fs/btrfs/relocation.c

   1 /*
   2  * Copyright (C) 2009 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/rbtree.h>
  24 #include <linux/slab.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "volumes.h"
  29 #include "locking.h"
  30 #include "btrfs_inode.h"
  31 #include "async-thread.h"
  32 #include "free-space-cache.h"
  33 #include "inode-map.h"
  34
  35 /*
  36  * backref_node, mapping_node and tree_block start with this
  37  */
  38 struct tree_entry {
  39         struct rb_node rb_node;
  40         u64 bytenr;
  41 };
  42
  43 /*
  44  * present a tree block in the backref cache
  45  */
  46 struct backref_node {
  47         struct rb_node rb_node;
  48         u64 bytenr;
  49
  50         u64 new_bytenr;
  51         /* objectid of tree block owner, can be not uptodate */
  52         u64 owner;
  53         /* link to pending, changed or detached list */
  54         struct list_head list;
  55         /* list of upper level blocks reference this block */
  56         struct list_head upper;
  57         /* list of child blocks in the cache */
  58         struct list_head lower;
  59         /* NULL if this node is not tree root */
  60         struct btrfs_root *root;
  61         /* extent buffer got by COW the block */
  62         struct extent_buffer *eb;
  63         /* level of tree block */
  64         unsigned int level:8;
  65         /* is the block in non-reference counted tree */
  66         unsigned int cowonly:1;
  67         /* 1 if no child node in the cache */
  68         unsigned int lowest:1;
  69         /* is the extent buffer locked */
  70         unsigned int locked:1;
  71         /* has the block been processed */
  72         unsigned int processed:1;
  73         /* have backrefs of this block been checked */
  74         unsigned int checked:1;
  75         /*
  76          * 1 if corresponding block has been cowed but some upper
  77          * level block pointers may not point to the new location
  78          */
  79         unsigned int pending:1;
  80         /*
  81          * 1 if the backref node isn't connected to any other
  82          * backref node.
  83          */
  84         unsigned int detached:1;
  85 };
  86
  87 /*
  88  * present a block pointer in the backref cache
  89  */
  90 struct backref_edge {
  91         struct list_head list[2];
  92         struct backref_node *node[2];
  93 };
  94
  95 #define LOWER   0
  96 #define UPPER   1
  97
  98 struct backref_cache {
  99         /* red black tree of all backref nodes in the cache */
 100         struct rb_root rb_root;
 101         /* for passing backref nodes to btrfs_reloc_cow_block */
 102         struct backref_node *path[BTRFS_MAX_LEVEL];
 103         /*
 104          * list of blocks that have been cowed but some block
 105          * pointers in upper level blocks may not reflect the
 106          * new location
 107          */
 108         struct list_head pending[BTRFS_MAX_LEVEL];
 109         /* list of backref nodes with no child node */
 110         struct list_head leaves;
 111         /* list of blocks that have been cowed in current transaction */
 112         struct list_head changed;
 113         /* list of detached backref node. */
 114         struct list_head detached;
 115
 116         u64 last_trans;
 117
 118         int nr_nodes;
 119         int nr_edges;
 120 };
 121
 122 /*
 123  * map address of tree root to tree
 124  */
 125 struct mapping_node {
 126         struct rb_node rb_node;
 127         u64 bytenr;
 128         void *data;
 129 };
 130
 131 struct mapping_tree {
 132         struct rb_root rb_root;
 133         spinlock_t lock;
 134 };
 135
 136 /*
 137  * present a tree block to process
 138  */
 139 struct tree_block {
 140         struct rb_node rb_node;
 141         u64 bytenr;
 142         struct btrfs_key key;
 143         unsigned int level:8;
 144         unsigned int key_ready:1;
 145 };
 146
 147 #define MAX_EXTENTS 128
 148
 149 struct file_extent_cluster {
 150         u64 start;
 151         u64 end;
 152         u64 boundary[MAX_EXTENTS];
 153         unsigned int nr;
 154 };
 155
 156 struct reloc_control {
 157         /* block group to relocate */
 158         struct btrfs_block_group_cache *block_group;
 159         /* extent tree */
 160         struct btrfs_root *extent_root;
 161         /* inode for moving data */
 162         struct inode *data_inode;
 163
 164         struct btrfs_block_rsv *block_rsv;
 165
 166         struct backref_cache backref_cache;
 167
 168         struct file_extent_cluster cluster;
 169         /* tree blocks have been processed */
 170         struct extent_io_tree processed_blocks;
 171         /* map start of tree root to corresponding reloc tree */
 172         struct mapping_tree reloc_root_tree;
 173         /* list of reloc trees */
 174         struct list_head reloc_roots;
 175         /* size of metadata reservation for merging reloc trees */
 176         u64 merging_rsv_size;
 177         /* size of relocated tree nodes */
 178         u64 nodes_relocated;
 179
 180         u64 search_start;
 181         u64 extents_found;
 182
 183         unsigned int stage:8;
 184         unsigned int create_reloc_tree:1;
 185         unsigned int merge_reloc_tree:1;
 186         unsigned int found_file_extent:1;
 187         unsigned int commit_transaction:1;
 188 };
 189
 190 /* stages of data relocation */
 191 #define MOVE_DATA_EXTENTS       0
 192 #define UPDATE_DATA_PTRS        1
 193
 194 static void remove_backref_node(struct backref_cache *cache,
 195                                 struct backref_node *node);
 196 static void __mark_block_processed(struct reloc_control *rc,
 197                                    struct backref_node *node);
 198
 199 static void mapping_tree_init(struct mapping_tree *tree)
 200 {
 201         tree->rb_root = RB_ROOT;
 202         spin_lock_init(&tree->lock);
 203 }
 204
 205 static void backref_cache_init(struct backref_cache *cache)
 206 {
 207         int i;
 208         cache->rb_root = RB_ROOT;
 209         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 210                 INIT_LIST_HEAD(&cache->pending[i]);
 211         INIT_LIST_HEAD(&cache->changed);
 212         INIT_LIST_HEAD(&cache->detached);
 213         INIT_LIST_HEAD(&cache->leaves);
 214 }
 215
 216 static void backref_cache_cleanup(struct backref_cache *cache)
 217 {
 218         struct backref_node *node;
 219         int i;
 220
 221         while (!list_empty(&cache->detached)) {
 222                 node = list_entry(cache->detached.next,
 223                                   struct backref_node, list);
 224                 remove_backref_node(cache, node);
 225         }
 226
 227         while (!list_empty(&cache->leaves)) {
 228                 node = list_entry(cache->leaves.next,
 229                                   struct backref_node, lower);
 230                 remove_backref_node(cache, node);
 231         }
 232
 233         cache->last_trans = 0;
 234
 235         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 236                 BUG_ON(!list_empty(&cache->pending[i]));
 237         BUG_ON(!list_empty(&cache->changed));
 238         BUG_ON(!list_empty(&cache->detached));
 239         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
 240         BUG_ON(cache->nr_nodes);
 241         BUG_ON(cache->nr_edges);
 242 }
 243
 244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 245 {
 246         struct backref_node *node;
 247
 248         node = kzalloc(sizeof(*node), GFP_NOFS);
 249         if (node) {
 250                 INIT_LIST_HEAD(&node->list);
 251                 INIT_LIST_HEAD(&node->upper);
 252                 INIT_LIST_HEAD(&node->lower);
 253                 RB_CLEAR_NODE(&node->rb_node);
 254                 cache->nr_nodes++;
 255         }
 256         return node;
 257 }
 258
 259 static void free_backref_node(struct backref_cache *cache,
 260                               struct backref_node *node)
 261 {
 262         if (node) {
 263                 cache->nr_nodes--;
 264                 kfree(node);
 265         }
 266 }
 267
 268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 269 {
 270         struct backref_edge *edge;
 271
 272         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 273         if (edge)
 274                 cache->nr_edges++;
 275         return edge;
 276 }
 277
 278 static void free_backref_edge(struct backref_cache *cache,
 279                               struct backref_edge *edge)
 280 {
 281         if (edge) {
 282                 cache->nr_edges--;
 283                 kfree(edge);
 284         }
 285 }
 286
 287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 288                                    struct rb_node *node)
 289 {
 290         struct rb_node **p = &root->rb_node;
 291         struct rb_node *parent = NULL;
 292         struct tree_entry *entry;
 293
 294         while (*p) {
 295                 parent = *p;
 296                 entry = rb_entry(parent, struct tree_entry, rb_node);
 297
 298                 if (bytenr < entry->bytenr)
 299                         p = &(*p)->rb_left;
 300                 else if (bytenr > entry->bytenr)
 301                         p = &(*p)->rb_right;
 302                 else
 303                         return parent;
 304         }
 305
 306         rb_link_node(node, parent, p);
 307         rb_insert_color(node, root);
 308         return NULL;
 309 }
 310
 311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 312 {
 313         struct rb_node *n = root->rb_node;
 314         struct tree_entry *entry;
 315
 316         while (n) {
 317                 entry = rb_entry(n, struct tree_entry, rb_node);
 318
 319                 if (bytenr < entry->bytenr)
 320                         n = n->rb_left;
 321                 else if (bytenr > entry->bytenr)
 322                         n = n->rb_right;
 323                 else
 324                         return n;
 325         }
 326         return NULL;
 327 }
 328
 329 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 330 {
 331
 332         struct btrfs_fs_info *fs_info = NULL;
 333         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 334                                               rb_node);
 335         if (bnode->root)
 336                 fs_info = bnode->root->fs_info;
 337         btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
 338                     "found at offset %llu\n", bytenr);
 339 }
 340
 341 /*
 342  * walk up backref nodes until reach node presents tree root
 343  */
 344 static struct backref_node *walk_up_backref(struct backref_node *node,
 345                                             struct backref_edge *edges[],
 346                                             int *index)
 347 {
 348         struct backref_edge *edge;
 349         int idx = *index;
 350
 351         while (!list_empty(&node->upper)) {
 352                 edge = list_entry(node->upper.next,
 353                                   struct backref_edge, list[LOWER]);
 354                 edges[idx++] = edge;
 355                 node = edge->node[UPPER];
 356         }
 357         BUG_ON(node->detached);
 358         *index = idx;
 359         return node;
 360 }
 361
 362 /*
 363  * walk down backref nodes to find start of next reference path
 364  */
 365 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 366                                               int *index)
 367 {
 368         struct backref_edge *edge;
 369         struct backref_node *lower;
 370         int idx = *index;
 371
 372         while (idx > 0) {
 373                 edge = edges[idx - 1];
 374                 lower = edge->node[LOWER];
 375                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 376                         idx--;
 377                         continue;
 378                 }
 379                 edge = list_entry(edge->list[LOWER].next,
 380                                   struct backref_edge, list[LOWER]);
 381                 edges[idx - 1] = edge;
 382                 *index = idx;
 383                 return edge->node[UPPER];
 384         }
 385         *index = 0;
 386         return NULL;
 387 }
 388
 389 static void unlock_node_buffer(struct backref_node *node)
 390 {
 391         if (node->locked) {
 392                 btrfs_tree_unlock(node->eb);
 393                 node->locked = 0;
 394         }
 395 }
 396
 397 static void drop_node_buffer(struct backref_node *node)
 398 {
 399         if (node->eb) {
 400                 unlock_node_buffer(node);
 401                 free_extent_buffer(node->eb);
 402                 node->eb = NULL;
 403         }
 404 }
 405
 406 static void drop_backref_node(struct backref_cache *tree,
 407                               struct backref_node *node)
 408 {
 409         BUG_ON(!list_empty(&node->upper));
 410
 411         drop_node_buffer(node);
 412         list_del(&node->list);
 413         list_del(&node->lower);
 414         if (!RB_EMPTY_NODE(&node->rb_node))
 415                 rb_erase(&node->rb_node, &tree->rb_root);
 416         free_backref_node(tree, node);
 417 }
 418
 419 /*
 420  * remove a backref node from the backref cache
 421  */
 422 static void remove_backref_node(struct backref_cache *cache,
 423                                 struct backref_node *node)
 424 {
 425         struct backref_node *upper;
 426         struct backref_edge *edge;
 427
 428         if (!node)
 429                 return;
 430
 431         BUG_ON(!node->lowest && !node->detached);
 432         while (!list_empty(&node->upper)) {
 433                 edge = list_entry(node->upper.next, struct backref_edge,
 434                                   list[LOWER]);
 435                 upper = edge->node[UPPER];
 436                 list_del(&edge->list[LOWER]);
 437                 list_del(&edge->list[UPPER]);
 438                 free_backref_edge(cache, edge);
 439
 440                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 441                         BUG_ON(!list_empty(&node->upper));
 442                         drop_backref_node(cache, node);
 443                         node = upper;
 444                         node->lowest = 1;
 445                         continue;
 446                 }
 447                 /*
 448                  * add the node to leaf node list if no other
 449                  * child block cached.
 450                  */
 451                 if (list_empty(&upper->lower)) {
 452                         list_add_tail(&upper->lower, &cache->leaves);
 453                         upper->lowest = 1;
 454                 }
 455         }
 456
 457         drop_backref_node(cache, node);
 458 }
 459
 460 static void update_backref_node(struct backref_cache *cache,
 461                                 struct backref_node *node, u64 bytenr)
 462 {
 463         struct rb_node *rb_node;
 464         rb_erase(&node->rb_node, &cache->rb_root);
 465         node->bytenr = bytenr;
 466         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 467         if (rb_node)
 468                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 469 }
 470
 471 /*
 472  * update backref cache after a transaction commit
 473  */
 474 static int update_backref_cache(struct btrfs_trans_handle *trans,
 475                                 struct backref_cache *cache)
 476 {
 477         struct backref_node *node;
 478         int level = 0;
 479
 480         if (cache->last_trans == 0) {
 481                 cache->last_trans = trans->transid;
 482                 return 0;
 483         }
 484
 485         if (cache->last_trans == trans->transid)
 486                 return 0;
 487
 488         /*
 489          * detached nodes are used to avoid unnecessary backref
 490          * lookup. transaction commit changes the extent tree.
 491          * so the detached nodes are no longer useful.
 492          */
 493         while (!list_empty(&cache->detached)) {
 494                 node = list_entry(cache->detached.next,
 495                                   struct backref_node, list);
 496                 remove_backref_node(cache, node);
 497         }
 498
 499         while (!list_empty(&cache->changed)) {
 500                 node = list_entry(cache->changed.next,
 501                                   struct backref_node, list);
 502                 list_del_init(&node->list);
 503                 BUG_ON(node->pending);
 504                 update_backref_node(cache, node, node->new_bytenr);
 505         }
 506
 507         /*
 508          * some nodes can be left in the pending list if there were
 509          * errors during processing the pending nodes.
 510          */
 511         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 512                 list_for_each_entry(node, &cache->pending[level], list) {
 513                         BUG_ON(!node->pending);
 514                         if (node->bytenr == node->new_bytenr)
 515                                 continue;
 516                         update_backref_node(cache, node, node->new_bytenr);
 517                 }
 518         }
 519
 520         cache->last_trans = 0;
 521         return 1;
 522 }
 523
 524
 525 static int should_ignore_root(struct btrfs_root *root)
 526 {
 527         struct btrfs_root *reloc_root;
 528
 529         if (!root->ref_cows)
 530                 return 0;
 531
 532         reloc_root = root->reloc_root;
 533         if (!reloc_root)
 534                 return 0;
 535
 536         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 537             root->fs_info->running_transaction->transid - 1)
 538                 return 0;
 539         /*
 540          * if there is reloc tree and it was created in previous
 541          * transaction backref lookup can find the reloc tree,
 542          * so backref node for the fs tree root is useless for
 543          * relocation.
 544          */
 545         return 1;
 546 }
 547 /*
 548  * find reloc tree by address of tree root
 549  */
 550 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 551                                           u64 bytenr)
 552 {
 553         struct rb_node *rb_node;
 554         struct mapping_node *node;
 555         struct btrfs_root *root = NULL;
 556
 557         spin_lock(&rc->reloc_root_tree.lock);
 558         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 559         if (rb_node) {
 560                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 561                 root = (struct btrfs_root *)node->data;
 562         }
 563         spin_unlock(&rc->reloc_root_tree.lock);
 564         return root;
 565 }
 566
 567 static int is_cowonly_root(u64 root_objectid)
 568 {
 569         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 570             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 571             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 572             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 573             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 574             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
 575                 return 1;
 576         return 0;
 577 }
 578
 579 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 580                                         u64 root_objectid)
 581 {
 582         struct btrfs_key key;
 583
 584         key.objectid = root_objectid;
 585         key.type = BTRFS_ROOT_ITEM_KEY;
 586         if (is_cowonly_root(root_objectid))
 587                 key.offset = 0;
 588         else
 589                 key.offset = (u64)-1;
 590
 591         return btrfs_read_fs_root_no_name(fs_info, &key);
 592 }
 593
 594 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 595 static noinline_for_stack
 596 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 597                                   struct extent_buffer *leaf,
 598                                   struct btrfs_extent_ref_v0 *ref0)
 599 {
 600         struct btrfs_root *root;
 601         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 602         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 603
 604         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 605
 606         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 607         BUG_ON(IS_ERR(root));
 608
 609         if (root->ref_cows &&
 610             generation != btrfs_root_generation(&root->root_item))
 611                 return NULL;
 612
 613         return root;
 614 }
 615 #endif
 616
 617 static noinline_for_stack
 618 int find_inline_backref(struct extent_buffer *leaf, int slot,
 619                         unsigned long *ptr, unsigned long *end)
 620 {
 621         struct btrfs_key key;
 622         struct btrfs_extent_item *ei;
 623         struct btrfs_tree_block_info *bi;
 624         u32 item_size;
 625
 626         btrfs_item_key_to_cpu(leaf, &key, slot);
 627
 628         item_size = btrfs_item_size_nr(leaf, slot);
 629 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 630         if (item_size < sizeof(*ei)) {
 631                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 632                 return 1;
 633         }
 634 #endif
 635         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 636         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 637                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 638
 639         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 640             item_size <= sizeof(*ei) + sizeof(*bi)) {
 641                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 642                 return 1;
 643         }
 644         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 645             item_size <= sizeof(*ei)) {
 646                 WARN_ON(item_size < sizeof(*ei));
 647                 return 1;
 648         }
 649
 650         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 651                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 652                 *ptr = (unsigned long)(bi + 1);
 653         } else {
 654                 *ptr = (unsigned long)(ei + 1);
 655         }
 656         *end = (unsigned long)ei + item_size;
 657         return 0;
 658 }
 659
 660 /*
 661  * build backref tree for a given tree block. root of the backref tree
 662  * corresponds the tree block, leaves of the backref tree correspond
 663  * roots of b-trees that reference the tree block.
 664  *
 665  * the basic idea of this function is check backrefs of a given block
 666  * to find upper level blocks that refernece the block, and then check
 667  * bakcrefs of these upper level blocks recursively. the recursion stop
 668  * when tree root is reached or backrefs for the block is cached.
 669  *
 670  * NOTE: if we find backrefs for a block are cached, we know backrefs
 671  * for all upper level blocks that directly/indirectly reference the
 672  * block are also cached.
 673  */
 674 static noinline_for_stack
 675 struct backref_node *build_backref_tree(struct reloc_control *rc,
 676                                         struct btrfs_key *node_key,
 677                                         int level, u64 bytenr)
 678 {
 679         struct backref_cache *cache = &rc->backref_cache;
 680         struct btrfs_path *path1;
 681         struct btrfs_path *path2;
 682         struct extent_buffer *eb;
 683         struct btrfs_root *root;
 684         struct backref_node *cur;
 685         struct backref_node *upper;
 686         struct backref_node *lower;
 687         struct backref_node *node = NULL;
 688         struct backref_node *exist = NULL;
 689         struct backref_edge *edge;
 690         struct rb_node *rb_node;
 691         struct btrfs_key key;
 692         unsigned long end;
 693         unsigned long ptr;
 694         LIST_HEAD(list);
 695         LIST_HEAD(useless);
 696         int cowonly;
 697         int ret;
 698         int err = 0;
 699         bool need_check = true;
 700
 701         path1 = btrfs_alloc_path();
 702         path2 = btrfs_alloc_path();
 703         if (!path1 || !path2) {
 704                 err = -ENOMEM;
 705                 goto out;
 706         }
 707         path1->reada = 1;
 708         path2->reada = 2;
 709
 710         node = alloc_backref_node(cache);
 711         if (!node) {
 712                 err = -ENOMEM;
 713                 goto out;
 714         }
 715
 716         node->bytenr = bytenr;
 717         node->level = level;
 718         node->lowest = 1;
 719         cur = node;
 720 again:
 721         end = 0;
 722         ptr = 0;
 723         key.objectid = cur->bytenr;
 724         key.type = BTRFS_METADATA_ITEM_KEY;
 725         key.offset = (u64)-1;
 726
 727         path1->search_commit_root = 1;
 728         path1->skip_locking = 1;
 729         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 730                                 0, 0);
 731         if (ret < 0) {
 732                 err = ret;
 733                 goto out;
 734         }
 735         BUG_ON(!ret || !path1->slots[0]);
 736
 737         path1->slots[0]--;
 738
 739         WARN_ON(cur->checked);
 740         if (!list_empty(&cur->upper)) {
 741                 /*
 742                  * the backref was added previously when processing
 743                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 744                  */
 745                 BUG_ON(!list_is_singular(&cur->upper));
 746                 edge = list_entry(cur->upper.next, struct backref_edge,
 747                                   list[LOWER]);
 748                 BUG_ON(!list_empty(&edge->list[UPPER]));
 749                 exist = edge->node[UPPER];
 750                 /*
 751                  * add the upper level block to pending list if we need
 752                  * check its backrefs
 753                  */
 754                 if (!exist->checked)
 755                         list_add_tail(&edge->list[UPPER], &list);
 756         } else {
 757                 exist = NULL;
 758         }
 759
 760         while (1) {
 761                 cond_resched();
 762                 eb = path1->nodes[0];
 763
 764                 if (ptr >= end) {
 765                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 766                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 767                                 if (ret < 0) {
 768                                         err = ret;
 769                                         goto out;
 770                                 }
 771                                 if (ret > 0)
 772                                         break;
 773                                 eb = path1->nodes[0];
 774                         }
 775
 776                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 777                         if (key.objectid != cur->bytenr) {
 778                                 WARN_ON(exist);
 779                                 break;
 780                         }
 781
 782                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 783                             key.type == BTRFS_METADATA_ITEM_KEY) {
 784                                 ret = find_inline_backref(eb, path1->slots[0],
 785                                                           &ptr, &end);
 786                                 if (ret)
 787                                         goto next;
 788                         }
 789                 }
 790
 791                 if (ptr < end) {
 792                         /* update key for inline back ref */
 793                         struct btrfs_extent_inline_ref *iref;
 794                         iref = (struct btrfs_extent_inline_ref *)ptr;
 795                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 796                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 797                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 798                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 799                 }
 800
 801                 if (exist &&
 802                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 803                       exist->owner == key.offset) ||
 804                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 805                       exist->bytenr == key.offset))) {
 806                         exist = NULL;
 807                         goto next;
 808                 }
 809
 810 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 811                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 812                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 813                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 814                                 struct btrfs_extent_ref_v0 *ref0;
 815                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 816                                                 struct btrfs_extent_ref_v0);
 817                                 if (key.objectid == key.offset) {
 818                                         root = find_tree_root(rc, eb, ref0);
 819                                         if (root && !should_ignore_root(root))
 820                                                 cur->root = root;
 821                                         else
 822                                                 list_add(&cur->list, &useless);
 823                                         break;
 824                                 }
 825                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 826                                                                       ref0)))
 827                                         cur->cowonly = 1;
 828                         }
 829 #else
 830                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 831                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 832 #endif
 833                         if (key.objectid == key.offset) {
 834                                 /*
 835                                  * only root blocks of reloc trees use
 836                                  * backref of this type.
 837                                  */
 838                                 root = find_reloc_root(rc, cur->bytenr);
 839                                 BUG_ON(!root);
 840                                 cur->root = root;
 841                                 break;
 842                         }
 843
 844                         edge = alloc_backref_edge(cache);
 845                         if (!edge) {
 846                                 err = -ENOMEM;
 847                                 goto out;
 848                         }
 849                         rb_node = tree_search(&cache->rb_root, key.offset);
 850                         if (!rb_node) {
 851                                 upper = alloc_backref_node(cache);
 852                                 if (!upper) {
 853                                         free_backref_edge(cache, edge);
 854                                         err = -ENOMEM;
 855                                         goto out;
 856                                 }
 857                                 upper->bytenr = key.offset;
 858                                 upper->level = cur->level + 1;
 859                                 /*
 860                                  *  backrefs for the upper level block isn't
 861                                  *  cached, add the block to pending list
 862                                  */
 863                                 list_add_tail(&edge->list[UPPER], &list);
 864                         } else {
 865                                 upper = rb_entry(rb_node, struct backref_node,
 866                                                  rb_node);
 867                                 BUG_ON(!upper->checked);
 868                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 869                         }
 870                         list_add_tail(&edge->list[LOWER], &cur->upper);
 871                         edge->node[LOWER] = cur;
 872                         edge->node[UPPER] = upper;
 873
 874                         goto next;
 875                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 876                         goto next;
 877                 }
 878
 879                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 880                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 881                 if (IS_ERR(root)) {
 882                         err = PTR_ERR(root);
 883                         goto out;
 884                 }
 885
 886                 if (!root->ref_cows)
 887                         cur->cowonly = 1;
 888
 889                 if (btrfs_root_level(&root->root_item) == cur->level) {
 890                         /* tree root */
 891                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 892                                cur->bytenr);
 893                         if (should_ignore_root(root))
 894                                 list_add(&cur->list, &useless);
 895                         else
 896                                 cur->root = root;
 897                         break;
 898                 }
 899
 900                 level = cur->level + 1;
 901
 902                 /*
 903                  * searching the tree to find upper level blocks
 904                  * reference the block.
 905                  */
 906                 path2->search_commit_root = 1;
 907                 path2->skip_locking = 1;
 908                 path2->lowest_level = level;
 909                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 910                 path2->lowest_level = 0;
 911                 if (ret < 0) {
 912                         err = ret;
 913                         goto out;
 914                 }
 915                 if (ret > 0 && path2->slots[level] > 0)
 916                         path2->slots[level]--;
 917
 918                 eb = path2->nodes[level];
 919                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 920                         cur->bytenr);
 921
 922                 lower = cur;
 923                 need_check = true;
 924                 for (; level < BTRFS_MAX_LEVEL; level++) {
 925                         if (!path2->nodes[level]) {
 926                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 927                                        lower->bytenr);
 928                                 if (should_ignore_root(root))
 929                                         list_add(&lower->list, &useless);
 930                                 else
 931                                         lower->root = root;
 932                                 break;
 933                         }
 934
 935                         edge = alloc_backref_edge(cache);
 936                         if (!edge) {
 937                                 err = -ENOMEM;
 938                                 goto out;
 939                         }
 940
 941                         eb = path2->nodes[level];
 942                         rb_node = tree_search(&cache->rb_root, eb->start);
 943                         if (!rb_node) {
 944                                 upper = alloc_backref_node(cache);
 945                                 if (!upper) {
 946                                         free_backref_edge(cache, edge);
 947                                         err = -ENOMEM;
 948                                         goto out;
 949                                 }
 950                                 upper->bytenr = eb->start;
 951                                 upper->owner = btrfs_header_owner(eb);
 952                                 upper->level = lower->level + 1;
 953                                 if (!root->ref_cows)
 954                                         upper->cowonly = 1;
 955
 956                                 /*
 957                                  * if we know the block isn't shared
 958                                  * we can void checking its backrefs.
 959                                  */
 960                                 if (btrfs_block_can_be_shared(root, eb))
 961                                         upper->checked = 0;
 962                                 else
 963                                         upper->checked = 1;
 964
 965                                 /*
 966                                  * add the block to pending list if we
 967                                  * need check its backrefs, we only do this once
 968                                  * while walking up a tree as we will catch
 969                                  * anything else later on.
 970                                  */
 971                                 if (!upper->checked && need_check) {
 972                                         need_check = false;
 973                                         list_add_tail(&edge->list[UPPER],
 974                                                       &list);
 975                                 } else
 976                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 977                         } else {
 978                                 upper = rb_entry(rb_node, struct backref_node,
 979                                                  rb_node);
 980                                 BUG_ON(!upper->checked);
 981                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 982                                 if (!upper->owner)
 983                                         upper->owner = btrfs_header_owner(eb);
 984                         }
 985                         list_add_tail(&edge->list[LOWER], &lower->upper);
 986                         edge->node[LOWER] = lower;
 987                         edge->node[UPPER] = upper;
 988
 989                         if (rb_node)
 990                                 break;
 991                         lower = upper;
 992                         upper = NULL;
 993                 }
 994                 btrfs_release_path(path2);
 995 next:
 996                 if (ptr < end) {
 997                         ptr += btrfs_extent_inline_ref_size(key.type);
 998                         if (ptr >= end) {
 999                                 WARN_ON(ptr > end);
1000                                 ptr = 0;
1001                                 end = 0;
1002                         }
1003                 }
1004                 if (ptr >= end)
1005                         path1->slots[0]++;
1006         }
1007         btrfs_release_path(path1);
1008
1009         cur->checked = 1;
1010         WARN_ON(exist);
1011
1012         /* the pending list isn't empty, take the first block to process */
1013         if (!list_empty(&list)) {
1014                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1015                 list_del_init(&edge->list[UPPER]);
1016                 cur = edge->node[UPPER];
1017                 goto again;
1018         }
1019
1020         /*
1021          * everything goes well, connect backref nodes and insert backref nodes
1022          * into the cache.
1023          */
1024         BUG_ON(!node->checked);
1025         cowonly = node->cowonly;
1026         if (!cowonly) {
1027                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1028                                       &node->rb_node);
1029                 if (rb_node)
1030                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1031                 list_add_tail(&node->lower, &cache->leaves);
1032         }
1033
1034         list_for_each_entry(edge, &node->upper, list[LOWER])
1035                 list_add_tail(&edge->list[UPPER], &list);
1036
1037         while (!list_empty(&list)) {
1038                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1039                 list_del_init(&edge->list[UPPER]);
1040                 upper = edge->node[UPPER];
1041                 if (upper->detached) {
1042                         list_del(&edge->list[LOWER]);
1043                         lower = edge->node[LOWER];
1044                         free_backref_edge(cache, edge);
1045                         if (list_empty(&lower->upper))
1046                                 list_add(&lower->list, &useless);
1047                         continue;
1048                 }
1049
1050                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1051                         if (upper->lowest) {
1052                                 list_del_init(&upper->lower);
1053                                 upper->lowest = 0;
1054                         }
1055
1056                         list_add_tail(&edge->list[UPPER], &upper->lower);
1057                         continue;
1058                 }
1059
1060                 BUG_ON(!upper->checked);
1061                 BUG_ON(cowonly != upper->cowonly);
1062                 if (!cowonly) {
1063                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1064                                               &upper->rb_node);
1065                         if (rb_node)
1066                                 backref_tree_panic(rb_node, -EEXIST,
1067                                                    upper->bytenr);
1068                 }
1069
1070                 list_add_tail(&edge->list[UPPER], &upper->lower);
1071
1072                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1073                         list_add_tail(&edge->list[UPPER], &list);
1074         }
1075         /*
1076          * process useless backref nodes. backref nodes for tree leaves
1077          * are deleted from the cache. backref nodes for upper level
1078          * tree blocks are left in the cache to avoid unnecessary backref
1079          * lookup.
1080          */
1081         while (!list_empty(&useless)) {
1082                 upper = list_entry(useless.next, struct backref_node, list);
1083                 list_del_init(&upper->list);
1084                 BUG_ON(!list_empty(&upper->upper));
1085                 if (upper == node)
1086                         node = NULL;
1087                 if (upper->lowest) {
1088                         list_del_init(&upper->lower);
1089                         upper->lowest = 0;
1090                 }
1091                 while (!list_empty(&upper->lower)) {
1092                         edge = list_entry(upper->lower.next,
1093                                           struct backref_edge, list[UPPER]);
1094                         list_del(&edge->list[UPPER]);
1095                         list_del(&edge->list[LOWER]);
1096                         lower = edge->node[LOWER];
1097                         free_backref_edge(cache, edge);
1098
1099                         if (list_empty(&lower->upper))
1100                                 list_add(&lower->list, &useless);
1101                 }
1102                 __mark_block_processed(rc, upper);
1103                 if (upper->level > 0) {
1104                         list_add(&upper->list, &cache->detached);
1105                         upper->detached = 1;
1106                 } else {
1107                         rb_erase(&upper->rb_node, &cache->rb_root);
1108                         free_backref_node(cache, upper);
1109                 }
1110         }
1111 out:
1112         btrfs_free_path(path1);
1113         btrfs_free_path(path2);
1114         if (err) {
1115                 while (!list_empty(&useless)) {
1116                         lower = list_entry(useless.next,
1117                                            struct backref_node, upper);
1118                         list_del_init(&lower->upper);
1119                 }
1120                 upper = node;
1121                 INIT_LIST_HEAD(&list);
1122                 while (upper) {
1123                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1124                                 list_splice_tail(&upper->upper, &list);
1125                                 free_backref_node(cache, upper);
1126                         }
1127
1128                         if (list_empty(&list))
1129                                 break;
1130
1131                         edge = list_entry(list.next, struct backref_edge,
1132                                           list[LOWER]);
1133                         list_del(&edge->list[LOWER]);
1134                         upper = edge->node[UPPER];
1135                         free_backref_edge(cache, edge);
1136                 }
1137                 return ERR_PTR(err);
1138         }
1139         BUG_ON(node && node->detached);
1140         return node;
1141 }
1142
1143 /*
1144  * helper to add backref node for the newly created snapshot.
1145  * the backref node is created by cloning backref node that
1146  * corresponds to root of source tree
1147  */
1148 static int clone_backref_node(struct btrfs_trans_handle *trans,
1149                               struct reloc_control *rc,
1150                               struct btrfs_root *src,
1151                               struct btrfs_root *dest)
1152 {
1153         struct btrfs_root *reloc_root = src->reloc_root;
1154         struct backref_cache *cache = &rc->backref_cache;
1155         struct backref_node *node = NULL;
1156         struct backref_node *new_node;
1157         struct backref_edge *edge;
1158         struct backref_edge *new_edge;
1159         struct rb_node *rb_node;
1160
1161         if (cache->last_trans > 0)
1162                 update_backref_cache(trans, cache);
1163
1164         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1165         if (rb_node) {
1166                 node = rb_entry(rb_node, struct backref_node, rb_node);
1167                 if (node->detached)
1168                         node = NULL;
1169                 else
1170                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1171         }
1172
1173         if (!node) {
1174                 rb_node = tree_search(&cache->rb_root,
1175                                       reloc_root->commit_root->start);
1176                 if (rb_node) {
1177                         node = rb_entry(rb_node, struct backref_node,
1178                                         rb_node);
1179                         BUG_ON(node->detached);
1180                 }
1181         }
1182
1183         if (!node)
1184                 return 0;
1185
1186         new_node = alloc_backref_node(cache);
1187         if (!new_node)
1188                 return -ENOMEM;
1189
1190         new_node->bytenr = dest->node->start;
1191         new_node->level = node->level;
1192         new_node->lowest = node->lowest;
1193         new_node->checked = 1;
1194         new_node->root = dest;
1195
1196         if (!node->lowest) {
1197                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1198                         new_edge = alloc_backref_edge(cache);
1199                         if (!new_edge)
1200                                 goto fail;
1201
1202                         new_edge->node[UPPER] = new_node;
1203                         new_edge->node[LOWER] = edge->node[LOWER];
1204                         list_add_tail(&new_edge->list[UPPER],
1205                                       &new_node->lower);
1206                 }
1207         } else {
1208                 list_add_tail(&new_node->lower, &cache->leaves);
1209         }
1210
1211         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1212                               &new_node->rb_node);
1213         if (rb_node)
1214                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1215
1216         if (!new_node->lowest) {
1217                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1218                         list_add_tail(&new_edge->list[LOWER],
1219                                       &new_edge->node[LOWER]->upper);
1220                 }
1221         }
1222         return 0;
1223 fail:
1224         while (!list_empty(&new_node->lower)) {
1225                 new_edge = list_entry(new_node->lower.next,
1226                                       struct backref_edge, list[UPPER]);
1227                 list_del(&new_edge->list[UPPER]);
1228                 free_backref_edge(cache, new_edge);
1229         }
1230         free_backref_node(cache, new_node);
1231         return -ENOMEM;
1232 }
1233
1234 /*
1235  * helper to add 'address of tree root -> reloc tree' mapping
1236  */
1237 static int __must_check __add_reloc_root(struct btrfs_root *root)
1238 {
1239         struct rb_node *rb_node;
1240         struct mapping_node *node;
1241         struct reloc_control *rc = root->fs_info->reloc_ctl;
1242
1243         node = kmalloc(sizeof(*node), GFP_NOFS);
1244         if (!node)
1245                 return -ENOMEM;
1246
1247         node->bytenr = root->node->start;
1248         node->data = root;
1249
1250         spin_lock(&rc->reloc_root_tree.lock);
1251         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1252                               node->bytenr, &node->rb_node);
1253         spin_unlock(&rc->reloc_root_tree.lock);
1254         if (rb_node) {
1255                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1256                             "for start=%llu while inserting into relocation "
1257                             "tree\n", node->bytenr);
1258                 kfree(node);
1259                 return -EEXIST;
1260         }
1261
1262         list_add_tail(&root->root_list, &rc->reloc_roots);
1263         return 0;
1264 }
1265
1266 /*
1267  * helper to update/delete the 'address of tree root -> reloc tree'
1268  * mapping
1269  */
1270 static int __update_reloc_root(struct btrfs_root *root, int del)
1271 {
1272         struct rb_node *rb_node;
1273         struct mapping_node *node = NULL;
1274         struct reloc_control *rc = root->fs_info->reloc_ctl;
1275
1276         spin_lock(&rc->reloc_root_tree.lock);
1277         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1278                               root->commit_root->start);
1279         if (rb_node) {
1280                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1281                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1282         }
1283         spin_unlock(&rc->reloc_root_tree.lock);
1284
1285         if (!node)
1286                 return 0;
1287         BUG_ON((struct btrfs_root *)node->data != root);
1288
1289         if (!del) {
1290                 spin_lock(&rc->reloc_root_tree.lock);
1291                 node->bytenr = root->node->start;
1292                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1293                                       node->bytenr, &node->rb_node);
1294                 spin_unlock(&rc->reloc_root_tree.lock);
1295                 if (rb_node)
1296                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1297         } else {
1298                 spin_lock(&root->fs_info->trans_lock);
1299                 list_del_init(&root->root_list);
1300                 spin_unlock(&root->fs_info->trans_lock);
1301                 kfree(node);
1302         }
1303         return 0;
1304 }
1305
1306 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1307                                         struct btrfs_root *root, u64 objectid)
1308 {
1309         struct btrfs_root *reloc_root;
1310         struct extent_buffer *eb;
1311         struct btrfs_root_item *root_item;
1312         struct btrfs_key root_key;
1313         u64 last_snap = 0;
1314         int ret;
1315
1316         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1317         BUG_ON(!root_item);
1318
1319         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1320         root_key.type = BTRFS_ROOT_ITEM_KEY;
1321         root_key.offset = objectid;
1322
1323         if (root->root_key.objectid == objectid) {
1324                 /* called by btrfs_init_reloc_root */
1325                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1326                                       BTRFS_TREE_RELOC_OBJECTID);
1327                 BUG_ON(ret);
1328
1329                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1330                 btrfs_set_root_last_snapshot(&root->root_item,
1331                                              trans->transid - 1);
1332         } else {
1333                 /*
1334                  * called by btrfs_reloc_post_snapshot_hook.
1335                  * the source tree is a reloc tree, all tree blocks
1336                  * modified after it was created have RELOC flag
1337                  * set in their headers. so it's OK to not update
1338                  * the 'last_snapshot'.
1339                  */
1340                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1341                                       BTRFS_TREE_RELOC_OBJECTID);
1342                 BUG_ON(ret);
1343         }
1344
1345         memcpy(root_item, &root->root_item, sizeof(*root_item));
1346         btrfs_set_root_bytenr(root_item, eb->start);
1347         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1348         btrfs_set_root_generation(root_item, trans->transid);
1349
1350         if (root->root_key.objectid == objectid) {
1351                 btrfs_set_root_refs(root_item, 0);
1352                 memset(&root_item->drop_progress, 0,
1353                        sizeof(struct btrfs_disk_key));
1354                 root_item->drop_level = 0;
1355                 /*
1356                  * abuse rtransid, it is safe because it is impossible to
1357                  * receive data into a relocation tree.
1358                  */
1359                 btrfs_set_root_rtransid(root_item, last_snap);
1360                 btrfs_set_root_otransid(root_item, trans->transid);
1361         }
1362
1363         btrfs_tree_unlock(eb);
1364         free_extent_buffer(eb);
1365
1366         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1367                                 &root_key, root_item);
1368         BUG_ON(ret);
1369         kfree(root_item);
1370
1371         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1372         BUG_ON(IS_ERR(reloc_root));
1373         reloc_root->last_trans = trans->transid;
1374         return reloc_root;
1375 }
1376
1377 /*
1378  * create reloc tree for a given fs tree. reloc tree is just a
1379  * snapshot of the fs tree with special root objectid.
1380  */
1381 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1382                           struct btrfs_root *root)
1383 {
1384         struct btrfs_root *reloc_root;
1385         struct reloc_control *rc = root->fs_info->reloc_ctl;
1386         int clear_rsv = 0;
1387         int ret;
1388
1389         if (root->reloc_root) {
1390                 reloc_root = root->reloc_root;
1391                 reloc_root->last_trans = trans->transid;
1392                 return 0;
1393         }
1394
1395         if (!rc || !rc->create_reloc_tree ||
1396             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1397                 return 0;
1398
1399         if (!trans->block_rsv) {
1400                 trans->block_rsv = rc->block_rsv;
1401                 clear_rsv = 1;
1402         }
1403         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1404         if (clear_rsv)
1405                 trans->block_rsv = NULL;
1406
1407         ret = __add_reloc_root(reloc_root);
1408         BUG_ON(ret < 0);
1409         root->reloc_root = reloc_root;
1410         return 0;
1411 }
1412
1413 /*
1414  * update root item of reloc tree
1415  */
1416 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1417                             struct btrfs_root *root)
1418 {
1419         struct btrfs_root *reloc_root;
1420         struct btrfs_root_item *root_item;
1421         int del = 0;
1422         int ret;
1423
1424         if (!root->reloc_root)
1425                 goto out;
1426
1427         reloc_root = root->reloc_root;
1428         root_item = &reloc_root->root_item;
1429
1430         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1431             btrfs_root_refs(root_item) == 0) {
1432                 root->reloc_root = NULL;
1433                 del = 1;
1434         }
1435
1436         __update_reloc_root(reloc_root, del);
1437
1438         if (reloc_root->commit_root != reloc_root->node) {
1439                 btrfs_set_root_node(root_item, reloc_root->node);
1440                 free_extent_buffer(reloc_root->commit_root);
1441                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1442         }
1443
1444         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1445                                 &reloc_root->root_key, root_item);
1446         BUG_ON(ret);
1447
1448 out:
1449         return 0;
1450 }
1451
1452 /*
1453  * helper to find first cached inode with inode number >= objectid
1454  * in a subvolume
1455  */
1456 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1457 {
1458         struct rb_node *node;
1459         struct rb_node *prev;
1460         struct btrfs_inode *entry;
1461         struct inode *inode;
1462
1463         spin_lock(&root->inode_lock);
1464 again:
1465         node = root->inode_tree.rb_node;
1466         prev = NULL;
1467         while (node) {
1468                 prev = node;
1469                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1470
1471                 if (objectid < btrfs_ino(&entry->vfs_inode))
1472                         node = node->rb_left;
1473                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1474                         node = node->rb_right;
1475                 else
1476                         break;
1477         }
1478         if (!node) {
1479                 while (prev) {
1480                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1481                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1482                                 node = prev;
1483                                 break;
1484                         }
1485                         prev = rb_next(prev);
1486                 }
1487         }
1488         while (node) {
1489                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1490                 inode = igrab(&entry->vfs_inode);
1491                 if (inode) {
1492                         spin_unlock(&root->inode_lock);
1493                         return inode;
1494                 }
1495
1496                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1497                 if (cond_resched_lock(&root->inode_lock))
1498                         goto again;
1499
1500                 node = rb_next(node);
1501         }
1502         spin_unlock(&root->inode_lock);
1503         return NULL;
1504 }
1505
1506 static int in_block_group(u64 bytenr,
1507                           struct btrfs_block_group_cache *block_group)
1508 {
1509         if (bytenr >= block_group->key.objectid &&
1510             bytenr < block_group->key.objectid + block_group->key.offset)
1511                 return 1;
1512         return 0;
1513 }
1514
1515 /*
1516  * get new location of data
1517  */
1518 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1519                             u64 bytenr, u64 num_bytes)
1520 {
1521         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1522         struct btrfs_path *path;
1523         struct btrfs_file_extent_item *fi;
1524         struct extent_buffer *leaf;
1525         int ret;
1526
1527         path = btrfs_alloc_path();
1528         if (!path)
1529                 return -ENOMEM;
1530
1531         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1532         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1533                                        bytenr, 0);
1534         if (ret < 0)
1535                 goto out;
1536         if (ret > 0) {
1537                 ret = -ENOENT;
1538                 goto out;
1539         }
1540
1541         leaf = path->nodes[0];
1542         fi = btrfs_item_ptr(leaf, path->slots[0],
1543                             struct btrfs_file_extent_item);
1544
1545         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1546                btrfs_file_extent_compression(leaf, fi) ||
1547                btrfs_file_extent_encryption(leaf, fi) ||
1548                btrfs_file_extent_other_encoding(leaf, fi));
1549
1550         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1551                 ret = 1;
1552                 goto out;
1553         }
1554
1555         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1556         ret = 0;
1557 out:
1558         btrfs_free_path(path);
1559         return ret;
1560 }
1561
1562 /*
1563  * update file extent items in the tree leaf to point to
1564  * the new locations.
1565  */
1566 static noinline_for_stack
1567 int replace_file_extents(struct btrfs_trans_handle *trans,
1568                          struct reloc_control *rc,
1569                          struct btrfs_root *root,
1570                          struct extent_buffer *leaf)
1571 {
1572         struct btrfs_key key;
1573         struct btrfs_file_extent_item *fi;
1574         struct inode *inode = NULL;
1575         u64 parent;
1576         u64 bytenr;
1577         u64 new_bytenr = 0;
1578         u64 num_bytes;
1579         u64 end;
1580         u32 nritems;
1581         u32 i;
1582         int ret;
1583         int first = 1;
1584         int dirty = 0;
1585
1586         if (rc->stage != UPDATE_DATA_PTRS)
1587                 return 0;
1588
1589         /* reloc trees always use full backref */
1590         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1591                 parent = leaf->start;
1592         else
1593                 parent = 0;
1594
1595         nritems = btrfs_header_nritems(leaf);
1596         for (i = 0; i < nritems; i++) {
1597                 cond_resched();
1598                 btrfs_item_key_to_cpu(leaf, &key, i);
1599                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1600                         continue;
1601                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1602                 if (btrfs_file_extent_type(leaf, fi) ==
1603                     BTRFS_FILE_EXTENT_INLINE)
1604                         continue;
1605                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1606                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1607                 if (bytenr == 0)
1608                         continue;
1609                 if (!in_block_group(bytenr, rc->block_group))
1610                         continue;
1611
1612                 /*
1613                  * if we are modifying block in fs tree, wait for readpage
1614                  * to complete and drop the extent cache
1615                  */
1616                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1617                         if (first) {
1618                                 inode = find_next_inode(root, key.objectid);
1619                                 first = 0;
1620                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1621                                 btrfs_add_delayed_iput(inode);
1622                                 inode = find_next_inode(root, key.objectid);
1623                         }
1624                         if (inode && btrfs_ino(inode) == key.objectid) {
1625                                 end = key.offset +
1626                                       btrfs_file_extent_num_bytes(leaf, fi);
1627                                 WARN_ON(!IS_ALIGNED(key.offset,
1628                                                     root->sectorsize));
1629                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1630                                 end--;
1631                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1632                                                       key.offset, end);
1633                                 if (!ret)
1634                                         continue;
1635
1636                                 btrfs_drop_extent_cache(inode, key.offset, end,
1637                                                         1);
1638                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1639                                               key.offset, end);
1640                         }
1641                 }
1642
1643                 ret = get_new_location(rc->data_inode, &new_bytenr,
1644                                        bytenr, num_bytes);
1645                 if (ret > 0) {
1646                         WARN_ON(1);
1647                         continue;
1648                 }
1649                 BUG_ON(ret < 0);
1650
1651                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1652                 dirty = 1;
1653
1654                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1655                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1656                                            num_bytes, parent,
1657                                            btrfs_header_owner(leaf),
1658                                            key.objectid, key.offset, 1);
1659                 BUG_ON(ret);
1660
1661                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1662                                         parent, btrfs_header_owner(leaf),
1663                                         key.objectid, key.offset, 1);
1664                 BUG_ON(ret);
1665         }
1666         if (dirty)
1667                 btrfs_mark_buffer_dirty(leaf);
1668         if (inode)
1669                 btrfs_add_delayed_iput(inode);
1670         return 0;
1671 }
1672
1673 static noinline_for_stack
1674 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1675                      struct btrfs_path *path, int level)
1676 {
1677         struct btrfs_disk_key key1;
1678         struct btrfs_disk_key key2;
1679         btrfs_node_key(eb, &key1, slot);
1680         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1681         return memcmp(&key1, &key2, sizeof(key1));
1682 }
1683
1684 /*
1685  * try to replace tree blocks in fs tree with the new blocks
1686  * in reloc tree. tree blocks haven't been modified since the
1687  * reloc tree was create can be replaced.
1688  *
1689  * if a block was replaced, level of the block + 1 is returned.
1690  * if no block got replaced, 0 is returned. if there are other
1691  * errors, a negative error number is returned.
1692  */
1693 static noinline_for_stack
1694 int replace_path(struct btrfs_trans_handle *trans,
1695                  struct btrfs_root *dest, struct btrfs_root *src,
1696                  struct btrfs_path *path, struct btrfs_key *next_key,
1697                  int lowest_level, int max_level)
1698 {
1699         struct extent_buffer *eb;
1700         struct extent_buffer *parent;
1701         struct btrfs_key key;
1702         u64 old_bytenr;
1703         u64 new_bytenr;
1704         u64 old_ptr_gen;
1705         u64 new_ptr_gen;
1706         u64 last_snapshot;
1707         u32 blocksize;
1708         int cow = 0;
1709         int level;
1710         int ret;
1711         int slot;
1712
1713         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1714         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1715
1716         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1717 again:
1718         slot = path->slots[lowest_level];
1719         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1720
1721         eb = btrfs_lock_root_node(dest);
1722         btrfs_set_lock_blocking(eb);
1723         level = btrfs_header_level(eb);
1724
1725         if (level < lowest_level) {
1726                 btrfs_tree_unlock(eb);
1727                 free_extent_buffer(eb);
1728                 return 0;
1729         }
1730
1731         if (cow) {
1732                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1733                 BUG_ON(ret);
1734         }
1735         btrfs_set_lock_blocking(eb);
1736
1737         if (next_key) {
1738                 next_key->objectid = (u64)-1;
1739                 next_key->type = (u8)-1;
1740                 next_key->offset = (u64)-1;
1741         }
1742
1743         parent = eb;
1744         while (1) {
1745                 level = btrfs_header_level(parent);
1746                 BUG_ON(level < lowest_level);
1747
1748                 ret = btrfs_bin_search(parent, &key, level, &slot);
1749                 if (ret && slot > 0)
1750                         slot--;
1751
1752                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1753                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1754
1755                 old_bytenr = btrfs_node_blockptr(parent, slot);
1756                 blocksize = btrfs_level_size(dest, level - 1);
1757                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1758
1759                 if (level <= max_level) {
1760                         eb = path->nodes[level];
1761                         new_bytenr = btrfs_node_blockptr(eb,
1762                                                         path->slots[level]);
1763                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1764                                                         path->slots[level]);
1765                 } else {
1766                         new_bytenr = 0;
1767                         new_ptr_gen = 0;
1768                 }
1769
1770                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1771                         WARN_ON(1);
1772                         ret = level;
1773                         break;
1774                 }
1775
1776                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1777                     memcmp_node_keys(parent, slot, path, level)) {
1778                         if (level <= lowest_level) {
1779                                 ret = 0;
1780                                 break;
1781                         }
1782
1783                         eb = read_tree_block(dest, old_bytenr, blocksize,
1784                                              old_ptr_gen);
1785                         if (!eb || !extent_buffer_uptodate(eb)) {
1786                                 ret = (!eb) ? -ENOMEM : -EIO;
1787                                 free_extent_buffer(eb);
1788                                 break;
1789                         }
1790                         btrfs_tree_lock(eb);
1791                         if (cow) {
1792                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1793                                                       slot, &eb);
1794                                 BUG_ON(ret);
1795                         }
1796                         btrfs_set_lock_blocking(eb);
1797
1798                         btrfs_tree_unlock(parent);
1799                         free_extent_buffer(parent);
1800
1801                         parent = eb;
1802                         continue;
1803                 }
1804
1805                 if (!cow) {
1806                         btrfs_tree_unlock(parent);
1807                         free_extent_buffer(parent);
1808                         cow = 1;
1809                         goto again;
1810                 }
1811
1812                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1813                                       path->slots[level]);
1814                 btrfs_release_path(path);
1815
1816                 path->lowest_level = level;
1817                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1818                 path->lowest_level = 0;
1819                 BUG_ON(ret);
1820
1821                 /*
1822                  * swap blocks in fs tree and reloc tree.
1823                  */
1824                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1825                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1826                 btrfs_mark_buffer_dirty(parent);
1827
1828                 btrfs_set_node_blockptr(path->nodes[level],
1829                                         path->slots[level], old_bytenr);
1830                 btrfs_set_node_ptr_generation(path->nodes[level],
1831                                               path->slots[level], old_ptr_gen);
1832                 btrfs_mark_buffer_dirty(path->nodes[level]);
1833
1834                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1835                                         path->nodes[level]->start,
1836                                         src->root_key.objectid, level - 1, 0,
1837                                         1);
1838                 BUG_ON(ret);
1839                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1840                                         0, dest->root_key.objectid, level - 1,
1841                                         0, 1);
1842                 BUG_ON(ret);
1843
1844                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1845                                         path->nodes[level]->start,
1846                                         src->root_key.objectid, level - 1, 0,
1847                                         1);
1848                 BUG_ON(ret);
1849
1850                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1851                                         0, dest->root_key.objectid, level - 1,
1852                                         0, 1);
1853                 BUG_ON(ret);
1854
1855                 btrfs_unlock_up_safe(path, 0);
1856
1857                 ret = level;
1858                 break;
1859         }
1860         btrfs_tree_unlock(parent);
1861         free_extent_buffer(parent);
1862         return ret;
1863 }
1864
1865 /*
1866  * helper to find next relocated block in reloc tree
1867  */
1868 static noinline_for_stack
1869 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1870                        int *level)
1871 {
1872         struct extent_buffer *eb;
1873         int i;
1874         u64 last_snapshot;
1875         u32 nritems;
1876
1877         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1878
1879         for (i = 0; i < *level; i++) {
1880                 free_extent_buffer(path->nodes[i]);
1881                 path->nodes[i] = NULL;
1882         }
1883
1884         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1885                 eb = path->nodes[i];
1886                 nritems = btrfs_header_nritems(eb);
1887                 while (path->slots[i] + 1 < nritems) {
1888                         path->slots[i]++;
1889                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1890                             last_snapshot)
1891                                 continue;
1892
1893                         *level = i;
1894                         return 0;
1895                 }
1896                 free_extent_buffer(path->nodes[i]);
1897                 path->nodes[i] = NULL;
1898         }
1899         return 1;
1900 }
1901
1902 /*
1903  * walk down reloc tree to find relocated block of lowest level
1904  */
1905 static noinline_for_stack
1906 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1907                          int *level)
1908 {
1909         struct extent_buffer *eb = NULL;
1910         int i;
1911         u64 bytenr;
1912         u64 ptr_gen = 0;
1913         u64 last_snapshot;
1914         u32 blocksize;
1915         u32 nritems;
1916
1917         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1918
1919         for (i = *level; i > 0; i--) {
1920                 eb = path->nodes[i];
1921                 nritems = btrfs_header_nritems(eb);
1922                 while (path->slots[i] < nritems) {
1923                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1924                         if (ptr_gen > last_snapshot)
1925                                 break;
1926                         path->slots[i]++;
1927                 }
1928                 if (path->slots[i] >= nritems) {
1929                         if (i == *level)
1930                                 break;
1931                         *level = i + 1;
1932                         return 0;
1933                 }
1934                 if (i == 1) {
1935                         *level = i;
1936                         return 0;
1937                 }
1938
1939                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1940                 blocksize = btrfs_level_size(root, i - 1);
1941                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1942                 if (!eb || !extent_buffer_uptodate(eb)) {
1943                         free_extent_buffer(eb);
1944                         return -EIO;
1945                 }
1946                 BUG_ON(btrfs_header_level(eb) != i - 1);
1947                 path->nodes[i - 1] = eb;
1948                 path->slots[i - 1] = 0;
1949         }
1950         return 1;
1951 }
1952
1953 /*
1954  * invalidate extent cache for file extents whose key in range of
1955  * [min_key, max_key)
1956  */
1957 static int invalidate_extent_cache(struct btrfs_root *root,
1958                                    struct btrfs_key *min_key,
1959                                    struct btrfs_key *max_key)
1960 {
1961         struct inode *inode = NULL;
1962         u64 objectid;
1963         u64 start, end;
1964         u64 ino;
1965
1966         objectid = min_key->objectid;
1967         while (1) {
1968                 cond_resched();
1969                 iput(inode);
1970
1971                 if (objectid > max_key->objectid)
1972                         break;
1973
1974                 inode = find_next_inode(root, objectid);
1975                 if (!inode)
1976                         break;
1977                 ino = btrfs_ino(inode);
1978
1979                 if (ino > max_key->objectid) {
1980                         iput(inode);
1981                         break;
1982                 }
1983
1984                 objectid = ino + 1;
1985                 if (!S_ISREG(inode->i_mode))
1986                         continue;
1987
1988                 if (unlikely(min_key->objectid == ino)) {
1989                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1990                                 continue;
1991                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1992                                 start = 0;
1993                         else {
1994                                 start = min_key->offset;
1995                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1996                         }
1997                 } else {
1998                         start = 0;
1999                 }
2000
2001                 if (unlikely(max_key->objectid == ino)) {
2002                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2003                                 continue;
2004                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2005                                 end = (u64)-1;
2006                         } else {
2007                                 if (max_key->offset == 0)
2008                                         continue;
2009                                 end = max_key->offset;
2010                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2011                                 end--;
2012                         }
2013                 } else {
2014                         end = (u64)-1;
2015                 }
2016
2017                 /* the lock_extent waits for readpage to complete */
2018                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2019                 btrfs_drop_extent_cache(inode, start, end, 1);
2020                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2021         }
2022         return 0;
2023 }
2024
2025 static int find_next_key(struct btrfs_path *path, int level,
2026                          struct btrfs_key *key)
2027
2028 {
2029         while (level < BTRFS_MAX_LEVEL) {
2030                 if (!path->nodes[level])
2031                         break;
2032                 if (path->slots[level] + 1 <
2033                     btrfs_header_nritems(path->nodes[level])) {
2034                         btrfs_node_key_to_cpu(path->nodes[level], key,
2035                                               path->slots[level] + 1);
2036                         return 0;
2037                 }
2038                 level++;
2039         }
2040         return 1;
2041 }
2042
2043 /*
2044  * merge the relocated tree blocks in reloc tree with corresponding
2045  * fs tree.
2046  */
2047 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2048                                                struct btrfs_root *root)
2049 {
2050         LIST_HEAD(inode_list);
2051         struct btrfs_key key;
2052         struct btrfs_key next_key;
2053         struct btrfs_trans_handle *trans;
2054         struct btrfs_root *reloc_root;
2055         struct btrfs_root_item *root_item;
2056         struct btrfs_path *path;
2057         struct extent_buffer *leaf;
2058         int level;
2059         int max_level;
2060         int replaced = 0;
2061         int ret;
2062         int err = 0;
2063         u32 min_reserved;
2064
2065         path = btrfs_alloc_path();
2066         if (!path)
2067                 return -ENOMEM;
2068         path->reada = 1;
2069
2070         reloc_root = root->reloc_root;
2071         root_item = &reloc_root->root_item;
2072
2073         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2074                 level = btrfs_root_level(root_item);
2075                 extent_buffer_get(reloc_root->node);
2076                 path->nodes[level] = reloc_root->node;
2077                 path->slots[level] = 0;
2078         } else {
2079                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2080
2081                 level = root_item->drop_level;
2082                 BUG_ON(level == 0);
2083                 path->lowest_level = level;
2084                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2085                 path->lowest_level = 0;
2086                 if (ret < 0) {
2087                         btrfs_free_path(path);
2088                         return ret;
2089                 }
2090
2091                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2092                                       path->slots[level]);
2093                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2094
2095                 btrfs_unlock_up_safe(path, 0);
2096         }
2097
2098         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2099         memset(&next_key, 0, sizeof(next_key));
2100
2101         while (1) {
2102                 trans = btrfs_start_transaction(root, 0);
2103                 BUG_ON(IS_ERR(trans));
2104                 trans->block_rsv = rc->block_rsv;
2105
2106                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2107                                              BTRFS_RESERVE_FLUSH_ALL);
2108                 if (ret) {
2109                         BUG_ON(ret != -EAGAIN);
2110                         ret = btrfs_commit_transaction(trans, root);
2111                         BUG_ON(ret);
2112                         continue;
2113                 }
2114
2115                 replaced = 0;
2116                 max_level = level;
2117
2118                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2119                 if (ret < 0) {
2120                         err = ret;
2121                         goto out;
2122                 }
2123                 if (ret > 0)
2124                         break;
2125
2126                 if (!find_next_key(path, level, &key) &&
2127                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2128                         ret = 0;
2129                 } else {
2130                         ret = replace_path(trans, root, reloc_root, path,
2131                                            &next_key, level, max_level);
2132                 }
2133                 if (ret < 0) {
2134                         err = ret;
2135                         goto out;
2136                 }
2137
2138                 if (ret > 0) {
2139                         level = ret;
2140                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2141                                               path->slots[level]);
2142                         replaced = 1;
2143                 }
2144
2145                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2146                 if (ret > 0)
2147                         break;
2148
2149                 BUG_ON(level == 0);
2150                 /*
2151                  * save the merging progress in the drop_progress.
2152                  * this is OK since root refs == 1 in this case.
2153                  */
2154                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2155                                path->slots[level]);
2156                 root_item->drop_level = level;
2157
2158                 btrfs_end_transaction_throttle(trans, root);
2159
2160                 btrfs_btree_balance_dirty(root);
2161
2162                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2163                         invalidate_extent_cache(root, &key, &next_key);
2164         }
2165
2166         /*
2167          * handle the case only one block in the fs tree need to be
2168          * relocated and the block is tree root.
2169          */
2170         leaf = btrfs_lock_root_node(root);
2171         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2172         btrfs_tree_unlock(leaf);
2173         free_extent_buffer(leaf);
2174         if (ret < 0)
2175                 err = ret;
2176 out:
2177         btrfs_free_path(path);
2178
2179         if (err == 0) {
2180                 memset(&root_item->drop_progress, 0,
2181                        sizeof(root_item->drop_progress));
2182                 root_item->drop_level = 0;
2183                 btrfs_set_root_refs(root_item, 0);
2184                 btrfs_update_reloc_root(trans, root);
2185         }
2186
2187         btrfs_end_transaction_throttle(trans, root);
2188
2189         btrfs_btree_balance_dirty(root);
2190
2191         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2192                 invalidate_extent_cache(root, &key, &next_key);
2193
2194         return err;
2195 }
2196
2197 static noinline_for_stack
2198 int prepare_to_merge(struct reloc_control *rc, int err)
2199 {
2200         struct btrfs_root *root = rc->extent_root;
2201         struct btrfs_root *reloc_root;
2202         struct btrfs_trans_handle *trans;
2203         LIST_HEAD(reloc_roots);
2204         u64 num_bytes = 0;
2205         int ret;
2206
2207         mutex_lock(&root->fs_info->reloc_mutex);
2208         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2209         rc->merging_rsv_size += rc->nodes_relocated * 2;
2210         mutex_unlock(&root->fs_info->reloc_mutex);
2211
2212 again:
2213         if (!err) {
2214                 num_bytes = rc->merging_rsv_size;
2215                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2216                                           BTRFS_RESERVE_FLUSH_ALL);
2217                 if (ret)
2218                         err = ret;
2219         }
2220
2221         trans = btrfs_join_transaction(rc->extent_root);
2222         if (IS_ERR(trans)) {
2223                 if (!err)
2224                         btrfs_block_rsv_release(rc->extent_root,
2225                                                 rc->block_rsv, num_bytes);
2226                 return PTR_ERR(trans);
2227         }
2228
2229         if (!err) {
2230                 if (num_bytes != rc->merging_rsv_size) {
2231                         btrfs_end_transaction(trans, rc->extent_root);
2232                         btrfs_block_rsv_release(rc->extent_root,
2233                                                 rc->block_rsv, num_bytes);
2234                         goto again;
2235                 }
2236         }
2237
2238         rc->merge_reloc_tree = 1;
2239
2240         while (!list_empty(&rc->reloc_roots)) {
2241                 reloc_root = list_entry(rc->reloc_roots.next,
2242                                         struct btrfs_root, root_list);
2243                 list_del_init(&reloc_root->root_list);
2244
2245                 root = read_fs_root(reloc_root->fs_info,
2246                                     reloc_root->root_key.offset);
2247                 BUG_ON(IS_ERR(root));
2248                 BUG_ON(root->reloc_root != reloc_root);
2249
2250                 /*
2251                  * set reference count to 1, so btrfs_recover_relocation
2252                  * knows it should resumes merging
2253                  */
2254                 if (!err)
2255                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2256                 btrfs_update_reloc_root(trans, root);
2257
2258                 list_add(&reloc_root->root_list, &reloc_roots);
2259         }
2260
2261         list_splice(&reloc_roots, &rc->reloc_roots);
2262
2263         if (!err)
2264                 btrfs_commit_transaction(trans, rc->extent_root);
2265         else
2266                 btrfs_end_transaction(trans, rc->extent_root);
2267         return err;
2268 }
2269
2270 static noinline_for_stack
2271 void free_reloc_roots(struct list_head *list)
2272 {
2273         struct btrfs_root *reloc_root;
2274
2275         while (!list_empty(list)) {
2276                 reloc_root = list_entry(list->next, struct btrfs_root,
2277                                         root_list);
2278                 __update_reloc_root(reloc_root, 1);
2279                 free_extent_buffer(reloc_root->node);
2280                 free_extent_buffer(reloc_root->commit_root);
2281                 kfree(reloc_root);
2282         }
2283 }
2284
2285 static noinline_for_stack
2286 int merge_reloc_roots(struct reloc_control *rc)
2287 {
2288         struct btrfs_trans_handle *trans;
2289         struct btrfs_root *root;
2290         struct btrfs_root *reloc_root;
2291         u64 last_snap;
2292         u64 otransid;
2293         u64 objectid;
2294         LIST_HEAD(reloc_roots);
2295         int found = 0;
2296         int ret = 0;
2297 again:
2298         root = rc->extent_root;
2299
2300         /*
2301          * this serializes us with btrfs_record_root_in_transaction,
2302          * we have to make sure nobody is in the middle of
2303          * adding their roots to the list while we are
2304          * doing this splice
2305          */
2306         mutex_lock(&root->fs_info->reloc_mutex);
2307         list_splice_init(&rc->reloc_roots, &reloc_roots);
2308         mutex_unlock(&root->fs_info->reloc_mutex);
2309
2310         while (!list_empty(&reloc_roots)) {
2311                 found = 1;
2312                 reloc_root = list_entry(reloc_roots.next,
2313                                         struct btrfs_root, root_list);
2314
2315                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2316                         root = read_fs_root(reloc_root->fs_info,
2317                                             reloc_root->root_key.offset);
2318                         BUG_ON(IS_ERR(root));
2319                         BUG_ON(root->reloc_root != reloc_root);
2320
2321                         ret = merge_reloc_root(rc, root);
2322                         if (ret) {
2323                                 __update_reloc_root(reloc_root, 1);
2324                                 free_extent_buffer(reloc_root->node);
2325                                 free_extent_buffer(reloc_root->commit_root);
2326                                 kfree(reloc_root);
2327                                 goto out;
2328                         }
2329                 } else {
2330                         list_del_init(&reloc_root->root_list);
2331                 }
2332
2333                 /*
2334                  * we keep the old last snapshod transid in rtranid when we
2335                  * created the relocation tree.
2336                  */
2337                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2338                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2339                 objectid = reloc_root->root_key.offset;
2340
2341                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2342                 if (ret < 0) {
2343                         if (list_empty(&reloc_root->root_list))
2344                                 list_add_tail(&reloc_root->root_list,
2345                                               &reloc_roots);
2346                         goto out;
2347                 } else if (!ret) {
2348                         /*
2349                          * recover the last snapshot tranid to avoid
2350                          * the space balance break NOCOW.
2351                          */
2352                         root = read_fs_root(rc->extent_root->fs_info,
2353                                             objectid);
2354                         if (IS_ERR(root))
2355                                 continue;
2356
2357                         trans = btrfs_join_transaction(root);
2358                         BUG_ON(IS_ERR(trans));
2359
2360                         /* Check if the fs/file tree was snapshoted or not. */
2361                         if (btrfs_root_last_snapshot(&root->root_item) ==
2362                             otransid - 1)
2363                                 btrfs_set_root_last_snapshot(&root->root_item,
2364                                                              last_snap);
2365
2366                         btrfs_end_transaction(trans, root);
2367                 }
2368         }
2369
2370         if (found) {
2371                 found = 0;
2372                 goto again;
2373         }
2374 out:
2375         if (ret) {
2376                 btrfs_std_error(root->fs_info, ret);
2377                 if (!list_empty(&reloc_roots))
2378                         free_reloc_roots(&reloc_roots);
2379         }
2380
2381         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2382         return ret;
2383 }
2384
2385 static void free_block_list(struct rb_root *blocks)
2386 {
2387         struct tree_block *block;
2388         struct rb_node *rb_node;
2389         while ((rb_node = rb_first(blocks))) {
2390                 block = rb_entry(rb_node, struct tree_block, rb_node);
2391                 rb_erase(rb_node, blocks);
2392                 kfree(block);
2393         }
2394 }
2395
2396 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2397                                       struct btrfs_root *reloc_root)
2398 {
2399         struct btrfs_root *root;
2400
2401         if (reloc_root->last_trans == trans->transid)
2402                 return 0;
2403
2404         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2405         BUG_ON(IS_ERR(root));
2406         BUG_ON(root->reloc_root != reloc_root);
2407
2408         return btrfs_record_root_in_trans(trans, root);
2409 }
2410
2411 static noinline_for_stack
2412 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2413                                      struct reloc_control *rc,
2414                                      struct backref_node *node,
2415                                      struct backref_edge *edges[], int *nr)
2416 {
2417         struct backref_node *next;
2418         struct btrfs_root *root;
2419         int index = 0;
2420
2421         next = node;
2422         while (1) {
2423                 cond_resched();
2424                 next = walk_up_backref(next, edges, &index);
2425                 root = next->root;
2426                 BUG_ON(!root);
2427                 BUG_ON(!root->ref_cows);
2428
2429                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2430                         record_reloc_root_in_trans(trans, root);
2431                         break;
2432                 }
2433
2434                 btrfs_record_root_in_trans(trans, root);
2435                 root = root->reloc_root;
2436
2437                 if (next->new_bytenr != root->node->start) {
2438                         BUG_ON(next->new_bytenr);
2439                         BUG_ON(!list_empty(&next->list));
2440                         next->new_bytenr = root->node->start;
2441                         next->root = root;
2442                         list_add_tail(&next->list,
2443                                       &rc->backref_cache.changed);
2444                         __mark_block_processed(rc, next);
2445                         break;
2446                 }
2447
2448                 WARN_ON(1);
2449                 root = NULL;
2450                 next = walk_down_backref(edges, &index);
2451                 if (!next || next->level <= node->level)
2452                         break;
2453         }
2454         if (!root)
2455                 return NULL;
2456
2457         *nr = index;
2458         next = node;
2459         /* setup backref node path for btrfs_reloc_cow_block */
2460         while (1) {
2461                 rc->backref_cache.path[next->level] = next;
2462                 if (--index < 0)
2463                         break;
2464                 next = edges[index]->node[UPPER];
2465         }
2466         return root;
2467 }
2468
2469 /*
2470  * select a tree root for relocation. return NULL if the block
2471  * is reference counted. we should use do_relocation() in this
2472  * case. return a tree root pointer if the block isn't reference
2473  * counted. return -ENOENT if the block is root of reloc tree.
2474  */
2475 static noinline_for_stack
2476 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2477                                    struct backref_node *node)
2478 {
2479         struct backref_node *next;
2480         struct btrfs_root *root;
2481         struct btrfs_root *fs_root = NULL;
2482         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2483         int index = 0;
2484
2485         next = node;
2486         while (1) {
2487                 cond_resched();
2488                 next = walk_up_backref(next, edges, &index);
2489                 root = next->root;
2490                 BUG_ON(!root);
2491
2492                 /* no other choice for non-references counted tree */
2493                 if (!root->ref_cows)
2494                         return root;
2495
2496                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2497                         fs_root = root;
2498
2499                 if (next != node)
2500                         return NULL;
2501
2502                 next = walk_down_backref(edges, &index);
2503                 if (!next || next->level <= node->level)
2504                         break;
2505         }
2506
2507         if (!fs_root)
2508                 return ERR_PTR(-ENOENT);
2509         return fs_root;
2510 }
2511
2512 static noinline_for_stack
2513 u64 calcu_metadata_size(struct reloc_control *rc,
2514                         struct backref_node *node, int reserve)
2515 {
2516         struct backref_node *next = node;
2517         struct backref_edge *edge;
2518         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2519         u64 num_bytes = 0;
2520         int index = 0;
2521
2522         BUG_ON(reserve && node->processed);
2523
2524         while (next) {
2525                 cond_resched();
2526                 while (1) {
2527                         if (next->processed && (reserve || next != node))
2528                                 break;
2529
2530                         num_bytes += btrfs_level_size(rc->extent_root,
2531                                                       next->level);
2532
2533                         if (list_empty(&next->upper))
2534                                 break;
2535
2536                         edge = list_entry(next->upper.next,
2537                                           struct backref_edge, list[LOWER]);
2538                         edges[index++] = edge;
2539                         next = edge->node[UPPER];
2540                 }
2541                 next = walk_down_backref(edges, &index);
2542         }
2543         return num_bytes;
2544 }
2545
2546 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2547                                   struct reloc_control *rc,
2548                                   struct backref_node *node)
2549 {
2550         struct btrfs_root *root = rc->extent_root;
2551         u64 num_bytes;
2552         int ret;
2553
2554         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2555
2556         trans->block_rsv = rc->block_rsv;
2557         ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2558                                   BTRFS_RESERVE_FLUSH_ALL);
2559         if (ret) {
2560                 if (ret == -EAGAIN)
2561                         rc->commit_transaction = 1;
2562                 return ret;
2563         }
2564
2565         return 0;
2566 }
2567
2568 static void release_metadata_space(struct reloc_control *rc,
2569                                    struct backref_node *node)
2570 {
2571         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2572         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2573 }
2574
2575 /*
2576  * relocate a block tree, and then update pointers in upper level
2577  * blocks that reference the block to point to the new location.
2578  *
2579  * if called by link_to_upper, the block has already been relocated.
2580  * in that case this function just updates pointers.
2581  */
2582 static int do_relocation(struct btrfs_trans_handle *trans,
2583                          struct reloc_control *rc,
2584                          struct backref_node *node,
2585                          struct btrfs_key *key,
2586                          struct btrfs_path *path, int lowest)
2587 {
2588         struct backref_node *upper;
2589         struct backref_edge *edge;
2590         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2591         struct btrfs_root *root;
2592         struct extent_buffer *eb;
2593         u32 blocksize;
2594         u64 bytenr;
2595         u64 generation;
2596         int nr;
2597         int slot;
2598         int ret;
2599         int err = 0;
2600
2601         BUG_ON(lowest && node->eb);
2602
2603         path->lowest_level = node->level + 1;
2604         rc->backref_cache.path[node->level] = node;
2605         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2606                 cond_resched();
2607
2608                 upper = edge->node[UPPER];
2609                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2610                 BUG_ON(!root);
2611
2612                 if (upper->eb && !upper->locked) {
2613                         if (!lowest) {
2614                                 ret = btrfs_bin_search(upper->eb, key,
2615                                                        upper->level, &slot);
2616                                 BUG_ON(ret);
2617                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2618                                 if (node->eb->start == bytenr)
2619                                         goto next;
2620                         }
2621                         drop_node_buffer(upper);
2622                 }
2623
2624                 if (!upper->eb) {
2625                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2626                         if (ret < 0) {
2627                                 err = ret;
2628                                 break;
2629                         }
2630                         BUG_ON(ret > 0);
2631
2632                         if (!upper->eb) {
2633                                 upper->eb = path->nodes[upper->level];
2634                                 path->nodes[upper->level] = NULL;
2635                         } else {
2636                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2637                         }
2638
2639                         upper->locked = 1;
2640                         path->locks[upper->level] = 0;
2641
2642                         slot = path->slots[upper->level];
2643                         btrfs_release_path(path);
2644                 } else {
2645                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2646                                                &slot);
2647                         BUG_ON(ret);
2648                 }
2649
2650                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2651                 if (lowest) {
2652                         BUG_ON(bytenr != node->bytenr);
2653                 } else {
2654                         if (node->eb->start == bytenr)
2655                                 goto next;
2656                 }
2657
2658                 blocksize = btrfs_level_size(root, node->level);
2659                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2660                 eb = read_tree_block(root, bytenr, blocksize, generation);
2661                 if (!eb || !extent_buffer_uptodate(eb)) {
2662                         free_extent_buffer(eb);
2663                         err = -EIO;
2664                         goto next;
2665                 }
2666                 btrfs_tree_lock(eb);
2667                 btrfs_set_lock_blocking(eb);
2668
2669                 if (!node->eb) {
2670                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2671                                               slot, &eb);
2672                         btrfs_tree_unlock(eb);
2673                         free_extent_buffer(eb);
2674                         if (ret < 0) {
2675                                 err = ret;
2676                                 goto next;
2677                         }
2678                         BUG_ON(node->eb != eb);
2679                 } else {
2680                         btrfs_set_node_blockptr(upper->eb, slot,
2681                                                 node->eb->start);
2682                         btrfs_set_node_ptr_generation(upper->eb, slot,
2683                                                       trans->transid);
2684                         btrfs_mark_buffer_dirty(upper->eb);
2685
2686                         ret = btrfs_inc_extent_ref(trans, root,
2687                                                 node->eb->start, blocksize,
2688                                                 upper->eb->start,
2689                                                 btrfs_header_owner(upper->eb),
2690                                                 node->level, 0, 1);
2691                         BUG_ON(ret);
2692
2693                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2694                         BUG_ON(ret);
2695                 }
2696 next:
2697                 if (!upper->pending)
2698                         drop_node_buffer(upper);
2699                 else
2700                         unlock_node_buffer(upper);
2701                 if (err)
2702                         break;
2703         }
2704
2705         if (!err && node->pending) {
2706                 drop_node_buffer(node);
2707                 list_move_tail(&node->list, &rc->backref_cache.changed);
2708                 node->pending = 0;
2709         }
2710
2711         path->lowest_level = 0;
2712         BUG_ON(err == -ENOSPC);
2713         return err;
2714 }
2715
2716 static int link_to_upper(struct btrfs_trans_handle *trans,
2717                          struct reloc_control *rc,
2718                          struct backref_node *node,
2719                          struct btrfs_path *path)
2720 {
2721         struct btrfs_key key;
2722
2723         btrfs_node_key_to_cpu(node->eb, &key, 0);
2724         return do_relocation(trans, rc, node, &key, path, 0);
2725 }
2726
2727 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2728                                 struct reloc_control *rc,
2729                                 struct btrfs_path *path, int err)
2730 {
2731         LIST_HEAD(list);
2732         struct backref_cache *cache = &rc->backref_cache;
2733         struct backref_node *node;
2734         int level;
2735         int ret;
2736
2737         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2738                 while (!list_empty(&cache->pending[level])) {
2739                         node = list_entry(cache->pending[level].next,
2740                                           struct backref_node, list);
2741                         list_move_tail(&node->list, &list);
2742                         BUG_ON(!node->pending);
2743
2744                         if (!err) {
2745                                 ret = link_to_upper(trans, rc, node, path);
2746                                 if (ret < 0)
2747                                         err = ret;
2748                         }
2749                 }
2750                 list_splice_init(&list, &cache->pending[level]);
2751         }
2752         return err;
2753 }
2754
2755 static void mark_block_processed(struct reloc_control *rc,
2756                                  u64 bytenr, u32 blocksize)
2757 {
2758         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2759                         EXTENT_DIRTY, GFP_NOFS);
2760 }
2761
2762 static void __mark_block_processed(struct reloc_control *rc,
2763                                    struct backref_node *node)
2764 {
2765         u32 blocksize;
2766         if (node->level == 0 ||
2767             in_block_group(node->bytenr, rc->block_group)) {
2768                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2769                 mark_block_processed(rc, node->bytenr, blocksize);
2770         }
2771         node->processed = 1;
2772 }
2773
2774 /*
2775  * mark a block and all blocks directly/indirectly reference the block
2776  * as processed.
2777  */
2778 static void update_processed_blocks(struct reloc_control *rc,
2779                                     struct backref_node *node)
2780 {
2781         struct backref_node *next = node;
2782         struct backref_edge *edge;
2783         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2784         int index = 0;
2785
2786         while (next) {
2787                 cond_resched();
2788                 while (1) {
2789                         if (next->processed)
2790                                 break;
2791
2792                         __mark_block_processed(rc, next);
2793
2794                         if (list_empty(&next->upper))
2795                                 break;
2796
2797                         edge = list_entry(next->upper.next,
2798                                           struct backref_edge, list[LOWER]);
2799                         edges[index++] = edge;
2800                         next = edge->node[UPPER];
2801                 }
2802                 next = walk_down_backref(edges, &index);
2803         }
2804 }
2805
2806 static int tree_block_processed(u64 bytenr, u32 blocksize,
2807                                 struct reloc_control *rc)
2808 {
2809         if (test_range_bit(&rc->processed_blocks, bytenr,
2810                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2811                 return 1;
2812         return 0;
2813 }
2814
2815 static int get_tree_block_key(struct reloc_control *rc,
2816                               struct tree_block *block)
2817 {
2818         struct extent_buffer *eb;
2819
2820         BUG_ON(block->key_ready);
2821         eb = read_tree_block(rc->extent_root, block->bytenr,
2822                              block->key.objectid, block->key.offset);
2823         if (!eb || !extent_buffer_uptodate(eb)) {
2824                 free_extent_buffer(eb);
2825                 return -EIO;
2826         }
2827         WARN_ON(btrfs_header_level(eb) != block->level);
2828         if (block->level == 0)
2829                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2830         else
2831                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2832         free_extent_buffer(eb);
2833         block->key_ready = 1;
2834         return 0;
2835 }
2836
2837 static int reada_tree_block(struct reloc_control *rc,
2838                             struct tree_block *block)
2839 {
2840         BUG_ON(block->key_ready);
2841         if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2842                 readahead_tree_block(rc->extent_root, block->bytenr,
2843                                      block->key.objectid,
2844                                      rc->extent_root->leafsize);
2845         else
2846                 readahead_tree_block(rc->extent_root, block->bytenr,
2847                                      block->key.objectid, block->key.offset);
2848         return 0;
2849 }
2850
2851 /*
2852  * helper function to relocate a tree block
2853  */
2854 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2855                                 struct reloc_control *rc,
2856                                 struct backref_node *node,
2857                                 struct btrfs_key *key,
2858                                 struct btrfs_path *path)
2859 {
2860         struct btrfs_root *root;
2861         int release = 0;
2862         int ret = 0;
2863
2864         if (!node)
2865                 return 0;
2866
2867         BUG_ON(node->processed);
2868         root = select_one_root(trans, node);
2869         if (root == ERR_PTR(-ENOENT)) {
2870                 update_processed_blocks(rc, node);
2871                 goto out;
2872         }
2873
2874         if (!root || root->ref_cows) {
2875                 ret = reserve_metadata_space(trans, rc, node);
2876                 if (ret)
2877                         goto out;
2878                 release = 1;
2879         }
2880
2881         if (root) {
2882                 if (root->ref_cows) {
2883                         BUG_ON(node->new_bytenr);
2884                         BUG_ON(!list_empty(&node->list));
2885                         btrfs_record_root_in_trans(trans, root);
2886                         root = root->reloc_root;
2887                         node->new_bytenr = root->node->start;
2888                         node->root = root;
2889                         list_add_tail(&node->list, &rc->backref_cache.changed);
2890                 } else {
2891                         path->lowest_level = node->level;
2892                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2893                         btrfs_release_path(path);
2894                         if (ret > 0)
2895                                 ret = 0;
2896                 }
2897                 if (!ret)
2898                         update_processed_blocks(rc, node);
2899         } else {
2900                 ret = do_relocation(trans, rc, node, key, path, 1);
2901         }
2902 out:
2903         if (ret || node->level == 0 || node->cowonly) {
2904                 if (release)
2905                         release_metadata_space(rc, node);
2906                 remove_backref_node(&rc->backref_cache, node);
2907         }
2908         return ret;
2909 }
2910
2911 /*
2912  * relocate a list of blocks
2913  */
2914 static noinline_for_stack
2915 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2916                          struct reloc_control *rc, struct rb_root *blocks)
2917 {
2918         struct backref_node *node;
2919         struct btrfs_path *path;
2920         struct tree_block *block;
2921         struct rb_node *rb_node;
2922         int ret;
2923         int err = 0;
2924
2925         path = btrfs_alloc_path();
2926         if (!path) {
2927                 err = -ENOMEM;
2928                 goto out_free_blocks;
2929         }
2930
2931         rb_node = rb_first(blocks);
2932         while (rb_node) {
2933                 block = rb_entry(rb_node, struct tree_block, rb_node);
2934                 if (!block->key_ready)
2935                         reada_tree_block(rc, block);
2936                 rb_node = rb_next(rb_node);
2937         }
2938
2939         rb_node = rb_first(blocks);
2940         while (rb_node) {
2941                 block = rb_entry(rb_node, struct tree_block, rb_node);
2942                 if (!block->key_ready) {
2943                         err = get_tree_block_key(rc, block);
2944                         if (err)
2945                                 goto out_free_path;
2946                 }
2947                 rb_node = rb_next(rb_node);
2948         }
2949
2950         rb_node = rb_first(blocks);
2951         while (rb_node) {
2952                 block = rb_entry(rb_node, struct tree_block, rb_node);
2953
2954                 node = build_backref_tree(rc, &block->key,
2955                                           block->level, block->bytenr);
2956                 if (IS_ERR(node)) {
2957                         err = PTR_ERR(node);
2958                         goto out;
2959                 }
2960
2961                 ret = relocate_tree_block(trans, rc, node, &block->key,
2962                                           path);
2963                 if (ret < 0) {
2964                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2965                                 err = ret;
2966                         goto out;
2967                 }
2968                 rb_node = rb_next(rb_node);
2969         }
2970 out:
2971         err = finish_pending_nodes(trans, rc, path, err);
2972
2973 out_free_path:
2974         btrfs_free_path(path);
2975 out_free_blocks:
2976         free_block_list(blocks);
2977         return err;
2978 }
2979
2980 static noinline_for_stack
2981 int prealloc_file_extent_cluster(struct inode *inode,
2982                                  struct file_extent_cluster *cluster)
2983 {
2984         u64 alloc_hint = 0;
2985         u64 start;
2986         u64 end;
2987         u64 offset = BTRFS_I(inode)->index_cnt;
2988         u64 num_bytes;
2989         int nr = 0;
2990         int ret = 0;
2991
2992         BUG_ON(cluster->start != cluster->boundary[0]);
2993         mutex_lock(&inode->i_mutex);
2994
2995         ret = btrfs_check_data_free_space(inode, cluster->end +
2996                                           1 - cluster->start);
2997         if (ret)
2998                 goto out;
2999
3000         while (nr < cluster->nr) {
3001                 start = cluster->boundary[nr] - offset;
3002                 if (nr + 1 < cluster->nr)
3003                         end = cluster->boundary[nr + 1] - 1 - offset;
3004                 else
3005                         end = cluster->end - offset;
3006
3007                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3008                 num_bytes = end + 1 - start;
3009                 ret = btrfs_prealloc_file_range(inode, 0, start,
3010                                                 num_bytes, num_bytes,
3011                                                 end + 1, &alloc_hint);
3012                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3013                 if (ret)
3014                         break;
3015                 nr++;
3016         }
3017         btrfs_free_reserved_data_space(inode, cluster->end +
3018                                        1 - cluster->start);
3019 out:
3020         mutex_unlock(&inode->i_mutex);
3021         return ret;
3022 }
3023
3024 static noinline_for_stack
3025 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3026                          u64 block_start)
3027 {
3028         struct btrfs_root *root = BTRFS_I(inode)->root;
3029         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3030         struct extent_map *em;
3031         int ret = 0;
3032
3033         em = alloc_extent_map();
3034         if (!em)
3035                 return -ENOMEM;
3036
3037         em->start = start;
3038         em->len = end + 1 - start;
3039         em->block_len = em->len;
3040         em->block_start = block_start;
3041         em->bdev = root->fs_info->fs_devices->latest_bdev;
3042         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3043
3044         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3045         while (1) {
3046                 write_lock(&em_tree->lock);
3047                 ret = add_extent_mapping(em_tree, em, 0);
3048                 write_unlock(&em_tree->lock);
3049                 if (ret != -EEXIST) {
3050                         free_extent_map(em);
3051                         break;
3052                 }
3053                 btrfs_drop_extent_cache(inode, start, end, 0);
3054         }
3055         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3056         return ret;
3057 }
3058
3059 static int relocate_file_extent_cluster(struct inode *inode,
3060                                         struct file_extent_cluster *cluster)
3061 {
3062         u64 page_start;
3063         u64 page_end;
3064         u64 offset = BTRFS_I(inode)->index_cnt;
3065         unsigned long index;
3066         unsigned long last_index;
3067         struct page *page;
3068         struct file_ra_state *ra;
3069         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3070         int nr = 0;
3071         int ret = 0;
3072
3073         if (!cluster->nr)
3074                 return 0;
3075
3076         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3077         if (!ra)
3078                 return -ENOMEM;
3079
3080         ret = prealloc_file_extent_cluster(inode, cluster);
3081         if (ret)
3082                 goto out;
3083
3084         file_ra_state_init(ra, inode->i_mapping);
3085
3086         ret = setup_extent_mapping(inode, cluster->start - offset,
3087                                    cluster->end - offset, cluster->start);
3088         if (ret)
3089                 goto out;
3090
3091         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3092         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3093         while (index <= last_index) {
3094                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3095                 if (ret)
3096                         goto out;
3097
3098                 page = find_lock_page(inode->i_mapping, index);
3099                 if (!page) {
3100                         page_cache_sync_readahead(inode->i_mapping,
3101                                                   ra, NULL, index,
3102                                                   last_index + 1 - index);
3103                         page = find_or_create_page(inode->i_mapping, index,
3104                                                    mask);
3105                         if (!page) {
3106                                 btrfs_delalloc_release_metadata(inode,
3107                                                         PAGE_CACHE_SIZE);
3108                                 ret = -ENOMEM;
3109                                 goto out;
3110                         }
3111                 }
3112
3113                 if (PageReadahead(page)) {
3114                         page_cache_async_readahead(inode->i_mapping,
3115                                                    ra, NULL, page, index,
3116                                                    last_index + 1 - index);
3117                 }
3118
3119                 if (!PageUptodate(page)) {
3120                         btrfs_readpage(NULL, page);
3121                         lock_page(page);
3122                         if (!PageUptodate(page)) {
3123                                 unlock_page(page);
3124                                 page_cache_release(page);
3125                                 btrfs_delalloc_release_metadata(inode,
3126                                                         PAGE_CACHE_SIZE);
3127                                 ret = -EIO;
3128                                 goto out;
3129                         }
3130                 }
3131
3132                 page_start = page_offset(page);
3133                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3134
3135                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3136
3137                 set_page_extent_mapped(page);
3138
3139                 if (nr < cluster->nr &&
3140                     page_start + offset == cluster->boundary[nr]) {
3141                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3142                                         page_start, page_end,
3143                                         EXTENT_BOUNDARY, GFP_NOFS);
3144                         nr++;
3145                 }
3146
3147                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3148                 set_page_dirty(page);
3149
3150                 unlock_extent(&BTRFS_I(inode)->io_tree,
3151                               page_start, page_end);
3152                 unlock_page(page);
3153                 page_cache_release(page);
3154
3155                 index++;
3156                 balance_dirty_pages_ratelimited(inode->i_mapping);
3157                 btrfs_throttle(BTRFS_I(inode)->root);
3158         }
3159         WARN_ON(nr != cluster->nr);
3160 out:
3161         kfree(ra);
3162         return ret;
3163 }
3164
3165 static noinline_for_stack
3166 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3167                          struct file_extent_cluster *cluster)
3168 {
3169         int ret;
3170
3171         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3172                 ret = relocate_file_extent_cluster(inode, cluster);
3173                 if (ret)
3174                         return ret;
3175                 cluster->nr = 0;
3176         }
3177
3178         if (!cluster->nr)
3179                 cluster->start = extent_key->objectid;
3180         else
3181                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3182         cluster->end = extent_key->objectid + extent_key->offset - 1;
3183         cluster->boundary[cluster->nr] = extent_key->objectid;
3184         cluster->nr++;
3185
3186         if (cluster->nr >= MAX_EXTENTS) {
3187                 ret = relocate_file_extent_cluster(inode, cluster);
3188                 if (ret)
3189                         return ret;
3190                 cluster->nr = 0;
3191         }
3192         return 0;
3193 }
3194
3195 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3196 static int get_ref_objectid_v0(struct reloc_control *rc,
3197                                struct btrfs_path *path,
3198                                struct btrfs_key *extent_key,
3199                                u64 *ref_objectid, int *path_change)
3200 {
3201         struct btrfs_key key;
3202         struct extent_buffer *leaf;
3203         struct btrfs_extent_ref_v0 *ref0;
3204         int ret;
3205         int slot;
3206
3207         leaf = path->nodes[0];
3208         slot = path->slots[0];
3209         while (1) {
3210                 if (slot >= btrfs_header_nritems(leaf)) {
3211                         ret = btrfs_next_leaf(rc->extent_root, path);
3212                         if (ret < 0)
3213                                 return ret;
3214                         BUG_ON(ret > 0);
3215                         leaf = path->nodes[0];
3216                         slot = path->slots[0];
3217                         if (path_change)
3218                                 *path_change = 1;
3219                 }
3220                 btrfs_item_key_to_cpu(leaf, &key, slot);
3221                 if (key.objectid != extent_key->objectid)
3222                         return -ENOENT;
3223
3224                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3225                         slot++;
3226                         continue;
3227                 }
3228                 ref0 = btrfs_item_ptr(leaf, slot,
3229                                 struct btrfs_extent_ref_v0);
3230                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3231                 break;
3232         }
3233         return 0;
3234 }
3235 #endif
3236
3237 /*
3238  * helper to add a tree block to the list.
3239  * the major work is getting the generation and level of the block
3240  */
3241 static int add_tree_block(struct reloc_control *rc,
3242                           struct btrfs_key *extent_key,
3243                           struct btrfs_path *path,
3244                           struct rb_root *blocks)
3245 {
3246         struct extent_buffer *eb;
3247         struct btrfs_extent_item *ei;
3248         struct btrfs_tree_block_info *bi;
3249         struct tree_block *block;
3250         struct rb_node *rb_node;
3251         u32 item_size;
3252         int level = -1;
3253         int generation;
3254
3255         eb =  path->nodes[0];
3256         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3257
3258         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3259             item_size >= sizeof(*ei) + sizeof(*bi)) {
3260                 ei = btrfs_item_ptr(eb, path->slots[0],
3261                                 struct btrfs_extent_item);
3262                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3263                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3264                         level = btrfs_tree_block_level(eb, bi);
3265                 } else {
3266                         level = (int)extent_key->offset;
3267                 }
3268                 generation = btrfs_extent_generation(eb, ei);
3269         } else {
3270 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3271                 u64 ref_owner;
3272                 int ret;
3273
3274                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3275                 ret = get_ref_objectid_v0(rc, path, extent_key,
3276                                           &ref_owner, NULL);
3277                 if (ret < 0)
3278                         return ret;
3279                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3280                 level = (int)ref_owner;
3281                 /* FIXME: get real generation */
3282                 generation = 0;
3283 #else
3284                 BUG();
3285 #endif
3286         }
3287
3288         btrfs_release_path(path);
3289
3290         BUG_ON(level == -1);
3291
3292         block = kmalloc(sizeof(*block), GFP_NOFS);
3293         if (!block)
3294                 return -ENOMEM;
3295
3296         block->bytenr = extent_key->objectid;
3297         block->key.objectid = rc->extent_root->leafsize;
3298         block->key.offset = generation;
3299         block->level = level;
3300         block->key_ready = 0;
3301
3302         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3303         if (rb_node)
3304                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3305
3306         return 0;
3307 }
3308
3309 /*
3310  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3311  */
3312 static int __add_tree_block(struct reloc_control *rc,
3313                             u64 bytenr, u32 blocksize,
3314                             struct rb_root *blocks)
3315 {
3316         struct btrfs_path *path;
3317         struct btrfs_key key;
3318         int ret;
3319         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3320                                         SKINNY_METADATA);
3321
3322         if (tree_block_processed(bytenr, blocksize, rc))
3323                 return 0;
3324
3325         if (tree_search(blocks, bytenr))
3326                 return 0;
3327
3328         path = btrfs_alloc_path();
3329         if (!path)
3330                 return -ENOMEM;
3331 again:
3332         key.objectid = bytenr;
3333         if (skinny) {
3334                 key.type = BTRFS_METADATA_ITEM_KEY;
3335                 key.offset = (u64)-1;
3336         } else {
3337                 key.type = BTRFS_EXTENT_ITEM_KEY;
3338                 key.offset = blocksize;
3339         }
3340
3341         path->search_commit_root = 1;
3342         path->skip_locking = 1;
3343         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3344         if (ret < 0)
3345                 goto out;
3346
3347         if (ret > 0 && skinny) {
3348                 if (path->slots[0]) {
3349                         path->slots[0]--;
3350                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3351                                               path->slots[0]);
3352                         if (key.objectid == bytenr &&
3353                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3354                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3355                               key.offset == blocksize)))
3356                                 ret = 0;
3357                 }
3358
3359                 if (ret) {
3360                         skinny = false;
3361                         btrfs_release_path(path);
3362                         goto again;
3363                 }
3364         }
3365         BUG_ON(ret);
3366
3367         ret = add_tree_block(rc, &key, path, blocks);
3368 out:
3369         btrfs_free_path(path);
3370         return ret;
3371 }
3372
3373 /*
3374  * helper to check if the block use full backrefs for pointers in it
3375  */
3376 static int block_use_full_backref(struct reloc_control *rc,
3377                                   struct extent_buffer *eb)
3378 {
3379         u64 flags;
3380         int ret;
3381
3382         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3383             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3384                 return 1;
3385
3386         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3387                                        eb->start, btrfs_header_level(eb), 1,
3388                                        NULL, &flags);
3389         BUG_ON(ret);
3390
3391         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3392                 ret = 1;
3393         else
3394                 ret = 0;
3395         return ret;
3396 }
3397
3398 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3399                                     struct inode *inode, u64 ino)
3400 {
3401         struct btrfs_key key;
3402         struct btrfs_path *path;
3403         struct btrfs_root *root = fs_info->tree_root;
3404         struct btrfs_trans_handle *trans;
3405         int ret = 0;
3406
3407         if (inode)
3408                 goto truncate;
3409
3410         key.objectid = ino;
3411         key.type = BTRFS_INODE_ITEM_KEY;
3412         key.offset = 0;
3413
3414         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3415         if (IS_ERR(inode) || is_bad_inode(inode)) {
3416                 if (!IS_ERR(inode))
3417                         iput(inode);
3418                 return -ENOENT;
3419         }
3420
3421 truncate:
3422         ret = btrfs_check_trunc_cache_free_space(root,
3423                                                  &fs_info->global_block_rsv);
3424         if (ret)
3425                 goto out;
3426
3427         path = btrfs_alloc_path();
3428         if (!path) {
3429                 ret = -ENOMEM;
3430                 goto out;
3431         }
3432
3433         trans = btrfs_join_transaction(root);
3434         if (IS_ERR(trans)) {
3435                 btrfs_free_path(path);
3436                 ret = PTR_ERR(trans);
3437                 goto out;
3438         }
3439
3440         ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3441
3442         btrfs_free_path(path);
3443         btrfs_end_transaction(trans, root);
3444         btrfs_btree_balance_dirty(root);
3445 out:
3446         iput(inode);
3447         return ret;
3448 }
3449
3450 /*
3451  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3452  * this function scans fs tree to find blocks reference the data extent
3453  */
3454 static int find_data_references(struct reloc_control *rc,
3455                                 struct btrfs_key *extent_key,
3456                                 struct extent_buffer *leaf,
3457                                 struct btrfs_extent_data_ref *ref,
3458                                 struct rb_root *blocks)
3459 {
3460         struct btrfs_path *path;
3461         struct tree_block *block;
3462         struct btrfs_root *root;
3463         struct btrfs_file_extent_item *fi;
3464         struct rb_node *rb_node;
3465         struct btrfs_key key;
3466         u64 ref_root;
3467         u64 ref_objectid;
3468         u64 ref_offset;
3469         u32 ref_count;
3470         u32 nritems;
3471         int err = 0;
3472         int added = 0;
3473         int counted;
3474         int ret;
3475
3476         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3477         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3478         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3479         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3480
3481         /*
3482          * This is an extent belonging to the free space cache, lets just delete
3483          * it and redo the search.
3484          */
3485         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3486                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3487                                                NULL, ref_objectid);
3488                 if (ret != -ENOENT)
3489                         return ret;
3490                 ret = 0;
3491         }
3492
3493         path = btrfs_alloc_path();
3494         if (!path)
3495                 return -ENOMEM;
3496         path->reada = 1;
3497
3498         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3499         if (IS_ERR(root)) {
3500                 err = PTR_ERR(root);
3501                 goto out;
3502         }
3503
3504         key.objectid = ref_objectid;
3505         key.type = BTRFS_EXTENT_DATA_KEY;
3506         if (ref_offset > ((u64)-1 << 32))
3507                 key.offset = 0;
3508         else
3509                 key.offset = ref_offset;
3510
3511         path->search_commit_root = 1;
3512         path->skip_locking = 1;
3513         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3514         if (ret < 0) {
3515                 err = ret;
3516                 goto out;
3517         }
3518
3519         leaf = path->nodes[0];
3520         nritems = btrfs_header_nritems(leaf);
3521         /*
3522          * the references in tree blocks that use full backrefs
3523          * are not counted in
3524          */
3525         if (block_use_full_backref(rc, leaf))
3526                 counted = 0;
3527         else
3528                 counted = 1;
3529         rb_node = tree_search(blocks, leaf->start);
3530         if (rb_node) {
3531                 if (counted)
3532                         added = 1;
3533                 else
3534                         path->slots[0] = nritems;
3535         }
3536
3537         while (ref_count > 0) {
3538                 while (path->slots[0] >= nritems) {
3539                         ret = btrfs_next_leaf(root, path);
3540                         if (ret < 0) {
3541                                 err = ret;
3542                                 goto out;
3543                         }
3544                         if (ret > 0) {
3545                                 WARN_ON(1);
3546                                 goto out;
3547                         }
3548
3549                         leaf = path->nodes[0];
3550                         nritems = btrfs_header_nritems(leaf);
3551                         added = 0;
3552
3553                         if (block_use_full_backref(rc, leaf))
3554                                 counted = 0;
3555                         else
3556                                 counted = 1;
3557                         rb_node = tree_search(blocks, leaf->start);
3558                         if (rb_node) {
3559                                 if (counted)
3560                                         added = 1;
3561                                 else
3562                                         path->slots[0] = nritems;
3563                         }
3564                 }
3565
3566                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3567                 if (key.objectid != ref_objectid ||
3568                     key.type != BTRFS_EXTENT_DATA_KEY) {
3569                         WARN_ON(1);
3570                         break;
3571                 }
3572
3573                 fi = btrfs_item_ptr(leaf, path->slots[0],
3574                                     struct btrfs_file_extent_item);
3575
3576                 if (btrfs_file_extent_type(leaf, fi) ==
3577                     BTRFS_FILE_EXTENT_INLINE)
3578                         goto next;
3579
3580                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3581                     extent_key->objectid)
3582                         goto next;
3583
3584                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3585                 if (key.offset != ref_offset)
3586                         goto next;
3587
3588                 if (counted)
3589                         ref_count--;
3590                 if (added)
3591                         goto next;
3592
3593                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3594                         block = kmalloc(sizeof(*block), GFP_NOFS);
3595                         if (!block) {
3596                                 err = -ENOMEM;
3597                                 break;
3598                         }
3599                         block->bytenr = leaf->start;
3600                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3601                         block->level = 0;
3602                         block->key_ready = 1;
3603                         rb_node = tree_insert(blocks, block->bytenr,
3604                                               &block->rb_node);
3605                         if (rb_node)
3606                                 backref_tree_panic(rb_node, -EEXIST,
3607                                                    block->bytenr);
3608                 }
3609                 if (counted)
3610                         added = 1;
3611                 else
3612                         path->slots[0] = nritems;
3613 next:
3614                 path->slots[0]++;
3615
3616         }
3617 out:
3618         btrfs_free_path(path);
3619         return err;
3620 }
3621
3622 /*
3623  * helper to find all tree blocks that reference a given data extent
3624  */
3625 static noinline_for_stack
3626 int add_data_references(struct reloc_control *rc,
3627                         struct btrfs_key *extent_key,
3628                         struct btrfs_path *path,
3629                         struct rb_root *blocks)
3630 {
3631         struct btrfs_key key;
3632         struct extent_buffer *eb;
3633         struct btrfs_extent_data_ref *dref;
3634         struct btrfs_extent_inline_ref *iref;
3635         unsigned long ptr;
3636         unsigned long end;
3637         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3638         int ret = 0;
3639         int err = 0;
3640
3641         eb = path->nodes[0];
3642         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3643         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3644 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3645         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3646                 ptr = end;
3647         else
3648 #endif
3649                 ptr += sizeof(struct btrfs_extent_item);
3650
3651         while (ptr < end) {
3652                 iref = (struct btrfs_extent_inline_ref *)ptr;
3653                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3654                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3655                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3656                         ret = __add_tree_block(rc, key.offset, blocksize,
3657                                                blocks);
3658                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3659                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3660                         ret = find_data_references(rc, extent_key,
3661                                                    eb, dref, blocks);
3662                 } else {
3663                         BUG();
3664                 }
3665                 if (ret) {
3666                         err = ret;
3667                         goto out;
3668                 }
3669                 ptr += btrfs_extent_inline_ref_size(key.type);
3670         }
3671         WARN_ON(ptr > end);
3672
3673         while (1) {
3674                 cond_resched();
3675                 eb = path->nodes[0];
3676                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3677                         ret = btrfs_next_leaf(rc->extent_root, path);
3678                         if (ret < 0) {
3679                                 err = ret;
3680                                 break;
3681                         }
3682                         if (ret > 0)
3683                                 break;
3684                         eb = path->nodes[0];
3685                 }
3686
3687                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3688                 if (key.objectid != extent_key->objectid)
3689                         break;
3690
3691 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3692                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3693                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3694 #else
3695                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3696                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3697 #endif
3698                         ret = __add_tree_block(rc, key.offset, blocksize,
3699                                                blocks);
3700                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3701                         dref = btrfs_item_ptr(eb, path->slots[0],
3702                                               struct btrfs_extent_data_ref);
3703                         ret = find_data_references(rc, extent_key,
3704                                                    eb, dref, blocks);
3705                 } else {
3706                         ret = 0;
3707                 }
3708                 if (ret) {
3709                         err = ret;
3710                         break;
3711                 }
3712                 path->slots[0]++;
3713         }
3714 out:
3715         btrfs_release_path(path);
3716         if (err)
3717                 free_block_list(blocks);
3718         return err;
3719 }
3720
3721 /*
3722  * helper to find next unprocessed extent
3723  */
3724 static noinline_for_stack
3725 int find_next_extent(struct btrfs_trans_handle *trans,
3726                      struct reloc_control *rc, struct btrfs_path *path,
3727                      struct btrfs_key *extent_key)
3728 {
3729         struct btrfs_key key;
3730         struct extent_buffer *leaf;
3731         u64 start, end, last;
3732         int ret;
3733
3734         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3735         while (1) {
3736                 cond_resched();
3737                 if (rc->search_start >= last) {
3738                         ret = 1;
3739                         break;
3740                 }
3741
3742                 key.objectid = rc->search_start;
3743                 key.type = BTRFS_EXTENT_ITEM_KEY;
3744                 key.offset = 0;
3745
3746                 path->search_commit_root = 1;
3747                 path->skip_locking = 1;
3748                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3749                                         0, 0);
3750                 if (ret < 0)
3751                         break;
3752 next:
3753                 leaf = path->nodes[0];
3754                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3755                         ret = btrfs_next_leaf(rc->extent_root, path);
3756                         if (ret != 0)
3757                                 break;
3758                         leaf = path->nodes[0];
3759                 }
3760
3761                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3762                 if (key.objectid >= last) {
3763                         ret = 1;
3764                         break;
3765                 }
3766
3767                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3768                     key.type != BTRFS_METADATA_ITEM_KEY) {
3769                         path->slots[0]++;
3770                         goto next;
3771                 }
3772
3773                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3774                     key.objectid + key.offset <= rc->search_start) {
3775                         path->slots[0]++;
3776                         goto next;
3777                 }
3778
3779                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3780                     key.objectid + rc->extent_root->leafsize <=
3781                     rc->search_start) {
3782                         path->slots[0]++;
3783                         goto next;
3784                 }
3785
3786                 ret = find_first_extent_bit(&rc->processed_blocks,
3787                                             key.objectid, &start, &end,
3788                                             EXTENT_DIRTY, NULL);
3789
3790                 if (ret == 0 && start <= key.objectid) {
3791                         btrfs_release_path(path);
3792                         rc->search_start = end + 1;
3793                 } else {
3794                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3795                                 rc->search_start = key.objectid + key.offset;
3796                         else
3797                                 rc->search_start = key.objectid +
3798                                         rc->extent_root->leafsize;
3799                         memcpy(extent_key, &key, sizeof(key));
3800                         return 0;
3801                 }
3802         }
3803         btrfs_release_path(path);
3804         return ret;
3805 }
3806
3807 static void set_reloc_control(struct reloc_control *rc)
3808 {
3809         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3810
3811         mutex_lock(&fs_info->reloc_mutex);
3812         fs_info->reloc_ctl = rc;
3813         mutex_unlock(&fs_info->reloc_mutex);
3814 }
3815
3816 static void unset_reloc_control(struct reloc_control *rc)
3817 {
3818         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3819
3820         mutex_lock(&fs_info->reloc_mutex);
3821         fs_info->reloc_ctl = NULL;
3822         mutex_unlock(&fs_info->reloc_mutex);
3823 }
3824
3825 static int check_extent_flags(u64 flags)
3826 {
3827         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3828             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3829                 return 1;
3830         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3831             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3832                 return 1;
3833         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3834             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3835                 return 1;
3836         return 0;
3837 }
3838
3839 static noinline_for_stack
3840 int prepare_to_relocate(struct reloc_control *rc)
3841 {
3842         struct btrfs_trans_handle *trans;
3843         int ret;
3844
3845         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3846                                               BTRFS_BLOCK_RSV_TEMP);
3847         if (!rc->block_rsv)
3848                 return -ENOMEM;
3849
3850         /*
3851          * reserve some space for creating reloc trees.
3852          * btrfs_init_reloc_root will use them when there
3853          * is no reservation in transaction handle.
3854          */
3855         ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3856                                   rc->extent_root->nodesize * 256,
3857                                   BTRFS_RESERVE_FLUSH_ALL);
3858         if (ret)
3859                 return ret;
3860
3861         memset(&rc->cluster, 0, sizeof(rc->cluster));
3862         rc->search_start = rc->block_group->key.objectid;
3863         rc->extents_found = 0;
3864         rc->nodes_relocated = 0;
3865         rc->merging_rsv_size = 0;
3866
3867         rc->create_reloc_tree = 1;
3868         set_reloc_control(rc);
3869
3870         trans = btrfs_join_transaction(rc->extent_root);
3871         if (IS_ERR(trans)) {
3872                 unset_reloc_control(rc);
3873                 /*
3874                  * extent tree is not a ref_cow tree and has no reloc_root to
3875                  * cleanup.  And callers are responsible to free the above
3876                  * block rsv.
3877                  */
3878                 return PTR_ERR(trans);
3879         }
3880         btrfs_commit_transaction(trans, rc->extent_root);
3881         return 0;
3882 }
3883
3884 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3885 {
3886         struct rb_root blocks = RB_ROOT;
3887         struct btrfs_key key;
3888         struct btrfs_trans_handle *trans = NULL;
3889         struct btrfs_path *path;
3890         struct btrfs_extent_item *ei;
3891         u64 flags;
3892         u32 item_size;
3893         int ret;
3894         int err = 0;
3895         int progress = 0;
3896
3897         path = btrfs_alloc_path();
3898         if (!path)
3899                 return -ENOMEM;
3900         path->reada = 1;
3901
3902         ret = prepare_to_relocate(rc);
3903         if (ret) {
3904                 err = ret;
3905                 goto out_free;
3906         }
3907
3908         while (1) {
3909                 progress++;
3910                 trans = btrfs_start_transaction(rc->extent_root, 0);
3911                 if (IS_ERR(trans)) {
3912                         err = PTR_ERR(trans);
3913                         trans = NULL;
3914                         break;
3915                 }
3916 restart:
3917                 if (update_backref_cache(trans, &rc->backref_cache)) {
3918                         btrfs_end_transaction(trans, rc->extent_root);
3919                         continue;
3920                 }
3921
3922                 ret = find_next_extent(trans, rc, path, &key);
3923                 if (ret < 0)
3924                         err = ret;
3925                 if (ret != 0)
3926                         break;
3927
3928                 rc->extents_found++;
3929
3930                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3931                                     struct btrfs_extent_item);
3932                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3933                 if (item_size >= sizeof(*ei)) {
3934                         flags = btrfs_extent_flags(path->nodes[0], ei);
3935                         ret = check_extent_flags(flags);
3936                         BUG_ON(ret);
3937
3938                 } else {
3939 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3940                         u64 ref_owner;
3941                         int path_change = 0;
3942
3943                         BUG_ON(item_size !=
3944                                sizeof(struct btrfs_extent_item_v0));
3945                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3946                                                   &path_change);
3947                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3948                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3949                         else
3950                                 flags = BTRFS_EXTENT_FLAG_DATA;
3951
3952                         if (path_change) {
3953                                 btrfs_release_path(path);
3954
3955                                 path->search_commit_root = 1;
3956                                 path->skip_locking = 1;
3957                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3958                                                         &key, path, 0, 0);
3959                                 if (ret < 0) {
3960                                         err = ret;
3961                                         break;
3962                                 }
3963                                 BUG_ON(ret > 0);
3964                         }
3965 #else
3966                         BUG();
3967 #endif
3968                 }
3969
3970                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3971                         ret = add_tree_block(rc, &key, path, &blocks);
3972                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3973                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3974                         ret = add_data_references(rc, &key, path, &blocks);
3975                 } else {
3976                         btrfs_release_path(path);
3977                         ret = 0;
3978                 }
3979                 if (ret < 0) {
3980                         err = ret;
3981                         break;
3982                 }
3983
3984                 if (!RB_EMPTY_ROOT(&blocks)) {
3985                         ret = relocate_tree_blocks(trans, rc, &blocks);
3986                         if (ret < 0) {
3987                                 if (ret != -EAGAIN) {
3988                                         err = ret;
3989                                         break;
3990                                 }
3991                                 rc->extents_found--;
3992                                 rc->search_start = key.objectid;
3993                         }
3994                 }
3995
3996                 ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5);
3997                 if (ret < 0) {
3998                         if (ret != -ENOSPC) {
3999                                 err = ret;
4000                                 WARN_ON(1);
4001                                 break;
4002                         }
4003                         rc->commit_transaction = 1;
4004                 }
4005
4006                 if (rc->commit_transaction) {
4007                         rc->commit_transaction = 0;
4008                         ret = btrfs_commit_transaction(trans, rc->extent_root);
4009                         BUG_ON(ret);
4010                 } else {
4011                         btrfs_end_transaction_throttle(trans, rc->extent_root);
4012                         btrfs_btree_balance_dirty(rc->extent_root);
4013                 }
4014                 trans = NULL;
4015
4016                 if (rc->stage == MOVE_DATA_EXTENTS &&
4017                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4018                         rc->found_file_extent = 1;
4019                         ret = relocate_data_extent(rc->data_inode,
4020                                                    &key, &rc->cluster);
4021                         if (ret < 0) {
4022                                 err = ret;
4023                                 break;
4024                         }
4025                 }
4026         }
4027         if (trans && progress && err == -ENOSPC) {
4028                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4029                                               rc->block_group->flags);
4030                 if (ret == 0) {
4031                         err = 0;
4032                         progress = 0;
4033                         goto restart;
4034                 }
4035         }
4036
4037         btrfs_release_path(path);
4038         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4039                           GFP_NOFS);
4040
4041         if (trans) {
4042                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4043                 btrfs_btree_balance_dirty(rc->extent_root);
4044         }
4045
4046         if (!err) {
4047                 ret = relocate_file_extent_cluster(rc->data_inode,
4048                                                    &rc->cluster);
4049                 if (ret < 0)
4050                         err = ret;
4051         }
4052
4053         rc->create_reloc_tree = 0;
4054         set_reloc_control(rc);
4055
4056         backref_cache_cleanup(&rc->backref_cache);
4057         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4058
4059         err = prepare_to_merge(rc, err);
4060
4061         merge_reloc_roots(rc);
4062
4063         rc->merge_reloc_tree = 0;
4064         unset_reloc_control(rc);
4065         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4066
4067         /* get rid of pinned extents */
4068         trans = btrfs_join_transaction(rc->extent_root);
4069         if (IS_ERR(trans))
4070                 err = PTR_ERR(trans);
4071         else
4072                 btrfs_commit_transaction(trans, rc->extent_root);
4073 out_free:
4074         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4075         btrfs_free_path(path);
4076         return err;
4077 }
4078
4079 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4080                                  struct btrfs_root *root, u64 objectid)
4081 {
4082         struct btrfs_path *path;
4083         struct btrfs_inode_item *item;
4084         struct extent_buffer *leaf;
4085         int ret;
4086
4087         path = btrfs_alloc_path();
4088         if (!path)
4089                 return -ENOMEM;
4090
4091         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4092         if (ret)
4093                 goto out;
4094
4095         leaf = path->nodes[0];
4096         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4097         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4098         btrfs_set_inode_generation(leaf, item, 1);
4099         btrfs_set_inode_size(leaf, item, 0);
4100         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4101         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4102                                           BTRFS_INODE_PREALLOC);
4103         btrfs_mark_buffer_dirty(leaf);
4104         btrfs_release_path(path);
4105 out:
4106         btrfs_free_path(path);
4107         return ret;
4108 }
4109
4110 /*
4111  * helper to create inode for data relocation.
4112  * the inode is in data relocation tree and its link count is 0
4113  */
4114 static noinline_for_stack
4115 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4116                                  struct btrfs_block_group_cache *group)
4117 {
4118         struct inode *inode = NULL;
4119         struct btrfs_trans_handle *trans;
4120         struct btrfs_root *root;
4121         struct btrfs_key key;
4122         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4123         int err = 0;
4124
4125         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4126         if (IS_ERR(root))
4127                 return ERR_CAST(root);
4128
4129         trans = btrfs_start_transaction(root, 6);
4130         if (IS_ERR(trans))
4131                 return ERR_CAST(trans);
4132
4133         err = btrfs_find_free_objectid(root, &objectid);
4134         if (err)
4135                 goto out;
4136
4137         err = __insert_orphan_inode(trans, root, objectid);
4138         BUG_ON(err);
4139
4140         key.objectid = objectid;
4141         key.type = BTRFS_INODE_ITEM_KEY;
4142         key.offset = 0;
4143         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4144         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4145         BTRFS_I(inode)->index_cnt = group->key.objectid;
4146
4147         err = btrfs_orphan_add(trans, inode);
4148 out:
4149         btrfs_end_transaction(trans, root);
4150         btrfs_btree_balance_dirty(root);
4151         if (err) {
4152                 if (inode)
4153                         iput(inode);
4154                 inode = ERR_PTR(err);
4155         }
4156         return inode;
4157 }
4158
4159 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4160 {
4161         struct reloc_control *rc;
4162
4163         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4164         if (!rc)
4165                 return NULL;
4166
4167         INIT_LIST_HEAD(&rc->reloc_roots);
4168         backref_cache_init(&rc->backref_cache);
4169         mapping_tree_init(&rc->reloc_root_tree);
4170         extent_io_tree_init(&rc->processed_blocks,
4171                             fs_info->btree_inode->i_mapping);
4172         return rc;
4173 }
4174
4175 /*
4176  * function to relocate all extents in a block group.
4177  */
4178 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4179 {
4180         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4181         struct reloc_control *rc;
4182         struct inode *inode;
4183         struct btrfs_path *path;
4184         int ret;
4185         int rw = 0;
4186         int err = 0;
4187
4188         rc = alloc_reloc_control(fs_info);
4189         if (!rc)
4190                 return -ENOMEM;
4191
4192         rc->extent_root = extent_root;
4193
4194         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4195         BUG_ON(!rc->block_group);
4196
4197         if (!rc->block_group->ro) {
4198                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4199                 if (ret) {
4200                         err = ret;
4201                         goto out;
4202                 }
4203                 rw = 1;
4204         }
4205
4206         path = btrfs_alloc_path();
4207         if (!path) {
4208                 err = -ENOMEM;
4209                 goto out;
4210         }
4211
4212         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4213                                         path);
4214         btrfs_free_path(path);
4215
4216         if (!IS_ERR(inode))
4217                 ret = delete_block_group_cache(fs_info, inode, 0);
4218         else
4219                 ret = PTR_ERR(inode);
4220
4221         if (ret && ret != -ENOENT) {
4222                 err = ret;
4223                 goto out;
4224         }
4225
4226         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4227         if (IS_ERR(rc->data_inode)) {
4228                 err = PTR_ERR(rc->data_inode);
4229                 rc->data_inode = NULL;
4230                 goto out;
4231         }
4232
4233         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4234                rc->block_group->key.objectid, rc->block_group->flags);
4235
4236         ret = btrfs_start_all_delalloc_inodes(fs_info, 0);
4237         if (ret < 0) {
4238                 err = ret;
4239                 goto out;
4240         }
4241         btrfs_wait_all_ordered_extents(fs_info, 0);
4242
4243         while (1) {
4244                 mutex_lock(&fs_info->cleaner_mutex);
4245                 ret = relocate_block_group(rc);
4246                 mutex_unlock(&fs_info->cleaner_mutex);
4247                 if (ret < 0) {
4248                         err = ret;
4249                         goto out;
4250                 }
4251
4252                 if (rc->extents_found == 0)
4253                         break;
4254
4255                 printk(KERN_INFO "btrfs: found %llu extents\n",
4256                         rc->extents_found);
4257
4258                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4259                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4260                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4261                                                  0, -1);
4262                         rc->stage = UPDATE_DATA_PTRS;
4263                 }
4264         }
4265
4266         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4267                                      rc->block_group->key.objectid,
4268                                      rc->block_group->key.objectid +
4269                                      rc->block_group->key.offset - 1);
4270
4271         WARN_ON(rc->block_group->pinned > 0);
4272         WARN_ON(rc->block_group->reserved > 0);
4273         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4274 out:
4275         if (err && rw)
4276                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4277         iput(rc->data_inode);
4278         btrfs_put_block_group(rc->block_group);
4279         kfree(rc);
4280         return err;
4281 }
4282
4283 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4284 {
4285         struct btrfs_trans_handle *trans;
4286         int ret, err;
4287
4288         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4289         if (IS_ERR(trans))
4290                 return PTR_ERR(trans);
4291
4292         memset(&root->root_item.drop_progress, 0,
4293                 sizeof(root->root_item.drop_progress));
4294         root->root_item.drop_level = 0;
4295         btrfs_set_root_refs(&root->root_item, 0);
4296         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4297                                 &root->root_key, &root->root_item);
4298
4299         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4300         if (err)
4301                 return err;
4302         return ret;
4303 }
4304
4305 /*
4306  * recover relocation interrupted by system crash.
4307  *
4308  * this function resumes merging reloc trees with corresponding fs trees.
4309  * this is important for keeping the sharing of tree blocks
4310  */
4311 int btrfs_recover_relocation(struct btrfs_root *root)
4312 {
4313         LIST_HEAD(reloc_roots);
4314         struct btrfs_key key;
4315         struct btrfs_root *fs_root;
4316         struct btrfs_root *reloc_root;
4317         struct btrfs_path *path;
4318         struct extent_buffer *leaf;
4319         struct reloc_control *rc = NULL;
4320         struct btrfs_trans_handle *trans;
4321         int ret;
4322         int err = 0;
4323
4324         path = btrfs_alloc_path();
4325         if (!path)
4326                 return -ENOMEM;
4327         path->reada = -1;
4328
4329         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4330         key.type = BTRFS_ROOT_ITEM_KEY;
4331         key.offset = (u64)-1;
4332
4333         while (1) {
4334                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4335                                         path, 0, 0);
4336                 if (ret < 0) {
4337                         err = ret;
4338                         goto out;
4339                 }
4340                 if (ret > 0) {
4341                         if (path->slots[0] == 0)
4342                                 break;
4343                         path->slots[0]--;
4344                 }
4345                 leaf = path->nodes[0];
4346                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4347                 btrfs_release_path(path);
4348
4349                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4350                     key.type != BTRFS_ROOT_ITEM_KEY)
4351                         break;
4352
4353                 reloc_root = btrfs_read_fs_root(root, &key);
4354                 if (IS_ERR(reloc_root)) {
4355                         err = PTR_ERR(reloc_root);
4356                         goto out;
4357                 }
4358
4359                 list_add(&reloc_root->root_list, &reloc_roots);
4360
4361                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4362                         fs_root = read_fs_root(root->fs_info,
4363                                                reloc_root->root_key.offset);
4364                         if (IS_ERR(fs_root)) {
4365                                 ret = PTR_ERR(fs_root);
4366                                 if (ret != -ENOENT) {
4367                                         err = ret;
4368                                         goto out;
4369                                 }
4370                                 ret = mark_garbage_root(reloc_root);
4371                                 if (ret < 0) {
4372                                         err = ret;
4373                                         goto out;
4374                                 }
4375                         }
4376                 }
4377
4378                 if (key.offset == 0)
4379                         break;
4380
4381                 key.offset--;
4382         }
4383         btrfs_release_path(path);
4384
4385         if (list_empty(&reloc_roots))
4386                 goto out;
4387
4388         rc = alloc_reloc_control(root->fs_info);
4389         if (!rc) {
4390                 err = -ENOMEM;
4391                 goto out;
4392         }
4393
4394         rc->extent_root = root->fs_info->extent_root;
4395
4396         set_reloc_control(rc);
4397
4398         trans = btrfs_join_transaction(rc->extent_root);
4399         if (IS_ERR(trans)) {
4400                 unset_reloc_control(rc);
4401                 err = PTR_ERR(trans);
4402                 goto out_free;
4403         }
4404
4405         rc->merge_reloc_tree = 1;
4406
4407         while (!list_empty(&reloc_roots)) {
4408                 reloc_root = list_entry(reloc_roots.next,
4409                                         struct btrfs_root, root_list);
4410                 list_del(&reloc_root->root_list);
4411
4412                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4413                         list_add_tail(&reloc_root->root_list,
4414                                       &rc->reloc_roots);
4415                         continue;
4416                 }
4417
4418                 fs_root = read_fs_root(root->fs_info,
4419                                        reloc_root->root_key.offset);
4420                 if (IS_ERR(fs_root)) {
4421                         err = PTR_ERR(fs_root);
4422                         goto out_free;
4423                 }
4424
4425                 err = __add_reloc_root(reloc_root);
4426                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4427                 fs_root->reloc_root = reloc_root;
4428         }
4429
4430         err = btrfs_commit_transaction(trans, rc->extent_root);
4431         if (err)
4432                 goto out_free;
4433
4434         merge_reloc_roots(rc);
4435
4436         unset_reloc_control(rc);
4437
4438         trans = btrfs_join_transaction(rc->extent_root);
4439         if (IS_ERR(trans))
4440                 err = PTR_ERR(trans);
4441         else
4442                 err = btrfs_commit_transaction(trans, rc->extent_root);
4443 out_free:
4444         kfree(rc);
4445 out:
4446         if (!list_empty(&reloc_roots))
4447                 free_reloc_roots(&reloc_roots);
4448
4449         btrfs_free_path(path);
4450
4451         if (err == 0) {
4452                 /* cleanup orphan inode in data relocation tree */
4453                 fs_root = read_fs_root(root->fs_info,
4454                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4455                 if (IS_ERR(fs_root))
4456                         err = PTR_ERR(fs_root);
4457                 else
4458                         err = btrfs_orphan_cleanup(fs_root);
4459         }
4460         return err;
4461 }
4462
4463 /*
4464  * helper to add ordered checksum for data relocation.
4465  *
4466  * cloning checksum properly handles the nodatasum extents.
4467  * it also saves CPU time to re-calculate the checksum.
4468  */
4469 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4470 {
4471         struct btrfs_ordered_sum *sums;
4472         struct btrfs_ordered_extent *ordered;
4473         struct btrfs_root *root = BTRFS_I(inode)->root;
4474         int ret;
4475         u64 disk_bytenr;
4476         LIST_HEAD(list);
4477
4478         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4479         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4480
4481         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4482         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4483                                        disk_bytenr + len - 1, &list, 0);
4484         if (ret)
4485                 goto out;
4486
4487         disk_bytenr = ordered->start;
4488         while (!list_empty(&list)) {
4489                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4490                 list_del_init(&sums->list);
4491
4492                 sums->bytenr = disk_bytenr;
4493                 disk_bytenr += sums->len;
4494
4495                 btrfs_add_ordered_sum(inode, ordered, sums);
4496         }
4497 out:
4498         btrfs_put_ordered_extent(ordered);
4499         return ret;
4500 }
4501
4502 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4503                            struct btrfs_root *root, struct extent_buffer *buf,
4504                            struct extent_buffer *cow)
4505 {
4506         struct reloc_control *rc;
4507         struct backref_node *node;
4508         int first_cow = 0;
4509         int level;
4510         int ret;
4511
4512         rc = root->fs_info->reloc_ctl;
4513         if (!rc)
4514                 return;
4515
4516         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4517                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4518
4519         level = btrfs_header_level(buf);
4520         if (btrfs_header_generation(buf) <=
4521             btrfs_root_last_snapshot(&root->root_item))
4522                 first_cow = 1;
4523
4524         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4525             rc->create_reloc_tree) {
4526                 WARN_ON(!first_cow && level == 0);
4527
4528                 node = rc->backref_cache.path[level];
4529                 BUG_ON(node->bytenr != buf->start &&
4530                        node->new_bytenr != buf->start);
4531
4532                 drop_node_buffer(node);
4533                 extent_buffer_get(cow);
4534                 node->eb = cow;
4535                 node->new_bytenr = cow->start;
4536
4537                 if (!node->pending) {
4538                         list_move_tail(&node->list,
4539                                        &rc->backref_cache.pending[level]);
4540                         node->pending = 1;
4541                 }
4542
4543                 if (first_cow)
4544                         __mark_block_processed(rc, node);
4545
4546                 if (first_cow && level > 0)
4547                         rc->nodes_relocated += buf->len;
4548         }
4549
4550         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4551                 ret = replace_file_extents(trans, rc, root, cow);
4552                 BUG_ON(ret);
4553         }
4554 }
4555
4556 /*
4557  * called before creating snapshot. it calculates metadata reservation
4558  * requried for relocating tree blocks in the snapshot
4559  */
4560 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4561                               struct btrfs_pending_snapshot *pending,
4562                               u64 *bytes_to_reserve)
4563 {
4564         struct btrfs_root *root;
4565         struct reloc_control *rc;
4566
4567         root = pending->root;
4568         if (!root->reloc_root)
4569                 return;
4570
4571         rc = root->fs_info->reloc_ctl;
4572         if (!rc->merge_reloc_tree)
4573                 return;
4574
4575         root = root->reloc_root;
4576         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4577         /*
4578          * relocation is in the stage of merging trees. the space
4579          * used by merging a reloc tree is twice the size of
4580          * relocated tree nodes in the worst case. half for cowing
4581          * the reloc tree, half for cowing the fs tree. the space
4582          * used by cowing the reloc tree will be freed after the
4583          * tree is dropped. if we create snapshot, cowing the fs
4584          * tree may use more space than it frees. so we need
4585          * reserve extra space.
4586          */
4587         *bytes_to_reserve += rc->nodes_relocated;
4588 }
4589
4590 /*
4591  * called after snapshot is created. migrate block reservation
4592  * and create reloc root for the newly created snapshot
4593  */
4594 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4595                                struct btrfs_pending_snapshot *pending)
4596 {
4597         struct btrfs_root *root = pending->root;
4598         struct btrfs_root *reloc_root;
4599         struct btrfs_root *new_root;
4600         struct reloc_control *rc;
4601         int ret;
4602
4603         if (!root->reloc_root)
4604                 return 0;
4605
4606         rc = root->fs_info->reloc_ctl;
4607         rc->merging_rsv_size += rc->nodes_relocated;
4608
4609         if (rc->merge_reloc_tree) {
4610                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4611                                               rc->block_rsv,
4612                                               rc->nodes_relocated);
4613                 if (ret)
4614                         return ret;
4615         }
4616
4617         new_root = pending->snap;
4618         reloc_root = create_reloc_root(trans, root->reloc_root,
4619                                        new_root->root_key.objectid);
4620         if (IS_ERR(reloc_root))
4621                 return PTR_ERR(reloc_root);
4622
4623         ret = __add_reloc_root(reloc_root);
4624         BUG_ON(ret < 0);
4625         new_root->reloc_root = reloc_root;
4626
4627         if (rc->create_reloc_tree)
4628                 ret = clone_backref_node(trans, rc, root, reloc_root);
4629         return ret;
4630 }