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 #define RELOCATION_RESERVED_NODES       256
  98
  99 struct backref_cache {
 100         /* red black tree of all backref nodes in the cache */
 101         struct rb_root rb_root;
 102         /* for passing backref nodes to btrfs_reloc_cow_block */
 103         struct backref_node *path[BTRFS_MAX_LEVEL];
 104         /*
 105          * list of blocks that have been cowed but some block
 106          * pointers in upper level blocks may not reflect the
 107          * new location
 108          */
 109         struct list_head pending[BTRFS_MAX_LEVEL];
 110         /* list of backref nodes with no child node */
 111         struct list_head leaves;
 112         /* list of blocks that have been cowed in current transaction */
 113         struct list_head changed;
 114         /* list of detached backref node. */
 115         struct list_head detached;
 116
 117         u64 last_trans;
 118
 119         int nr_nodes;
 120         int nr_edges;
 121 };
 122
 123 /*
 124  * map address of tree root to tree
 125  */
 126 struct mapping_node {
 127         struct rb_node rb_node;
 128         u64 bytenr;
 129         void *data;
 130 };
 131
 132 struct mapping_tree {
 133         struct rb_root rb_root;
 134         spinlock_t lock;
 135 };
 136
 137 /*
 138  * present a tree block to process
 139  */
 140 struct tree_block {
 141         struct rb_node rb_node;
 142         u64 bytenr;
 143         struct btrfs_key key;
 144         unsigned int level:8;
 145         unsigned int key_ready:1;
 146 };
 147
 148 #define MAX_EXTENTS 128
 149
 150 struct file_extent_cluster {
 151         u64 start;
 152         u64 end;
 153         u64 boundary[MAX_EXTENTS];
 154         unsigned int nr;
 155 };
 156
 157 struct reloc_control {
 158         /* block group to relocate */
 159         struct btrfs_block_group_cache *block_group;
 160         /* extent tree */
 161         struct btrfs_root *extent_root;
 162         /* inode for moving data */
 163         struct inode *data_inode;
 164
 165         struct btrfs_block_rsv *block_rsv;
 166
 167         struct backref_cache backref_cache;
 168
 169         struct file_extent_cluster cluster;
 170         /* tree blocks have been processed */
 171         struct extent_io_tree processed_blocks;
 172         /* map start of tree root to corresponding reloc tree */
 173         struct mapping_tree reloc_root_tree;
 174         /* list of reloc trees */
 175         struct list_head reloc_roots;
 176         /* size of metadata reservation for merging reloc trees */
 177         u64 merging_rsv_size;
 178         /* size of relocated tree nodes */
 179         u64 nodes_relocated;
 180         /* reserved size for block group relocation*/
 181         u64 reserved_bytes;
 182
 183         u64 search_start;
 184         u64 extents_found;
 185
 186         unsigned int stage:8;
 187         unsigned int create_reloc_tree:1;
 188         unsigned int merge_reloc_tree:1;
 189         unsigned int found_file_extent:1;
 190 };
 191
 192 /* stages of data relocation */
 193 #define MOVE_DATA_EXTENTS       0
 194 #define UPDATE_DATA_PTRS        1
 195
 196 static void remove_backref_node(struct backref_cache *cache,
 197                                 struct backref_node *node);
 198 static void __mark_block_processed(struct reloc_control *rc,
 199                                    struct backref_node *node);
 200
 201 static void mapping_tree_init(struct mapping_tree *tree)
 202 {
 203         tree->rb_root = RB_ROOT;
 204         spin_lock_init(&tree->lock);
 205 }
 206
 207 static void backref_cache_init(struct backref_cache *cache)
 208 {
 209         int i;
 210         cache->rb_root = RB_ROOT;
 211         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 212                 INIT_LIST_HEAD(&cache->pending[i]);
 213         INIT_LIST_HEAD(&cache->changed);
 214         INIT_LIST_HEAD(&cache->detached);
 215         INIT_LIST_HEAD(&cache->leaves);
 216 }
 217
 218 static void backref_cache_cleanup(struct backref_cache *cache)
 219 {
 220         struct backref_node *node;
 221         int i;
 222
 223         while (!list_empty(&cache->detached)) {
 224                 node = list_entry(cache->detached.next,
 225                                   struct backref_node, list);
 226                 remove_backref_node(cache, node);
 227         }
 228
 229         while (!list_empty(&cache->leaves)) {
 230                 node = list_entry(cache->leaves.next,
 231                                   struct backref_node, lower);
 232                 remove_backref_node(cache, node);
 233         }
 234
 235         cache->last_trans = 0;
 236
 237         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 238                 BUG_ON(!list_empty(&cache->pending[i]));
 239         BUG_ON(!list_empty(&cache->changed));
 240         BUG_ON(!list_empty(&cache->detached));
 241         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
 242         BUG_ON(cache->nr_nodes);
 243         BUG_ON(cache->nr_edges);
 244 }
 245
 246 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 247 {
 248         struct backref_node *node;
 249
 250         node = kzalloc(sizeof(*node), GFP_NOFS);
 251         if (node) {
 252                 INIT_LIST_HEAD(&node->list);
 253                 INIT_LIST_HEAD(&node->upper);
 254                 INIT_LIST_HEAD(&node->lower);
 255                 RB_CLEAR_NODE(&node->rb_node);
 256                 cache->nr_nodes++;
 257         }
 258         return node;
 259 }
 260
 261 static void free_backref_node(struct backref_cache *cache,
 262                               struct backref_node *node)
 263 {
 264         if (node) {
 265                 cache->nr_nodes--;
 266                 kfree(node);
 267         }
 268 }
 269
 270 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 271 {
 272         struct backref_edge *edge;
 273
 274         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 275         if (edge)
 276                 cache->nr_edges++;
 277         return edge;
 278 }
 279
 280 static void free_backref_edge(struct backref_cache *cache,
 281                               struct backref_edge *edge)
 282 {
 283         if (edge) {
 284                 cache->nr_edges--;
 285                 kfree(edge);
 286         }
 287 }
 288
 289 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 290                                    struct rb_node *node)
 291 {
 292         struct rb_node **p = &root->rb_node;
 293         struct rb_node *parent = NULL;
 294         struct tree_entry *entry;
 295
 296         while (*p) {
 297                 parent = *p;
 298                 entry = rb_entry(parent, struct tree_entry, rb_node);
 299
 300                 if (bytenr < entry->bytenr)
 301                         p = &(*p)->rb_left;
 302                 else if (bytenr > entry->bytenr)
 303                         p = &(*p)->rb_right;
 304                 else
 305                         return parent;
 306         }
 307
 308         rb_link_node(node, parent, p);
 309         rb_insert_color(node, root);
 310         return NULL;
 311 }
 312
 313 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 314 {
 315         struct rb_node *n = root->rb_node;
 316         struct tree_entry *entry;
 317
 318         while (n) {
 319                 entry = rb_entry(n, struct tree_entry, rb_node);
 320
 321                 if (bytenr < entry->bytenr)
 322                         n = n->rb_left;
 323                 else if (bytenr > entry->bytenr)
 324                         n = n->rb_right;
 325                 else
 326                         return n;
 327         }
 328         return NULL;
 329 }
 330
 331 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 332 {
 333
 334         struct btrfs_fs_info *fs_info = NULL;
 335         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 336                                               rb_node);
 337         if (bnode->root)
 338                 fs_info = bnode->root->fs_info;
 339         btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
 340                     "found at offset %llu", bytenr);
 341 }
 342
 343 /*
 344  * walk up backref nodes until reach node presents tree root
 345  */
 346 static struct backref_node *walk_up_backref(struct backref_node *node,
 347                                             struct backref_edge *edges[],
 348                                             int *index)
 349 {
 350         struct backref_edge *edge;
 351         int idx = *index;
 352
 353         while (!list_empty(&node->upper)) {
 354                 edge = list_entry(node->upper.next,
 355                                   struct backref_edge, list[LOWER]);
 356                 edges[idx++] = edge;
 357                 node = edge->node[UPPER];
 358         }
 359         BUG_ON(node->detached);
 360         *index = idx;
 361         return node;
 362 }
 363
 364 /*
 365  * walk down backref nodes to find start of next reference path
 366  */
 367 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 368                                               int *index)
 369 {
 370         struct backref_edge *edge;
 371         struct backref_node *lower;
 372         int idx = *index;
 373
 374         while (idx > 0) {
 375                 edge = edges[idx - 1];
 376                 lower = edge->node[LOWER];
 377                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 378                         idx--;
 379                         continue;
 380                 }
 381                 edge = list_entry(edge->list[LOWER].next,
 382                                   struct backref_edge, list[LOWER]);
 383                 edges[idx - 1] = edge;
 384                 *index = idx;
 385                 return edge->node[UPPER];
 386         }
 387         *index = 0;
 388         return NULL;
 389 }
 390
 391 static void unlock_node_buffer(struct backref_node *node)
 392 {
 393         if (node->locked) {
 394                 btrfs_tree_unlock(node->eb);
 395                 node->locked = 0;
 396         }
 397 }
 398
 399 static void drop_node_buffer(struct backref_node *node)
 400 {
 401         if (node->eb) {
 402                 unlock_node_buffer(node);
 403                 free_extent_buffer(node->eb);
 404                 node->eb = NULL;
 405         }
 406 }
 407
 408 static void drop_backref_node(struct backref_cache *tree,
 409                               struct backref_node *node)
 410 {
 411         BUG_ON(!list_empty(&node->upper));
 412
 413         drop_node_buffer(node);
 414         list_del(&node->list);
 415         list_del(&node->lower);
 416         if (!RB_EMPTY_NODE(&node->rb_node))
 417                 rb_erase(&node->rb_node, &tree->rb_root);
 418         free_backref_node(tree, node);
 419 }
 420
 421 /*
 422  * remove a backref node from the backref cache
 423  */
 424 static void remove_backref_node(struct backref_cache *cache,
 425                                 struct backref_node *node)
 426 {
 427         struct backref_node *upper;
 428         struct backref_edge *edge;
 429
 430         if (!node)
 431                 return;
 432
 433         BUG_ON(!node->lowest && !node->detached);
 434         while (!list_empty(&node->upper)) {
 435                 edge = list_entry(node->upper.next, struct backref_edge,
 436                                   list[LOWER]);
 437                 upper = edge->node[UPPER];
 438                 list_del(&edge->list[LOWER]);
 439                 list_del(&edge->list[UPPER]);
 440                 free_backref_edge(cache, edge);
 441
 442                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 443                         BUG_ON(!list_empty(&node->upper));
 444                         drop_backref_node(cache, node);
 445                         node = upper;
 446                         node->lowest = 1;
 447                         continue;
 448                 }
 449                 /*
 450                  * add the node to leaf node list if no other
 451                  * child block cached.
 452                  */
 453                 if (list_empty(&upper->lower)) {
 454                         list_add_tail(&upper->lower, &cache->leaves);
 455                         upper->lowest = 1;
 456                 }
 457         }
 458
 459         drop_backref_node(cache, node);
 460 }
 461
 462 static void update_backref_node(struct backref_cache *cache,
 463                                 struct backref_node *node, u64 bytenr)
 464 {
 465         struct rb_node *rb_node;
 466         rb_erase(&node->rb_node, &cache->rb_root);
 467         node->bytenr = bytenr;
 468         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 469         if (rb_node)
 470                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 471 }
 472
 473 /*
 474  * update backref cache after a transaction commit
 475  */
 476 static int update_backref_cache(struct btrfs_trans_handle *trans,
 477                                 struct backref_cache *cache)
 478 {
 479         struct backref_node *node;
 480         int level = 0;
 481
 482         if (cache->last_trans == 0) {
 483                 cache->last_trans = trans->transid;
 484                 return 0;
 485         }
 486
 487         if (cache->last_trans == trans->transid)
 488                 return 0;
 489
 490         /*
 491          * detached nodes are used to avoid unnecessary backref
 492          * lookup. transaction commit changes the extent tree.
 493          * so the detached nodes are no longer useful.
 494          */
 495         while (!list_empty(&cache->detached)) {
 496                 node = list_entry(cache->detached.next,
 497                                   struct backref_node, list);
 498                 remove_backref_node(cache, node);
 499         }
 500
 501         while (!list_empty(&cache->changed)) {
 502                 node = list_entry(cache->changed.next,
 503                                   struct backref_node, list);
 504                 list_del_init(&node->list);
 505                 BUG_ON(node->pending);
 506                 update_backref_node(cache, node, node->new_bytenr);
 507         }
 508
 509         /*
 510          * some nodes can be left in the pending list if there were
 511          * errors during processing the pending nodes.
 512          */
 513         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 514                 list_for_each_entry(node, &cache->pending[level], list) {
 515                         BUG_ON(!node->pending);
 516                         if (node->bytenr == node->new_bytenr)
 517                                 continue;
 518                         update_backref_node(cache, node, node->new_bytenr);
 519                 }
 520         }
 521
 522         cache->last_trans = 0;
 523         return 1;
 524 }
 525
 526
 527 static int should_ignore_root(struct btrfs_root *root)
 528 {
 529         struct btrfs_root *reloc_root;
 530
 531         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 532                 return 0;
 533
 534         reloc_root = root->reloc_root;
 535         if (!reloc_root)
 536                 return 0;
 537
 538         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 539             root->fs_info->running_transaction->transid - 1)
 540                 return 0;
 541         /*
 542          * if there is reloc tree and it was created in previous
 543          * transaction backref lookup can find the reloc tree,
 544          * so backref node for the fs tree root is useless for
 545          * relocation.
 546          */
 547         return 1;
 548 }
 549 /*
 550  * find reloc tree by address of tree root
 551  */
 552 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 553                                           u64 bytenr)
 554 {
 555         struct rb_node *rb_node;
 556         struct mapping_node *node;
 557         struct btrfs_root *root = NULL;
 558
 559         spin_lock(&rc->reloc_root_tree.lock);
 560         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 561         if (rb_node) {
 562                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 563                 root = (struct btrfs_root *)node->data;
 564         }
 565         spin_unlock(&rc->reloc_root_tree.lock);
 566         return root;
 567 }
 568
 569 static int is_cowonly_root(u64 root_objectid)
 570 {
 571         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 572             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 573             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 574             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 575             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 576             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 577             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 578             root_objectid == BTRFS_QUOTA_TREE_OBJECTID)
 579                 return 1;
 580         return 0;
 581 }
 582
 583 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 584                                         u64 root_objectid)
 585 {
 586         struct btrfs_key key;
 587
 588         key.objectid = root_objectid;
 589         key.type = BTRFS_ROOT_ITEM_KEY;
 590         if (is_cowonly_root(root_objectid))
 591                 key.offset = 0;
 592         else
 593                 key.offset = (u64)-1;
 594
 595         return btrfs_get_fs_root(fs_info, &key, false);
 596 }
 597
 598 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 599 static noinline_for_stack
 600 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 601                                   struct extent_buffer *leaf,
 602                                   struct btrfs_extent_ref_v0 *ref0)
 603 {
 604         struct btrfs_root *root;
 605         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 606         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 607
 608         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 609
 610         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 611         BUG_ON(IS_ERR(root));
 612
 613         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 614             generation != btrfs_root_generation(&root->root_item))
 615                 return NULL;
 616
 617         return root;
 618 }
 619 #endif
 620
 621 static noinline_for_stack
 622 int find_inline_backref(struct extent_buffer *leaf, int slot,
 623                         unsigned long *ptr, unsigned long *end)
 624 {
 625         struct btrfs_key key;
 626         struct btrfs_extent_item *ei;
 627         struct btrfs_tree_block_info *bi;
 628         u32 item_size;
 629
 630         btrfs_item_key_to_cpu(leaf, &key, slot);
 631
 632         item_size = btrfs_item_size_nr(leaf, slot);
 633 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 634         if (item_size < sizeof(*ei)) {
 635                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 636                 return 1;
 637         }
 638 #endif
 639         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 640         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 641                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 642
 643         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 644             item_size <= sizeof(*ei) + sizeof(*bi)) {
 645                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 646                 return 1;
 647         }
 648         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 649             item_size <= sizeof(*ei)) {
 650                 WARN_ON(item_size < sizeof(*ei));
 651                 return 1;
 652         }
 653
 654         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 655                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 656                 *ptr = (unsigned long)(bi + 1);
 657         } else {
 658                 *ptr = (unsigned long)(ei + 1);
 659         }
 660         *end = (unsigned long)ei + item_size;
 661         return 0;
 662 }
 663
 664 /*
 665  * build backref tree for a given tree block. root of the backref tree
 666  * corresponds the tree block, leaves of the backref tree correspond
 667  * roots of b-trees that reference the tree block.
 668  *
 669  * the basic idea of this function is check backrefs of a given block
 670  * to find upper level blocks that refernece the block, and then check
 671  * bakcrefs of these upper level blocks recursively. the recursion stop
 672  * when tree root is reached or backrefs for the block is cached.
 673  *
 674  * NOTE: if we find backrefs for a block are cached, we know backrefs
 675  * for all upper level blocks that directly/indirectly reference the
 676  * block are also cached.
 677  */
 678 static noinline_for_stack
 679 struct backref_node *build_backref_tree(struct reloc_control *rc,
 680                                         struct btrfs_key *node_key,
 681                                         int level, u64 bytenr)
 682 {
 683         struct backref_cache *cache = &rc->backref_cache;
 684         struct btrfs_path *path1;
 685         struct btrfs_path *path2;
 686         struct extent_buffer *eb;
 687         struct btrfs_root *root;
 688         struct backref_node *cur;
 689         struct backref_node *upper;
 690         struct backref_node *lower;
 691         struct backref_node *node = NULL;
 692         struct backref_node *exist = NULL;
 693         struct backref_edge *edge;
 694         struct rb_node *rb_node;
 695         struct btrfs_key key;
 696         unsigned long end;
 697         unsigned long ptr;
 698         LIST_HEAD(list);
 699         LIST_HEAD(useless);
 700         int cowonly;
 701         int ret;
 702         int err = 0;
 703         bool need_check = true;
 704
 705         path1 = btrfs_alloc_path();
 706         path2 = btrfs_alloc_path();
 707         if (!path1 || !path2) {
 708                 err = -ENOMEM;
 709                 goto out;
 710         }
 711         path1->reada = 1;
 712         path2->reada = 2;
 713
 714         node = alloc_backref_node(cache);
 715         if (!node) {
 716                 err = -ENOMEM;
 717                 goto out;
 718         }
 719
 720         node->bytenr = bytenr;
 721         node->level = level;
 722         node->lowest = 1;
 723         cur = node;
 724 again:
 725         end = 0;
 726         ptr = 0;
 727         key.objectid = cur->bytenr;
 728         key.type = BTRFS_METADATA_ITEM_KEY;
 729         key.offset = (u64)-1;
 730
 731         path1->search_commit_root = 1;
 732         path1->skip_locking = 1;
 733         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 734                                 0, 0);
 735         if (ret < 0) {
 736                 err = ret;
 737                 goto out;
 738         }
 739         ASSERT(ret);
 740         ASSERT(path1->slots[0]);
 741
 742         path1->slots[0]--;
 743
 744         WARN_ON(cur->checked);
 745         if (!list_empty(&cur->upper)) {
 746                 /*
 747                  * the backref was added previously when processing
 748                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 749                  */
 750                 ASSERT(list_is_singular(&cur->upper));
 751                 edge = list_entry(cur->upper.next, struct backref_edge,
 752                                   list[LOWER]);
 753                 ASSERT(list_empty(&edge->list[UPPER]));
 754                 exist = edge->node[UPPER];
 755                 /*
 756                  * add the upper level block to pending list if we need
 757                  * check its backrefs
 758                  */
 759                 if (!exist->checked)
 760                         list_add_tail(&edge->list[UPPER], &list);
 761         } else {
 762                 exist = NULL;
 763         }
 764
 765         while (1) {
 766                 cond_resched();
 767                 eb = path1->nodes[0];
 768
 769                 if (ptr >= end) {
 770                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 771                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 772                                 if (ret < 0) {
 773                                         err = ret;
 774                                         goto out;
 775                                 }
 776                                 if (ret > 0)
 777                                         break;
 778                                 eb = path1->nodes[0];
 779                         }
 780
 781                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 782                         if (key.objectid != cur->bytenr) {
 783                                 WARN_ON(exist);
 784                                 break;
 785                         }
 786
 787                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 788                             key.type == BTRFS_METADATA_ITEM_KEY) {
 789                                 ret = find_inline_backref(eb, path1->slots[0],
 790                                                           &ptr, &end);
 791                                 if (ret)
 792                                         goto next;
 793                         }
 794                 }
 795
 796                 if (ptr < end) {
 797                         /* update key for inline back ref */
 798                         struct btrfs_extent_inline_ref *iref;
 799                         iref = (struct btrfs_extent_inline_ref *)ptr;
 800                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 801                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 802                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 803                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 804                 }
 805
 806                 if (exist &&
 807                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 808                       exist->owner == key.offset) ||
 809                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 810                       exist->bytenr == key.offset))) {
 811                         exist = NULL;
 812                         goto next;
 813                 }
 814
 815 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 816                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 817                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 818                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 819                                 struct btrfs_extent_ref_v0 *ref0;
 820                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 821                                                 struct btrfs_extent_ref_v0);
 822                                 if (key.objectid == key.offset) {
 823                                         root = find_tree_root(rc, eb, ref0);
 824                                         if (root && !should_ignore_root(root))
 825                                                 cur->root = root;
 826                                         else
 827                                                 list_add(&cur->list, &useless);
 828                                         break;
 829                                 }
 830                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 831                                                                       ref0)))
 832                                         cur->cowonly = 1;
 833                         }
 834 #else
 835                 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
 836                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 837 #endif
 838                         if (key.objectid == key.offset) {
 839                                 /*
 840                                  * only root blocks of reloc trees use
 841                                  * backref of this type.
 842                                  */
 843                                 root = find_reloc_root(rc, cur->bytenr);
 844                                 ASSERT(root);
 845                                 cur->root = root;
 846                                 break;
 847                         }
 848
 849                         edge = alloc_backref_edge(cache);
 850                         if (!edge) {
 851                                 err = -ENOMEM;
 852                                 goto out;
 853                         }
 854                         rb_node = tree_search(&cache->rb_root, key.offset);
 855                         if (!rb_node) {
 856                                 upper = alloc_backref_node(cache);
 857                                 if (!upper) {
 858                                         free_backref_edge(cache, edge);
 859                                         err = -ENOMEM;
 860                                         goto out;
 861                                 }
 862                                 upper->bytenr = key.offset;
 863                                 upper->level = cur->level + 1;
 864                                 /*
 865                                  *  backrefs for the upper level block isn't
 866                                  *  cached, add the block to pending list
 867                                  */
 868                                 list_add_tail(&edge->list[UPPER], &list);
 869                         } else {
 870                                 upper = rb_entry(rb_node, struct backref_node,
 871                                                  rb_node);
 872                                 ASSERT(upper->checked);
 873                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 874                         }
 875                         list_add_tail(&edge->list[LOWER], &cur->upper);
 876                         edge->node[LOWER] = cur;
 877                         edge->node[UPPER] = upper;
 878
 879                         goto next;
 880                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 881                         goto next;
 882                 }
 883
 884                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 885                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 886                 if (IS_ERR(root)) {
 887                         err = PTR_ERR(root);
 888                         goto out;
 889                 }
 890
 891                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 892                         cur->cowonly = 1;
 893
 894                 if (btrfs_root_level(&root->root_item) == cur->level) {
 895                         /* tree root */
 896                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
 897                                cur->bytenr);
 898                         if (should_ignore_root(root))
 899                                 list_add(&cur->list, &useless);
 900                         else
 901                                 cur->root = root;
 902                         break;
 903                 }
 904
 905                 level = cur->level + 1;
 906
 907                 /*
 908                  * searching the tree to find upper level blocks
 909                  * reference the block.
 910                  */
 911                 path2->search_commit_root = 1;
 912                 path2->skip_locking = 1;
 913                 path2->lowest_level = level;
 914                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 915                 path2->lowest_level = 0;
 916                 if (ret < 0) {
 917                         err = ret;
 918                         goto out;
 919                 }
 920                 if (ret > 0 && path2->slots[level] > 0)
 921                         path2->slots[level]--;
 922
 923                 eb = path2->nodes[level];
 924                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 925                         cur->bytenr);
 926
 927                 lower = cur;
 928                 need_check = true;
 929                 for (; level < BTRFS_MAX_LEVEL; level++) {
 930                         if (!path2->nodes[level]) {
 931                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
 932                                        lower->bytenr);
 933                                 if (should_ignore_root(root))
 934                                         list_add(&lower->list, &useless);
 935                                 else
 936                                         lower->root = root;
 937                                 break;
 938                         }
 939
 940                         edge = alloc_backref_edge(cache);
 941                         if (!edge) {
 942                                 err = -ENOMEM;
 943                                 goto out;
 944                         }
 945
 946                         eb = path2->nodes[level];
 947                         rb_node = tree_search(&cache->rb_root, eb->start);
 948                         if (!rb_node) {
 949                                 upper = alloc_backref_node(cache);
 950                                 if (!upper) {
 951                                         free_backref_edge(cache, edge);
 952                                         err = -ENOMEM;
 953                                         goto out;
 954                                 }
 955                                 upper->bytenr = eb->start;
 956                                 upper->owner = btrfs_header_owner(eb);
 957                                 upper->level = lower->level + 1;
 958                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
 959                                               &root->state))
 960                                         upper->cowonly = 1;
 961
 962                                 /*
 963                                  * if we know the block isn't shared
 964                                  * we can void checking its backrefs.
 965                                  */
 966                                 if (btrfs_block_can_be_shared(root, eb))
 967                                         upper->checked = 0;
 968                                 else
 969                                         upper->checked = 1;
 970
 971                                 /*
 972                                  * add the block to pending list if we
 973                                  * need check its backrefs, we only do this once
 974                                  * while walking up a tree as we will catch
 975                                  * anything else later on.
 976                                  */
 977                                 if (!upper->checked && need_check) {
 978                                         need_check = false;
 979                                         list_add_tail(&edge->list[UPPER],
 980                                                       &list);
 981                                 } else
 982                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 983                         } else {
 984                                 upper = rb_entry(rb_node, struct backref_node,
 985                                                  rb_node);
 986                                 ASSERT(upper->checked);
 987                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 988                                 if (!upper->owner)
 989                                         upper->owner = btrfs_header_owner(eb);
 990                         }
 991                         list_add_tail(&edge->list[LOWER], &lower->upper);
 992                         edge->node[LOWER] = lower;
 993                         edge->node[UPPER] = upper;
 994
 995                         if (rb_node)
 996                                 break;
 997                         lower = upper;
 998                         upper = NULL;
 999                 }
1000                 btrfs_release_path(path2);
1001 next:
1002                 if (ptr < end) {
1003                         ptr += btrfs_extent_inline_ref_size(key.type);
1004                         if (ptr >= end) {
1005                                 WARN_ON(ptr > end);
1006                                 ptr = 0;
1007                                 end = 0;
1008                         }
1009                 }
1010                 if (ptr >= end)
1011                         path1->slots[0]++;
1012         }
1013         btrfs_release_path(path1);
1014
1015         cur->checked = 1;
1016         WARN_ON(exist);
1017
1018         /* the pending list isn't empty, take the first block to process */
1019         if (!list_empty(&list)) {
1020                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1021                 list_del_init(&edge->list[UPPER]);
1022                 cur = edge->node[UPPER];
1023                 goto again;
1024         }
1025
1026         /*
1027          * everything goes well, connect backref nodes and insert backref nodes
1028          * into the cache.
1029          */
1030         ASSERT(node->checked);
1031         cowonly = node->cowonly;
1032         if (!cowonly) {
1033                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1034                                       &node->rb_node);
1035                 if (rb_node)
1036                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1037                 list_add_tail(&node->lower, &cache->leaves);
1038         }
1039
1040         list_for_each_entry(edge, &node->upper, list[LOWER])
1041                 list_add_tail(&edge->list[UPPER], &list);
1042
1043         while (!list_empty(&list)) {
1044                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1045                 list_del_init(&edge->list[UPPER]);
1046                 upper = edge->node[UPPER];
1047                 if (upper->detached) {
1048                         list_del(&edge->list[LOWER]);
1049                         lower = edge->node[LOWER];
1050                         free_backref_edge(cache, edge);
1051                         if (list_empty(&lower->upper))
1052                                 list_add(&lower->list, &useless);
1053                         continue;
1054                 }
1055
1056                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1057                         if (upper->lowest) {
1058                                 list_del_init(&upper->lower);
1059                                 upper->lowest = 0;
1060                         }
1061
1062                         list_add_tail(&edge->list[UPPER], &upper->lower);
1063                         continue;
1064                 }
1065
1066                 if (!upper->checked) {
1067                         /*
1068                          * Still want to blow up for developers since this is a
1069                          * logic bug.
1070                          */
1071                         ASSERT(0);
1072                         err = -EINVAL;
1073                         goto out;
1074                 }
1075                 if (cowonly != upper->cowonly) {
1076                         ASSERT(0);
1077                         err = -EINVAL;
1078                         goto out;
1079                 }
1080
1081                 if (!cowonly) {
1082                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1083                                               &upper->rb_node);
1084                         if (rb_node)
1085                                 backref_tree_panic(rb_node, -EEXIST,
1086                                                    upper->bytenr);
1087                 }
1088
1089                 list_add_tail(&edge->list[UPPER], &upper->lower);
1090
1091                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1092                         list_add_tail(&edge->list[UPPER], &list);
1093         }
1094         /*
1095          * process useless backref nodes. backref nodes for tree leaves
1096          * are deleted from the cache. backref nodes for upper level
1097          * tree blocks are left in the cache to avoid unnecessary backref
1098          * lookup.
1099          */
1100         while (!list_empty(&useless)) {
1101                 upper = list_entry(useless.next, struct backref_node, list);
1102                 list_del_init(&upper->list);
1103                 ASSERT(list_empty(&upper->upper));
1104                 if (upper == node)
1105                         node = NULL;
1106                 if (upper->lowest) {
1107                         list_del_init(&upper->lower);
1108                         upper->lowest = 0;
1109                 }
1110                 while (!list_empty(&upper->lower)) {
1111                         edge = list_entry(upper->lower.next,
1112                                           struct backref_edge, list[UPPER]);
1113                         list_del(&edge->list[UPPER]);
1114                         list_del(&edge->list[LOWER]);
1115                         lower = edge->node[LOWER];
1116                         free_backref_edge(cache, edge);
1117
1118                         if (list_empty(&lower->upper))
1119                                 list_add(&lower->list, &useless);
1120                 }
1121                 __mark_block_processed(rc, upper);
1122                 if (upper->level > 0) {
1123                         list_add(&upper->list, &cache->detached);
1124                         upper->detached = 1;
1125                 } else {
1126                         rb_erase(&upper->rb_node, &cache->rb_root);
1127                         free_backref_node(cache, upper);
1128                 }
1129         }
1130 out:
1131         btrfs_free_path(path1);
1132         btrfs_free_path(path2);
1133         if (err) {
1134                 while (!list_empty(&useless)) {
1135                         lower = list_entry(useless.next,
1136                                            struct backref_node, list);
1137                         list_del_init(&lower->list);
1138                 }
1139                 while (!list_empty(&list)) {
1140                         edge = list_first_entry(&list, struct backref_edge,
1141                                                 list[UPPER]);
1142                         list_del(&edge->list[UPPER]);
1143                         list_del(&edge->list[LOWER]);
1144                         lower = edge->node[LOWER];
1145                         upper = edge->node[UPPER];
1146                         free_backref_edge(cache, edge);
1147
1148                         /*
1149                          * Lower is no longer linked to any upper backref nodes
1150                          * and isn't in the cache, we can free it ourselves.
1151                          */
1152                         if (list_empty(&lower->upper) &&
1153                             RB_EMPTY_NODE(&lower->rb_node))
1154                                 list_add(&lower->list, &useless);
1155
1156                         if (!RB_EMPTY_NODE(&upper->rb_node))
1157                                 continue;
1158
1159                         /* Add this guy's upper edges to the list to proces */
1160                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1161                                 list_add_tail(&edge->list[UPPER], &list);
1162                         if (list_empty(&upper->upper))
1163                                 list_add(&upper->list, &useless);
1164                 }
1165
1166                 while (!list_empty(&useless)) {
1167                         lower = list_entry(useless.next,
1168                                            struct backref_node, list);
1169                         list_del_init(&lower->list);
1170                         free_backref_node(cache, lower);
1171                 }
1172                 return ERR_PTR(err);
1173         }
1174         ASSERT(!node || !node->detached);
1175         return node;
1176 }
1177
1178 /*
1179  * helper to add backref node for the newly created snapshot.
1180  * the backref node is created by cloning backref node that
1181  * corresponds to root of source tree
1182  */
1183 static int clone_backref_node(struct btrfs_trans_handle *trans,
1184                               struct reloc_control *rc,
1185                               struct btrfs_root *src,
1186                               struct btrfs_root *dest)
1187 {
1188         struct btrfs_root *reloc_root = src->reloc_root;
1189         struct backref_cache *cache = &rc->backref_cache;
1190         struct backref_node *node = NULL;
1191         struct backref_node *new_node;
1192         struct backref_edge *edge;
1193         struct backref_edge *new_edge;
1194         struct rb_node *rb_node;
1195
1196         if (cache->last_trans > 0)
1197                 update_backref_cache(trans, cache);
1198
1199         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1200         if (rb_node) {
1201                 node = rb_entry(rb_node, struct backref_node, rb_node);
1202                 if (node->detached)
1203                         node = NULL;
1204                 else
1205                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1206         }
1207
1208         if (!node) {
1209                 rb_node = tree_search(&cache->rb_root,
1210                                       reloc_root->commit_root->start);
1211                 if (rb_node) {
1212                         node = rb_entry(rb_node, struct backref_node,
1213                                         rb_node);
1214                         BUG_ON(node->detached);
1215                 }
1216         }
1217
1218         if (!node)
1219                 return 0;
1220
1221         new_node = alloc_backref_node(cache);
1222         if (!new_node)
1223                 return -ENOMEM;
1224
1225         new_node->bytenr = dest->node->start;
1226         new_node->level = node->level;
1227         new_node->lowest = node->lowest;
1228         new_node->checked = 1;
1229         new_node->root = dest;
1230
1231         if (!node->lowest) {
1232                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1233                         new_edge = alloc_backref_edge(cache);
1234                         if (!new_edge)
1235                                 goto fail;
1236
1237                         new_edge->node[UPPER] = new_node;
1238                         new_edge->node[LOWER] = edge->node[LOWER];
1239                         list_add_tail(&new_edge->list[UPPER],
1240                                       &new_node->lower);
1241                 }
1242         } else {
1243                 list_add_tail(&new_node->lower, &cache->leaves);
1244         }
1245
1246         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1247                               &new_node->rb_node);
1248         if (rb_node)
1249                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1250
1251         if (!new_node->lowest) {
1252                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1253                         list_add_tail(&new_edge->list[LOWER],
1254                                       &new_edge->node[LOWER]->upper);
1255                 }
1256         }
1257         return 0;
1258 fail:
1259         while (!list_empty(&new_node->lower)) {
1260                 new_edge = list_entry(new_node->lower.next,
1261                                       struct backref_edge, list[UPPER]);
1262                 list_del(&new_edge->list[UPPER]);
1263                 free_backref_edge(cache, new_edge);
1264         }
1265         free_backref_node(cache, new_node);
1266         return -ENOMEM;
1267 }
1268
1269 /*
1270  * helper to add 'address of tree root -> reloc tree' mapping
1271  */
1272 static int __must_check __add_reloc_root(struct btrfs_root *root)
1273 {
1274         struct rb_node *rb_node;
1275         struct mapping_node *node;
1276         struct reloc_control *rc = root->fs_info->reloc_ctl;
1277
1278         node = kmalloc(sizeof(*node), GFP_NOFS);
1279         if (!node)
1280                 return -ENOMEM;
1281
1282         node->bytenr = root->node->start;
1283         node->data = root;
1284
1285         spin_lock(&rc->reloc_root_tree.lock);
1286         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1287                               node->bytenr, &node->rb_node);
1288         spin_unlock(&rc->reloc_root_tree.lock);
1289         if (rb_node) {
1290                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1291                             "for start=%llu while inserting into relocation "
1292                             "tree", node->bytenr);
1293                 kfree(node);
1294                 return -EEXIST;
1295         }
1296
1297         list_add_tail(&root->root_list, &rc->reloc_roots);
1298         return 0;
1299 }
1300
1301 /*
1302  * helper to delete the 'address of tree root -> reloc tree'
1303  * mapping
1304  */
1305 static void __del_reloc_root(struct btrfs_root *root)
1306 {
1307         struct rb_node *rb_node;
1308         struct mapping_node *node = NULL;
1309         struct reloc_control *rc = root->fs_info->reloc_ctl;
1310
1311         spin_lock(&rc->reloc_root_tree.lock);
1312         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1313                               root->node->start);
1314         if (rb_node) {
1315                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1316                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1317         }
1318         spin_unlock(&rc->reloc_root_tree.lock);
1319
1320         if (!node)
1321                 return;
1322         BUG_ON((struct btrfs_root *)node->data != root);
1323
1324         spin_lock(&root->fs_info->trans_lock);
1325         list_del_init(&root->root_list);
1326         spin_unlock(&root->fs_info->trans_lock);
1327         kfree(node);
1328 }
1329
1330 /*
1331  * helper to update the 'address of tree root -> reloc tree'
1332  * mapping
1333  */
1334 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1335 {
1336         struct rb_node *rb_node;
1337         struct mapping_node *node = NULL;
1338         struct reloc_control *rc = root->fs_info->reloc_ctl;
1339
1340         spin_lock(&rc->reloc_root_tree.lock);
1341         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1342                               root->node->start);
1343         if (rb_node) {
1344                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1345                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1346         }
1347         spin_unlock(&rc->reloc_root_tree.lock);
1348
1349         if (!node)
1350                 return 0;
1351         BUG_ON((struct btrfs_root *)node->data != root);
1352
1353         spin_lock(&rc->reloc_root_tree.lock);
1354         node->bytenr = new_bytenr;
1355         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1356                               node->bytenr, &node->rb_node);
1357         spin_unlock(&rc->reloc_root_tree.lock);
1358         if (rb_node)
1359                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1360         return 0;
1361 }
1362
1363 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1364                                         struct btrfs_root *root, u64 objectid)
1365 {
1366         struct btrfs_root *reloc_root;
1367         struct extent_buffer *eb;
1368         struct btrfs_root_item *root_item;
1369         struct btrfs_key root_key;
1370         u64 last_snap = 0;
1371         int ret;
1372
1373         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1374         BUG_ON(!root_item);
1375
1376         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1377         root_key.type = BTRFS_ROOT_ITEM_KEY;
1378         root_key.offset = objectid;
1379
1380         if (root->root_key.objectid == objectid) {
1381                 /* called by btrfs_init_reloc_root */
1382                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1383                                       BTRFS_TREE_RELOC_OBJECTID);
1384                 BUG_ON(ret);
1385
1386                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1387                 btrfs_set_root_last_snapshot(&root->root_item,
1388                                              trans->transid - 1);
1389         } else {
1390                 /*
1391                  * called by btrfs_reloc_post_snapshot_hook.
1392                  * the source tree is a reloc tree, all tree blocks
1393                  * modified after it was created have RELOC flag
1394                  * set in their headers. so it's OK to not update
1395                  * the 'last_snapshot'.
1396                  */
1397                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1398                                       BTRFS_TREE_RELOC_OBJECTID);
1399                 BUG_ON(ret);
1400         }
1401
1402         memcpy(root_item, &root->root_item, sizeof(*root_item));
1403         btrfs_set_root_bytenr(root_item, eb->start);
1404         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1405         btrfs_set_root_generation(root_item, trans->transid);
1406
1407         if (root->root_key.objectid == objectid) {
1408                 btrfs_set_root_refs(root_item, 0);
1409                 memset(&root_item->drop_progress, 0,
1410                        sizeof(struct btrfs_disk_key));
1411                 root_item->drop_level = 0;
1412                 /*
1413                  * abuse rtransid, it is safe because it is impossible to
1414                  * receive data into a relocation tree.
1415                  */
1416                 btrfs_set_root_rtransid(root_item, last_snap);
1417                 btrfs_set_root_otransid(root_item, trans->transid);
1418         }
1419
1420         btrfs_tree_unlock(eb);
1421         free_extent_buffer(eb);
1422
1423         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1424                                 &root_key, root_item);
1425         BUG_ON(ret);
1426         kfree(root_item);
1427
1428         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1429         BUG_ON(IS_ERR(reloc_root));
1430         reloc_root->last_trans = trans->transid;
1431         return reloc_root;
1432 }
1433
1434 /*
1435  * create reloc tree for a given fs tree. reloc tree is just a
1436  * snapshot of the fs tree with special root objectid.
1437  */
1438 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1439                           struct btrfs_root *root)
1440 {
1441         struct btrfs_root *reloc_root;
1442         struct reloc_control *rc = root->fs_info->reloc_ctl;
1443         struct btrfs_block_rsv *rsv;
1444         int clear_rsv = 0;
1445         int ret;
1446
1447         if (root->reloc_root) {
1448                 reloc_root = root->reloc_root;
1449                 reloc_root->last_trans = trans->transid;
1450                 return 0;
1451         }
1452
1453         if (!rc || !rc->create_reloc_tree ||
1454             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1455                 return 0;
1456
1457         if (!trans->reloc_reserved) {
1458                 rsv = trans->block_rsv;
1459                 trans->block_rsv = rc->block_rsv;
1460                 clear_rsv = 1;
1461         }
1462         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1463         if (clear_rsv)
1464                 trans->block_rsv = rsv;
1465
1466         ret = __add_reloc_root(reloc_root);
1467         BUG_ON(ret < 0);
1468         root->reloc_root = reloc_root;
1469         return 0;
1470 }
1471
1472 /*
1473  * update root item of reloc tree
1474  */
1475 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1476                             struct btrfs_root *root)
1477 {
1478         struct btrfs_root *reloc_root;
1479         struct btrfs_root_item *root_item;
1480         int ret;
1481
1482         if (!root->reloc_root)
1483                 goto out;
1484
1485         reloc_root = root->reloc_root;
1486         root_item = &reloc_root->root_item;
1487
1488         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1489             btrfs_root_refs(root_item) == 0) {
1490                 root->reloc_root = NULL;
1491                 __del_reloc_root(reloc_root);
1492         }
1493
1494         if (reloc_root->commit_root != reloc_root->node) {
1495                 btrfs_set_root_node(root_item, reloc_root->node);
1496                 free_extent_buffer(reloc_root->commit_root);
1497                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1498         }
1499
1500         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1501                                 &reloc_root->root_key, root_item);
1502         BUG_ON(ret);
1503
1504 out:
1505         return 0;
1506 }
1507
1508 /*
1509  * helper to find first cached inode with inode number >= objectid
1510  * in a subvolume
1511  */
1512 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1513 {
1514         struct rb_node *node;
1515         struct rb_node *prev;
1516         struct btrfs_inode *entry;
1517         struct inode *inode;
1518
1519         spin_lock(&root->inode_lock);
1520 again:
1521         node = root->inode_tree.rb_node;
1522         prev = NULL;
1523         while (node) {
1524                 prev = node;
1525                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1526
1527                 if (objectid < btrfs_ino(&entry->vfs_inode))
1528                         node = node->rb_left;
1529                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1530                         node = node->rb_right;
1531                 else
1532                         break;
1533         }
1534         if (!node) {
1535                 while (prev) {
1536                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1537                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1538                                 node = prev;
1539                                 break;
1540                         }
1541                         prev = rb_next(prev);
1542                 }
1543         }
1544         while (node) {
1545                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1546                 inode = igrab(&entry->vfs_inode);
1547                 if (inode) {
1548                         spin_unlock(&root->inode_lock);
1549                         return inode;
1550                 }
1551
1552                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1553                 if (cond_resched_lock(&root->inode_lock))
1554                         goto again;
1555
1556                 node = rb_next(node);
1557         }
1558         spin_unlock(&root->inode_lock);
1559         return NULL;
1560 }
1561
1562 static int in_block_group(u64 bytenr,
1563                           struct btrfs_block_group_cache *block_group)
1564 {
1565         if (bytenr >= block_group->key.objectid &&
1566             bytenr < block_group->key.objectid + block_group->key.offset)
1567                 return 1;
1568         return 0;
1569 }
1570
1571 /*
1572  * get new location of data
1573  */
1574 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1575                             u64 bytenr, u64 num_bytes)
1576 {
1577         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1578         struct btrfs_path *path;
1579         struct btrfs_file_extent_item *fi;
1580         struct extent_buffer *leaf;
1581         int ret;
1582
1583         path = btrfs_alloc_path();
1584         if (!path)
1585                 return -ENOMEM;
1586
1587         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1588         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1589                                        bytenr, 0);
1590         if (ret < 0)
1591                 goto out;
1592         if (ret > 0) {
1593                 ret = -ENOENT;
1594                 goto out;
1595         }
1596
1597         leaf = path->nodes[0];
1598         fi = btrfs_item_ptr(leaf, path->slots[0],
1599                             struct btrfs_file_extent_item);
1600
1601         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1602                btrfs_file_extent_compression(leaf, fi) ||
1603                btrfs_file_extent_encryption(leaf, fi) ||
1604                btrfs_file_extent_other_encoding(leaf, fi));
1605
1606         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1607                 ret = -EINVAL;
1608                 goto out;
1609         }
1610
1611         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1612         ret = 0;
1613 out:
1614         btrfs_free_path(path);
1615         return ret;
1616 }
1617
1618 /*
1619  * update file extent items in the tree leaf to point to
1620  * the new locations.
1621  */
1622 static noinline_for_stack
1623 int replace_file_extents(struct btrfs_trans_handle *trans,
1624                          struct reloc_control *rc,
1625                          struct btrfs_root *root,
1626                          struct extent_buffer *leaf)
1627 {
1628         struct btrfs_key key;
1629         struct btrfs_file_extent_item *fi;
1630         struct inode *inode = NULL;
1631         u64 parent;
1632         u64 bytenr;
1633         u64 new_bytenr = 0;
1634         u64 num_bytes;
1635         u64 end;
1636         u32 nritems;
1637         u32 i;
1638         int ret = 0;
1639         int first = 1;
1640         int dirty = 0;
1641
1642         if (rc->stage != UPDATE_DATA_PTRS)
1643                 return 0;
1644
1645         /* reloc trees always use full backref */
1646         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1647                 parent = leaf->start;
1648         else
1649                 parent = 0;
1650
1651         nritems = btrfs_header_nritems(leaf);
1652         for (i = 0; i < nritems; i++) {
1653                 cond_resched();
1654                 btrfs_item_key_to_cpu(leaf, &key, i);
1655                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1656                         continue;
1657                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1658                 if (btrfs_file_extent_type(leaf, fi) ==
1659                     BTRFS_FILE_EXTENT_INLINE)
1660                         continue;
1661                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1662                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1663                 if (bytenr == 0)
1664                         continue;
1665                 if (!in_block_group(bytenr, rc->block_group))
1666                         continue;
1667
1668                 /*
1669                  * if we are modifying block in fs tree, wait for readpage
1670                  * to complete and drop the extent cache
1671                  */
1672                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1673                         if (first) {
1674                                 inode = find_next_inode(root, key.objectid);
1675                                 first = 0;
1676                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1677                                 btrfs_add_delayed_iput(inode);
1678                                 inode = find_next_inode(root, key.objectid);
1679                         }
1680                         if (inode && btrfs_ino(inode) == key.objectid) {
1681                                 end = key.offset +
1682                                       btrfs_file_extent_num_bytes(leaf, fi);
1683                                 WARN_ON(!IS_ALIGNED(key.offset,
1684                                                     root->sectorsize));
1685                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1686                                 end--;
1687                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1688                                                       key.offset, end);
1689                                 if (!ret)
1690                                         continue;
1691
1692                                 btrfs_drop_extent_cache(inode, key.offset, end,
1693                                                         1);
1694                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1695                                               key.offset, end);
1696                         }
1697                 }
1698
1699                 ret = get_new_location(rc->data_inode, &new_bytenr,
1700                                        bytenr, num_bytes);
1701                 if (ret) {
1702                         /*
1703                          * Don't have to abort since we've not changed anything
1704                          * in the file extent yet.
1705                          */
1706                         break;
1707                 }
1708
1709                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1710                 dirty = 1;
1711
1712                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1713                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1714                                            num_bytes, parent,
1715                                            btrfs_header_owner(leaf),
1716                                            key.objectid, key.offset, 1);
1717                 if (ret) {
1718                         btrfs_abort_transaction(trans, root, ret);
1719                         break;
1720                 }
1721
1722                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1723                                         parent, btrfs_header_owner(leaf),
1724                                         key.objectid, key.offset, 1);
1725                 if (ret) {
1726                         btrfs_abort_transaction(trans, root, ret);
1727                         break;
1728                 }
1729         }
1730         if (dirty)
1731                 btrfs_mark_buffer_dirty(leaf);
1732         if (inode)
1733                 btrfs_add_delayed_iput(inode);
1734         return ret;
1735 }
1736
1737 static noinline_for_stack
1738 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1739                      struct btrfs_path *path, int level)
1740 {
1741         struct btrfs_disk_key key1;
1742         struct btrfs_disk_key key2;
1743         btrfs_node_key(eb, &key1, slot);
1744         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1745         return memcmp(&key1, &key2, sizeof(key1));
1746 }
1747
1748 /*
1749  * try to replace tree blocks in fs tree with the new blocks
1750  * in reloc tree. tree blocks haven't been modified since the
1751  * reloc tree was create can be replaced.
1752  *
1753  * if a block was replaced, level of the block + 1 is returned.
1754  * if no block got replaced, 0 is returned. if there are other
1755  * errors, a negative error number is returned.
1756  */
1757 static noinline_for_stack
1758 int replace_path(struct btrfs_trans_handle *trans,
1759                  struct btrfs_root *dest, struct btrfs_root *src,
1760                  struct btrfs_path *path, struct btrfs_key *next_key,
1761                  int lowest_level, int max_level)
1762 {
1763         struct extent_buffer *eb;
1764         struct extent_buffer *parent;
1765         struct btrfs_key key;
1766         u64 old_bytenr;
1767         u64 new_bytenr;
1768         u64 old_ptr_gen;
1769         u64 new_ptr_gen;
1770         u64 last_snapshot;
1771         u32 blocksize;
1772         int cow = 0;
1773         int level;
1774         int ret;
1775         int slot;
1776
1777         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1778         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1779
1780         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1781 again:
1782         slot = path->slots[lowest_level];
1783         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1784
1785         eb = btrfs_lock_root_node(dest);
1786         btrfs_set_lock_blocking(eb);
1787         level = btrfs_header_level(eb);
1788
1789         if (level < lowest_level) {
1790                 btrfs_tree_unlock(eb);
1791                 free_extent_buffer(eb);
1792                 return 0;
1793         }
1794
1795         if (cow) {
1796                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1797                 BUG_ON(ret);
1798         }
1799         btrfs_set_lock_blocking(eb);
1800
1801         if (next_key) {
1802                 next_key->objectid = (u64)-1;
1803                 next_key->type = (u8)-1;
1804                 next_key->offset = (u64)-1;
1805         }
1806
1807         parent = eb;
1808         while (1) {
1809                 level = btrfs_header_level(parent);
1810                 BUG_ON(level < lowest_level);
1811
1812                 ret = btrfs_bin_search(parent, &key, level, &slot);
1813                 if (ret && slot > 0)
1814                         slot--;
1815
1816                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1817                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1818
1819                 old_bytenr = btrfs_node_blockptr(parent, slot);
1820                 blocksize = dest->nodesize;
1821                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1822
1823                 if (level <= max_level) {
1824                         eb = path->nodes[level];
1825                         new_bytenr = btrfs_node_blockptr(eb,
1826                                                         path->slots[level]);
1827                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1828                                                         path->slots[level]);
1829                 } else {
1830                         new_bytenr = 0;
1831                         new_ptr_gen = 0;
1832                 }
1833
1834                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1835                         ret = level;
1836                         break;
1837                 }
1838
1839                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1840                     memcmp_node_keys(parent, slot, path, level)) {
1841                         if (level <= lowest_level) {
1842                                 ret = 0;
1843                                 break;
1844                         }
1845
1846                         eb = read_tree_block(dest, old_bytenr, blocksize,
1847                                              old_ptr_gen);
1848                         if (!eb || !extent_buffer_uptodate(eb)) {
1849                                 ret = (!eb) ? -ENOMEM : -EIO;
1850                                 free_extent_buffer(eb);
1851                                 break;
1852                         }
1853                         btrfs_tree_lock(eb);
1854                         if (cow) {
1855                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1856                                                       slot, &eb);
1857                                 BUG_ON(ret);
1858                         }
1859                         btrfs_set_lock_blocking(eb);
1860
1861                         btrfs_tree_unlock(parent);
1862                         free_extent_buffer(parent);
1863
1864                         parent = eb;
1865                         continue;
1866                 }
1867
1868                 if (!cow) {
1869                         btrfs_tree_unlock(parent);
1870                         free_extent_buffer(parent);
1871                         cow = 1;
1872                         goto again;
1873                 }
1874
1875                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1876                                       path->slots[level]);
1877                 btrfs_release_path(path);
1878
1879                 path->lowest_level = level;
1880                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1881                 path->lowest_level = 0;
1882                 BUG_ON(ret);
1883
1884                 /*
1885                  * swap blocks in fs tree and reloc tree.
1886                  */
1887                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1888                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1889                 btrfs_mark_buffer_dirty(parent);
1890
1891                 btrfs_set_node_blockptr(path->nodes[level],
1892                                         path->slots[level], old_bytenr);
1893                 btrfs_set_node_ptr_generation(path->nodes[level],
1894                                               path->slots[level], old_ptr_gen);
1895                 btrfs_mark_buffer_dirty(path->nodes[level]);
1896
1897                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1898                                         path->nodes[level]->start,
1899                                         src->root_key.objectid, level - 1, 0,
1900                                         1);
1901                 BUG_ON(ret);
1902                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1903                                         0, dest->root_key.objectid, level - 1,
1904                                         0, 1);
1905                 BUG_ON(ret);
1906
1907                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1908                                         path->nodes[level]->start,
1909                                         src->root_key.objectid, level - 1, 0,
1910                                         1);
1911                 BUG_ON(ret);
1912
1913                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1914                                         0, dest->root_key.objectid, level - 1,
1915                                         0, 1);
1916                 BUG_ON(ret);
1917
1918                 btrfs_unlock_up_safe(path, 0);
1919
1920                 ret = level;
1921                 break;
1922         }
1923         btrfs_tree_unlock(parent);
1924         free_extent_buffer(parent);
1925         return ret;
1926 }
1927
1928 /*
1929  * helper to find next relocated block in reloc tree
1930  */
1931 static noinline_for_stack
1932 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1933                        int *level)
1934 {
1935         struct extent_buffer *eb;
1936         int i;
1937         u64 last_snapshot;
1938         u32 nritems;
1939
1940         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1941
1942         for (i = 0; i < *level; i++) {
1943                 free_extent_buffer(path->nodes[i]);
1944                 path->nodes[i] = NULL;
1945         }
1946
1947         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1948                 eb = path->nodes[i];
1949                 nritems = btrfs_header_nritems(eb);
1950                 while (path->slots[i] + 1 < nritems) {
1951                         path->slots[i]++;
1952                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1953                             last_snapshot)
1954                                 continue;
1955
1956                         *level = i;
1957                         return 0;
1958                 }
1959                 free_extent_buffer(path->nodes[i]);
1960                 path->nodes[i] = NULL;
1961         }
1962         return 1;
1963 }
1964
1965 /*
1966  * walk down reloc tree to find relocated block of lowest level
1967  */
1968 static noinline_for_stack
1969 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1970                          int *level)
1971 {
1972         struct extent_buffer *eb = NULL;
1973         int i;
1974         u64 bytenr;
1975         u64 ptr_gen = 0;
1976         u64 last_snapshot;
1977         u32 blocksize;
1978         u32 nritems;
1979
1980         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1981
1982         for (i = *level; i > 0; i--) {
1983                 eb = path->nodes[i];
1984                 nritems = btrfs_header_nritems(eb);
1985                 while (path->slots[i] < nritems) {
1986                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1987                         if (ptr_gen > last_snapshot)
1988                                 break;
1989                         path->slots[i]++;
1990                 }
1991                 if (path->slots[i] >= nritems) {
1992                         if (i == *level)
1993                                 break;
1994                         *level = i + 1;
1995                         return 0;
1996                 }
1997                 if (i == 1) {
1998                         *level = i;
1999                         return 0;
2000                 }
2001
2002                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2003                 blocksize = root->nodesize;
2004                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
2005                 if (!eb || !extent_buffer_uptodate(eb)) {
2006                         free_extent_buffer(eb);
2007                         return -EIO;
2008                 }
2009                 BUG_ON(btrfs_header_level(eb) != i - 1);
2010                 path->nodes[i - 1] = eb;
2011                 path->slots[i - 1] = 0;
2012         }
2013         return 1;
2014 }
2015
2016 /*
2017  * invalidate extent cache for file extents whose key in range of
2018  * [min_key, max_key)
2019  */
2020 static int invalidate_extent_cache(struct btrfs_root *root,
2021                                    struct btrfs_key *min_key,
2022                                    struct btrfs_key *max_key)
2023 {
2024         struct inode *inode = NULL;
2025         u64 objectid;
2026         u64 start, end;
2027         u64 ino;
2028
2029         objectid = min_key->objectid;
2030         while (1) {
2031                 cond_resched();
2032                 iput(inode);
2033
2034                 if (objectid > max_key->objectid)
2035                         break;
2036
2037                 inode = find_next_inode(root, objectid);
2038                 if (!inode)
2039                         break;
2040                 ino = btrfs_ino(inode);
2041
2042                 if (ino > max_key->objectid) {
2043                         iput(inode);
2044                         break;
2045                 }
2046
2047                 objectid = ino + 1;
2048                 if (!S_ISREG(inode->i_mode))
2049                         continue;
2050
2051                 if (unlikely(min_key->objectid == ino)) {
2052                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2053                                 continue;
2054                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2055                                 start = 0;
2056                         else {
2057                                 start = min_key->offset;
2058                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2059                         }
2060                 } else {
2061                         start = 0;
2062                 }
2063
2064                 if (unlikely(max_key->objectid == ino)) {
2065                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2066                                 continue;
2067                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2068                                 end = (u64)-1;
2069                         } else {
2070                                 if (max_key->offset == 0)
2071                                         continue;
2072                                 end = max_key->offset;
2073                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2074                                 end--;
2075                         }
2076                 } else {
2077                         end = (u64)-1;
2078                 }
2079
2080                 /* the lock_extent waits for readpage to complete */
2081                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2082                 btrfs_drop_extent_cache(inode, start, end, 1);
2083                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2084         }
2085         return 0;
2086 }
2087
2088 static int find_next_key(struct btrfs_path *path, int level,
2089                          struct btrfs_key *key)
2090
2091 {
2092         while (level < BTRFS_MAX_LEVEL) {
2093                 if (!path->nodes[level])
2094                         break;
2095                 if (path->slots[level] + 1 <
2096                     btrfs_header_nritems(path->nodes[level])) {
2097                         btrfs_node_key_to_cpu(path->nodes[level], key,
2098                                               path->slots[level] + 1);
2099                         return 0;
2100                 }
2101                 level++;
2102         }
2103         return 1;
2104 }
2105
2106 /*
2107  * merge the relocated tree blocks in reloc tree with corresponding
2108  * fs tree.
2109  */
2110 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2111                                                struct btrfs_root *root)
2112 {
2113         LIST_HEAD(inode_list);
2114         struct btrfs_key key;
2115         struct btrfs_key next_key;
2116         struct btrfs_trans_handle *trans = NULL;
2117         struct btrfs_root *reloc_root;
2118         struct btrfs_root_item *root_item;
2119         struct btrfs_path *path;
2120         struct extent_buffer *leaf;
2121         int level;
2122         int max_level;
2123         int replaced = 0;
2124         int ret;
2125         int err = 0;
2126         u32 min_reserved;
2127
2128         path = btrfs_alloc_path();
2129         if (!path)
2130                 return -ENOMEM;
2131         path->reada = 1;
2132
2133         reloc_root = root->reloc_root;
2134         root_item = &reloc_root->root_item;
2135
2136         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2137                 level = btrfs_root_level(root_item);
2138                 extent_buffer_get(reloc_root->node);
2139                 path->nodes[level] = reloc_root->node;
2140                 path->slots[level] = 0;
2141         } else {
2142                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2143
2144                 level = root_item->drop_level;
2145                 BUG_ON(level == 0);
2146                 path->lowest_level = level;
2147                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2148                 path->lowest_level = 0;
2149                 if (ret < 0) {
2150                         btrfs_free_path(path);
2151                         return ret;
2152                 }
2153
2154                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2155                                       path->slots[level]);
2156                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2157
2158                 btrfs_unlock_up_safe(path, 0);
2159         }
2160
2161         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2162         memset(&next_key, 0, sizeof(next_key));
2163
2164         while (1) {
2165                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2166                                              BTRFS_RESERVE_FLUSH_ALL);
2167                 if (ret) {
2168                         err = ret;
2169                         goto out;
2170                 }
2171                 trans = btrfs_start_transaction(root, 0);
2172                 if (IS_ERR(trans)) {
2173                         err = PTR_ERR(trans);
2174                         trans = NULL;
2175                         goto out;
2176                 }
2177                 trans->block_rsv = rc->block_rsv;
2178
2179                 replaced = 0;
2180                 max_level = level;
2181
2182                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2183                 if (ret < 0) {
2184                         err = ret;
2185                         goto out;
2186                 }
2187                 if (ret > 0)
2188                         break;
2189
2190                 if (!find_next_key(path, level, &key) &&
2191                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2192                         ret = 0;
2193                 } else {
2194                         ret = replace_path(trans, root, reloc_root, path,
2195                                            &next_key, level, max_level);
2196                 }
2197                 if (ret < 0) {
2198                         err = ret;
2199                         goto out;
2200                 }
2201
2202                 if (ret > 0) {
2203                         level = ret;
2204                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2205                                               path->slots[level]);
2206                         replaced = 1;
2207                 }
2208
2209                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2210                 if (ret > 0)
2211                         break;
2212
2213                 BUG_ON(level == 0);
2214                 /*
2215                  * save the merging progress in the drop_progress.
2216                  * this is OK since root refs == 1 in this case.
2217                  */
2218                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2219                                path->slots[level]);
2220                 root_item->drop_level = level;
2221
2222                 btrfs_end_transaction_throttle(trans, root);
2223                 trans = NULL;
2224
2225                 btrfs_btree_balance_dirty(root);
2226
2227                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2228                         invalidate_extent_cache(root, &key, &next_key);
2229         }
2230
2231         /*
2232          * handle the case only one block in the fs tree need to be
2233          * relocated and the block is tree root.
2234          */
2235         leaf = btrfs_lock_root_node(root);
2236         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2237         btrfs_tree_unlock(leaf);
2238         free_extent_buffer(leaf);
2239         if (ret < 0)
2240                 err = ret;
2241 out:
2242         btrfs_free_path(path);
2243
2244         if (err == 0) {
2245                 memset(&root_item->drop_progress, 0,
2246                        sizeof(root_item->drop_progress));
2247                 root_item->drop_level = 0;
2248                 btrfs_set_root_refs(root_item, 0);
2249                 btrfs_update_reloc_root(trans, root);
2250         }
2251
2252         if (trans)
2253                 btrfs_end_transaction_throttle(trans, root);
2254
2255         btrfs_btree_balance_dirty(root);
2256
2257         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2258                 invalidate_extent_cache(root, &key, &next_key);
2259
2260         return err;
2261 }
2262
2263 static noinline_for_stack
2264 int prepare_to_merge(struct reloc_control *rc, int err)
2265 {
2266         struct btrfs_root *root = rc->extent_root;
2267         struct btrfs_root *reloc_root;
2268         struct btrfs_trans_handle *trans;
2269         LIST_HEAD(reloc_roots);
2270         u64 num_bytes = 0;
2271         int ret;
2272
2273         mutex_lock(&root->fs_info->reloc_mutex);
2274         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2275         rc->merging_rsv_size += rc->nodes_relocated * 2;
2276         mutex_unlock(&root->fs_info->reloc_mutex);
2277
2278 again:
2279         if (!err) {
2280                 num_bytes = rc->merging_rsv_size;
2281                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2282                                           BTRFS_RESERVE_FLUSH_ALL);
2283                 if (ret)
2284                         err = ret;
2285         }
2286
2287         trans = btrfs_join_transaction(rc->extent_root);
2288         if (IS_ERR(trans)) {
2289                 if (!err)
2290                         btrfs_block_rsv_release(rc->extent_root,
2291                                                 rc->block_rsv, num_bytes);
2292                 return PTR_ERR(trans);
2293         }
2294
2295         if (!err) {
2296                 if (num_bytes != rc->merging_rsv_size) {
2297                         btrfs_end_transaction(trans, rc->extent_root);
2298                         btrfs_block_rsv_release(rc->extent_root,
2299                                                 rc->block_rsv, num_bytes);
2300                         goto again;
2301                 }
2302         }
2303
2304         rc->merge_reloc_tree = 1;
2305
2306         while (!list_empty(&rc->reloc_roots)) {
2307                 reloc_root = list_entry(rc->reloc_roots.next,
2308                                         struct btrfs_root, root_list);
2309                 list_del_init(&reloc_root->root_list);
2310
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                 /*
2317                  * set reference count to 1, so btrfs_recover_relocation
2318                  * knows it should resumes merging
2319                  */
2320                 if (!err)
2321                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2322                 btrfs_update_reloc_root(trans, root);
2323
2324                 list_add(&reloc_root->root_list, &reloc_roots);
2325         }
2326
2327         list_splice(&reloc_roots, &rc->reloc_roots);
2328
2329         if (!err)
2330                 btrfs_commit_transaction(trans, rc->extent_root);
2331         else
2332                 btrfs_end_transaction(trans, rc->extent_root);
2333         return err;
2334 }
2335
2336 static noinline_for_stack
2337 void free_reloc_roots(struct list_head *list)
2338 {
2339         struct btrfs_root *reloc_root;
2340
2341         while (!list_empty(list)) {
2342                 reloc_root = list_entry(list->next, struct btrfs_root,
2343                                         root_list);
2344                 __del_reloc_root(reloc_root);
2345         }
2346 }
2347
2348 static noinline_for_stack
2349 int merge_reloc_roots(struct reloc_control *rc)
2350 {
2351         struct btrfs_root *root;
2352         struct btrfs_root *reloc_root;
2353         u64 last_snap;
2354         u64 otransid;
2355         u64 objectid;
2356         LIST_HEAD(reloc_roots);
2357         int found = 0;
2358         int ret = 0;
2359 again:
2360         root = rc->extent_root;
2361
2362         /*
2363          * this serializes us with btrfs_record_root_in_transaction,
2364          * we have to make sure nobody is in the middle of
2365          * adding their roots to the list while we are
2366          * doing this splice
2367          */
2368         mutex_lock(&root->fs_info->reloc_mutex);
2369         list_splice_init(&rc->reloc_roots, &reloc_roots);
2370         mutex_unlock(&root->fs_info->reloc_mutex);
2371
2372         while (!list_empty(&reloc_roots)) {
2373                 found = 1;
2374                 reloc_root = list_entry(reloc_roots.next,
2375                                         struct btrfs_root, root_list);
2376
2377                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2378                         root = read_fs_root(reloc_root->fs_info,
2379                                             reloc_root->root_key.offset);
2380                         BUG_ON(IS_ERR(root));
2381                         BUG_ON(root->reloc_root != reloc_root);
2382
2383                         ret = merge_reloc_root(rc, root);
2384                         if (ret) {
2385                                 if (list_empty(&reloc_root->root_list))
2386                                         list_add_tail(&reloc_root->root_list,
2387                                                       &reloc_roots);
2388                                 goto out;
2389                         }
2390                 } else {
2391                         list_del_init(&reloc_root->root_list);
2392                 }
2393
2394                 /*
2395                  * we keep the old last snapshod transid in rtranid when we
2396                  * created the relocation tree.
2397                  */
2398                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2399                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2400                 objectid = reloc_root->root_key.offset;
2401
2402                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2403                 if (ret < 0) {
2404                         if (list_empty(&reloc_root->root_list))
2405                                 list_add_tail(&reloc_root->root_list,
2406                                               &reloc_roots);
2407                         goto out;
2408                 }
2409         }
2410
2411         if (found) {
2412                 found = 0;
2413                 goto again;
2414         }
2415 out:
2416         if (ret) {
2417                 btrfs_std_error(root->fs_info, ret);
2418                 if (!list_empty(&reloc_roots))
2419                         free_reloc_roots(&reloc_roots);
2420
2421                 /* new reloc root may be added */
2422                 mutex_lock(&root->fs_info->reloc_mutex);
2423                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2424                 mutex_unlock(&root->fs_info->reloc_mutex);
2425                 if (!list_empty(&reloc_roots))
2426                         free_reloc_roots(&reloc_roots);
2427         }
2428
2429         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2430         return ret;
2431 }
2432
2433 static void free_block_list(struct rb_root *blocks)
2434 {
2435         struct tree_block *block;
2436         struct rb_node *rb_node;
2437         while ((rb_node = rb_first(blocks))) {
2438                 block = rb_entry(rb_node, struct tree_block, rb_node);
2439                 rb_erase(rb_node, blocks);
2440                 kfree(block);
2441         }
2442 }
2443
2444 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2445                                       struct btrfs_root *reloc_root)
2446 {
2447         struct btrfs_root *root;
2448
2449         if (reloc_root->last_trans == trans->transid)
2450                 return 0;
2451
2452         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2453         BUG_ON(IS_ERR(root));
2454         BUG_ON(root->reloc_root != reloc_root);
2455
2456         return btrfs_record_root_in_trans(trans, root);
2457 }
2458
2459 static noinline_for_stack
2460 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2461                                      struct reloc_control *rc,
2462                                      struct backref_node *node,
2463                                      struct backref_edge *edges[])
2464 {
2465         struct backref_node *next;
2466         struct btrfs_root *root;
2467         int index = 0;
2468
2469         next = node;
2470         while (1) {
2471                 cond_resched();
2472                 next = walk_up_backref(next, edges, &index);
2473                 root = next->root;
2474                 BUG_ON(!root);
2475                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2476
2477                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2478                         record_reloc_root_in_trans(trans, root);
2479                         break;
2480                 }
2481
2482                 btrfs_record_root_in_trans(trans, root);
2483                 root = root->reloc_root;
2484
2485                 if (next->new_bytenr != root->node->start) {
2486                         BUG_ON(next->new_bytenr);
2487                         BUG_ON(!list_empty(&next->list));
2488                         next->new_bytenr = root->node->start;
2489                         next->root = root;
2490                         list_add_tail(&next->list,
2491                                       &rc->backref_cache.changed);
2492                         __mark_block_processed(rc, next);
2493                         break;
2494                 }
2495
2496                 WARN_ON(1);
2497                 root = NULL;
2498                 next = walk_down_backref(edges, &index);
2499                 if (!next || next->level <= node->level)
2500                         break;
2501         }
2502         if (!root)
2503                 return NULL;
2504
2505         next = node;
2506         /* setup backref node path for btrfs_reloc_cow_block */
2507         while (1) {
2508                 rc->backref_cache.path[next->level] = next;
2509                 if (--index < 0)
2510                         break;
2511                 next = edges[index]->node[UPPER];
2512         }
2513         return root;
2514 }
2515
2516 /*
2517  * select a tree root for relocation. return NULL if the block
2518  * is reference counted. we should use do_relocation() in this
2519  * case. return a tree root pointer if the block isn't reference
2520  * counted. return -ENOENT if the block is root of reloc tree.
2521  */
2522 static noinline_for_stack
2523 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2524                                    struct backref_node *node)
2525 {
2526         struct backref_node *next;
2527         struct btrfs_root *root;
2528         struct btrfs_root *fs_root = NULL;
2529         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2530         int index = 0;
2531
2532         next = node;
2533         while (1) {
2534                 cond_resched();
2535                 next = walk_up_backref(next, edges, &index);
2536                 root = next->root;
2537                 BUG_ON(!root);
2538
2539                 /* no other choice for non-references counted tree */
2540                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2541                         return root;
2542
2543                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2544                         fs_root = root;
2545
2546                 if (next != node)
2547                         return NULL;
2548
2549                 next = walk_down_backref(edges, &index);
2550                 if (!next || next->level <= node->level)
2551                         break;
2552         }
2553
2554         if (!fs_root)
2555                 return ERR_PTR(-ENOENT);
2556         return fs_root;
2557 }
2558
2559 static noinline_for_stack
2560 u64 calcu_metadata_size(struct reloc_control *rc,
2561                         struct backref_node *node, int reserve)
2562 {
2563         struct backref_node *next = node;
2564         struct backref_edge *edge;
2565         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2566         u64 num_bytes = 0;
2567         int index = 0;
2568
2569         BUG_ON(reserve && node->processed);
2570
2571         while (next) {
2572                 cond_resched();
2573                 while (1) {
2574                         if (next->processed && (reserve || next != node))
2575                                 break;
2576
2577                         num_bytes += rc->extent_root->nodesize;
2578
2579                         if (list_empty(&next->upper))
2580                                 break;
2581
2582                         edge = list_entry(next->upper.next,
2583                                           struct backref_edge, list[LOWER]);
2584                         edges[index++] = edge;
2585                         next = edge->node[UPPER];
2586                 }
2587                 next = walk_down_backref(edges, &index);
2588         }
2589         return num_bytes;
2590 }
2591
2592 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2593                                   struct reloc_control *rc,
2594                                   struct backref_node *node)
2595 {
2596         struct btrfs_root *root = rc->extent_root;
2597         u64 num_bytes;
2598         int ret;
2599         u64 tmp;
2600
2601         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2602
2603         trans->block_rsv = rc->block_rsv;
2604         rc->reserved_bytes += num_bytes;
2605         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2606                                 BTRFS_RESERVE_FLUSH_ALL);
2607         if (ret) {
2608                 if (ret == -EAGAIN) {
2609                         tmp = rc->extent_root->nodesize *
2610                                 RELOCATION_RESERVED_NODES;
2611                         while (tmp <= rc->reserved_bytes)
2612                                 tmp <<= 1;
2613                         /*
2614                          * only one thread can access block_rsv at this point,
2615                          * so we don't need hold lock to protect block_rsv.
2616                          * we expand more reservation size here to allow enough
2617                          * space for relocation and we will return eailer in
2618                          * enospc case.
2619                          */
2620                         rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2621                                               RELOCATION_RESERVED_NODES;
2622                 }
2623                 return ret;
2624         }
2625
2626         return 0;
2627 }
2628
2629 /*
2630  * relocate a block tree, and then update pointers in upper level
2631  * blocks that reference the block to point to the new location.
2632  *
2633  * if called by link_to_upper, the block has already been relocated.
2634  * in that case this function just updates pointers.
2635  */
2636 static int do_relocation(struct btrfs_trans_handle *trans,
2637                          struct reloc_control *rc,
2638                          struct backref_node *node,
2639                          struct btrfs_key *key,
2640                          struct btrfs_path *path, int lowest)
2641 {
2642         struct backref_node *upper;
2643         struct backref_edge *edge;
2644         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2645         struct btrfs_root *root;
2646         struct extent_buffer *eb;
2647         u32 blocksize;
2648         u64 bytenr;
2649         u64 generation;
2650         int slot;
2651         int ret;
2652         int err = 0;
2653
2654         BUG_ON(lowest && node->eb);
2655
2656         path->lowest_level = node->level + 1;
2657         rc->backref_cache.path[node->level] = node;
2658         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2659                 cond_resched();
2660
2661                 upper = edge->node[UPPER];
2662                 root = select_reloc_root(trans, rc, upper, edges);
2663                 BUG_ON(!root);
2664
2665                 if (upper->eb && !upper->locked) {
2666                         if (!lowest) {
2667                                 ret = btrfs_bin_search(upper->eb, key,
2668                                                        upper->level, &slot);
2669                                 BUG_ON(ret);
2670                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2671                                 if (node->eb->start == bytenr)
2672                                         goto next;
2673                         }
2674                         drop_node_buffer(upper);
2675                 }
2676
2677                 if (!upper->eb) {
2678                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2679                         if (ret < 0) {
2680                                 err = ret;
2681                                 break;
2682                         }
2683                         BUG_ON(ret > 0);
2684
2685                         if (!upper->eb) {
2686                                 upper->eb = path->nodes[upper->level];
2687                                 path->nodes[upper->level] = NULL;
2688                         } else {
2689                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2690                         }
2691
2692                         upper->locked = 1;
2693                         path->locks[upper->level] = 0;
2694
2695                         slot = path->slots[upper->level];
2696                         btrfs_release_path(path);
2697                 } else {
2698                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2699                                                &slot);
2700                         BUG_ON(ret);
2701                 }
2702
2703                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2704                 if (lowest) {
2705                         BUG_ON(bytenr != node->bytenr);
2706                 } else {
2707                         if (node->eb->start == bytenr)
2708                                 goto next;
2709                 }
2710
2711                 blocksize = root->nodesize;
2712                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2713                 eb = read_tree_block(root, bytenr, blocksize, generation);
2714                 if (!eb || !extent_buffer_uptodate(eb)) {
2715                         free_extent_buffer(eb);
2716                         err = -EIO;
2717                         goto next;
2718                 }
2719                 btrfs_tree_lock(eb);
2720                 btrfs_set_lock_blocking(eb);
2721
2722                 if (!node->eb) {
2723                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2724                                               slot, &eb);
2725                         btrfs_tree_unlock(eb);
2726                         free_extent_buffer(eb);
2727                         if (ret < 0) {
2728                                 err = ret;
2729                                 goto next;
2730                         }
2731                         BUG_ON(node->eb != eb);
2732                 } else {
2733                         btrfs_set_node_blockptr(upper->eb, slot,
2734                                                 node->eb->start);
2735                         btrfs_set_node_ptr_generation(upper->eb, slot,
2736                                                       trans->transid);
2737                         btrfs_mark_buffer_dirty(upper->eb);
2738
2739                         ret = btrfs_inc_extent_ref(trans, root,
2740                                                 node->eb->start, blocksize,
2741                                                 upper->eb->start,
2742                                                 btrfs_header_owner(upper->eb),
2743                                                 node->level, 0, 1);
2744                         BUG_ON(ret);
2745
2746                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2747                         BUG_ON(ret);
2748                 }
2749 next:
2750                 if (!upper->pending)
2751                         drop_node_buffer(upper);
2752                 else
2753                         unlock_node_buffer(upper);
2754                 if (err)
2755                         break;
2756         }
2757
2758         if (!err && node->pending) {
2759                 drop_node_buffer(node);
2760                 list_move_tail(&node->list, &rc->backref_cache.changed);
2761                 node->pending = 0;
2762         }
2763
2764         path->lowest_level = 0;
2765         BUG_ON(err == -ENOSPC);
2766         return err;
2767 }
2768
2769 static int link_to_upper(struct btrfs_trans_handle *trans,
2770                          struct reloc_control *rc,
2771                          struct backref_node *node,
2772                          struct btrfs_path *path)
2773 {
2774         struct btrfs_key key;
2775
2776         btrfs_node_key_to_cpu(node->eb, &key, 0);
2777         return do_relocation(trans, rc, node, &key, path, 0);
2778 }
2779
2780 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2781                                 struct reloc_control *rc,
2782                                 struct btrfs_path *path, int err)
2783 {
2784         LIST_HEAD(list);
2785         struct backref_cache *cache = &rc->backref_cache;
2786         struct backref_node *node;
2787         int level;
2788         int ret;
2789
2790         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2791                 while (!list_empty(&cache->pending[level])) {
2792                         node = list_entry(cache->pending[level].next,
2793                                           struct backref_node, list);
2794                         list_move_tail(&node->list, &list);
2795                         BUG_ON(!node->pending);
2796
2797                         if (!err) {
2798                                 ret = link_to_upper(trans, rc, node, path);
2799                                 if (ret < 0)
2800                                         err = ret;
2801                         }
2802                 }
2803                 list_splice_init(&list, &cache->pending[level]);
2804         }
2805         return err;
2806 }
2807
2808 static void mark_block_processed(struct reloc_control *rc,
2809                                  u64 bytenr, u32 blocksize)
2810 {
2811         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2812                         EXTENT_DIRTY, GFP_NOFS);
2813 }
2814
2815 static void __mark_block_processed(struct reloc_control *rc,
2816                                    struct backref_node *node)
2817 {
2818         u32 blocksize;
2819         if (node->level == 0 ||
2820             in_block_group(node->bytenr, rc->block_group)) {
2821                 blocksize = rc->extent_root->nodesize;
2822                 mark_block_processed(rc, node->bytenr, blocksize);
2823         }
2824         node->processed = 1;
2825 }
2826
2827 /*
2828  * mark a block and all blocks directly/indirectly reference the block
2829  * as processed.
2830  */
2831 static void update_processed_blocks(struct reloc_control *rc,
2832                                     struct backref_node *node)
2833 {
2834         struct backref_node *next = node;
2835         struct backref_edge *edge;
2836         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2837         int index = 0;
2838
2839         while (next) {
2840                 cond_resched();
2841                 while (1) {
2842                         if (next->processed)
2843                                 break;
2844
2845                         __mark_block_processed(rc, next);
2846
2847                         if (list_empty(&next->upper))
2848                                 break;
2849
2850                         edge = list_entry(next->upper.next,
2851                                           struct backref_edge, list[LOWER]);
2852                         edges[index++] = edge;
2853                         next = edge->node[UPPER];
2854                 }
2855                 next = walk_down_backref(edges, &index);
2856         }
2857 }
2858
2859 static int tree_block_processed(u64 bytenr, u32 blocksize,
2860                                 struct reloc_control *rc)
2861 {
2862         if (test_range_bit(&rc->processed_blocks, bytenr,
2863                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2864                 return 1;
2865         return 0;
2866 }
2867
2868 static int get_tree_block_key(struct reloc_control *rc,
2869                               struct tree_block *block)
2870 {
2871         struct extent_buffer *eb;
2872
2873         BUG_ON(block->key_ready);
2874         eb = read_tree_block(rc->extent_root, block->bytenr,
2875                              block->key.objectid, block->key.offset);
2876         if (!eb || !extent_buffer_uptodate(eb)) {
2877                 free_extent_buffer(eb);
2878                 return -EIO;
2879         }
2880         WARN_ON(btrfs_header_level(eb) != block->level);
2881         if (block->level == 0)
2882                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2883         else
2884                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2885         free_extent_buffer(eb);
2886         block->key_ready = 1;
2887         return 0;
2888 }
2889
2890 static int reada_tree_block(struct reloc_control *rc,
2891                             struct tree_block *block)
2892 {
2893         BUG_ON(block->key_ready);
2894         if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2895                 readahead_tree_block(rc->extent_root, block->bytenr,
2896                                      block->key.objectid,
2897                                      rc->extent_root->nodesize);
2898         else
2899                 readahead_tree_block(rc->extent_root, block->bytenr,
2900                                      block->key.objectid, block->key.offset);
2901         return 0;
2902 }
2903
2904 /*
2905  * helper function to relocate a tree block
2906  */
2907 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2908                                 struct reloc_control *rc,
2909                                 struct backref_node *node,
2910                                 struct btrfs_key *key,
2911                                 struct btrfs_path *path)
2912 {
2913         struct btrfs_root *root;
2914         int ret = 0;
2915
2916         if (!node)
2917                 return 0;
2918
2919         BUG_ON(node->processed);
2920         root = select_one_root(trans, node);
2921         if (root == ERR_PTR(-ENOENT)) {
2922                 update_processed_blocks(rc, node);
2923                 goto out;
2924         }
2925
2926         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2927                 ret = reserve_metadata_space(trans, rc, node);
2928                 if (ret)
2929                         goto out;
2930         }
2931
2932         if (root) {
2933                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2934                         BUG_ON(node->new_bytenr);
2935                         BUG_ON(!list_empty(&node->list));
2936                         btrfs_record_root_in_trans(trans, root);
2937                         root = root->reloc_root;
2938                         node->new_bytenr = root->node->start;
2939                         node->root = root;
2940                         list_add_tail(&node->list, &rc->backref_cache.changed);
2941                 } else {
2942                         path->lowest_level = node->level;
2943                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2944                         btrfs_release_path(path);
2945                         if (ret > 0)
2946                                 ret = 0;
2947                 }
2948                 if (!ret)
2949                         update_processed_blocks(rc, node);
2950         } else {
2951                 ret = do_relocation(trans, rc, node, key, path, 1);
2952         }
2953 out:
2954         if (ret || node->level == 0 || node->cowonly)
2955                 remove_backref_node(&rc->backref_cache, node);
2956         return ret;
2957 }
2958
2959 /*
2960  * relocate a list of blocks
2961  */
2962 static noinline_for_stack
2963 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2964                          struct reloc_control *rc, struct rb_root *blocks)
2965 {
2966         struct backref_node *node;
2967         struct btrfs_path *path;
2968         struct tree_block *block;
2969         struct rb_node *rb_node;
2970         int ret;
2971         int err = 0;
2972
2973         path = btrfs_alloc_path();
2974         if (!path) {
2975                 err = -ENOMEM;
2976                 goto out_free_blocks;
2977         }
2978
2979         rb_node = rb_first(blocks);
2980         while (rb_node) {
2981                 block = rb_entry(rb_node, struct tree_block, rb_node);
2982                 if (!block->key_ready)
2983                         reada_tree_block(rc, block);
2984                 rb_node = rb_next(rb_node);
2985         }
2986
2987         rb_node = rb_first(blocks);
2988         while (rb_node) {
2989                 block = rb_entry(rb_node, struct tree_block, rb_node);
2990                 if (!block->key_ready) {
2991                         err = get_tree_block_key(rc, block);
2992                         if (err)
2993                                 goto out_free_path;
2994                 }
2995                 rb_node = rb_next(rb_node);
2996         }
2997
2998         rb_node = rb_first(blocks);
2999         while (rb_node) {
3000                 block = rb_entry(rb_node, struct tree_block, rb_node);
3001
3002                 node = build_backref_tree(rc, &block->key,
3003                                           block->level, block->bytenr);
3004                 if (IS_ERR(node)) {
3005                         err = PTR_ERR(node);
3006                         goto out;
3007                 }
3008
3009                 ret = relocate_tree_block(trans, rc, node, &block->key,
3010                                           path);
3011                 if (ret < 0) {
3012                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3013                                 err = ret;
3014                         goto out;
3015                 }
3016                 rb_node = rb_next(rb_node);
3017         }
3018 out:
3019         err = finish_pending_nodes(trans, rc, path, err);
3020
3021 out_free_path:
3022         btrfs_free_path(path);
3023 out_free_blocks:
3024         free_block_list(blocks);
3025         return err;
3026 }
3027
3028 static noinline_for_stack
3029 int prealloc_file_extent_cluster(struct inode *inode,
3030                                  struct file_extent_cluster *cluster)
3031 {
3032         u64 alloc_hint = 0;
3033         u64 start;
3034         u64 end;
3035         u64 offset = BTRFS_I(inode)->index_cnt;
3036         u64 num_bytes;
3037         int nr = 0;
3038         int ret = 0;
3039
3040         BUG_ON(cluster->start != cluster->boundary[0]);
3041         mutex_lock(&inode->i_mutex);
3042
3043         ret = btrfs_check_data_free_space(inode, cluster->end +
3044                                           1 - cluster->start);
3045         if (ret)
3046                 goto out;
3047
3048         while (nr < cluster->nr) {
3049                 start = cluster->boundary[nr] - offset;
3050                 if (nr + 1 < cluster->nr)
3051                         end = cluster->boundary[nr + 1] - 1 - offset;
3052                 else
3053                         end = cluster->end - offset;
3054
3055                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3056                 num_bytes = end + 1 - start;
3057                 ret = btrfs_prealloc_file_range(inode, 0, start,
3058                                                 num_bytes, num_bytes,
3059                                                 end + 1, &alloc_hint);
3060                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3061                 if (ret)
3062                         break;
3063                 nr++;
3064         }
3065         btrfs_free_reserved_data_space(inode, cluster->end +
3066                                        1 - cluster->start);
3067 out:
3068         mutex_unlock(&inode->i_mutex);
3069         return ret;
3070 }
3071
3072 static noinline_for_stack
3073 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3074                          u64 block_start)
3075 {
3076         struct btrfs_root *root = BTRFS_I(inode)->root;
3077         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3078         struct extent_map *em;
3079         int ret = 0;
3080
3081         em = alloc_extent_map();
3082         if (!em)
3083                 return -ENOMEM;
3084
3085         em->start = start;
3086         em->len = end + 1 - start;
3087         em->block_len = em->len;
3088         em->block_start = block_start;
3089         em->bdev = root->fs_info->fs_devices->latest_bdev;
3090         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3091
3092         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3093         while (1) {
3094                 write_lock(&em_tree->lock);
3095                 ret = add_extent_mapping(em_tree, em, 0);
3096                 write_unlock(&em_tree->lock);
3097                 if (ret != -EEXIST) {
3098                         free_extent_map(em);
3099                         break;
3100                 }
3101                 btrfs_drop_extent_cache(inode, start, end, 0);
3102         }
3103         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3104         return ret;
3105 }
3106
3107 static int relocate_file_extent_cluster(struct inode *inode,
3108                                         struct file_extent_cluster *cluster)
3109 {
3110         u64 page_start;
3111         u64 page_end;
3112         u64 offset = BTRFS_I(inode)->index_cnt;
3113         unsigned long index;
3114         unsigned long last_index;
3115         struct page *page;
3116         struct file_ra_state *ra;
3117         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3118         int nr = 0;
3119         int ret = 0;
3120
3121         if (!cluster->nr)
3122                 return 0;
3123
3124         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3125         if (!ra)
3126                 return -ENOMEM;
3127
3128         ret = prealloc_file_extent_cluster(inode, cluster);
3129         if (ret)
3130                 goto out;
3131
3132         file_ra_state_init(ra, inode->i_mapping);
3133
3134         ret = setup_extent_mapping(inode, cluster->start - offset,
3135                                    cluster->end - offset, cluster->start);
3136         if (ret)
3137                 goto out;
3138
3139         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3140         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3141         while (index <= last_index) {
3142                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3143                 if (ret)
3144                         goto out;
3145
3146                 page = find_lock_page(inode->i_mapping, index);
3147                 if (!page) {
3148                         page_cache_sync_readahead(inode->i_mapping,
3149                                                   ra, NULL, index,
3150                                                   last_index + 1 - index);
3151                         page = find_or_create_page(inode->i_mapping, index,
3152                                                    mask);
3153                         if (!page) {
3154                                 btrfs_delalloc_release_metadata(inode,
3155                                                         PAGE_CACHE_SIZE);
3156                                 ret = -ENOMEM;
3157                                 goto out;
3158                         }
3159                 }
3160
3161                 if (PageReadahead(page)) {
3162                         page_cache_async_readahead(inode->i_mapping,
3163                                                    ra, NULL, page, index,
3164                                                    last_index + 1 - index);
3165                 }
3166
3167                 if (!PageUptodate(page)) {
3168                         btrfs_readpage(NULL, page);
3169                         lock_page(page);
3170                         if (!PageUptodate(page)) {
3171                                 unlock_page(page);
3172                                 page_cache_release(page);
3173                                 btrfs_delalloc_release_metadata(inode,
3174                                                         PAGE_CACHE_SIZE);
3175                                 ret = -EIO;
3176                                 goto out;
3177                         }
3178                 }
3179
3180                 page_start = page_offset(page);
3181                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3182
3183                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3184
3185                 set_page_extent_mapped(page);
3186
3187                 if (nr < cluster->nr &&
3188                     page_start + offset == cluster->boundary[nr]) {
3189                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3190                                         page_start, page_end,
3191                                         EXTENT_BOUNDARY, GFP_NOFS);
3192                         nr++;
3193                 }
3194
3195                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3196                 set_page_dirty(page);
3197
3198                 unlock_extent(&BTRFS_I(inode)->io_tree,
3199                               page_start, page_end);
3200                 unlock_page(page);
3201                 page_cache_release(page);
3202
3203                 index++;
3204                 balance_dirty_pages_ratelimited(inode->i_mapping);
3205                 btrfs_throttle(BTRFS_I(inode)->root);
3206         }
3207         WARN_ON(nr != cluster->nr);
3208 out:
3209         kfree(ra);
3210         return ret;
3211 }
3212
3213 static noinline_for_stack
3214 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3215                          struct file_extent_cluster *cluster)
3216 {
3217         int ret;
3218
3219         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3220                 ret = relocate_file_extent_cluster(inode, cluster);
3221                 if (ret)
3222                         return ret;
3223                 cluster->nr = 0;
3224         }
3225
3226         if (!cluster->nr)
3227                 cluster->start = extent_key->objectid;
3228         else
3229                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3230         cluster->end = extent_key->objectid + extent_key->offset - 1;
3231         cluster->boundary[cluster->nr] = extent_key->objectid;
3232         cluster->nr++;
3233
3234         if (cluster->nr >= MAX_EXTENTS) {
3235                 ret = relocate_file_extent_cluster(inode, cluster);
3236                 if (ret)
3237                         return ret;
3238                 cluster->nr = 0;
3239         }
3240         return 0;
3241 }
3242
3243 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3244 static int get_ref_objectid_v0(struct reloc_control *rc,
3245                                struct btrfs_path *path,
3246                                struct btrfs_key *extent_key,
3247                                u64 *ref_objectid, int *path_change)
3248 {
3249         struct btrfs_key key;
3250         struct extent_buffer *leaf;
3251         struct btrfs_extent_ref_v0 *ref0;
3252         int ret;
3253         int slot;
3254
3255         leaf = path->nodes[0];
3256         slot = path->slots[0];
3257         while (1) {
3258                 if (slot >= btrfs_header_nritems(leaf)) {
3259                         ret = btrfs_next_leaf(rc->extent_root, path);
3260                         if (ret < 0)
3261                                 return ret;
3262                         BUG_ON(ret > 0);
3263                         leaf = path->nodes[0];
3264                         slot = path->slots[0];
3265                         if (path_change)
3266                                 *path_change = 1;
3267                 }
3268                 btrfs_item_key_to_cpu(leaf, &key, slot);
3269                 if (key.objectid != extent_key->objectid)
3270                         return -ENOENT;
3271
3272                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3273                         slot++;
3274                         continue;
3275                 }
3276                 ref0 = btrfs_item_ptr(leaf, slot,
3277                                 struct btrfs_extent_ref_v0);
3278                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3279                 break;
3280         }
3281         return 0;
3282 }
3283 #endif
3284
3285 /*
3286  * helper to add a tree block to the list.
3287  * the major work is getting the generation and level of the block
3288  */
3289 static int add_tree_block(struct reloc_control *rc,
3290                           struct btrfs_key *extent_key,
3291                           struct btrfs_path *path,
3292                           struct rb_root *blocks)
3293 {
3294         struct extent_buffer *eb;
3295         struct btrfs_extent_item *ei;
3296         struct btrfs_tree_block_info *bi;
3297         struct tree_block *block;
3298         struct rb_node *rb_node;
3299         u32 item_size;
3300         int level = -1;
3301         u64 generation;
3302
3303         eb =  path->nodes[0];
3304         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3305
3306         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3307             item_size >= sizeof(*ei) + sizeof(*bi)) {
3308                 ei = btrfs_item_ptr(eb, path->slots[0],
3309                                 struct btrfs_extent_item);
3310                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3311                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3312                         level = btrfs_tree_block_level(eb, bi);
3313                 } else {
3314                         level = (int)extent_key->offset;
3315                 }
3316                 generation = btrfs_extent_generation(eb, ei);
3317         } else {
3318 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3319                 u64 ref_owner;
3320                 int ret;
3321
3322                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3323                 ret = get_ref_objectid_v0(rc, path, extent_key,
3324                                           &ref_owner, NULL);
3325                 if (ret < 0)
3326                         return ret;
3327                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3328                 level = (int)ref_owner;
3329                 /* FIXME: get real generation */
3330                 generation = 0;
3331 #else
3332                 BUG();
3333 #endif
3334         }
3335
3336         btrfs_release_path(path);
3337
3338         BUG_ON(level == -1);
3339
3340         block = kmalloc(sizeof(*block), GFP_NOFS);
3341         if (!block)
3342                 return -ENOMEM;
3343
3344         block->bytenr = extent_key->objectid;
3345         block->key.objectid = rc->extent_root->nodesize;
3346         block->key.offset = generation;
3347         block->level = level;
3348         block->key_ready = 0;
3349
3350         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3351         if (rb_node)
3352                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3353
3354         return 0;
3355 }
3356
3357 /*
3358  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3359  */
3360 static int __add_tree_block(struct reloc_control *rc,
3361                             u64 bytenr, u32 blocksize,
3362                             struct rb_root *blocks)
3363 {
3364         struct btrfs_path *path;
3365         struct btrfs_key key;
3366         int ret;
3367         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3368                                         SKINNY_METADATA);
3369
3370         if (tree_block_processed(bytenr, blocksize, rc))
3371                 return 0;
3372
3373         if (tree_search(blocks, bytenr))
3374                 return 0;
3375
3376         path = btrfs_alloc_path();
3377         if (!path)
3378                 return -ENOMEM;
3379 again:
3380         key.objectid = bytenr;
3381         if (skinny) {
3382                 key.type = BTRFS_METADATA_ITEM_KEY;
3383                 key.offset = (u64)-1;
3384         } else {
3385                 key.type = BTRFS_EXTENT_ITEM_KEY;
3386                 key.offset = blocksize;
3387         }
3388
3389         path->search_commit_root = 1;
3390         path->skip_locking = 1;
3391         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3392         if (ret < 0)
3393                 goto out;
3394
3395         if (ret > 0 && skinny) {
3396                 if (path->slots[0]) {
3397                         path->slots[0]--;
3398                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3399                                               path->slots[0]);
3400                         if (key.objectid == bytenr &&
3401                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3402                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3403                               key.offset == blocksize)))
3404                                 ret = 0;
3405                 }
3406
3407                 if (ret) {
3408                         skinny = false;
3409                         btrfs_release_path(path);
3410                         goto again;
3411                 }
3412         }
3413         BUG_ON(ret);
3414
3415         ret = add_tree_block(rc, &key, path, blocks);
3416 out:
3417         btrfs_free_path(path);
3418         return ret;
3419 }
3420
3421 /*
3422  * helper to check if the block use full backrefs for pointers in it
3423  */
3424 static int block_use_full_backref(struct reloc_control *rc,
3425                                   struct extent_buffer *eb)
3426 {
3427         u64 flags;
3428         int ret;
3429
3430         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3431             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3432                 return 1;
3433
3434         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3435                                        eb->start, btrfs_header_level(eb), 1,
3436                                        NULL, &flags);
3437         BUG_ON(ret);
3438
3439         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3440                 ret = 1;
3441         else
3442                 ret = 0;
3443         return ret;
3444 }
3445
3446 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3447                                     struct inode *inode, u64 ino)
3448 {
3449         struct btrfs_key key;
3450         struct btrfs_root *root = fs_info->tree_root;
3451         struct btrfs_trans_handle *trans;
3452         int ret = 0;
3453
3454         if (inode)
3455                 goto truncate;
3456
3457         key.objectid = ino;
3458         key.type = BTRFS_INODE_ITEM_KEY;
3459         key.offset = 0;
3460
3461         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3462         if (IS_ERR(inode) || is_bad_inode(inode)) {
3463                 if (!IS_ERR(inode))
3464                         iput(inode);
3465                 return -ENOENT;
3466         }
3467
3468 truncate:
3469         ret = btrfs_check_trunc_cache_free_space(root,
3470                                                  &fs_info->global_block_rsv);
3471         if (ret)
3472                 goto out;
3473
3474         trans = btrfs_join_transaction(root);
3475         if (IS_ERR(trans)) {
3476                 ret = PTR_ERR(trans);
3477                 goto out;
3478         }
3479
3480         ret = btrfs_truncate_free_space_cache(root, trans, inode);
3481
3482         btrfs_end_transaction(trans, root);
3483         btrfs_btree_balance_dirty(root);
3484 out:
3485         iput(inode);
3486         return ret;
3487 }
3488
3489 /*
3490  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3491  * this function scans fs tree to find blocks reference the data extent
3492  */
3493 static int find_data_references(struct reloc_control *rc,
3494                                 struct btrfs_key *extent_key,
3495                                 struct extent_buffer *leaf,
3496                                 struct btrfs_extent_data_ref *ref,
3497                                 struct rb_root *blocks)
3498 {
3499         struct btrfs_path *path;
3500         struct tree_block *block;
3501         struct btrfs_root *root;
3502         struct btrfs_file_extent_item *fi;
3503         struct rb_node *rb_node;
3504         struct btrfs_key key;
3505         u64 ref_root;
3506         u64 ref_objectid;
3507         u64 ref_offset;
3508         u32 ref_count;
3509         u32 nritems;
3510         int err = 0;
3511         int added = 0;
3512         int counted;
3513         int ret;
3514
3515         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3516         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3517         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3518         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3519
3520         /*
3521          * This is an extent belonging to the free space cache, lets just delete
3522          * it and redo the search.
3523          */
3524         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3525                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3526                                                NULL, ref_objectid);
3527                 if (ret != -ENOENT)
3528                         return ret;
3529                 ret = 0;
3530         }
3531
3532         path = btrfs_alloc_path();
3533         if (!path)
3534                 return -ENOMEM;
3535         path->reada = 1;
3536
3537         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3538         if (IS_ERR(root)) {
3539                 err = PTR_ERR(root);
3540                 goto out;
3541         }
3542
3543         key.objectid = ref_objectid;
3544         key.type = BTRFS_EXTENT_DATA_KEY;
3545         if (ref_offset > ((u64)-1 << 32))
3546                 key.offset = 0;
3547         else
3548                 key.offset = ref_offset;
3549
3550         path->search_commit_root = 1;
3551         path->skip_locking = 1;
3552         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3553         if (ret < 0) {
3554                 err = ret;
3555                 goto out;
3556         }
3557
3558         leaf = path->nodes[0];
3559         nritems = btrfs_header_nritems(leaf);
3560         /*
3561          * the references in tree blocks that use full backrefs
3562          * are not counted in
3563          */
3564         if (block_use_full_backref(rc, leaf))
3565                 counted = 0;
3566         else
3567                 counted = 1;
3568         rb_node = tree_search(blocks, leaf->start);
3569         if (rb_node) {
3570                 if (counted)
3571                         added = 1;
3572                 else
3573                         path->slots[0] = nritems;
3574         }
3575
3576         while (ref_count > 0) {
3577                 while (path->slots[0] >= nritems) {
3578                         ret = btrfs_next_leaf(root, path);
3579                         if (ret < 0) {
3580                                 err = ret;
3581                                 goto out;
3582                         }
3583                         if (WARN_ON(ret > 0))
3584                                 goto out;
3585
3586                         leaf = path->nodes[0];
3587                         nritems = btrfs_header_nritems(leaf);
3588                         added = 0;
3589
3590                         if (block_use_full_backref(rc, leaf))
3591                                 counted = 0;
3592                         else
3593                                 counted = 1;
3594                         rb_node = tree_search(blocks, leaf->start);
3595                         if (rb_node) {
3596                                 if (counted)
3597                                         added = 1;
3598                                 else
3599                                         path->slots[0] = nritems;
3600                         }
3601                 }
3602
3603                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3604                 if (WARN_ON(key.objectid != ref_objectid ||
3605                     key.type != BTRFS_EXTENT_DATA_KEY))
3606                         break;
3607
3608                 fi = btrfs_item_ptr(leaf, path->slots[0],
3609                                     struct btrfs_file_extent_item);
3610
3611                 if (btrfs_file_extent_type(leaf, fi) ==
3612                     BTRFS_FILE_EXTENT_INLINE)
3613                         goto next;
3614
3615                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3616                     extent_key->objectid)
3617                         goto next;
3618
3619                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3620                 if (key.offset != ref_offset)
3621                         goto next;
3622
3623                 if (counted)
3624                         ref_count--;
3625                 if (added)
3626                         goto next;
3627
3628                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3629                         block = kmalloc(sizeof(*block), GFP_NOFS);
3630                         if (!block) {
3631                                 err = -ENOMEM;
3632                                 break;
3633                         }
3634                         block->bytenr = leaf->start;
3635                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3636                         block->level = 0;
3637                         block->key_ready = 1;
3638                         rb_node = tree_insert(blocks, block->bytenr,
3639                                               &block->rb_node);
3640                         if (rb_node)
3641                                 backref_tree_panic(rb_node, -EEXIST,
3642                                                    block->bytenr);
3643                 }
3644                 if (counted)
3645                         added = 1;
3646                 else
3647                         path->slots[0] = nritems;
3648 next:
3649                 path->slots[0]++;
3650
3651         }
3652 out:
3653         btrfs_free_path(path);
3654         return err;
3655 }
3656
3657 /*
3658  * helper to find all tree blocks that reference a given data extent
3659  */
3660 static noinline_for_stack
3661 int add_data_references(struct reloc_control *rc,
3662                         struct btrfs_key *extent_key,
3663                         struct btrfs_path *path,
3664                         struct rb_root *blocks)
3665 {
3666         struct btrfs_key key;
3667         struct extent_buffer *eb;
3668         struct btrfs_extent_data_ref *dref;
3669         struct btrfs_extent_inline_ref *iref;
3670         unsigned long ptr;
3671         unsigned long end;
3672         u32 blocksize = rc->extent_root->nodesize;
3673         int ret = 0;
3674         int err = 0;
3675
3676         eb = path->nodes[0];
3677         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3678         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3679 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3680         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3681                 ptr = end;
3682         else
3683 #endif
3684                 ptr += sizeof(struct btrfs_extent_item);
3685
3686         while (ptr < end) {
3687                 iref = (struct btrfs_extent_inline_ref *)ptr;
3688                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3689                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3690                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3691                         ret = __add_tree_block(rc, key.offset, blocksize,
3692                                                blocks);
3693                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3694                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3695                         ret = find_data_references(rc, extent_key,
3696                                                    eb, dref, blocks);
3697                 } else {
3698                         BUG();
3699                 }
3700                 if (ret) {
3701                         err = ret;
3702                         goto out;
3703                 }
3704                 ptr += btrfs_extent_inline_ref_size(key.type);
3705         }
3706         WARN_ON(ptr > end);
3707
3708         while (1) {
3709                 cond_resched();
3710                 eb = path->nodes[0];
3711                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3712                         ret = btrfs_next_leaf(rc->extent_root, path);
3713                         if (ret < 0) {
3714                                 err = ret;
3715                                 break;
3716                         }
3717                         if (ret > 0)
3718                                 break;
3719                         eb = path->nodes[0];
3720                 }
3721
3722                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3723                 if (key.objectid != extent_key->objectid)
3724                         break;
3725
3726 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3727                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3728                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3729 #else
3730                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3731                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3732 #endif
3733                         ret = __add_tree_block(rc, key.offset, blocksize,
3734                                                blocks);
3735                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3736                         dref = btrfs_item_ptr(eb, path->slots[0],
3737                                               struct btrfs_extent_data_ref);
3738                         ret = find_data_references(rc, extent_key,
3739                                                    eb, dref, blocks);
3740                 } else {
3741                         ret = 0;
3742                 }
3743                 if (ret) {
3744                         err = ret;
3745                         break;
3746                 }
3747                 path->slots[0]++;
3748         }
3749 out:
3750         btrfs_release_path(path);
3751         if (err)
3752                 free_block_list(blocks);
3753         return err;
3754 }
3755
3756 /*
3757  * helper to find next unprocessed extent
3758  */
3759 static noinline_for_stack
3760 int find_next_extent(struct btrfs_trans_handle *trans,
3761                      struct reloc_control *rc, struct btrfs_path *path,
3762                      struct btrfs_key *extent_key)
3763 {
3764         struct btrfs_key key;
3765         struct extent_buffer *leaf;
3766         u64 start, end, last;
3767         int ret;
3768
3769         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3770         while (1) {
3771                 cond_resched();
3772                 if (rc->search_start >= last) {
3773                         ret = 1;
3774                         break;
3775                 }
3776
3777                 key.objectid = rc->search_start;
3778                 key.type = BTRFS_EXTENT_ITEM_KEY;
3779                 key.offset = 0;
3780
3781                 path->search_commit_root = 1;
3782                 path->skip_locking = 1;
3783                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3784                                         0, 0);
3785                 if (ret < 0)
3786                         break;
3787 next:
3788                 leaf = path->nodes[0];
3789                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3790                         ret = btrfs_next_leaf(rc->extent_root, path);
3791                         if (ret != 0)
3792                                 break;
3793                         leaf = path->nodes[0];
3794                 }
3795
3796                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3797                 if (key.objectid >= last) {
3798                         ret = 1;
3799                         break;
3800                 }
3801
3802                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3803                     key.type != BTRFS_METADATA_ITEM_KEY) {
3804                         path->slots[0]++;
3805                         goto next;
3806                 }
3807
3808                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3809                     key.objectid + key.offset <= rc->search_start) {
3810                         path->slots[0]++;
3811                         goto next;
3812                 }
3813
3814                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3815                     key.objectid + rc->extent_root->nodesize <=
3816                     rc->search_start) {
3817                         path->slots[0]++;
3818                         goto next;
3819                 }
3820
3821                 ret = find_first_extent_bit(&rc->processed_blocks,
3822                                             key.objectid, &start, &end,
3823                                             EXTENT_DIRTY, NULL);
3824
3825                 if (ret == 0 && start <= key.objectid) {
3826                         btrfs_release_path(path);
3827                         rc->search_start = end + 1;
3828                 } else {
3829                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3830                                 rc->search_start = key.objectid + key.offset;
3831                         else
3832                                 rc->search_start = key.objectid +
3833                                         rc->extent_root->nodesize;
3834                         memcpy(extent_key, &key, sizeof(key));
3835                         return 0;
3836                 }
3837         }
3838         btrfs_release_path(path);
3839         return ret;
3840 }
3841
3842 static void set_reloc_control(struct reloc_control *rc)
3843 {
3844         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3845
3846         mutex_lock(&fs_info->reloc_mutex);
3847         fs_info->reloc_ctl = rc;
3848         mutex_unlock(&fs_info->reloc_mutex);
3849 }
3850
3851 static void unset_reloc_control(struct reloc_control *rc)
3852 {
3853         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3854
3855         mutex_lock(&fs_info->reloc_mutex);
3856         fs_info->reloc_ctl = NULL;
3857         mutex_unlock(&fs_info->reloc_mutex);
3858 }
3859
3860 static int check_extent_flags(u64 flags)
3861 {
3862         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3863             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3864                 return 1;
3865         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3866             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3867                 return 1;
3868         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3869             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3870                 return 1;
3871         return 0;
3872 }
3873
3874 static noinline_for_stack
3875 int prepare_to_relocate(struct reloc_control *rc)
3876 {
3877         struct btrfs_trans_handle *trans;
3878
3879         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3880                                               BTRFS_BLOCK_RSV_TEMP);
3881         if (!rc->block_rsv)
3882                 return -ENOMEM;
3883
3884         memset(&rc->cluster, 0, sizeof(rc->cluster));
3885         rc->search_start = rc->block_group->key.objectid;
3886         rc->extents_found = 0;
3887         rc->nodes_relocated = 0;
3888         rc->merging_rsv_size = 0;
3889         rc->reserved_bytes = 0;
3890         rc->block_rsv->size = rc->extent_root->nodesize *
3891                               RELOCATION_RESERVED_NODES;
3892
3893         rc->create_reloc_tree = 1;
3894         set_reloc_control(rc);
3895
3896         trans = btrfs_join_transaction(rc->extent_root);
3897         if (IS_ERR(trans)) {
3898                 unset_reloc_control(rc);
3899                 /*
3900                  * extent tree is not a ref_cow tree and has no reloc_root to
3901                  * cleanup.  And callers are responsible to free the above
3902                  * block rsv.
3903                  */
3904                 return PTR_ERR(trans);
3905         }
3906         btrfs_commit_transaction(trans, rc->extent_root);
3907         return 0;
3908 }
3909
3910 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3911 {
3912         struct rb_root blocks = RB_ROOT;
3913         struct btrfs_key key;
3914         struct btrfs_trans_handle *trans = NULL;
3915         struct btrfs_path *path;
3916         struct btrfs_extent_item *ei;
3917         u64 flags;
3918         u32 item_size;
3919         int ret;
3920         int err = 0;
3921         int progress = 0;
3922
3923         path = btrfs_alloc_path();
3924         if (!path)
3925                 return -ENOMEM;
3926         path->reada = 1;
3927
3928         ret = prepare_to_relocate(rc);
3929         if (ret) {
3930                 err = ret;
3931                 goto out_free;
3932         }
3933
3934         while (1) {
3935                 rc->reserved_bytes = 0;
3936                 ret = btrfs_block_rsv_refill(rc->extent_root,
3937                                         rc->block_rsv, rc->block_rsv->size,
3938                                         BTRFS_RESERVE_FLUSH_ALL);
3939                 if (ret) {
3940                         err = ret;
3941                         break;
3942                 }
3943                 progress++;
3944                 trans = btrfs_start_transaction(rc->extent_root, 0);
3945                 if (IS_ERR(trans)) {
3946                         err = PTR_ERR(trans);
3947                         trans = NULL;
3948                         break;
3949                 }
3950 restart:
3951                 if (update_backref_cache(trans, &rc->backref_cache)) {
3952                         btrfs_end_transaction(trans, rc->extent_root);
3953                         continue;
3954                 }
3955
3956                 ret = find_next_extent(trans, rc, path, &key);
3957                 if (ret < 0)
3958                         err = ret;
3959                 if (ret != 0)
3960                         break;
3961
3962                 rc->extents_found++;
3963
3964                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3965                                     struct btrfs_extent_item);
3966                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3967                 if (item_size >= sizeof(*ei)) {
3968                         flags = btrfs_extent_flags(path->nodes[0], ei);
3969                         ret = check_extent_flags(flags);
3970                         BUG_ON(ret);
3971
3972                 } else {
3973 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3974                         u64 ref_owner;
3975                         int path_change = 0;
3976
3977                         BUG_ON(item_size !=
3978                                sizeof(struct btrfs_extent_item_v0));
3979                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3980                                                   &path_change);
3981                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3982                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3983                         else
3984                                 flags = BTRFS_EXTENT_FLAG_DATA;
3985
3986                         if (path_change) {
3987                                 btrfs_release_path(path);
3988
3989                                 path->search_commit_root = 1;
3990                                 path->skip_locking = 1;
3991                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3992                                                         &key, path, 0, 0);
3993                                 if (ret < 0) {
3994                                         err = ret;
3995                                         break;
3996                                 }
3997                                 BUG_ON(ret > 0);
3998                         }
3999 #else
4000                         BUG();
4001 #endif
4002                 }
4003
4004                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4005                         ret = add_tree_block(rc, &key, path, &blocks);
4006                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4007                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4008                         ret = add_data_references(rc, &key, path, &blocks);
4009                 } else {
4010                         btrfs_release_path(path);
4011                         ret = 0;
4012                 }
4013                 if (ret < 0) {
4014                         err = ret;
4015                         break;
4016                 }
4017
4018                 if (!RB_EMPTY_ROOT(&blocks)) {
4019                         ret = relocate_tree_blocks(trans, rc, &blocks);
4020                         if (ret < 0) {
4021                                 /*
4022                                  * if we fail to relocate tree blocks, force to update
4023                                  * backref cache when committing transaction.
4024                                  */
4025                                 rc->backref_cache.last_trans = trans->transid - 1;
4026
4027                                 if (ret != -EAGAIN) {
4028                                         err = ret;
4029                                         break;
4030                                 }
4031                                 rc->extents_found--;
4032                                 rc->search_start = key.objectid;
4033                         }
4034                 }
4035
4036                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4037                 btrfs_btree_balance_dirty(rc->extent_root);
4038                 trans = NULL;
4039
4040                 if (rc->stage == MOVE_DATA_EXTENTS &&
4041                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4042                         rc->found_file_extent = 1;
4043                         ret = relocate_data_extent(rc->data_inode,
4044                                                    &key, &rc->cluster);
4045                         if (ret < 0) {
4046                                 err = ret;
4047                                 break;
4048                         }
4049                 }
4050         }
4051         if (trans && progress && err == -ENOSPC) {
4052                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4053                                               rc->block_group->flags);
4054                 if (ret == 0) {
4055                         err = 0;
4056                         progress = 0;
4057                         goto restart;
4058                 }
4059         }
4060
4061         btrfs_release_path(path);
4062         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4063                           GFP_NOFS);
4064
4065         if (trans) {
4066                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4067                 btrfs_btree_balance_dirty(rc->extent_root);
4068         }
4069
4070         if (!err) {
4071                 ret = relocate_file_extent_cluster(rc->data_inode,
4072                                                    &rc->cluster);
4073                 if (ret < 0)
4074                         err = ret;
4075         }
4076
4077         rc->create_reloc_tree = 0;
4078         set_reloc_control(rc);
4079
4080         backref_cache_cleanup(&rc->backref_cache);
4081         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4082
4083         err = prepare_to_merge(rc, err);
4084
4085         merge_reloc_roots(rc);
4086
4087         rc->merge_reloc_tree = 0;
4088         unset_reloc_control(rc);
4089         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4090
4091         /* get rid of pinned extents */
4092         trans = btrfs_join_transaction(rc->extent_root);
4093         if (IS_ERR(trans))
4094                 err = PTR_ERR(trans);
4095         else
4096                 btrfs_commit_transaction(trans, rc->extent_root);
4097 out_free:
4098         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4099         btrfs_free_path(path);
4100         return err;
4101 }
4102
4103 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4104                                  struct btrfs_root *root, u64 objectid)
4105 {
4106         struct btrfs_path *path;
4107         struct btrfs_inode_item *item;
4108         struct extent_buffer *leaf;
4109         int ret;
4110
4111         path = btrfs_alloc_path();
4112         if (!path)
4113                 return -ENOMEM;
4114
4115         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4116         if (ret)
4117                 goto out;
4118
4119         leaf = path->nodes[0];
4120         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4121         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4122         btrfs_set_inode_generation(leaf, item, 1);
4123         btrfs_set_inode_size(leaf, item, 0);
4124         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4125         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4126                                           BTRFS_INODE_PREALLOC);
4127         btrfs_mark_buffer_dirty(leaf);
4128 out:
4129         btrfs_free_path(path);
4130         return ret;
4131 }
4132
4133 /*
4134  * helper to create inode for data relocation.
4135  * the inode is in data relocation tree and its link count is 0
4136  */
4137 static noinline_for_stack
4138 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4139                                  struct btrfs_block_group_cache *group)
4140 {
4141         struct inode *inode = NULL;
4142         struct btrfs_trans_handle *trans;
4143         struct btrfs_root *root;
4144         struct btrfs_key key;
4145         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4146         int err = 0;
4147
4148         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4149         if (IS_ERR(root))
4150                 return ERR_CAST(root);
4151
4152         trans = btrfs_start_transaction(root, 6);
4153         if (IS_ERR(trans))
4154                 return ERR_CAST(trans);
4155
4156         err = btrfs_find_free_objectid(root, &objectid);
4157         if (err)
4158                 goto out;
4159
4160         err = __insert_orphan_inode(trans, root, objectid);
4161         BUG_ON(err);
4162
4163         key.objectid = objectid;
4164         key.type = BTRFS_INODE_ITEM_KEY;
4165         key.offset = 0;
4166         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4167         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4168         BTRFS_I(inode)->index_cnt = group->key.objectid;
4169
4170         err = btrfs_orphan_add(trans, inode);
4171 out:
4172         btrfs_end_transaction(trans, root);
4173         btrfs_btree_balance_dirty(root);
4174         if (err) {
4175                 if (inode)
4176                         iput(inode);
4177                 inode = ERR_PTR(err);
4178         }
4179         return inode;
4180 }
4181
4182 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4183 {
4184         struct reloc_control *rc;
4185
4186         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4187         if (!rc)
4188                 return NULL;
4189
4190         INIT_LIST_HEAD(&rc->reloc_roots);
4191         backref_cache_init(&rc->backref_cache);
4192         mapping_tree_init(&rc->reloc_root_tree);
4193         extent_io_tree_init(&rc->processed_blocks,
4194                             fs_info->btree_inode->i_mapping);
4195         return rc;
4196 }
4197
4198 /*
4199  * function to relocate all extents in a block group.
4200  */
4201 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4202 {
4203         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4204         struct reloc_control *rc;
4205         struct inode *inode;
4206         struct btrfs_path *path;
4207         int ret;
4208         int rw = 0;
4209         int err = 0;
4210
4211         rc = alloc_reloc_control(fs_info);
4212         if (!rc)
4213                 return -ENOMEM;
4214
4215         rc->extent_root = extent_root;
4216
4217         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4218         BUG_ON(!rc->block_group);
4219
4220         if (!rc->block_group->ro) {
4221                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4222                 if (ret) {
4223                         err = ret;
4224                         goto out;
4225                 }
4226                 rw = 1;
4227         }
4228
4229         path = btrfs_alloc_path();
4230         if (!path) {
4231                 err = -ENOMEM;
4232                 goto out;
4233         }
4234
4235         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4236                                         path);
4237         btrfs_free_path(path);
4238
4239         if (!IS_ERR(inode))
4240                 ret = delete_block_group_cache(fs_info, inode, 0);
4241         else
4242                 ret = PTR_ERR(inode);
4243
4244         if (ret && ret != -ENOENT) {
4245                 err = ret;
4246                 goto out;
4247         }
4248
4249         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4250         if (IS_ERR(rc->data_inode)) {
4251                 err = PTR_ERR(rc->data_inode);
4252                 rc->data_inode = NULL;
4253                 goto out;
4254         }
4255
4256         btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
4257                rc->block_group->key.objectid, rc->block_group->flags);
4258
4259         ret = btrfs_start_delalloc_roots(fs_info, 0, -1);
4260         if (ret < 0) {
4261                 err = ret;
4262                 goto out;
4263         }
4264         btrfs_wait_ordered_roots(fs_info, -1);
4265
4266         while (1) {
4267                 mutex_lock(&fs_info->cleaner_mutex);
4268                 ret = relocate_block_group(rc);
4269                 mutex_unlock(&fs_info->cleaner_mutex);
4270                 if (ret < 0) {
4271                         err = ret;
4272                         goto out;
4273                 }
4274
4275                 if (rc->extents_found == 0)
4276                         break;
4277
4278                 btrfs_info(extent_root->fs_info, "found %llu extents",
4279                         rc->extents_found);
4280
4281                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4282                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4283                                                        (u64)-1);
4284                         if (ret) {
4285                                 err = ret;
4286                                 goto out;
4287                         }
4288                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4289                                                  0, -1);
4290                         rc->stage = UPDATE_DATA_PTRS;
4291                 }
4292         }
4293
4294         WARN_ON(rc->block_group->pinned > 0);
4295         WARN_ON(rc->block_group->reserved > 0);
4296         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4297 out:
4298         if (err && rw)
4299                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4300         iput(rc->data_inode);
4301         btrfs_put_block_group(rc->block_group);
4302         kfree(rc);
4303         return err;
4304 }
4305
4306 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4307 {
4308         struct btrfs_trans_handle *trans;
4309         int ret, err;
4310
4311         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4312         if (IS_ERR(trans))
4313                 return PTR_ERR(trans);
4314
4315         memset(&root->root_item.drop_progress, 0,
4316                 sizeof(root->root_item.drop_progress));
4317         root->root_item.drop_level = 0;
4318         btrfs_set_root_refs(&root->root_item, 0);
4319         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4320                                 &root->root_key, &root->root_item);
4321
4322         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4323         if (err)
4324                 return err;
4325         return ret;
4326 }
4327
4328 /*
4329  * recover relocation interrupted by system crash.
4330  *
4331  * this function resumes merging reloc trees with corresponding fs trees.
4332  * this is important for keeping the sharing of tree blocks
4333  */
4334 int btrfs_recover_relocation(struct btrfs_root *root)
4335 {
4336         LIST_HEAD(reloc_roots);
4337         struct btrfs_key key;
4338         struct btrfs_root *fs_root;
4339         struct btrfs_root *reloc_root;
4340         struct btrfs_path *path;
4341         struct extent_buffer *leaf;
4342         struct reloc_control *rc = NULL;
4343         struct btrfs_trans_handle *trans;
4344         int ret;
4345         int err = 0;
4346
4347         path = btrfs_alloc_path();
4348         if (!path)
4349                 return -ENOMEM;
4350         path->reada = -1;
4351
4352         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4353         key.type = BTRFS_ROOT_ITEM_KEY;
4354         key.offset = (u64)-1;
4355
4356         while (1) {
4357                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4358                                         path, 0, 0);
4359                 if (ret < 0) {
4360                         err = ret;
4361                         goto out;
4362                 }
4363                 if (ret > 0) {
4364                         if (path->slots[0] == 0)
4365                                 break;
4366                         path->slots[0]--;
4367                 }
4368                 leaf = path->nodes[0];
4369                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4370                 btrfs_release_path(path);
4371
4372                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4373                     key.type != BTRFS_ROOT_ITEM_KEY)
4374                         break;
4375
4376                 reloc_root = btrfs_read_fs_root(root, &key);
4377                 if (IS_ERR(reloc_root)) {
4378                         err = PTR_ERR(reloc_root);
4379                         goto out;
4380                 }
4381
4382                 list_add(&reloc_root->root_list, &reloc_roots);
4383
4384                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4385                         fs_root = read_fs_root(root->fs_info,
4386                                                reloc_root->root_key.offset);
4387                         if (IS_ERR(fs_root)) {
4388                                 ret = PTR_ERR(fs_root);
4389                                 if (ret != -ENOENT) {
4390                                         err = ret;
4391                                         goto out;
4392                                 }
4393                                 ret = mark_garbage_root(reloc_root);
4394                                 if (ret < 0) {
4395                                         err = ret;
4396                                         goto out;
4397                                 }
4398                         }
4399                 }
4400
4401                 if (key.offset == 0)
4402                         break;
4403
4404                 key.offset--;
4405         }
4406         btrfs_release_path(path);
4407
4408         if (list_empty(&reloc_roots))
4409                 goto out;
4410
4411         rc = alloc_reloc_control(root->fs_info);
4412         if (!rc) {
4413                 err = -ENOMEM;
4414                 goto out;
4415         }
4416
4417         rc->extent_root = root->fs_info->extent_root;
4418
4419         set_reloc_control(rc);
4420
4421         trans = btrfs_join_transaction(rc->extent_root);
4422         if (IS_ERR(trans)) {
4423                 unset_reloc_control(rc);
4424                 err = PTR_ERR(trans);
4425                 goto out_free;
4426         }
4427
4428         rc->merge_reloc_tree = 1;
4429
4430         while (!list_empty(&reloc_roots)) {
4431                 reloc_root = list_entry(reloc_roots.next,
4432                                         struct btrfs_root, root_list);
4433                 list_del(&reloc_root->root_list);
4434
4435                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4436                         list_add_tail(&reloc_root->root_list,
4437                                       &rc->reloc_roots);
4438                         continue;
4439                 }
4440
4441                 fs_root = read_fs_root(root->fs_info,
4442                                        reloc_root->root_key.offset);
4443                 if (IS_ERR(fs_root)) {
4444                         err = PTR_ERR(fs_root);
4445                         goto out_free;
4446                 }
4447
4448                 err = __add_reloc_root(reloc_root);
4449                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4450                 fs_root->reloc_root = reloc_root;
4451         }
4452
4453         err = btrfs_commit_transaction(trans, rc->extent_root);
4454         if (err)
4455                 goto out_free;
4456
4457         merge_reloc_roots(rc);
4458
4459         unset_reloc_control(rc);
4460
4461         trans = btrfs_join_transaction(rc->extent_root);
4462         if (IS_ERR(trans))
4463                 err = PTR_ERR(trans);
4464         else
4465                 err = btrfs_commit_transaction(trans, rc->extent_root);
4466 out_free:
4467         kfree(rc);
4468 out:
4469         if (!list_empty(&reloc_roots))
4470                 free_reloc_roots(&reloc_roots);
4471
4472         btrfs_free_path(path);
4473
4474         if (err == 0) {
4475                 /* cleanup orphan inode in data relocation tree */
4476                 fs_root = read_fs_root(root->fs_info,
4477                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4478                 if (IS_ERR(fs_root))
4479                         err = PTR_ERR(fs_root);
4480                 else
4481                         err = btrfs_orphan_cleanup(fs_root);
4482         }
4483         return err;
4484 }
4485
4486 /*
4487  * helper to add ordered checksum for data relocation.
4488  *
4489  * cloning checksum properly handles the nodatasum extents.
4490  * it also saves CPU time to re-calculate the checksum.
4491  */
4492 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4493 {
4494         struct btrfs_ordered_sum *sums;
4495         struct btrfs_ordered_extent *ordered;
4496         struct btrfs_root *root = BTRFS_I(inode)->root;
4497         int ret;
4498         u64 disk_bytenr;
4499         u64 new_bytenr;
4500         LIST_HEAD(list);
4501
4502         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4503         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4504
4505         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4506         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4507                                        disk_bytenr + len - 1, &list, 0);
4508         if (ret)
4509                 goto out;
4510
4511         while (!list_empty(&list)) {
4512                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4513                 list_del_init(&sums->list);
4514
4515                 /*
4516                  * We need to offset the new_bytenr based on where the csum is.
4517                  * We need to do this because we will read in entire prealloc
4518                  * extents but we may have written to say the middle of the
4519                  * prealloc extent, so we need to make sure the csum goes with
4520                  * the right disk offset.
4521                  *
4522                  * We can do this because the data reloc inode refers strictly
4523                  * to the on disk bytes, so we don't have to worry about
4524                  * disk_len vs real len like with real inodes since it's all
4525                  * disk length.
4526                  */
4527                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4528                 sums->bytenr = new_bytenr;
4529
4530                 btrfs_add_ordered_sum(inode, ordered, sums);
4531         }
4532 out:
4533         btrfs_put_ordered_extent(ordered);
4534         return ret;
4535 }
4536
4537 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4538                           struct btrfs_root *root, struct extent_buffer *buf,
4539                           struct extent_buffer *cow)
4540 {
4541         struct reloc_control *rc;
4542         struct backref_node *node;
4543         int first_cow = 0;
4544         int level;
4545         int ret = 0;
4546
4547         rc = root->fs_info->reloc_ctl;
4548         if (!rc)
4549                 return 0;
4550
4551         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4552                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4553
4554         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4555                 if (buf == root->node)
4556                         __update_reloc_root(root, cow->start);
4557         }
4558
4559         level = btrfs_header_level(buf);
4560         if (btrfs_header_generation(buf) <=
4561             btrfs_root_last_snapshot(&root->root_item))
4562                 first_cow = 1;
4563
4564         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4565             rc->create_reloc_tree) {
4566                 WARN_ON(!first_cow && level == 0);
4567
4568                 node = rc->backref_cache.path[level];
4569                 BUG_ON(node->bytenr != buf->start &&
4570                        node->new_bytenr != buf->start);
4571
4572                 drop_node_buffer(node);
4573                 extent_buffer_get(cow);
4574                 node->eb = cow;
4575                 node->new_bytenr = cow->start;
4576
4577                 if (!node->pending) {
4578                         list_move_tail(&node->list,
4579                                        &rc->backref_cache.pending[level]);
4580                         node->pending = 1;
4581                 }
4582
4583                 if (first_cow)
4584                         __mark_block_processed(rc, node);
4585
4586                 if (first_cow && level > 0)
4587                         rc->nodes_relocated += buf->len;
4588         }
4589
4590         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4591                 ret = replace_file_extents(trans, rc, root, cow);
4592         return ret;
4593 }
4594
4595 /*
4596  * called before creating snapshot. it calculates metadata reservation
4597  * requried for relocating tree blocks in the snapshot
4598  */
4599 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4600                               struct btrfs_pending_snapshot *pending,
4601                               u64 *bytes_to_reserve)
4602 {
4603         struct btrfs_root *root;
4604         struct reloc_control *rc;
4605
4606         root = pending->root;
4607         if (!root->reloc_root)
4608                 return;
4609
4610         rc = root->fs_info->reloc_ctl;
4611         if (!rc->merge_reloc_tree)
4612                 return;
4613
4614         root = root->reloc_root;
4615         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4616         /*
4617          * relocation is in the stage of merging trees. the space
4618          * used by merging a reloc tree is twice the size of
4619          * relocated tree nodes in the worst case. half for cowing
4620          * the reloc tree, half for cowing the fs tree. the space
4621          * used by cowing the reloc tree will be freed after the
4622          * tree is dropped. if we create snapshot, cowing the fs
4623          * tree may use more space than it frees. so we need
4624          * reserve extra space.
4625          */
4626         *bytes_to_reserve += rc->nodes_relocated;
4627 }
4628
4629 /*
4630  * called after snapshot is created. migrate block reservation
4631  * and create reloc root for the newly created snapshot
4632  */
4633 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4634                                struct btrfs_pending_snapshot *pending)
4635 {
4636         struct btrfs_root *root = pending->root;
4637         struct btrfs_root *reloc_root;
4638         struct btrfs_root *new_root;
4639         struct reloc_control *rc;
4640         int ret;
4641
4642         if (!root->reloc_root)
4643                 return 0;
4644
4645         rc = root->fs_info->reloc_ctl;
4646         rc->merging_rsv_size += rc->nodes_relocated;
4647
4648         if (rc->merge_reloc_tree) {
4649                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4650                                               rc->block_rsv,
4651                                               rc->nodes_relocated);
4652                 if (ret)
4653                         return ret;
4654         }
4655
4656         new_root = pending->snap;
4657         reloc_root = create_reloc_root(trans, root->reloc_root,
4658                                        new_root->root_key.objectid);
4659         if (IS_ERR(reloc_root))
4660                 return PTR_ERR(reloc_root);
4661
4662         ret = __add_reloc_root(reloc_root);
4663         BUG_ON(ret < 0);
4664         new_root->reloc_root = reloc_root;
4665
4666         if (rc->create_reloc_tree)
4667                 ret = clone_backref_node(trans, rc, root, reloc_root);
4668         return ret;
4669 }