Pileus Git - ~andy/linux/blob - fs/btrfs/relocation.c

   1 /*
   2  * Copyright (C) 2009 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/rbtree.h>
  24 #include <linux/slab.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "volumes.h"
  29 #include "locking.h"
  30 #include "btrfs_inode.h"
  31 #include "async-thread.h"
  32 #include "free-space-cache.h"
  33 #include "inode-map.h"
  34
  35 /*
  36  * backref_node, mapping_node and tree_block start with this
  37  */
  38 struct tree_entry {
  39         struct rb_node rb_node;
  40         u64 bytenr;
  41 };
  42
  43 /*
  44  * present a tree block in the backref cache
  45  */
  46 struct backref_node {
  47         struct rb_node rb_node;
  48         u64 bytenr;
  49
  50         u64 new_bytenr;
  51         /* objectid of tree block owner, can be not uptodate */
  52         u64 owner;
  53         /* link to pending, changed or detached list */
  54         struct list_head list;
  55         /* list of upper level blocks reference this block */
  56         struct list_head upper;
  57         /* list of child blocks in the cache */
  58         struct list_head lower;
  59         /* NULL if this node is not tree root */
  60         struct btrfs_root *root;
  61         /* extent buffer got by COW the block */
  62         struct extent_buffer *eb;
  63         /* level of tree block */
  64         unsigned int level:8;
  65         /* is the block in non-reference counted tree */
  66         unsigned int cowonly:1;
  67         /* 1 if no child node in the cache */
  68         unsigned int lowest:1;
  69         /* is the extent buffer locked */
  70         unsigned int locked:1;
  71         /* has the block been processed */
  72         unsigned int processed:1;
  73         /* have backrefs of this block been checked */
  74         unsigned int checked:1;
  75         /*
  76          * 1 if corresponding block has been cowed but some upper
  77          * level block pointers may not point to the new location
  78          */
  79         unsigned int pending:1;
  80         /*
  81          * 1 if the backref node isn't connected to any other
  82          * backref node.
  83          */
  84         unsigned int detached:1;
  85 };
  86
  87 /*
  88  * present a block pointer in the backref cache
  89  */
  90 struct backref_edge {
  91         struct list_head list[2];
  92         struct backref_node *node[2];
  93 };
  94
  95 #define LOWER   0
  96 #define UPPER   1
  97
  98 struct backref_cache {
  99         /* red black tree of all backref nodes in the cache */
 100         struct rb_root rb_root;
 101         /* for passing backref nodes to btrfs_reloc_cow_block */
 102         struct backref_node *path[BTRFS_MAX_LEVEL];
 103         /*
 104          * list of blocks that have been cowed but some block
 105          * pointers in upper level blocks may not reflect the
 106          * new location
 107          */
 108         struct list_head pending[BTRFS_MAX_LEVEL];
 109         /* list of backref nodes with no child node */
 110         struct list_head leaves;
 111         /* list of blocks that have been cowed in current transaction */
 112         struct list_head changed;
 113         /* list of detached backref node. */
 114         struct list_head detached;
 115
 116         u64 last_trans;
 117
 118         int nr_nodes;
 119         int nr_edges;
 120 };
 121
 122 /*
 123  * map address of tree root to tree
 124  */
 125 struct mapping_node {
 126         struct rb_node rb_node;
 127         u64 bytenr;
 128         void *data;
 129 };
 130
 131 struct mapping_tree {
 132         struct rb_root rb_root;
 133         spinlock_t lock;
 134 };
 135
 136 /*
 137  * present a tree block to process
 138  */
 139 struct tree_block {
 140         struct rb_node rb_node;
 141         u64 bytenr;
 142         struct btrfs_key key;
 143         unsigned int level:8;
 144         unsigned int key_ready:1;
 145 };
 146
 147 #define MAX_EXTENTS 128
 148
 149 struct file_extent_cluster {
 150         u64 start;
 151         u64 end;
 152         u64 boundary[MAX_EXTENTS];
 153         unsigned int nr;
 154 };
 155
 156 struct reloc_control {
 157         /* block group to relocate */
 158         struct btrfs_block_group_cache *block_group;
 159         /* extent tree */
 160         struct btrfs_root *extent_root;
 161         /* inode for moving data */
 162         struct inode *data_inode;
 163
 164         struct btrfs_block_rsv *block_rsv;
 165
 166         struct backref_cache backref_cache;
 167
 168         struct file_extent_cluster cluster;
 169         /* tree blocks have been processed */
 170         struct extent_io_tree processed_blocks;
 171         /* map start of tree root to corresponding reloc tree */
 172         struct mapping_tree reloc_root_tree;
 173         /* list of reloc trees */
 174         struct list_head reloc_roots;
 175         /* size of metadata reservation for merging reloc trees */
 176         u64 merging_rsv_size;
 177         /* size of relocated tree nodes */
 178         u64 nodes_relocated;
 179
 180         u64 search_start;
 181         u64 extents_found;
 182
 183         unsigned int stage:8;
 184         unsigned int create_reloc_tree:1;
 185         unsigned int merge_reloc_tree:1;
 186         unsigned int found_file_extent:1;
 187         unsigned int commit_transaction:1;
 188 };
 189
 190 /* stages of data relocation */
 191 #define MOVE_DATA_EXTENTS       0
 192 #define UPDATE_DATA_PTRS        1
 193
 194 static void remove_backref_node(struct backref_cache *cache,
 195                                 struct backref_node *node);
 196 static void __mark_block_processed(struct reloc_control *rc,
 197                                    struct backref_node *node);
 198
 199 static void mapping_tree_init(struct mapping_tree *tree)
 200 {
 201         tree->rb_root = RB_ROOT;
 202         spin_lock_init(&tree->lock);
 203 }
 204
 205 static void backref_cache_init(struct backref_cache *cache)
 206 {
 207         int i;
 208         cache->rb_root = RB_ROOT;
 209         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 210                 INIT_LIST_HEAD(&cache->pending[i]);
 211         INIT_LIST_HEAD(&cache->changed);
 212         INIT_LIST_HEAD(&cache->detached);
 213         INIT_LIST_HEAD(&cache->leaves);
 214 }
 215
 216 static void backref_cache_cleanup(struct backref_cache *cache)
 217 {
 218         struct backref_node *node;
 219         int i;
 220
 221         while (!list_empty(&cache->detached)) {
 222                 node = list_entry(cache->detached.next,
 223                                   struct backref_node, list);
 224                 remove_backref_node(cache, node);
 225         }
 226
 227         while (!list_empty(&cache->leaves)) {
 228                 node = list_entry(cache->leaves.next,
 229                                   struct backref_node, lower);
 230                 remove_backref_node(cache, node);
 231         }
 232
 233         cache->last_trans = 0;
 234
 235         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 236                 BUG_ON(!list_empty(&cache->pending[i]));
 237         BUG_ON(!list_empty(&cache->changed));
 238         BUG_ON(!list_empty(&cache->detached));
 239         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
 240         BUG_ON(cache->nr_nodes);
 241         BUG_ON(cache->nr_edges);
 242 }
 243
 244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 245 {
 246         struct backref_node *node;
 247
 248         node = kzalloc(sizeof(*node), GFP_NOFS);
 249         if (node) {
 250                 INIT_LIST_HEAD(&node->list);
 251                 INIT_LIST_HEAD(&node->upper);
 252                 INIT_LIST_HEAD(&node->lower);
 253                 RB_CLEAR_NODE(&node->rb_node);
 254                 cache->nr_nodes++;
 255         }
 256         return node;
 257 }
 258
 259 static void free_backref_node(struct backref_cache *cache,
 260                               struct backref_node *node)
 261 {
 262         if (node) {
 263                 cache->nr_nodes--;
 264                 kfree(node);
 265         }
 266 }
 267
 268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 269 {
 270         struct backref_edge *edge;
 271
 272         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 273         if (edge)
 274                 cache->nr_edges++;
 275         return edge;
 276 }
 277
 278 static void free_backref_edge(struct backref_cache *cache,
 279                               struct backref_edge *edge)
 280 {
 281         if (edge) {
 282                 cache->nr_edges--;
 283                 kfree(edge);
 284         }
 285 }
 286
 287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 288                                    struct rb_node *node)
 289 {
 290         struct rb_node **p = &root->rb_node;
 291         struct rb_node *parent = NULL;
 292         struct tree_entry *entry;
 293
 294         while (*p) {
 295                 parent = *p;
 296                 entry = rb_entry(parent, struct tree_entry, rb_node);
 297
 298                 if (bytenr < entry->bytenr)
 299                         p = &(*p)->rb_left;
 300                 else if (bytenr > entry->bytenr)
 301                         p = &(*p)->rb_right;
 302                 else
 303                         return parent;
 304         }
 305
 306         rb_link_node(node, parent, p);
 307         rb_insert_color(node, root);
 308         return NULL;
 309 }
 310
 311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 312 {
 313         struct rb_node *n = root->rb_node;
 314         struct tree_entry *entry;
 315
 316         while (n) {
 317                 entry = rb_entry(n, struct tree_entry, rb_node);
 318
 319                 if (bytenr < entry->bytenr)
 320                         n = n->rb_left;
 321                 else if (bytenr > entry->bytenr)
 322                         n = n->rb_right;
 323                 else
 324                         return n;
 325         }
 326         return NULL;
 327 }
 328
 329 void backref_tree_panic(struct rb_node *rb_node, int errno,
 330                                           u64 bytenr)
 331 {
 332
 333         struct btrfs_fs_info *fs_info = NULL;
 334         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 335                                               rb_node);
 336         if (bnode->root)
 337                 fs_info = bnode->root->fs_info;
 338         btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
 339                     "found at offset %llu\n", (unsigned long long)bytenr);
 340 }
 341
 342 /*
 343  * walk up backref nodes until reach node presents tree root
 344  */
 345 static struct backref_node *walk_up_backref(struct backref_node *node,
 346                                             struct backref_edge *edges[],
 347                                             int *index)
 348 {
 349         struct backref_edge *edge;
 350         int idx = *index;
 351
 352         while (!list_empty(&node->upper)) {
 353                 edge = list_entry(node->upper.next,
 354                                   struct backref_edge, list[LOWER]);
 355                 edges[idx++] = edge;
 356                 node = edge->node[UPPER];
 357         }
 358         BUG_ON(node->detached);
 359         *index = idx;
 360         return node;
 361 }
 362
 363 /*
 364  * walk down backref nodes to find start of next reference path
 365  */
 366 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 367                                               int *index)
 368 {
 369         struct backref_edge *edge;
 370         struct backref_node *lower;
 371         int idx = *index;
 372
 373         while (idx > 0) {
 374                 edge = edges[idx - 1];
 375                 lower = edge->node[LOWER];
 376                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 377                         idx--;
 378                         continue;
 379                 }
 380                 edge = list_entry(edge->list[LOWER].next,
 381                                   struct backref_edge, list[LOWER]);
 382                 edges[idx - 1] = edge;
 383                 *index = idx;
 384                 return edge->node[UPPER];
 385         }
 386         *index = 0;
 387         return NULL;
 388 }
 389
 390 static void unlock_node_buffer(struct backref_node *node)
 391 {
 392         if (node->locked) {
 393                 btrfs_tree_unlock(node->eb);
 394                 node->locked = 0;
 395         }
 396 }
 397
 398 static void drop_node_buffer(struct backref_node *node)
 399 {
 400         if (node->eb) {
 401                 unlock_node_buffer(node);
 402                 free_extent_buffer(node->eb);
 403                 node->eb = NULL;
 404         }
 405 }
 406
 407 static void drop_backref_node(struct backref_cache *tree,
 408                               struct backref_node *node)
 409 {
 410         BUG_ON(!list_empty(&node->upper));
 411
 412         drop_node_buffer(node);
 413         list_del(&node->list);
 414         list_del(&node->lower);
 415         if (!RB_EMPTY_NODE(&node->rb_node))
 416                 rb_erase(&node->rb_node, &tree->rb_root);
 417         free_backref_node(tree, node);
 418 }
 419
 420 /*
 421  * remove a backref node from the backref cache
 422  */
 423 static void remove_backref_node(struct backref_cache *cache,
 424                                 struct backref_node *node)
 425 {
 426         struct backref_node *upper;
 427         struct backref_edge *edge;
 428
 429         if (!node)
 430                 return;
 431
 432         BUG_ON(!node->lowest && !node->detached);
 433         while (!list_empty(&node->upper)) {
 434                 edge = list_entry(node->upper.next, struct backref_edge,
 435                                   list[LOWER]);
 436                 upper = edge->node[UPPER];
 437                 list_del(&edge->list[LOWER]);
 438                 list_del(&edge->list[UPPER]);
 439                 free_backref_edge(cache, edge);
 440
 441                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 442                         BUG_ON(!list_empty(&node->upper));
 443                         drop_backref_node(cache, node);
 444                         node = upper;
 445                         node->lowest = 1;
 446                         continue;
 447                 }
 448                 /*
 449                  * add the node to leaf node list if no other
 450                  * child block cached.
 451                  */
 452                 if (list_empty(&upper->lower)) {
 453                         list_add_tail(&upper->lower, &cache->leaves);
 454                         upper->lowest = 1;
 455                 }
 456         }
 457
 458         drop_backref_node(cache, node);
 459 }
 460
 461 static void update_backref_node(struct backref_cache *cache,
 462                                 struct backref_node *node, u64 bytenr)
 463 {
 464         struct rb_node *rb_node;
 465         rb_erase(&node->rb_node, &cache->rb_root);
 466         node->bytenr = bytenr;
 467         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 468         if (rb_node)
 469                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 470 }
 471
 472 /*
 473  * update backref cache after a transaction commit
 474  */
 475 static int update_backref_cache(struct btrfs_trans_handle *trans,
 476                                 struct backref_cache *cache)
 477 {
 478         struct backref_node *node;
 479         int level = 0;
 480
 481         if (cache->last_trans == 0) {
 482                 cache->last_trans = trans->transid;
 483                 return 0;
 484         }
 485
 486         if (cache->last_trans == trans->transid)
 487                 return 0;
 488
 489         /*
 490          * detached nodes are used to avoid unnecessary backref
 491          * lookup. transaction commit changes the extent tree.
 492          * so the detached nodes are no longer useful.
 493          */
 494         while (!list_empty(&cache->detached)) {
 495                 node = list_entry(cache->detached.next,
 496                                   struct backref_node, list);
 497                 remove_backref_node(cache, node);
 498         }
 499
 500         while (!list_empty(&cache->changed)) {
 501                 node = list_entry(cache->changed.next,
 502                                   struct backref_node, list);
 503                 list_del_init(&node->list);
 504                 BUG_ON(node->pending);
 505                 update_backref_node(cache, node, node->new_bytenr);
 506         }
 507
 508         /*
 509          * some nodes can be left in the pending list if there were
 510          * errors during processing the pending nodes.
 511          */
 512         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 513                 list_for_each_entry(node, &cache->pending[level], list) {
 514                         BUG_ON(!node->pending);
 515                         if (node->bytenr == node->new_bytenr)
 516                                 continue;
 517                         update_backref_node(cache, node, node->new_bytenr);
 518                 }
 519         }
 520
 521         cache->last_trans = 0;
 522         return 1;
 523 }
 524
 525
 526 static int should_ignore_root(struct btrfs_root *root)
 527 {
 528         struct btrfs_root *reloc_root;
 529
 530         if (!root->ref_cows)
 531                 return 0;
 532
 533         reloc_root = root->reloc_root;
 534         if (!reloc_root)
 535                 return 0;
 536
 537         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 538             root->fs_info->running_transaction->transid - 1)
 539                 return 0;
 540         /*
 541          * if there is reloc tree and it was created in previous
 542          * transaction backref lookup can find the reloc tree,
 543          * so backref node for the fs tree root is useless for
 544          * relocation.
 545          */
 546         return 1;
 547 }
 548 /*
 549  * find reloc tree by address of tree root
 550  */
 551 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 552                                           u64 bytenr)
 553 {
 554         struct rb_node *rb_node;
 555         struct mapping_node *node;
 556         struct btrfs_root *root = NULL;
 557
 558         spin_lock(&rc->reloc_root_tree.lock);
 559         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 560         if (rb_node) {
 561                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 562                 root = (struct btrfs_root *)node->data;
 563         }
 564         spin_unlock(&rc->reloc_root_tree.lock);
 565         return root;
 566 }
 567
 568 static int is_cowonly_root(u64 root_objectid)
 569 {
 570         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 571             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 572             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 573             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 574             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 575             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
 576                 return 1;
 577         return 0;
 578 }
 579
 580 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 581                                         u64 root_objectid)
 582 {
 583         struct btrfs_key key;
 584
 585         key.objectid = root_objectid;
 586         key.type = BTRFS_ROOT_ITEM_KEY;
 587         if (is_cowonly_root(root_objectid))
 588                 key.offset = 0;
 589         else
 590                 key.offset = (u64)-1;
 591
 592         return btrfs_read_fs_root_no_name(fs_info, &key);
 593 }
 594
 595 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 596 static noinline_for_stack
 597 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 598                                   struct extent_buffer *leaf,
 599                                   struct btrfs_extent_ref_v0 *ref0)
 600 {
 601         struct btrfs_root *root;
 602         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 603         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 604
 605         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 606
 607         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 608         BUG_ON(IS_ERR(root));
 609
 610         if (root->ref_cows &&
 611             generation != btrfs_root_generation(&root->root_item))
 612                 return NULL;
 613
 614         return root;
 615 }
 616 #endif
 617
 618 static noinline_for_stack
 619 int find_inline_backref(struct extent_buffer *leaf, int slot,
 620                         unsigned long *ptr, unsigned long *end)
 621 {
 622         struct btrfs_extent_item *ei;
 623         struct btrfs_tree_block_info *bi;
 624         u32 item_size;
 625
 626         item_size = btrfs_item_size_nr(leaf, slot);
 627 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 628         if (item_size < sizeof(*ei)) {
 629                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 630                 return 1;
 631         }
 632 #endif
 633         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 634         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 635                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 636
 637         if (item_size <= sizeof(*ei) + sizeof(*bi)) {
 638                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 639                 return 1;
 640         }
 641
 642         bi = (struct btrfs_tree_block_info *)(ei + 1);
 643         *ptr = (unsigned long)(bi + 1);
 644         *end = (unsigned long)ei + item_size;
 645         return 0;
 646 }
 647
 648 /*
 649  * build backref tree for a given tree block. root of the backref tree
 650  * corresponds the tree block, leaves of the backref tree correspond
 651  * roots of b-trees that reference the tree block.
 652  *
 653  * the basic idea of this function is check backrefs of a given block
 654  * to find upper level blocks that refernece the block, and then check
 655  * bakcrefs of these upper level blocks recursively. the recursion stop
 656  * when tree root is reached or backrefs for the block is cached.
 657  *
 658  * NOTE: if we find backrefs for a block are cached, we know backrefs
 659  * for all upper level blocks that directly/indirectly reference the
 660  * block are also cached.
 661  */
 662 static noinline_for_stack
 663 struct backref_node *build_backref_tree(struct reloc_control *rc,
 664                                         struct btrfs_key *node_key,
 665                                         int level, u64 bytenr)
 666 {
 667         struct backref_cache *cache = &rc->backref_cache;
 668         struct btrfs_path *path1;
 669         struct btrfs_path *path2;
 670         struct extent_buffer *eb;
 671         struct btrfs_root *root;
 672         struct backref_node *cur;
 673         struct backref_node *upper;
 674         struct backref_node *lower;
 675         struct backref_node *node = NULL;
 676         struct backref_node *exist = NULL;
 677         struct backref_edge *edge;
 678         struct rb_node *rb_node;
 679         struct btrfs_key key;
 680         unsigned long end;
 681         unsigned long ptr;
 682         LIST_HEAD(list);
 683         LIST_HEAD(useless);
 684         int cowonly;
 685         int ret;
 686         int err = 0;
 687
 688         path1 = btrfs_alloc_path();
 689         path2 = btrfs_alloc_path();
 690         if (!path1 || !path2) {
 691                 err = -ENOMEM;
 692                 goto out;
 693         }
 694         path1->reada = 1;
 695         path2->reada = 2;
 696
 697         node = alloc_backref_node(cache);
 698         if (!node) {
 699                 err = -ENOMEM;
 700                 goto out;
 701         }
 702
 703         node->bytenr = bytenr;
 704         node->level = level;
 705         node->lowest = 1;
 706         cur = node;
 707 again:
 708         end = 0;
 709         ptr = 0;
 710         key.objectid = cur->bytenr;
 711         key.type = BTRFS_EXTENT_ITEM_KEY;
 712         key.offset = (u64)-1;
 713
 714         path1->search_commit_root = 1;
 715         path1->skip_locking = 1;
 716         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 717                                 0, 0);
 718         if (ret < 0) {
 719                 err = ret;
 720                 goto out;
 721         }
 722         BUG_ON(!ret || !path1->slots[0]);
 723
 724         path1->slots[0]--;
 725
 726         WARN_ON(cur->checked);
 727         if (!list_empty(&cur->upper)) {
 728                 /*
 729                  * the backref was added previously when processing
 730                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 731                  */
 732                 BUG_ON(!list_is_singular(&cur->upper));
 733                 edge = list_entry(cur->upper.next, struct backref_edge,
 734                                   list[LOWER]);
 735                 BUG_ON(!list_empty(&edge->list[UPPER]));
 736                 exist = edge->node[UPPER];
 737                 /*
 738                  * add the upper level block to pending list if we need
 739                  * check its backrefs
 740                  */
 741                 if (!exist->checked)
 742                         list_add_tail(&edge->list[UPPER], &list);
 743         } else {
 744                 exist = NULL;
 745         }
 746
 747         while (1) {
 748                 cond_resched();
 749                 eb = path1->nodes[0];
 750
 751                 if (ptr >= end) {
 752                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 753                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 754                                 if (ret < 0) {
 755                                         err = ret;
 756                                         goto out;
 757                                 }
 758                                 if (ret > 0)
 759                                         break;
 760                                 eb = path1->nodes[0];
 761                         }
 762
 763                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 764                         if (key.objectid != cur->bytenr) {
 765                                 WARN_ON(exist);
 766                                 break;
 767                         }
 768
 769                         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 770                                 ret = find_inline_backref(eb, path1->slots[0],
 771                                                           &ptr, &end);
 772                                 if (ret)
 773                                         goto next;
 774                         }
 775                 }
 776
 777                 if (ptr < end) {
 778                         /* update key for inline back ref */
 779                         struct btrfs_extent_inline_ref *iref;
 780                         iref = (struct btrfs_extent_inline_ref *)ptr;
 781                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 782                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 783                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 784                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 785                 }
 786
 787                 if (exist &&
 788                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 789                       exist->owner == key.offset) ||
 790                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 791                       exist->bytenr == key.offset))) {
 792                         exist = NULL;
 793                         goto next;
 794                 }
 795
 796 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 797                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 798                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 799                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 800                                 struct btrfs_extent_ref_v0 *ref0;
 801                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 802                                                 struct btrfs_extent_ref_v0);
 803                                 if (key.objectid == key.offset) {
 804                                         root = find_tree_root(rc, eb, ref0);
 805                                         if (root && !should_ignore_root(root))
 806                                                 cur->root = root;
 807                                         else
 808                                                 list_add(&cur->list, &useless);
 809                                         break;
 810                                 }
 811                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 812                                                                       ref0)))
 813                                         cur->cowonly = 1;
 814                         }
 815 #else
 816                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 817                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 818 #endif
 819                         if (key.objectid == key.offset) {
 820                                 /*
 821                                  * only root blocks of reloc trees use
 822                                  * backref of this type.
 823                                  */
 824                                 root = find_reloc_root(rc, cur->bytenr);
 825                                 BUG_ON(!root);
 826                                 cur->root = root;
 827                                 break;
 828                         }
 829
 830                         edge = alloc_backref_edge(cache);
 831                         if (!edge) {
 832                                 err = -ENOMEM;
 833                                 goto out;
 834                         }
 835                         rb_node = tree_search(&cache->rb_root, key.offset);
 836                         if (!rb_node) {
 837                                 upper = alloc_backref_node(cache);
 838                                 if (!upper) {
 839                                         free_backref_edge(cache, edge);
 840                                         err = -ENOMEM;
 841                                         goto out;
 842                                 }
 843                                 upper->bytenr = key.offset;
 844                                 upper->level = cur->level + 1;
 845                                 /*
 846                                  *  backrefs for the upper level block isn't
 847                                  *  cached, add the block to pending list
 848                                  */
 849                                 list_add_tail(&edge->list[UPPER], &list);
 850                         } else {
 851                                 upper = rb_entry(rb_node, struct backref_node,
 852                                                  rb_node);
 853                                 BUG_ON(!upper->checked);
 854                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 855                         }
 856                         list_add_tail(&edge->list[LOWER], &cur->upper);
 857                         edge->node[LOWER] = cur;
 858                         edge->node[UPPER] = upper;
 859
 860                         goto next;
 861                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 862                         goto next;
 863                 }
 864
 865                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 866                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 867                 if (IS_ERR(root)) {
 868                         err = PTR_ERR(root);
 869                         goto out;
 870                 }
 871
 872                 if (!root->ref_cows)
 873                         cur->cowonly = 1;
 874
 875                 if (btrfs_root_level(&root->root_item) == cur->level) {
 876                         /* tree root */
 877                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 878                                cur->bytenr);
 879                         if (should_ignore_root(root))
 880                                 list_add(&cur->list, &useless);
 881                         else
 882                                 cur->root = root;
 883                         break;
 884                 }
 885
 886                 level = cur->level + 1;
 887
 888                 /*
 889                  * searching the tree to find upper level blocks
 890                  * reference the block.
 891                  */
 892                 path2->search_commit_root = 1;
 893                 path2->skip_locking = 1;
 894                 path2->lowest_level = level;
 895                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 896                 path2->lowest_level = 0;
 897                 if (ret < 0) {
 898                         err = ret;
 899                         goto out;
 900                 }
 901                 if (ret > 0 && path2->slots[level] > 0)
 902                         path2->slots[level]--;
 903
 904                 eb = path2->nodes[level];
 905                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 906                         cur->bytenr);
 907
 908                 lower = cur;
 909                 for (; level < BTRFS_MAX_LEVEL; level++) {
 910                         if (!path2->nodes[level]) {
 911                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 912                                        lower->bytenr);
 913                                 if (should_ignore_root(root))
 914                                         list_add(&lower->list, &useless);
 915                                 else
 916                                         lower->root = root;
 917                                 break;
 918                         }
 919
 920                         edge = alloc_backref_edge(cache);
 921                         if (!edge) {
 922                                 err = -ENOMEM;
 923                                 goto out;
 924                         }
 925
 926                         eb = path2->nodes[level];
 927                         rb_node = tree_search(&cache->rb_root, eb->start);
 928                         if (!rb_node) {
 929                                 upper = alloc_backref_node(cache);
 930                                 if (!upper) {
 931                                         free_backref_edge(cache, edge);
 932                                         err = -ENOMEM;
 933                                         goto out;
 934                                 }
 935                                 upper->bytenr = eb->start;
 936                                 upper->owner = btrfs_header_owner(eb);
 937                                 upper->level = lower->level + 1;
 938                                 if (!root->ref_cows)
 939                                         upper->cowonly = 1;
 940
 941                                 /*
 942                                  * if we know the block isn't shared
 943                                  * we can void checking its backrefs.
 944                                  */
 945                                 if (btrfs_block_can_be_shared(root, eb))
 946                                         upper->checked = 0;
 947                                 else
 948                                         upper->checked = 1;
 949
 950                                 /*
 951                                  * add the block to pending list if we
 952                                  * need check its backrefs. only block
 953                                  * at 'cur->level + 1' is added to the
 954                                  * tail of pending list. this guarantees
 955                                  * we check backrefs from lower level
 956                                  * blocks to upper level blocks.
 957                                  */
 958                                 if (!upper->checked &&
 959                                     level == cur->level + 1) {
 960                                         list_add_tail(&edge->list[UPPER],
 961                                                       &list);
 962                                 } else
 963                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 964                         } else {
 965                                 upper = rb_entry(rb_node, struct backref_node,
 966                                                  rb_node);
 967                                 BUG_ON(!upper->checked);
 968                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 969                                 if (!upper->owner)
 970                                         upper->owner = btrfs_header_owner(eb);
 971                         }
 972                         list_add_tail(&edge->list[LOWER], &lower->upper);
 973                         edge->node[LOWER] = lower;
 974                         edge->node[UPPER] = upper;
 975
 976                         if (rb_node)
 977                                 break;
 978                         lower = upper;
 979                         upper = NULL;
 980                 }
 981                 btrfs_release_path(path2);
 982 next:
 983                 if (ptr < end) {
 984                         ptr += btrfs_extent_inline_ref_size(key.type);
 985                         if (ptr >= end) {
 986                                 WARN_ON(ptr > end);
 987                                 ptr = 0;
 988                                 end = 0;
 989                         }
 990                 }
 991                 if (ptr >= end)
 992                         path1->slots[0]++;
 993         }
 994         btrfs_release_path(path1);
 995
 996         cur->checked = 1;
 997         WARN_ON(exist);
 998
 999         /* the pending list isn't empty, take the first block to process */
1000         if (!list_empty(&list)) {
1001                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1002                 list_del_init(&edge->list[UPPER]);
1003                 cur = edge->node[UPPER];
1004                 goto again;
1005         }
1006
1007         /*
1008          * everything goes well, connect backref nodes and insert backref nodes
1009          * into the cache.
1010          */
1011         BUG_ON(!node->checked);
1012         cowonly = node->cowonly;
1013         if (!cowonly) {
1014                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1015                                       &node->rb_node);
1016                 if (rb_node)
1017                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1018                 list_add_tail(&node->lower, &cache->leaves);
1019         }
1020
1021         list_for_each_entry(edge, &node->upper, list[LOWER])
1022                 list_add_tail(&edge->list[UPPER], &list);
1023
1024         while (!list_empty(&list)) {
1025                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1026                 list_del_init(&edge->list[UPPER]);
1027                 upper = edge->node[UPPER];
1028                 if (upper->detached) {
1029                         list_del(&edge->list[LOWER]);
1030                         lower = edge->node[LOWER];
1031                         free_backref_edge(cache, edge);
1032                         if (list_empty(&lower->upper))
1033                                 list_add(&lower->list, &useless);
1034                         continue;
1035                 }
1036
1037                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1038                         if (upper->lowest) {
1039                                 list_del_init(&upper->lower);
1040                                 upper->lowest = 0;
1041                         }
1042
1043                         list_add_tail(&edge->list[UPPER], &upper->lower);
1044                         continue;
1045                 }
1046
1047                 BUG_ON(!upper->checked);
1048                 BUG_ON(cowonly != upper->cowonly);
1049                 if (!cowonly) {
1050                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1051                                               &upper->rb_node);
1052                         if (rb_node)
1053                                 backref_tree_panic(rb_node, -EEXIST,
1054                                                    upper->bytenr);
1055                 }
1056
1057                 list_add_tail(&edge->list[UPPER], &upper->lower);
1058
1059                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1060                         list_add_tail(&edge->list[UPPER], &list);
1061         }
1062         /*
1063          * process useless backref nodes. backref nodes for tree leaves
1064          * are deleted from the cache. backref nodes for upper level
1065          * tree blocks are left in the cache to avoid unnecessary backref
1066          * lookup.
1067          */
1068         while (!list_empty(&useless)) {
1069                 upper = list_entry(useless.next, struct backref_node, list);
1070                 list_del_init(&upper->list);
1071                 BUG_ON(!list_empty(&upper->upper));
1072                 if (upper == node)
1073                         node = NULL;
1074                 if (upper->lowest) {
1075                         list_del_init(&upper->lower);
1076                         upper->lowest = 0;
1077                 }
1078                 while (!list_empty(&upper->lower)) {
1079                         edge = list_entry(upper->lower.next,
1080                                           struct backref_edge, list[UPPER]);
1081                         list_del(&edge->list[UPPER]);
1082                         list_del(&edge->list[LOWER]);
1083                         lower = edge->node[LOWER];
1084                         free_backref_edge(cache, edge);
1085
1086                         if (list_empty(&lower->upper))
1087                                 list_add(&lower->list, &useless);
1088                 }
1089                 __mark_block_processed(rc, upper);
1090                 if (upper->level > 0) {
1091                         list_add(&upper->list, &cache->detached);
1092                         upper->detached = 1;
1093                 } else {
1094                         rb_erase(&upper->rb_node, &cache->rb_root);
1095                         free_backref_node(cache, upper);
1096                 }
1097         }
1098 out:
1099         btrfs_free_path(path1);
1100         btrfs_free_path(path2);
1101         if (err) {
1102                 while (!list_empty(&useless)) {
1103                         lower = list_entry(useless.next,
1104                                            struct backref_node, upper);
1105                         list_del_init(&lower->upper);
1106                 }
1107                 upper = node;
1108                 INIT_LIST_HEAD(&list);
1109                 while (upper) {
1110                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1111                                 list_splice_tail(&upper->upper, &list);
1112                                 free_backref_node(cache, upper);
1113                         }
1114
1115                         if (list_empty(&list))
1116                                 break;
1117
1118                         edge = list_entry(list.next, struct backref_edge,
1119                                           list[LOWER]);
1120                         list_del(&edge->list[LOWER]);
1121                         upper = edge->node[UPPER];
1122                         free_backref_edge(cache, edge);
1123                 }
1124                 return ERR_PTR(err);
1125         }
1126         BUG_ON(node && node->detached);
1127         return node;
1128 }
1129
1130 /*
1131  * helper to add backref node for the newly created snapshot.
1132  * the backref node is created by cloning backref node that
1133  * corresponds to root of source tree
1134  */
1135 static int clone_backref_node(struct btrfs_trans_handle *trans,
1136                               struct reloc_control *rc,
1137                               struct btrfs_root *src,
1138                               struct btrfs_root *dest)
1139 {
1140         struct btrfs_root *reloc_root = src->reloc_root;
1141         struct backref_cache *cache = &rc->backref_cache;
1142         struct backref_node *node = NULL;
1143         struct backref_node *new_node;
1144         struct backref_edge *edge;
1145         struct backref_edge *new_edge;
1146         struct rb_node *rb_node;
1147
1148         if (cache->last_trans > 0)
1149                 update_backref_cache(trans, cache);
1150
1151         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1152         if (rb_node) {
1153                 node = rb_entry(rb_node, struct backref_node, rb_node);
1154                 if (node->detached)
1155                         node = NULL;
1156                 else
1157                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1158         }
1159
1160         if (!node) {
1161                 rb_node = tree_search(&cache->rb_root,
1162                                       reloc_root->commit_root->start);
1163                 if (rb_node) {
1164                         node = rb_entry(rb_node, struct backref_node,
1165                                         rb_node);
1166                         BUG_ON(node->detached);
1167                 }
1168         }
1169
1170         if (!node)
1171                 return 0;
1172
1173         new_node = alloc_backref_node(cache);
1174         if (!new_node)
1175                 return -ENOMEM;
1176
1177         new_node->bytenr = dest->node->start;
1178         new_node->level = node->level;
1179         new_node->lowest = node->lowest;
1180         new_node->checked = 1;
1181         new_node->root = dest;
1182
1183         if (!node->lowest) {
1184                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1185                         new_edge = alloc_backref_edge(cache);
1186                         if (!new_edge)
1187                                 goto fail;
1188
1189                         new_edge->node[UPPER] = new_node;
1190                         new_edge->node[LOWER] = edge->node[LOWER];
1191                         list_add_tail(&new_edge->list[UPPER],
1192                                       &new_node->lower);
1193                 }
1194         } else {
1195                 list_add_tail(&new_node->lower, &cache->leaves);
1196         }
1197
1198         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1199                               &new_node->rb_node);
1200         if (rb_node)
1201                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1202
1203         if (!new_node->lowest) {
1204                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1205                         list_add_tail(&new_edge->list[LOWER],
1206                                       &new_edge->node[LOWER]->upper);
1207                 }
1208         }
1209         return 0;
1210 fail:
1211         while (!list_empty(&new_node->lower)) {
1212                 new_edge = list_entry(new_node->lower.next,
1213                                       struct backref_edge, list[UPPER]);
1214                 list_del(&new_edge->list[UPPER]);
1215                 free_backref_edge(cache, new_edge);
1216         }
1217         free_backref_node(cache, new_node);
1218         return -ENOMEM;
1219 }
1220
1221 /*
1222  * helper to add 'address of tree root -> reloc tree' mapping
1223  */
1224 static int __must_check __add_reloc_root(struct btrfs_root *root)
1225 {
1226         struct rb_node *rb_node;
1227         struct mapping_node *node;
1228         struct reloc_control *rc = root->fs_info->reloc_ctl;
1229
1230         node = kmalloc(sizeof(*node), GFP_NOFS);
1231         if (!node)
1232                 return -ENOMEM;
1233
1234         node->bytenr = root->node->start;
1235         node->data = root;
1236
1237         spin_lock(&rc->reloc_root_tree.lock);
1238         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1239                               node->bytenr, &node->rb_node);
1240         spin_unlock(&rc->reloc_root_tree.lock);
1241         if (rb_node) {
1242                 kfree(node);
1243                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1244                             "for start=%llu while inserting into relocation "
1245                             "tree\n");
1246         }
1247
1248         list_add_tail(&root->root_list, &rc->reloc_roots);
1249         return 0;
1250 }
1251
1252 /*
1253  * helper to update/delete the 'address of tree root -> reloc tree'
1254  * mapping
1255  */
1256 static int __update_reloc_root(struct btrfs_root *root, int del)
1257 {
1258         struct rb_node *rb_node;
1259         struct mapping_node *node = NULL;
1260         struct reloc_control *rc = root->fs_info->reloc_ctl;
1261
1262         spin_lock(&rc->reloc_root_tree.lock);
1263         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1264                               root->commit_root->start);
1265         if (rb_node) {
1266                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1267                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1268         }
1269         spin_unlock(&rc->reloc_root_tree.lock);
1270
1271         BUG_ON((struct btrfs_root *)node->data != root);
1272
1273         if (!del) {
1274                 spin_lock(&rc->reloc_root_tree.lock);
1275                 node->bytenr = root->node->start;
1276                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1277                                       node->bytenr, &node->rb_node);
1278                 spin_unlock(&rc->reloc_root_tree.lock);
1279                 if (rb_node)
1280                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1281         } else {
1282                 list_del_init(&root->root_list);
1283                 kfree(node);
1284         }
1285         return 0;
1286 }
1287
1288 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1289                                         struct btrfs_root *root, u64 objectid)
1290 {
1291         struct btrfs_root *reloc_root;
1292         struct extent_buffer *eb;
1293         struct btrfs_root_item *root_item;
1294         struct btrfs_key root_key;
1295         int ret;
1296
1297         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1298         BUG_ON(!root_item);
1299
1300         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1301         root_key.type = BTRFS_ROOT_ITEM_KEY;
1302         root_key.offset = objectid;
1303
1304         if (root->root_key.objectid == objectid) {
1305                 /* called by btrfs_init_reloc_root */
1306                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1307                                       BTRFS_TREE_RELOC_OBJECTID);
1308                 BUG_ON(ret);
1309
1310                 btrfs_set_root_last_snapshot(&root->root_item,
1311                                              trans->transid - 1);
1312         } else {
1313                 /*
1314                  * called by btrfs_reloc_post_snapshot_hook.
1315                  * the source tree is a reloc tree, all tree blocks
1316                  * modified after it was created have RELOC flag
1317                  * set in their headers. so it's OK to not update
1318                  * the 'last_snapshot'.
1319                  */
1320                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1321                                       BTRFS_TREE_RELOC_OBJECTID);
1322                 BUG_ON(ret);
1323         }
1324
1325         memcpy(root_item, &root->root_item, sizeof(*root_item));
1326         btrfs_set_root_bytenr(root_item, eb->start);
1327         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1328         btrfs_set_root_generation(root_item, trans->transid);
1329
1330         if (root->root_key.objectid == objectid) {
1331                 btrfs_set_root_refs(root_item, 0);
1332                 memset(&root_item->drop_progress, 0,
1333                        sizeof(struct btrfs_disk_key));
1334                 root_item->drop_level = 0;
1335         }
1336
1337         btrfs_tree_unlock(eb);
1338         free_extent_buffer(eb);
1339
1340         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1341                                 &root_key, root_item);
1342         BUG_ON(ret);
1343         kfree(root_item);
1344
1345         reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
1346                                                  &root_key);
1347         BUG_ON(IS_ERR(reloc_root));
1348         reloc_root->last_trans = trans->transid;
1349         return reloc_root;
1350 }
1351
1352 /*
1353  * create reloc tree for a given fs tree. reloc tree is just a
1354  * snapshot of the fs tree with special root objectid.
1355  */
1356 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1357                           struct btrfs_root *root)
1358 {
1359         struct btrfs_root *reloc_root;
1360         struct reloc_control *rc = root->fs_info->reloc_ctl;
1361         int clear_rsv = 0;
1362         int ret;
1363
1364         if (root->reloc_root) {
1365                 reloc_root = root->reloc_root;
1366                 reloc_root->last_trans = trans->transid;
1367                 return 0;
1368         }
1369
1370         if (!rc || !rc->create_reloc_tree ||
1371             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1372                 return 0;
1373
1374         if (!trans->block_rsv) {
1375                 trans->block_rsv = rc->block_rsv;
1376                 clear_rsv = 1;
1377         }
1378         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1379         if (clear_rsv)
1380                 trans->block_rsv = NULL;
1381
1382         ret = __add_reloc_root(reloc_root);
1383         BUG_ON(ret < 0);
1384         root->reloc_root = reloc_root;
1385         return 0;
1386 }
1387
1388 /*
1389  * update root item of reloc tree
1390  */
1391 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1392                             struct btrfs_root *root)
1393 {
1394         struct btrfs_root *reloc_root;
1395         struct btrfs_root_item *root_item;
1396         int del = 0;
1397         int ret;
1398
1399         if (!root->reloc_root)
1400                 goto out;
1401
1402         reloc_root = root->reloc_root;
1403         root_item = &reloc_root->root_item;
1404
1405         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1406             btrfs_root_refs(root_item) == 0) {
1407                 root->reloc_root = NULL;
1408                 del = 1;
1409         }
1410
1411         __update_reloc_root(reloc_root, del);
1412
1413         if (reloc_root->commit_root != reloc_root->node) {
1414                 btrfs_set_root_node(root_item, reloc_root->node);
1415                 free_extent_buffer(reloc_root->commit_root);
1416                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1417         }
1418
1419         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1420                                 &reloc_root->root_key, root_item);
1421         BUG_ON(ret);
1422
1423 out:
1424         return 0;
1425 }
1426
1427 /*
1428  * helper to find first cached inode with inode number >= objectid
1429  * in a subvolume
1430  */
1431 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1432 {
1433         struct rb_node *node;
1434         struct rb_node *prev;
1435         struct btrfs_inode *entry;
1436         struct inode *inode;
1437
1438         spin_lock(&root->inode_lock);
1439 again:
1440         node = root->inode_tree.rb_node;
1441         prev = NULL;
1442         while (node) {
1443                 prev = node;
1444                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1445
1446                 if (objectid < btrfs_ino(&entry->vfs_inode))
1447                         node = node->rb_left;
1448                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1449                         node = node->rb_right;
1450                 else
1451                         break;
1452         }
1453         if (!node) {
1454                 while (prev) {
1455                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1456                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1457                                 node = prev;
1458                                 break;
1459                         }
1460                         prev = rb_next(prev);
1461                 }
1462         }
1463         while (node) {
1464                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1465                 inode = igrab(&entry->vfs_inode);
1466                 if (inode) {
1467                         spin_unlock(&root->inode_lock);
1468                         return inode;
1469                 }
1470
1471                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1472                 if (cond_resched_lock(&root->inode_lock))
1473                         goto again;
1474
1475                 node = rb_next(node);
1476         }
1477         spin_unlock(&root->inode_lock);
1478         return NULL;
1479 }
1480
1481 static int in_block_group(u64 bytenr,
1482                           struct btrfs_block_group_cache *block_group)
1483 {
1484         if (bytenr >= block_group->key.objectid &&
1485             bytenr < block_group->key.objectid + block_group->key.offset)
1486                 return 1;
1487         return 0;
1488 }
1489
1490 /*
1491  * get new location of data
1492  */
1493 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1494                             u64 bytenr, u64 num_bytes)
1495 {
1496         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1497         struct btrfs_path *path;
1498         struct btrfs_file_extent_item *fi;
1499         struct extent_buffer *leaf;
1500         int ret;
1501
1502         path = btrfs_alloc_path();
1503         if (!path)
1504                 return -ENOMEM;
1505
1506         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1507         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1508                                        bytenr, 0);
1509         if (ret < 0)
1510                 goto out;
1511         if (ret > 0) {
1512                 ret = -ENOENT;
1513                 goto out;
1514         }
1515
1516         leaf = path->nodes[0];
1517         fi = btrfs_item_ptr(leaf, path->slots[0],
1518                             struct btrfs_file_extent_item);
1519
1520         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1521                btrfs_file_extent_compression(leaf, fi) ||
1522                btrfs_file_extent_encryption(leaf, fi) ||
1523                btrfs_file_extent_other_encoding(leaf, fi));
1524
1525         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1526                 ret = 1;
1527                 goto out;
1528         }
1529
1530         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1531         ret = 0;
1532 out:
1533         btrfs_free_path(path);
1534         return ret;
1535 }
1536
1537 /*
1538  * update file extent items in the tree leaf to point to
1539  * the new locations.
1540  */
1541 static noinline_for_stack
1542 int replace_file_extents(struct btrfs_trans_handle *trans,
1543                          struct reloc_control *rc,
1544                          struct btrfs_root *root,
1545                          struct extent_buffer *leaf)
1546 {
1547         struct btrfs_key key;
1548         struct btrfs_file_extent_item *fi;
1549         struct inode *inode = NULL;
1550         u64 parent;
1551         u64 bytenr;
1552         u64 new_bytenr = 0;
1553         u64 num_bytes;
1554         u64 end;
1555         u32 nritems;
1556         u32 i;
1557         int ret;
1558         int first = 1;
1559         int dirty = 0;
1560
1561         if (rc->stage != UPDATE_DATA_PTRS)
1562                 return 0;
1563
1564         /* reloc trees always use full backref */
1565         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1566                 parent = leaf->start;
1567         else
1568                 parent = 0;
1569
1570         nritems = btrfs_header_nritems(leaf);
1571         for (i = 0; i < nritems; i++) {
1572                 cond_resched();
1573                 btrfs_item_key_to_cpu(leaf, &key, i);
1574                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1575                         continue;
1576                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1577                 if (btrfs_file_extent_type(leaf, fi) ==
1578                     BTRFS_FILE_EXTENT_INLINE)
1579                         continue;
1580                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1581                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1582                 if (bytenr == 0)
1583                         continue;
1584                 if (!in_block_group(bytenr, rc->block_group))
1585                         continue;
1586
1587                 /*
1588                  * if we are modifying block in fs tree, wait for readpage
1589                  * to complete and drop the extent cache
1590                  */
1591                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1592                         if (first) {
1593                                 inode = find_next_inode(root, key.objectid);
1594                                 first = 0;
1595                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1596                                 btrfs_add_delayed_iput(inode);
1597                                 inode = find_next_inode(root, key.objectid);
1598                         }
1599                         if (inode && btrfs_ino(inode) == key.objectid) {
1600                                 end = key.offset +
1601                                       btrfs_file_extent_num_bytes(leaf, fi);
1602                                 WARN_ON(!IS_ALIGNED(key.offset,
1603                                                     root->sectorsize));
1604                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1605                                 end--;
1606                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1607                                                       key.offset, end);
1608                                 if (!ret)
1609                                         continue;
1610
1611                                 btrfs_drop_extent_cache(inode, key.offset, end,
1612                                                         1);
1613                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1614                                               key.offset, end);
1615                         }
1616                 }
1617
1618                 ret = get_new_location(rc->data_inode, &new_bytenr,
1619                                        bytenr, num_bytes);
1620                 if (ret > 0) {
1621                         WARN_ON(1);
1622                         continue;
1623                 }
1624                 BUG_ON(ret < 0);
1625
1626                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1627                 dirty = 1;
1628
1629                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1630                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1631                                            num_bytes, parent,
1632                                            btrfs_header_owner(leaf),
1633                                            key.objectid, key.offset, 1);
1634                 BUG_ON(ret);
1635
1636                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1637                                         parent, btrfs_header_owner(leaf),
1638                                         key.objectid, key.offset, 1);
1639                 BUG_ON(ret);
1640         }
1641         if (dirty)
1642                 btrfs_mark_buffer_dirty(leaf);
1643         if (inode)
1644                 btrfs_add_delayed_iput(inode);
1645         return 0;
1646 }
1647
1648 static noinline_for_stack
1649 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1650                      struct btrfs_path *path, int level)
1651 {
1652         struct btrfs_disk_key key1;
1653         struct btrfs_disk_key key2;
1654         btrfs_node_key(eb, &key1, slot);
1655         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1656         return memcmp(&key1, &key2, sizeof(key1));
1657 }
1658
1659 /*
1660  * try to replace tree blocks in fs tree with the new blocks
1661  * in reloc tree. tree blocks haven't been modified since the
1662  * reloc tree was create can be replaced.
1663  *
1664  * if a block was replaced, level of the block + 1 is returned.
1665  * if no block got replaced, 0 is returned. if there are other
1666  * errors, a negative error number is returned.
1667  */
1668 static noinline_for_stack
1669 int replace_path(struct btrfs_trans_handle *trans,
1670                  struct btrfs_root *dest, struct btrfs_root *src,
1671                  struct btrfs_path *path, struct btrfs_key *next_key,
1672                  int lowest_level, int max_level)
1673 {
1674         struct extent_buffer *eb;
1675         struct extent_buffer *parent;
1676         struct btrfs_key key;
1677         u64 old_bytenr;
1678         u64 new_bytenr;
1679         u64 old_ptr_gen;
1680         u64 new_ptr_gen;
1681         u64 last_snapshot;
1682         u32 blocksize;
1683         int cow = 0;
1684         int level;
1685         int ret;
1686         int slot;
1687
1688         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1689         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1690
1691         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1692 again:
1693         slot = path->slots[lowest_level];
1694         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1695
1696         eb = btrfs_lock_root_node(dest);
1697         btrfs_set_lock_blocking(eb);
1698         level = btrfs_header_level(eb);
1699
1700         if (level < lowest_level) {
1701                 btrfs_tree_unlock(eb);
1702                 free_extent_buffer(eb);
1703                 return 0;
1704         }
1705
1706         if (cow) {
1707                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1708                 BUG_ON(ret);
1709         }
1710         btrfs_set_lock_blocking(eb);
1711
1712         if (next_key) {
1713                 next_key->objectid = (u64)-1;
1714                 next_key->type = (u8)-1;
1715                 next_key->offset = (u64)-1;
1716         }
1717
1718         parent = eb;
1719         while (1) {
1720                 level = btrfs_header_level(parent);
1721                 BUG_ON(level < lowest_level);
1722
1723                 ret = btrfs_bin_search(parent, &key, level, &slot);
1724                 if (ret && slot > 0)
1725                         slot--;
1726
1727                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1728                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1729
1730                 old_bytenr = btrfs_node_blockptr(parent, slot);
1731                 blocksize = btrfs_level_size(dest, level - 1);
1732                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1733
1734                 if (level <= max_level) {
1735                         eb = path->nodes[level];
1736                         new_bytenr = btrfs_node_blockptr(eb,
1737                                                         path->slots[level]);
1738                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1739                                                         path->slots[level]);
1740                 } else {
1741                         new_bytenr = 0;
1742                         new_ptr_gen = 0;
1743                 }
1744
1745                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1746                         WARN_ON(1);
1747                         ret = level;
1748                         break;
1749                 }
1750
1751                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1752                     memcmp_node_keys(parent, slot, path, level)) {
1753                         if (level <= lowest_level) {
1754                                 ret = 0;
1755                                 break;
1756                         }
1757
1758                         eb = read_tree_block(dest, old_bytenr, blocksize,
1759                                              old_ptr_gen);
1760                         BUG_ON(!eb);
1761                         btrfs_tree_lock(eb);
1762                         if (cow) {
1763                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1764                                                       slot, &eb);
1765                                 BUG_ON(ret);
1766                         }
1767                         btrfs_set_lock_blocking(eb);
1768
1769                         btrfs_tree_unlock(parent);
1770                         free_extent_buffer(parent);
1771
1772                         parent = eb;
1773                         continue;
1774                 }
1775
1776                 if (!cow) {
1777                         btrfs_tree_unlock(parent);
1778                         free_extent_buffer(parent);
1779                         cow = 1;
1780                         goto again;
1781                 }
1782
1783                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1784                                       path->slots[level]);
1785                 btrfs_release_path(path);
1786
1787                 path->lowest_level = level;
1788                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1789                 path->lowest_level = 0;
1790                 BUG_ON(ret);
1791
1792                 /*
1793                  * swap blocks in fs tree and reloc tree.
1794                  */
1795                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1796                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1797                 btrfs_mark_buffer_dirty(parent);
1798
1799                 btrfs_set_node_blockptr(path->nodes[level],
1800                                         path->slots[level], old_bytenr);
1801                 btrfs_set_node_ptr_generation(path->nodes[level],
1802                                               path->slots[level], old_ptr_gen);
1803                 btrfs_mark_buffer_dirty(path->nodes[level]);
1804
1805                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1806                                         path->nodes[level]->start,
1807                                         src->root_key.objectid, level - 1, 0,
1808                                         1);
1809                 BUG_ON(ret);
1810                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1811                                         0, dest->root_key.objectid, level - 1,
1812                                         0, 1);
1813                 BUG_ON(ret);
1814
1815                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1816                                         path->nodes[level]->start,
1817                                         src->root_key.objectid, level - 1, 0,
1818                                         1);
1819                 BUG_ON(ret);
1820
1821                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1822                                         0, dest->root_key.objectid, level - 1,
1823                                         0, 1);
1824                 BUG_ON(ret);
1825
1826                 btrfs_unlock_up_safe(path, 0);
1827
1828                 ret = level;
1829                 break;
1830         }
1831         btrfs_tree_unlock(parent);
1832         free_extent_buffer(parent);
1833         return ret;
1834 }
1835
1836 /*
1837  * helper to find next relocated block in reloc tree
1838  */
1839 static noinline_for_stack
1840 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1841                        int *level)
1842 {
1843         struct extent_buffer *eb;
1844         int i;
1845         u64 last_snapshot;
1846         u32 nritems;
1847
1848         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1849
1850         for (i = 0; i < *level; i++) {
1851                 free_extent_buffer(path->nodes[i]);
1852                 path->nodes[i] = NULL;
1853         }
1854
1855         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1856                 eb = path->nodes[i];
1857                 nritems = btrfs_header_nritems(eb);
1858                 while (path->slots[i] + 1 < nritems) {
1859                         path->slots[i]++;
1860                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1861                             last_snapshot)
1862                                 continue;
1863
1864                         *level = i;
1865                         return 0;
1866                 }
1867                 free_extent_buffer(path->nodes[i]);
1868                 path->nodes[i] = NULL;
1869         }
1870         return 1;
1871 }
1872
1873 /*
1874  * walk down reloc tree to find relocated block of lowest level
1875  */
1876 static noinline_for_stack
1877 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1878                          int *level)
1879 {
1880         struct extent_buffer *eb = NULL;
1881         int i;
1882         u64 bytenr;
1883         u64 ptr_gen = 0;
1884         u64 last_snapshot;
1885         u32 blocksize;
1886         u32 nritems;
1887
1888         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1889
1890         for (i = *level; i > 0; i--) {
1891                 eb = path->nodes[i];
1892                 nritems = btrfs_header_nritems(eb);
1893                 while (path->slots[i] < nritems) {
1894                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1895                         if (ptr_gen > last_snapshot)
1896                                 break;
1897                         path->slots[i]++;
1898                 }
1899                 if (path->slots[i] >= nritems) {
1900                         if (i == *level)
1901                                 break;
1902                         *level = i + 1;
1903                         return 0;
1904                 }
1905                 if (i == 1) {
1906                         *level = i;
1907                         return 0;
1908                 }
1909
1910                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1911                 blocksize = btrfs_level_size(root, i - 1);
1912                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1913                 BUG_ON(btrfs_header_level(eb) != i - 1);
1914                 path->nodes[i - 1] = eb;
1915                 path->slots[i - 1] = 0;
1916         }
1917         return 1;
1918 }
1919
1920 /*
1921  * invalidate extent cache for file extents whose key in range of
1922  * [min_key, max_key)
1923  */
1924 static int invalidate_extent_cache(struct btrfs_root *root,
1925                                    struct btrfs_key *min_key,
1926                                    struct btrfs_key *max_key)
1927 {
1928         struct inode *inode = NULL;
1929         u64 objectid;
1930         u64 start, end;
1931         u64 ino;
1932
1933         objectid = min_key->objectid;
1934         while (1) {
1935                 cond_resched();
1936                 iput(inode);
1937
1938                 if (objectid > max_key->objectid)
1939                         break;
1940
1941                 inode = find_next_inode(root, objectid);
1942                 if (!inode)
1943                         break;
1944                 ino = btrfs_ino(inode);
1945
1946                 if (ino > max_key->objectid) {
1947                         iput(inode);
1948                         break;
1949                 }
1950
1951                 objectid = ino + 1;
1952                 if (!S_ISREG(inode->i_mode))
1953                         continue;
1954
1955                 if (unlikely(min_key->objectid == ino)) {
1956                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1957                                 continue;
1958                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1959                                 start = 0;
1960                         else {
1961                                 start = min_key->offset;
1962                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1963                         }
1964                 } else {
1965                         start = 0;
1966                 }
1967
1968                 if (unlikely(max_key->objectid == ino)) {
1969                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1970                                 continue;
1971                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1972                                 end = (u64)-1;
1973                         } else {
1974                                 if (max_key->offset == 0)
1975                                         continue;
1976                                 end = max_key->offset;
1977                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1978                                 end--;
1979                         }
1980                 } else {
1981                         end = (u64)-1;
1982                 }
1983
1984                 /* the lock_extent waits for readpage to complete */
1985                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
1986                 btrfs_drop_extent_cache(inode, start, end, 1);
1987                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1988         }
1989         return 0;
1990 }
1991
1992 static int find_next_key(struct btrfs_path *path, int level,
1993                          struct btrfs_key *key)
1994
1995 {
1996         while (level < BTRFS_MAX_LEVEL) {
1997                 if (!path->nodes[level])
1998                         break;
1999                 if (path->slots[level] + 1 <
2000                     btrfs_header_nritems(path->nodes[level])) {
2001                         btrfs_node_key_to_cpu(path->nodes[level], key,
2002                                               path->slots[level] + 1);
2003                         return 0;
2004                 }
2005                 level++;
2006         }
2007         return 1;
2008 }
2009
2010 /*
2011  * merge the relocated tree blocks in reloc tree with corresponding
2012  * fs tree.
2013  */
2014 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2015                                                struct btrfs_root *root)
2016 {
2017         LIST_HEAD(inode_list);
2018         struct btrfs_key key;
2019         struct btrfs_key next_key;
2020         struct btrfs_trans_handle *trans;
2021         struct btrfs_root *reloc_root;
2022         struct btrfs_root_item *root_item;
2023         struct btrfs_path *path;
2024         struct extent_buffer *leaf;
2025         unsigned long nr;
2026         int level;
2027         int max_level;
2028         int replaced = 0;
2029         int ret;
2030         int err = 0;
2031         u32 min_reserved;
2032
2033         path = btrfs_alloc_path();
2034         if (!path)
2035                 return -ENOMEM;
2036         path->reada = 1;
2037
2038         reloc_root = root->reloc_root;
2039         root_item = &reloc_root->root_item;
2040
2041         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2042                 level = btrfs_root_level(root_item);
2043                 extent_buffer_get(reloc_root->node);
2044                 path->nodes[level] = reloc_root->node;
2045                 path->slots[level] = 0;
2046         } else {
2047                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2048
2049                 level = root_item->drop_level;
2050                 BUG_ON(level == 0);
2051                 path->lowest_level = level;
2052                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2053                 path->lowest_level = 0;
2054                 if (ret < 0) {
2055                         btrfs_free_path(path);
2056                         return ret;
2057                 }
2058
2059                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2060                                       path->slots[level]);
2061                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2062
2063                 btrfs_unlock_up_safe(path, 0);
2064         }
2065
2066         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2067         memset(&next_key, 0, sizeof(next_key));
2068
2069         while (1) {
2070                 trans = btrfs_start_transaction(root, 0);
2071                 BUG_ON(IS_ERR(trans));
2072                 trans->block_rsv = rc->block_rsv;
2073
2074                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved);
2075                 if (ret) {
2076                         BUG_ON(ret != -EAGAIN);
2077                         ret = btrfs_commit_transaction(trans, root);
2078                         BUG_ON(ret);
2079                         continue;
2080                 }
2081
2082                 replaced = 0;
2083                 max_level = level;
2084
2085                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2086                 if (ret < 0) {
2087                         err = ret;
2088                         goto out;
2089                 }
2090                 if (ret > 0)
2091                         break;
2092
2093                 if (!find_next_key(path, level, &key) &&
2094                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2095                         ret = 0;
2096                 } else {
2097                         ret = replace_path(trans, root, reloc_root, path,
2098                                            &next_key, level, max_level);
2099                 }
2100                 if (ret < 0) {
2101                         err = ret;
2102                         goto out;
2103                 }
2104
2105                 if (ret > 0) {
2106                         level = ret;
2107                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2108                                               path->slots[level]);
2109                         replaced = 1;
2110                 }
2111
2112                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2113                 if (ret > 0)
2114                         break;
2115
2116                 BUG_ON(level == 0);
2117                 /*
2118                  * save the merging progress in the drop_progress.
2119                  * this is OK since root refs == 1 in this case.
2120                  */
2121                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2122                                path->slots[level]);
2123                 root_item->drop_level = level;
2124
2125                 nr = trans->blocks_used;
2126                 btrfs_end_transaction_throttle(trans, root);
2127
2128                 btrfs_btree_balance_dirty(root, nr);
2129
2130                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2131                         invalidate_extent_cache(root, &key, &next_key);
2132         }
2133
2134         /*
2135          * handle the case only one block in the fs tree need to be
2136          * relocated and the block is tree root.
2137          */
2138         leaf = btrfs_lock_root_node(root);
2139         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2140         btrfs_tree_unlock(leaf);
2141         free_extent_buffer(leaf);
2142         if (ret < 0)
2143                 err = ret;
2144 out:
2145         btrfs_free_path(path);
2146
2147         if (err == 0) {
2148                 memset(&root_item->drop_progress, 0,
2149                        sizeof(root_item->drop_progress));
2150                 root_item->drop_level = 0;
2151                 btrfs_set_root_refs(root_item, 0);
2152                 btrfs_update_reloc_root(trans, root);
2153         }
2154
2155         nr = trans->blocks_used;
2156         btrfs_end_transaction_throttle(trans, root);
2157
2158         btrfs_btree_balance_dirty(root, nr);
2159
2160         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2161                 invalidate_extent_cache(root, &key, &next_key);
2162
2163         return err;
2164 }
2165
2166 static noinline_for_stack
2167 int prepare_to_merge(struct reloc_control *rc, int err)
2168 {
2169         struct btrfs_root *root = rc->extent_root;
2170         struct btrfs_root *reloc_root;
2171         struct btrfs_trans_handle *trans;
2172         LIST_HEAD(reloc_roots);
2173         u64 num_bytes = 0;
2174         int ret;
2175
2176         mutex_lock(&root->fs_info->reloc_mutex);
2177         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2178         rc->merging_rsv_size += rc->nodes_relocated * 2;
2179         mutex_unlock(&root->fs_info->reloc_mutex);
2180
2181 again:
2182         if (!err) {
2183                 num_bytes = rc->merging_rsv_size;
2184                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes);
2185                 if (ret)
2186                         err = ret;
2187         }
2188
2189         trans = btrfs_join_transaction(rc->extent_root);
2190         if (IS_ERR(trans)) {
2191                 if (!err)
2192                         btrfs_block_rsv_release(rc->extent_root,
2193                                                 rc->block_rsv, num_bytes);
2194                 return PTR_ERR(trans);
2195         }
2196
2197         if (!err) {
2198                 if (num_bytes != rc->merging_rsv_size) {
2199                         btrfs_end_transaction(trans, rc->extent_root);
2200                         btrfs_block_rsv_release(rc->extent_root,
2201                                                 rc->block_rsv, num_bytes);
2202                         goto again;
2203                 }
2204         }
2205
2206         rc->merge_reloc_tree = 1;
2207
2208         while (!list_empty(&rc->reloc_roots)) {
2209                 reloc_root = list_entry(rc->reloc_roots.next,
2210                                         struct btrfs_root, root_list);
2211                 list_del_init(&reloc_root->root_list);
2212
2213                 root = read_fs_root(reloc_root->fs_info,
2214                                     reloc_root->root_key.offset);
2215                 BUG_ON(IS_ERR(root));
2216                 BUG_ON(root->reloc_root != reloc_root);
2217
2218                 /*
2219                  * set reference count to 1, so btrfs_recover_relocation
2220                  * knows it should resumes merging
2221                  */
2222                 if (!err)
2223                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2224                 btrfs_update_reloc_root(trans, root);
2225
2226                 list_add(&reloc_root->root_list, &reloc_roots);
2227         }
2228
2229         list_splice(&reloc_roots, &rc->reloc_roots);
2230
2231         if (!err)
2232                 btrfs_commit_transaction(trans, rc->extent_root);
2233         else
2234                 btrfs_end_transaction(trans, rc->extent_root);
2235         return err;
2236 }
2237
2238 static noinline_for_stack
2239 int merge_reloc_roots(struct reloc_control *rc)
2240 {
2241         struct btrfs_root *root;
2242         struct btrfs_root *reloc_root;
2243         LIST_HEAD(reloc_roots);
2244         int found = 0;
2245         int ret;
2246 again:
2247         root = rc->extent_root;
2248
2249         /*
2250          * this serializes us with btrfs_record_root_in_transaction,
2251          * we have to make sure nobody is in the middle of
2252          * adding their roots to the list while we are
2253          * doing this splice
2254          */
2255         mutex_lock(&root->fs_info->reloc_mutex);
2256         list_splice_init(&rc->reloc_roots, &reloc_roots);
2257         mutex_unlock(&root->fs_info->reloc_mutex);
2258
2259         while (!list_empty(&reloc_roots)) {
2260                 found = 1;
2261                 reloc_root = list_entry(reloc_roots.next,
2262                                         struct btrfs_root, root_list);
2263
2264                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2265                         root = read_fs_root(reloc_root->fs_info,
2266                                             reloc_root->root_key.offset);
2267                         BUG_ON(IS_ERR(root));
2268                         BUG_ON(root->reloc_root != reloc_root);
2269
2270                         ret = merge_reloc_root(rc, root);
2271                         BUG_ON(ret);
2272                 } else {
2273                         list_del_init(&reloc_root->root_list);
2274                 }
2275                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2276                 BUG_ON(ret < 0);
2277         }
2278
2279         if (found) {
2280                 found = 0;
2281                 goto again;
2282         }
2283         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2284         return 0;
2285 }
2286
2287 static void free_block_list(struct rb_root *blocks)
2288 {
2289         struct tree_block *block;
2290         struct rb_node *rb_node;
2291         while ((rb_node = rb_first(blocks))) {
2292                 block = rb_entry(rb_node, struct tree_block, rb_node);
2293                 rb_erase(rb_node, blocks);
2294                 kfree(block);
2295         }
2296 }
2297
2298 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2299                                       struct btrfs_root *reloc_root)
2300 {
2301         struct btrfs_root *root;
2302
2303         if (reloc_root->last_trans == trans->transid)
2304                 return 0;
2305
2306         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2307         BUG_ON(IS_ERR(root));
2308         BUG_ON(root->reloc_root != reloc_root);
2309
2310         return btrfs_record_root_in_trans(trans, root);
2311 }
2312
2313 static noinline_for_stack
2314 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2315                                      struct reloc_control *rc,
2316                                      struct backref_node *node,
2317                                      struct backref_edge *edges[], int *nr)
2318 {
2319         struct backref_node *next;
2320         struct btrfs_root *root;
2321         int index = 0;
2322
2323         next = node;
2324         while (1) {
2325                 cond_resched();
2326                 next = walk_up_backref(next, edges, &index);
2327                 root = next->root;
2328                 BUG_ON(!root);
2329                 BUG_ON(!root->ref_cows);
2330
2331                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2332                         record_reloc_root_in_trans(trans, root);
2333                         break;
2334                 }
2335
2336                 btrfs_record_root_in_trans(trans, root);
2337                 root = root->reloc_root;
2338
2339                 if (next->new_bytenr != root->node->start) {
2340                         BUG_ON(next->new_bytenr);
2341                         BUG_ON(!list_empty(&next->list));
2342                         next->new_bytenr = root->node->start;
2343                         next->root = root;
2344                         list_add_tail(&next->list,
2345                                       &rc->backref_cache.changed);
2346                         __mark_block_processed(rc, next);
2347                         break;
2348                 }
2349
2350                 WARN_ON(1);
2351                 root = NULL;
2352                 next = walk_down_backref(edges, &index);
2353                 if (!next || next->level <= node->level)
2354                         break;
2355         }
2356         if (!root)
2357                 return NULL;
2358
2359         *nr = index;
2360         next = node;
2361         /* setup backref node path for btrfs_reloc_cow_block */
2362         while (1) {
2363                 rc->backref_cache.path[next->level] = next;
2364                 if (--index < 0)
2365                         break;
2366                 next = edges[index]->node[UPPER];
2367         }
2368         return root;
2369 }
2370
2371 /*
2372  * select a tree root for relocation. return NULL if the block
2373  * is reference counted. we should use do_relocation() in this
2374  * case. return a tree root pointer if the block isn't reference
2375  * counted. return -ENOENT if the block is root of reloc tree.
2376  */
2377 static noinline_for_stack
2378 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2379                                    struct backref_node *node)
2380 {
2381         struct backref_node *next;
2382         struct btrfs_root *root;
2383         struct btrfs_root *fs_root = NULL;
2384         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2385         int index = 0;
2386
2387         next = node;
2388         while (1) {
2389                 cond_resched();
2390                 next = walk_up_backref(next, edges, &index);
2391                 root = next->root;
2392                 BUG_ON(!root);
2393
2394                 /* no other choice for non-references counted tree */
2395                 if (!root->ref_cows)
2396                         return root;
2397
2398                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2399                         fs_root = root;
2400
2401                 if (next != node)
2402                         return NULL;
2403
2404                 next = walk_down_backref(edges, &index);
2405                 if (!next || next->level <= node->level)
2406                         break;
2407         }
2408
2409         if (!fs_root)
2410                 return ERR_PTR(-ENOENT);
2411         return fs_root;
2412 }
2413
2414 static noinline_for_stack
2415 u64 calcu_metadata_size(struct reloc_control *rc,
2416                         struct backref_node *node, int reserve)
2417 {
2418         struct backref_node *next = node;
2419         struct backref_edge *edge;
2420         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2421         u64 num_bytes = 0;
2422         int index = 0;
2423
2424         BUG_ON(reserve && node->processed);
2425
2426         while (next) {
2427                 cond_resched();
2428                 while (1) {
2429                         if (next->processed && (reserve || next != node))
2430                                 break;
2431
2432                         num_bytes += btrfs_level_size(rc->extent_root,
2433                                                       next->level);
2434
2435                         if (list_empty(&next->upper))
2436                                 break;
2437
2438                         edge = list_entry(next->upper.next,
2439                                           struct backref_edge, list[LOWER]);
2440                         edges[index++] = edge;
2441                         next = edge->node[UPPER];
2442                 }
2443                 next = walk_down_backref(edges, &index);
2444         }
2445         return num_bytes;
2446 }
2447
2448 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2449                                   struct reloc_control *rc,
2450                                   struct backref_node *node)
2451 {
2452         struct btrfs_root *root = rc->extent_root;
2453         u64 num_bytes;
2454         int ret;
2455
2456         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2457
2458         trans->block_rsv = rc->block_rsv;
2459         ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes);
2460         if (ret) {
2461                 if (ret == -EAGAIN)
2462                         rc->commit_transaction = 1;
2463                 return ret;
2464         }
2465
2466         return 0;
2467 }
2468
2469 static void release_metadata_space(struct reloc_control *rc,
2470                                    struct backref_node *node)
2471 {
2472         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2473         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2474 }
2475
2476 /*
2477  * relocate a block tree, and then update pointers in upper level
2478  * blocks that reference the block to point to the new location.
2479  *
2480  * if called by link_to_upper, the block has already been relocated.
2481  * in that case this function just updates pointers.
2482  */
2483 static int do_relocation(struct btrfs_trans_handle *trans,
2484                          struct reloc_control *rc,
2485                          struct backref_node *node,
2486                          struct btrfs_key *key,
2487                          struct btrfs_path *path, int lowest)
2488 {
2489         struct backref_node *upper;
2490         struct backref_edge *edge;
2491         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2492         struct btrfs_root *root;
2493         struct extent_buffer *eb;
2494         u32 blocksize;
2495         u64 bytenr;
2496         u64 generation;
2497         int nr;
2498         int slot;
2499         int ret;
2500         int err = 0;
2501
2502         BUG_ON(lowest && node->eb);
2503
2504         path->lowest_level = node->level + 1;
2505         rc->backref_cache.path[node->level] = node;
2506         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2507                 cond_resched();
2508
2509                 upper = edge->node[UPPER];
2510                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2511                 BUG_ON(!root);
2512
2513                 if (upper->eb && !upper->locked) {
2514                         if (!lowest) {
2515                                 ret = btrfs_bin_search(upper->eb, key,
2516                                                        upper->level, &slot);
2517                                 BUG_ON(ret);
2518                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2519                                 if (node->eb->start == bytenr)
2520                                         goto next;
2521                         }
2522                         drop_node_buffer(upper);
2523                 }
2524
2525                 if (!upper->eb) {
2526                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2527                         if (ret < 0) {
2528                                 err = ret;
2529                                 break;
2530                         }
2531                         BUG_ON(ret > 0);
2532
2533                         if (!upper->eb) {
2534                                 upper->eb = path->nodes[upper->level];
2535                                 path->nodes[upper->level] = NULL;
2536                         } else {
2537                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2538                         }
2539
2540                         upper->locked = 1;
2541                         path->locks[upper->level] = 0;
2542
2543                         slot = path->slots[upper->level];
2544                         btrfs_release_path(path);
2545                 } else {
2546                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2547                                                &slot);
2548                         BUG_ON(ret);
2549                 }
2550
2551                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2552                 if (lowest) {
2553                         BUG_ON(bytenr != node->bytenr);
2554                 } else {
2555                         if (node->eb->start == bytenr)
2556                                 goto next;
2557                 }
2558
2559                 blocksize = btrfs_level_size(root, node->level);
2560                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2561                 eb = read_tree_block(root, bytenr, blocksize, generation);
2562                 if (!eb) {
2563                         err = -EIO;
2564                         goto next;
2565                 }
2566                 btrfs_tree_lock(eb);
2567                 btrfs_set_lock_blocking(eb);
2568
2569                 if (!node->eb) {
2570                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2571                                               slot, &eb);
2572                         btrfs_tree_unlock(eb);
2573                         free_extent_buffer(eb);
2574                         if (ret < 0) {
2575                                 err = ret;
2576                                 goto next;
2577                         }
2578                         BUG_ON(node->eb != eb);
2579                 } else {
2580                         btrfs_set_node_blockptr(upper->eb, slot,
2581                                                 node->eb->start);
2582                         btrfs_set_node_ptr_generation(upper->eb, slot,
2583                                                       trans->transid);
2584                         btrfs_mark_buffer_dirty(upper->eb);
2585
2586                         ret = btrfs_inc_extent_ref(trans, root,
2587                                                 node->eb->start, blocksize,
2588                                                 upper->eb->start,
2589                                                 btrfs_header_owner(upper->eb),
2590                                                 node->level, 0, 1);
2591                         BUG_ON(ret);
2592
2593                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2594                         BUG_ON(ret);
2595                 }
2596 next:
2597                 if (!upper->pending)
2598                         drop_node_buffer(upper);
2599                 else
2600                         unlock_node_buffer(upper);
2601                 if (err)
2602                         break;
2603         }
2604
2605         if (!err && node->pending) {
2606                 drop_node_buffer(node);
2607                 list_move_tail(&node->list, &rc->backref_cache.changed);
2608                 node->pending = 0;
2609         }
2610
2611         path->lowest_level = 0;
2612         BUG_ON(err == -ENOSPC);
2613         return err;
2614 }
2615
2616 static int link_to_upper(struct btrfs_trans_handle *trans,
2617                          struct reloc_control *rc,
2618                          struct backref_node *node,
2619                          struct btrfs_path *path)
2620 {
2621         struct btrfs_key key;
2622
2623         btrfs_node_key_to_cpu(node->eb, &key, 0);
2624         return do_relocation(trans, rc, node, &key, path, 0);
2625 }
2626
2627 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2628                                 struct reloc_control *rc,
2629                                 struct btrfs_path *path, int err)
2630 {
2631         LIST_HEAD(list);
2632         struct backref_cache *cache = &rc->backref_cache;
2633         struct backref_node *node;
2634         int level;
2635         int ret;
2636
2637         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2638                 while (!list_empty(&cache->pending[level])) {
2639                         node = list_entry(cache->pending[level].next,
2640                                           struct backref_node, list);
2641                         list_move_tail(&node->list, &list);
2642                         BUG_ON(!node->pending);
2643
2644                         if (!err) {
2645                                 ret = link_to_upper(trans, rc, node, path);
2646                                 if (ret < 0)
2647                                         err = ret;
2648                         }
2649                 }
2650                 list_splice_init(&list, &cache->pending[level]);
2651         }
2652         return err;
2653 }
2654
2655 static void mark_block_processed(struct reloc_control *rc,
2656                                  u64 bytenr, u32 blocksize)
2657 {
2658         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2659                         EXTENT_DIRTY, GFP_NOFS);
2660 }
2661
2662 static void __mark_block_processed(struct reloc_control *rc,
2663                                    struct backref_node *node)
2664 {
2665         u32 blocksize;
2666         if (node->level == 0 ||
2667             in_block_group(node->bytenr, rc->block_group)) {
2668                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2669                 mark_block_processed(rc, node->bytenr, blocksize);
2670         }
2671         node->processed = 1;
2672 }
2673
2674 /*
2675  * mark a block and all blocks directly/indirectly reference the block
2676  * as processed.
2677  */
2678 static void update_processed_blocks(struct reloc_control *rc,
2679                                     struct backref_node *node)
2680 {
2681         struct backref_node *next = node;
2682         struct backref_edge *edge;
2683         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2684         int index = 0;
2685
2686         while (next) {
2687                 cond_resched();
2688                 while (1) {
2689                         if (next->processed)
2690                                 break;
2691
2692                         __mark_block_processed(rc, next);
2693
2694                         if (list_empty(&next->upper))
2695                                 break;
2696
2697                         edge = list_entry(next->upper.next,
2698                                           struct backref_edge, list[LOWER]);
2699                         edges[index++] = edge;
2700                         next = edge->node[UPPER];
2701                 }
2702                 next = walk_down_backref(edges, &index);
2703         }
2704 }
2705
2706 static int tree_block_processed(u64 bytenr, u32 blocksize,
2707                                 struct reloc_control *rc)
2708 {
2709         if (test_range_bit(&rc->processed_blocks, bytenr,
2710                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2711                 return 1;
2712         return 0;
2713 }
2714
2715 static int get_tree_block_key(struct reloc_control *rc,
2716                               struct tree_block *block)
2717 {
2718         struct extent_buffer *eb;
2719
2720         BUG_ON(block->key_ready);
2721         eb = read_tree_block(rc->extent_root, block->bytenr,
2722                              block->key.objectid, block->key.offset);
2723         BUG_ON(!eb);
2724         WARN_ON(btrfs_header_level(eb) != block->level);
2725         if (block->level == 0)
2726                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2727         else
2728                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2729         free_extent_buffer(eb);
2730         block->key_ready = 1;
2731         return 0;
2732 }
2733
2734 static int reada_tree_block(struct reloc_control *rc,
2735                             struct tree_block *block)
2736 {
2737         BUG_ON(block->key_ready);
2738         readahead_tree_block(rc->extent_root, block->bytenr,
2739                              block->key.objectid, block->key.offset);
2740         return 0;
2741 }
2742
2743 /*
2744  * helper function to relocate a tree block
2745  */
2746 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2747                                 struct reloc_control *rc,
2748                                 struct backref_node *node,
2749                                 struct btrfs_key *key,
2750                                 struct btrfs_path *path)
2751 {
2752         struct btrfs_root *root;
2753         int release = 0;
2754         int ret = 0;
2755
2756         if (!node)
2757                 return 0;
2758
2759         BUG_ON(node->processed);
2760         root = select_one_root(trans, node);
2761         if (root == ERR_PTR(-ENOENT)) {
2762                 update_processed_blocks(rc, node);
2763                 goto out;
2764         }
2765
2766         if (!root || root->ref_cows) {
2767                 ret = reserve_metadata_space(trans, rc, node);
2768                 if (ret)
2769                         goto out;
2770                 release = 1;
2771         }
2772
2773         if (root) {
2774                 if (root->ref_cows) {
2775                         BUG_ON(node->new_bytenr);
2776                         BUG_ON(!list_empty(&node->list));
2777                         btrfs_record_root_in_trans(trans, root);
2778                         root = root->reloc_root;
2779                         node->new_bytenr = root->node->start;
2780                         node->root = root;
2781                         list_add_tail(&node->list, &rc->backref_cache.changed);
2782                 } else {
2783                         path->lowest_level = node->level;
2784                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2785                         btrfs_release_path(path);
2786                         if (ret > 0)
2787                                 ret = 0;
2788                 }
2789                 if (!ret)
2790                         update_processed_blocks(rc, node);
2791         } else {
2792                 ret = do_relocation(trans, rc, node, key, path, 1);
2793         }
2794 out:
2795         if (ret || node->level == 0 || node->cowonly) {
2796                 if (release)
2797                         release_metadata_space(rc, node);
2798                 remove_backref_node(&rc->backref_cache, node);
2799         }
2800         return ret;
2801 }
2802
2803 /*
2804  * relocate a list of blocks
2805  */
2806 static noinline_for_stack
2807 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2808                          struct reloc_control *rc, struct rb_root *blocks)
2809 {
2810         struct backref_node *node;
2811         struct btrfs_path *path;
2812         struct tree_block *block;
2813         struct rb_node *rb_node;
2814         int ret;
2815         int err = 0;
2816
2817         path = btrfs_alloc_path();
2818         if (!path)
2819                 return -ENOMEM;
2820
2821         rb_node = rb_first(blocks);
2822         while (rb_node) {
2823                 block = rb_entry(rb_node, struct tree_block, rb_node);
2824                 if (!block->key_ready)
2825                         reada_tree_block(rc, block);
2826                 rb_node = rb_next(rb_node);
2827         }
2828
2829         rb_node = rb_first(blocks);
2830         while (rb_node) {
2831                 block = rb_entry(rb_node, struct tree_block, rb_node);
2832                 if (!block->key_ready)
2833                         get_tree_block_key(rc, block);
2834                 rb_node = rb_next(rb_node);
2835         }
2836
2837         rb_node = rb_first(blocks);
2838         while (rb_node) {
2839                 block = rb_entry(rb_node, struct tree_block, rb_node);
2840
2841                 node = build_backref_tree(rc, &block->key,
2842                                           block->level, block->bytenr);
2843                 if (IS_ERR(node)) {
2844                         err = PTR_ERR(node);
2845                         goto out;
2846                 }
2847
2848                 ret = relocate_tree_block(trans, rc, node, &block->key,
2849                                           path);
2850                 if (ret < 0) {
2851                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2852                                 err = ret;
2853                         goto out;
2854                 }
2855                 rb_node = rb_next(rb_node);
2856         }
2857 out:
2858         free_block_list(blocks);
2859         err = finish_pending_nodes(trans, rc, path, err);
2860
2861         btrfs_free_path(path);
2862         return err;
2863 }
2864
2865 static noinline_for_stack
2866 int prealloc_file_extent_cluster(struct inode *inode,
2867                                  struct file_extent_cluster *cluster)
2868 {
2869         u64 alloc_hint = 0;
2870         u64 start;
2871         u64 end;
2872         u64 offset = BTRFS_I(inode)->index_cnt;
2873         u64 num_bytes;
2874         int nr = 0;
2875         int ret = 0;
2876
2877         BUG_ON(cluster->start != cluster->boundary[0]);
2878         mutex_lock(&inode->i_mutex);
2879
2880         ret = btrfs_check_data_free_space(inode, cluster->end +
2881                                           1 - cluster->start);
2882         if (ret)
2883                 goto out;
2884
2885         while (nr < cluster->nr) {
2886                 start = cluster->boundary[nr] - offset;
2887                 if (nr + 1 < cluster->nr)
2888                         end = cluster->boundary[nr + 1] - 1 - offset;
2889                 else
2890                         end = cluster->end - offset;
2891
2892                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2893                 num_bytes = end + 1 - start;
2894                 ret = btrfs_prealloc_file_range(inode, 0, start,
2895                                                 num_bytes, num_bytes,
2896                                                 end + 1, &alloc_hint);
2897                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2898                 if (ret)
2899                         break;
2900                 nr++;
2901         }
2902         btrfs_free_reserved_data_space(inode, cluster->end +
2903                                        1 - cluster->start);
2904 out:
2905         mutex_unlock(&inode->i_mutex);
2906         return ret;
2907 }
2908
2909 static noinline_for_stack
2910 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2911                          u64 block_start)
2912 {
2913         struct btrfs_root *root = BTRFS_I(inode)->root;
2914         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2915         struct extent_map *em;
2916         int ret = 0;
2917
2918         em = alloc_extent_map();
2919         if (!em)
2920                 return -ENOMEM;
2921
2922         em->start = start;
2923         em->len = end + 1 - start;
2924         em->block_len = em->len;
2925         em->block_start = block_start;
2926         em->bdev = root->fs_info->fs_devices->latest_bdev;
2927         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2928
2929         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2930         while (1) {
2931                 write_lock(&em_tree->lock);
2932                 ret = add_extent_mapping(em_tree, em);
2933                 write_unlock(&em_tree->lock);
2934                 if (ret != -EEXIST) {
2935                         free_extent_map(em);
2936                         break;
2937                 }
2938                 btrfs_drop_extent_cache(inode, start, end, 0);
2939         }
2940         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2941         return ret;
2942 }
2943
2944 static int relocate_file_extent_cluster(struct inode *inode,
2945                                         struct file_extent_cluster *cluster)
2946 {
2947         u64 page_start;
2948         u64 page_end;
2949         u64 offset = BTRFS_I(inode)->index_cnt;
2950         unsigned long index;
2951         unsigned long last_index;
2952         struct page *page;
2953         struct file_ra_state *ra;
2954         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2955         int nr = 0;
2956         int ret = 0;
2957
2958         if (!cluster->nr)
2959                 return 0;
2960
2961         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2962         if (!ra)
2963                 return -ENOMEM;
2964
2965         ret = prealloc_file_extent_cluster(inode, cluster);
2966         if (ret)
2967                 goto out;
2968
2969         file_ra_state_init(ra, inode->i_mapping);
2970
2971         ret = setup_extent_mapping(inode, cluster->start - offset,
2972                                    cluster->end - offset, cluster->start);
2973         if (ret)
2974                 goto out;
2975
2976         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
2977         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
2978         while (index <= last_index) {
2979                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
2980                 if (ret)
2981                         goto out;
2982
2983                 page = find_lock_page(inode->i_mapping, index);
2984                 if (!page) {
2985                         page_cache_sync_readahead(inode->i_mapping,
2986                                                   ra, NULL, index,
2987                                                   last_index + 1 - index);
2988                         page = find_or_create_page(inode->i_mapping, index,
2989                                                    mask);
2990                         if (!page) {
2991                                 btrfs_delalloc_release_metadata(inode,
2992                                                         PAGE_CACHE_SIZE);
2993                                 ret = -ENOMEM;
2994                                 goto out;
2995                         }
2996                 }
2997
2998                 if (PageReadahead(page)) {
2999                         page_cache_async_readahead(inode->i_mapping,
3000                                                    ra, NULL, page, index,
3001                                                    last_index + 1 - index);
3002                 }
3003
3004                 if (!PageUptodate(page)) {
3005                         btrfs_readpage(NULL, page);
3006                         lock_page(page);
3007                         if (!PageUptodate(page)) {
3008                                 unlock_page(page);
3009                                 page_cache_release(page);
3010                                 btrfs_delalloc_release_metadata(inode,
3011                                                         PAGE_CACHE_SIZE);
3012                                 ret = -EIO;
3013                                 goto out;
3014                         }
3015                 }
3016
3017                 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
3018                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3019
3020                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3021
3022                 set_page_extent_mapped(page);
3023
3024                 if (nr < cluster->nr &&
3025                     page_start + offset == cluster->boundary[nr]) {
3026                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3027                                         page_start, page_end,
3028                                         EXTENT_BOUNDARY, GFP_NOFS);
3029                         nr++;
3030                 }
3031
3032                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3033                 set_page_dirty(page);
3034
3035                 unlock_extent(&BTRFS_I(inode)->io_tree,
3036                               page_start, page_end);
3037                 unlock_page(page);
3038                 page_cache_release(page);
3039
3040                 index++;
3041                 balance_dirty_pages_ratelimited(inode->i_mapping);
3042                 btrfs_throttle(BTRFS_I(inode)->root);
3043         }
3044         WARN_ON(nr != cluster->nr);
3045 out:
3046         kfree(ra);
3047         return ret;
3048 }
3049
3050 static noinline_for_stack
3051 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3052                          struct file_extent_cluster *cluster)
3053 {
3054         int ret;
3055
3056         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3057                 ret = relocate_file_extent_cluster(inode, cluster);
3058                 if (ret)
3059                         return ret;
3060                 cluster->nr = 0;
3061         }
3062
3063         if (!cluster->nr)
3064                 cluster->start = extent_key->objectid;
3065         else
3066                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3067         cluster->end = extent_key->objectid + extent_key->offset - 1;
3068         cluster->boundary[cluster->nr] = extent_key->objectid;
3069         cluster->nr++;
3070
3071         if (cluster->nr >= MAX_EXTENTS) {
3072                 ret = relocate_file_extent_cluster(inode, cluster);
3073                 if (ret)
3074                         return ret;
3075                 cluster->nr = 0;
3076         }
3077         return 0;
3078 }
3079
3080 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3081 static int get_ref_objectid_v0(struct reloc_control *rc,
3082                                struct btrfs_path *path,
3083                                struct btrfs_key *extent_key,
3084                                u64 *ref_objectid, int *path_change)
3085 {
3086         struct btrfs_key key;
3087         struct extent_buffer *leaf;
3088         struct btrfs_extent_ref_v0 *ref0;
3089         int ret;
3090         int slot;
3091
3092         leaf = path->nodes[0];
3093         slot = path->slots[0];
3094         while (1) {
3095                 if (slot >= btrfs_header_nritems(leaf)) {
3096                         ret = btrfs_next_leaf(rc->extent_root, path);
3097                         if (ret < 0)
3098                                 return ret;
3099                         BUG_ON(ret > 0);
3100                         leaf = path->nodes[0];
3101                         slot = path->slots[0];
3102                         if (path_change)
3103                                 *path_change = 1;
3104                 }
3105                 btrfs_item_key_to_cpu(leaf, &key, slot);
3106                 if (key.objectid != extent_key->objectid)
3107                         return -ENOENT;
3108
3109                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3110                         slot++;
3111                         continue;
3112                 }
3113                 ref0 = btrfs_item_ptr(leaf, slot,
3114                                 struct btrfs_extent_ref_v0);
3115                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3116                 break;
3117         }
3118         return 0;
3119 }
3120 #endif
3121
3122 /*
3123  * helper to add a tree block to the list.
3124  * the major work is getting the generation and level of the block
3125  */
3126 static int add_tree_block(struct reloc_control *rc,
3127                           struct btrfs_key *extent_key,
3128                           struct btrfs_path *path,
3129                           struct rb_root *blocks)
3130 {
3131         struct extent_buffer *eb;
3132         struct btrfs_extent_item *ei;
3133         struct btrfs_tree_block_info *bi;
3134         struct tree_block *block;
3135         struct rb_node *rb_node;
3136         u32 item_size;
3137         int level = -1;
3138         int generation;
3139
3140         eb =  path->nodes[0];
3141         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3142
3143         if (item_size >= sizeof(*ei) + sizeof(*bi)) {
3144                 ei = btrfs_item_ptr(eb, path->slots[0],
3145                                 struct btrfs_extent_item);
3146                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3147                 generation = btrfs_extent_generation(eb, ei);
3148                 level = btrfs_tree_block_level(eb, bi);
3149         } else {
3150 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3151                 u64 ref_owner;
3152                 int ret;
3153
3154                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3155                 ret = get_ref_objectid_v0(rc, path, extent_key,
3156                                           &ref_owner, NULL);
3157                 if (ret < 0)
3158                         return ret;
3159                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3160                 level = (int)ref_owner;
3161                 /* FIXME: get real generation */
3162                 generation = 0;
3163 #else
3164                 BUG();
3165 #endif
3166         }
3167
3168         btrfs_release_path(path);
3169
3170         BUG_ON(level == -1);
3171
3172         block = kmalloc(sizeof(*block), GFP_NOFS);
3173         if (!block)
3174                 return -ENOMEM;
3175
3176         block->bytenr = extent_key->objectid;
3177         block->key.objectid = extent_key->offset;
3178         block->key.offset = generation;
3179         block->level = level;
3180         block->key_ready = 0;
3181
3182         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3183         if (rb_node)
3184                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3185
3186         return 0;
3187 }
3188
3189 /*
3190  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3191  */
3192 static int __add_tree_block(struct reloc_control *rc,
3193                             u64 bytenr, u32 blocksize,
3194                             struct rb_root *blocks)
3195 {
3196         struct btrfs_path *path;
3197         struct btrfs_key key;
3198         int ret;
3199
3200         if (tree_block_processed(bytenr, blocksize, rc))
3201                 return 0;
3202
3203         if (tree_search(blocks, bytenr))
3204                 return 0;
3205
3206         path = btrfs_alloc_path();
3207         if (!path)
3208                 return -ENOMEM;
3209
3210         key.objectid = bytenr;
3211         key.type = BTRFS_EXTENT_ITEM_KEY;
3212         key.offset = blocksize;
3213
3214         path->search_commit_root = 1;
3215         path->skip_locking = 1;
3216         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3217         if (ret < 0)
3218                 goto out;
3219         BUG_ON(ret);
3220
3221         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3222         ret = add_tree_block(rc, &key, path, blocks);
3223 out:
3224         btrfs_free_path(path);
3225         return ret;
3226 }
3227
3228 /*
3229  * helper to check if the block use full backrefs for pointers in it
3230  */
3231 static int block_use_full_backref(struct reloc_control *rc,
3232                                   struct extent_buffer *eb)
3233 {
3234         u64 flags;
3235         int ret;
3236
3237         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3238             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3239                 return 1;
3240
3241         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3242                                        eb->start, eb->len, NULL, &flags);
3243         BUG_ON(ret);
3244
3245         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3246                 ret = 1;
3247         else
3248                 ret = 0;
3249         return ret;
3250 }
3251
3252 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3253                                     struct inode *inode, u64 ino)
3254 {
3255         struct btrfs_key key;
3256         struct btrfs_path *path;
3257         struct btrfs_root *root = fs_info->tree_root;
3258         struct btrfs_trans_handle *trans;
3259         unsigned long nr;
3260         int ret = 0;
3261
3262         if (inode)
3263                 goto truncate;
3264
3265         key.objectid = ino;
3266         key.type = BTRFS_INODE_ITEM_KEY;
3267         key.offset = 0;
3268
3269         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3270         if (IS_ERR_OR_NULL(inode) || is_bad_inode(inode)) {
3271                 if (inode && !IS_ERR(inode))
3272                         iput(inode);
3273                 return -ENOENT;
3274         }
3275
3276 truncate:
3277         path = btrfs_alloc_path();
3278         if (!path) {
3279                 ret = -ENOMEM;
3280                 goto out;
3281         }
3282
3283         trans = btrfs_join_transaction(root);
3284         if (IS_ERR(trans)) {
3285                 btrfs_free_path(path);
3286                 ret = PTR_ERR(trans);
3287                 goto out;
3288         }
3289
3290         ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3291
3292         btrfs_free_path(path);
3293         nr = trans->blocks_used;
3294         btrfs_end_transaction(trans, root);
3295         btrfs_btree_balance_dirty(root, nr);
3296 out:
3297         iput(inode);
3298         return ret;
3299 }
3300
3301 /*
3302  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3303  * this function scans fs tree to find blocks reference the data extent
3304  */
3305 static int find_data_references(struct reloc_control *rc,
3306                                 struct btrfs_key *extent_key,
3307                                 struct extent_buffer *leaf,
3308                                 struct btrfs_extent_data_ref *ref,
3309                                 struct rb_root *blocks)
3310 {
3311         struct btrfs_path *path;
3312         struct tree_block *block;
3313         struct btrfs_root *root;
3314         struct btrfs_file_extent_item *fi;
3315         struct rb_node *rb_node;
3316         struct btrfs_key key;
3317         u64 ref_root;
3318         u64 ref_objectid;
3319         u64 ref_offset;
3320         u32 ref_count;
3321         u32 nritems;
3322         int err = 0;
3323         int added = 0;
3324         int counted;
3325         int ret;
3326
3327         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3328         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3329         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3330         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3331
3332         /*
3333          * This is an extent belonging to the free space cache, lets just delete
3334          * it and redo the search.
3335          */
3336         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3337                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3338                                                NULL, ref_objectid);
3339                 if (ret != -ENOENT)
3340                         return ret;
3341                 ret = 0;
3342         }
3343
3344         path = btrfs_alloc_path();
3345         if (!path)
3346                 return -ENOMEM;
3347         path->reada = 1;
3348
3349         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3350         if (IS_ERR(root)) {
3351                 err = PTR_ERR(root);
3352                 goto out;
3353         }
3354
3355         key.objectid = ref_objectid;
3356         key.type = BTRFS_EXTENT_DATA_KEY;
3357         if (ref_offset > ((u64)-1 << 32))
3358                 key.offset = 0;
3359         else
3360                 key.offset = ref_offset;
3361
3362         path->search_commit_root = 1;
3363         path->skip_locking = 1;
3364         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3365         if (ret < 0) {
3366                 err = ret;
3367                 goto out;
3368         }
3369
3370         leaf = path->nodes[0];
3371         nritems = btrfs_header_nritems(leaf);
3372         /*
3373          * the references in tree blocks that use full backrefs
3374          * are not counted in
3375          */
3376         if (block_use_full_backref(rc, leaf))
3377                 counted = 0;
3378         else
3379                 counted = 1;
3380         rb_node = tree_search(blocks, leaf->start);
3381         if (rb_node) {
3382                 if (counted)
3383                         added = 1;
3384                 else
3385                         path->slots[0] = nritems;
3386         }
3387
3388         while (ref_count > 0) {
3389                 while (path->slots[0] >= nritems) {
3390                         ret = btrfs_next_leaf(root, path);
3391                         if (ret < 0) {
3392                                 err = ret;
3393                                 goto out;
3394                         }
3395                         if (ret > 0) {
3396                                 WARN_ON(1);
3397                                 goto out;
3398                         }
3399
3400                         leaf = path->nodes[0];
3401                         nritems = btrfs_header_nritems(leaf);
3402                         added = 0;
3403
3404                         if (block_use_full_backref(rc, leaf))
3405                                 counted = 0;
3406                         else
3407                                 counted = 1;
3408                         rb_node = tree_search(blocks, leaf->start);
3409                         if (rb_node) {
3410                                 if (counted)
3411                                         added = 1;
3412                                 else
3413                                         path->slots[0] = nritems;
3414                         }
3415                 }
3416
3417                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3418                 if (key.objectid != ref_objectid ||
3419                     key.type != BTRFS_EXTENT_DATA_KEY) {
3420                         WARN_ON(1);
3421                         break;
3422                 }
3423
3424                 fi = btrfs_item_ptr(leaf, path->slots[0],
3425                                     struct btrfs_file_extent_item);
3426
3427                 if (btrfs_file_extent_type(leaf, fi) ==
3428                     BTRFS_FILE_EXTENT_INLINE)
3429                         goto next;
3430
3431                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3432                     extent_key->objectid)
3433                         goto next;
3434
3435                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3436                 if (key.offset != ref_offset)
3437                         goto next;
3438
3439                 if (counted)
3440                         ref_count--;
3441                 if (added)
3442                         goto next;
3443
3444                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3445                         block = kmalloc(sizeof(*block), GFP_NOFS);
3446                         if (!block) {
3447                                 err = -ENOMEM;
3448                                 break;
3449                         }
3450                         block->bytenr = leaf->start;
3451                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3452                         block->level = 0;
3453                         block->key_ready = 1;
3454                         rb_node = tree_insert(blocks, block->bytenr,
3455                                               &block->rb_node);
3456                         if (rb_node)
3457                                 backref_tree_panic(rb_node, -EEXIST,
3458                                                    block->bytenr);
3459                 }
3460                 if (counted)
3461                         added = 1;
3462                 else
3463                         path->slots[0] = nritems;
3464 next:
3465                 path->slots[0]++;
3466
3467         }
3468 out:
3469         btrfs_free_path(path);
3470         return err;
3471 }
3472
3473 /*
3474  * hepler to find all tree blocks that reference a given data extent
3475  */
3476 static noinline_for_stack
3477 int add_data_references(struct reloc_control *rc,
3478                         struct btrfs_key *extent_key,
3479                         struct btrfs_path *path,
3480                         struct rb_root *blocks)
3481 {
3482         struct btrfs_key key;
3483         struct extent_buffer *eb;
3484         struct btrfs_extent_data_ref *dref;
3485         struct btrfs_extent_inline_ref *iref;
3486         unsigned long ptr;
3487         unsigned long end;
3488         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3489         int ret;
3490         int err = 0;
3491
3492         eb = path->nodes[0];
3493         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3494         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3495 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3496         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3497                 ptr = end;
3498         else
3499 #endif
3500                 ptr += sizeof(struct btrfs_extent_item);
3501
3502         while (ptr < end) {
3503                 iref = (struct btrfs_extent_inline_ref *)ptr;
3504                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3505                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3506                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3507                         ret = __add_tree_block(rc, key.offset, blocksize,
3508                                                blocks);
3509                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3510                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3511                         ret = find_data_references(rc, extent_key,
3512                                                    eb, dref, blocks);
3513                 } else {
3514                         BUG();
3515                 }
3516                 ptr += btrfs_extent_inline_ref_size(key.type);
3517         }
3518         WARN_ON(ptr > end);
3519
3520         while (1) {
3521                 cond_resched();
3522                 eb = path->nodes[0];
3523                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3524                         ret = btrfs_next_leaf(rc->extent_root, path);
3525                         if (ret < 0) {
3526                                 err = ret;
3527                                 break;
3528                         }
3529                         if (ret > 0)
3530                                 break;
3531                         eb = path->nodes[0];
3532                 }
3533
3534                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3535                 if (key.objectid != extent_key->objectid)
3536                         break;
3537
3538 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3539                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3540                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3541 #else
3542                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3543                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3544 #endif
3545                         ret = __add_tree_block(rc, key.offset, blocksize,
3546                                                blocks);
3547                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3548                         dref = btrfs_item_ptr(eb, path->slots[0],
3549                                               struct btrfs_extent_data_ref);
3550                         ret = find_data_references(rc, extent_key,
3551                                                    eb, dref, blocks);
3552                 } else {
3553                         ret = 0;
3554                 }
3555                 if (ret) {
3556                         err = ret;
3557                         break;
3558                 }
3559                 path->slots[0]++;
3560         }
3561         btrfs_release_path(path);
3562         if (err)
3563                 free_block_list(blocks);
3564         return err;
3565 }
3566
3567 /*
3568  * hepler to find next unprocessed extent
3569  */
3570 static noinline_for_stack
3571 int find_next_extent(struct btrfs_trans_handle *trans,
3572                      struct reloc_control *rc, struct btrfs_path *path,
3573                      struct btrfs_key *extent_key)
3574 {
3575         struct btrfs_key key;
3576         struct extent_buffer *leaf;
3577         u64 start, end, last;
3578         int ret;
3579
3580         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3581         while (1) {
3582                 cond_resched();
3583                 if (rc->search_start >= last) {
3584                         ret = 1;
3585                         break;
3586                 }
3587
3588                 key.objectid = rc->search_start;
3589                 key.type = BTRFS_EXTENT_ITEM_KEY;
3590                 key.offset = 0;
3591
3592                 path->search_commit_root = 1;
3593                 path->skip_locking = 1;
3594                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3595                                         0, 0);
3596                 if (ret < 0)
3597                         break;
3598 next:
3599                 leaf = path->nodes[0];
3600                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3601                         ret = btrfs_next_leaf(rc->extent_root, path);
3602                         if (ret != 0)
3603                                 break;
3604                         leaf = path->nodes[0];
3605                 }
3606
3607                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3608                 if (key.objectid >= last) {
3609                         ret = 1;
3610                         break;
3611                 }
3612
3613                 if (key.type != BTRFS_EXTENT_ITEM_KEY ||
3614                     key.objectid + key.offset <= rc->search_start) {
3615                         path->slots[0]++;
3616                         goto next;
3617                 }
3618
3619                 ret = find_first_extent_bit(&rc->processed_blocks,
3620                                             key.objectid, &start, &end,
3621                                             EXTENT_DIRTY);
3622
3623                 if (ret == 0 && start <= key.objectid) {
3624                         btrfs_release_path(path);
3625                         rc->search_start = end + 1;
3626                 } else {
3627                         rc->search_start = key.objectid + key.offset;
3628                         memcpy(extent_key, &key, sizeof(key));
3629                         return 0;
3630                 }
3631         }
3632         btrfs_release_path(path);
3633         return ret;
3634 }
3635
3636 static void set_reloc_control(struct reloc_control *rc)
3637 {
3638         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3639
3640         mutex_lock(&fs_info->reloc_mutex);
3641         fs_info->reloc_ctl = rc;
3642         mutex_unlock(&fs_info->reloc_mutex);
3643 }
3644
3645 static void unset_reloc_control(struct reloc_control *rc)
3646 {
3647         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3648
3649         mutex_lock(&fs_info->reloc_mutex);
3650         fs_info->reloc_ctl = NULL;
3651         mutex_unlock(&fs_info->reloc_mutex);
3652 }
3653
3654 static int check_extent_flags(u64 flags)
3655 {
3656         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3657             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3658                 return 1;
3659         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3660             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3661                 return 1;
3662         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3663             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3664                 return 1;
3665         return 0;
3666 }
3667
3668 static noinline_for_stack
3669 int prepare_to_relocate(struct reloc_control *rc)
3670 {
3671         struct btrfs_trans_handle *trans;
3672         int ret;
3673
3674         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root);
3675         if (!rc->block_rsv)
3676                 return -ENOMEM;
3677
3678         /*
3679          * reserve some space for creating reloc trees.
3680          * btrfs_init_reloc_root will use them when there
3681          * is no reservation in transaction handle.
3682          */
3683         ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3684                                   rc->extent_root->nodesize * 256);
3685         if (ret)
3686                 return ret;
3687
3688         memset(&rc->cluster, 0, sizeof(rc->cluster));
3689         rc->search_start = rc->block_group->key.objectid;
3690         rc->extents_found = 0;
3691         rc->nodes_relocated = 0;
3692         rc->merging_rsv_size = 0;
3693
3694         rc->create_reloc_tree = 1;
3695         set_reloc_control(rc);
3696
3697         trans = btrfs_join_transaction(rc->extent_root);
3698         BUG_ON(IS_ERR(trans));
3699         btrfs_commit_transaction(trans, rc->extent_root);
3700         return 0;
3701 }
3702
3703 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3704 {
3705         struct rb_root blocks = RB_ROOT;
3706         struct btrfs_key key;
3707         struct btrfs_trans_handle *trans = NULL;
3708         struct btrfs_path *path;
3709         struct btrfs_extent_item *ei;
3710         unsigned long nr;
3711         u64 flags;
3712         u32 item_size;
3713         int ret;
3714         int err = 0;
3715         int progress = 0;
3716
3717         path = btrfs_alloc_path();
3718         if (!path)
3719                 return -ENOMEM;
3720         path->reada = 1;
3721
3722         ret = prepare_to_relocate(rc);
3723         if (ret) {
3724                 err = ret;
3725                 goto out_free;
3726         }
3727
3728         while (1) {
3729                 progress++;
3730                 trans = btrfs_start_transaction(rc->extent_root, 0);
3731                 BUG_ON(IS_ERR(trans));
3732 restart:
3733                 if (update_backref_cache(trans, &rc->backref_cache)) {
3734                         btrfs_end_transaction(trans, rc->extent_root);
3735                         continue;
3736                 }
3737
3738                 ret = find_next_extent(trans, rc, path, &key);
3739                 if (ret < 0)
3740                         err = ret;
3741                 if (ret != 0)
3742                         break;
3743
3744                 rc->extents_found++;
3745
3746                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3747                                     struct btrfs_extent_item);
3748                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3749                 if (item_size >= sizeof(*ei)) {
3750                         flags = btrfs_extent_flags(path->nodes[0], ei);
3751                         ret = check_extent_flags(flags);
3752                         BUG_ON(ret);
3753
3754                 } else {
3755 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3756                         u64 ref_owner;
3757                         int path_change = 0;
3758
3759                         BUG_ON(item_size !=
3760                                sizeof(struct btrfs_extent_item_v0));
3761                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3762                                                   &path_change);
3763                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3764                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3765                         else
3766                                 flags = BTRFS_EXTENT_FLAG_DATA;
3767
3768                         if (path_change) {
3769                                 btrfs_release_path(path);
3770
3771                                 path->search_commit_root = 1;
3772                                 path->skip_locking = 1;
3773                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3774                                                         &key, path, 0, 0);
3775                                 if (ret < 0) {
3776                                         err = ret;
3777                                         break;
3778                                 }
3779                                 BUG_ON(ret > 0);
3780                         }
3781 #else
3782                         BUG();
3783 #endif
3784                 }
3785
3786                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3787                         ret = add_tree_block(rc, &key, path, &blocks);
3788                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3789                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3790                         ret = add_data_references(rc, &key, path, &blocks);
3791                 } else {
3792                         btrfs_release_path(path);
3793                         ret = 0;
3794                 }
3795                 if (ret < 0) {
3796                         err = ret;
3797                         break;
3798                 }
3799
3800                 if (!RB_EMPTY_ROOT(&blocks)) {
3801                         ret = relocate_tree_blocks(trans, rc, &blocks);
3802                         if (ret < 0) {
3803                                 if (ret != -EAGAIN) {
3804                                         err = ret;
3805                                         break;
3806                                 }
3807                                 rc->extents_found--;
3808                                 rc->search_start = key.objectid;
3809                         }
3810                 }
3811
3812                 ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5);
3813                 if (ret < 0) {
3814                         if (ret != -EAGAIN) {
3815                                 err = ret;
3816                                 WARN_ON(1);
3817                                 break;
3818                         }
3819                         rc->commit_transaction = 1;
3820                 }
3821
3822                 if (rc->commit_transaction) {
3823                         rc->commit_transaction = 0;
3824                         ret = btrfs_commit_transaction(trans, rc->extent_root);
3825                         BUG_ON(ret);
3826                 } else {
3827                         nr = trans->blocks_used;
3828                         btrfs_end_transaction_throttle(trans, rc->extent_root);
3829                         btrfs_btree_balance_dirty(rc->extent_root, nr);
3830                 }
3831                 trans = NULL;
3832
3833                 if (rc->stage == MOVE_DATA_EXTENTS &&
3834                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3835                         rc->found_file_extent = 1;
3836                         ret = relocate_data_extent(rc->data_inode,
3837                                                    &key, &rc->cluster);
3838                         if (ret < 0) {
3839                                 err = ret;
3840                                 break;
3841                         }
3842                 }
3843         }
3844         if (trans && progress && err == -ENOSPC) {
3845                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
3846                                               rc->block_group->flags);
3847                 if (ret == 0) {
3848                         err = 0;
3849                         progress = 0;
3850                         goto restart;
3851                 }
3852         }
3853
3854         btrfs_release_path(path);
3855         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3856                           GFP_NOFS);
3857
3858         if (trans) {
3859                 nr = trans->blocks_used;
3860                 btrfs_end_transaction_throttle(trans, rc->extent_root);
3861                 btrfs_btree_balance_dirty(rc->extent_root, nr);
3862         }
3863
3864         if (!err) {
3865                 ret = relocate_file_extent_cluster(rc->data_inode,
3866                                                    &rc->cluster);
3867                 if (ret < 0)
3868                         err = ret;
3869         }
3870
3871         rc->create_reloc_tree = 0;
3872         set_reloc_control(rc);
3873
3874         backref_cache_cleanup(&rc->backref_cache);
3875         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3876
3877         err = prepare_to_merge(rc, err);
3878
3879         merge_reloc_roots(rc);
3880
3881         rc->merge_reloc_tree = 0;
3882         unset_reloc_control(rc);
3883         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3884
3885         /* get rid of pinned extents */
3886         trans = btrfs_join_transaction(rc->extent_root);
3887         if (IS_ERR(trans))
3888                 err = PTR_ERR(trans);
3889         else
3890                 btrfs_commit_transaction(trans, rc->extent_root);
3891 out_free:
3892         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
3893         btrfs_free_path(path);
3894         return err;
3895 }
3896
3897 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3898                                  struct btrfs_root *root, u64 objectid)
3899 {
3900         struct btrfs_path *path;
3901         struct btrfs_inode_item *item;
3902         struct extent_buffer *leaf;
3903         int ret;
3904
3905         path = btrfs_alloc_path();
3906         if (!path)
3907                 return -ENOMEM;
3908
3909         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3910         if (ret)
3911                 goto out;
3912
3913         leaf = path->nodes[0];
3914         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3915         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
3916         btrfs_set_inode_generation(leaf, item, 1);
3917         btrfs_set_inode_size(leaf, item, 0);
3918         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3919         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3920                                           BTRFS_INODE_PREALLOC);
3921         btrfs_mark_buffer_dirty(leaf);
3922         btrfs_release_path(path);
3923 out:
3924         btrfs_free_path(path);
3925         return ret;
3926 }
3927
3928 /*
3929  * helper to create inode for data relocation.
3930  * the inode is in data relocation tree and its link count is 0
3931  */
3932 static noinline_for_stack
3933 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3934                                  struct btrfs_block_group_cache *group)
3935 {
3936         struct inode *inode = NULL;
3937         struct btrfs_trans_handle *trans;
3938         struct btrfs_root *root;
3939         struct btrfs_key key;
3940         unsigned long nr;
3941         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
3942         int err = 0;
3943
3944         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
3945         if (IS_ERR(root))
3946                 return ERR_CAST(root);
3947
3948         trans = btrfs_start_transaction(root, 6);
3949         if (IS_ERR(trans))
3950                 return ERR_CAST(trans);
3951
3952         err = btrfs_find_free_objectid(root, &objectid);
3953         if (err)
3954                 goto out;
3955
3956         err = __insert_orphan_inode(trans, root, objectid);
3957         BUG_ON(err);
3958
3959         key.objectid = objectid;
3960         key.type = BTRFS_INODE_ITEM_KEY;
3961         key.offset = 0;
3962         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
3963         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
3964         BTRFS_I(inode)->index_cnt = group->key.objectid;
3965
3966         err = btrfs_orphan_add(trans, inode);
3967 out:
3968         nr = trans->blocks_used;
3969         btrfs_end_transaction(trans, root);
3970         btrfs_btree_balance_dirty(root, nr);
3971         if (err) {
3972                 if (inode)
3973                         iput(inode);
3974                 inode = ERR_PTR(err);
3975         }
3976         return inode;
3977 }
3978
3979 static struct reloc_control *alloc_reloc_control(void)
3980 {
3981         struct reloc_control *rc;
3982
3983         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3984         if (!rc)
3985                 return NULL;
3986
3987         INIT_LIST_HEAD(&rc->reloc_roots);
3988         backref_cache_init(&rc->backref_cache);
3989         mapping_tree_init(&rc->reloc_root_tree);
3990         extent_io_tree_init(&rc->processed_blocks, NULL);
3991         return rc;
3992 }
3993
3994 /*
3995  * function to relocate all extents in a block group.
3996  */
3997 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3998 {
3999         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4000         struct reloc_control *rc;
4001         struct inode *inode;
4002         struct btrfs_path *path;
4003         int ret;
4004         int rw = 0;
4005         int err = 0;
4006
4007         rc = alloc_reloc_control();
4008         if (!rc)
4009                 return -ENOMEM;
4010
4011         rc->extent_root = extent_root;
4012
4013         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4014         BUG_ON(!rc->block_group);
4015
4016         if (!rc->block_group->ro) {
4017                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4018                 if (ret) {
4019                         err = ret;
4020                         goto out;
4021                 }
4022                 rw = 1;
4023         }
4024
4025         path = btrfs_alloc_path();
4026         if (!path) {
4027                 err = -ENOMEM;
4028                 goto out;
4029         }
4030
4031         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4032                                         path);
4033         btrfs_free_path(path);
4034
4035         if (!IS_ERR(inode))
4036                 ret = delete_block_group_cache(fs_info, inode, 0);
4037         else
4038                 ret = PTR_ERR(inode);
4039
4040         if (ret && ret != -ENOENT) {
4041                 err = ret;
4042                 goto out;
4043         }
4044
4045         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4046         if (IS_ERR(rc->data_inode)) {
4047                 err = PTR_ERR(rc->data_inode);
4048                 rc->data_inode = NULL;
4049                 goto out;
4050         }
4051
4052         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4053                (unsigned long long)rc->block_group->key.objectid,
4054                (unsigned long long)rc->block_group->flags);
4055
4056         btrfs_start_delalloc_inodes(fs_info->tree_root, 0);
4057         btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0);
4058
4059         while (1) {
4060                 mutex_lock(&fs_info->cleaner_mutex);
4061
4062                 btrfs_clean_old_snapshots(fs_info->tree_root);
4063                 ret = relocate_block_group(rc);
4064
4065                 mutex_unlock(&fs_info->cleaner_mutex);
4066                 if (ret < 0) {
4067                         err = ret;
4068                         goto out;
4069                 }
4070
4071                 if (rc->extents_found == 0)
4072                         break;
4073
4074                 printk(KERN_INFO "btrfs: found %llu extents\n",
4075                         (unsigned long long)rc->extents_found);
4076
4077                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4078                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4079                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4080                                                  0, -1);
4081                         rc->stage = UPDATE_DATA_PTRS;
4082                 }
4083         }
4084
4085         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4086                                      rc->block_group->key.objectid,
4087                                      rc->block_group->key.objectid +
4088                                      rc->block_group->key.offset - 1);
4089
4090         WARN_ON(rc->block_group->pinned > 0);
4091         WARN_ON(rc->block_group->reserved > 0);
4092         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4093 out:
4094         if (err && rw)
4095                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4096         iput(rc->data_inode);
4097         btrfs_put_block_group(rc->block_group);
4098         kfree(rc);
4099         return err;
4100 }
4101
4102 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4103 {
4104         struct btrfs_trans_handle *trans;
4105         int ret, err;
4106
4107         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4108         if (IS_ERR(trans))
4109                 return PTR_ERR(trans);
4110
4111         memset(&root->root_item.drop_progress, 0,
4112                 sizeof(root->root_item.drop_progress));
4113         root->root_item.drop_level = 0;
4114         btrfs_set_root_refs(&root->root_item, 0);
4115         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4116                                 &root->root_key, &root->root_item);
4117
4118         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4119         if (err)
4120                 return err;
4121         return ret;
4122 }
4123
4124 /*
4125  * recover relocation interrupted by system crash.
4126  *
4127  * this function resumes merging reloc trees with corresponding fs trees.
4128  * this is important for keeping the sharing of tree blocks
4129  */
4130 int btrfs_recover_relocation(struct btrfs_root *root)
4131 {
4132         LIST_HEAD(reloc_roots);
4133         struct btrfs_key key;
4134         struct btrfs_root *fs_root;
4135         struct btrfs_root *reloc_root;
4136         struct btrfs_path *path;
4137         struct extent_buffer *leaf;
4138         struct reloc_control *rc = NULL;
4139         struct btrfs_trans_handle *trans;
4140         int ret;
4141         int err = 0;
4142
4143         path = btrfs_alloc_path();
4144         if (!path)
4145                 return -ENOMEM;
4146         path->reada = -1;
4147
4148         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4149         key.type = BTRFS_ROOT_ITEM_KEY;
4150         key.offset = (u64)-1;
4151
4152         while (1) {
4153                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4154                                         path, 0, 0);
4155                 if (ret < 0) {
4156                         err = ret;
4157                         goto out;
4158                 }
4159                 if (ret > 0) {
4160                         if (path->slots[0] == 0)
4161                                 break;
4162                         path->slots[0]--;
4163                 }
4164                 leaf = path->nodes[0];
4165                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4166                 btrfs_release_path(path);
4167
4168                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4169                     key.type != BTRFS_ROOT_ITEM_KEY)
4170                         break;
4171
4172                 reloc_root = btrfs_read_fs_root_no_radix(root, &key);
4173                 if (IS_ERR(reloc_root)) {
4174                         err = PTR_ERR(reloc_root);
4175                         goto out;
4176                 }
4177
4178                 list_add(&reloc_root->root_list, &reloc_roots);
4179
4180                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4181                         fs_root = read_fs_root(root->fs_info,
4182                                                reloc_root->root_key.offset);
4183                         if (IS_ERR(fs_root)) {
4184                                 ret = PTR_ERR(fs_root);
4185                                 if (ret != -ENOENT) {
4186                                         err = ret;
4187                                         goto out;
4188                                 }
4189                                 ret = mark_garbage_root(reloc_root);
4190                                 if (ret < 0) {
4191                                         err = ret;
4192                                         goto out;
4193                                 }
4194                         }
4195                 }
4196
4197                 if (key.offset == 0)
4198                         break;
4199
4200                 key.offset--;
4201         }
4202         btrfs_release_path(path);
4203
4204         if (list_empty(&reloc_roots))
4205                 goto out;
4206
4207         rc = alloc_reloc_control();
4208         if (!rc) {
4209                 err = -ENOMEM;
4210                 goto out;
4211         }
4212
4213         rc->extent_root = root->fs_info->extent_root;
4214
4215         set_reloc_control(rc);
4216
4217         trans = btrfs_join_transaction(rc->extent_root);
4218         if (IS_ERR(trans)) {
4219                 unset_reloc_control(rc);
4220                 err = PTR_ERR(trans);
4221                 goto out_free;
4222         }
4223
4224         rc->merge_reloc_tree = 1;
4225
4226         while (!list_empty(&reloc_roots)) {
4227                 reloc_root = list_entry(reloc_roots.next,
4228                                         struct btrfs_root, root_list);
4229                 list_del(&reloc_root->root_list);
4230
4231                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4232                         list_add_tail(&reloc_root->root_list,
4233                                       &rc->reloc_roots);
4234                         continue;
4235                 }
4236
4237                 fs_root = read_fs_root(root->fs_info,
4238                                        reloc_root->root_key.offset);
4239                 if (IS_ERR(fs_root)) {
4240                         err = PTR_ERR(fs_root);
4241                         goto out_free;
4242                 }
4243
4244                 err = __add_reloc_root(reloc_root);
4245                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4246                 fs_root->reloc_root = reloc_root;
4247         }
4248
4249         err = btrfs_commit_transaction(trans, rc->extent_root);
4250         if (err)
4251                 goto out_free;
4252
4253         merge_reloc_roots(rc);
4254
4255         unset_reloc_control(rc);
4256
4257         trans = btrfs_join_transaction(rc->extent_root);
4258         if (IS_ERR(trans))
4259                 err = PTR_ERR(trans);
4260         else
4261                 err = btrfs_commit_transaction(trans, rc->extent_root);
4262 out_free:
4263         kfree(rc);
4264 out:
4265         while (!list_empty(&reloc_roots)) {
4266                 reloc_root = list_entry(reloc_roots.next,
4267                                         struct btrfs_root, root_list);
4268                 list_del(&reloc_root->root_list);
4269                 free_extent_buffer(reloc_root->node);
4270                 free_extent_buffer(reloc_root->commit_root);
4271                 kfree(reloc_root);
4272         }
4273         btrfs_free_path(path);
4274
4275         if (err == 0) {
4276                 /* cleanup orphan inode in data relocation tree */
4277                 fs_root = read_fs_root(root->fs_info,
4278                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4279                 if (IS_ERR(fs_root))
4280                         err = PTR_ERR(fs_root);
4281                 else
4282                         err = btrfs_orphan_cleanup(fs_root);
4283         }
4284         return err;
4285 }
4286
4287 /*
4288  * helper to add ordered checksum for data relocation.
4289  *
4290  * cloning checksum properly handles the nodatasum extents.
4291  * it also saves CPU time to re-calculate the checksum.
4292  */
4293 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4294 {
4295         struct btrfs_ordered_sum *sums;
4296         struct btrfs_sector_sum *sector_sum;
4297         struct btrfs_ordered_extent *ordered;
4298         struct btrfs_root *root = BTRFS_I(inode)->root;
4299         size_t offset;
4300         int ret;
4301         u64 disk_bytenr;
4302         LIST_HEAD(list);
4303
4304         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4305         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4306
4307         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4308         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4309                                        disk_bytenr + len - 1, &list, 0);
4310         if (ret)
4311                 goto out;
4312
4313         while (!list_empty(&list)) {
4314                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4315                 list_del_init(&sums->list);
4316
4317                 sector_sum = sums->sums;
4318                 sums->bytenr = ordered->start;
4319
4320                 offset = 0;
4321                 while (offset < sums->len) {
4322                         sector_sum->bytenr += ordered->start - disk_bytenr;
4323                         sector_sum++;
4324                         offset += root->sectorsize;
4325                 }
4326
4327                 btrfs_add_ordered_sum(inode, ordered, sums);
4328         }
4329 out:
4330         btrfs_put_ordered_extent(ordered);
4331         return ret;
4332 }
4333
4334 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4335                            struct btrfs_root *root, struct extent_buffer *buf,
4336                            struct extent_buffer *cow)
4337 {
4338         struct reloc_control *rc;
4339         struct backref_node *node;
4340         int first_cow = 0;
4341         int level;
4342         int ret;
4343
4344         rc = root->fs_info->reloc_ctl;
4345         if (!rc)
4346                 return;
4347
4348         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4349                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4350
4351         level = btrfs_header_level(buf);
4352         if (btrfs_header_generation(buf) <=
4353             btrfs_root_last_snapshot(&root->root_item))
4354                 first_cow = 1;
4355
4356         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4357             rc->create_reloc_tree) {
4358                 WARN_ON(!first_cow && level == 0);
4359
4360                 node = rc->backref_cache.path[level];
4361                 BUG_ON(node->bytenr != buf->start &&
4362                        node->new_bytenr != buf->start);
4363
4364                 drop_node_buffer(node);
4365                 extent_buffer_get(cow);
4366                 node->eb = cow;
4367                 node->new_bytenr = cow->start;
4368
4369                 if (!node->pending) {
4370                         list_move_tail(&node->list,
4371                                        &rc->backref_cache.pending[level]);
4372                         node->pending = 1;
4373                 }
4374
4375                 if (first_cow)
4376                         __mark_block_processed(rc, node);
4377
4378                 if (first_cow && level > 0)
4379                         rc->nodes_relocated += buf->len;
4380         }
4381
4382         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4383                 ret = replace_file_extents(trans, rc, root, cow);
4384                 BUG_ON(ret);
4385         }
4386 }
4387
4388 /*
4389  * called before creating snapshot. it calculates metadata reservation
4390  * requried for relocating tree blocks in the snapshot
4391  */
4392 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4393                               struct btrfs_pending_snapshot *pending,
4394                               u64 *bytes_to_reserve)
4395 {
4396         struct btrfs_root *root;
4397         struct reloc_control *rc;
4398
4399         root = pending->root;
4400         if (!root->reloc_root)
4401                 return;
4402
4403         rc = root->fs_info->reloc_ctl;
4404         if (!rc->merge_reloc_tree)
4405                 return;
4406
4407         root = root->reloc_root;
4408         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4409         /*
4410          * relocation is in the stage of merging trees. the space
4411          * used by merging a reloc tree is twice the size of
4412          * relocated tree nodes in the worst case. half for cowing
4413          * the reloc tree, half for cowing the fs tree. the space
4414          * used by cowing the reloc tree will be freed after the
4415          * tree is dropped. if we create snapshot, cowing the fs
4416          * tree may use more space than it frees. so we need
4417          * reserve extra space.
4418          */
4419         *bytes_to_reserve += rc->nodes_relocated;
4420 }
4421
4422 /*
4423  * called after snapshot is created. migrate block reservation
4424  * and create reloc root for the newly created snapshot
4425  */
4426 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4427                                struct btrfs_pending_snapshot *pending)
4428 {
4429         struct btrfs_root *root = pending->root;
4430         struct btrfs_root *reloc_root;
4431         struct btrfs_root *new_root;
4432         struct reloc_control *rc;
4433         int ret;
4434
4435         if (!root->reloc_root)
4436                 return 0;
4437
4438         rc = root->fs_info->reloc_ctl;
4439         rc->merging_rsv_size += rc->nodes_relocated;
4440
4441         if (rc->merge_reloc_tree) {
4442                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4443                                               rc->block_rsv,
4444                                               rc->nodes_relocated);
4445                 if (ret)
4446                         return ret;
4447         }
4448
4449         new_root = pending->snap;
4450         reloc_root = create_reloc_root(trans, root->reloc_root,
4451                                        new_root->root_key.objectid);
4452         if (IS_ERR(reloc_root))
4453                 return PTR_ERR(reloc_root);
4454
4455         ret = __add_reloc_root(reloc_root);
4456         BUG_ON(ret < 0);
4457         new_root->reloc_root = reloc_root;
4458
4459         if (rc->create_reloc_tree)
4460                 ret = clone_backref_node(trans, rc, root, reloc_root);
4461         return ret;
4462 }