttle(BTRFS_I(inode)->root);
3189         }
3190         WARN_ON(nr != cluster->nr);
3191 out:
3192         kfree(ra);
3193         return ret;
3194 }
3195
3196 static noinline_for_stack
3197 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3198                          struct file_extent_cluster *cluster)
3199 {
3200         int ret;
3201
3202         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3203                 ret = relocate_file_extent_cluster(inode, cluster);
3204                 if (ret)
3205                         return ret;
3206                 cluster->nr = 0;
3207         }
3208
3209         if (!cluster->nr)
3210                 cluster->start = extent_key->objectid;
3211         else
3212                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3213         cluster->end = extent_key->objectid + extent_key->offset - 1;
3214         cluster->boundary[cluster->nr] = extent_key->objectid;
3215         cluster->nr++;
3216
3217         if (cluster->nr >= MAX_EXTENTS) {
3218                 ret = relocate_file_extent_cluster(inode, cluster);
3219                 if (ret)
3220                         return ret;
3221                 cluster->nr = 0;
3222         }
3223         return 0;
3224 }
3225
3226 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3227 static int get_ref_objectid_v0(struct reloc_control *rc,
3228                                struct btrfs_path *path,
3229                                struct btrfs_key *extent_key,
3230                                u64 *ref_objectid, int *path_change)
3231 {
3232         struct btrfs_key key;
3233         struct extent_buffer *leaf;
3234         struct btrfs_extent_ref_v0 *ref0;
3235         int ret;
3236         int slot;
3237
3238         leaf = path->nodes[0];
3239         slot = path->slots[0];
3240         while (1) {
3241                 if (slot >= btrfs_header_nritems(leaf)) {
3242                         ret = btrfs_next_leaf(rc->extent_root, path);
3243                         if (ret < 0)
3244                                 return ret;
3245                         BUG_ON(ret > 0);
3246                         leaf = path->nodes[0];
3247                         slot = path->slots[0];
3248                         if (path_change)
3249                                 *path_change = 1;
3250                 }
3251                 btrfs_item_key_to_cpu(leaf, &key, slot);
3252                 if (key.objectid != extent_key->objectid)
3253                         return -ENOENT;
3254
3255                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3256                         slot++;
3257                         continue;
3258                 }
3259                 ref0 = btrfs_item_ptr(leaf, slot,
3260                                 struct btrfs_extent_ref_v0);
3261                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3262                 break;
3263         }
3264         return 0;
3265 }
3266 #endif
3267
3268 /*
3269  * helper to add a tree block to the list.
3270  * the major work is getting the generation and level of the block
3271  */
3272 static int add_tree_block(struct reloc_control *rc,
3273                           struct btrfs_key *extent_key,
3274                           struct btrfs_path *path,
3275                           struct rb_root *blocks)
3276 {
3277         struct extent_buffer *eb;
3278         struct btrfs_extent_item *ei;
3279         struct btrfs_tree_block_info *bi;
3280         struct tree_block *block;
3281         struct rb_node *rb_node;
3282         u32 item_size;
3283         int level = -1;
3284         u64 generation;
3285
3286         eb =  path->nodes[0];
3287         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3288
3289         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3290             item_size >= sizeof(*ei) + sizeof(*bi)) {
3291                 ei = btrfs_item_ptr(eb, path->slots[0],
3292                                 struct btrfs_extent_item);
3293                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3294                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3295                         level = btrfs_tree_block_level(eb, bi);
3296                 } else {
3297                         level = (int)extent_key->offset;
3298                 }
3299                 generation = btrfs_extent_generation(eb, ei);
3300         } else {
3301 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3302                 u64 ref_owner;
3303                 int ret;
3304
3305                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3306                 ret = get_ref_objectid_v0(rc, path, extent_key,
3307                                           &ref_owner, NULL);
3308                 if (ret < 0)
3309                         return ret;
3310                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3311                 level = (int)ref_owner;
3312                 /* FIXME: get real generation */
3313                 generation = 0;
3314 #else
3315                 BUG();
3316 #endif
3317         }
3318
3319         btrfs_release_path(path);
3320
3321         BUG_ON(level == -1);
3322
3323         block = kmalloc(sizeof(*block), GFP_NOFS);
3324         if (!block)
3325                 return -ENOMEM;
3326
3327         block->bytenr = extent_key->objectid;
3328         block->key.objectid = rc->extent_root->leafsize;
3329         block->key.offset = generation;
3330         block->level = level;
3331         block->key_ready = 0;
3332
3333         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3334         if (rb_node)
3335                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3336
3337         return 0;
3338 }
3339
3340 /*
3341  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3342  */
3343 static int __add_tree_block(struct reloc_control *rc,
3344                             u64 bytenr, u32 blocksize,
3345                             struct rb_root *blocks)
3346 {
3347         struct btrfs_path *path;
3348         struct btrfs_key key;
3349         int ret;
3350         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3351                                         SKINNY_METADATA);
3352
3353         if (tree_block_processed(bytenr, blocksize, rc))
3354                 return 0;
3355
3356         if (tree_search(blocks, bytenr))
3357                 return 0;
3358
3359         path = btrfs_alloc_path();
3360         if (!path)
3361                 return -ENOMEM;
3362 again:
3363         key.objectid = bytenr;
3364         if (skinny) {
3365                 key.type = BTRFS_METADATA_ITEM_KEY;
3366                 key.offset = (u64)-1;
3367         } else {
3368                 key.type = BTRFS_EXTENT_ITEM_KEY;
3369                 key.offset = blocksize;
3370         }
3371
3372         path->search_commit_root = 1;
3373         path->skip_locking = 1;
3374         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3375         if (ret < 0)
3376                 goto out;
3377
3378         if (ret > 0 && skinny) {
3379                 if (path->slots[0]) {
3380                         path->slots[0]--;
3381                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3382                                               path->slots[0]);
3383                         if (key.objectid == bytenr &&
3384                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3385                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3386                               key.offset == blocksize)))
3387                                 ret = 0;
3388                 }
3389
3390                 if (ret) {
3391                         skinny = false;
3392                         btrfs_release_path(path);
3393                         goto again;
3394                 }
3395         }
3396         BUG_ON(ret);
3397
3398         ret = add_tree_block(rc, &key, path, blocks);
3399 out:
3400         btrfs_free_path(path);
3401         return ret;
3402 }
3403
3404 /*
3405  * helper to check if the block use full backrefs for pointers in it
3406  */
3407 static int block_use_full_backref(struct reloc_control *rc,
3408                                   struct extent_buffer *eb)
3409 {
3410         u64 flags;
3411         int ret;
3412
3413         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3414             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3415                 return 1;
3416
3417         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3418                                        eb->start, btrfs_header_level(eb), 1,
3419                                        NULL, &flags);
3420         BUG_ON(ret);
3421
3422         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3423                 ret = 1;
3424         else
3425                 ret = 0;
3426         return ret;
3427 }
3428
3429 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3430                                     struct inode *inode, u64 ino)
3431 {
3432         struct btrfs_key key;
3433         struct btrfs_root *root = fs_info->tree_root;
3434         struct btrfs_trans_handle *trans;
3435         int ret = 0;
3436
3437         if (inode)
3438                 goto truncate;
3439
3440         key.objectid = ino;
3441         key.type = BTRFS_INODE_ITEM_KEY;
3442         key.offset = 0;
3443
3444         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3445         if (IS_ERR(inode) || is_bad_inode(inode)) {
3446                 if (!IS_ERR(inode))
3447                         iput(inode);
3448                 return -ENOENT;
3449         }
3450
3451 truncate:
3452         ret = btrfs_check_trunc_cache_free_space(root,
3453                                                  &fs_info->global_block_rsv);
3454         if (ret)
3455                 goto out;
3456
3457         trans = btrfs_join_transaction(root);
3458         if (IS_ERR(trans)) {
3459                 ret = PTR_ERR(trans);
3460                 goto out;
3461         }
3462
3463         ret = btrfs_truncate_free_space_cache(root, trans, inode);
3464
3465         btrfs_end_transaction(trans, root);
3466         btrfs_btree_balance_dirty(root);
3467 out:
3468         iput(inode);
3469         return ret;
3470 }
3471
3472 /*
3473  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3474  * this function scans fs tree to find blocks reference the data extent
3475  */
3476 static int find_data_references(struct reloc_control *rc,
3477                                 struct btrfs_key *extent_key,
3478                                 struct extent_buffer *leaf,
3479                                 struct btrfs_extent_data_ref *ref,
3480                                 struct rb_root *blocks)
3481 {
3482         struct btrfs_path *path;
3483         struct tree_block *block;
3484         struct btrfs_root *root;
3485         struct btrfs_file_extent_item *fi;
3486         struct rb_node *rb_node;
3487         struct btrfs_key key;
3488         u64 ref_root;
3489         u64 ref_objectid;
3490         u64 ref_offset;
3491         u32 ref_count;
3492         u32 nritems;
3493         int err = 0;
3494         int added = 0;
3495         int counted;
3496         int ret;
3497
3498         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3499         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3500         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3501         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3502
3503         /*
3504          * This is an extent belonging to the free space cache, lets just delete
3505          * it and redo the search.
3506          */
3507         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3508                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3509                                                NULL, ref_objectid);
3510                 if (ret != -ENOENT)
3511                         return ret;
3512                 ret = 0;
3513         }
3514
3515         path = btrfs_alloc_path();
3516         if (!path)
3517                 return -ENOMEM;
3518         path->reada = 1;
3519
3520         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3521         if (IS_ERR(root)) {
3522                 err = PTR_ERR(root);
3523                 goto out;
3524         }
3525
3526         key.objectid = ref_objectid;
3527         key.type = BTRFS_EXTENT_DATA_KEY;
3528         if (ref_offset > ((u64)-1 << 32))
3529                 key.offset = 0;
3530         else
3531                 key.offset = ref_offset;
3532
3533         path->search_commit_root = 1;
3534         path->skip_locking = 1;
3535         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3536         if (ret < 0) {
3537                 err = ret;
3538                 goto out;
3539         }
3540
3541         leaf = path->nodes[0];
3542         nritems = btrfs_header_nritems(leaf);
3543         /*
3544          * the references in tree blocks that use full backrefs
3545          * are not counted in
3546          */
3547         if (block_use_full_backref(rc, leaf))
3548                 counted = 0;
3549         else
3550                 counted = 1;
3551         rb_node = tree_search(blocks, leaf->start);
3552         if (rb_node) {
3553                 if (counted)
3554                         added = 1;
3555                 else
3556                         path->slots[0] = nritems;
3557         }
3558
3559         while (ref_count > 0) {
3560                 while (path->slots[0] >= nritems) {
3561                         ret = btrfs_next_leaf(root, path);
3562                         if (ret < 0) {
3563                                 err = ret;
3564                                 goto out;
3565                         }
3566                         if (WARN_ON(ret > 0))
3567                                 goto out;
3568
3569                         leaf = path->nodes[0];
3570                         nritems = btrfs_header_nritems(leaf);
3571                         added = 0;
3572
3573                         if (block_use_full_backref(rc, leaf))
3574                                 counted = 0;
3575                         else
3576                                 counted = 1;
3577                         rb_node = tree_search(blocks, leaf->start);
3578                         if (rb_node) {
3579                                 if (counted)
3580                                         added = 1;
3581                                 else
3582                                         path->slots[0] = nritems;
3583                         }
3584                 }
3585
3586                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3587                 if (WARN_ON(key.objectid != ref_objectid ||
3588                     key.type != BTRFS_EXTENT_DATA_KEY))
3589                         break;
3590
3591                 fi = btrfs_item_ptr(leaf, path->slots[0],
3592                                     struct btrfs_file_extent_item);
3593
3594                 if (btrfs_file_extent_type(leaf, fi) ==
3595                     BTRFS_FILE_EXTENT_INLINE)
3596                         goto next;
3597
3598                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3599                     extent_key->objectid)
3600                         goto next;
3601
3602                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3603                 if (key.offset != ref_offset)
3604                         goto next;
3605
3606                 if (counted)
3607                         ref_count--;
3608                 if (added)
3609                         goto next;
3610
3611                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3612                         block = kmalloc(sizeof(*block), GFP_NOFS);
3613                         if (!block) {
3614                                 err = -ENOMEM;
3615                                 break;
3616                         }
3617                         block->bytenr = leaf->start;
3618                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3619                         block->level = 0;
3620                         block->key_ready = 1;
3621                         rb_node = tree_insert(blocks, block->bytenr,
3622                                               &block->rb_node);
3623                         if (rb_node)
3624                                 backref_tree_panic(rb_node, -EEXIST,
3625                                                    block->bytenr);
3626                 }
3627                 if (counted)
3628                         added = 1;
3629                 else
3630                         path->slots[0] = nritems;
3631 next:
3632                 path->slots[0]++;
3633
3634         }
3635 out:
3636         btrfs_free_path(path);
3637         return err;
3638 }
3639
3640 /*
3641  * helper to find all tree blocks that reference a given data extent
3642  */
3643 static noinline_for_stack
3644 int add_data_references(struct reloc_control *rc,
3645                         struct btrfs_key *extent_key,
3646                         struct btrfs_path *path,
3647                         struct rb_root *blocks)
3648 {
3649         struct btrfs_key key;
3650         struct extent_buffer *eb;
3651         struct btrfs_extent_data_ref *dref;
3652         struct btrfs_extent_inline_ref *iref;
3653         unsigned long ptr;
3654         unsigned long end;
3655         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3656         int ret = 0;
3657         int err = 0;
3658
3659         eb = path->nodes[0];
3660         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3661         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3662 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3663         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3664                 ptr = end;
3665         else
3666 #endif
3667                 ptr += sizeof(struct btrfs_extent_item);
3668
3669         while (ptr < end) {
3670                 iref = (struct btrfs_extent_inline_ref *)ptr;
3671                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3672                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3673                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3674                         ret = __add_tree_block(rc, key.offset, blocksize,
3675                                                blocks);
3676                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3677                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3678                         ret = find_data_references(rc, extent_key,
3679                                                    eb, dref, blocks);
3680                 } else {
3681                         BUG();
3682                 }
3683                 if (ret) {
3684                         err = ret;
3685                         goto out;
3686                 }
3687                 ptr += btrfs_extent_inline_ref_size(key.type);
3688         }
3689         WARN_ON(ptr > end);
3690
3691         while (1) {
3692                 cond_resched();
3693                 eb = path->nodes[0];
3694                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3695                         ret = btrfs_next_leaf(rc->extent_root, path);
3696                         if (ret < 0) {
3697                                 err = ret;
3698                                 break;
3699                         }
3700                         if (ret > 0)
3701                                 break;
3702                         eb = path->nodes[0];
3703                 }
3704
3705                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3706                 if (key.objectid != extent_key->objectid)
3707                         break;
3708
3709 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3710                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3711                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3712 #else
3713                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3714                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3715 #endif
3716                         ret = __add_tree_block(rc, key.offset, blocksize,
3717                                                blocks);
3718                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3719                         dref = btrfs_item_ptr(eb, path->slots[0],
3720                                               struct btrfs_extent_data_ref);
3721                         ret = find_data_references(rc, extent_key,
3722                                                    eb, dref, blocks);
3723                 } else {
3724                         ret = 0;
3725                 }
3726                 if (ret) {
3727                         err = ret;
3728                         break;
3729                 }
3730                 path->slots[0]++;
3731         }
3732 out:
3733         btrfs_release_path(path);
3734         if (err)
3735                 free_block_list(blocks);
3736         return err;
3737 }
3738
3739 /*
3740  * helper to find next unprocessed extent
3741  */
3742 static noinline_for_stack
3743 int find_next_extent(struct btrfs_trans_handle *trans,
3744                      struct reloc_control *rc, struct btrfs_path *path,
3745                      struct btrfs_key *extent_key)
3746 {
3747         struct btrfs_key key;
3748         struct extent_buffer *leaf;
3749         u64 start, end, last;
3750         int ret;
3751
3752         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3753         while (1) {
3754                 cond_resched();
3755                 if (rc->search_start >= last) {
3756                         ret = 1;
3757                         break;
3758                 }
3759
3760                 key.objectid = rc->search_start;
3761                 key.type = BTRFS_EXTENT_ITEM_KEY;
3762                 key.offset = 0;
3763
3764                 path->search_commit_root = 1;
3765                 path->skip_locking = 1;
3766                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3767                                         0, 0);
3768                 if (ret < 0)
3769                         break;
3770 next:
3771                 leaf = path->nodes[0];
3772                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3773                         ret = btrfs_next_leaf(rc->extent_root, path);
3774                         if (ret != 0)
3775                                 break;
3776                         leaf = path->nodes[0];
3777                 }
3778
3779                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3780                 if (key.objectid >= last) {
3781                         ret = 1;
3782                         break;
3783                 }
3784
3785                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3786                     key.type != BTRFS_METADATA_ITEM_KEY) {
3787                         path->slots[0]++;
3788                         goto next;
3789                 }
3790
3791                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3792                     key.objectid + key.offset <= rc->search_start) {
3793                         path->slots[0]++;
3794                         goto next;
3795                 }
3796
3797                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3798                     key.objectid + rc->extent_root->leafsize <=
3799                     rc->search_start) {
3800                         path->slots[0]++;
3801                         goto next;
3802                 }
3803
3804                 ret = find_first_extent_bit(&rc->processed_blocks,
3805                                             key.objectid, &start, &end,
3806                                             EXTENT_DIRTY, NULL);
3807
3808                 if (ret == 0 && start <= key.objectid) {
3809                         btrfs_release_path(path);
3810                         rc->search_start = end + 1;
3811                 } else {
3812                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3813                                 rc->search_start = key.objectid + key.offset;
3814                         else
3815                                 rc->search_start = key.objectid +
3816                                         rc->extent_root->leafsize;
3817                         memcpy(extent_key, &key, sizeof(key));
3818                         return 0;
3819                 }
3820         }
3821         btrfs_release_path(path);
3822         return ret;
3823 }
3824
3825 static void set_reloc_control(struct reloc_control *rc)
3826 {
3827         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3828
3829         mutex_lock(&fs_info->reloc_mutex);
3830         fs_info->reloc_ctl = rc;
3831         mutex_unlock(&fs_info->reloc_mutex);
3832 }
3833
3834 static void unset_reloc_control(struct reloc_control *rc)
3835 {
3836         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3837
3838         mutex_lock(&fs_info->reloc_mutex);
3839         fs_info->reloc_ctl = NULL;
3840         mutex_unlock(&fs_info->reloc_mutex);
3841 }
3842
3843 static int check_extent_flags(u64 flags)
3844 {
3845         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3846             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3847                 return 1;
3848         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3849             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3850                 return 1;
3851         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3852             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3853                 return 1;
3854         return 0;
3855 }
3856
3857 static noinline_for_stack
3858 int prepare_to_relocate(struct reloc_control *rc)
3859 {
3860         struct btrfs_trans_handle *trans;
3861         int ret;
3862
3863         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3864                                               BTRFS_BLOCK_RSV_TEMP);
3865         if (!rc->block_rsv)
3866                 return -ENOMEM;
3867
3868         /*
3869          * reserve some space for creating reloc trees.
3870          * btrfs_init_reloc_root will use them when there
3871          * is no reservation in transaction handle.
3872          */
3873         ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3874                                   rc->extent_root->nodesize * 256,
3875                                   BTRFS_RESERVE_FLUSH_ALL);
3876         if (ret)
3877                 return ret;
3878
3879         memset(&rc->cluster, 0, sizeof(rc->cluster));
3880         rc->search_start = rc->block_group->key.objectid;
3881         rc->extents_found = 0;
3882         rc->nodes_relocated = 0;
3883         rc->merging_rsv_size = 0;
3884
3885         rc->create_reloc_tree = 1;
3886         set_reloc_control(rc);
3887
3888         trans = btrfs_join_transaction(rc->extent_root);
3889         if (IS_ERR(trans)) {
3890                 unset_reloc_control(rc);
3891                 /*
3892                  * extent tree is not a ref_cow tree and has no reloc_root to
3893                  * cleanup.  And callers are responsible to free the above
3894                  * block rsv.
3895                  */
3896                 return PTR_ERR(trans);
3897         }
3898         btrfs_commit_transaction(trans, rc->extent_root);
3899         return 0;
3900 }
3901
3902 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3903 {
3904         struct rb_root blocks = RB_ROOT;
3905         struct btrfs_key key;
3906         struct btrfs_trans_handle *trans = NULL;
3907         struct btrfs_path *path;
3908         struct btrfs_extent_item *ei;
3909         u64 flags;
3910         u32 item_size;
3911         int ret;
3912         int err = 0;
3913         int progress = 0;
3914
3915         path = btrfs_alloc_path();
3916         if (!path)
3917                 return -ENOMEM;
3918         path->reada = 1;
3919
3920         ret = prepare_to_relocate(rc);
3921         if (ret) {
3922                 err = ret;
3923                 goto out_free;
3924         }
3925
3926         while (1) {
3927                 progress++;
3928                 trans = btrfs_start_transaction(rc->extent_root, 0);
3929                 if (IS_ERR(trans)) {
3930                         err = PTR_ERR(trans);
3931                         trans = NULL;
3932                         break;
3933                 }
3934 restart:
3935                 if (update_backref_cache(trans, &rc->backref_cache)) {
3936                         btrfs_end_transaction(trans, rc->extent_root);
3937                         continue;
3938                 }
3939
3940                 ret = find_next_extent(trans, rc, path, &key);
3941                 if (ret < 0)
3942                         err = ret;
3943                 if (ret != 0)
3944                         break;
3945
3946                 rc->extents_found++;
3947
3948                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3949                                     struct btrfs_extent_item);
3950                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3951                 if (item_size >= sizeof(*ei)) {
3952                         flags = btrfs_extent_flags(path->nodes[0], ei);
3953                         ret = check_extent_flags(flags);
3954                         BUG_ON(ret);
3955
3956                 } else {
3957 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3958                         u64 ref_owner;
3959                         int path_change = 0;
3960
3961                         BUG_ON(item_size !=
3962                                sizeof(struct btrfs_extent_item_v0));
3963                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3964                                                   &path_change);
3965                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3966                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3967                         else
3968                                 flags = BTRFS_EXTENT_FLAG_DATA;
3969
3970                         if (path_change) {
3971                                 btrfs_release_path(path);
3972
3973                                 path->search_commit_root = 1;
3974                                 path->skip_locking = 1;
3975                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3976                                                         &key, path, 0, 0);
3977                                 if (ret < 0) {
3978                                         err = ret;
3979                                         break;
3980                                 }
3981                                 BUG_ON(ret > 0);
3982                         }
3983 #else
3984                         BUG();
3985 #endif
3986                 }
3987
3988                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3989                         ret = add_tree_block(rc, &key, path, &blocks);
3990                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3991                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3992                         ret = add_data_references(rc, &key, path, &blocks);
3993                 } else {
3994                         btrfs_release_path(path);
3995                         ret = 0;
3996                 }
3997                 if (ret < 0) {
3998                         err = ret;
3999                         break;
4000                 }
4001
4002                 if (!RB_EMPTY_ROOT(&blocks)) {
4003                         ret = relocate_tree_blocks(trans, rc, &blocks);
4004                         if (ret < 0) {
4005                                 if (ret != -EAGAIN) {
4006                                         err = ret;
4007                                         break;
4008                                 }
4009                                 rc->extents_found--;
4010                                 rc->search_start = key.objectid;
4011                         }
4012                 }
4013
4014                 if (rc->commit_transaction) {
4015                         rc->commit_transaction = 0;
4016                         ret = btrfs_commit_transaction(trans, rc->extent_root);
4017                         BUG_ON(ret);
4018                 } else {
4019                         btrfs_end_transaction_throttle(trans, rc->extent_root);
4020                         btrfs_btree_balance_dirty(rc->extent_root);
4021                 }
4022                 trans = NULL;
4023
4024                 if (rc->stage == MOVE_DATA_EXTENTS &&
4025                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4026                         rc->found_file_extent = 1;
4027                         ret = relocate_data_extent(rc->data_inode,
4028                                                    &key, &rc->cluster);
4029                         if (ret < 0) {
4030                                 err = ret;
4031                                 break;
4032                         }
4033                 }
4034         }
4035         if (trans && progress && err == -ENOSPC) {
4036                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4037                                               rc->block_group->flags);
4038                 if (ret == 0) {
4039                         err = 0;
4040                         progress = 0;
4041                         goto restart;
4042                 }
4043         }
4044
4045         btrfs_release_path(path);
4046         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4047                           GFP_NOFS);
4048
4049         if (trans) {
4050                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4051                 btrfs_btree_balance_dirty(rc->extent_root);
4052         }
4053
4054         if (!err) {
4055                 ret = relocate_file_extent_cluster(rc->data_inode,
4056                                                    &rc->cluster);
4057                 if (ret < 0)
4058                         err = ret;
4059         }
4060
4061         rc->create_reloc_tree = 0;
4062         set_reloc_control(rc);
4063
4064         backref_cache_cleanup(&rc->backref_cache);
4065         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4066
4067         err = prepare_to_merge(rc, err);
4068
4069         merge_reloc_roots(rc);
4070
4071         rc->merge_reloc_tree = 0;
4072         unset_reloc_control(rc);
4073         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4074
4075         /* get rid of pinned extents */
4076         trans = btrfs_join_transaction(rc->extent_root);
4077         if (IS_ERR(trans))
4078                 err = PTR_ERR(trans);
4079         else
4080                 btrfs_commit_transaction(trans, rc->extent_root);
4081 out_free:
4082         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4083         btrfs_free_path(path);
4084         return err;
4085 }
4086
4087 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4088                                  struct btrfs_root *root, u64 objectid)
4089 {
4090         struct btrfs_path *path;
4091         struct btrfs_inode_item *item;
4092         struct extent_buffer *leaf;
4093         int ret;
4094
4095         path = btrfs_alloc_path();
4096         if (!path)
4097                 return -ENOMEM;
4098
4099         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4100         if (ret)
4101                 goto out;
4102
4103         leaf = path->nodes[0];
4104         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4105         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4106         btrfs_set_inode_generation(leaf, item, 1);
4107         btrfs_set_inode_size(leaf, item, 0);
4108         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4109         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4110                                           BTRFS_INODE_PREALLOC);
4111         btrfs_mark_buffer_dirty(leaf);
4112         btrfs_release_path(path);
4113 out:
4114         btrfs_free_path(path);
4115         return ret;
4116 }
4117
4118 /*
4119  * helper to create inode for data relocation.
4120  * the inode is in data relocation tree and its link count is 0
4121  */
4122 static noinline_for_stack
4123 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4124                                  struct btrfs_block_group_cache *group)
4125 {
4126         struct inode *inode = NULL;
4127         struct btrfs_trans_handle *trans;
4128         struct btrfs_root *root;
4129         struct btrfs_key key;
4130         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4131         int err = 0;
4132
4133         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4134         if (IS_ERR(root))
4135                 return ERR_CAST(root);
4136
4137         trans = btrfs_start_transaction(root, 6);
4138         if (IS_ERR(trans))
4139                 return ERR_CAST(trans);
4140
4141         err = btrfs_find_free_objectid(root, &objectid);
4142         if (err)
4143                 goto out;
4144
4145         err = __insert_orphan_inode(trans, root, objectid);
4146         BUG_ON(err);
4147
4148         key.objectid = objectid;
4149         key.type = BTRFS_INODE_ITEM_KEY;
4150         key.offset = 0;
4151         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4152         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4153         BTRFS_I(inode)->index_cnt = group->key.objectid;
4154
4155         err = btrfs_orphan_add(trans, inode);
4156 out:
4157         btrfs_end_transaction(trans, root);
4158         btrfs_btree_balance_dirty(root);
4159         if (err) {
4160                 if (inode)
4161                         iput(inode);
4162                 inode = ERR_PTR(err);
4163         }
4164         return inode;
4165 }
4166
4167 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4168 {
4169         struct reloc_control *rc;
4170
4171         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4172         if (!rc)
4173                 return NULL;
4174
4175         INIT_LIST_HEAD(&rc->reloc_roots);
4176         backref_cache_init(&rc->backref_cache);
4177         mapping_tree_init(&rc->reloc_root_tree);
4178         extent_io_tree_init(&rc->processed_blocks,
4179                             fs_info->btree_inode->i_mapping);
4180         return rc;
4181 }
4182
4183 /*
4184  * function to relocate all extents in a block group.
4185  */
4186 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4187 {
4188         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4189         struct reloc_control *rc;
4190         struct inode *inode;
4191         struct btrfs_path *path;
4192         int ret;
4193         int rw = 0;
4194         int err = 0;
4195
4196         rc = alloc_reloc_control(fs_info);
4197         if (!rc)
4198                 return -ENOMEM;
4199
4200         rc->extent_root = extent_root;
4201
4202         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4203         BUG_ON(!rc->block_group);
4204
4205         if (!rc->block_group->ro) {
4206                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4207                 if (ret) {
4208                         err = ret;
4209                         goto out;
4210                 }
4211                 rw = 1;
4212         }
4213
4214         path = btrfs_alloc_path();
4215         if (!path) {
4216                 err = -ENOMEM;
4217                 goto out;
4218         }
4219
4220         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4221                                         path);
4222         btrfs_free_path(path);
4223
4224         if (!IS_ERR(inode))
4225                 ret = delete_block_group_cache(fs_info, inode, 0);
4226         else
4227                 ret = PTR_ERR(inode);
4228
4229         if (ret && ret != -ENOENT) {
4230                 err = ret;
4231                 goto out;
4232         }
4233
4234         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4235         if (IS_ERR(rc->data_inode)) {
4236                 err = PTR_ERR(rc->data_inode);
4237                 rc->data_inode = NULL;
4238                 goto out;
4239         }
4240
4241         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4242                rc->block_group->key.objectid, rc->block_group->flags);
4243
4244         ret = btrfs_start_delalloc_roots(fs_info, 0);
4245         if (ret < 0) {
4246                 err = ret;
4247                 goto out;
4248         }
4249         btrfs_wait_ordered_roots(fs_info, -1);
4250
4251         while (1) {
4252                 mutex_lock(&fs_info->cleaner_mutex);
4253                 ret = relocate_block_group(rc);
4254                 mutex_unlock(&fs_info->cleaner_mutex);
4255                 if (ret < 0) {
4256                         err = ret;
4257                         goto out;
4258                 }
4259
4260                 if (rc->extents_found == 0)
4261                         break;
4262
4263                 printk(KERN_INFO "btrfs: found %llu extents\n",
4264                         rc->extents_found);
4265
4266                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4267                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4268                                                        (u64)-1);
4269                         if (ret) {
4270                                 err = ret;
4271                                 goto out;
4272                         }
4273                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4274                                                  0, -1);
4275                         rc->stage = UPDATE_DATA_PTRS;
4276                 }
4277         }
4278
4279         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4280                                      rc->block_group->key.objectid,
4281                                      rc->block_group->key.objectid +
4282                                      rc->block_group->key.offset - 1);
4283
4284         WARN_ON(rc->block_group->pinned > 0);
4285         WARN_ON(rc->block_group->reserved > 0);
4286         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4287 out:
4288         if (err && rw)
4289                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4290         iput(rc->data_inode);
4291         btrfs_put_block_group(rc->block_group);
4292         kfree(rc);
4293         return err;
4294 }
4295
4296 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4297 {
4298         struct btrfs_trans_handle *trans;
4299         int ret, err;
4300
4301         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4302         if (IS_ERR(trans))
4303                 return PTR_ERR(trans);
4304
4305         memset(&root->root_item.drop_progress, 0,
4306                 sizeof(root->root_item.drop_progress));
4307         root->root_item.drop_level = 0;
4308         btrfs_set_root_refs(&root->root_item, 0);
4309         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4310                                 &root->root_key, &root->root_item);
4311
4312         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4313         if (err)
4314                 return err;
4315         return ret;
4316 }
4317
4318 /*
4319  * recover relocation interrupted by system crash.
4320  *
4321  * this function resumes merging reloc trees with corresponding fs trees.
4322  * this is important for keeping the sharing of tree blocks
4323  */
4324 int btrfs_recover_relocation(struct btrfs_root *root)
4325 {
4326         LIST_HEAD(reloc_roots);
4327         struct btrfs_key key;
4328         struct btrfs_root *fs_root;
4329         struct btrfs_root *reloc_root;
4330         struct btrfs_path *path;
4331         struct extent_buffer *leaf;
4332         struct reloc_control *rc = NULL;
4333         struct btrfs_trans_handle *trans;
4334         int ret;
4335         int err = 0;
4336
4337         path = btrfs_alloc_path();
4338         if (!path)
4339                 return -ENOMEM;
4340         path->reada = -1;
4341
4342         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4343         key.type = BTRFS_ROOT_ITEM_KEY;
4344         key.offset = (u64)-1;
4345
4346         while (1) {
4347                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4348                                         path, 0, 0);
4349                 if (ret < 0) {
4350                         err = ret;
4351                         goto out;
4352                 }
4353                 if (ret > 0) {
4354                         if (path->slots[0] == 0)
4355                                 break;
4356                         path->slots[0]--;
4357                 }
4358                 leaf = path->nodes[0];
4359                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4360                 btrfs_release_path(path);
4361
4362                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4363                     key.type != BTRFS_ROOT_ITEM_KEY)
4364                         break;
4365
4366                 reloc_root = btrfs_read_fs_root(root, &key);
4367                 if (IS_ERR(reloc_root)) {
4368                         err = PTR_ERR(reloc_root);
4369                         goto out;
4370                 }
4371
4372                 list_add(&reloc_root->root_list, &reloc_roots);
4373
4374                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4375                         fs_root = read_fs_root(root->fs_info,
4376                                                reloc_root->root_key.offset);
4377                         if (IS_ERR(fs_root)) {
4378                                 ret = PTR_ERR(fs_root);
4379                                 if (ret != -ENOENT) {
4380                                         err = ret;
4381                                         goto out;
4382                                 }
4383                                 ret = mark_garbage_root(reloc_root);
4384                                 if (ret < 0) {
4385                                         err = ret;
4386                                         goto out;
4387                                 }
4388                         }
4389                 }
4390
4391                 if (key.offset == 0)
4392                         break;
4393
4394                 key.offset--;
4395         }
4396         btrfs_release_path(path);
4397
4398         if (list_empty(&reloc_roots))
4399                 goto out;
4400
4401         rc = alloc_reloc_control(root->fs_info);
4402         if (!rc) {
4403                 err = -ENOMEM;
4404                 goto out;
4405         }
4406
4407         rc->extent_root = root->fs_info->extent_root;
4408
4409         set_reloc_control(rc);
4410
4411         trans = btrfs_join_transaction(rc->extent_root);
4412         if (IS_ERR(trans)) {
4413                 unset_reloc_control(rc);
4414                 err = PTR_ERR(trans);
4415                 goto out_free;
4416         }
4417
4418         rc->merge_reloc_tree = 1;
4419
4420         while (!list_empty(&reloc_roots)) {
4421                 reloc_root = list_entry(reloc_roots.next,
4422                                         struct btrfs_root, root_list);
4423                 list_del(&reloc_root->root_list);
4424
4425                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4426                         list_add_tail(&reloc_root->root_list,
4427                                       &rc->reloc_roots);
4428                         continue;
4429                 }
4430
4431                 fs_root = read_fs_root(root->fs_info,
4432                                        reloc_root->root_key.offset);
4433                 if (IS_ERR(fs_root)) {
4434                         err = PTR_ERR(fs_root);
4435                         goto out_free;
4436                 }
4437
4438                 err = __add_reloc_root(reloc_root);
4439                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4440                 fs_root->reloc_root = reloc_root;
4441         }
4442
4443         err = btrfs_commit_transaction(trans, rc->extent_root);
4444         if (err)
4445                 goto out_free;
4446
4447         merge_reloc_roots(rc);
4448
4449         unset_reloc_control(rc);
4450
4451         trans = btrfs_join_transaction(rc->extent_root);
4452         if (IS_ERR(trans))
4453                 err = PTR_ERR(trans);
4454         else
4455                 err = btrfs_commit_transaction(trans, rc->extent_root);
4456 out_free:
4457         kfree(rc);
4458 out:
4459         if (!list_empty(&reloc_roots))
4460                 free_reloc_roots(&reloc_roots);
4461
4462         btrfs_free_path(path);
4463
4464         if (err == 0) {
4465                 /* cleanup orphan inode in data relocation tree */
4466                 fs_root = read_fs_root(root->fs_info,
4467                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4468                 if (IS_ERR(fs_root))
4469                         err = PTR_ERR(fs_root);
4470                 else
4471                         err = btrfs_orphan_cleanup(fs_root);
4472         }
4473         return err;
4474 }
4475
4476 /*
4477  * helper to add ordered checksum for data relocation.
4478  *
4479  * cloning checksum properly handles the nodatasum extents.
4480  * it also saves CPU time to re-calculate the checksum.
4481  */
4482 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4483 {
4484         struct btrfs_ordered_sum *sums;
4485         struct btrfs_ordered_extent *ordered;
4486         struct btrfs_root *root = BTRFS_I(inode)->root;
4487         int ret;
4488         u64 disk_bytenr;
4489         u64 new_bytenr;
4490         LIST_HEAD(list);
4491
4492         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4493         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4494
4495         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4496         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4497                                        disk_bytenr + len - 1, &list, 0);
4498         if (ret)
4499                 goto out;
4500
4501         while (!list_empty(&list)) {
4502                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4503                 list_del_init(&sums->list);
4504
4505                 /*
4506                  * We need to offset the new_bytenr based on where the csum is.
4507                  * We need to do this because we will read in entire prealloc
4508                  * extents but we may have written to say the middle of the
4509                  * prealloc extent, so we need to make sure the csum goes with
4510                  * the right disk offset.
4511                  *
4512                  * We can do this because the data reloc inode refers strictly
4513                  * to the on disk bytes, so we don't have to worry about
4514                  * disk_len vs real len like with real inodes since it's all
4515                  * disk length.
4516                  */
4517                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4518                 sums->bytenr = new_bytenr;
4519
4520                 btrfs_add_ordered_sum(inode, ordered, sums);
4521         }
4522 out:
4523         btrfs_put_ordered_extent(ordered);
4524         return ret;
4525 }
4526
4527 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4528                           struct btrfs_root *root, struct extent_buffer *buf,
4529                           struct extent_buffer *cow)
4530 {
4531         struct reloc_control *rc;
4532         struct backref_node *node;
4533         int first_cow = 0;
4534         int level;
4535         int ret = 0;
4536
4537         rc = root->fs_info->reloc_ctl;
4538         if (!rc)
4539                 return 0;
4540
4541         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4542                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4543
4544         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4545                 if (buf == root->node)
4546                         __update_reloc_root(root, cow->start);
4547         }
4548
4549         level = btrfs_header_level(buf);
4550         if (btrfs_header_generation(buf) <=
4551             btrfs_root_last_snapshot(&root->root_item))
4552                 first_cow = 1;
4553
4554         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4555             rc->create_reloc_tree) {
4556                 WARN_ON(!first_cow && level == 0);
4557
4558                 node = rc->backref_cache.path[level];
4559                 BUG_ON(node->bytenr != buf->start &&
4560                        node->new_bytenr != buf->start);
4561
4562                 drop_node_buffer(node);
4563                 extent_buffer_get(cow);
4564                 node->eb = cow;
4565                 node->new_bytenr = cow->start;
4566
4567                 if (!node->pending) {
4568                         list_move_tail(&node->list,
4569                                        &rc->backref_cache.pending[level]);
4570                         node->pending = 1;
4571                 }
4572
4573                 if (first_cow)
4574                         __mark_block_processed(rc, node);
4575
4576                 if (first_cow && level > 0)
4577                         rc->nodes_relocated += buf->len;
4578         }
4579
4580         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4581                 ret = replace_file_extents(trans, rc, root, cow);
4582         return ret;
4583 }
4584
4585 /*
4586  * called before creating snapshot. it calculates metadata reservation
4587  * requried for relocating tree blocks in the snapshot
4588  */
4589 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4590                               struct btrfs_pending_snapshot *pending,
4591                               u64 *bytes_to_reserve)
4592 {
4593         struct btrfs_root *root;
4594         struct reloc_control *rc;
4595
4596         root = pending->root;
4597         if (!root->reloc_root)
4598                 return;
4599
4600         rc = root->fs_info->reloc_ctl;
4601         if (!rc->merge_reloc_tree)
4602                 return;
4603
4604         root = root->reloc_root;
4605         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4606         /*
4607          * relocation is in the stage of merging trees. the space
4608          * used by merging a reloc tree is twice the size of
4609          * relocated tree nodes in the worst case. half for cowing
4610          * the reloc tree, half for cowing the fs tree. the space
4611          * used by cowing the reloc tree will be freed after the
4612          * tree is dropped. if we create snapshot, cowing the fs
4613          * tree may use more space than it frees. so we need
4614          * reserve extra space.
4615          */
4616         *bytes_to_reserve += rc->nodes_relocated;
4617 }
4618
4619 /*
4620  * called after snapshot is created. migrate block reservation
4621  * and create reloc root for the newly created snapshot
4622  */
4623 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4624                                struct btrfs_pending_snapshot *pending)
4625 {
4626         struct btrfs_root *root = pending->root;
4627         struct btrfs_root *reloc_root;
4628         struct btrfs_root *new_root;
4629         struct reloc_control *rc;
4630         int ret;
4631
4632         if (!root->reloc_root)
4633                 return 0;
4634
4635         rc = root->fs_info->reloc_ctl;
4636         rc->merging_rsv_size += rc->nodes_relocated;
4637
4638         if (rc->merge_reloc_tree) {
4639                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4640                                               rc->block_rsv,
4641                                               rc->nodes_relocated);
4642                 if (ret)
4643                         return ret;
4644         }
4645
4646         new_root = pending->snap;
4647         reloc_root = create_reloc_root(trans, root->reloc_root,
4648                                        new_root->root_key.objectid);
4649         if (IS_ERR(reloc_root))
4650                 return PTR_ERR(reloc_root);
4651
4652         ret = __add_reloc_root(reloc_root);
4653         BUG_ON(ret < 0);
4654         new_root->reloc_root = reloc_root;
4655
4656         if (rc->create_reloc_tree)
4657                 ret = clone_backref_node(trans, rc, root, reloc_root);
4658         return ret;
4659 }