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