fs/btrfs/relocation.c

   1 /*
   2  * Copyright (C) 2009 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/rbtree.h>
  24 #include <linux/slab.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "volumes.h"
  29 #include "locking.h"
  30 #include "btrfs_inode.h"
  31 #include "async-thread.h"
  32 #include "free-space-cache.h"
  33 #include "inode-map.h"
  34
  35 /*
  36  * backref_node, mapping_node and tree_block start with this
  37  */
  38 struct tree_entry {
  39         struct rb_node rb_node;
  40         u64 bytenr;
  41 };
  42
  43 /*
  44  * present a tree block in the backref cache
  45  */
  46 struct backref_node {
  47         struct rb_node rb_node;
  48         u64 bytenr;
  49
  50         u64 new_bytenr;
  51         /* objectid of tree block owner, can be not uptodate */
  52         u64 owner;
  53         /* link to pending, changed or detached list */
  54         struct list_head list;
  55         /* list of upper level blocks reference this block */
  56         struct list_head upper;
  57         /* list of child blocks in the cache */
  58         struct list_head lower;
  59         /* NULL if this node is not tree root */
  60         struct btrfs_root *root;
  61         /* extent buffer got by COW the block */
  62         struct extent_buffer *eb;
  63         /* level of tree block */
  64         unsigned int level:8;
  65         /* is the block in non-reference counted tree */
  66         unsigned int cowonly:1;
  67         /* 1 if no child node in the cache */
  68         unsigned int lowest:1;
  69         /* is the extent buffer locked */
  70         unsigned int locked:1;
  71         /* has the block been processed */
  72         unsigned int processed:1;
  73         /* have backrefs of this block been checked */
  74         unsigned int checked:1;
  75         /*
  76          * 1 if corresponding block has been cowed but some upper
  77          * level block pointers may not point to the new location
  78          */
  79         unsigned int pending:1;
  80         /*
  81          * 1 if the backref node isn't connected to any other
  82          * backref node.
  83          */
  84         unsigned int detached:1;
  85 };
  86
  87 /*
  88  * present a block pointer in the backref cache
  89  */
  90 struct backref_edge {
  91         struct list_head list[2];
  92         struct backref_node *node[2];
  93 };
  94
  95 #define LOWER   0
  96 #define UPPER   1
  97
  98 struct backref_cache {
  99         /* red black tree of all backref nodes in the cache */
 100         struct rb_root rb_root;
 101         /* for passing backref nodes to btrfs_reloc_cow_block */
 102         struct backref_node *path[BTRFS_MAX_LEVEL];
 103         /*
 104          * list of blocks that have been cowed but some block
 105          * pointers in upper level blocks may not reflect the
 106          * new location
 107          */
 108         struct list_head pending[BTRFS_MAX_LEVEL];
 109         /* list of backref nodes with no child node */
 110         struct list_head leaves;
 111         /* list of blocks that have been cowed in current transaction */
 112         struct list_head changed;
 113         /* list of detached backref node. */
 114         struct list_head detached;
 115
 116         u64 last_trans;
 117
 118         int nr_nodes;
 119         int nr_edges;
 120 };
 121
 122 /*
 123  * map address of tree root to tree
 124  */
 125 struct mapping_node {
 126         struct rb_node rb_node;
 127         u64 bytenr;
 128         void *data;
 129 };
 130
 131 struct mapping_tree {
 132         struct rb_root rb_root;
 133         spinlock_t lock;
 134 };
 135
 136 /*
 137  * present a tree block to process
 138  */
 139 struct tree_block {
 140         struct rb_node rb_node;
 141         u64 bytenr;
 142         struct btrfs_key key;
 143         unsigned int level:8;
 144         unsigned int key_ready:1;
 145 };
 146
 147 #define MAX_EXTENTS 128
 148
 149 struct file_extent_cluster {
 150         u64 start;
 151         u64 end;
 152         u64 boundary[MAX_EXTENTS];
 153         unsigned int nr;
 154 };
 155
 156 struct reloc_control {
 157         /* block group to relocate */
 158         struct btrfs_block_group_cache *block_group;
 159         /* extent tree */
 160         struct btrfs_root *extent_root;
 161         /* inode for moving data */
 162         struct inode *data_inode;
 163
 164         struct btrfs_block_rsv *block_rsv;
 165
 166         struct backref_cache backref_cache;
 167
 168         struct file_extent_cluster cluster;
 169         /* tree blocks have been processed */
 170         struct extent_io_tree processed_blocks;
 171         /* map start of tree root to corresponding reloc tree */
 172         struct mapping_tree reloc_root_tree;
 173         /* list of reloc trees */
 174         struct list_head reloc_roots;
 175         /* size of metadata reservation for merging reloc trees */
 176         u64 merging_rsv_size;
 177         /* size of relocated tree nodes */
 178         u64 nodes_relocated;
 179
 180         u64 search_start;
 181         u64 extents_found;
 182
 183         unsigned int stage:8;
 184         unsigned int create_reloc_tree:1;
 185         unsigned int merge_reloc_tree:1;
 186         unsigned int found_file_extent:1;
 187         unsigned int commit_transaction:1;
 188 };
 189
 190 /* stages of data relocation */
 191 #define MOVE_DATA_EXTENTS       0
 192 #define UPDATE_DATA_PTRS        1
 193
 194 static void remove_backref_node(struct backref_cache *cache,
 195                                 struct backref_node *node);
 196 static void __mark_block_processed(struct reloc_control *rc,
 197                                    struct backref_node *node);
 198
 199 static void mapping_tree_init(struct mapping_tree *tree)
 200 {
 201         tree->rb_root = RB_ROOT;
 202         spin_lock_init(&tree->lock);
 203 }
 204
 205 static void backref_cache_init(struct backref_cache *cache)
 206 {
 207         int i;
 208         cache->rb_root = RB_ROOT;
 209         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 210                 INIT_LIST_HEAD(&cache->pending[i]);
 211         INIT_LIST_HEAD(&cache->changed);
 212         INIT_LIST_HEAD(&cache->detached);
 213         INIT_LIST_HEAD(&cache->leaves);
 214 }
 215
 216 static void backref_cache_cleanup(struct backref_cache *cache)
 217 {
 218         struct backref_node *node;
 219         int i;
 220
 221         while (!list_empty(&cache->detached)) {
 222                 node = list_entry(cache->detached.next,
 223                                   struct backref_node, list);
 224                 remove_backref_node(cache, node);
 225         }
 226
 227         while (!list_empty(&cache->leaves)) {
 228                 node = list_entry(cache->leaves.next,
 229                                   struct backref_node, lower);
 230                 remove_backref_node(cache, node);
 231         }
 232
 233         cache->last_trans = 0;
 234
 235         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 236                 BUG_ON(!list_empty(&cache->pending[i]));
 237         BUG_ON(!list_empty(&cache->changed));
 238         BUG_ON(!list_empty(&cache->detached));
 239         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
 240         BUG_ON(cache->nr_nodes);
 241         BUG_ON(cache->nr_edges);
 242 }
 243
 244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 245 {
 246         struct backref_node *node;
 247
 248         node = kzalloc(sizeof(*node), GFP_NOFS);
 249         if (node) {
 250                 INIT_LIST_HEAD(&node->list);
 251                 INIT_LIST_HEAD(&node->upper);
 252                 INIT_LIST_HEAD(&node->lower);
 253                 RB_CLEAR_NODE(&node->rb_node);
 254                 cache->nr_nodes++;
 255         }
 256         return node;
 257 }
 258
 259 static void free_backref_node(struct backref_cache *cache,
 260                               struct backref_node *node)
 261 {
 262         if (node) {
 263                 cache->nr_nodes--;
 264                 kfree(node);
 265         }
 266 }
 267
 268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 269 {
 270         struct backref_edge *edge;
 271
 272         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 273         if (edge)
 274                 cache->nr_edges++;
 275         return edge;
 276 }
 277
 278 static void free_backref_edge(struct backref_cache *cache,
 279                               struct backref_edge *edge)
 280 {
 281         if (edge) {
 282                 cache->nr_edges--;
 283                 kfree(edge);
 284         }
 285 }
 286
 287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 288                                    struct rb_node *node)
 289 {
 290         struct rb_node **p = &root->rb_node;
 291         struct rb_node *parent = NULL;
 292         struct tree_entry *entry;
 293
 294         while (*p) {
 295                 parent = *p;
 296                 entry = rb_entry(parent, struct tree_entry, rb_node);
 297
 298                 if (bytenr < entry->bytenr)
 299                         p = &(*p)->rb_left;
 300                 else if (bytenr > entry->bytenr)
 301                         p = &(*p)->rb_right;
 302                 else
 303                         return parent;
 304         }
 305
 306         rb_link_node(node, parent, p);
 307         rb_insert_color(node, root);
 308         return NULL;
 309 }
 310
 311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 312 {
 313         struct rb_node *n = root->rb_node;
 314         struct tree_entry *entry;
 315
 316         while (n) {
 317                 entry = rb_entry(n, struct tree_entry, rb_node);
 318
 319                 if (bytenr < entry->bytenr)
 320                         n = n->rb_left;
 321                 else if (bytenr > entry->bytenr)
 322                         n = n->rb_right;
 323                 else
 324                         return n;
 325         }
 326         return NULL;
 327 }
 328
 329 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 330 {
 331
 332         struct btrfs_fs_info *fs_info = NULL;
 333         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 334                                               rb_node);
 335         if (bnode->root)
 336                 fs_info = bnode->root->fs_info;
 337         btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
 338                     "found at offset %llu\n", bytenr);
 339 }
 340
 341 /*
 342  * walk up backref nodes until reach node presents tree root
 343  */
 344 static struct backref_node *walk_up_backref(struct backref_node *node,
 345                                             struct backref_edge *edges[],
 346                                             int *index)
 347 {
 348         struct backref_edge *edge;
 349         int idx = *index;
 350
 351         while (!list_empty(&node->upper)) {
 352                 edge = list_entry(node->upper.next,
 353                                   struct backref_edge, list[LOWER]);
 354                 edges[idx++] = edge;
 355                 node = edge->node[UPPER];
 356         }
 357         BUG_ON(node->detached);
 358         *index = idx;
 359         return node;
 360 }
 361
 362 /*
 363  * walk down backref nodes to find start of next reference path
 364  */
 365 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 366                                               int *index)
 367 {
 368         struct backref_edge *edge;
 369         struct backref_node *lower;
 370         int idx = *index;
 371
 372         while (idx > 0) {
 373                 edge = edges[idx - 1];
 374                 lower = edge->node[LOWER];
 375                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 376                         idx--;
 377                         continue;
 378                 }
 379                 edge = list_entry(edge->list[LOWER].next,
 380                                   struct backref_edge, list[LOWER]);
 381                 edges[idx - 1] = edge;
 382                 *index = idx;
 383                 return edge->node[UPPER];
 384         }
 385         *index = 0;
 386         return NULL;
 387 }
 388
 389 static void unlock_node_buffer(struct backref_node *node)
 390 {
 391         if (node->locked) {
 392                 btrfs_tree_unlock(node->eb);
 393                 node->locked = 0;
 394         }
 395 }
 396
 397 static void drop_node_buffer(struct backref_node *node)
 398 {
 399         if (node->eb) {
 400                 unlock_node_buffer(node);
 401                 free_extent_buffer(node->eb);
 402                 node->eb = NULL;
 403         }
 404 }
 405
 406 static void drop_backref_node(struct backref_cache *tree,
 407                               struct backref_node *node)
 408 {
 409         BUG_ON(!list_empty(&node->upper));
 410
 411         drop_node_buffer(node);
 412         list_del(&node->list);
 413         list_del(&node->lower);
 414         if (!RB_EMPTY_NODE(&node->rb_node))
 415                 rb_erase(&node->rb_node, &tree->rb_root);
 416         free_backref_node(tree, node);
 417 }
 418
 419 /*
 420  * remove a backref node from the backref cache
 421  */
 422 static void remove_backref_node(struct backref_cache *cache,
 423                                 struct backref_node *node)
 424 {
 425         struct backref_node *upper;
 426         struct backref_edge *edge;
 427
 428         if (!node)
 429                 return;
 430
 431         BUG_ON(!node->lowest && !node->detached);
 432         while (!list_empty(&node->upper)) {
 433                 edge = list_entry(node->upper.next, struct backref_edge,
 434                                   list[LOWER]);
 435                 upper = edge->node[UPPER];
 436                 list_del(&edge->list[LOWER]);
 437                 list_del(&edge->list[UPPER]);
 438                 free_backref_edge(cache, edge);
 439
 440                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 441                         BUG_ON(!list_empty(&node->upper));
 442                         drop_backref_node(cache, node);
 443                         node = upper;
 444                         node->lowest = 1;
 445                         continue;
 446                 }
 447                 /*
 448                  * add the node to leaf node list if no other
 449                  * child block cached.
 450                  */
 451                 if (list_empty(&upper->lower)) {
 452                         list_add_tail(&upper->lower, &cache->leaves);
 453                         upper->lowest = 1;
 454                 }
 455         }
 456
 457         drop_backref_node(cache, node);
 458 }
 459
 460 static void update_backref_node(struct backref_cache *cache,
 461                                 struct backref_node *node, u64 bytenr)
 462 {
 463         struct rb_node *rb_node;
 464         rb_erase(&node->rb_node, &cache->rb_root);
 465         node->bytenr = bytenr;
 466         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 467         if (rb_node)
 468                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 469 }
 470
 471 /*
 472  * update backref cache after a transaction commit
 473  */
 474 static int update_backref_cache(struct btrfs_trans_handle *trans,
 475                                 struct backref_cache *cache)
 476 {
 477         struct backref_node *node;
 478         int level = 0;
 479
 480         if (cache->last_trans == 0) {
 481                 cache->last_trans = trans->transid;
 482                 return 0;
 483         }
 484
 485         if (cache->last_trans == trans->transid)
 486                 return 0;
 487
 488         /*
 489          * detached nodes are used to avoid unnecessary backref
 490          * lookup. transaction commit changes the extent tree.
 491          * so the detached nodes are no longer useful.
 492          */
 493         while (!list_empty(&cache->detached)) {
 494                 node = list_entry(cache->detached.next,
 495                                   struct backref_node, list);
 496                 remove_backref_node(cache, node);
 497         }
 498
 499         while (!list_empty(&cache->changed)) {
 500                 node = list_entry(cache->changed.next,
 501                                   struct backref_node, list);
 502                 list_del_init(&node->list);
 503                 BUG_ON(node->pending);
 504                 update_backref_node(cache, node, node->new_bytenr);
 505         }
 506
 507         /*
 508          * some nodes can be left in the pending list if there were
 509          * errors during processing the pending nodes.
 510          */
 511         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 512                 list_for_each_entry(node, &cache->pending[level], list) {
 513                         BUG_ON(!node->pending);
 514                         if (node->bytenr == node->new_bytenr)
 515                                 continue;
 516                         update_backref_node(cache, node, node->new_bytenr);
 517                 }
 518         }
 519
 520         cache->last_trans = 0;
 521         return 1;
 522 }
 523
 524
 525 static int should_ignore_root(struct btrfs_root *root)
 526 {
 527         struct btrfs_root *reloc_root;
 528
 529         if (!root->ref_cows)
 530                 return 0;
 531
 532         reloc_root = root->reloc_root;
 533         if (!reloc_root)
 534                 return 0;
 535
 536         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 537             root->fs_info->running_transaction->transid - 1)
 538                 return 0;
 539         /*
 540          * if there is reloc tree and it was created in previous
 541          * transaction backref lookup can find the reloc tree,
 542          * so backref node for the fs tree root is useless for
 543          * relocation.
 544          */
 545         return 1;
 546 }
 547 /*
 548  * find reloc tree by address of tree root
 549  */
 550 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 551                                           u64 bytenr)
 552 {
 553         struct rb_node *rb_node;
 554         struct mapping_node *node;
 555         struct btrfs_root *root = NULL;
 556
 557         spin_lock(&rc->reloc_root_tree.lock);
 558         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 559         if (rb_node) {
 560                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 561                 root = (struct btrfs_root *)node->data;
 562         }
 563         spin_unlock(&rc->reloc_root_tree.lock);
 564         return root;
 565 }
 566
 567 static int is_cowonly_root(u64 root_objectid)
 568 {
 569         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 570             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 571             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 572             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 573             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 574             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
 575                 return 1;
 576         return 0;
 577 }
 578
 579 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 580                                         u64 root_objectid)
 581 {
 582         struct btrfs_key key;
 583
 584         key.objectid = root_objectid;
 585         key.type = BTRFS_ROOT_ITEM_KEY;
 586         if (is_cowonly_root(root_objectid))
 587                 key.offset = 0;
 588         else
 589                 key.offset = (u64)-1;
 590
 591         return btrfs_get_fs_root(fs_info, &key, false);
 592 }
 593
 594 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 595 static noinline_for_stack
 596 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 597                                   struct extent_buffer *leaf,
 598                                   struct btrfs_extent_ref_v0 *ref0)
 599 {
 600         struct btrfs_root *root;
 601         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 602         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 603
 604         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 605
 606         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 607         BUG_ON(IS_ERR(root));
 608
 609         if (root->ref_cows &&
 610             generation != btrfs_root_generation(&root->root_item))
 611                 return NULL;
 612
 613         return root;
 614 }
 615 #endif
 616
 617 static noinline_for_stack
 618 int find_inline_backref(struct extent_buffer *leaf, int slot,
 619                         unsigned long *ptr, unsigned long *end)
 620 {
 621         struct btrfs_key key;
 622         struct btrfs_extent_item *ei;
 623         struct btrfs_tree_block_info *bi;
 624         u32 item_size;
 625
 626         btrfs_item_key_to_cpu(leaf, &key, slot);
 627
 628         item_size = btrfs_item_size_nr(leaf, slot);
 629 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 630         if (item_size < sizeof(*ei)) {
 631                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 632                 return 1;
 633         }
 634 #endif
 635         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 636         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 637                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 638
 639         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 640             item_size <= sizeof(*ei) + sizeof(*bi)) {
 641                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 642                 return 1;
 643         }
 644         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 645             item_size <= sizeof(*ei)) {
 646                 WARN_ON(item_size < sizeof(*ei));
 647                 return 1;
 648         }
 649
 650         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 651                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 652                 *ptr = (unsigned long)(bi + 1);
 653         } else {
 654                 *ptr = (unsigned long)(ei + 1);
 655         }
 656         *end = (unsigned long)ei + item_size;
 657         return 0;
 658 }
 659
 660 /*
 661  * build backref tree for a given tree block. root of the backref tree
 662  * corresponds the tree block, leaves of the backref tree correspond
 663  * roots of b-trees that reference the tree block.
 664  *
 665  * the basic idea of this function is check backrefs of a given block
 666  * to find upper level blocks that refernece the block, and then check
 667  * bakcrefs of these upper level blocks recursively. the recursion stop
 668  * when tree root is reached or backrefs for the block is cached.
 669  *
 670  * NOTE: if we find backrefs for a block are cached, we know backrefs
 671  * for all upper level blocks that directly/indirectly reference the
 672  * block are also cached.
 673  */
 674 static noinline_for_stack
 675 struct backref_node *build_backref_tree(struct reloc_control *rc,
 676                                         struct btrfs_key *node_key,
 677                                         int level, u64 bytenr)
 678 {
 679         struct backref_cache *cache = &rc->backref_cache;
 680         struct btrfs_path *path1;
 681         struct btrfs_path *path2;
 682         struct extent_buffer *eb;
 683         struct btrfs_root *root;
 684         struct backref_node *cur;
 685         struct backref_node *upper;
 686         struct backref_node *lower;
 687         struct backref_node *node = NULL;
 688         struct backref_node *exist = NULL;
 689         struct backref_edge *edge;
 690         struct rb_node *rb_node;
 691         struct btrfs_key key;
 692         unsigned long end;
 693         unsigned long ptr;
 694         LIST_HEAD(list);
 695         LIST_HEAD(useless);
 696         int cowonly;
 697         int ret;
 698         int err = 0;
 699         bool need_check = true;
 700
 701         path1 = btrfs_alloc_path();
 702         path2 = btrfs_alloc_path();
 703         if (!path1 || !path2) {
 704                 err = -ENOMEM;
 705                 goto out;
 706         }
 707         path1->reada = 1;
 708         path2->reada = 2;
 709
 710         node = alloc_backref_node(cache);
 711         if (!node) {
 712                 err = -ENOMEM;
 713                 goto out;
 714         }
 715
 716         node->bytenr = bytenr;
 717         node->level = level;
 718         node->lowest = 1;
 719         cur = node;
 720 again:
 721         end = 0;
 722         ptr = 0;
 723         key.objectid = cur->bytenr;
 724         key.type = BTRFS_METADATA_ITEM_KEY;
 725         key.offset = (u64)-1;
 726
 727         path1->search_commit_root = 1;
 728         path1->skip_locking = 1;
 729         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 730                                 0, 0);
 731         if (ret < 0) {
 732                 err = ret;
 733                 goto out;
 734         }
 735         BUG_ON(!ret || !path1->slots[0]);
 736
 737         path1->slots[0]--;
 738
 739         WARN_ON(cur->checked);
 740         if (!list_empty(&cur->upper)) {
 741                 /*
 742                  * the backref was added previously when processing
 743                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 744                  */
 745                 BUG_ON(!list_is_singular(&cur->upper));
 746                 edge = list_entry(cur->upper.next, struct backref_edge,
 747                                   list[LOWER]);
 748                 BUG_ON(!list_empty(&edge->list[UPPER]));
 749                 exist = edge->node[UPPER];
 750                 /*
 751                  * add the upper level block to pending list if we need
 752                  * check its backrefs
 753                  */
 754                 if (!exist->checked)
 755                         list_add_tail(&edge->list[UPPER], &list);
 756         } else {
 757                 exist = NULL;
 758         }
 759
 760         while (1) {
 761                 cond_resched();
 762                 eb = path1->nodes[0];
 763
 764                 if (ptr >= end) {
 765                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 766                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 767                                 if (ret < 0) {
 768                                         err = ret;
 769                                         goto out;
 770                                 }
 771                                 if (ret > 0)
 772                                         break;
 773                                 eb = path1->nodes[0];
 774                         }
 775
 776                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 777                         if (key.objectid != cur->bytenr) {
 778                                 WARN_ON(exist);
 779                                 break;
 780                         }
 781
 782                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 783                             key.type == BTRFS_METADATA_ITEM_KEY) {
 784                                 ret = find_inline_backref(eb, path1->slots[0],
 785                                                           &ptr, &end);
 786                                 if (ret)
 787                                         goto next;
 788                         }
 789                 }
 790
 791                 if (ptr < end) {
 792                         /* update key for inline back ref */
 793                         struct btrfs_extent_inline_ref *iref;
 794                         iref = (struct btrfs_extent_inline_ref *)ptr;
 795                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 796                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 797                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 798                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 799                 }
 800
 801                 if (exist &&
 802                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 803                       exist->owner == key.offset) ||
 804                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 805                       exist->bytenr == key.offset))) {
 806                         exist = NULL;
 807                         goto next;
 808                 }
 809
 810 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 811                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 812                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 813                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 814                                 struct btrfs_extent_ref_v0 *ref0;
 815                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 816                                                 struct btrfs_extent_ref_v0);
 817                                 if (key.objectid == key.offset) {
 818                                         root = find_tree_root(rc, eb, ref0);
 819                                         if (root && !should_ignore_root(root))
 820                                                 cur->root = root;
 821                                         else
 822                                                 list_add(&cur->list, &useless);
 823                                         break;
 824                                 }
 825                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 826                                                                       ref0)))
 827                                         cur->cowonly = 1;
 828                         }
 829 #else
 830                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 831                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 832 #endif
 833                         if (key.objectid == key.offset) {
 834                                 /*
 835                                  * only root blocks of reloc trees use
 836                                  * backref of this type.
 837                                  */
 838                                 root = find_reloc_root(rc, cur->bytenr);
 839                                 BUG_ON(!root);
 840                                 cur->root = root;
 841                                 break;
 842                         }
 843
 844                         edge = alloc_backref_edge(cache);
 845                         if (!edge) {
 846                                 err = -ENOMEM;
 847                                 goto out;
 848                         }
 849                         rb_node = tree_search(&cache->rb_root, key.offset);
 850                         if (!rb_node) {
 851                                 upper = alloc_backref_node(cache);
 852                                 if (!upper) {
 853                                         free_backref_edge(cache, edge);
 854                                         err = -ENOMEM;
 855                                         goto out;
 856                                 }
 857                                 upper->bytenr = key.offset;
 858                                 upper->level = cur->level + 1;
 859                                 /*
 860                                  *  backrefs for the upper level block isn't
 861                                  *  cached, add the block to pending list
 862                                  */
 863                                 list_add_tail(&edge->list[UPPER], &list);
 864                         } else {
 865                                 upper = rb_entry(rb_node, struct backref_node,
 866                                                  rb_node);
 867                                 BUG_ON(!upper->checked);
 868                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 869                         }
 870                         list_add_tail(&edge->list[LOWER], &cur->upper);
 871                         edge->node[LOWER] = cur;
 872                         edge->node[UPPER] = upper;
 873
 874                         goto next;
 875                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 876                         goto next;
 877                 }
 878
 879                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 880                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 881                 if (IS_ERR(root)) {
 882                         err = PTR_ERR(root);
 883                         goto out;
 884                 }
 885
 886                 if (!root->ref_cows)
 887                         cur->cowonly = 1;
 888
 889                 if (btrfs_root_level(&root->root_item) == cur->level) {
 890                         /* tree root */
 891                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 892                                cur->bytenr);
 893                         if (should_ignore_root(root))
 894                                 list_add(&cur->list, &useless);
 895                         else
 896                                 cur->root = root;
 897                         break;
 898                 }
 899
 900                 level = cur->level + 1;
 901
 902                 /*
 903                  * searching the tree to find upper level blocks
 904                  * reference the block.
 905                  */
 906                 path2->search_commit_root = 1;
 907                 path2->skip_locking = 1;
 908                 path2->lowest_level = level;
 909                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 910                 path2->lowest_level = 0;
 911                 if (ret < 0) {
 912                         err = ret;
 913                         goto out;
 914                 }
 915                 if (ret > 0 && path2->slots[level] > 0)
 916                         path2->slots[level]--;
 917
 918                 eb = path2->nodes[level];
 919                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 920                         cur->bytenr);
 921
 922                 lower = cur;
 923                 need_check = true;
 924                 for (; level < BTRFS_MAX_LEVEL; level++) {
 925                         if (!path2->nodes[level]) {
 926                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 927                                        lower->bytenr);
 928                                 if (should_ignore_root(root))
 929                                         list_add(&lower->list, &useless);
 930                                 else
 931                                         lower->root = root;
 932                                 break;
 933                         }
 934
 935                         edge = alloc_backref_edge(cache);
 936                         if (!edge) {
 937                                 err = -ENOMEM;
 938                                 goto out;
 939                         }
 940
 941                         eb = path2->nodes[level];
 942                         rb_node = tree_search(&cache->rb_root, eb->start);
 943                         if (!rb_node) {
 944                                 upper = alloc_backref_node(cache);
 945                                 if (!upper) {
 946                                         free_backref_edge(cache, edge);
 947                                         err = -ENOMEM;
 948                                         goto out;
 949                                 }
 950                                 upper->bytenr = eb->start;
 951                                 upper->owner = btrfs_header_owner(eb);
 952                                 upper->level = lower->level + 1;
 953                                 if (!root->ref_cows)
 954                                         upper->cowonly = 1;
 955
 956                                 /*
 957                                  * if we know the block isn't shared
 958                                  * we can void checking its backrefs.
 959                                  */
 960                                 if (btrfs_block_can_be_shared(root, eb))
 961                                         upper->checked = 0;
 962                                 else
 963                                         upper->checked = 1;
 964
 965                                 /*
 966                                  * add the block to pending list if we
 967                                  * need check its backrefs, we only do this once
 968                                  * while walking up a tree as we will catch
 969                                  * anything else later on.
 970                                  */
 971                                 if (!upper->checked && need_check) {
 972                                         need_check = false;
 973                                         list_add_tail(&edge->list[UPPER],
 974                                                       &list);
 975                                 } else
 976                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 977                         } else {
 978                                 upper = rb_entry(rb_node, struct backref_node,
 979                                                  rb_node);
 980                                 BUG_ON(!upper->checked);
 981                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 982                                 if (!upper->owner)
 983                                         upper->owner = btrfs_header_owner(eb);
 984                         }
 985                         list_add_tail(&edge->list[LOWER], &lower->upper);
 986                         edge->node[LOWER] = lower;
 987                         edge->node[UPPER] = upper;
 988
 989                         if (rb_node)
 990                                 break;
 991                         lower = upper;
 992                         upper = NULL;
 993                 }
 994                 btrfs_release_path(path2);
 995 next:
 996                 if (ptr < end) {
 997                         ptr += btrfs_extent_inline_ref_size(key.type);
 998                         if (ptr >= end) {
 999                                 WARN_ON(ptr > end);
1000                                 ptr = 0;
1001                                 end = 0;
1002                         }
1003                 }
1004                 if (ptr >= end)
1005                         path1->slots[0]++;
1006         }
1007         btrfs_release_path(path1);
1008
1009         cur->checked = 1;
1010         WARN_ON(exist);
1011
1012         /* the pending list isn't empty, take the first block to process */
1013         if (!list_empty(&list)) {
1014                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1015                 list_del_init(&edge->list[UPPER]);
1016                 cur = edge->node[UPPER];
1017                 goto again;
1018         }
1019
1020         /*
1021          * everything goes well, connect backref nodes and insert backref nodes
1022          * into the cache.
1023          */
1024         BUG_ON(!node->checked);
1025         cowonly = node->cowonly;
1026         if (!cowonly) {
1027                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1028                                       &node->rb_node);
1029                 if (rb_node)
1030                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1031                 list_add_tail(&node->lower, &cache->leaves);
1032         }
1033
1034         list_for_each_entry(edge, &node->upper, list[LOWER])
1035                 list_add_tail(&edge->list[UPPER], &list);
1036
1037         while (!list_empty(&list)) {
1038                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1039                 list_del_init(&edge->list[UPPER]);
1040                 upper = edge->node[UPPER];
1041                 if (upper->detached) {
1042                         list_del(&edge->list[LOWER]);
1043                         lower = edge->node[LOWER];
1044                         free_backref_edge(cache, edge);
1045                         if (list_empty(&lower->upper))
1046                                 list_add(&lower->list, &useless);
1047                         continue;
1048                 }
1049
1050                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1051                         if (upper->lowest) {
1052                                 list_del_init(&upper->lower);
1053                                 upper->lowest = 0;
1054                         }
1055
1056                         list_add_tail(&edge->list[UPPER], &upper->lower);
1057                         continue;
1058                 }
1059
1060                 BUG_ON(!upper->checked);
1061                 BUG_ON(cowonly != upper->cowonly);
1062                 if (!cowonly) {
1063                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1064                                               &upper->rb_node);
1065                         if (rb_node)
1066                                 backref_tree_panic(rb_node, -EEXIST,
1067                                                    upper->bytenr);
1068                 }
1069
1070                 list_add_tail(&edge->list[UPPER], &upper->lower);
1071
1072                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1073                         list_add_tail(&edge->list[UPPER], &list);
1074         }
1075         /*
1076          * process useless backref nodes. backref nodes for tree leaves
1077          * are deleted from the cache. backref nodes for upper level
1078          * tree blocks are left in the cache to avoid unnecessary backref
1079          * lookup.
1080          */
1081         while (!list_empty(&useless)) {
1082                 upper = list_entry(useless.next, struct backref_node, list);
1083                 list_del_init(&upper->list);
1084                 BUG_ON(!list_empty(&upper->upper));
1085                 if (upper == node)
1086                         node = NULL;
1087                 if (upper->lowest) {
1088                         list_del_init(&upper->lower);
1089                         upper->lowest = 0;
1090                 }
1091                 while (!list_empty(&upper->lower)) {
1092                         edge = list_entry(upper->lower.next,
1093                                           struct backref_edge, list[UPPER]);
1094                         list_del(&edge->list[UPPER]);
1095                         list_del(&edge->list[LOWER]);
1096                         lower = edge->node[LOWER];
1097                         free_backref_edge(cache, edge);
1098
1099                         if (list_empty(&lower->upper))
1100                                 list_add(&lower->list, &useless);
1101                 }
1102                 __mark_block_processed(rc, upper);
1103                 if (upper->level > 0) {
1104                         list_add(&upper->list, &cache->detached);
1105                         upper->detached = 1;
1106                 } else {
1107                         rb_erase(&upper->rb_node, &cache->rb_root);
1108                         free_backref_node(cache, upper);
1109                 }
1110         }
1111 out:
1112         btrfs_free_path(path1);
1113         btrfs_free_path(path2);
1114         if (err) {
1115                 while (!list_empty(&useless)) {
1116                         lower = list_entry(useless.next,
1117                                            struct backref_node, upper);
1118                         list_del_init(&lower->upper);
1119                 }
1120                 upper = node;
1121                 INIT_LIST_HEAD(&list);
1122                 while (upper) {
1123                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1124                                 list_splice_tail(&upper->upper, &list);
1125                                 free_backref_node(cache, upper);
1126                         }
1127
1128                         if (list_empty(&list))
1129                                 break;
1130
1131                         edge = list_entry(list.next, struct backref_edge,
1132                                           list[LOWER]);
1133                         list_del(&edge->list[LOWER]);
1134                         upper = edge->node[UPPER];
1135                         free_backref_edge(cache, edge);
1136                 }
1137                 return ERR_PTR(err);
1138         }
1139         BUG_ON(node && node->detached);
1140         return node;
1141 }
1142
1143 /*
1144  * helper to add backref node for the newly created snapshot.
1145  * the backref node is created by cloning backref node that
1146  * corresponds to root of source tree
1147  */
1148 static int clone_backref_node(struct btrfs_trans_handle *trans,
1149                               struct reloc_control *rc,
1150                               struct btrfs_root *src,
1151                               struct btrfs_root *dest)
1152 {
1153         struct btrfs_root *reloc_root = src->reloc_root;
1154         struct backref_cache *cache = &rc->backref_cache;
1155         struct backref_node *node = NULL;
1156         struct backref_node *new_node;
1157         struct backref_edge *edge;
1158         struct backref_edge *new_edge;
1159         struct rb_node *rb_node;
1160
1161         if (cache->last_trans > 0)
1162                 update_backref_cache(trans, cache);
1163
1164         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1165         if (rb_node) {
1166                 node = rb_entry(rb_node, struct backref_node, rb_node);
1167                 if (node->detached)
1168                         node = NULL;
1169                 else
1170                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1171         }
1172
1173         if (!node) {
1174                 rb_node = tree_search(&cache->rb_root,
1175                                       reloc_root->commit_root->start);
1176                 if (rb_node) {
1177                         node = rb_entry(rb_node, struct backref_node,
1178                                         rb_node);
1179                         BUG_ON(node->detached);
1180                 }
1181         }
1182
1183         if (!node)
1184                 return 0;
1185
1186         new_node = alloc_backref_node(cache);
1187         if (!new_node)
1188                 return -ENOMEM;
1189
1190         new_node->bytenr = dest->node->start;
1191         new_node->level = node->level;
1192         new_node->lowest = node->lowest;
1193         new_node->checked = 1;
1194         new_node->root = dest;
1195
1196         if (!node->lowest) {
1197                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1198                         new_edge = alloc_backref_edge(cache);
1199                         if (!new_edge)
1200                                 goto fail;
1201
1202                         new_edge->node[UPPER] = new_node;
1203                         new_edge->node[LOWER] = edge->node[LOWER];
1204                         list_add_tail(&new_edge->list[UPPER],
1205                                       &new_node->lower);
1206                 }
1207         } else {
1208                 list_add_tail(&new_node->lower, &cache->leaves);
1209         }
1210
1211         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1212                               &new_node->rb_node);
1213         if (rb_node)
1214                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1215
1216         if (!new_node->lowest) {
1217                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1218                         list_add_tail(&new_edge->list[LOWER],
1219                                       &new_edge->node[LOWER]->upper);
1220                 }
1221         }
1222         return 0;
1223 fail:
1224         while (!list_empty(&new_node->lower)) {
1225                 new_edge = list_entry(new_node->lower.next,
1226                                       struct backref_edge, list[UPPER]);
1227                 list_del(&new_edge->list[UPPER]);
1228                 free_backref_edge(cache, new_edge);
1229         }
1230         free_backref_node(cache, new_node);
1231         return -ENOMEM;
1232 }
1233
1234 /*
1235  * helper to add 'address of tree root -> reloc tree' mapping
1236  */
1237 static int __must_check __add_reloc_root(struct btrfs_root *root)
1238 {
1239         struct rb_node *rb_node;
1240         struct mapping_node *node;
1241         struct reloc_control *rc = root->fs_info->reloc_ctl;
1242
1243         node = kmalloc(sizeof(*node), GFP_NOFS);
1244         if (!node)
1245                 return -ENOMEM;
1246
1247         node->bytenr = root->node->start;
1248         node->data = root;
1249
1250         spin_lock(&rc->reloc_root_tree.lock);
1251         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1252                               node->bytenr, &node->rb_node);
1253         spin_unlock(&rc->reloc_root_tree.lock);
1254         if (rb_node) {
1255                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1256                             "for start=%llu while inserting into relocation "
1257                             "tree\n", node->bytenr);
1258                 kfree(node);
1259                 return -EEXIST;
1260         }
1261
1262         list_add_tail(&root->root_list, &rc->reloc_roots);
1263         return 0;
1264 }
1265
1266 /*
1267  * helper to update/delete the 'address of tree root -> reloc tree'
1268  * mapping
1269  */
1270 static int __update_reloc_root(struct btrfs_root *root, int del)
1271 {
1272         struct rb_node *rb_node;
1273         struct mapping_node *node = NULL;
1274         struct reloc_control *rc = root->fs_info->reloc_ctl;
1275
1276         spin_lock(&rc->reloc_root_tree.lock);
1277         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1278                               root->commit_root->start);
1279         if (rb_node) {
1280                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1281                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1282         }
1283         spin_unlock(&rc->reloc_root_tree.lock);
1284
1285         if (!node)
1286                 return 0;
1287         BUG_ON((struct btrfs_root *)node->data != root);
1288
1289         if (!del) {
1290                 spin_lock(&rc->reloc_root_tree.lock);
1291                 node->bytenr = root->node->start;
1292                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1293                                       node->bytenr, &node->rb_node);
1294                 spin_unlock(&rc->reloc_root_tree.lock);
1295                 if (rb_node)
1296                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1297         } else {
1298                 spin_lock(&root->fs_info->trans_lock);
1299                 list_del_init(&root->root_list);
1300                 spin_unlock(&root->fs_info->trans_lock);
1301                 kfree(node);
1302         }
1303         return 0;
1304 }
1305
1306 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1307                                         struct btrfs_root *root, u64 objectid)
1308 {
1309         struct btrfs_root *reloc_root;
1310         struct extent_buffer *eb;
1311         struct btrfs_root_item *root_item;
1312         struct btrfs_key root_key;
1313         u64 last_snap = 0;
1314         int ret;
1315
1316         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1317         BUG_ON(!root_item);
1318
1319         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1320         root_key.type = BTRFS_ROOT_ITEM_KEY;
1321         root_key.offset = objectid;
1322
1323         if (root->root_key.objectid == objectid) {
1324                 /* called by btrfs_init_reloc_root */
1325                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1326                                       BTRFS_TREE_RELOC_OBJECTID);
1327                 BUG_ON(ret);
1328
1329                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1330                 btrfs_set_root_last_snapshot(&root->root_item,
1331                                              trans->transid - 1);
1332         } else {
1333                 /*
1334                  * called by btrfs_reloc_post_snapshot_hook.
1335                  * the source tree is a reloc tree, all tree blocks
1336                  * modified after it was created have RELOC flag
1337                  * set in their headers. so it's OK to not update
1338                  * the 'last_snapshot'.
1339                  */
1340                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1341                                       BTRFS_TREE_RELOC_OBJECTID);
1342                 BUG_ON(ret);
1343         }
1344
1345         memcpy(root_item, &root->root_item, sizeof(*root_item));
1346         btrfs_set_root_bytenr(root_item, eb->start);
1347         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1348         btrfs_set_root_generation(root_item, trans->transid);
1349
1350         if (root->root_key.objectid == objectid) {
1351                 btrfs_set_root_refs(root_item, 0);
1352                 memset(&root_item->drop_progress, 0,
1353                        sizeof(struct btrfs_disk_key));
1354                 root_item->drop_level = 0;
1355                 /*
1356                  * abuse rtransid, it is safe because it is impossible to
1357                  * receive data into a relocation tree.
1358                  */
1359                 btrfs_set_root_rtransid(root_item, last_snap);
1360                 btrfs_set_root_otransid(root_item, trans->transid);
1361         }
1362
1363         btrfs_tree_unlock(eb);
1364         free_extent_buffer(eb);
1365
1366         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1367                                 &root_key, root_item);
1368         BUG_ON(ret);
1369         kfree(root_item);
1370
1371         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1372         BUG_ON(IS_ERR(reloc_root));
1373         reloc_root->last_trans = trans->transid;
1374         return reloc_root;
1375 }
1376
1377 /*
1378  * create reloc tree for a given fs tree. reloc tree is just a
1379  * snapshot of the fs tree with special root objectid.
1380  */
1381 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1382                           struct btrfs_root *root)
1383 {
1384         struct btrfs_root *reloc_root;
1385         struct reloc_control *rc = root->fs_info->reloc_ctl;
1386         struct btrfs_block_rsv *rsv;
1387         int clear_rsv = 0;
1388         int ret;
1389
1390         if (root->reloc_root) {
1391                 reloc_root = root->reloc_root;
1392                 reloc_root->last_trans = trans->transid;
1393                 return 0;
1394         }
1395
1396         if (!rc || !rc->create_reloc_tree ||
1397             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1398                 return 0;
1399
1400         if (!trans->reloc_reserved) {
1401                 rsv = trans->block_rsv;
1402                 trans->block_rsv = rc->block_rsv;
1403                 clear_rsv = 1;
1404         }
1405         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1406         if (clear_rsv)
1407                 trans->block_rsv = rsv;
1408
1409         ret = __add_reloc_root(reloc_root);
1410         BUG_ON(ret < 0);
1411         root->reloc_root = reloc_root;
1412         return 0;
1413 }
1414
1415 /*
1416  * update root item of reloc tree
1417  */
1418 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1419                             struct btrfs_root *root)
1420 {
1421         struct btrfs_root *reloc_root;
1422         struct btrfs_root_item *root_item;
1423         int del = 0;
1424         int ret;
1425
1426         if (!root->reloc_root)
1427                 goto out;
1428
1429         reloc_root = root->reloc_root;
1430         root_item = &reloc_root->root_item;
1431
1432         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1433             btrfs_root_refs(root_item) == 0) {
1434                 root->reloc_root = NULL;
1435                 del = 1;
1436         }
1437
1438         __update_reloc_root(reloc_root, del);
1439
1440         if (reloc_root->commit_root != reloc_root->node) {
1441                 btrfs_set_root_node(root_item, reloc_root->node);
1442                 free_extent_buffer(reloc_root->commit_root);
1443                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1444         }
1445
1446         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1447                                 &reloc_root->root_key, root_item);
1448         BUG_ON(ret);
1449
1450 out:
1451         return 0;
1452 }
1453
1454 /*
1455  * helper to find first cached inode with inode number >= objectid
1456  * in a subvolume
1457  */
1458 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1459 {
1460         struct rb_node *node;
1461         struct rb_node *prev;
1462         struct btrfs_inode *entry;
1463         struct inode *inode;
1464
1465         spin_lock(&root->inode_lock);
1466 again:
1467         node = root->inode_tree.rb_node;
1468         prev = NULL;
1469         while (node) {
1470                 prev = node;
1471                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1472
1473                 if (objectid < btrfs_ino(&entry->vfs_inode))
1474                         node = node->rb_left;
1475                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1476                         node = node->rb_right;
1477                 else
1478                         break;
1479         }
1480         if (!node) {
1481                 while (prev) {
1482                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1483                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1484                                 node = prev;
1485                                 break;
1486                         }
1487                         prev = rb_next(prev);
1488                 }
1489         }
1490         while (node) {
1491                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1492                 inode = igrab(&entry->vfs_inode);
1493                 if (inode) {
1494                         spin_unlock(&root->inode_lock);
1495                         return inode;
1496                 }
1497
1498                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1499                 if (cond_resched_lock(&root->inode_lock))
1500                         goto again;
1501
1502                 node = rb_next(node);
1503         }
1504         spin_unlock(&root->inode_lock);
1505         return NULL;
1506 }
1507
1508 static int in_block_group(u64 bytenr,
1509                           struct btrfs_block_group_cache *block_group)
1510 {
1511         if (bytenr >= block_group->key.objectid &&
1512             bytenr < block_group->key.objectid + block_group->key.offset)
1513                 return 1;
1514         return 0;
1515 }
1516
1517 /*
1518  * get new location of data
1519  */
1520 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1521                             u64 bytenr, u64 num_bytes)
1522 {
1523         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1524         struct btrfs_path *path;
1525         struct btrfs_file_extent_item *fi;
1526         struct extent_buffer *leaf;
1527         int ret;
1528
1529         path = btrfs_alloc_path();
1530         if (!path)
1531                 return -ENOMEM;
1532
1533         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1534         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1535                                        bytenr, 0);
1536         if (ret < 0)
1537                 goto out;
1538         if (ret > 0) {
1539                 ret = -ENOENT;
1540                 goto out;
1541         }
1542
1543         leaf = path->nodes[0];
1544         fi = btrfs_item_ptr(leaf, path->slots[0],
1545                             struct btrfs_file_extent_item);
1546
1547         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1548                btrfs_file_extent_compression(leaf, fi) ||
1549                btrfs_file_extent_encryption(leaf, fi) ||
1550                btrfs_file_extent_other_encoding(leaf, fi));
1551
1552         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1553                 ret = -EINVAL;
1554                 goto out;
1555         }
1556
1557         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1558         ret = 0;
1559 out:
1560         btrfs_free_path(path);
1561         return ret;
1562 }
1563
1564 /*
1565  * update file extent items in the tree leaf to point to
1566  * the new locations.
1567  */
1568 static noinline_for_stack
1569 int replace_file_extents(struct btrfs_trans_handle *trans,
1570                          struct reloc_control *rc,
1571                          struct btrfs_root *root,
1572                          struct extent_buffer *leaf)
1573 {
1574         struct btrfs_key key;
1575         struct btrfs_file_extent_item *fi;
1576         struct inode *inode = NULL;
1577         u64 parent;
1578         u64 bytenr;
1579         u64 new_bytenr = 0;
1580         u64 num_bytes;
1581         u64 end;
1582         u32 nritems;
1583         u32 i;
1584         int ret = 0;
1585         int first = 1;
1586         int dirty = 0;
1587
1588         if (rc->stage != UPDATE_DATA_PTRS)
1589                 return 0;
1590
1591         /* reloc trees always use full backref */
1592         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1593                 parent = leaf->start;
1594         else
1595                 parent = 0;
1596
1597         nritems = btrfs_header_nritems(leaf);
1598         for (i = 0; i < nritems; i++) {
1599                 cond_resched();
1600                 btrfs_item_key_to_cpu(leaf, &key, i);
1601                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1602                         continue;
1603                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1604                 if (btrfs_file_extent_type(leaf, fi) ==
1605                     BTRFS_FILE_EXTENT_INLINE)
1606                         continue;
1607                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1608                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1609                 if (bytenr == 0)
1610                         continue;
1611                 if (!in_block_group(bytenr, rc->block_group))
1612                         continue;
1613
1614                 /*
1615                  * if we are modifying block in fs tree, wait for readpage
1616                  * to complete and drop the extent cache
1617                  */
1618                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1619                         if (first) {
1620                                 inode = find_next_inode(root, key.objectid);
1621                                 first = 0;
1622                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1623                                 btrfs_add_delayed_iput(inode);
1624                                 inode = find_next_inode(root, key.objectid);
1625                         }
1626                         if (inode && btrfs_ino(inode) == key.objectid) {
1627                                 end = key.offset +
1628                                       btrfs_file_extent_num_bytes(leaf, fi);
1629                                 WARN_ON(!IS_ALIGNED(key.offset,
1630                                                     root->sectorsize));
1631                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1632                                 end--;
1633                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1634                                                       key.offset, end);
1635                                 if (!ret)
1636                                         continue;
1637
1638                                 btrfs_drop_extent_cache(inode, key.offset, end,
1639                                                         1);
1640                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1641                                               key.offset, end);
1642                         }
1643                 }
1644
1645                 ret = get_new_location(rc->data_inode, &new_bytenr,
1646                                        bytenr, num_bytes);
1647                 if (ret) {
1648                         /*
1649                          * Don't have to abort since we've not changed anything
1650                          * in the file extent yet.
1651                          */
1652                         break;
1653                 }
1654
1655                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1656                 dirty = 1;
1657
1658                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1659                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1660                                            num_bytes, parent,
1661                                            btrfs_header_owner(leaf),
1662                                            key.objectid, key.offset, 1);
1663                 if (ret) {
1664                         btrfs_abort_transaction(trans, root, ret);
1665                         break;
1666                 }
1667
1668                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1669                                         parent, btrfs_header_owner(leaf),
1670                                         key.objectid, key.offset, 1);
1671                 if (ret) {
1672                         btrfs_abort_transaction(trans, root, ret);
1673                         break;
1674                 }
1675         }
1676         if (dirty)
1677                 btrfs_mark_buffer_dirty(leaf);
1678         if (inode)
1679                 btrfs_add_delayed_iput(inode);
1680         return ret;
1681 }
1682
1683 static noinline_for_stack
1684 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1685                      struct btrfs_path *path, int level)
1686 {
1687         struct btrfs_disk_key key1;
1688         struct btrfs_disk_key key2;
1689         btrfs_node_key(eb, &key1, slot);
1690         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1691         return memcmp(&key1, &key2, sizeof(key1));
1692 }
1693
1694 /*
1695  * try to replace tree blocks in fs tree with the new blocks
1696  * in reloc tree. tree blocks haven't been modified since the
1697  * reloc tree was create can be replaced.
1698  *
1699  * if a block was replaced, level of the block + 1 is returned.
1700  * if no block got replaced, 0 is returned. if there are other
1701  * errors, a negative error number is returned.
1702  */
1703 static noinline_for_stack
1704 int replace_path(struct btrfs_trans_handle *trans,
1705                  struct btrfs_root *dest, struct btrfs_root *src,
1706                  struct btrfs_path *path, struct btrfs_key *next_key,
1707                  int lowest_level, int max_level)
1708 {
1709         struct extent_buffer *eb;
1710         struct extent_buffer *parent;
1711         struct btrfs_key key;
1712         u64 old_bytenr;
1713         u64 new_bytenr;
1714         u64 old_ptr_gen;
1715         u64 new_ptr_gen;
1716         u64 last_snapshot;
1717         u32 blocksize;
1718         int cow = 0;
1719         int level;
1720         int ret;
1721         int slot;
1722
1723         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1724         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1725
1726         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1727 again:
1728         slot = path->slots[lowest_level];
1729         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1730
1731         eb = btrfs_lock_root_node(dest);
1732         btrfs_set_lock_blocking(eb);
1733         level = btrfs_header_level(eb);
1734
1735         if (level < lowest_level) {
1736                 btrfs_tree_unlock(eb);
1737                 free_extent_buffer(eb);
1738                 return 0;
1739         }
1740
1741         if (cow) {
1742                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1743                 BUG_ON(ret);
1744         }
1745         btrfs_set_lock_blocking(eb);
1746
1747         if (next_key) {
1748                 next_key->objectid = (u64)-1;
1749                 next_key->type = (u8)-1;
1750                 next_key->offset = (u64)-1;
1751         }
1752
1753         parent = eb;
1754         while (1) {
1755                 level = btrfs_header_level(parent);
1756                 BUG_ON(level < lowest_level);
1757
1758                 ret = btrfs_bin_search(parent, &key, level, &slot);
1759                 if (ret && slot > 0)
1760                         slot--;
1761
1762                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1763                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1764
1765                 old_bytenr = btrfs_node_blockptr(parent, slot);
1766                 blocksize = btrfs_level_size(dest, level - 1);
1767                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1768
1769                 if (level <= max_level) {
1770                         eb = path->nodes[level];
1771                         new_bytenr = btrfs_node_blockptr(eb,
1772                                                         path->slots[level]);
1773                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1774                                                         path->slots[level]);
1775                 } else {
1776                         new_bytenr = 0;
1777                         new_ptr_gen = 0;
1778                 }
1779
1780                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1781                         ret = level;
1782                         break;
1783                 }
1784
1785                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1786                     memcmp_node_keys(parent, slot, path, level)) {
1787                         if (level <= lowest_level) {
1788                                 ret = 0;
1789                                 break;
1790                         }
1791
1792                         eb = read_tree_block(dest, old_bytenr, blocksize,
1793                                              old_ptr_gen);
1794                         if (!eb || !extent_buffer_uptodate(eb)) {
1795                                 ret = (!eb) ? -ENOMEM : -EIO;
1796                                 free_extent_buffer(eb);
1797                                 break;
1798                         }
1799                         btrfs_tree_lock(eb);
1800                         if (cow) {
1801                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1802                                                       slot, &eb);
1803                                 BUG_ON(ret);
1804                         }
1805                         btrfs_set_lock_blocking(eb);
1806
1807                         btrfs_tree_unlock(parent);
1808                         free_extent_buffer(parent);
1809
1810                         parent = eb;
1811                         continue;
1812                 }
1813
1814                 if (!cow) {
1815                         btrfs_tree_unlock(parent);
1816                         free_extent_buffer(parent);
1817                         cow = 1;
1818                         goto again;
1819                 }
1820
1821                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1822                                       path->slots[level]);
1823                 btrfs_release_path(path);
1824
1825                 path->lowest_level = level;
1826                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1827                 path->lowest_level = 0;
1828                 BUG_ON(ret);
1829
1830                 /*
1831                  * swap blocks in fs tree and reloc tree.
1832                  */
1833                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1834                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1835                 btrfs_mark_buffer_dirty(parent);
1836
1837                 btrfs_set_node_blockptr(path->nodes[level],
1838                                         path->slots[level], old_bytenr);
1839                 btrfs_set_node_ptr_generation(path->nodes[level],
1840                                               path->slots[level], old_ptr_gen);
1841                 btrfs_mark_buffer_dirty(path->nodes[level]);
1842
1843                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1844                                         path->nodes[level]->start,
1845                                         src->root_key.objectid, level - 1, 0,
1846                                         1);
1847                 BUG_ON(ret);
1848                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1849                                         0, dest->root_key.objectid, level - 1,
1850                                         0, 1);
1851                 BUG_ON(ret);
1852
1853                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1854                                         path->nodes[level]->start,
1855                                         src->root_key.objectid, level - 1, 0,
1856                                         1);
1857                 BUG_ON(ret);
1858
1859                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1860                                         0, dest->root_key.objectid, level - 1,
1861                                         0, 1);
1862                 BUG_ON(ret);
1863
1864                 btrfs_unlock_up_safe(path, 0);
1865
1866                 ret = level;
1867                 break;
1868         }
1869         btrfs_tree_unlock(parent);
1870         free_extent_buffer(parent);
1871         return ret;
1872 }
1873
1874 /*
1875  * helper to find next relocated block in reloc tree
1876  */
1877 static noinline_for_stack
1878 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1879                        int *level)
1880 {
1881         struct extent_buffer *eb;
1882         int i;
1883         u64 last_snapshot;
1884         u32 nritems;
1885
1886         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1887
1888         for (i = 0; i < *level; i++) {
1889                 free_extent_buffer(path->nodes[i]);
1890                 path->nodes[i] = NULL;
1891         }
1892
1893         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1894                 eb = path->nodes[i];
1895                 nritems = btrfs_header_nritems(eb);
1896                 while (path->slots[i] + 1 < nritems) {
1897                         path->slots[i]++;
1898                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1899                             last_snapshot)
1900                                 continue;
1901
1902                         *level = i;
1903                         return 0;
1904                 }
1905                 free_extent_buffer(path->nodes[i]);
1906                 path->nodes[i] = NULL;
1907         }
1908         return 1;
1909 }
1910
1911 /*
1912  * walk down reloc tree to find relocated block of lowest level
1913  */
1914 static noinline_for_stack
1915 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1916                          int *level)
1917 {
1918         struct extent_buffer *eb = NULL;
1919         int i;
1920         u64 bytenr;
1921         u64 ptr_gen = 0;
1922         u64 last_snapshot;
1923         u32 blocksize;
1924         u32 nritems;
1925
1926         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1927
1928         for (i = *level; i > 0; i--) {
1929                 eb = path->nodes[i];
1930                 nritems = btrfs_header_nritems(eb);
1931                 while (path->slots[i] < nritems) {
1932                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1933                         if (ptr_gen > last_snapshot)
1934                                 break;
1935                         path->slots[i]++;
1936                 }
1937                 if (path->slots[i] >= nritems) {
1938                         if (i == *level)
1939                                 break;
1940                         *level = i + 1;
1941                         return 0;
1942                 }
1943                 if (i == 1) {
1944                         *level = i;
1945                         return 0;
1946                 }
1947
1948                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1949                 blocksize = btrfs_level_size(root, i - 1);
1950                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1951                 if (!eb || !extent_buffer_uptodate(eb)) {
1952                         free_extent_buffer(eb);
1953                         return -EIO;
1954                 }
1955                 BUG_ON(btrfs_header_level(eb) != i - 1);
1956                 path->nodes[i - 1] = eb;
1957                 path->slots[i - 1] = 0;
1958         }
1959         return 1;
1960 }
1961
1962 /*
1963  * invalidate extent cache for file extents whose key in range of
1964  * [min_key, max_key)
1965  */
1966 static int invalidate_extent_cache(struct btrfs_root *root,
1967                                    struct btrfs_key *min_key,
1968                                    struct btrfs_key *max_key)
1969 {
1970         struct inode *inode = NULL;
1971         u64 objectid;
1972         u64 start, end;
1973         u64 ino;
1974
1975         objectid = min_key->objectid;
1976         while (1) {
1977                 cond_resched();
1978                 iput(inode);
1979
1980                 if (objectid > max_key->objectid)
1981                         break;
1982
1983                 inode = find_next_inode(root, objectid);
1984                 if (!inode)
1985                         break;
1986                 ino = btrfs_ino(inode);
1987
1988                 if (ino > max_key->objectid) {
1989                         iput(inode);
1990                         break;
1991                 }
1992
1993                 objectid = ino + 1;
1994                 if (!S_ISREG(inode->i_mode))
1995                         continue;
1996
1997                 if (unlikely(min_key->objectid == ino)) {
1998                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1999                                 continue;
2000                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2001                                 start = 0;
2002                         else {
2003                                 start = min_key->offset;
2004                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2005                         }
2006                 } else {
2007                         start = 0;
2008                 }
2009
2010                 if (unlikely(max_key->objectid == ino)) {
2011                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2012                                 continue;
2013                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2014                                 end = (u64)-1;
2015                         } else {
2016                                 if (max_key->offset == 0)
2017                                         continue;
2018                                 end = max_key->offset;
2019                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2020                                 end--;
2021                         }
2022                 } else {
2023                         end = (u64)-1;
2024                 }
2025
2026                 /* the lock_extent waits for readpage to complete */
2027                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2028                 btrfs_drop_extent_cache(inode, start, end, 1);
2029                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2030         }
2031         return 0;
2032 }
2033
2034 static int find_next_key(struct btrfs_path *path, int level,
2035                          struct btrfs_key *key)
2036
2037 {
2038         while (level < BTRFS_MAX_LEVEL) {
2039                 if (!path->nodes[level])
2040                         break;
2041                 if (path->slots[level] + 1 <
2042                     btrfs_header_nritems(path->nodes[level])) {
2043                         btrfs_node_key_to_cpu(path->nodes[level], key,
2044                                               path->slots[level] + 1);
2045                         return 0;
2046                 }
2047                 level++;
2048         }
2049         return 1;
2050 }
2051
2052 /*
2053  * merge the relocated tree blocks in reloc tree with corresponding
2054  * fs tree.
2055  */
2056 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2057                                                struct btrfs_root *root)
2058 {
2059         LIST_HEAD(inode_list);
2060         struct btrfs_key key;
2061         struct btrfs_key next_key;
2062         struct btrfs_trans_handle *trans = NULL;
2063         struct btrfs_root *reloc_root;
2064         struct btrfs_root_item *root_item;
2065         struct btrfs_path *path;
2066         struct extent_buffer *leaf;
2067         int level;
2068         int max_level;
2069         int replaced = 0;
2070         int ret;
2071         int err = 0;
2072         u32 min_reserved;
2073
2074         path = btrfs_alloc_path();
2075         if (!path)
2076                 return -ENOMEM;
2077         path->reada = 1;
2078
2079         reloc_root = root->reloc_root;
2080         root_item = &reloc_root->root_item;
2081
2082         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2083                 level = btrfs_root_level(root_item);
2084                 extent_buffer_get(reloc_root->node);
2085                 path->nodes[level] = reloc_root->node;
2086                 path->slots[level] = 0;
2087         } else {
2088                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2089
2090                 level = root_item->drop_level;
2091                 BUG_ON(level == 0);
2092                 path->lowest_level = level;
2093                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2094                 path->lowest_level = 0;
2095                 if (ret < 0) {
2096                         btrfs_free_path(path);
2097                         return ret;
2098                 }
2099
2100                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2101                                       path->slots[level]);
2102                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2103
2104                 btrfs_unlock_up_safe(path, 0);
2105         }
2106
2107         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2108         memset(&next_key, 0, sizeof(next_key));
2109
2110         while (1) {
2111                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2112                                              BTRFS_RESERVE_FLUSH_ALL);
2113                 if (ret) {
2114                         err = ret;
2115                         goto out;
2116                 }
2117                 trans = btrfs_start_transaction(root, 0);
2118                 if (IS_ERR(trans)) {
2119                         err = PTR_ERR(trans);
2120                         trans = NULL;
2121                         goto out;
2122                 }
2123                 trans->block_rsv = rc->block_rsv;
2124
2125                 replaced = 0;
2126                 max_level = level;
2127
2128                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2129                 if (ret < 0) {
2130                         err = ret;
2131                         goto out;
2132                 }
2133                 if (ret > 0)
2134                         break;
2135
2136                 if (!find_next_key(path, level, &key) &&
2137                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2138                         ret = 0;
2139                 } else {
2140                         ret = replace_path(trans, root, reloc_root, path,
2141                                            &next_key, level, max_level);
2142                 }
2143                 if (ret < 0) {
2144                         err = ret;
2145                         goto out;
2146                 }
2147
2148                 if (ret > 0) {
2149                         level = ret;
2150                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2151                                               path->slots[level]);
2152                         replaced = 1;
2153                 }
2154
2155                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2156                 if (ret > 0)
2157                         break;
2158
2159                 BUG_ON(level == 0);
2160                 /*
2161                  * save the merging progress in the drop_progress.
2162                  * this is OK since root refs == 1 in this case.
2163                  */
2164                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2165                                path->slots[level]);
2166                 root_item->drop_level = level;
2167
2168                 btrfs_end_transaction_throttle(trans, root);
2169                 trans = NULL;
2170
2171                 btrfs_btree_balance_dirty(root);
2172
2173                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2174                         invalidate_extent_cache(root, &key, &next_key);
2175         }
2176
2177         /*
2178          * handle the case only one block in the fs tree need to be
2179          * relocated and the block is tree root.
2180          */
2181         leaf = btrfs_lock_root_node(root);
2182         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2183         btrfs_tree_unlock(leaf);
2184         free_extent_buffer(leaf);
2185         if (ret < 0)
2186                 err = ret;
2187 out:
2188         btrfs_free_path(path);
2189
2190         if (err == 0) {
2191                 memset(&root_item->drop_progress, 0,
2192                        sizeof(root_item->drop_progress));
2193                 root_item->drop_level = 0;
2194                 btrfs_set_root_refs(root_item, 0);
2195                 btrfs_update_reloc_root(trans, root);
2196         }
2197
2198         if (trans)
2199                 btrfs_end_transaction_throttle(trans, root);
2200
2201         btrfs_btree_balance_dirty(root);
2202
2203         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2204                 invalidate_extent_cache(root, &key, &next_key);
2205
2206         return err;
2207 }
2208
2209 static noinline_for_stack
2210 int prepare_to_merge(struct reloc_control *rc, int err)
2211 {
2212         struct btrfs_root *root = rc->extent_root;
2213         struct btrfs_root *reloc_root;
2214         struct btrfs_trans_handle *trans;
2215         LIST_HEAD(reloc_roots);
2216         u64 num_bytes = 0;
2217         int ret;
2218
2219         mutex_lock(&root->fs_info->reloc_mutex);
2220         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2221         rc->merging_rsv_size += rc->nodes_relocated * 2;
2222         mutex_unlock(&root->fs_info->reloc_mutex);
2223
2224 again:
2225         if (!err) {
2226                 num_bytes = rc->merging_rsv_size;
2227                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2228                                           BTRFS_RESERVE_FLUSH_ALL);
2229                 if (ret)
2230                         err = ret;
2231         }
2232
2233         trans = btrfs_join_transaction(rc->extent_root);
2234         if (IS_ERR(trans)) {
2235                 if (!err)
2236                         btrfs_block_rsv_release(rc->extent_root,
2237                                                 rc->block_rsv, num_bytes);
2238                 return PTR_ERR(trans);
2239         }
2240
2241         if (!err) {
2242                 if (num_bytes != rc->merging_rsv_size) {
2243                         btrfs_end_transaction(trans, rc->extent_root);
2244                         btrfs_block_rsv_release(rc->extent_root,
2245                                                 rc->block_rsv, num_bytes);
2246                         goto again;
2247                 }
2248         }
2249
2250         rc->merge_reloc_tree = 1;
2251
2252         while (!list_empty(&rc->reloc_roots)) {
2253                 reloc_root = list_entry(rc->reloc_roots.next,
2254                                         struct btrfs_root, root_list);
2255                 list_del_init(&reloc_root->root_list);
2256
2257                 root = read_fs_root(reloc_root->fs_info,
2258                                     reloc_root->root_key.offset);
2259                 BUG_ON(IS_ERR(root));
2260                 BUG_ON(root->reloc_root != reloc_root);
2261
2262                 /*
2263                  * set reference count to 1, so btrfs_recover_relocation
2264                  * knows it should resumes merging
2265                  */
2266                 if (!err)
2267                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2268                 btrfs_update_reloc_root(trans, root);
2269
2270                 list_add(&reloc_root->root_list, &reloc_roots);
2271         }
2272
2273         list_splice(&reloc_roots, &rc->reloc_roots);
2274
2275         if (!err)
2276                 btrfs_commit_transaction(trans, rc->extent_root);
2277         else
2278                 btrfs_end_transaction(trans, rc->extent_root);
2279         return err;
2280 }
2281
2282 static noinline_for_stack
2283 void free_reloc_roots(struct list_head *list)
2284 {
2285         struct btrfs_root *reloc_root;
2286
2287         while (!list_empty(list)) {
2288                 reloc_root = list_entry(list->next, struct btrfs_root,
2289                                         root_list);
2290                 __update_reloc_root(reloc_root, 1);
2291                 free_extent_buffer(reloc_root->node);
2292                 free_extent_buffer(reloc_root->commit_root);
2293                 kfree(reloc_root);
2294         }
2295 }
2296
2297 static noinline_for_stack
2298 int merge_reloc_roots(struct reloc_control *rc)
2299 {
2300         struct btrfs_trans_handle *trans;
2301         struct btrfs_root *root;
2302         struct btrfs_root *reloc_root;
2303         u64 last_snap;
2304         u64 otransid;
2305         u64 objectid;
2306         LIST_HEAD(reloc_roots);
2307         int found = 0;
2308         int ret = 0;
2309 again:
2310         root = rc->extent_root;
2311
2312         /*
2313          * this serializes us with btrfs_record_root_in_transaction,
2314          * we have to make sure nobody is in the middle of
2315          * adding their roots to the list while we are
2316          * doing this splice
2317          */
2318         mutex_lock(&root->fs_info->reloc_mutex);
2319         list_splice_init(&rc->reloc_roots, &reloc_roots);
2320         mutex_unlock(&root->fs_info->reloc_mutex);
2321
2322         while (!list_empty(&reloc_roots)) {
2323                 found = 1;
2324                 reloc_root = list_entry(reloc_roots.next,
2325                                         struct btrfs_root, root_list);
2326
2327                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2328                         root = read_fs_root(reloc_root->fs_info,
2329                                             reloc_root->root_key.offset);
2330                         BUG_ON(IS_ERR(root));
2331                         BUG_ON(root->reloc_root != reloc_root);
2332
2333                         ret = merge_reloc_root(rc, root);
2334                         if (ret) {
2335                                 __update_reloc_root(reloc_root, 1);
2336                                 free_extent_buffer(reloc_root->node);
2337                                 free_extent_buffer(reloc_root->commit_root);
2338                                 kfree(reloc_root);
2339                                 goto out;
2340                         }
2341                 } else {
2342                         list_del_init(&reloc_root->root_list);
2343                 }
2344
2345                 /*
2346                  * we keep the old last snapshod transid in rtranid when we
2347                  * created the relocation tree.
2348                  */
2349                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2350                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2351                 objectid = reloc_root->root_key.offset;
2352
2353                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2354                 if (ret < 0) {
2355                         if (list_empty(&reloc_root->root_list))
2356                                 list_add_tail(&reloc_root->root_list,
2357                                               &reloc_roots);
2358                         goto out;
2359                 } else if (!ret) {
2360                         /*
2361                          * recover the last snapshot tranid to avoid
2362                          * the space balance break NOCOW.
2363                          */
2364                         root = read_fs_root(rc->extent_root->fs_info,
2365                                             objectid);
2366                         if (IS_ERR(root))
2367                                 continue;
2368
2369                         trans = btrfs_join_transaction(root);
2370                         BUG_ON(IS_ERR(trans));
2371
2372                         /* Check if the fs/file tree was snapshoted or not. */
2373                         if (btrfs_root_last_snapshot(&root->root_item) ==
2374                             otransid - 1)
2375                                 btrfs_set_root_last_snapshot(&root->root_item,
2376                                                              last_snap);
2377
2378                         btrfs_end_transaction(trans, root);
2379                 }
2380         }
2381
2382         if (found) {
2383                 found = 0;
2384                 goto again;
2385         }
2386 out:
2387         if (ret) {
2388                 btrfs_std_error(root->fs_info, ret);
2389                 if (!list_empty(&reloc_roots))
2390                         free_reloc_roots(&reloc_roots);
2391         }
2392
2393         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2394         return ret;
2395 }
2396
2397 static void free_block_list(struct rb_root *blocks)
2398 {
2399         struct tree_block *block;
2400         struct rb_node *rb_node;
2401         while ((rb_node = rb_first(blocks))) {
2402                 block = rb_entry(rb_node, struct tree_block, rb_node);
2403                 rb_erase(rb_node, blocks);
2404                 kfree(block);
2405         }
2406 }
2407
2408 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2409                                       struct btrfs_root *reloc_root)
2410 {
2411         struct btrfs_root *root;
2412
2413         if (reloc_root->last_trans == trans->transid)
2414                 return 0;
2415
2416         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2417         BUG_ON(IS_ERR(root));
2418         BUG_ON(root->reloc_root != reloc_root);
2419
2420         return btrfs_record_root_in_trans(trans, root);
2421 }
2422
2423 static noinline_for_stack
2424 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2425                                      struct reloc_control *rc,
2426                                      struct backref_node *node,
2427                                      struct backref_edge *edges[], int *nr)
2428 {
2429         struct backref_node *next;
2430         struct btrfs_root *root;
2431         int index = 0;
2432
2433         next = node;
2434         while (1) {
2435                 cond_resched();
2436                 next = walk_up_backref(next, edges, &index);
2437                 root = next->root;
2438                 BUG_ON(!root);
2439                 BUG_ON(!root->ref_cows);
2440
2441                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2442                         record_reloc_root_in_trans(trans, root);
2443                         break;
2444                 }
2445
2446                 btrfs_record_root_in_trans(trans, root);
2447                 root = root->reloc_root;
2448
2449                 if (next->new_bytenr != root->node->start) {
2450                         BUG_ON(next->new_bytenr);
2451                         BUG_ON(!list_empty(&next->list));
2452                         next->new_bytenr = root->node->start;
2453                         next->root = root;
2454                         list_add_tail(&next->list,
2455                                       &rc->backref_cache.changed);
2456                         __mark_block_processed(rc, next);
2457                         break;
2458                 }
2459
2460                 WARN_ON(1);
2461                 root = NULL;
2462                 next = walk_down_backref(edges, &index);
2463                 if (!next || next->level <= node->level)
2464                         break;
2465         }
2466         if (!root)
2467                 return NULL;
2468
2469         *nr = index;
2470         next = node;
2471         /* setup backref node path for btrfs_reloc_cow_block */
2472         while (1) {
2473                 rc->backref_cache.path[next->level] = next;
2474                 if (--index < 0)
2475                         break;
2476                 next = edges[index]->node[UPPER];
2477         }
2478         return root;
2479 }
2480
2481 /*
2482  * select a tree root for relocation. return NULL if the block
2483  * is reference counted. we should use do_relocation() in this
2484  * case. return a tree root pointer if the block isn't reference
2485  * counted. return -ENOENT if the block is root of reloc tree.
2486  */
2487 static noinline_for_stack
2488 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2489                                    struct backref_node *node)
2490 {
2491         struct backref_node *next;
2492         struct btrfs_root *root;
2493         struct btrfs_root *fs_root = NULL;
2494         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2495         int index = 0;
2496
2497         next = node;
2498         while (1) {
2499                 cond_resched();
2500                 next = walk_up_backref(next, edges, &index);
2501                 root = next->root;
2502                 BUG_ON(!root);
2503
2504                 /* no other choice for non-references counted tree */
2505                 if (!root->ref_cows)
2506                         return root;
2507
2508                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2509                         fs_root = root;
2510
2511                 if (next != node)
2512                         return NULL;
2513
2514                 next = walk_down_backref(edges, &index);
2515                 if (!next || next->level <= node->level)
2516                         break;
2517         }
2518
2519         if (!fs_root)
2520                 return ERR_PTR(-ENOENT);
2521         return fs_root;
2522 }
2523
2524 static noinline_for_stack
2525 u64 calcu_metadata_size(struct reloc_control *rc,
2526                         struct backref_node *node, int reserve)
2527 {
2528         struct backref_node *next = node;
2529         struct backref_edge *edge;
2530         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2531         u64 num_bytes = 0;
2532         int index = 0;
2533
2534         BUG_ON(reserve && node->processed);
2535
2536         while (next) {
2537                 cond_resched();
2538                 while (1) {
2539                         if (next->processed && (reserve || next != node))
2540                                 break;
2541
2542                         num_bytes += btrfs_level_size(rc->extent_root,
2543                                                       next->level);
2544
2545                         if (list_empty(&next->upper))
2546                                 break;
2547
2548                         edge = list_entry(next->upper.next,
2549                                           struct backref_edge, list[LOWER]);
2550                         edges[index++] = edge;
2551                         next = edge->node[UPPER];
2552                 }
2553                 next = walk_down_backref(edges, &index);
2554         }
2555         return num_bytes;
2556 }
2557
2558 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2559                                   struct reloc_control *rc,
2560                                   struct backref_node *node)
2561 {
2562         struct btrfs_root *root = rc->extent_root;
2563         u64 num_bytes;
2564         int ret;
2565
2566         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2567
2568         trans->block_rsv = rc->block_rsv;
2569         ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2570                                   BTRFS_RESERVE_FLUSH_ALL);
2571         if (ret) {
2572                 if (ret == -EAGAIN)
2573                         rc->commit_transaction = 1;
2574                 return ret;
2575         }
2576
2577         return 0;
2578 }
2579
2580 static void release_metadata_space(struct reloc_control *rc,
2581                                    struct backref_node *node)
2582 {
2583         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2584         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2585 }
2586
2587 /*
2588  * relocate a block tree, and then update pointers in upper level
2589  * blocks that reference the block to point to the new location.
2590  *
2591  * if called by link_to_upper, the block has already been relocated.
2592  * in that case this function just updates pointers.
2593  */
2594 static int do_relocation(struct btrfs_trans_handle *trans,
2595                          struct reloc_control *rc,
2596                          struct backref_node *node,
2597                          struct btrfs_key *key,
2598                          struct btrfs_path *path, int lowest)
2599 {
2600         struct backref_node *upper;
2601         struct backref_edge *edge;
2602         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2603         struct btrfs_root *root;
2604         struct extent_buffer *eb;
2605         u32 blocksize;
2606         u64 bytenr;
2607         u64 generation;
2608         int nr;
2609         int slot;
2610         int ret;
2611         int err = 0;
2612
2613         BUG_ON(lowest && node->eb);
2614
2615         path->lowest_level = node->level + 1;
2616         rc->backref_cache.path[node->level] = node;
2617         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2618                 cond_resched();
2619
2620                 upper = edge->node[UPPER];
2621                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2622                 BUG_ON(!root);
2623
2624                 if (upper->eb && !upper->locked) {
2625                         if (!lowest) {
2626                                 ret = btrfs_bin_search(upper->eb, key,
2627                                                        upper->level, &slot);
2628                                 BUG_ON(ret);
2629                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2630                                 if (node->eb->start == bytenr)
2631                                         goto next;
2632                         }
2633                         drop_node_buffer(upper);
2634                 }
2635
2636                 if (!upper->eb) {
2637                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2638                         if (ret < 0) {
2639                                 err = ret;
2640                                 break;
2641                         }
2642                         BUG_ON(ret > 0);
2643
2644                         if (!upper->eb) {
2645                                 upper->eb = path->nodes[upper->level];
2646                                 path->nodes[upper->level] = NULL;
2647                         } else {
2648                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2649                         }
2650
2651                         upper->locked = 1;
2652                         path->locks[upper->level] = 0;
2653
2654                         slot = path->slots[upper->level];
2655                         btrfs_release_path(path);
2656                 } else {
2657                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2658                                                &slot);
2659                         BUG_ON(ret);
2660                 }
2661
2662                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2663                 if (lowest) {
2664                         BUG_ON(bytenr != node->bytenr);
2665                 } else {
2666                         if (node->eb->start == bytenr)
2667                                 goto next;
2668                 }
2669
2670                 blocksize = btrfs_level_size(root, node->level);
2671                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2672                 eb = read_tree_block(root, bytenr, blocksize, generation);
2673                 if (!eb || !extent_buffer_uptodate(eb)) {
2674                         free_extent_buffer(eb);
2675                         err = -EIO;
2676                         goto next;
2677                 }
2678                 btrfs_tree_lock(eb);
2679                 btrfs_set_lock_blocking(eb);
2680
2681                 if (!node->eb) {
2682                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2683                                               slot, &eb);
2684                         btrfs_tree_unlock(eb);
2685                         free_extent_buffer(eb);
2686                         if (ret < 0) {
2687                                 err = ret;
2688                                 goto next;
2689                         }
2690                         BUG_ON(node->eb != eb);
2691                 } else {
2692                         btrfs_set_node_blockptr(upper->eb, slot,
2693                                                 node->eb->start);
2694                         btrfs_set_node_ptr_generation(upper->eb, slot,
2695                                                       trans->transid);
2696                         btrfs_mark_buffer_dirty(upper->eb);
2697
2698                         ret = btrfs_inc_extent_ref(trans, root,
2699                                                 node->eb->start, blocksize,
2700                                                 upper->eb->start,
2701                                                 btrfs_header_owner(upper->eb),
2702                                                 node->level, 0, 1);
2703                         BUG_ON(ret);
2704
2705                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2706                         BUG_ON(ret);
2707                 }
2708 next:
2709                 if (!upper->pending)
2710                         drop_node_buffer(upper);
2711                 else
2712                         unlock_node_buffer(upper);
2713                 if (err)
2714                         break;
2715         }
2716
2717         if (!err && node->pending) {
2718                 drop_node_buffer(node);
2719                 list_move_tail(&node->list, &rc->backref_cache.changed);
2720                 node->pending = 0;
2721         }
2722
2723         path->lowest_level = 0;
2724         BUG_ON(err == -ENOSPC);
2725         return err;
2726 }
2727
2728 static int link_to_upper(struct btrfs_trans_handle *trans,
2729                          struct reloc_control *rc,
2730                          struct backref_node *node,
2731                          struct btrfs_path *path)
2732 {
2733         struct btrfs_key key;
2734
2735         btrfs_node_key_to_cpu(node->eb, &key, 0);
2736         return do_relocation(trans, rc, node, &key, path, 0);
2737 }
2738
2739 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2740                                 struct reloc_control *rc,
2741                                 struct btrfs_path *path, int err)
2742 {
2743         LIST_HEAD(list);
2744         struct backref_cache *cache = &rc->backref_cache;
2745         struct backref_node *node;
2746         int level;
2747         int ret;
2748
2749         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2750                 while (!list_empty(&cache->pending[level])) {
2751                         node = list_entry(cache->pending[level].next,
2752                                           struct backref_node, list);
2753                         list_move_tail(&node->list, &list);
2754                         BUG_ON(!node->pending);
2755
2756                         if (!err) {
2757                                 ret = link_to_upper(trans, rc, node, path);
2758                                 if (ret < 0)
2759                                         err = ret;
2760                         }
2761                 }
2762                 list_splice_init(&list, &cache->pending[level]);
2763         }
2764         return err;
2765 }
2766
2767 static void mark_block_processed(struct reloc_control *rc,
2768                                  u64 bytenr, u32 blocksize)
2769 {
2770         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2771                         EXTENT_DIRTY, GFP_NOFS);
2772 }
2773
2774 static void __mark_block_processed(struct reloc_control *rc,
2775                                    struct backref_node *node)
2776 {
2777         u32 blocksize;
2778         if (node->level == 0 ||
2779             in_block_group(node->bytenr, rc->block_group)) {
2780                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2781                 mark_block_processed(rc, node->bytenr, blocksize);
2782         }
2783         node->processed = 1;
2784 }
2785
2786 /*
2787  * mark a block and all blocks directly/indirectly reference the block
2788  * as processed.
2789  */
2790 static void update_processed_blocks(struct reloc_control *rc,
2791                                     struct backref_node *node)
2792 {
2793         struct backref_node *next = node;
2794         struct backref_edge *edge;
2795         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2796         int index = 0;
2797
2798         while (next) {
2799                 cond_resched();
2800                 while (1) {
2801                         if (next->processed)
2802                                 break;
2803
2804                         __mark_block_processed(rc, next);
2805
2806                         if (list_empty(&next->upper))
2807                                 break;
2808
2809                         edge = list_entry(next->upper.next,
2810                                           struct backref_edge, list[LOWER]);
2811                         edges[index++] = edge;
2812                         next = edge->node[UPPER];
2813                 }
2814                 next = walk_down_backref(edges, &index);
2815         }
2816 }
2817
2818 static int tree_block_processed(u64 bytenr, u32 blocksize,
2819                                 struct reloc_control *rc)
2820 {
2821         if (test_range_bit(&rc->processed_blocks, bytenr,
2822                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2823                 return 1;
2824         return 0;
2825 }
2826
2827 static int get_tree_block_key(struct reloc_control *rc,
2828                               struct tree_block *block)
2829 {
2830         struct extent_buffer *eb;
2831
2832         BUG_ON(block->key_ready);
2833         eb = read_tree_block(rc->extent_root, block->bytenr,
2834                              block->key.objectid, block->key.offset);
2835         if (!eb || !extent_buffer_uptodate(eb)) {
2836                 free_extent_buffer(eb);
2837                 return -EIO;
2838         }
2839         WARN_ON(btrfs_header_level(eb) != block->level);
2840         if (block->level == 0)
2841                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2842         else
2843                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2844         free_extent_buffer(eb);
2845         block->key_ready = 1;
2846         return 0;
2847 }
2848
2849 static int reada_tree_block(struct reloc_control *rc,
2850                             struct tree_block *block)
2851 {
2852         BUG_ON(block->key_ready);
2853         if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2854                 readahead_tree_block(rc->extent_root, block->bytenr,
2855                                      block->key.objectid,
2856                                      rc->extent_root->leafsize);
2857         else
2858                 readahead_tree_block(rc->extent_root, block->bytenr,
2859                                      block->key.objectid, block->key.offset);
2860         return 0;
2861 }
2862
2863 /*
2864  * helper function to relocate a tree block
2865  */
2866 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2867                                 struct reloc_control *rc,
2868                                 struct backref_node *node,
2869                                 struct btrfs_key *key,
2870                                 struct btrfs_path *path)
2871 {
2872         struct btrfs_root *root;
2873         int release = 0;
2874         int ret = 0;
2875
2876         if (!node)
2877                 return 0;
2878
2879         BUG_ON(node->processed);
2880         root = select_one_root(trans, node);
2881         if (root == ERR_PTR(-ENOENT)) {
2882                 update_processed_blocks(rc, node);
2883                 goto out;
2884         }
2885
2886         if (!root || root->ref_cows) {
2887                 ret = reserve_metadata_space(trans, rc, node);
2888                 if (ret)
2889                         goto out;
2890                 release = 1;
2891         }
2892
2893         if (root) {
2894                 if (root->ref_cows) {
2895                         BUG_ON(node->new_bytenr);
2896                         BUG_ON(!list_empty(&node->list));
2897                         btrfs_record_root_in_trans(trans, root);
2898                         root = root->reloc_root;
2899                         node->new_bytenr = root->node->start;
2900                         node->root = root;
2901                         list_add_tail(&node->list, &rc->backref_cache.changed);
2902                 } else {
2903                         path->lowest_level = node->level;
2904                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2905                         btrfs_release_path(path);
2906                         if (ret > 0)
2907                                 ret = 0;
2908                 }
2909                 if (!ret)
2910                         update_processed_blocks(rc, node);
2911         } else {
2912                 ret = do_relocation(trans, rc, node, key, path, 1);
2913         }
2914 out:
2915         if (ret || node->level == 0 || node->cowonly) {
2916                 if (release)
2917                         release_metadata_space(rc, node);
2918                 remove_backref_node(&rc->backref_cache, node);
2919         }
2920         return ret;
2921 }
2922
2923 /*
2924  * relocate a list of blocks
2925  */
2926 static noinline_for_stack
2927 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2928                          struct reloc_control *rc, struct rb_root *blocks)
2929 {
2930         struct backref_node *node;
2931         struct btrfs_path *path;
2932         struct tree_block *block;
2933         struct rb_node *rb_node;
2934         int ret;
2935         int err = 0;
2936
2937         path = btrfs_alloc_path();
2938         if (!path) {
2939                 err = -ENOMEM;
2940                 goto out_free_blocks;
2941         }
2942
2943         rb_node = rb_first(blocks);
2944         while (rb_node) {
2945                 block = rb_entry(rb_node, struct tree_block, rb_node);
2946                 if (!block->key_ready)
2947                         reada_tree_block(rc, block);
2948                 rb_node = rb_next(rb_node);
2949         }
2950
2951         rb_node = rb_first(blocks);
2952         while (rb_node) {
2953                 block = rb_entry(rb_node, struct tree_block, rb_node);
2954                 if (!block->key_ready) {
2955                         err = get_tree_block_key(rc, block);
2956                         if (err)
2957                                 goto out_free_path;
2958                 }
2959                 rb_node = rb_next(rb_node);
2960         }
2961
2962         rb_node = rb_first(blocks);
2963         while (rb_node) {
2964                 block = rb_entry(rb_node, struct tree_block, rb_node);
2965
2966                 node = build_backref_tree(rc, &block->key,
2967                                           block->level, block->bytenr);
2968                 if (IS_ERR(node)) {
2969                         err = PTR_ERR(node);
2970                         goto out;
2971                 }
2972
2973                 ret = relocate_tree_block(trans, rc, node, &block->key,
2974                                           path);
2975                 if (ret < 0) {
2976                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2977                                 err = ret;
2978                         goto out;
2979                 }
2980                 rb_node = rb_next(rb_node);
2981         }
2982 out:
2983         err = finish_pending_nodes(trans, rc, path, err);
2984
2985 out_free_path:
2986         btrfs_free_path(path);
2987 out_free_blocks:
2988         free_block_list(blocks);
2989         return err;
2990 }
2991
2992 static noinline_for_stack
2993 int prealloc_file_extent_cluster(struct inode *inode,
2994                                  struct file_extent_cluster *cluster)
2995 {
2996         u64 alloc_hint = 0;
2997         u64 start;
2998         u64 end;
2999         u64 offset = BTRFS_I(inode)->index_cnt;
3000         u64 num_bytes;
3001         int nr = 0;
3002         int ret = 0;
3003
3004         BUG_ON(cluster->start != cluster->boundary[0]);
3005         mutex_lock(&inode->i_mutex);
3006
3007         ret = btrfs_check_data_free_space(inode, cluster->end +
3008                                           1 - cluster->start);
3009         if (ret)
3010                 goto out;
3011
3012         while (nr < cluster->nr) {
3013                 start = cluster->boundary[nr] - offset;
3014                 if (nr + 1 < cluster->nr)
3015                         end = cluster->boundary[nr + 1] - 1 - offset;
3016                 else
3017                         end = cluster->end - offset;
3018
3019                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3020                 num_bytes = end + 1 - start;
3021                 ret = btrfs_prealloc_file_range(inode, 0, start,
3022                                                 num_bytes, num_bytes,
3023                                                 end + 1, &alloc_hint);
3024                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3025                 if (ret)
3026                         break;
3027                 nr++;
3028         }
3029         btrfs_free_reserved_data_space(inode, cluster->end +
3030                                        1 - cluster->start);
3031 out:
3032         mutex_unlock(&inode->i_mutex);
3033         return ret;
3034 }
3035
3036 static noinline_for_stack
3037 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3038                          u64 block_start)
3039 {
3040         struct btrfs_root *root = BTRFS_I(inode)->root;
3041         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3042         struct extent_map *em;
3043         int ret = 0;
3044
3045         em = alloc_extent_map();
3046         if (!em)
3047                 return -ENOMEM;
3048
3049         em->start = start;
3050         em->len = end + 1 - start;
3051         em->block_len = em->len;
3052         em->block_start = block_start;
3053         em->bdev = root->fs_info->fs_devices->latest_bdev;
3054         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3055
3056         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3057         while (1) {
3058                 write_lock(&em_tree->lock);
3059                 ret = add_extent_mapping(em_tree, em, 0);
3060                 write_unlock(&em_tree->lock);
3061                 if (ret != -EEXIST) {
3062                         free_extent_map(em);
3063                         break;
3064                 }
3065                 btrfs_drop_extent_cache(inode, start, end, 0);
3066         }
3067         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3068         return ret;
3069 }
3070
3071 static int relocate_file_extent_cluster(struct inode *inode,
3072                                         struct file_extent_cluster *cluster)
3073 {
3074         u64 page_start;
3075         u64 page_end;
3076         u64 offset = BTRFS_I(inode)->index_cnt;
3077         unsigned long index;
3078         unsigned long last_index;
3079         struct page *page;
3080         struct file_ra_state *ra;
3081         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3082         int nr = 0;
3083         int ret = 0;
3084
3085         if (!cluster->nr)
3086                 return 0;
3087
3088         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3089         if (!ra)
3090                 return -ENOMEM;
3091
3092         ret = prealloc_file_extent_cluster(inode, cluster);
3093         if (ret)
3094                 goto out;
3095
3096         file_ra_state_init(ra, inode->i_mapping);
3097
3098         ret = setup_extent_mapping(inode, cluster->start - offset,
3099                                    cluster->end - offset, cluster->start);
3100         if (ret)
3101                 goto out;
3102
3103         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3104         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3105         while (index <= last_index) {
3106                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3107                 if (ret)
3108                         goto out;
3109
3110                 page = find_lock_page(inode->i_mapping, index);
3111                 if (!page) {
3112                         page_cache_sync_readahead(inode->i_mapping,
3113                                                   ra, NULL, index,
3114                                                   last_index + 1 - index);
3115                         page = find_or_create_page(inode->i_mapping, index,
3116                                                    mask);
3117                         if (!page) {
3118                                 btrfs_delalloc_release_metadata(inode,
3119                                                         PAGE_CACHE_SIZE);
3120                                 ret = -ENOMEM;
3121                                 goto out;
3122                         }
3123                 }
3124
3125                 if (PageReadahead(page)) {
3126                         page_cache_async_readahead(inode->i_mapping,
3127                                                    ra, NULL, page, index,
3128                                                    last_index + 1 - index);
3129                 }
3130
3131                 if (!PageUptodate(page)) {
3132                         btrfs_readpage(NULL, page);
3133                         lock_page(page);
3134                         if (!PageUptodate(page)) {
3135                                 unlock_page(page);
3136                                 page_cache_release(page);
3137                                 btrfs_delalloc_release_metadata(inode,
3138                                                         PAGE_CACHE_SIZE);
3139                                 ret = -EIO;
3140                                 goto out;
3141                         }
3142                 }
3143
3144                 page_start = page_offset(page);
3145                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3146
3147                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3148
3149                 set_page_extent_mapped(page);
3150
3151                 if (nr < cluster->nr &&
3152                     page_start + offset == cluster->boundary[nr]) {
3153                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3154                                         page_start, page_end,
3155                                         EXTENT_BOUNDARY, GFP_NOFS);
3156                         nr++;
3157                 }
3158
3159                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3160                 set_page_dirty(page);
3161
3162                 unlock_extent(&BTRFS_I(inode)->io_tree,
3163                               page_start, page_end);
3164                 unlock_page(page);
3165                 page_cache_release(page);
3166
3167                 index++;
3168                 balance_dirty_pages_ratelimited(inode->i_mapping);
3169                 btrfs_throttle(BTRFS_I(inode)->root);
3170         }
3171         WARN_ON(nr != cluster->nr);
3172 out:
3173         kfree(ra);
3174         return ret;
3175 }
3176
3177 static noinline_for_stack
3178 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3179                          struct file_extent_cluster *cluster)
3180 {
3181         int ret;
3182
3183         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3184                 ret = relocate_file_extent_cluster(inode, cluster);
3185                 if (ret)
3186                         return ret;
3187                 cluster->nr = 0;
3188         }
3189
3190         if (!cluster->nr)
3191                 cluster->start = extent_key->objectid;
3192         else
3193                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3194         cluster->end = extent_key->objectid + extent_key->offset - 1;
3195         cluster->boundary[cluster->nr] = extent_key->objectid;
3196         cluster->nr++;
3197
3198         if (cluster->nr >= MAX_EXTENTS) {
3199                 ret = relocate_file_extent_cluster(inode, cluster);
3200                 if (ret)
3201                         return ret;
3202                 cluster->nr = 0;
3203         }
3204         return 0;
3205 }
3206
3207 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3208 static int get_ref_objectid_v0(struct reloc_control *rc,
3209                                struct btrfs_path *path,
3210                                struct btrfs_key *extent_key,
3211                                u64 *ref_objectid, int *path_change)
3212 {
3213         struct btrfs_key key;
3214         struct extent_buffer *leaf;
3215         struct btrfs_extent_ref_v0 *ref0;
3216         int ret;
3217         int slot;
3218
3219         leaf = path->nodes[0];
3220         slot = path->slots[0];
3221         while (1) {
3222                 if (slot >= btrfs_header_nritems(leaf)) {
3223                         ret = btrfs_next_leaf(rc->extent_root, path);
3224                         if (ret < 0)
3225                                 return ret;
3226                         BUG_ON(ret > 0);
3227                         leaf = path->nodes[0];
3228                         slot = path->slots[0];
3229                         if (path_change)
3230                                 *path_change = 1;
3231                 }
3232                 btrfs_item_key_to_cpu(leaf, &key, slot);
3233                 if (key.objectid != extent_key->objectid)
3234                         return -ENOENT;
3235
3236                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3237                         slot++;
3238                         continue;
3239                 }
3240                 ref0 = btrfs_item_ptr(leaf, slot,
3241                                 struct btrfs_extent_ref_v0);
3242                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3243                 break;
3244         }
3245         return 0;
3246 }
3247 #endif
3248
3249 /*
3250  * helper to add a tree block to the list.
3251  * the major work is getting the generation and level of the block
3252  */
3253 static int add_tree_block(struct reloc_control *rc,
3254                           struct btrfs_key *extent_key,
3255                           struct btrfs_path *path,
3256                           struct rb_root *blocks)
3257 {
3258         struct extent_buffer *eb;
3259         struct btrfs_extent_item *ei;
3260         struct btrfs_tree_block_info *bi;
3261         struct tree_block *block;
3262         struct rb_node *rb_node;
3263         u32 item_size;
3264         int level = -1;
3265         u64 generation;
3266
3267         eb =  path->nodes[0];
3268         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3269
3270         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3271             item_size >= sizeof(*ei) + sizeof(*bi)) {
3272                 ei = btrfs_item_ptr(eb, path->slots[0],
3273                                 struct btrfs_extent_item);
3274                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3275                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3276                         level = btrfs_tree_block_level(eb, bi);
3277                 } else {
3278                         level = (int)extent_key->offset;
3279                 }
3280                 generation = btrfs_extent_generation(eb, ei);
3281         } else {
3282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3283                 u64 ref_owner;
3284                 int ret;
3285
3286                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3287                 ret = get_ref_objectid_v0(rc, path, extent_key,
3288                                           &ref_owner, NULL);
3289                 if (ret < 0)
3290                         return ret;
3291                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3292                 level = (int)ref_owner;
3293                 /* FIXME: get real generation */
3294                 generation = 0;
3295 #else
3296                 BUG();
3297 #endif
3298         }
3299
3300         btrfs_release_path(path);
3301
3302         BUG_ON(level == -1);
3303
3304         block = kmalloc(sizeof(*block), GFP_NOFS);
3305         if (!block)
3306                 return -ENOMEM;
3307
3308         block->bytenr = extent_key->objectid;
3309         block->key.objectid = rc->extent_root->leafsize;
3310         block->key.offset = generation;
3311         block->level = level;
3312         block->key_ready = 0;
3313
3314         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3315         if (rb_node)
3316                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3317
3318         return 0;
3319 }
3320
3321 /*
3322  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3323  */
3324 static int __add_tree_block(struct reloc_control *rc,
3325                             u64 bytenr, u32 blocksize,
3326                             struct rb_root *blocks)
3327 {
3328         struct btrfs_path *path;
3329         struct btrfs_key key;
3330         int ret;
3331         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3332                                         SKINNY_METADATA);
3333
3334         if (tree_block_processed(bytenr, blocksize, rc))
3335                 return 0;
3336
3337         if (tree_search(blocks, bytenr))
3338                 return 0;
3339
3340         path = btrfs_alloc_path();
3341         if (!path)
3342                 return -ENOMEM;
3343 again:
3344         key.objectid = bytenr;
3345         if (skinny) {
3346                 key.type = BTRFS_METADATA_ITEM_KEY;
3347                 key.offset = (u64)-1;
3348         } else {
3349                 key.type = BTRFS_EXTENT_ITEM_KEY;
3350                 key.offset = blocksize;
3351         }
3352
3353         path->search_commit_root = 1;
3354         path->skip_locking = 1;
3355         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3356         if (ret < 0)
3357                 goto out;
3358
3359         if (ret > 0 && skinny) {
3360                 if (path->slots[0]) {
3361                         path->slots[0]--;
3362                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3363                                               path->slots[0]);
3364                         if (key.objectid == bytenr &&
3365                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3366                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3367                               key.offset == blocksize)))
3368                                 ret = 0;
3369                 }
3370
3371                 if (ret) {
3372                         skinny = false;
3373                         btrfs_release_path(path);
3374                         goto again;
3375                 }
3376         }
3377         BUG_ON(ret);
3378
3379         ret = add_tree_block(rc, &key, path, blocks);
3380 out:
3381         btrfs_free_path(path);
3382         return ret;
3383 }
3384
3385 /*
3386  * helper to check if the block use full backrefs for pointers in it
3387  */
3388 static int block_use_full_backref(struct reloc_control *rc,
3389                                   struct extent_buffer *eb)
3390 {
3391         u64 flags;
3392         int ret;
3393
3394         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3395             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3396                 return 1;
3397
3398         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3399                                        eb->start, btrfs_header_level(eb), 1,
3400                                        NULL, &flags);
3401         BUG_ON(ret);
3402
3403         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3404                 ret = 1;
3405         else
3406                 ret = 0;
3407         return ret;
3408 }
3409
3410 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3411                                     struct inode *inode, u64 ino)
3412 {
3413         struct btrfs_key key;
3414         struct btrfs_root *root = fs_info->tree_root;
3415         struct btrfs_trans_handle *trans;
3416         int ret = 0;
3417
3418         if (inode)
3419                 goto truncate;
3420
3421         key.objectid = ino;
3422         key.type = BTRFS_INODE_ITEM_KEY;
3423         key.offset = 0;
3424
3425         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3426         if (IS_ERR(inode) || is_bad_inode(inode)) {
3427                 if (!IS_ERR(inode))
3428                         iput(inode);
3429                 return -ENOENT;
3430         }
3431
3432 truncate:
3433         ret = btrfs_check_trunc_cache_free_space(root,
3434                                                  &fs_info->global_block_rsv);
3435         if (ret)
3436                 goto out;
3437
3438         trans = btrfs_join_transaction(root);
3439         if (IS_ERR(trans)) {
3440                 ret = PTR_ERR(trans);
3441                 goto out;
3442         }
3443
3444         ret = btrfs_truncate_free_space_cache(root, trans, inode);
3445
3446         btrfs_end_transaction(trans, root);
3447         btrfs_btree_balance_dirty(root);
3448 out:
3449         iput(inode);
3450         return ret;
3451 }
3452
3453 /*
3454  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3455  * this function scans fs tree to find blocks reference the data extent
3456  */
3457 static int find_data_references(struct reloc_control *rc,
3458                                 struct btrfs_key *extent_key,
3459                                 struct extent_buffer *leaf,
3460                                 struct btrfs_extent_data_ref *ref,
3461                                 struct rb_root *blocks)
3462 {
3463         struct btrfs_path *path;
3464         struct tree_block *block;
3465         struct btrfs_root *root;
3466         struct btrfs_file_extent_item *fi;
3467         struct rb_node *rb_node;
3468         struct btrfs_key key;
3469         u64 ref_root;
3470         u64 ref_objectid;
3471         u64 ref_offset;
3472         u32 ref_count;
3473         u32 nritems;
3474         int err = 0;
3475         int added = 0;
3476         int counted;
3477         int ret;
3478
3479         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3480         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3481         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3482         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3483
3484         /*
3485          * This is an extent belonging to the free space cache, lets just delete
3486          * it and redo the search.
3487          */
3488         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3489                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3490                                                NULL, ref_objectid);
3491                 if (ret != -ENOENT)
3492                         return ret;
3493                 ret = 0;
3494         }
3495
3496         path = btrfs_alloc_path();
3497         if (!path)
3498                 return -ENOMEM;
3499         path->reada = 1;
3500
3501         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3502         if (IS_ERR(root)) {
3503                 err = PTR_ERR(root);
3504                 goto out;
3505         }
3506
3507         key.objectid = ref_objectid;
3508         key.type = BTRFS_EXTENT_DATA_KEY;
3509         if (ref_offset > ((u64)-1 << 32))
3510                 key.offset = 0;
3511         else
3512                 key.offset = ref_offset;
3513
3514         path->search_commit_root = 1;
3515         path->skip_locking = 1;
3516         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3517         if (ret < 0) {
3518                 err = ret;
3519                 goto out;
3520         }
3521
3522         leaf = path->nodes[0];
3523         nritems = btrfs_header_nritems(leaf);
3524         /*
3525          * the references in tree blocks that use full backrefs
3526          * are not counted in
3527          */
3528         if (block_use_full_backref(rc, leaf))
3529                 counted = 0;
3530         else
3531                 counted = 1;
3532         rb_node = tree_search(blocks, leaf->start);
3533         if (rb_node) {
3534                 if (counted)
3535                         added = 1;
3536                 else
3537                         path->slots[0] = nritems;
3538         }
3539
3540         while (ref_count > 0) {
3541                 while (path->slots[0] >= nritems) {
3542                         ret = btrfs_next_leaf(root, path);
3543                         if (ret < 0) {
3544                                 err = ret;
3545                                 goto out;
3546                         }
3547                         if (WARN_ON(ret > 0))
3548                                 goto out;
3549
3550                         leaf = path->nodes[0];
3551                         nritems = btrfs_header_nritems(leaf);
3552                         added = 0;
3553
3554                         if (block_use_full_backref(rc, leaf))
3555                                 counted = 0;
3556                         else
3557                                 counted = 1;
3558                         rb_node = tree_search(blocks, leaf->start);
3559                         if (rb_node) {
3560                                 if (counted)
3561                                         added = 1;
3562                                 else
3563                                         path->slots[0] = nritems;
3564                         }
3565                 }
3566
3567                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3568                 if (WARN_ON(key.objectid != ref_objectid ||
3569                     key.type != BTRFS_EXTENT_DATA_KEY))
3570                         break;
3571
3572                 fi = btrfs_item_ptr(leaf, path->slots[0],
3573                                     struct btrfs_file_extent_item);
3574
3575                 if (btrfs_file_extent_type(leaf, fi) ==
3576                     BTRFS_FILE_EXTENT_INLINE)
3577                         goto next;
3578
3579                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3580                     extent_key->objectid)
3581                         goto next;
3582
3583                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3584                 if (key.offset != ref_offset)
3585                         goto next;
3586
3587                 if (counted)
3588                         ref_count--;
3589                 if (added)
3590                         goto next;
3591
3592                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3593                         block = kmalloc(sizeof(*block), GFP_NOFS);
3594                         if (!block) {
3595                                 err = -ENOMEM;
3596                                 break;
3597                         }
3598                         block->bytenr = leaf->start;
3599                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3600                         block->level = 0;
3601                         block->key_ready = 1;
3602                         rb_node = tree_insert(blocks, block->bytenr,
3603                                               &block->rb_node);
3604                         if (rb_node)
3605                                 backref_tree_panic(rb_node, -EEXIST,
3606                                                    block->bytenr);
3607                 }
3608                 if (counted)
3609                         added = 1;
3610                 else
3611                         path->slots[0] = nritems;
3612 next:
3613                 path->slots[0]++;
3614
3615         }
3616 out:
3617         btrfs_free_path(path);
3618         return err;
3619 }
3620
3621 /*
3622  * helper to find all tree blocks that reference a given data extent
3623  */
3624 static noinline_for_stack
3625 int add_data_references(struct reloc_control *rc,
3626                         struct btrfs_key *extent_key,
3627                         struct btrfs_path *path,
3628                         struct rb_root *blocks)
3629 {
3630         struct btrfs_key key;
3631         struct extent_buffer *eb;
3632         struct btrfs_extent_data_ref *dref;
3633         struct btrfs_extent_inline_ref *iref;
3634         unsigned long ptr;
3635         unsigned long end;
3636         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3637         int ret = 0;
3638         int err = 0;
3639
3640         eb = path->nodes[0];
3641         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3642         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3643 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3644         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3645                 ptr = end;
3646         else
3647 #endif
3648                 ptr += sizeof(struct btrfs_extent_item);
3649
3650         while (ptr < end) {
3651                 iref = (struct btrfs_extent_inline_ref *)ptr;
3652                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3653                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3654                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3655                         ret = __add_tree_block(rc, key.offset, blocksize,
3656                                                blocks);
3657                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3658                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3659                         ret = find_data_references(rc, extent_key,
3660                                                    eb, dref, blocks);
3661                 } else {
3662                         BUG();
3663                 }
3664                 if (ret) {
3665                         err = ret;
3666                         goto out;
3667                 }
3668                 ptr += btrfs_extent_inline_ref_size(key.type);
3669         }
3670         WARN_ON(ptr > end);
3671
3672         while (1) {
3673                 cond_resched();
3674                 eb = path->nodes[0];
3675                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3676                         ret = btrfs_next_leaf(rc->extent_root, path);
3677                         if (ret < 0) {
3678                                 err = ret;
3679                                 break;
3680                         }
3681                         if (ret > 0)
3682                                 break;
3683                         eb = path->nodes[0];
3684                 }
3685
3686                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3687                 if (key.objectid != extent_key->objectid)
3688                         break;
3689
3690 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3691                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3692                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3693 #else
3694                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3695                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3696 #endif
3697                         ret = __add_tree_block(rc, key.offset, blocksize,
3698                                                blocks);
3699                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3700                         dref = btrfs_item_ptr(eb, path->slots[0],
3701                                               struct btrfs_extent_data_ref);
3702                         ret = find_data_references(rc, extent_key,
3703                                                    eb, dref, blocks);
3704                 } else {
3705                         ret = 0;
3706                 }
3707                 if (ret) {
3708                         err = ret;
3709                         break;
3710                 }
3711                 path->slots[0]++;
3712         }
3713 out:
3714         btrfs_release_path(path);
3715         if (err)
3716                 free_block_list(blocks);
3717         return err;
3718 }
3719
3720 /*
3721  * helper to find next unprocessed extent
3722  */
3723 static noinline_for_stack
3724 int find_next_extent(struct btrfs_trans_handle *trans,
3725                      struct reloc_control *rc, struct btrfs_path *path,
3726                      struct btrfs_key *extent_key)
3727 {
3728         struct btrfs_key key;
3729         struct extent_buffer *leaf;
3730         u64 start, end, last;
3731         int ret;
3732
3733         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3734         while (1) {
3735                 cond_resched();
3736                 if (rc->search_start >= last) {
3737                         ret = 1;
3738                         break;
3739                 }
3740
3741                 key.objectid = rc->search_start;
3742                 key.type = BTRFS_EXTENT_ITEM_KEY;
3743                 key.offset = 0;
3744
3745                 path->search_commit_root = 1;
3746                 path->skip_locking = 1;
3747                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3748                                         0, 0);
3749                 if (ret < 0)
3750                         break;
3751 next:
3752                 leaf = path->nodes[0];
3753                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3754                         ret = btrfs_next_leaf(rc->extent_root, path);
3755                         if (ret != 0)
3756                                 break;
3757                         leaf = path->nodes[0];
3758                 }
3759
3760                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3761                 if (key.objectid >= last) {
3762                         ret = 1;
3763                         break;
3764                 }
3765
3766                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3767                     key.type != BTRFS_METADATA_ITEM_KEY) {
3768                         path->slots[0]++;
3769                         goto next;
3770                 }
3771
3772                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3773                     key.objectid + key.offset <= rc->search_start) {
3774                         path->slots[0]++;
3775                         goto next;
3776                 }
3777
3778                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3779                     key.objectid + rc->extent_root->leafsize <=
3780                     rc->search_start) {
3781                         path->slots[0]++;
3782                         goto next;
3783                 }
3784
3785                 ret = find_first_extent_bit(&rc->processed_blocks,
3786                                             key.objectid, &start, &end,
3787                                             EXTENT_DIRTY, NULL);
3788
3789                 if (ret == 0 && start <= key.objectid) {
3790                         btrfs_release_path(path);
3791                         rc->search_start = end + 1;
3792                 } else {
3793                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3794                                 rc->search_start = key.objectid + key.offset;
3795                         else
3796                                 rc->search_start = key.objectid +
3797                                         rc->extent_root->leafsize;
3798                         memcpy(extent_key, &key, sizeof(key));
3799                         return 0;
3800                 }
3801         }
3802         btrfs_release_path(path);
3803         return ret;
3804 }
3805
3806 static void set_reloc_control(struct reloc_control *rc)
3807 {
3808         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3809
3810         mutex_lock(&fs_info->reloc_mutex);
3811         fs_info->reloc_ctl = rc;
3812         mutex_unlock(&fs_info->reloc_mutex);
3813 }
3814
3815 static void unset_reloc_control(struct reloc_control *rc)
3816 {
3817         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3818
3819         mutex_lock(&fs_info->reloc_mutex);
3820         fs_info->reloc_ctl = NULL;
3821         mutex_unlock(&fs_info->reloc_mutex);
3822 }
3823
3824 static int check_extent_flags(u64 flags)
3825 {
3826         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3827             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3828                 return 1;
3829         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3830             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3831                 return 1;
3832         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3833             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3834                 return 1;
3835         return 0;
3836 }
3837
3838 static noinline_for_stack
3839 int prepare_to_relocate(struct reloc_control *rc)
3840 {
3841         struct btrfs_trans_handle *trans;
3842         int ret;
3843
3844         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3845                                               BTRFS_BLOCK_RSV_TEMP);
3846         if (!rc->block_rsv)
3847                 return -ENOMEM;
3848
3849         /*
3850          * reserve some space for creating reloc trees.
3851          * btrfs_init_reloc_root will use them when there
3852          * is no reservation in transaction handle.
3853          */
3854         ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3855                                   rc->extent_root->nodesize * 256,
3856                                   BTRFS_RESERVE_FLUSH_ALL);
3857         if (ret)
3858                 return ret;
3859
3860         memset(&rc->cluster, 0, sizeof(rc->cluster));
3861         rc->search_start = rc->block_group->key.objectid;
3862         rc->extents_found = 0;
3863         rc->nodes_relocated = 0;
3864         rc->merging_rsv_size = 0;
3865
3866         rc->create_reloc_tree = 1;
3867         set_reloc_control(rc);
3868
3869         trans = btrfs_join_transaction(rc->extent_root);
3870         if (IS_ERR(trans)) {
3871                 unset_reloc_control(rc);
3872                 /*
3873                  * extent tree is not a ref_cow tree and has no reloc_root to
3874                  * cleanup.  And callers are responsible to free the above
3875                  * block rsv.
3876                  */
3877                 return PTR_ERR(trans);
3878         }
3879         btrfs_commit_transaction(trans, rc->extent_root);
3880         return 0;
3881 }
3882
3883 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3884 {
3885         struct rb_root blocks = RB_ROOT;
3886         struct btrfs_key key;
3887         struct btrfs_trans_handle *trans = NULL;
3888         struct btrfs_path *path;
3889         struct btrfs_extent_item *ei;
3890         u64 flags;
3891         u32 item_size;
3892         int ret;
3893         int err = 0;
3894         int progress = 0;
3895
3896         path = btrfs_alloc_path();
3897         if (!path)
3898                 return -ENOMEM;
3899         path->reada = 1;
3900
3901         ret = prepare_to_relocate(rc);
3902         if (ret) {
3903                 err = ret;
3904                 goto out_free;
3905         }
3906
3907         while (1) {
3908                 progress++;
3909                 trans = btrfs_start_transaction(rc->extent_root, 0);
3910                 if (IS_ERR(trans)) {
3911                         err = PTR_ERR(trans);
3912                         trans = NULL;
3913                         break;
3914                 }
3915 restart:
3916                 if (update_backref_cache(trans, &rc->backref_cache)) {
3917                         btrfs_end_transaction(trans, rc->extent_root);
3918                         continue;
3919                 }
3920
3921                 ret = find_next_extent(trans, rc, path, &key);
3922                 if (ret < 0)
3923                         err = ret;
3924                 if (ret != 0)
3925                         break;
3926
3927                 rc->extents_found++;
3928
3929                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3930                                     struct btrfs_extent_item);
3931                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3932                 if (item_size >= sizeof(*ei)) {
3933                         flags = btrfs_extent_flags(path->nodes[0], ei);
3934                         ret = check_extent_flags(flags);
3935                         BUG_ON(ret);
3936
3937                 } else {
3938 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3939                         u64 ref_owner;
3940                         int path_change = 0;
3941
3942                         BUG_ON(item_size !=
3943                                sizeof(struct btrfs_extent_item_v0));
3944                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3945                                                   &path_change);
3946                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3947                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3948                         else
3949                                 flags = BTRFS_EXTENT_FLAG_DATA;
3950
3951                         if (path_change) {
3952                                 btrfs_release_path(path);
3953
3954                                 path->search_commit_root = 1;
3955                                 path->skip_locking = 1;
3956                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3957                                                         &key, path, 0, 0);
3958                                 if (ret < 0) {
3959                                         err = ret;
3960                                         break;
3961                                 }
3962                                 BUG_ON(ret > 0);
3963                         }
3964 #else
3965                         BUG();
3966 #endif
3967                 }
3968
3969                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3970                         ret = add_tree_block(rc, &key, path, &blocks);
3971                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3972                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3973                         ret = add_data_references(rc, &key, path, &blocks);
3974                 } else {
3975                         btrfs_release_path(path);
3976                         ret = 0;
3977                 }
3978                 if (ret < 0) {
3979                         err = ret;
3980                         break;
3981                 }
3982
3983                 if (!RB_EMPTY_ROOT(&blocks)) {
3984                         ret = relocate_tree_blocks(trans, rc, &blocks);
3985                         if (ret < 0) {
3986                                 if (ret != -EAGAIN) {
3987                                         err = ret;
3988                                         break;
3989                                 }
3990                                 rc->extents_found--;
3991                                 rc->search_start = key.objectid;
3992                         }
3993                 }
3994
3995                 if (rc->commit_transaction) {
3996                         rc->commit_transaction = 0;
3997                         ret = btrfs_commit_transaction(trans, rc->extent_root);
3998                         BUG_ON(ret);
3999                 } else {
4000                         btrfs_end_transaction_throttle(trans, rc->extent_root);
4001                         btrfs_btree_balance_dirty(rc->extent_root);
4002                 }
4003                 trans = NULL;
4004
4005                 if (rc->stage == MOVE_DATA_EXTENTS &&
4006                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4007                         rc->found_file_extent = 1;
4008                         ret = relocate_data_extent(rc->data_inode,
4009                                                    &key, &rc->cluster);
4010                         if (ret < 0) {
4011                                 err = ret;
4012                                 break;
4013                         }
4014                 }
4015         }
4016         if (trans && progress && err == -ENOSPC) {
4017                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4018                                               rc->block_group->flags);
4019                 if (ret == 0) {
4020                         err = 0;
4021                         progress = 0;
4022                         goto restart;
4023                 }
4024         }
4025
4026         btrfs_release_path(path);
4027         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4028                           GFP_NOFS);
4029
4030         if (trans) {
4031                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4032                 btrfs_btree_balance_dirty(rc->extent_root);
4033         }
4034
4035         if (!err) {
4036                 ret = relocate_file_extent_cluster(rc->data_inode,
4037                                                    &rc->cluster);
4038                 if (ret < 0)
4039                         err = ret;
4040         }
4041
4042         rc->create_reloc_tree = 0;
4043         set_reloc_control(rc);
4044
4045         backref_cache_cleanup(&rc->backref_cache);
4046         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4047
4048         err = prepare_to_merge(rc, err);
4049
4050         merge_reloc_roots(rc);
4051
4052         rc->merge_reloc_tree = 0;
4053         unset_reloc_control(rc);
4054         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4055
4056         /* get rid of pinned extents */
4057         trans = btrfs_join_transaction(rc->extent_root);
4058         if (IS_ERR(trans))
4059                 err = PTR_ERR(trans);
4060         else
4061                 btrfs_commit_transaction(trans, rc->extent_root);
4062 out_free:
4063         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4064         btrfs_free_path(path);
4065         return err;
4066 }
4067
4068 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4069                                  struct btrfs_root *root, u64 objectid)
4070 {
4071         struct btrfs_path *path;
4072         struct btrfs_inode_item *item;
4073         struct extent_buffer *leaf;
4074         int ret;
4075
4076         path = btrfs_alloc_path();
4077         if (!path)
4078                 return -ENOMEM;
4079
4080         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4081         if (ret)
4082                 goto out;
4083
4084         leaf = path->nodes[0];
4085         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4086         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4087         btrfs_set_inode_generation(leaf, item, 1);
4088         btrfs_set_inode_size(leaf, item, 0);
4089         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4090         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4091                                           BTRFS_INODE_PREALLOC);
4092         btrfs_mark_buffer_dirty(leaf);
4093         btrfs_release_path(path);
4094 out:
4095         btrfs_free_path(path);
4096         return ret;
4097 }
4098
4099 /*
4100  * helper to create inode for data relocation.
4101  * the inode is in data relocation tree and its link count is 0
4102  */
4103 static noinline_for_stack
4104 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4105                                  struct btrfs_block_group_cache *group)
4106 {
4107         struct inode *inode = NULL;
4108         struct btrfs_trans_handle *trans;
4109         struct btrfs_root *root;
4110         struct btrfs_key key;
4111         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4112         int err = 0;
4113
4114         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4115         if (IS_ERR(root))
4116                 return ERR_CAST(root);
4117
4118         trans = btrfs_start_transaction(root, 6);
4119         if (IS_ERR(trans))
4120                 return ERR_CAST(trans);
4121
4122         err = btrfs_find_free_objectid(root, &objectid);
4123         if (err)
4124                 goto out;
4125
4126         err = __insert_orphan_inode(trans, root, objectid);
4127         BUG_ON(err);
4128
4129         key.objectid = objectid;
4130         key.type = BTRFS_INODE_ITEM_KEY;
4131         key.offset = 0;
4132         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4133         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4134         BTRFS_I(inode)->index_cnt = group->key.objectid;
4135
4136         err = btrfs_orphan_add(trans, inode);
4137 out:
4138         btrfs_end_transaction(trans, root);
4139         btrfs_btree_balance_dirty(root);
4140         if (err) {
4141                 if (inode)
4142                         iput(inode);
4143                 inode = ERR_PTR(err);
4144         }
4145         return inode;
4146 }
4147
4148 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4149 {
4150         struct reloc_control *rc;
4151
4152         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4153         if (!rc)
4154                 return NULL;
4155
4156         INIT_LIST_HEAD(&rc->reloc_roots);
4157         backref_cache_init(&rc->backref_cache);
4158         mapping_tree_init(&rc->reloc_root_tree);
4159         extent_io_tree_init(&rc->processed_blocks,
4160                             fs_info->btree_inode->i_mapping);
4161         return rc;
4162 }
4163
4164 /*
4165  * function to relocate all extents in a block group.
4166  */
4167 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4168 {
4169         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4170         struct reloc_control *rc;
4171         struct inode *inode;
4172         struct btrfs_path *path;
4173         int ret;
4174         int rw = 0;
4175         int err = 0;
4176
4177         rc = alloc_reloc_control(fs_info);
4178         if (!rc)
4179                 return -ENOMEM;
4180
4181         rc->extent_root = extent_root;
4182
4183         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4184         BUG_ON(!rc->block_group);
4185
4186         if (!rc->block_group->ro) {
4187                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4188                 if (ret) {
4189                         err = ret;
4190                         goto out;
4191                 }
4192                 rw = 1;
4193         }
4194
4195         path = btrfs_alloc_path();
4196         if (!path) {
4197                 err = -ENOMEM;
4198                 goto out;
4199         }
4200
4201         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4202                                         path);
4203         btrfs_free_path(path);
4204
4205         if (!IS_ERR(inode))
4206                 ret = delete_block_group_cache(fs_info, inode, 0);
4207         else
4208                 ret = PTR_ERR(inode);
4209
4210         if (ret && ret != -ENOENT) {
4211                 err = ret;
4212                 goto out;
4213         }
4214
4215         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4216         if (IS_ERR(rc->data_inode)) {
4217                 err = PTR_ERR(rc->data_inode);
4218                 rc->data_inode = NULL;
4219                 goto out;
4220         }
4221
4222         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4223                rc->block_group->key.objectid, rc->block_group->flags);
4224
4225         ret = btrfs_start_delalloc_roots(fs_info, 0);
4226         if (ret < 0) {
4227                 err = ret;
4228                 goto out;
4229         }
4230         btrfs_wait_ordered_roots(fs_info, -1);
4231
4232         while (1) {
4233                 mutex_lock(&fs_info->cleaner_mutex);
4234                 ret = relocate_block_group(rc);
4235                 mutex_unlock(&fs_info->cleaner_mutex);
4236                 if (ret < 0) {
4237                         err = ret;
4238                         goto out;
4239                 }
4240
4241                 if (rc->extents_found == 0)
4242                         break;
4243
4244                 printk(KERN_INFO "btrfs: found %llu extents\n",
4245                         rc->extents_found);
4246
4247                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4248                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4249                                                        (u64)-1);
4250                         if (ret) {
4251                                 err = ret;
4252                                 goto out;
4253                         }
4254                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4255                                                  0, -1);
4256                         rc->stage = UPDATE_DATA_PTRS;
4257                 }
4258         }
4259
4260         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4261                                      rc->block_group->key.objectid,
4262                                      rc->block_group->key.objectid +
4263                                      rc->block_group->key.offset - 1);
4264
4265         WARN_ON(rc->block_group->pinned > 0);
4266         WARN_ON(rc->block_group->reserved > 0);
4267         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4268 out:
4269         if (err && rw)
4270                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4271         iput(rc->data_inode);
4272         btrfs_put_block_group(rc->block_group);
4273         kfree(rc);
4274         return err;
4275 }
4276
4277 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4278 {
4279         struct btrfs_trans_handle *trans;
4280         int ret, err;
4281
4282         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4283         if (IS_ERR(trans))
4284                 return PTR_ERR(trans);
4285
4286         memset(&root->root_item.drop_progress, 0,
4287                 sizeof(root->root_item.drop_progress));
4288         root->root_item.drop_level = 0;
4289         btrfs_set_root_refs(&root->root_item, 0);
4290         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4291                                 &root->root_key, &root->root_item);
4292
4293         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4294         if (err)
4295                 return err;
4296         return ret;
4297 }
4298
4299 /*
4300  * recover relocation interrupted by system crash.
4301  *
4302  * this function resumes merging reloc trees with corresponding fs trees.
4303  * this is important for keeping the sharing of tree blocks
4304  */
4305 int btrfs_recover_relocation(struct btrfs_root *root)
4306 {
4307         LIST_HEAD(reloc_roots);
4308         struct btrfs_key key;
4309         struct btrfs_root *fs_root;
4310         struct btrfs_root *reloc_root;
4311         struct btrfs_path *path;
4312         struct extent_buffer *leaf;
4313         struct reloc_control *rc = NULL;
4314         struct btrfs_trans_handle *trans;
4315         int ret;
4316         int err = 0;
4317
4318         path = btrfs_alloc_path();
4319         if (!path)
4320                 return -ENOMEM;
4321         path->reada = -1;
4322
4323         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4324         key.type = BTRFS_ROOT_ITEM_KEY;
4325         key.offset = (u64)-1;
4326
4327         while (1) {
4328                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4329                                         path, 0, 0);
4330                 if (ret < 0) {
4331                         err = ret;
4332                         goto out;
4333                 }
4334                 if (ret > 0) {
4335                         if (path->slots[0] == 0)
4336                                 break;
4337                         path->slots[0]--;
4338                 }
4339                 leaf = path->nodes[0];
4340                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4341                 btrfs_release_path(path);
4342
4343                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4344                     key.type != BTRFS_ROOT_ITEM_KEY)
4345                         break;
4346
4347                 reloc_root = btrfs_read_fs_root(root, &key);
4348                 if (IS_ERR(reloc_root)) {
4349                         err = PTR_ERR(reloc_root);
4350                         goto out;
4351                 }
4352
4353                 list_add(&reloc_root->root_list, &reloc_roots);
4354
4355                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4356                         fs_root = read_fs_root(root->fs_info,
4357                                                reloc_root->root_key.offset);
4358                         if (IS_ERR(fs_root)) {
4359                                 ret = PTR_ERR(fs_root);
4360                                 if (ret != -ENOENT) {
4361                                         err = ret;
4362                                         goto out;
4363                                 }
4364                                 ret = mark_garbage_root(reloc_root);
4365                                 if (ret < 0) {
4366                                         err = ret;
4367                                         goto out;
4368                                 }
4369                         }
4370                 }
4371
4372                 if (key.offset == 0)
4373                         break;
4374
4375                 key.offset--;
4376         }
4377         btrfs_release_path(path);
4378
4379         if (list_empty(&reloc_roots))
4380                 goto out;
4381
4382         rc = alloc_reloc_control(root->fs_info);
4383         if (!rc) {
4384                 err = -ENOMEM;
4385                 goto out;
4386         }
4387
4388         rc->extent_root = root->fs_info->extent_root;
4389
4390         set_reloc_control(rc);
4391
4392         trans = btrfs_join_transaction(rc->extent_root);
4393         if (IS_ERR(trans)) {
4394                 unset_reloc_control(rc);
4395                 err = PTR_ERR(trans);
4396                 goto out_free;
4397         }
4398
4399         rc->merge_reloc_tree = 1;
4400
4401         while (!list_empty(&reloc_roots)) {
4402                 reloc_root = list_entry(reloc_roots.next,
4403                                         struct btrfs_root, root_list);
4404                 list_del(&reloc_root->root_list);
4405
4406                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4407                         list_add_tail(&reloc_root->root_list,
4408                                       &rc->reloc_roots);
4409                         continue;
4410                 }
4411
4412                 fs_root = read_fs_root(root->fs_info,
4413                                        reloc_root->root_key.offset);
4414                 if (IS_ERR(fs_root)) {
4415                         err = PTR_ERR(fs_root);
4416                         goto out_free;
4417                 }
4418
4419                 err = __add_reloc_root(reloc_root);
4420                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4421                 fs_root->reloc_root = reloc_root;
4422         }
4423
4424         err = btrfs_commit_transaction(trans, rc->extent_root);
4425         if (err)
4426                 goto out_free;
4427
4428         merge_reloc_roots(rc);
4429
4430         unset_reloc_control(rc);
4431
4432         trans = btrfs_join_transaction(rc->extent_root);
4433         if (IS_ERR(trans))
4434                 err = PTR_ERR(trans);
4435         else
4436                 err = btrfs_commit_transaction(trans, rc->extent_root);
4437 out_free:
4438         kfree(rc);
4439 out:
4440         if (!list_empty(&reloc_roots))
4441                 free_reloc_roots(&reloc_roots);
4442
4443         btrfs_free_path(path);
4444
4445         if (err == 0) {
4446                 /* cleanup orphan inode in data relocation tree */
4447                 fs_root = read_fs_root(root->fs_info,
4448                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4449                 if (IS_ERR(fs_root))
4450                         err = PTR_ERR(fs_root);
4451                 else
4452                         err = btrfs_orphan_cleanup(fs_root);
4453         }
4454         return err;
4455 }
4456
4457 /*
4458  * helper to add ordered checksum for data relocation.
4459  *
4460  * cloning checksum properly handles the nodatasum extents.
4461  * it also saves CPU time to re-calculate the checksum.
4462  */
4463 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4464 {
4465         struct btrfs_ordered_sum *sums;
4466         struct btrfs_ordered_extent *ordered;
4467         struct btrfs_root *root = BTRFS_I(inode)->root;
4468         int ret;
4469         u64 disk_bytenr;
4470         u64 new_bytenr;
4471         LIST_HEAD(list);
4472
4473         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4474         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4475
4476         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4477         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4478                                        disk_bytenr + len - 1, &list, 0);
4479         if (ret)
4480                 goto out;
4481
4482         while (!list_empty(&list)) {
4483                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4484                 list_del_init(&sums->list);
4485
4486                 /*
4487                  * We need to offset the new_bytenr based on where the csum is.
4488                  * We need to do this because we will read in entire prealloc
4489                  * extents but we may have written to say the middle of the
4490                  * prealloc extent, so we need to make sure the csum goes with
4491                  * the right disk offset.
4492                  *
4493                  * We can do this because the data reloc inode refers strictly
4494                  * to the on disk bytes, so we don't have to worry about
4495                  * disk_len vs real len like with real inodes since it's all
4496                  * disk length.
4497                  */
4498                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4499                 sums->bytenr = new_bytenr;
4500
4501                 btrfs_add_ordered_sum(inode, ordered, sums);
4502         }
4503 out:
4504         btrfs_put_ordered_extent(ordered);
4505         return ret;
4506 }
4507
4508 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4509                           struct btrfs_root *root, struct extent_buffer *buf,
4510                           struct extent_buffer *cow)
4511 {
4512         struct reloc_control *rc;
4513         struct backref_node *node;
4514         int first_cow = 0;
4515         int level;
4516         int ret = 0;
4517
4518         rc = root->fs_info->reloc_ctl;
4519         if (!rc)
4520                 return 0;
4521
4522         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4523                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4524
4525         level = btrfs_header_level(buf);
4526         if (btrfs_header_generation(buf) <=
4527             btrfs_root_last_snapshot(&root->root_item))
4528                 first_cow = 1;
4529
4530         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4531             rc->create_reloc_tree) {
4532                 WARN_ON(!first_cow && level == 0);
4533
4534                 node = rc->backref_cache.path[level];
4535                 BUG_ON(node->bytenr != buf->start &&
4536                        node->new_bytenr != buf->start);
4537
4538                 drop_node_buffer(node);
4539                 extent_buffer_get(cow);
4540                 node->eb = cow;
4541                 node->new_bytenr = cow->start;
4542
4543                 if (!node->pending) {
4544                         list_move_tail(&node->list,
4545                                        &rc->backref_cache.pending[level]);
4546                         node->pending = 1;
4547                 }
4548
4549                 if (first_cow)
4550                         __mark_block_processed(rc, node);
4551
4552                 if (first_cow && level > 0)
4553                         rc->nodes_relocated += buf->len;
4554         }
4555
4556         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4557                 ret = replace_file_extents(trans, rc, root, cow);
4558         return ret;
4559 }
4560
4561 /*
4562  * called before creating snapshot. it calculates metadata reservation
4563  * requried for relocating tree blocks in the snapshot
4564  */
4565 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4566                               struct btrfs_pending_snapshot *pending,
4567                               u64 *bytes_to_reserve)
4568 {
4569         struct btrfs_root *root;
4570         struct reloc_control *rc;
4571
4572         root = pending->root;
4573         if (!root->reloc_root)
4574                 return;
4575
4576         rc = root->fs_info->reloc_ctl;
4577         if (!rc->merge_reloc_tree)
4578                 return;
4579
4580         root = root->reloc_root;
4581         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4582         /*
4583          * relocation is in the stage of merging trees. the space
4584          * used by merging a reloc tree is twice the size of
4585          * relocated tree nodes in the worst case. half for cowing
4586          * the reloc tree, half for cowing the fs tree. the space
4587          * used by cowing the reloc tree will be freed after the
4588          * tree is dropped. if we create snapshot, cowing the fs
4589          * tree may use more space than it frees. so we need
4590          * reserve extra space.
4591          */
4592         *bytes_to_reserve += rc->nodes_relocated;
4593 }
4594
4595 /*
4596  * called after snapshot is created. migrate block reservation
4597  * and create reloc root for the newly created snapshot
4598  */
4599 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4600                                struct btrfs_pending_snapshot *pending)
4601 {
4602         struct btrfs_root *root = pending->root;
4603         struct btrfs_root *reloc_root;
4604         struct btrfs_root *new_root;
4605         struct reloc_control *rc;
4606         int ret;
4607
4608         if (!root->reloc_root)
4609                 return 0;
4610
4611         rc = root->fs_info->reloc_ctl;
4612         rc->merging_rsv_size += rc->nodes_relocated;
4613
4614         if (rc->merge_reloc_tree) {
4615                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4616                                               rc->block_rsv,
4617                                               rc->nodes_relocated);
4618                 if (ret)
4619                         return ret;
4620         }
4621
4622         new_root = pending->snap;
4623         reloc_root = create_reloc_root(trans, root->reloc_root,
4624                                        new_root->root_key.objectid);
4625         if (IS_ERR(reloc_root))
4626                 return PTR_ERR(reloc_root);
4627
4628         ret = __add_reloc_root(reloc_root);
4629         BUG_ON(ret < 0);
4630         new_root->reloc_root = reloc_root;
4631
4632         if (rc->create_reloc_tree)
4633                 ret = clone_backref_node(trans, rc, root, reloc_root);
4634         return ret;
4635 }