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 update/delete the 'address of tree root -> reloc tree'
1268  * mapping
1269  */
1270 static int __update_reloc_root(struct btrfs_root *root, int del)
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->commit_root->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 0;
1287         BUG_ON((struct btrfs_root *)node->data != root);
1288
1289         if (!del) {
1290                 spin_lock(&rc->reloc_root_tree.lock);
1291                 node->bytenr = root->node->start;
1292                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1293                                       node->bytenr, &node->rb_node);
1294                 spin_unlock(&rc->reloc_root_tree.lock);
1295                 if (rb_node)
1296                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1297         } else {
1298                 spin_lock(&root->fs_info->trans_lock);
1299                 list_del_init(&root->root_list);
1300                 spin_unlock(&root->fs_info->trans_lock);
1301                 kfree(node);
1302         }
1303         return 0;
1304 }
1305
1306 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1307                                         struct btrfs_root *root, u64 objectid)
1308 {
1309         struct btrfs_root *reloc_root;
1310         struct extent_buffer *eb;
1311         struct btrfs_root_item *root_item;
1312         struct btrfs_key root_key;
1313         u64 last_snap = 0;
1314         int ret;
1315
1316         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1317         BUG_ON(!root_item);
1318
1319         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1320         root_key.type = BTRFS_ROOT_ITEM_KEY;
1321         root_key.offset = objectid;
1322
1323         if (root->root_key.objectid == objectid) {
1324                 /* called by btrfs_init_reloc_root */
1325                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1326                                       BTRFS_TREE_RELOC_OBJECTID);
1327                 BUG_ON(ret);
1328
1329                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1330                 btrfs_set_root_last_snapshot(&root->root_item,
1331                                              trans->transid - 1);
1332         } else {
1333                 /*
1334                  * called by btrfs_reloc_post_snapshot_hook.
1335                  * the source tree is a reloc tree, all tree blocks
1336                  * modified after it was created have RELOC flag
1337                  * set in their headers. so it's OK to not update
1338                  * the 'last_snapshot'.
1339                  */
1340                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1341                                       BTRFS_TREE_RELOC_OBJECTID);
1342                 BUG_ON(ret);
1343         }
1344
1345         memcpy(root_item, &root->root_item, sizeof(*root_item));
1346         btrfs_set_root_bytenr(root_item, eb->start);
1347         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1348         btrfs_set_root_generation(root_item, trans->transid);
1349
1350         if (root->root_key.objectid == objectid) {
1351                 btrfs_set_root_refs(root_item, 0);
1352                 memset(&root_item->drop_progress, 0,
1353                        sizeof(struct btrfs_disk_key));
1354                 root_item->drop_level = 0;
1355                 /*
1356                  * abuse rtransid, it is safe because it is impossible to
1357                  * receive data into a relocation tree.
1358                  */
1359                 btrfs_set_root_rtransid(root_item, last_snap);
1360                 btrfs_set_root_otransid(root_item, trans->transid);
1361         }
1362
1363         btrfs_tree_unlock(eb);
1364         free_extent_buffer(eb);
1365
1366         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1367                                 &root_key, root_item);
1368         BUG_ON(ret);
1369         kfree(root_item);
1370
1371         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1372         BUG_ON(IS_ERR(reloc_root));
1373         reloc_root->last_trans = trans->transid;
1374         return reloc_root;
1375 }
1376
1377 /*
1378  * create reloc tree for a given fs tree. reloc tree is just a
1379  * snapshot of the fs tree with special root objectid.
1380  */
1381 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1382                           struct btrfs_root *root)
1383 {
1384         struct btrfs_root *reloc_root;
1385         struct reloc_control *rc = root->fs_info->reloc_ctl;
1386         struct btrfs_block_rsv *rsv;
1387         int clear_rsv = 0;
1388         int ret;
1389
1390         if (root->reloc_root) {
1391                 reloc_root = root->reloc_root;
1392                 reloc_root->last_trans = trans->transid;
1393                 return 0;
1394         }
1395
1396         if (!rc || !rc->create_reloc_tree ||
1397             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1398                 return 0;
1399
1400         if (!trans->reloc_reserved) {
1401                 rsv = trans->block_rsv;
1402                 trans->block_rsv = rc->block_rsv;
1403                 clear_rsv = 1;
1404         }
1405         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1406         if (clear_rsv)
1407                 trans->block_rsv = rsv;
1408
1409         ret = __add_reloc_root(reloc_root);
1410         BUG_ON(ret < 0);
1411         root->reloc_root = reloc_root;
1412         return 0;
1413 }
1414
1415 /*
1416  * update root item of reloc tree
1417  */
1418 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1419                             struct btrfs_root *root)
1420 {
1421         struct btrfs_root *reloc_root;
1422         struct btrfs_root_item *root_item;
1423         int del = 0;
1424         int ret;
1425
1426         if (!root->reloc_root)
1427                 goto out;
1428
1429         reloc_root = root->reloc_root;
1430         root_item = &reloc_root->root_item;
1431
1432         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1433             btrfs_root_refs(root_item) == 0) {
1434                 root->reloc_root = NULL;
1435                 del = 1;
1436         }
1437
1438         __update_reloc_root(reloc_root, del);
1439
1440         if (reloc_root->commit_root != reloc_root->node) {
1441                 btrfs_set_root_node(root_item, reloc_root->node);
1442                 free_extent_buffer(reloc_root->commit_root);
1443                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1444         }
1445
1446         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1447                                 &reloc_root->root_key, root_item);
1448         BUG_ON(ret);
1449
1450 out:
1451         return 0;
1452 }
1453
1454 /*
1455  * helper to find first cached inode with inode number >= objectid
1456  * in a subvolume
1457  */
1458 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1459 {
1460         struct rb_node *node;
1461         struct rb_node *prev;
1462         struct btrfs_inode *entry;
1463         struct inode *inode;
1464
1465         spin_lock(&root->inode_lock);
1466 again:
1467         node = root->inode_tree.rb_node;
1468         prev = NULL;
1469         while (node) {
1470                 prev = node;
1471                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1472
1473                 if (objectid < btrfs_ino(&entry->vfs_inode))
1474                         node = node->rb_left;
1475                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1476                         node = node->rb_right;
1477                 else
1478                         break;
1479         }
1480         if (!node) {
1481                 while (prev) {
1482                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1483                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1484                                 node = prev;
1485                                 break;
1486                         }
1487                         prev = rb_next(prev);
1488                 }
1489         }
1490         while (node) {
1491                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1492                 inode = igrab(&entry->vfs_inode);
1493                 if (inode) {
1494                         spin_unlock(&root->inode_lock);
1495                         return inode;
1496                 }
1497
1498                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1499                 if (cond_resched_lock(&root->inode_lock))
1500                         goto again;
1501
1502                 node = rb_next(node);
1503         }
1504         spin_unlock(&root->inode_lock);
1505         return NULL;
1506 }
1507
1508 static int in_block_group(u64 bytenr,
1509                           struct btrfs_block_group_cache *block_group)
1510 {
1511         if (bytenr >= block_group->key.objectid &&
1512             bytenr < block_group->key.objectid + block_group->key.offset)
1513                 return 1;
1514         return 0;
1515 }
1516
1517 /*
1518  * get new location of data
1519  */
1520 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1521                             u64 bytenr, u64 num_bytes)
1522 {
1523         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1524         struct btrfs_path *path;
1525         struct btrfs_file_extent_item *fi;
1526         struct extent_buffer *leaf;
1527         int ret;
1528
1529         path = btrfs_alloc_path();
1530         if (!path)
1531                 return -ENOMEM;
1532
1533         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1534         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1535                                        bytenr, 0);
1536         if (ret < 0)
1537                 goto out;
1538         if (ret > 0) {
1539                 ret = -ENOENT;
1540                 goto out;
1541         }
1542
1543         leaf = path->nodes[0];
1544         fi = btrfs_item_ptr(leaf, path->slots[0],
1545                             struct btrfs_file_extent_item);
1546
1547         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1548                btrfs_file_extent_compression(leaf, fi) ||
1549                btrfs_file_extent_encryption(leaf, fi) ||
1550                btrfs_file_extent_other_encoding(leaf, fi));
1551
1552         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1553                 ret = -EINVAL;
1554                 goto out;
1555         }
1556
1557         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1558         ret = 0;
1559 out:
1560         btrfs_free_path(path);
1561         return ret;
1562 }
1563
1564 /*
1565  * update file extent items in the tree leaf to point to
1566  * the new locations.
1567  */
1568 static noinline_for_stack
1569 int replace_file_extents(struct btrfs_trans_handle *trans,
1570                          struct reloc_control *rc,
1571                          struct btrfs_root *root,
1572                          struct extent_buffer *leaf)
1573 {
1574         struct btrfs_key key;
1575         struct btrfs_file_extent_item *fi;
1576         struct inode *inode = NULL;
1577         u64 parent;
1578         u64 bytenr;
1579         u64 new_bytenr = 0;
1580         u64 num_bytes;
1581         u64 end;
1582         u32 nritems;
1583         u32 i;
1584         int ret = 0;
1585         int first = 1;
1586         int dirty = 0;
1587
1588         if (rc->stage != UPDATE_DATA_PTRS)
1589                 return 0;
1590
1591         /* reloc trees always use full backref */
1592         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1593                 parent = leaf->start;
1594         else
1595                 parent = 0;
1596
1597         nritems = btrfs_header_nritems(leaf);
1598         for (i = 0; i < nritems; i++) {
1599                 cond_resched();
1600                 btrfs_item_key_to_cpu(leaf, &key, i);
1601                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1602                         continue;
1603                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1604                 if (btrfs_file_extent_type(leaf, fi) ==
1605                     BTRFS_FILE_EXTENT_INLINE)
1606                         continue;
1607                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1608                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1609                 if (bytenr == 0)
1610                         continue;
1611                 if (!in_block_group(bytenr, rc->block_group))
1612                         continue;
1613
1614                 /*
1615                  * if we are modifying block in fs tree, wait for readpage
1616                  * to complete and drop the extent cache
1617                  */
1618                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1619                         if (first) {
1620                                 inode = find_next_inode(root, key.objectid);
1621                                 first = 0;
1622                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1623                                 btrfs_add_delayed_iput(inode);
1624                                 inode = find_next_inode(root, key.objectid);
1625                         }
1626                         if (inode && btrfs_ino(inode) == key.objectid) {
1627                                 end = key.offset +
1628                                       btrfs_file_extent_num_bytes(leaf, fi);
1629                                 WARN_ON(!IS_ALIGNED(key.offset,
1630                                                     root->sectorsize));
1631                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1632                                 end--;
1633                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1634                                                       key.offset, end);
1635                                 if (!ret)
1636                                         continue;
1637
1638                                 btrfs_drop_extent_cache(inode, key.offset, end,
1639                                                         1);
1640                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1641                                               key.offset, end);
1642                         }
1643                 }
1644
1645                 ret = get_new_location(rc->data_inode, &new_bytenr,
1646                                        bytenr, num_bytes);
1647                 if (ret) {
1648                         /*
1649                          * Don't have to abort since we've not changed anything
1650                          * in the file extent yet.
1651                          */
1652                         break;
1653                 }
1654
1655                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1656                 dirty = 1;
1657
1658                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1659                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1660                                            num_bytes, parent,
1661                                            btrfs_header_owner(leaf),
1662                                            key.objectid, key.offset, 1);
1663                 if (ret) {
1664                         btrfs_abort_transaction(trans, root, ret);
1665                         break;
1666                 }
1667
1668                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1669                                         parent, btrfs_header_owner(leaf),
1670                                         key.objectid, key.offset, 1);
1671                 if (ret) {
1672                         btrfs_abort_transaction(trans, root, ret);
1673                         break;
1674                 }
1675         }
1676         if (dirty)
1677                 btrfs_mark_buffer_dirty(leaf);
1678         if (inode)
1679                 btrfs_add_delayed_iput(inode);
1680         return ret;
1681 }
1682
1683 static noinline_for_stack
1684 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1685                      struct btrfs_path *path, int level)
1686 {
1687         struct btrfs_disk_key key1;
1688         struct btrfs_disk_key key2;
1689         btrfs_node_key(eb, &key1, slot);
1690         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1691         return memcmp(&key1, &key2, sizeof(key1));
1692 }
1693
1694 /*
1695  * try to replace tree blocks in fs tree with the new blocks
1696  * in reloc tree. tree blocks haven't been modified since the
1697  * reloc tree was create can be replaced.
1698  *
1699  * if a block was replaced, level of the block + 1 is returned.
1700  * if no block got replaced, 0 is returned. if there are other
1701  * errors, a negative error number is returned.
1702  */
1703 static noinline_for_stack
1704 int replace_path(struct btrfs_trans_handle *trans,
1705                  struct btrfs_root *dest, struct btrfs_root *src,
1706                  struct btrfs_path *path, struct btrfs_key *next_key,
1707                  int lowest_level, int max_level)
1708 {
1709         struct extent_buffer *eb;
1710         struct extent_buffer *parent;
1711         struct btrfs_key key;
1712         u64 old_bytenr;
1713         u64 new_bytenr;
1714         u64 old_ptr_gen;
1715         u64 new_ptr_gen;
1716         u64 last_snapshot;
1717         u32 blocksize;
1718         int cow = 0;
1719         int level;
1720         int ret;
1721         int slot;
1722
1723         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1724         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1725
1726         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1727 again:
1728         slot = path->slots[lowest_level];
1729         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1730
1731         eb = btrfs_lock_root_node(dest);
1732         btrfs_set_lock_blocking(eb);
1733         level = btrfs_header_level(eb);
1734
1735         if (level < lowest_level) {
1736                 btrfs_tree_unlock(eb);
1737                 free_extent_buffer(eb);
1738                 return 0;
1739         }
1740
1741         if (cow) {
1742                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1743                 BUG_ON(ret);
1744         }
1745         btrfs_set_lock_blocking(eb);
1746
1747         if (next_key) {
1748                 next_key->objectid = (u64)-1;
1749                 next_key->type = (u8)-1;
1750                 next_key->offset = (u64)-1;
1751         }
1752
1753         parent = eb;
1754         while (1) {
1755                 level = btrfs_header_level(parent);
1756                 BUG_ON(level < lowest_level);
1757
1758                 ret = btrfs_bin_search(parent, &key, level, &slot);
1759                 if (ret && slot > 0)
1760                         slot--;
1761
1762                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1763                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1764
1765                 old_bytenr = btrfs_node_blockptr(parent, slot);
1766                 blocksize = btrfs_level_size(dest, level - 1);
1767                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1768
1769                 if (level <= max_level) {
1770                         eb = path->nodes[level];
1771                         new_bytenr = btrfs_node_blockptr(eb,
1772                                                         path->slots[level]);
1773                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1774                                                         path->slots[level]);
1775                 } else {
1776                         new_bytenr = 0;
1777                         new_ptr_gen = 0;
1778                 }
1779
1780                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1781                         WARN_ON(1);
1782                         ret = level;
1783                         break;
1784                 }
1785
1786                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1787                     memcmp_node_keys(parent, slot, path, level)) {
1788                         if (level <= lowest_level) {
1789                                 ret = 0;
1790                                 break;
1791                         }
1792
1793                         eb = read_tree_block(dest, old_bytenr, blocksize,
1794                                              old_ptr_gen);
1795                         if (!eb || !extent_buffer_uptodate(eb)) {
1796                                 ret = (!eb) ? -ENOMEM : -EIO;
1797                                 free_extent_buffer(eb);
1798                                 break;
1799                         }
1800                         btrfs_tree_lock(eb);
1801                         if (cow) {
1802                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1803                                                       slot, &eb);
1804                                 BUG_ON(ret);
1805                         }
1806                         btrfs_set_lock_blocking(eb);
1807
1808                         btrfs_tree_unlock(parent);
1809                         free_extent_buffer(parent);
1810
1811                         parent = eb;
1812                         continue;
1813                 }
1814
1815                 if (!cow) {
1816                         btrfs_tree_unlock(parent);
1817                         free_extent_buffer(parent);
1818                         cow = 1;
1819                         goto again;
1820                 }
1821
1822                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1823                                       path->slots[level]);
1824                 btrfs_release_path(path);
1825
1826                 path->lowest_level = level;
1827                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1828                 path->lowest_level = 0;
1829                 BUG_ON(ret);
1830
1831                 /*
1832                  * swap blocks in fs tree and reloc tree.
1833                  */
1834                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1835                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1836                 btrfs_mark_buffer_dirty(parent);
1837
1838                 btrfs_set_node_blockptr(path->nodes[level],
1839                                         path->slots[level], old_bytenr);
1840                 btrfs_set_node_ptr_generation(path->nodes[level],
1841                                               path->slots[level], old_ptr_gen);
1842                 btrfs_mark_buffer_dirty(path->nodes[level]);
1843
1844                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1845                                         path->nodes[level]->start,
1846                                         src->root_key.objectid, level - 1, 0,
1847                                         1);
1848                 BUG_ON(ret);
1849                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1850                                         0, dest->root_key.objectid, level - 1,
1851                                         0, 1);
1852                 BUG_ON(ret);
1853
1854                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1855                                         path->nodes[level]->start,
1856                                         src->root_key.objectid, level - 1, 0,
1857                                         1);
1858                 BUG_ON(ret);
1859
1860                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1861                                         0, dest->root_key.objectid, level - 1,
1862                                         0, 1);
1863                 BUG_ON(ret);
1864
1865                 btrfs_unlock_up_safe(path, 0);
1866
1867                 ret = level;
1868                 break;
1869         }
1870         btrfs_tree_unlock(parent);
1871         free_extent_buffer(parent);
1872         return ret;
1873 }
1874
1875 /*
1876  * helper to find next relocated block in reloc tree
1877  */
1878 static noinline_for_stack
1879 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1880                        int *level)
1881 {
1882         struct extent_buffer *eb;
1883         int i;
1884         u64 last_snapshot;
1885         u32 nritems;
1886
1887         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1888
1889         for (i = 0; i < *level; i++) {
1890                 free_extent_buffer(path->nodes[i]);
1891                 path->nodes[i] = NULL;
1892         }
1893
1894         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1895                 eb = path->nodes[i];
1896                 nritems = btrfs_header_nritems(eb);
1897                 while (path->slots[i] + 1 < nritems) {
1898                         path->slots[i]++;
1899                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1900                             last_snapshot)
1901                                 continue;
1902
1903                         *level = i;
1904                         return 0;
1905                 }
1906                 free_extent_buffer(path->nodes[i]);
1907                 path->nodes[i] = NULL;
1908         }
1909         return 1;
1910 }
1911
1912 /*
1913  * walk down reloc tree to find relocated block of lowest level
1914  */
1915 static noinline_for_stack
1916 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1917                          int *level)
1918 {
1919         struct extent_buffer *eb = NULL;
1920         int i;
1921         u64 bytenr;
1922         u64 ptr_gen = 0;
1923         u64 last_snapshot;
1924         u32 blocksize;
1925         u32 nritems;
1926
1927         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1928
1929         for (i = *level; i > 0; i--) {
1930                 eb = path->nodes[i];
1931                 nritems = btrfs_header_nritems(eb);
1932                 while (path->slots[i] < nritems) {
1933                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1934                         if (ptr_gen > last_snapshot)
1935                                 break;
1936                         path->slots[i]++;
1937                 }
1938                 if (path->slots[i] >= nritems) {
1939                         if (i == *level)
1940                                 break;
1941                         *level = i + 1;
1942                         return 0;
1943                 }
1944                 if (i == 1) {
1945                         *level = i;
1946                         return 0;
1947                 }
1948
1949                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1950                 blocksize = btrfs_level_size(root, i - 1);
1951                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1952                 if (!eb || !extent_buffer_uptodate(eb)) {
1953                         free_extent_buffer(eb);
1954                         return -EIO;
1955                 }
1956                 BUG_ON(btrfs_header_level(eb) != i - 1);
1957                 path->nodes[i - 1] = eb;
1958                 path->slots[i - 1] = 0;
1959         }
1960         return 1;
1961 }
1962
1963 /*
1964  * invalidate extent cache for file extents whose key in range of
1965  * [min_key, max_key)
1966  */
1967 static int invalidate_extent_cache(struct btrfs_root *root,
1968                                    struct btrfs_key *min_key,
1969                                    struct btrfs_key *max_key)
1970 {
1971         struct inode *inode = NULL;
1972         u64 objectid;
1973         u64 start, end;
1974         u64 ino;
1975
1976         objectid = min_key->objectid;
1977         while (1) {
1978                 cond_resched();
1979                 iput(inode);
1980
1981                 if (objectid > max_key->objectid)
1982                         break;
1983
1984                 inode = find_next_inode(root, objectid);
1985                 if (!inode)
1986                         break;
1987                 ino = btrfs_ino(inode);
1988
1989                 if (ino > max_key->objectid) {
1990                         iput(inode);
1991                         break;
1992                 }
1993
1994                 objectid = ino + 1;
1995                 if (!S_ISREG(inode->i_mode))
1996                         continue;
1997
1998                 if (unlikely(min_key->objectid == ino)) {
1999                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2000                                 continue;
2001                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2002                                 start = 0;
2003                         else {
2004                                 start = min_key->offset;
2005                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2006                         }
2007                 } else {
2008                         start = 0;
2009                 }
2010
2011                 if (unlikely(max_key->objectid == ino)) {
2012                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2013                                 continue;
2014                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2015                                 end = (u64)-1;
2016                         } else {
2017                                 if (max_key->offset == 0)
2018                                         continue;
2019                                 end = max_key->offset;
2020                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2021                                 end--;
2022                         }
2023                 } else {
2024                         end = (u64)-1;
2025                 }
2026
2027                 /* the lock_extent waits for readpage to complete */
2028                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2029                 btrfs_drop_extent_cache(inode, start, end, 1);
2030                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2031         }
2032         return 0;
2033 }
2034
2035 static int find_next_key(struct btrfs_path *path, int level,
2036                          struct btrfs_key *key)
2037
2038 {
2039         while (level < BTRFS_MAX_LEVEL) {
2040                 if (!path->nodes[level])
2041                         break;
2042                 if (path->slots[level] + 1 <
2043                     btrfs_header_nritems(path->nodes[level])) {
2044                         btrfs_node_key_to_cpu(path->nodes[level], key,
2045                                               path->slots[level] + 1);
2046                         return 0;
2047                 }
2048                 level++;
2049         }
2050         return 1;
2051 }
2052
2053 /*
2054  * merge the relocated tree blocks in reloc tree with corresponding
2055  * fs tree.
2056  */
2057 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2058                                                struct btrfs_root *root)
2059 {
2060         LIST_HEAD(inode_list);
2061         struct btrfs_key key;
2062         struct btrfs_key next_key;
2063         struct btrfs_trans_handle *trans;
2064         struct btrfs_root *reloc_root;
2065         struct btrfs_root_item *root_item;
2066         struct btrfs_path *path;
2067         struct extent_buffer *leaf;
2068         int level;
2069         int max_level;
2070         int replaced = 0;
2071         int ret;
2072         int err = 0;
2073         u32 min_reserved;
2074
2075         path = btrfs_alloc_path();
2076         if (!path)
2077                 return -ENOMEM;
2078         path->reada = 1;
2079
2080         reloc_root = root->reloc_root;
2081         root_item = &reloc_root->root_item;
2082
2083         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2084                 level = btrfs_root_level(root_item);
2085                 extent_buffer_get(reloc_root->node);
2086                 path->nodes[level] = reloc_root->node;
2087                 path->slots[level] = 0;
2088         } else {
2089                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2090
2091                 level = root_item->drop_level;
2092                 BUG_ON(level == 0);
2093                 path->lowest_level = level;
2094                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2095                 path->lowest_level = 0;
2096                 if (ret < 0) {
2097                         btrfs_free_path(path);
2098                         return ret;
2099                 }
2100
2101                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2102                                       path->slots[level]);
2103                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2104
2105                 btrfs_unlock_up_safe(path, 0);
2106         }
2107
2108         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2109         memset(&next_key, 0, sizeof(next_key));
2110
2111         while (1) {
2112                 trans = btrfs_start_transaction(root, 0);
2113                 BUG_ON(IS_ERR(trans));
2114                 trans->block_rsv = rc->block_rsv;
2115
2116                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2117                                              BTRFS_RESERVE_FLUSH_ALL);
2118                 if (ret) {
2119                         BUG_ON(ret != -EAGAIN);
2120                         ret = btrfs_commit_transaction(trans, root);
2121                         BUG_ON(ret);
2122                         continue;
2123                 }
2124
2125                 replaced = 0;
2126                 max_level = level;
2127
2128                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2129                 if (ret < 0) {
2130                         err = ret;
2131                         goto out;
2132                 }
2133                 if (ret > 0)
2134                         break;
2135
2136                 if (!find_next_key(path, level, &key) &&
2137                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2138                         ret = 0;
2139                 } else {
2140                         ret = replace_path(trans, root, reloc_root, path,
2141                                            &next_key, level, max_level);
2142                 }
2143                 if (ret < 0) {
2144                         err = ret;
2145                         goto out;
2146                 }
2147
2148                 if (ret > 0) {
2149                         level = ret;
2150                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2151                                               path->slots[level]);
2152                         replaced = 1;
2153                 }
2154
2155                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2156                 if (ret > 0)
2157                         break;
2158
2159                 BUG_ON(level == 0);
2160                 /*
2161                  * save the merging progress in the drop_progress.
2162                  * this is OK since root refs == 1 in this case.
2163                  */
2164                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2165                                path->slots[level]);
2166                 root_item->drop_level = level;
2167
2168                 btrfs_end_transaction_throttle(trans, root);
2169
2170                 btrfs_btree_balance_dirty(root);
2171
2172                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2173                         invalidate_extent_cache(root, &key, &next_key);
2174         }
2175
2176         /*
2177          * handle the case only one block in the fs tree need to be
2178          * relocated and the block is tree root.
2179          */
2180         leaf = btrfs_lock_root_node(root);
2181         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2182         btrfs_tree_unlock(leaf);
2183         free_extent_buffer(leaf);
2184         if (ret < 0)
2185                 err = ret;
2186 out:
2187         btrfs_free_path(path);
2188
2189         if (err == 0) {
2190                 memset(&root_item->drop_progress, 0,
2191                        sizeof(root_item->drop_progress));
2192                 root_item->drop_level = 0;
2193                 btrfs_set_root_refs(root_item, 0);
2194                 btrfs_update_reloc_root(trans, root);
2195         }
2196
2197         btrfs_end_transaction_throttle(trans, root);
2198
2199         btrfs_btree_balance_dirty(root);
2200
2201         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2202                 invalidate_extent_cache(root, &key, &next_key);
2203
2204         return err;
2205 }
2206
2207 static noinline_for_stack
2208 int prepare_to_merge(struct reloc_control *rc, int err)
2209 {
2210         struct btrfs_root *root = rc->extent_root;
2211         struct btrfs_root *reloc_root;
2212         struct btrfs_trans_handle *trans;
2213         LIST_HEAD(reloc_roots);
2214         u64 num_bytes = 0;
2215         int ret;
2216
2217         mutex_lock(&root->fs_info->reloc_mutex);
2218         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2219         rc->merging_rsv_size += rc->nodes_relocated * 2;
2220         mutex_unlock(&root->fs_info->reloc_mutex);
2221
2222 again:
2223         if (!err) {
2224                 num_bytes = rc->merging_rsv_size;
2225                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2226                                           BTRFS_RESERVE_FLUSH_ALL);
2227                 if (ret)
2228                         err = ret;
2229         }
2230
2231         trans = btrfs_join_transaction(rc->extent_root);
2232         if (IS_ERR(trans)) {
2233                 if (!err)
2234                         btrfs_block_rsv_release(rc->extent_root,
2235                                                 rc->block_rsv, num_bytes);
2236                 return PTR_ERR(trans);
2237         }
2238
2239         if (!err) {
2240                 if (num_bytes != rc->merging_rsv_size) {
2241                         btrfs_end_transaction(trans, rc->extent_root);
2242                         btrfs_block_rsv_release(rc->extent_root,
2243                                                 rc->block_rsv, num_bytes);
2244                         goto again;
2245                 }
2246         }
2247
2248         rc->merge_reloc_tree = 1;
2249
2250         while (!list_empty(&rc->reloc_roots)) {
2251                 reloc_root = list_entry(rc->reloc_roots.next,
2252                                         struct btrfs_root, root_list);
2253                 list_del_init(&reloc_root->root_list);
2254
2255                 root = read_fs_root(reloc_root->fs_info,
2256                                     reloc_root->root_key.offset);
2257                 BUG_ON(IS_ERR(root));
2258                 BUG_ON(root->reloc_root != reloc_root);
2259
2260                 /*
2261                  * set reference count to 1, so btrfs_recover_relocation
2262                  * knows it should resumes merging
2263                  */
2264                 if (!err)
2265                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2266                 btrfs_update_reloc_root(trans, root);
2267
2268                 list_add(&reloc_root->root_list, &reloc_roots);
2269         }
2270
2271         list_splice(&reloc_roots, &rc->reloc_roots);
2272
2273         if (!err)
2274                 btrfs_commit_transaction(trans, rc->extent_root);
2275         else
2276                 btrfs_end_transaction(trans, rc->extent_root);
2277         return err;
2278 }
2279
2280 static noinline_for_stack
2281 void free_reloc_roots(struct list_head *list)
2282 {
2283         struct btrfs_root *reloc_root;
2284
2285         while (!list_empty(list)) {
2286                 reloc_root = list_entry(list->next, struct btrfs_root,
2287                                         root_list);
2288                 __update_reloc_root(reloc_root, 1);
2289                 free_extent_buffer(reloc_root->node);
2290                 free_extent_buffer(reloc_root->commit_root);
2291                 kfree(reloc_root);
2292         }
2293 }
2294
2295 static noinline_for_stack
2296 int merge_reloc_roots(struct reloc_control *rc)
2297 {
2298         struct btrfs_trans_handle *trans;
2299         struct btrfs_root *root;
2300         struct btrfs_root *reloc_root;
2301         u64 last_snap;
2302         u64 otransid;
2303         u64 objectid;
2304         LIST_HEAD(reloc_roots);
2305         int found = 0;
2306         int ret = 0;
2307 again:
2308         root = rc->extent_root;
2309
2310         /*
2311          * this serializes us with btrfs_record_root_in_transaction,
2312          * we have to make sure nobody is in the middle of
2313          * adding their roots to the list while we are
2314          * doing this splice
2315          */
2316         mutex_lock(&root->fs_info->reloc_mutex);
2317         list_splice_init(&rc->reloc_roots, &reloc_roots);
2318         mutex_unlock(&root->fs_info->reloc_mutex);
2319
2320         while (!list_empty(&reloc_roots)) {
2321                 found = 1;
2322                 reloc_root = list_entry(reloc_roots.next,
2323                                         struct btrfs_root, root_list);
2324
2325                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2326                         root = read_fs_root(reloc_root->fs_info,
2327                                             reloc_root->root_key.offset);
2328                         BUG_ON(IS_ERR(root));
2329                         BUG_ON(root->reloc_root != reloc_root);
2330
2331                         ret = merge_reloc_root(rc, root);
2332                         if (ret) {
2333                                 __update_reloc_root(reloc_root, 1);
2334                                 free_extent_buffer(reloc_root->node);
2335                                 free_extent_buffer(reloc_root->commit_root);
2336                                 kfree(reloc_root);
2337                                 goto out;
2338                         }
2339                 } else {
2340                         list_del_init(&reloc_root->root_list);
2341                 }
2342
2343                 /*
2344                  * we keep the old last snapshod transid in rtranid when we
2345                  * created the relocation tree.
2346                  */
2347                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2348                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2349                 objectid = reloc_root->root_key.offset;
2350
2351                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2352                 if (ret < 0) {
2353                         if (list_empty(&reloc_root->root_list))
2354                                 list_add_tail(&reloc_root->root_list,
2355                                               &reloc_roots);
2356                         goto out;
2357                 } else if (!ret) {
2358                         /*
2359                          * recover the last snapshot tranid to avoid
2360                          * the space balance break NOCOW.
2361                          */
2362                         root = read_fs_root(rc->extent_root->fs_info,
2363                                             objectid);
2364                         if (IS_ERR(root))
2365                                 continue;
2366
2367                         trans = btrfs_join_transaction(root);
2368                         BUG_ON(IS_ERR(trans));
2369
2370                         /* Check if the fs/file tree was snapshoted or not. */
2371                         if (btrfs_root_last_snapshot(&root->root_item) ==
2372                             otransid - 1)
2373                                 btrfs_set_root_last_snapshot(&root->root_item,
2374                                                              last_snap);
2375
2376                         btrfs_end_transaction(trans, root);
2377                 }
2378         }
2379
2380         if (found) {
2381                 found = 0;
2382                 goto again;
2383         }
2384 out:
2385         if (ret) {
2386                 btrfs_std_error(root->fs_info, ret);
2387                 if (!list_empty(&reloc_roots))
2388                         free_reloc_roots(&reloc_roots);
2389         }
2390
2391         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2392         return ret;
2393 }
2394
2395 static void free_block_list(struct rb_root *blocks)
2396 {
2397         struct tree_block *block;
2398         struct rb_node *rb_node;
2399         while ((rb_node = rb_first(blocks))) {
2400                 block = rb_entry(rb_node, struct tree_block, rb_node);
2401                 rb_erase(rb_node, blocks);
2402                 kfree(block);
2403         }
2404 }
2405
2406 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2407                                       struct btrfs_root *reloc_root)
2408 {
2409         struct btrfs_root *root;
2410
2411         if (reloc_root->last_trans == trans->transid)
2412                 return 0;
2413
2414         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2415         BUG_ON(IS_ERR(root));
2416         BUG_ON(root->reloc_root != reloc_root);
2417
2418         return btrfs_record_root_in_trans(trans, root);
2419 }
2420
2421 static noinline_for_stack
2422 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2423                                      struct reloc_control *rc,
2424                                      struct backref_node *node,
2425                                      struct backref_edge *edges[], int *nr)
2426 {
2427         struct backref_node *next;
2428         struct btrfs_root *root;
2429         int index = 0;
2430
2431         next = node;
2432         while (1) {
2433                 cond_resched();
2434                 next = walk_up_backref(next, edges, &index);
2435                 root = next->root;
2436                 BUG_ON(!root);
2437                 BUG_ON(!root->ref_cows);
2438
2439                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2440                         record_reloc_root_in_trans(trans, root);
2441                         break;
2442                 }
2443
2444                 btrfs_record_root_in_trans(trans, root);
2445                 root = root->reloc_root;
2446
2447                 if (next->new_bytenr != root->node->start) {
2448                         BUG_ON(next->new_bytenr);
2449                         BUG_ON(!list_empty(&next->list));
2450                         next->new_bytenr = root->node->start;
2451                         next->root = root;
2452                         list_add_tail(&next->list,
2453                                       &rc->backref_cache.changed);
2454                         __mark_block_processed(rc, next);
2455                         break;
2456                 }
2457
2458                 WARN_ON(1);
2459                 root = NULL;
2460                 next = walk_down_backref(edges, &index);
2461                 if (!next || next->level <= node->level)
2462                         break;
2463         }
2464         if (!root)
2465                 return NULL;
2466
2467         *nr = index;
2468         next = node;
2469         /* setup backref node path for btrfs_reloc_cow_block */
2470         while (1) {
2471                 rc->backref_cache.path[next->level] = next;
2472                 if (--index < 0)
2473                         break;
2474                 next = edges[index]->node[UPPER];
2475         }
2476         return root;
2477 }
2478
2479 /*
2480  * select a tree root for relocation. return NULL if the block
2481  * is reference counted. we should use do_relocation() in this
2482  * case. return a tree root pointer if the block isn't reference
2483  * counted. return -ENOENT if the block is root of reloc tree.
2484  */
2485 static noinline_for_stack
2486 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2487                                    struct backref_node *node)
2488 {
2489         struct backref_node *next;
2490         struct btrfs_root *root;
2491         struct btrfs_root *fs_root = NULL;
2492         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2493         int index = 0;
2494
2495         next = node;
2496         while (1) {
2497                 cond_resched();
2498                 next = walk_up_backref(next, edges, &index);
2499                 root = next->root;
2500                 BUG_ON(!root);
2501
2502                 /* no other choice for non-references counted tree */
2503                 if (!root->ref_cows)
2504                         return root;
2505
2506                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2507                         fs_root = root;
2508
2509                 if (next != node)
2510                         return NULL;
2511
2512                 next = walk_down_backref(edges, &index);
2513                 if (!next || next->level <= node->level)
2514                         break;
2515         }
2516
2517         if (!fs_root)
2518                 return ERR_PTR(-ENOENT);
2519         return fs_root;
2520 }
2521
2522 static noinline_for_stack
2523 u64 calcu_metadata_size(struct reloc_control *rc,
2524                         struct backref_node *node, int reserve)
2525 {
2526         struct backref_node *next = node;
2527         struct backref_edge *edge;
2528         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2529         u64 num_bytes = 0;
2530         int index = 0;
2531
2532         BUG_ON(reserve && node->processed);
2533
2534         while (next) {
2535                 cond_resched();
2536                 while (1) {
2537                         if (next->processed && (reserve || next != node))
2538                                 break;
2539
2540                         num_bytes += btrfs_level_size(rc->extent_root,
2541                                                       next->level);
2542
2543                         if (list_empty(&next->upper))
2544                                 break;
2545
2546                         edge = list_entry(next->upper.next,
2547                                           struct backref_edge, list[LOWER]);
2548                         edges[index++] = edge;
2549                         next = edge->node[UPPER];
2550                 }
2551                 next = walk_down_backref(edges, &index);
2552         }
2553         return num_bytes;
2554 }
2555
2556 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2557                                   struct reloc_control *rc,
2558                                   struct backref_node *node)
2559 {
2560         struct btrfs_root *root = rc->extent_root;
2561         u64 num_bytes;
2562         int ret;
2563
2564         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2565
2566         trans->block_rsv = rc->block_rsv;
2567         ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2568                                   BTRFS_RESERVE_FLUSH_ALL);
2569         if (ret) {
2570                 if (ret == -EAGAIN)
2571                         rc->commit_transaction = 1;
2572                 return ret;
2573         }
2574
2575         return 0;
2576 }
2577
2578 static void release_metadata_space(struct reloc_control *rc,
2579                                    struct backref_node *node)
2580 {
2581         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2582         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2583 }
2584
2585 /*
2586  * relocate a block tree, and then update pointers in upper level
2587  * blocks that reference the block to point to the new location.
2588  *
2589  * if called by link_to_upper, the block has already been relocated.
2590  * in that case this function just updates pointers.
2591  */
2592 static int do_relocation(struct btrfs_trans_handle *trans,
2593                          struct reloc_control *rc,
2594                          struct backref_node *node,
2595                          struct btrfs_key *key,
2596                          struct btrfs_path *path, int lowest)
2597 {
2598         struct backref_node *upper;
2599         struct backref_edge *edge;
2600         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2601         struct btrfs_root *root;
2602         struct extent_buffer *eb;
2603         u32 blocksize;
2604         u64 bytenr;
2605         u64 generation;
2606         int nr;
2607         int slot;
2608         int ret;
2609         int err = 0;
2610
2611         BUG_ON(lowest && node->eb);
2612
2613         path->lowest_level = node->level + 1;
2614         rc->backref_cache.path[node->level] = node;
2615         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2616                 cond_resched();
2617
2618                 upper = edge->node[UPPER];
2619                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2620                 BUG_ON(!root);
2621
2622                 if (upper->eb && !upper->locked) {
2623                         if (!lowest) {
2624                                 ret = btrfs_bin_search(upper->eb, key,
2625                                                        upper->level, &slot);
2626                                 BUG_ON(ret);
2627                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2628                                 if (node->eb->start == bytenr)
2629                                         goto next;
2630                         }
2631                         drop_node_buffer(upper);
2632                 }
2633
2634                 if (!upper->eb) {
2635                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2636                         if (ret < 0) {
2637                                 err = ret;
2638                                 break;
2639                         }
2640                         BUG_ON(ret > 0);
2641
2642                         if (!upper->eb) {
2643                                 upper->eb = path->nodes[upper->level];
2644                                 path->nodes[upper->level] = NULL;
2645                         } else {
2646                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2647                         }
2648
2649                         upper->locked = 1;
2650                         path->locks[upper->level] = 0;
2651
2652                         slot = path->slots[upper->level];
2653                         btrfs_release_path(path);
2654                 } else {
2655                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2656                                                &slot);
2657                         BUG_ON(ret);
2658                 }
2659
2660                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2661                 if (lowest) {
2662                         BUG_ON(bytenr != node->bytenr);
2663                 } else {
2664                         if (node->eb->start == bytenr)
2665                                 goto next;
2666                 }
2667
2668                 blocksize = btrfs_level_size(root, node->level);
2669                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2670                 eb = read_tree_block(root, bytenr, blocksize, generation);
2671                 if (!eb || !extent_buffer_uptodate(eb)) {
2672                         free_extent_buffer(eb);
2673                         err = -EIO;
2674                         goto next;
2675                 }
2676                 btrfs_tree_lock(eb);
2677                 btrfs_set_lock_blocking(eb);
2678
2679                 if (!node->eb) {
2680                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2681                                               slot, &eb);
2682                         btrfs_tree_unlock(eb);
2683                         free_extent_buffer(eb);
2684                         if (ret < 0) {
2685                                 err = ret;
2686                                 goto next;
2687                         }
2688                         BUG_ON(node->eb != eb);
2689                 } else {
2690                         btrfs_set_node_blockptr(upper->eb, slot,
2691                                                 node->eb->start);
2692                         btrfs_set_node_ptr_generation(upper->eb, slot,
2693                                                       trans->transid);
2694                         btrfs_mark_buffer_dirty(upper->eb);
2695
2696                         ret = btrfs_inc_extent_ref(trans, root,
2697                                                 node->eb->start, blocksize,
2698                                                 upper->eb->start,
2699                                                 btrfs_header_owner(upper->eb),
2700                                                 node->level, 0, 1);
2701                         BUG_ON(ret);
2702
2703                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2704                         BUG_ON(ret);
2705                 }
2706 next:
2707                 if (!upper->pending)
2708                         drop_node_buffer(upper);
2709                 else
2710                         unlock_node_buffer(upper);
2711                 if (err)
2712                         break;
2713         }
2714
2715         if (!err && node->pending) {
2716                 drop_node_buffer(node);
2717                 list_move_tail(&node->list, &rc->backref_cache.changed);
2718                 node->pending = 0;
2719         }
2720
2721         path->lowest_level = 0;
2722         BUG_ON(err == -ENOSPC);
2723         return err;
2724 }
2725
2726 static int link_to_upper(struct btrfs_trans_handle *trans,
2727                          struct reloc_control *rc,
2728                          struct backref_node *node,
2729                          struct btrfs_path *path)
2730 {
2731         struct btrfs_key key;
2732
2733         btrfs_node_key_to_cpu(node->eb, &key, 0);
2734         return do_relocation(trans, rc, node, &key, path, 0);
2735 }
2736
2737 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2738                                 struct reloc_control *rc,
2739                                 struct btrfs_path *path, int err)
2740 {
2741         LIST_HEAD(list);
2742         struct backref_cache *cache = &rc->backref_cache;
2743         struct backref_node *node;
2744         int level;
2745         int ret;
2746
2747         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2748                 while (!list_empty(&cache->pending[level])) {
2749                         node = list_entry(cache->pending[level].next,
2750                                           struct backref_node, list);
2751                         list_move_tail(&node->list, &list);
2752                         BUG_ON(!node->pending);
2753
2754                         if (!err) {
2755                                 ret = link_to_upper(trans, rc, node, path);
2756                                 if (ret < 0)
2757                                         err = ret;
2758                         }
2759                 }
2760                 list_splice_init(&list, &cache->pending[level]);
2761         }
2762         return err;
2763 }
2764
2765 static void mark_block_processed(struct reloc_control *rc,
2766                                  u64 bytenr, u32 blocksize)
2767 {
2768         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2769                         EXTENT_DIRTY, GFP_NOFS);
2770 }
2771
2772 static void __mark_block_processed(struct reloc_control *rc,
2773                                    struct backref_node *node)
2774 {
2775         u32 blocksize;
2776         if (node->level == 0 ||
2777             in_block_group(node->bytenr, rc->block_group)) {
2778                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2779                 mark_block_processed(rc, node->bytenr, blocksize);
2780         }
2781         node->processed = 1;
2782 }
2783
2784 /*
2785  * mark a block and all blocks directly/indirectly reference the block
2786  * as processed.
2787  */
2788 static void update_processed_blocks(struct reloc_control *rc,
2789                                     struct backref_node *node)
2790 {
2791         struct backref_node *next = node;
2792         struct backref_edge *edge;
2793         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2794         int index = 0;
2795
2796         while (next) {
2797                 cond_resched();
2798                 while (1) {
2799                         if (next->processed)
2800                                 break;
2801
2802                         __mark_block_processed(rc, next);
2803
2804                         if (list_empty(&next->upper))
2805                                 break;
2806
2807                         edge = list_entry(next->upper.next,
2808                                           struct backref_edge, list[LOWER]);
2809                         edges[index++] = edge;
2810                         next = edge->node[UPPER];
2811                 }
2812                 next = walk_down_backref(edges, &index);
2813         }
2814 }
2815
2816 static int tree_block_processed(u64 bytenr, u32 blocksize,
2817                                 struct reloc_control *rc)
2818 {
2819         if (test_range_bit(&rc->processed_blocks, bytenr,
2820                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2821                 return 1;
2822         return 0;
2823 }
2824
2825 static int get_tree_block_key(struct reloc_control *rc,
2826                               struct tree_block *block)
2827 {
2828         struct extent_buffer *eb;
2829
2830         BUG_ON(block->key_ready);
2831         eb = read_tree_block(rc->extent_root, block->bytenr,
2832                              block->key.objectid, block->key.offset);
2833         if (!eb || !extent_buffer_uptodate(eb)) {
2834                 free_extent_buffer(eb);
2835                 return -EIO;
2836         }
2837         WARN_ON(btrfs_header_level(eb) != block->level);
2838         if (block->level == 0)
2839                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2840         else
2841                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2842         free_extent_buffer(eb);
2843         block->key_ready = 1;
2844         return 0;
2845 }
2846
2847 static int reada_tree_block(struct reloc_control *rc,
2848                             struct tree_block *block)
2849 {
2850         BUG_ON(block->key_ready);
2851         if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2852                 readahead_tree_block(rc->extent_root, block->bytenr,
2853                                      block->key.objectid,
2854                                      rc->extent_root->leafsize);
2855         else
2856                 readahead_tree_block(rc->extent_root, block->bytenr,
2857                                      block->key.objectid, block->key.offset);
2858         return 0;
2859 }
2860
2861 /*
2862  * helper function to relocate a tree block
2863  */
2864 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2865                                 struct reloc_control *rc,
2866                                 struct backref_node *node,
2867                                 struct btrfs_key *key,
2868                                 struct btrfs_path *path)
2869 {
2870         struct btrfs_root *root;
2871         int release = 0;
2872         int ret = 0;
2873
2874         if (!node)
2875                 return 0;
2876
2877         BUG_ON(node->processed);
2878         root = select_one_root(trans, node);
2879         if (root == ERR_PTR(-ENOENT)) {
2880                 update_processed_blocks(rc, node);
2881                 goto out;
2882         }
2883
2884         if (!root || root->ref_cows) {
2885                 ret = reserve_metadata_space(trans, rc, node);
2886                 if (ret)
2887                         goto out;
2888                 release = 1;
2889         }
2890
2891         if (root) {
2892                 if (root->ref_cows) {
2893                         BUG_ON(node->new_bytenr);
2894                         BUG_ON(!list_empty(&node->list));
2895                         btrfs_record_root_in_trans(trans, root);
2896                         root = root->reloc_root;
2897                         node->new_bytenr = root->node->start;
2898                         node->root = root;
2899                         list_add_tail(&node->list, &rc->backref_cache.changed);
2900                 } else {
2901                         path->lowest_level = node->level;
2902                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2903                         btrfs_release_path(path);
2904                         if (ret > 0)
2905                                 ret = 0;
2906                 }
2907                 if (!ret)
2908                         update_processed_blocks(rc, node);
2909         } else {
2910                 ret = do_relocation(trans, rc, node, key, path, 1);
2911         }
2912 out:
2913         if (ret || node->level == 0 || node->cowonly) {
2914                 if (release)
2915                         release_metadata_space(rc, node);
2916                 remove_backref_node(&rc->backref_cache, node);
2917         }
2918         return ret;
2919 }
2920
2921 /*
2922  * relocate a list of blocks
2923  */
2924 static noinline_for_stack
2925 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2926                          struct reloc_control *rc, struct rb_root *blocks)
2927 {
2928         struct backref_node *node;
2929         struct btrfs_path *path;
2930         struct tree_block *block;
2931         struct rb_node *rb_node;
2932         int ret;
2933         int err = 0;
2934
2935         path = btrfs_alloc_path();
2936         if (!path) {
2937                 err = -ENOMEM;
2938                 goto out_free_blocks;
2939         }
2940
2941         rb_node = rb_first(blocks);
2942         while (rb_node) {
2943                 block = rb_entry(rb_node, struct tree_block, rb_node);
2944                 if (!block->key_ready)
2945                         reada_tree_block(rc, block);
2946                 rb_node = rb_next(rb_node);
2947         }
2948
2949         rb_node = rb_first(blocks);
2950         while (rb_node) {
2951                 block = rb_entry(rb_node, struct tree_block, rb_node);
2952                 if (!block->key_ready) {
2953                         err = get_tree_block_key(rc, block);
2954                         if (err)
2955                                 goto out_free_path;
2956                 }
2957                 rb_node = rb_next(rb_node);
2958         }
2959
2960         rb_node = rb_first(blocks);
2961         while (rb_node) {
2962                 block = rb_entry(rb_node, struct tree_block, rb_node);
2963
2964                 node = build_backref_tree(rc, &block->key,
2965                                           block->level, block->bytenr);
2966                 if (IS_ERR(node)) {
2967                         err = PTR_ERR(node);
2968                         goto out;
2969                 }
2970
2971                 ret = relocate_tree_block(trans, rc, node, &block->key,
2972                                           path);
2973                 if (ret < 0) {
2974                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2975                                 err = ret;
2976                         goto out;
2977                 }
2978                 rb_node = rb_next(rb_node);
2979         }
2980 out:
2981         err = finish_pending_nodes(trans, rc, path, err);
2982
2983 out_free_path:
2984         btrfs_free_path(path);
2985 out_free_blocks:
2986         free_block_list(blocks);
2987         return err;
2988 }
2989
2990 static noinline_for_stack
2991 int prealloc_file_extent_cluster(struct inode *inode,
2992                                  struct file_extent_cluster *cluster)
2993 {
2994         u64 alloc_hint = 0;
2995         u64 start;
2996         u64 end;
2997         u64 offset = BTRFS_I(inode)->index_cnt;
2998         u64 num_bytes;
2999         int nr = 0;
3000         int ret = 0;
3001
3002         BUG_ON(cluster->start != cluster->boundary[0]);
3003         mutex_lock(&inode->i_mutex);
3004
3005         ret = btrfs_check_data_free_space(inode, cluster->end +
3006                                           1 - cluster->start);
3007         if (ret)
3008                 goto out;
3009
3010         while (nr < cluster->nr) {
3011                 start = cluster->boundary[nr] - offset;
3012                 if (nr + 1 < cluster->nr)
3013                         end = cluster->boundary[nr + 1] - 1 - offset;
3014                 else
3015                         end = cluster->end - offset;
3016
3017                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3018                 num_bytes = end + 1 - start;
3019                 ret = btrfs_prealloc_file_range(inode, 0, start,
3020                                                 num_bytes, num_bytes,
3021                                                 end + 1, &alloc_hint);
3022                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3023                 if (ret)
3024                         break;
3025                 nr++;
3026         }
3027         btrfs_free_reserved_data_space(inode, cluster->end +
3028                                        1 - cluster->start);
3029 out:
3030         mutex_unlock(&inode->i_mutex);
3031         return ret;
3032 }
3033
3034 static noinline_for_stack
3035 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3036                          u64 block_start)
3037 {
3038         struct btrfs_root *root = BTRFS_I(inode)->root;
3039         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3040         struct extent_map *em;
3041         int ret = 0;
3042
3043         em = alloc_extent_map();
3044         if (!em)
3045                 return -ENOMEM;
3046
3047         em->start = start;
3048         em->len = end + 1 - start;
3049         em->block_len = em->len;
3050         em->block_start = block_start;
3051         em->bdev = root->fs_info->fs_devices->latest_bdev;
3052         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3053
3054         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3055         while (1) {
3056                 write_lock(&em_tree->lock);
3057                 ret = add_extent_mapping(em_tree, em, 0);
3058                 write_unlock(&em_tree->lock);
3059                 if (ret != -EEXIST) {
3060                         free_extent_map(em);
3061                         break;
3062                 }
3063                 btrfs_drop_extent_cache(inode, start, end, 0);
3064         }
3065         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3066         return ret;
3067 }
3068
3069 static int relocate_file_extent_cluster(struct inode *inode,
3070                                         struct file_extent_cluster *cluster)
3071 {
3072         u64 page_start;
3073         u64 page_end;
3074         u64 offset = BTRFS_I(inode)->index_cnt;
3075         unsigned long index;
3076         unsigned long last_index;
3077         struct page *page;
3078         struct file_ra_state *ra;
3079         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3080         int nr = 0;
3081         int ret = 0;
3082
3083         if (!cluster->nr)
3084                 return 0;
3085
3086         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3087         if (!ra)
3088                 return -ENOMEM;
3089
3090         ret = prealloc_file_extent_cluster(inode, cluster);
3091         if (ret)
3092                 goto out;
3093
3094         file_ra_state_init(ra, inode->i_mapping);
3095
3096         ret = setup_extent_mapping(inode, cluster->start - offset,
3097                                    cluster->end - offset, cluster->start);
3098         if (ret)
3099                 goto out;
3100
3101         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3102         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3103         while (index <= last_index) {
3104                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3105                 if (ret)
3106                         goto out;
3107
3108                 page = find_lock_page(inode->i_mapping, index);
3109                 if (!page) {
3110                         page_cache_sync_readahead(inode->i_mapping,
3111                                                   ra, NULL, index,
3112                                                   last_index + 1 - index);
3113                         page = find_or_create_page(inode->i_mapping, index,
3114                                                    mask);
3115                         if (!page) {
3116                                 btrfs_delalloc_release_metadata(inode,
3117                                                         PAGE_CACHE_SIZE);
3118                                 ret = -ENOMEM;
3119                                 goto out;
3120                         }
3121                 }
3122
3123                 if (PageReadahead(page)) {
3124                         page_cache_async_readahead(inode->i_mapping,
3125                                                    ra, NULL, page, index,
3126                                                    last_index + 1 - index);
3127                 }
3128
3129                 if (!PageUptodate(page)) {
3130                         btrfs_readpage(NULL, page);
3131                         lock_page(page);
3132                         if (!PageUptodate(page)) {
3133                                 unlock_page(page);
3134                                 page_cache_release(page);
3135                                 btrfs_delalloc_release_metadata(inode,
3136                                                         PAGE_CACHE_SIZE);
3137                                 ret = -EIO;
3138                                 goto out;
3139                         }
3140                 }
3141
3142                 page_start = page_offset(page);
3143                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3144
3145                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3146
3147                 set_page_extent_mapped(page);
3148
3149                 if (nr < cluster->nr &&
3150                     page_start + offset == cluster->boundary[nr]) {
3151                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3152                                         page_start, page_end,
3153                                         EXTENT_BOUNDARY, GFP_NOFS);
3154                         nr++;
3155                 }
3156
3157                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3158                 set_page_dirty(page);
3159
3160                 unlock_extent(&BTRFS_I(inode)->io_tree,
3161                               page_start, page_end);
3162                 unlock_page(page);
3163                 page_cache_release(page);
3164
3165                 index++;
3166                 balance_dirty_pages_ratelimited(inode->i_mapping);
3167                 btrfs_throttle(BTRFS_I(inode)->root);
3168         }
3169         WARN_ON(nr != cluster->nr);
3170 out:
3171         kfree(ra);
3172         return ret;
3173 }
3174
3175 static noinline_for_stack
3176 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3177                          struct file_extent_cluster *cluster)
3178 {
3179         int ret;
3180
3181         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3182                 ret = relocate_file_extent_cluster(inode, cluster);
3183                 if (ret)
3184                         return ret;
3185                 cluster->nr = 0;
3186         }
3187
3188         if (!cluster->nr)
3189                 cluster->start = extent_key->objectid;
3190         else
3191                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3192         cluster->end = extent_key->objectid + extent_key->offset - 1;
3193         cluster->boundary[cluster->nr] = extent_key->objectid;
3194         cluster->nr++;
3195
3196         if (cluster->nr >= MAX_EXTENTS) {
3197                 ret = relocate_file_extent_cluster(inode, cluster);
3198                 if (ret)
3199                         return ret;
3200                 cluster->nr = 0;
3201         }
3202         return 0;
3203 }
3204
3205 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3206 static int get_ref_objectid_v0(struct reloc_control *rc,
3207                                struct btrfs_path *path,
3208                                struct btrfs_key *extent_key,
3209                                u64 *ref_objectid, int *path_change)
3210 {
3211         struct btrfs_key key;
3212         struct extent_buffer *leaf;
3213         struct btrfs_extent_ref_v0 *ref0;
3214         int ret;
3215         int slot;
3216
3217         leaf = path->nodes[0];
3218         slot = path->slots[0];
3219         while (1) {
3220                 if (slot >= btrfs_header_nritems(leaf)) {
3221                         ret = btrfs_next_leaf(rc->extent_root, path);
3222                         if (ret < 0)
3223                                 return ret;
3224                         BUG_ON(ret > 0);
3225                         leaf = path->nodes[0];
3226                         slot = path->slots[0];
3227                         if (path_change)
3228                                 *path_change = 1;
3229                 }
3230                 btrfs_item_key_to_cpu(leaf, &key, slot);
3231                 if (key.objectid != extent_key->objectid)
3232                         return -ENOENT;
3233
3234                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3235                         slot++;
3236                         continue;
3237                 }
3238                 ref0 = btrfs_item_ptr(leaf, slot,
3239                                 struct btrfs_extent_ref_v0);
3240                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3241                 break;
3242         }
3243         return 0;
3244 }
3245 #endif
3246
3247 /*
3248  * helper to add a tree block to the list.
3249  * the major work is getting the generation and level of the block
3250  */
3251 static int add_tree_block(struct reloc_control *rc,
3252                           struct btrfs_key *extent_key,
3253                           struct btrfs_path *path,
3254                           struct rb_root *blocks)
3255 {
3256         struct extent_buffer *eb;
3257         struct btrfs_extent_item *ei;
3258         struct btrfs_tree_block_info *bi;
3259         struct tree_block *block;
3260         struct rb_node *rb_node;
3261         u32 item_size;
3262         int level = -1;
3263         int generation;
3264
3265         eb =  path->nodes[0];
3266         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3267
3268         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3269             item_size >= sizeof(*ei) + sizeof(*bi)) {
3270                 ei = btrfs_item_ptr(eb, path->slots[0],
3271                                 struct btrfs_extent_item);
3272                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3273                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3274                         level = btrfs_tree_block_level(eb, bi);
3275                 } else {
3276                         level = (int)extent_key->offset;
3277                 }
3278                 generation = btrfs_extent_generation(eb, ei);
3279         } else {
3280 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3281                 u64 ref_owner;
3282                 int ret;
3283
3284                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3285                 ret = get_ref_objectid_v0(rc, path, extent_key,
3286                                           &ref_owner, NULL);
3287                 if (ret < 0)
3288                         return ret;
3289                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3290                 level = (int)ref_owner;
3291                 /* FIXME: get real generation */
3292                 generation = 0;
3293 #else
3294                 BUG();
3295 #endif
3296         }
3297
3298         btrfs_release_path(path);
3299
3300         BUG_ON(level == -1);
3301
3302         block = kmalloc(sizeof(*block), GFP_NOFS);
3303         if (!block)
3304                 return -ENOMEM;
3305
3306         block->bytenr = extent_key->objectid;
3307         block->key.objectid = rc->extent_root->leafsize;
3308         block->key.offset = generation;
3309         block->level = level;
3310         block->key_ready = 0;
3311
3312         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3313         if (rb_node)
3314                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3315
3316         return 0;
3317 }
3318
3319 /*
3320  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3321  */
3322 static int __add_tree_block(struct reloc_control *rc,
3323                             u64 bytenr, u32 blocksize,
3324                             struct rb_root *blocks)
3325 {
3326         struct btrfs_path *path;
3327         struct btrfs_key key;
3328         int ret;
3329         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3330                                         SKINNY_METADATA);
3331
3332         if (tree_block_processed(bytenr, blocksize, rc))
3333                 return 0;
3334
3335         if (tree_search(blocks, bytenr))
3336                 return 0;
3337
3338         path = btrfs_alloc_path();
3339         if (!path)
3340                 return -ENOMEM;
3341 again:
3342         key.objectid = bytenr;
3343         if (skinny) {
3344                 key.type = BTRFS_METADATA_ITEM_KEY;
3345                 key.offset = (u64)-1;
3346         } else {
3347                 key.type = BTRFS_EXTENT_ITEM_KEY;
3348                 key.offset = blocksize;
3349         }
3350
3351         path->search_commit_root = 1;
3352         path->skip_locking = 1;
3353         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3354         if (ret < 0)
3355                 goto out;
3356
3357         if (ret > 0 && skinny) {
3358                 if (path->slots[0]) {
3359                         path->slots[0]--;
3360                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3361                                               path->slots[0]);
3362                         if (key.objectid == bytenr &&
3363                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3364                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3365                               key.offset == blocksize)))
3366                                 ret = 0;
3367                 }
3368
3369                 if (ret) {
3370                         skinny = false;
3371                         btrfs_release_path(path);
3372                         goto again;
3373                 }
3374         }
3375         BUG_ON(ret);
3376
3377         ret = add_tree_block(rc, &key, path, blocks);
3378 out:
3379         btrfs_free_path(path);
3380         return ret;
3381 }
3382
3383 /*
3384  * helper to check if the block use full backrefs for pointers in it
3385  */
3386 static int block_use_full_backref(struct reloc_control *rc,
3387                                   struct extent_buffer *eb)
3388 {
3389         u64 flags;
3390         int ret;
3391
3392         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3393             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3394                 return 1;
3395
3396         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3397                                        eb->start, btrfs_header_level(eb), 1,
3398                                        NULL, &flags);
3399         BUG_ON(ret);
3400
3401         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3402                 ret = 1;
3403         else
3404                 ret = 0;
3405         return ret;
3406 }
3407
3408 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3409                                     struct inode *inode, u64 ino)
3410 {
3411         struct btrfs_key key;
3412         struct btrfs_root *root = fs_info->tree_root;
3413         struct btrfs_trans_handle *trans;
3414         int ret = 0;
3415
3416         if (inode)
3417                 goto truncate;
3418
3419         key.objectid = ino;
3420         key.type = BTRFS_INODE_ITEM_KEY;
3421         key.offset = 0;
3422
3423         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3424         if (IS_ERR(inode) || is_bad_inode(inode)) {
3425                 if (!IS_ERR(inode))
3426                         iput(inode);
3427                 return -ENOENT;
3428         }
3429
3430 truncate:
3431         ret = btrfs_check_trunc_cache_free_space(root,
3432                                                  &fs_info->global_block_rsv);
3433         if (ret)
3434                 goto out;
3435
3436         trans = btrfs_join_transaction(root);
3437         if (IS_ERR(trans)) {
3438                 ret = PTR_ERR(trans);
3439                 goto out;
3440         }
3441
3442         ret = btrfs_truncate_free_space_cache(root, trans, inode);
3443
3444         btrfs_end_transaction(trans, root);
3445         btrfs_btree_balance_dirty(root);
3446 out:
3447         iput(inode);
3448         return ret;
3449 }
3450
3451 /*
3452  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3453  * this function scans fs tree to find blocks reference the data extent
3454  */
3455 static int find_data_references(struct reloc_control *rc,
3456                                 struct btrfs_key *extent_key,
3457                                 struct extent_buffer *leaf,
3458                                 struct btrfs_extent_data_ref *ref,
3459                                 struct rb_root *blocks)
3460 {
3461         struct btrfs_path *path;
3462         struct tree_block *block;
3463         struct btrfs_root *root;
3464         struct btrfs_file_extent_item *fi;
3465         struct rb_node *rb_node;
3466         struct btrfs_key key;
3467         u64 ref_root;
3468         u64 ref_objectid;
3469         u64 ref_offset;
3470         u32 ref_count;
3471         u32 nritems;
3472         int err = 0;
3473         int added = 0;
3474         int counted;
3475         int ret;
3476
3477         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3478         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3479         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3480         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3481
3482         /*
3483          * This is an extent belonging to the free space cache, lets just delete
3484          * it and redo the search.
3485          */
3486         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3487                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3488                                                NULL, ref_objectid);
3489                 if (ret != -ENOENT)
3490                         return ret;
3491                 ret = 0;
3492         }
3493
3494         path = btrfs_alloc_path();
3495         if (!path)
3496                 return -ENOMEM;
3497         path->reada = 1;
3498
3499         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3500         if (IS_ERR(root)) {
3501                 err = PTR_ERR(root);
3502                 goto out;
3503         }
3504
3505         key.objectid = ref_objectid;
3506         key.type = BTRFS_EXTENT_DATA_KEY;
3507         if (ref_offset > ((u64)-1 << 32))
3508                 key.offset = 0;
3509         else
3510                 key.offset = ref_offset;
3511
3512         path->search_commit_root = 1;
3513         path->skip_locking = 1;
3514         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3515         if (ret < 0) {
3516                 err = ret;
3517                 goto out;
3518         }
3519
3520         leaf = path->nodes[0];
3521         nritems = btrfs_header_nritems(leaf);
3522         /*
3523          * the references in tree blocks that use full backrefs
3524          * are not counted in
3525          */
3526         if (block_use_full_backref(rc, leaf))
3527                 counted = 0;
3528         else
3529                 counted = 1;
3530         rb_node = tree_search(blocks, leaf->start);
3531         if (rb_node) {
3532                 if (counted)
3533                         added = 1;
3534                 else
3535                         path->slots[0] = nritems;
3536         }
3537
3538         while (ref_count > 0) {
3539                 while (path->slots[0] >= nritems) {
3540                         ret = btrfs_next_leaf(root, path);
3541                         if (ret < 0) {
3542                                 err = ret;
3543                                 goto out;
3544                         }
3545                         if (ret > 0) {
3546                                 WARN_ON(1);
3547                                 goto out;
3548                         }
3549
3550                         leaf = path->nodes[0];
3551                         nritems = btrfs_header_nritems(leaf);
3552                         added = 0;
3553
3554                         if (block_use_full_backref(rc, leaf))
3555                                 counted = 0;
3556                         else
3557                                 counted = 1;
3558                         rb_node = tree_search(blocks, leaf->start);
3559                         if (rb_node) {
3560                                 if (counted)
3561                                         added = 1;
3562                                 else
3563                                         path->slots[0] = nritems;
3564                         }
3565                 }
3566
3567                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3568                 if (key.objectid != ref_objectid ||
3569                     key.type != BTRFS_EXTENT_DATA_KEY) {
3570                         WARN_ON(1);
3571                         break;
3572                 }
3573
3574                 fi = btrfs_item_ptr(leaf, path->slots[0],
3575                                     struct btrfs_file_extent_item);
3576
3577                 if (btrfs_file_extent_type(leaf, fi) ==
3578                     BTRFS_FILE_EXTENT_INLINE)
3579                         goto next;
3580
3581                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3582                     extent_key->objectid)
3583                         goto next;
3584
3585                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3586                 if (key.offset != ref_offset)
3587                         goto next;
3588
3589                 if (counted)
3590                         ref_count--;
3591                 if (added)
3592                         goto next;
3593
3594                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3595                         block = kmalloc(sizeof(*block), GFP_NOFS);
3596                         if (!block) {
3597                                 err = -ENOMEM;
3598                                 break;
3599                         }
3600                         block->bytenr = leaf->start;
3601                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3602                         block->level = 0;
3603                         block->key_ready = 1;
3604                         rb_node = tree_insert(blocks, block->bytenr,
3605                                               &block->rb_node);
3606                         if (rb_node)
3607                                 backref_tree_panic(rb_node, -EEXIST,
3608                                                    block->bytenr);
3609                 }
3610                 if (counted)
3611                         added = 1;
3612                 else
3613                         path->slots[0] = nritems;
3614 next:
3615                 path->slots[0]++;
3616
3617         }
3618 out:
3619         btrfs_free_path(path);
3620         return err;
3621 }
3622
3623 /*
3624  * helper to find all tree blocks that reference a given data extent
3625  */
3626 static noinline_for_stack
3627 int add_data_references(struct reloc_control *rc,
3628                         struct btrfs_key *extent_key,
3629                         struct btrfs_path *path,
3630                         struct rb_root *blocks)
3631 {
3632         struct btrfs_key key;
3633         struct extent_buffer *eb;
3634         struct btrfs_extent_data_ref *dref;
3635         struct btrfs_extent_inline_ref *iref;
3636         unsigned long ptr;
3637         unsigned long end;
3638         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3639         int ret = 0;
3640         int err = 0;
3641
3642         eb = path->nodes[0];
3643         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3644         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3645 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3646         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3647                 ptr = end;
3648         else
3649 #endif
3650                 ptr += sizeof(struct btrfs_extent_item);
3651
3652         while (ptr < end) {
3653                 iref = (struct btrfs_extent_inline_ref *)ptr;
3654                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3655                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3656                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3657                         ret = __add_tree_block(rc, key.offset, blocksize,
3658                                                blocks);
3659                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3660                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3661                         ret = find_data_references(rc, extent_key,
3662                                                    eb, dref, blocks);
3663                 } else {
3664                         BUG();
3665                 }
3666                 if (ret) {
3667                         err = ret;
3668                         goto out;
3669                 }
3670                 ptr += btrfs_extent_inline_ref_size(key.type);
3671         }
3672         WARN_ON(ptr > end);
3673
3674         while (1) {
3675                 cond_resched();
3676                 eb = path->nodes[0];
3677                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3678                         ret = btrfs_next_leaf(rc->extent_root, path);
3679                         if (ret < 0) {
3680                                 err = ret;
3681                                 break;
3682                         }
3683                         if (ret > 0)
3684                                 break;
3685                         eb = path->nodes[0];
3686                 }
3687
3688                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3689                 if (key.objectid != extent_key->objectid)
3690                         break;
3691
3692 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3693                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3694                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3695 #else
3696                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3697                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3698 #endif
3699                         ret = __add_tree_block(rc, key.offset, blocksize,
3700                                                blocks);
3701                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3702                         dref = btrfs_item_ptr(eb, path->slots[0],
3703                                               struct btrfs_extent_data_ref);
3704                         ret = find_data_references(rc, extent_key,
3705                                                    eb, dref, blocks);
3706                 } else {
3707                         ret = 0;
3708                 }
3709                 if (ret) {
3710                         err = ret;
3711                         break;
3712                 }
3713                 path->slots[0]++;
3714         }
3715 out:
3716         btrfs_release_path(path);
3717         if (err)
3718                 free_block_list(blocks);
3719         return err;
3720 }
3721
3722 /*
3723  * helper to find next unprocessed extent
3724  */
3725 static noinline_for_stack
3726 int find_next_extent(struct btrfs_trans_handle *trans,
3727                      struct reloc_control *rc, struct btrfs_path *path,
3728                      struct btrfs_key *extent_key)
3729 {
3730         struct btrfs_key key;
3731         struct extent_buffer *leaf;
3732         u64 start, end, last;
3733         int ret;
3734
3735         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3736         while (1) {
3737                 cond_resched();
3738                 if (rc->search_start >= last) {
3739                         ret = 1;
3740                         break;
3741                 }
3742
3743                 key.objectid = rc->search_start;
3744                 key.type = BTRFS_EXTENT_ITEM_KEY;
3745                 key.offset = 0;
3746
3747                 path->search_commit_root = 1;
3748                 path->skip_locking = 1;
3749                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3750                                         0, 0);
3751                 if (ret < 0)
3752                         break;
3753 next:
3754                 leaf = path->nodes[0];
3755                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3756                         ret = btrfs_next_leaf(rc->extent_root, path);
3757                         if (ret != 0)
3758                                 break;
3759                         leaf = path->nodes[0];
3760                 }
3761
3762                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3763                 if (key.objectid >= last) {
3764                         ret = 1;
3765                         break;
3766                 }
3767
3768                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3769                     key.type != BTRFS_METADATA_ITEM_KEY) {
3770                         path->slots[0]++;
3771                         goto next;
3772                 }
3773
3774                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3775                     key.objectid + key.offset <= rc->search_start) {
3776                         path->slots[0]++;
3777                         goto next;
3778                 }
3779
3780                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3781                     key.objectid + rc->extent_root->leafsize <=
3782                     rc->search_start) {
3783                         path->slots[0]++;
3784                         goto next;
3785                 }
3786
3787                 ret = find_first_extent_bit(&rc->processed_blocks,
3788                                             key.objectid, &start, &end,
3789                                             EXTENT_DIRTY, NULL);
3790
3791                 if (ret == 0 && start <= key.objectid) {
3792                         btrfs_release_path(path);
3793                         rc->search_start = end + 1;
3794                 } else {
3795                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3796                                 rc->search_start = key.objectid + key.offset;
3797                         else
3798                                 rc->search_start = key.objectid +
3799                                         rc->extent_root->leafsize;
3800                         memcpy(extent_key, &key, sizeof(key));
3801                         return 0;
3802                 }
3803         }
3804         btrfs_release_path(path);
3805         return ret;
3806 }
3807
3808 static void set_reloc_control(struct reloc_control *rc)
3809 {
3810         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3811
3812         mutex_lock(&fs_info->reloc_mutex);
3813         fs_info->reloc_ctl = rc;
3814         mutex_unlock(&fs_info->reloc_mutex);
3815 }
3816
3817 static void unset_reloc_control(struct reloc_control *rc)
3818 {
3819         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3820
3821         mutex_lock(&fs_info->reloc_mutex);
3822         fs_info->reloc_ctl = NULL;
3823         mutex_unlock(&fs_info->reloc_mutex);
3824 }
3825
3826 static int check_extent_flags(u64 flags)
3827 {
3828         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3829             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3830                 return 1;
3831         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3832             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3833                 return 1;
3834         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3835             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3836                 return 1;
3837         return 0;
3838 }
3839
3840 static noinline_for_stack
3841 int prepare_to_relocate(struct reloc_control *rc)
3842 {
3843         struct btrfs_trans_handle *trans;
3844         int ret;
3845
3846         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3847                                               BTRFS_BLOCK_RSV_TEMP);
3848         if (!rc->block_rsv)
3849                 return -ENOMEM;
3850
3851         /*
3852          * reserve some space for creating reloc trees.
3853          * btrfs_init_reloc_root will use them when there
3854          * is no reservation in transaction handle.
3855          */
3856         ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3857                                   rc->extent_root->nodesize * 256,
3858                                   BTRFS_RESERVE_FLUSH_ALL);
3859         if (ret)
3860                 return ret;
3861
3862         memset(&rc->cluster, 0, sizeof(rc->cluster));
3863         rc->search_start = rc->block_group->key.objectid;
3864         rc->extents_found = 0;
3865         rc->nodes_relocated = 0;
3866         rc->merging_rsv_size = 0;
3867
3868         rc->create_reloc_tree = 1;
3869         set_reloc_control(rc);
3870
3871         trans = btrfs_join_transaction(rc->extent_root);
3872         if (IS_ERR(trans)) {
3873                 unset_reloc_control(rc);
3874                 /*
3875                  * extent tree is not a ref_cow tree and has no reloc_root to
3876                  * cleanup.  And callers are responsible to free the above
3877                  * block rsv.
3878                  */
3879                 return PTR_ERR(trans);
3880         }
3881         btrfs_commit_transaction(trans, rc->extent_root);
3882         return 0;
3883 }
3884
3885 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3886 {
3887         struct rb_root blocks = RB_ROOT;
3888         struct btrfs_key key;
3889         struct btrfs_trans_handle *trans = NULL;
3890         struct btrfs_path *path;
3891         struct btrfs_extent_item *ei;
3892         u64 flags;
3893         u32 item_size;
3894         int ret;
3895         int err = 0;
3896         int progress = 0;
3897
3898         path = btrfs_alloc_path();
3899         if (!path)
3900                 return -ENOMEM;
3901         path->reada = 1;
3902
3903         ret = prepare_to_relocate(rc);
3904         if (ret) {
3905                 err = ret;
3906                 goto out_free;
3907         }
3908
3909         while (1) {
3910                 progress++;
3911                 trans = btrfs_start_transaction(rc->extent_root, 0);
3912                 if (IS_ERR(trans)) {
3913                         err = PTR_ERR(trans);
3914                         trans = NULL;
3915                         break;
3916                 }
3917 restart:
3918                 if (update_backref_cache(trans, &rc->backref_cache)) {
3919                         btrfs_end_transaction(trans, rc->extent_root);
3920                         continue;
3921                 }
3922
3923                 ret = find_next_extent(trans, rc, path, &key);
3924                 if (ret < 0)
3925                         err = ret;
3926                 if (ret != 0)
3927                         break;
3928
3929                 rc->extents_found++;
3930
3931                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3932                                     struct btrfs_extent_item);
3933                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3934                 if (item_size >= sizeof(*ei)) {
3935                         flags = btrfs_extent_flags(path->nodes[0], ei);
3936                         ret = check_extent_flags(flags);
3937                         BUG_ON(ret);
3938
3939                 } else {
3940 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3941                         u64 ref_owner;
3942                         int path_change = 0;
3943
3944                         BUG_ON(item_size !=
3945                                sizeof(struct btrfs_extent_item_v0));
3946                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3947                                                   &path_change);
3948                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3949                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3950                         else
3951                                 flags = BTRFS_EXTENT_FLAG_DATA;
3952
3953                         if (path_change) {
3954                                 btrfs_release_path(path);
3955
3956                                 path->search_commit_root = 1;
3957                                 path->skip_locking = 1;
3958                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3959                                                         &key, path, 0, 0);
3960                                 if (ret < 0) {
3961                                         err = ret;
3962                                         break;
3963                                 }
3964                                 BUG_ON(ret > 0);
3965                         }
3966 #else
3967                         BUG();
3968 #endif
3969                 }
3970
3971                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3972                         ret = add_tree_block(rc, &key, path, &blocks);
3973                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3974                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3975                         ret = add_data_references(rc, &key, path, &blocks);
3976                 } else {
3977                         btrfs_release_path(path);
3978                         ret = 0;
3979                 }
3980                 if (ret < 0) {
3981                         err = ret;
3982                         break;
3983                 }
3984
3985                 if (!RB_EMPTY_ROOT(&blocks)) {
3986                         ret = relocate_tree_blocks(trans, rc, &blocks);
3987                         if (ret < 0) {
3988                                 if (ret != -EAGAIN) {
3989                                         err = ret;
3990                                         break;
3991                                 }
3992                                 rc->extents_found--;
3993                                 rc->search_start = key.objectid;
3994                         }
3995                 }
3996
3997                 ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5);
3998                 if (ret < 0) {
3999                         if (ret != -ENOSPC) {
4000                                 err = ret;
4001                                 WARN_ON(1);
4002                                 break;
4003                         }
4004                         rc->commit_transaction = 1;
4005                 }
4006
4007                 if (rc->commit_transaction) {
4008                         rc->commit_transaction = 0;
4009                         ret = btrfs_commit_transaction(trans, rc->extent_root);
4010                         BUG_ON(ret);
4011                 } else {
4012                         btrfs_end_transaction_throttle(trans, rc->extent_root);
4013                         btrfs_btree_balance_dirty(rc->extent_root);
4014                 }
4015                 trans = NULL;
4016
4017                 if (rc->stage == MOVE_DATA_EXTENTS &&
4018                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4019                         rc->found_file_extent = 1;
4020                         ret = relocate_data_extent(rc->data_inode,
4021                                                    &key, &rc->cluster);
4022                         if (ret < 0) {
4023                                 err = ret;
4024                                 break;
4025                         }
4026                 }
4027         }
4028         if (trans && progress && err == -ENOSPC) {
4029                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4030                                               rc->block_group->flags);
4031                 if (ret == 0) {
4032                         err = 0;
4033                         progress = 0;
4034                         goto restart;
4035                 }
4036         }
4037
4038         btrfs_release_path(path);
4039         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4040                           GFP_NOFS);
4041
4042         if (trans) {
4043                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4044                 btrfs_btree_balance_dirty(rc->extent_root);
4045         }
4046
4047         if (!err) {
4048                 ret = relocate_file_extent_cluster(rc->data_inode,
4049                                                    &rc->cluster);
4050                 if (ret < 0)
4051                         err = ret;
4052         }
4053
4054         rc->create_reloc_tree = 0;
4055         set_reloc_control(rc);
4056
4057         backref_cache_cleanup(&rc->backref_cache);
4058         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4059
4060         err = prepare_to_merge(rc, err);
4061
4062         merge_reloc_roots(rc);
4063
4064         rc->merge_reloc_tree = 0;
4065         unset_reloc_control(rc);
4066         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4067
4068         /* get rid of pinned extents */
4069         trans = btrfs_join_transaction(rc->extent_root);
4070         if (IS_ERR(trans))
4071                 err = PTR_ERR(trans);
4072         else
4073                 btrfs_commit_transaction(trans, rc->extent_root);
4074 out_free:
4075         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4076         btrfs_free_path(path);
4077         return err;
4078 }
4079
4080 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4081                                  struct btrfs_root *root, u64 objectid)
4082 {
4083         struct btrfs_path *path;
4084         struct btrfs_inode_item *item;
4085         struct extent_buffer *leaf;
4086         int ret;
4087
4088         path = btrfs_alloc_path();
4089         if (!path)
4090                 return -ENOMEM;
4091
4092         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4093         if (ret)
4094                 goto out;
4095
4096         leaf = path->nodes[0];
4097         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4098         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4099         btrfs_set_inode_generation(leaf, item, 1);
4100         btrfs_set_inode_size(leaf, item, 0);
4101         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4102         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4103                                           BTRFS_INODE_PREALLOC);
4104         btrfs_mark_buffer_dirty(leaf);
4105         btrfs_release_path(path);
4106 out:
4107         btrfs_free_path(path);
4108         return ret;
4109 }
4110
4111 /*
4112  * helper to create inode for data relocation.
4113  * the inode is in data relocation tree and its link count is 0
4114  */
4115 static noinline_for_stack
4116 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4117                                  struct btrfs_block_group_cache *group)
4118 {
4119         struct inode *inode = NULL;
4120         struct btrfs_trans_handle *trans;
4121         struct btrfs_root *root;
4122         struct btrfs_key key;
4123         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4124         int err = 0;
4125
4126         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4127         if (IS_ERR(root))
4128                 return ERR_CAST(root);
4129
4130         trans = btrfs_start_transaction(root, 6);
4131         if (IS_ERR(trans))
4132                 return ERR_CAST(trans);
4133
4134         err = btrfs_find_free_objectid(root, &objectid);
4135         if (err)
4136                 goto out;
4137
4138         err = __insert_orphan_inode(trans, root, objectid);
4139         BUG_ON(err);
4140
4141         key.objectid = objectid;
4142         key.type = BTRFS_INODE_ITEM_KEY;
4143         key.offset = 0;
4144         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4145         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4146         BTRFS_I(inode)->index_cnt = group->key.objectid;
4147
4148         err = btrfs_orphan_add(trans, inode);
4149 out:
4150         btrfs_end_transaction(trans, root);
4151         btrfs_btree_balance_dirty(root);
4152         if (err) {
4153                 if (inode)
4154                         iput(inode);
4155                 inode = ERR_PTR(err);
4156         }
4157         return inode;
4158 }
4159
4160 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4161 {
4162         struct reloc_control *rc;
4163
4164         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4165         if (!rc)
4166                 return NULL;
4167
4168         INIT_LIST_HEAD(&rc->reloc_roots);
4169         backref_cache_init(&rc->backref_cache);
4170         mapping_tree_init(&rc->reloc_root_tree);
4171         extent_io_tree_init(&rc->processed_blocks,
4172                             fs_info->btree_inode->i_mapping);
4173         return rc;
4174 }
4175
4176 /*
4177  * function to relocate all extents in a block group.
4178  */
4179 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4180 {
4181         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4182         struct reloc_control *rc;
4183         struct inode *inode;
4184         struct btrfs_path *path;
4185         int ret;
4186         int rw = 0;
4187         int err = 0;
4188
4189         rc = alloc_reloc_control(fs_info);
4190         if (!rc)
4191                 return -ENOMEM;
4192
4193         rc->extent_root = extent_root;
4194
4195         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4196         BUG_ON(!rc->block_group);
4197
4198         if (!rc->block_group->ro) {
4199                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4200                 if (ret) {
4201                         err = ret;
4202                         goto out;
4203                 }
4204                 rw = 1;
4205         }
4206
4207         path = btrfs_alloc_path();
4208         if (!path) {
4209                 err = -ENOMEM;
4210                 goto out;
4211         }
4212
4213         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4214                                         path);
4215         btrfs_free_path(path);
4216
4217         if (!IS_ERR(inode))
4218                 ret = delete_block_group_cache(fs_info, inode, 0);
4219         else
4220                 ret = PTR_ERR(inode);
4221
4222         if (ret && ret != -ENOENT) {
4223                 err = ret;
4224                 goto out;
4225         }
4226
4227         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4228         if (IS_ERR(rc->data_inode)) {
4229                 err = PTR_ERR(rc->data_inode);
4230                 rc->data_inode = NULL;
4231                 goto out;
4232         }
4233
4234         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4235                rc->block_group->key.objectid, rc->block_group->flags);
4236
4237         ret = btrfs_start_all_delalloc_inodes(fs_info, 0);
4238         if (ret < 0) {
4239                 err = ret;
4240                 goto out;
4241         }
4242         btrfs_wait_all_ordered_extents(fs_info);
4243
4244         while (1) {
4245                 mutex_lock(&fs_info->cleaner_mutex);
4246                 ret = relocate_block_group(rc);
4247                 mutex_unlock(&fs_info->cleaner_mutex);
4248                 if (ret < 0) {
4249                         err = ret;
4250                         goto out;
4251                 }
4252
4253                 if (rc->extents_found == 0)
4254                         break;
4255
4256                 printk(KERN_INFO "btrfs: found %llu extents\n",
4257                         rc->extents_found);
4258
4259                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4260                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4261                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4262                                                  0, -1);
4263                         rc->stage = UPDATE_DATA_PTRS;
4264                 }
4265         }
4266
4267         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4268                                      rc->block_group->key.objectid,
4269                                      rc->block_group->key.objectid +
4270                                      rc->block_group->key.offset - 1);
4271
4272         WARN_ON(rc->block_group->pinned > 0);
4273         WARN_ON(rc->block_group->reserved > 0);
4274         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4275 out:
4276         if (err && rw)
4277                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4278         iput(rc->data_inode);
4279         btrfs_put_block_group(rc->block_group);
4280         kfree(rc);
4281         return err;
4282 }
4283
4284 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4285 {
4286         struct btrfs_trans_handle *trans;
4287         int ret, err;
4288
4289         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4290         if (IS_ERR(trans))
4291                 return PTR_ERR(trans);
4292
4293         memset(&root->root_item.drop_progress, 0,
4294                 sizeof(root->root_item.drop_progress));
4295         root->root_item.drop_level = 0;
4296         btrfs_set_root_refs(&root->root_item, 0);
4297         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4298                                 &root->root_key, &root->root_item);
4299
4300         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4301         if (err)
4302                 return err;
4303         return ret;
4304 }
4305
4306 /*
4307  * recover relocation interrupted by system crash.
4308  *
4309  * this function resumes merging reloc trees with corresponding fs trees.
4310  * this is important for keeping the sharing of tree blocks
4311  */
4312 int btrfs_recover_relocation(struct btrfs_root *root)
4313 {
4314         LIST_HEAD(reloc_roots);
4315         struct btrfs_key key;
4316         struct btrfs_root *fs_root;
4317         struct btrfs_root *reloc_root;
4318         struct btrfs_path *path;
4319         struct extent_buffer *leaf;
4320         struct reloc_control *rc = NULL;
4321         struct btrfs_trans_handle *trans;
4322         int ret;
4323         int err = 0;
4324
4325         path = btrfs_alloc_path();
4326         if (!path)
4327                 return -ENOMEM;
4328         path->reada = -1;
4329
4330         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4331         key.type = BTRFS_ROOT_ITEM_KEY;
4332         key.offset = (u64)-1;
4333
4334         while (1) {
4335                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4336                                         path, 0, 0);
4337                 if (ret < 0) {
4338                         err = ret;
4339                         goto out;
4340                 }
4341                 if (ret > 0) {
4342                         if (path->slots[0] == 0)
4343                                 break;
4344                         path->slots[0]--;
4345                 }
4346                 leaf = path->nodes[0];
4347                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4348                 btrfs_release_path(path);
4349
4350                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4351                     key.type != BTRFS_ROOT_ITEM_KEY)
4352                         break;
4353
4354                 reloc_root = btrfs_read_fs_root(root, &key);
4355                 if (IS_ERR(reloc_root)) {
4356                         err = PTR_ERR(reloc_root);
4357                         goto out;
4358                 }
4359
4360                 list_add(&reloc_root->root_list, &reloc_roots);
4361
4362                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4363                         fs_root = read_fs_root(root->fs_info,
4364                                                reloc_root->root_key.offset);
4365                         if (IS_ERR(fs_root)) {
4366                                 ret = PTR_ERR(fs_root);
4367                                 if (ret != -ENOENT) {
4368                                         err = ret;
4369                                         goto out;
4370                                 }
4371                                 ret = mark_garbage_root(reloc_root);
4372                                 if (ret < 0) {
4373                                         err = ret;
4374                                         goto out;
4375                                 }
4376                         }
4377                 }
4378
4379                 if (key.offset == 0)
4380                         break;
4381
4382                 key.offset--;
4383         }
4384         btrfs_release_path(path);
4385
4386         if (list_empty(&reloc_roots))
4387                 goto out;
4388
4389         rc = alloc_reloc_control(root->fs_info);
4390         if (!rc) {
4391                 err = -ENOMEM;
4392                 goto out;
4393         }
4394
4395         rc->extent_root = root->fs_info->extent_root;
4396
4397         set_reloc_control(rc);
4398
4399         trans = btrfs_join_transaction(rc->extent_root);
4400         if (IS_ERR(trans)) {
4401                 unset_reloc_control(rc);
4402                 err = PTR_ERR(trans);
4403                 goto out_free;
4404         }
4405
4406         rc->merge_reloc_tree = 1;
4407
4408         while (!list_empty(&reloc_roots)) {
4409                 reloc_root = list_entry(reloc_roots.next,
4410                                         struct btrfs_root, root_list);
4411                 list_del(&reloc_root->root_list);
4412
4413                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4414                         list_add_tail(&reloc_root->root_list,
4415                                       &rc->reloc_roots);
4416                         continue;
4417                 }
4418
4419                 fs_root = read_fs_root(root->fs_info,
4420                                        reloc_root->root_key.offset);
4421                 if (IS_ERR(fs_root)) {
4422                         err = PTR_ERR(fs_root);
4423                         goto out_free;
4424                 }
4425
4426                 err = __add_reloc_root(reloc_root);
4427                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4428                 fs_root->reloc_root = reloc_root;
4429         }
4430
4431         err = btrfs_commit_transaction(trans, rc->extent_root);
4432         if (err)
4433                 goto out_free;
4434
4435         merge_reloc_roots(rc);
4436
4437         unset_reloc_control(rc);
4438
4439         trans = btrfs_join_transaction(rc->extent_root);
4440         if (IS_ERR(trans))
4441                 err = PTR_ERR(trans);
4442         else
4443                 err = btrfs_commit_transaction(trans, rc->extent_root);
4444 out_free:
4445         kfree(rc);
4446 out:
4447         if (!list_empty(&reloc_roots))
4448                 free_reloc_roots(&reloc_roots);
4449
4450         btrfs_free_path(path);
4451
4452         if (err == 0) {
4453                 /* cleanup orphan inode in data relocation tree */
4454                 fs_root = read_fs_root(root->fs_info,
4455                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4456                 if (IS_ERR(fs_root))
4457                         err = PTR_ERR(fs_root);
4458                 else
4459                         err = btrfs_orphan_cleanup(fs_root);
4460         }
4461         return err;
4462 }
4463
4464 /*
4465  * helper to add ordered checksum for data relocation.
4466  *
4467  * cloning checksum properly handles the nodatasum extents.
4468  * it also saves CPU time to re-calculate the checksum.
4469  */
4470 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4471 {
4472         struct btrfs_ordered_sum *sums;
4473         struct btrfs_ordered_extent *ordered;
4474         struct btrfs_root *root = BTRFS_I(inode)->root;
4475         int ret;
4476         u64 disk_bytenr;
4477         u64 new_bytenr;
4478         LIST_HEAD(list);
4479
4480         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4481         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4482
4483         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4484         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4485                                        disk_bytenr + len - 1, &list, 0);
4486         if (ret)
4487                 goto out;
4488
4489         while (!list_empty(&list)) {
4490                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4491                 list_del_init(&sums->list);
4492
4493                 /*
4494                  * We need to offset the new_bytenr based on where the csum is.
4495                  * We need to do this because we will read in entire prealloc
4496                  * extents but we may have written to say the middle of the
4497                  * prealloc extent, so we need to make sure the csum goes with
4498                  * the right disk offset.
4499                  *
4500                  * We can do this because the data reloc inode refers strictly
4501                  * to the on disk bytes, so we don't have to worry about
4502                  * disk_len vs real len like with real inodes since it's all
4503                  * disk length.
4504                  */
4505                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4506                 sums->bytenr = new_bytenr;
4507
4508                 btrfs_add_ordered_sum(inode, ordered, sums);
4509         }
4510 out:
4511         btrfs_put_ordered_extent(ordered);
4512         return ret;
4513 }
4514
4515 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4516                           struct btrfs_root *root, struct extent_buffer *buf,
4517                           struct extent_buffer *cow)
4518 {
4519         struct reloc_control *rc;
4520         struct backref_node *node;
4521         int first_cow = 0;
4522         int level;
4523         int ret = 0;
4524
4525         rc = root->fs_info->reloc_ctl;
4526         if (!rc)
4527                 return 0;
4528
4529         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4530                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4531
4532         level = btrfs_header_level(buf);
4533         if (btrfs_header_generation(buf) <=
4534             btrfs_root_last_snapshot(&root->root_item))
4535                 first_cow = 1;
4536
4537         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4538             rc->create_reloc_tree) {
4539                 WARN_ON(!first_cow && level == 0);
4540
4541                 node = rc->backref_cache.path[level];
4542                 BUG_ON(node->bytenr != buf->start &&
4543                        node->new_bytenr != buf->start);
4544
4545                 drop_node_buffer(node);
4546                 extent_buffer_get(cow);
4547                 node->eb = cow;
4548                 node->new_bytenr = cow->start;
4549
4550                 if (!node->pending) {
4551                         list_move_tail(&node->list,
4552                                        &rc->backref_cache.pending[level]);
4553                         node->pending = 1;
4554                 }
4555
4556                 if (first_cow)
4557                         __mark_block_processed(rc, node);
4558
4559                 if (first_cow && level > 0)
4560                         rc->nodes_relocated += buf->len;
4561         }
4562
4563         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4564                 ret = replace_file_extents(trans, rc, root, cow);
4565         return ret;
4566 }
4567
4568 /*
4569  * called before creating snapshot. it calculates metadata reservation
4570  * requried for relocating tree blocks in the snapshot
4571  */
4572 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4573                               struct btrfs_pending_snapshot *pending,
4574                               u64 *bytes_to_reserve)
4575 {
4576         struct btrfs_root *root;
4577         struct reloc_control *rc;
4578
4579         root = pending->root;
4580         if (!root->reloc_root)
4581                 return;
4582
4583         rc = root->fs_info->reloc_ctl;
4584         if (!rc->merge_reloc_tree)
4585                 return;
4586
4587         root = root->reloc_root;
4588         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4589         /*
4590          * relocation is in the stage of merging trees. the space
4591          * used by merging a reloc tree is twice the size of
4592          * relocated tree nodes in the worst case. half for cowing
4593          * the reloc tree, half for cowing the fs tree. the space
4594          * used by cowing the reloc tree will be freed after the
4595          * tree is dropped. if we create snapshot, cowing the fs
4596          * tree may use more space than it frees. so we need
4597          * reserve extra space.
4598          */
4599         *bytes_to_reserve += rc->nodes_relocated;
4600 }
4601
4602 /*
4603  * called after snapshot is created. migrate block reservation
4604  * and create reloc root for the newly created snapshot
4605  */
4606 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4607                                struct btrfs_pending_snapshot *pending)
4608 {
4609         struct btrfs_root *root = pending->root;
4610         struct btrfs_root *reloc_root;
4611         struct btrfs_root *new_root;
4612         struct reloc_control *rc;
4613         int ret;
4614
4615         if (!root->reloc_root)
4616                 return 0;
4617
4618         rc = root->fs_info->reloc_ctl;
4619         rc->merging_rsv_size += rc->nodes_relocated;
4620
4621         if (rc->merge_reloc_tree) {
4622                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4623                                               rc->block_rsv,
4624                                               rc->nodes_relocated);
4625                 if (ret)
4626                         return ret;
4627         }
4628
4629         new_root = pending->snap;
4630         reloc_root = create_reloc_root(trans, root->reloc_root,
4631                                        new_root->root_key.objectid);
4632         if (IS_ERR(reloc_root))
4633                 return PTR_ERR(reloc_root);
4634
4635         ret = __add_reloc_root(reloc_root);
4636         BUG_ON(ret < 0);
4637         new_root->reloc_root = reloc_root;
4638
4639         if (rc->create_reloc_tree)
4640                 ret = clone_backref_node(trans, rc, root, reloc_root);
4641         return ret;
4642 }