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 #include "qgroup.h"
  35 #include "print-tree.h"
  36
  37 /*
  38  * backref_node, mapping_node and tree_block start with this
  39  */
  40 struct tree_entry {
  41         struct rb_node rb_node;
  42         u64 bytenr;
  43 };
  44
  45 /*
  46  * present a tree block in the backref cache
  47  */
  48 struct backref_node {
  49         struct rb_node rb_node;
  50         u64 bytenr;
  51
  52         u64 new_bytenr;
  53         /* objectid of tree block owner, can be not uptodate */
  54         u64 owner;
  55         /* link to pending, changed or detached list */
  56         struct list_head list;
  57         /* list of upper level blocks reference this block */
  58         struct list_head upper;
  59         /* list of child blocks in the cache */
  60         struct list_head lower;
  61         /* NULL if this node is not tree root */
  62         struct btrfs_root *root;
  63         /* extent buffer got by COW the block */
  64         struct extent_buffer *eb;
  65         /* level of tree block */
  66         unsigned int level:8;
  67         /* is the block in non-reference counted tree */
  68         unsigned int cowonly:1;
  69         /* 1 if no child node in the cache */
  70         unsigned int lowest:1;
  71         /* is the extent buffer locked */
  72         unsigned int locked:1;
  73         /* has the block been processed */
  74         unsigned int processed:1;
  75         /* have backrefs of this block been checked */
  76         unsigned int checked:1;
  77         /*
  78          * 1 if corresponding block has been cowed but some upper
  79          * level block pointers may not point to the new location
  80          */
  81         unsigned int pending:1;
  82         /*
  83          * 1 if the backref node isn't connected to any other
  84          * backref node.
  85          */
  86         unsigned int detached:1;
  87 };
  88
  89 /*
  90  * present a block pointer in the backref cache
  91  */
  92 struct backref_edge {
  93         struct list_head list[2];
  94         struct backref_node *node[2];
  95 };
  96
  97 #define LOWER   0
  98 #define UPPER   1
  99 #define RELOCATION_RESERVED_NODES       256
 100
 101 struct backref_cache {
 102         /* red black tree of all backref nodes in the cache */
 103         struct rb_root rb_root;
 104         /* for passing backref nodes to btrfs_reloc_cow_block */
 105         struct backref_node *path[BTRFS_MAX_LEVEL];
 106         /*
 107          * list of blocks that have been cowed but some block
 108          * pointers in upper level blocks may not reflect the
 109          * new location
 110          */
 111         struct list_head pending[BTRFS_MAX_LEVEL];
 112         /* list of backref nodes with no child node */
 113         struct list_head leaves;
 114         /* list of blocks that have been cowed in current transaction */
 115         struct list_head changed;
 116         /* list of detached backref node. */
 117         struct list_head detached;
 118
 119         u64 last_trans;
 120
 121         int nr_nodes;
 122         int nr_edges;
 123 };
 124
 125 /*
 126  * map address of tree root to tree
 127  */
 128 struct mapping_node {
 129         struct rb_node rb_node;
 130         u64 bytenr;
 131         void *data;
 132 };
 133
 134 struct mapping_tree {
 135         struct rb_root rb_root;
 136         spinlock_t lock;
 137 };
 138
 139 /*
 140  * present a tree block to process
 141  */
 142 struct tree_block {
 143         struct rb_node rb_node;
 144         u64 bytenr;
 145         struct btrfs_key key;
 146         unsigned int level:8;
 147         unsigned int key_ready:1;
 148 };
 149
 150 #define MAX_EXTENTS 128
 151
 152 struct file_extent_cluster {
 153         u64 start;
 154         u64 end;
 155         u64 boundary[MAX_EXTENTS];
 156         unsigned int nr;
 157 };
 158
 159 struct reloc_control {
 160         /* block group to relocate */
 161         struct btrfs_block_group_cache *block_group;
 162         /* extent tree */
 163         struct btrfs_root *extent_root;
 164         /* inode for moving data */
 165         struct inode *data_inode;
 166
 167         struct btrfs_block_rsv *block_rsv;
 168
 169         struct backref_cache backref_cache;
 170
 171         struct file_extent_cluster cluster;
 172         /* tree blocks have been processed */
 173         struct extent_io_tree processed_blocks;
 174         /* map start of tree root to corresponding reloc tree */
 175         struct mapping_tree reloc_root_tree;
 176         /* list of reloc trees */
 177         struct list_head reloc_roots;
 178         /* size of metadata reservation for merging reloc trees */
 179         u64 merging_rsv_size;
 180         /* size of relocated tree nodes */
 181         u64 nodes_relocated;
 182         /* reserved size for block group relocation*/
 183         u64 reserved_bytes;
 184
 185         u64 search_start;
 186         u64 extents_found;
 187
 188         unsigned int stage:8;
 189         unsigned int create_reloc_tree:1;
 190         unsigned int merge_reloc_tree:1;
 191         unsigned int found_file_extent:1;
 192 };
 193
 194 /* stages of data relocation */
 195 #define MOVE_DATA_EXTENTS       0
 196 #define UPDATE_DATA_PTRS        1
 197
 198 static void remove_backref_node(struct backref_cache *cache,
 199                                 struct backref_node *node);
 200 static void __mark_block_processed(struct reloc_control *rc,
 201                                    struct backref_node *node);
 202
 203 static void mapping_tree_init(struct mapping_tree *tree)
 204 {
 205         tree->rb_root = RB_ROOT;
 206         spin_lock_init(&tree->lock);
 207 }
 208
 209 static void backref_cache_init(struct backref_cache *cache)
 210 {
 211         int i;
 212         cache->rb_root = RB_ROOT;
 213         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 214                 INIT_LIST_HEAD(&cache->pending[i]);
 215         INIT_LIST_HEAD(&cache->changed);
 216         INIT_LIST_HEAD(&cache->detached);
 217         INIT_LIST_HEAD(&cache->leaves);
 218 }
 219
 220 static void backref_cache_cleanup(struct backref_cache *cache)
 221 {
 222         struct backref_node *node;
 223         int i;
 224
 225         while (!list_empty(&cache->detached)) {
 226                 node = list_entry(cache->detached.next,
 227                                   struct backref_node, list);
 228                 remove_backref_node(cache, node);
 229         }
 230
 231         while (!list_empty(&cache->leaves)) {
 232                 node = list_entry(cache->leaves.next,
 233                                   struct backref_node, lower);
 234                 remove_backref_node(cache, node);
 235         }
 236
 237         cache->last_trans = 0;
 238
 239         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 240                 ASSERT(list_empty(&cache->pending[i]));
 241         ASSERT(list_empty(&cache->changed));
 242         ASSERT(list_empty(&cache->detached));
 243         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
 244         ASSERT(!cache->nr_nodes);
 245         ASSERT(!cache->nr_edges);
 246 }
 247
 248 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 249 {
 250         struct backref_node *node;
 251
 252         node = kzalloc(sizeof(*node), GFP_NOFS);
 253         if (node) {
 254                 INIT_LIST_HEAD(&node->list);
 255                 INIT_LIST_HEAD(&node->upper);
 256                 INIT_LIST_HEAD(&node->lower);
 257                 RB_CLEAR_NODE(&node->rb_node);
 258                 cache->nr_nodes++;
 259         }
 260         return node;
 261 }
 262
 263 static void free_backref_node(struct backref_cache *cache,
 264                               struct backref_node *node)
 265 {
 266         if (node) {
 267                 cache->nr_nodes--;
 268                 kfree(node);
 269         }
 270 }
 271
 272 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 273 {
 274         struct backref_edge *edge;
 275
 276         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 277         if (edge)
 278                 cache->nr_edges++;
 279         return edge;
 280 }
 281
 282 static void free_backref_edge(struct backref_cache *cache,
 283                               struct backref_edge *edge)
 284 {
 285         if (edge) {
 286                 cache->nr_edges--;
 287                 kfree(edge);
 288         }
 289 }
 290
 291 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 292                                    struct rb_node *node)
 293 {
 294         struct rb_node **p = &root->rb_node;
 295         struct rb_node *parent = NULL;
 296         struct tree_entry *entry;
 297
 298         while (*p) {
 299                 parent = *p;
 300                 entry = rb_entry(parent, struct tree_entry, rb_node);
 301
 302                 if (bytenr < entry->bytenr)
 303                         p = &(*p)->rb_left;
 304                 else if (bytenr > entry->bytenr)
 305                         p = &(*p)->rb_right;
 306                 else
 307                         return parent;
 308         }
 309
 310         rb_link_node(node, parent, p);
 311         rb_insert_color(node, root);
 312         return NULL;
 313 }
 314
 315 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 316 {
 317         struct rb_node *n = root->rb_node;
 318         struct tree_entry *entry;
 319
 320         while (n) {
 321                 entry = rb_entry(n, struct tree_entry, rb_node);
 322
 323                 if (bytenr < entry->bytenr)
 324                         n = n->rb_left;
 325                 else if (bytenr > entry->bytenr)
 326                         n = n->rb_right;
 327                 else
 328                         return n;
 329         }
 330         return NULL;
 331 }
 332
 333 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 334 {
 335
 336         struct btrfs_fs_info *fs_info = NULL;
 337         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 338                                               rb_node);
 339         if (bnode->root)
 340                 fs_info = bnode->root->fs_info;
 341         btrfs_panic(fs_info, errno,
 342                     "Inconsistency in backref cache found at offset %llu",
 343                     bytenr);
 344 }
 345
 346 /*
 347  * walk up backref nodes until reach node presents tree root
 348  */
 349 static struct backref_node *walk_up_backref(struct backref_node *node,
 350                                             struct backref_edge *edges[],
 351                                             int *index)
 352 {
 353         struct backref_edge *edge;
 354         int idx = *index;
 355
 356         while (!list_empty(&node->upper)) {
 357                 edge = list_entry(node->upper.next,
 358                                   struct backref_edge, list[LOWER]);
 359                 edges[idx++] = edge;
 360                 node = edge->node[UPPER];
 361         }
 362         BUG_ON(node->detached);
 363         *index = idx;
 364         return node;
 365 }
 366
 367 /*
 368  * walk down backref nodes to find start of next reference path
 369  */
 370 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 371                                               int *index)
 372 {
 373         struct backref_edge *edge;
 374         struct backref_node *lower;
 375         int idx = *index;
 376
 377         while (idx > 0) {
 378                 edge = edges[idx - 1];
 379                 lower = edge->node[LOWER];
 380                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 381                         idx--;
 382                         continue;
 383                 }
 384                 edge = list_entry(edge->list[LOWER].next,
 385                                   struct backref_edge, list[LOWER]);
 386                 edges[idx - 1] = edge;
 387                 *index = idx;
 388                 return edge->node[UPPER];
 389         }
 390         *index = 0;
 391         return NULL;
 392 }
 393
 394 static void unlock_node_buffer(struct backref_node *node)
 395 {
 396         if (node->locked) {
 397                 btrfs_tree_unlock(node->eb);
 398                 node->locked = 0;
 399         }
 400 }
 401
 402 static void drop_node_buffer(struct backref_node *node)
 403 {
 404         if (node->eb) {
 405                 unlock_node_buffer(node);
 406                 free_extent_buffer(node->eb);
 407                 node->eb = NULL;
 408         }
 409 }
 410
 411 static void drop_backref_node(struct backref_cache *tree,
 412                               struct backref_node *node)
 413 {
 414         BUG_ON(!list_empty(&node->upper));
 415
 416         drop_node_buffer(node);
 417         list_del(&node->list);
 418         list_del(&node->lower);
 419         if (!RB_EMPTY_NODE(&node->rb_node))
 420                 rb_erase(&node->rb_node, &tree->rb_root);
 421         free_backref_node(tree, node);
 422 }
 423
 424 /*
 425  * remove a backref node from the backref cache
 426  */
 427 static void remove_backref_node(struct backref_cache *cache,
 428                                 struct backref_node *node)
 429 {
 430         struct backref_node *upper;
 431         struct backref_edge *edge;
 432
 433         if (!node)
 434                 return;
 435
 436         BUG_ON(!node->lowest && !node->detached);
 437         while (!list_empty(&node->upper)) {
 438                 edge = list_entry(node->upper.next, struct backref_edge,
 439                                   list[LOWER]);
 440                 upper = edge->node[UPPER];
 441                 list_del(&edge->list[LOWER]);
 442                 list_del(&edge->list[UPPER]);
 443                 free_backref_edge(cache, edge);
 444
 445                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 446                         BUG_ON(!list_empty(&node->upper));
 447                         drop_backref_node(cache, node);
 448                         node = upper;
 449                         node->lowest = 1;
 450                         continue;
 451                 }
 452                 /*
 453                  * add the node to leaf node list if no other
 454                  * child block cached.
 455                  */
 456                 if (list_empty(&upper->lower)) {
 457                         list_add_tail(&upper->lower, &cache->leaves);
 458                         upper->lowest = 1;
 459                 }
 460         }
 461
 462         drop_backref_node(cache, node);
 463 }
 464
 465 static void update_backref_node(struct backref_cache *cache,
 466                                 struct backref_node *node, u64 bytenr)
 467 {
 468         struct rb_node *rb_node;
 469         rb_erase(&node->rb_node, &cache->rb_root);
 470         node->bytenr = bytenr;
 471         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 472         if (rb_node)
 473                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 474 }
 475
 476 /*
 477  * update backref cache after a transaction commit
 478  */
 479 static int update_backref_cache(struct btrfs_trans_handle *trans,
 480                                 struct backref_cache *cache)
 481 {
 482         struct backref_node *node;
 483         int level = 0;
 484
 485         if (cache->last_trans == 0) {
 486                 cache->last_trans = trans->transid;
 487                 return 0;
 488         }
 489
 490         if (cache->last_trans == trans->transid)
 491                 return 0;
 492
 493         /*
 494          * detached nodes are used to avoid unnecessary backref
 495          * lookup. transaction commit changes the extent tree.
 496          * so the detached nodes are no longer useful.
 497          */
 498         while (!list_empty(&cache->detached)) {
 499                 node = list_entry(cache->detached.next,
 500                                   struct backref_node, list);
 501                 remove_backref_node(cache, node);
 502         }
 503
 504         while (!list_empty(&cache->changed)) {
 505                 node = list_entry(cache->changed.next,
 506                                   struct backref_node, list);
 507                 list_del_init(&node->list);
 508                 BUG_ON(node->pending);
 509                 update_backref_node(cache, node, node->new_bytenr);
 510         }
 511
 512         /*
 513          * some nodes can be left in the pending list if there were
 514          * errors during processing the pending nodes.
 515          */
 516         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 517                 list_for_each_entry(node, &cache->pending[level], list) {
 518                         BUG_ON(!node->pending);
 519                         if (node->bytenr == node->new_bytenr)
 520                                 continue;
 521                         update_backref_node(cache, node, node->new_bytenr);
 522                 }
 523         }
 524
 525         cache->last_trans = 0;
 526         return 1;
 527 }
 528
 529
 530 static int should_ignore_root(struct btrfs_root *root)
 531 {
 532         struct btrfs_root *reloc_root;
 533
 534         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 535                 return 0;
 536
 537         reloc_root = root->reloc_root;
 538         if (!reloc_root)
 539                 return 0;
 540
 541         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 542             root->fs_info->running_transaction->transid - 1)
 543                 return 0;
 544         /*
 545          * if there is reloc tree and it was created in previous
 546          * transaction backref lookup can find the reloc tree,
 547          * so backref node for the fs tree root is useless for
 548          * relocation.
 549          */
 550         return 1;
 551 }
 552 /*
 553  * find reloc tree by address of tree root
 554  */
 555 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 556                                           u64 bytenr)
 557 {
 558         struct rb_node *rb_node;
 559         struct mapping_node *node;
 560         struct btrfs_root *root = NULL;
 561
 562         spin_lock(&rc->reloc_root_tree.lock);
 563         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 564         if (rb_node) {
 565                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 566                 root = (struct btrfs_root *)node->data;
 567         }
 568         spin_unlock(&rc->reloc_root_tree.lock);
 569         return root;
 570 }
 571
 572 static int is_cowonly_root(u64 root_objectid)
 573 {
 574         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 575             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 576             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 577             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 578             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 579             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 580             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 581             root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
 582             root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
 583                 return 1;
 584         return 0;
 585 }
 586
 587 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 588                                         u64 root_objectid)
 589 {
 590         struct btrfs_key key;
 591
 592         key.objectid = root_objectid;
 593         key.type = BTRFS_ROOT_ITEM_KEY;
 594         if (is_cowonly_root(root_objectid))
 595                 key.offset = 0;
 596         else
 597                 key.offset = (u64)-1;
 598
 599         return btrfs_get_fs_root(fs_info, &key, false);
 600 }
 601
 602 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 603 static noinline_for_stack
 604 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 605                                   struct extent_buffer *leaf,
 606                                   struct btrfs_extent_ref_v0 *ref0)
 607 {
 608         struct btrfs_root *root;
 609         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 610         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 611
 612         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 613
 614         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 615         BUG_ON(IS_ERR(root));
 616
 617         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 618             generation != btrfs_root_generation(&root->root_item))
 619                 return NULL;
 620
 621         return root;
 622 }
 623 #endif
 624
 625 static noinline_for_stack
 626 int find_inline_backref(struct extent_buffer *leaf, int slot,
 627                         unsigned long *ptr, unsigned long *end)
 628 {
 629         struct btrfs_key key;
 630         struct btrfs_extent_item *ei;
 631         struct btrfs_tree_block_info *bi;
 632         u32 item_size;
 633
 634         btrfs_item_key_to_cpu(leaf, &key, slot);
 635
 636         item_size = btrfs_item_size_nr(leaf, slot);
 637 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 638         if (item_size < sizeof(*ei)) {
 639                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 640                 return 1;
 641         }
 642 #endif
 643         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 644         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 645                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 646
 647         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 648             item_size <= sizeof(*ei) + sizeof(*bi)) {
 649                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 650                 return 1;
 651         }
 652         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 653             item_size <= sizeof(*ei)) {
 654                 WARN_ON(item_size < sizeof(*ei));
 655                 return 1;
 656         }
 657
 658         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 659                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 660                 *ptr = (unsigned long)(bi + 1);
 661         } else {
 662                 *ptr = (unsigned long)(ei + 1);
 663         }
 664         *end = (unsigned long)ei + item_size;
 665         return 0;
 666 }
 667
 668 /*
 669  * build backref tree for a given tree block. root of the backref tree
 670  * corresponds the tree block, leaves of the backref tree correspond
 671  * roots of b-trees that reference the tree block.
 672  *
 673  * the basic idea of this function is check backrefs of a given block
 674  * to find upper level blocks that reference the block, and then check
 675  * backrefs of these upper level blocks recursively. the recursion stop
 676  * when tree root is reached or backrefs for the block is cached.
 677  *
 678  * NOTE: if we find backrefs for a block are cached, we know backrefs
 679  * for all upper level blocks that directly/indirectly reference the
 680  * block are also cached.
 681  */
 682 static noinline_for_stack
 683 struct backref_node *build_backref_tree(struct reloc_control *rc,
 684                                         struct btrfs_key *node_key,
 685                                         int level, u64 bytenr)
 686 {
 687         struct backref_cache *cache = &rc->backref_cache;
 688         struct btrfs_path *path1;
 689         struct btrfs_path *path2;
 690         struct extent_buffer *eb;
 691         struct btrfs_root *root;
 692         struct backref_node *cur;
 693         struct backref_node *upper;
 694         struct backref_node *lower;
 695         struct backref_node *node = NULL;
 696         struct backref_node *exist = NULL;
 697         struct backref_edge *edge;
 698         struct rb_node *rb_node;
 699         struct btrfs_key key;
 700         unsigned long end;
 701         unsigned long ptr;
 702         LIST_HEAD(list);
 703         LIST_HEAD(useless);
 704         int cowonly;
 705         int ret;
 706         int err = 0;
 707         bool need_check = true;
 708
 709         path1 = btrfs_alloc_path();
 710         path2 = btrfs_alloc_path();
 711         if (!path1 || !path2) {
 712                 err = -ENOMEM;
 713                 goto out;
 714         }
 715         path1->reada = READA_FORWARD;
 716         path2->reada = READA_FORWARD;
 717
 718         node = alloc_backref_node(cache);
 719         if (!node) {
 720                 err = -ENOMEM;
 721                 goto out;
 722         }
 723
 724         node->bytenr = bytenr;
 725         node->level = level;
 726         node->lowest = 1;
 727         cur = node;
 728 again:
 729         end = 0;
 730         ptr = 0;
 731         key.objectid = cur->bytenr;
 732         key.type = BTRFS_METADATA_ITEM_KEY;
 733         key.offset = (u64)-1;
 734
 735         path1->search_commit_root = 1;
 736         path1->skip_locking = 1;
 737         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 738                                 0, 0);
 739         if (ret < 0) {
 740                 err = ret;
 741                 goto out;
 742         }
 743         ASSERT(ret);
 744         ASSERT(path1->slots[0]);
 745
 746         path1->slots[0]--;
 747
 748         WARN_ON(cur->checked);
 749         if (!list_empty(&cur->upper)) {
 750                 /*
 751                  * the backref was added previously when processing
 752                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 753                  */
 754                 ASSERT(list_is_singular(&cur->upper));
 755                 edge = list_entry(cur->upper.next, struct backref_edge,
 756                                   list[LOWER]);
 757                 ASSERT(list_empty(&edge->list[UPPER]));
 758                 exist = edge->node[UPPER];
 759                 /*
 760                  * add the upper level block to pending list if we need
 761                  * check its backrefs
 762                  */
 763                 if (!exist->checked)
 764                         list_add_tail(&edge->list[UPPER], &list);
 765         } else {
 766                 exist = NULL;
 767         }
 768
 769         while (1) {
 770                 cond_resched();
 771                 eb = path1->nodes[0];
 772
 773                 if (ptr >= end) {
 774                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 775                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 776                                 if (ret < 0) {
 777                                         err = ret;
 778                                         goto out;
 779                                 }
 780                                 if (ret > 0)
 781                                         break;
 782                                 eb = path1->nodes[0];
 783                         }
 784
 785                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 786                         if (key.objectid != cur->bytenr) {
 787                                 WARN_ON(exist);
 788                                 break;
 789                         }
 790
 791                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 792                             key.type == BTRFS_METADATA_ITEM_KEY) {
 793                                 ret = find_inline_backref(eb, path1->slots[0],
 794                                                           &ptr, &end);
 795                                 if (ret)
 796                                         goto next;
 797                         }
 798                 }
 799
 800                 if (ptr < end) {
 801                         /* update key for inline back ref */
 802                         struct btrfs_extent_inline_ref *iref;
 803                         int type;
 804                         iref = (struct btrfs_extent_inline_ref *)ptr;
 805                         type = btrfs_get_extent_inline_ref_type(eb, iref,
 806                                                         BTRFS_REF_TYPE_BLOCK);
 807                         if (type == BTRFS_REF_TYPE_INVALID) {
 808                                 err = -EINVAL;
 809                                 goto out;
 810                         }
 811                         key.type = type;
 812                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 813
 814                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 815                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 816                 }
 817
 818                 if (exist &&
 819                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 820                       exist->owner == key.offset) ||
 821                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 822                       exist->bytenr == key.offset))) {
 823                         exist = NULL;
 824                         goto next;
 825                 }
 826
 827 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 828                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 829                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 830                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 831                                 struct btrfs_extent_ref_v0 *ref0;
 832                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 833                                                 struct btrfs_extent_ref_v0);
 834                                 if (key.objectid == key.offset) {
 835                                         root = find_tree_root(rc, eb, ref0);
 836                                         if (root && !should_ignore_root(root))
 837                                                 cur->root = root;
 838                                         else
 839                                                 list_add(&cur->list, &useless);
 840                                         break;
 841                                 }
 842                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 843                                                                       ref0)))
 844                                         cur->cowonly = 1;
 845                         }
 846 #else
 847                 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
 848                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 849 #endif
 850                         if (key.objectid == key.offset) {
 851                                 /*
 852                                  * only root blocks of reloc trees use
 853                                  * backref of this type.
 854                                  */
 855                                 root = find_reloc_root(rc, cur->bytenr);
 856                                 ASSERT(root);
 857                                 cur->root = root;
 858                                 break;
 859                         }
 860
 861                         edge = alloc_backref_edge(cache);
 862                         if (!edge) {
 863                                 err = -ENOMEM;
 864                                 goto out;
 865                         }
 866                         rb_node = tree_search(&cache->rb_root, key.offset);
 867                         if (!rb_node) {
 868                                 upper = alloc_backref_node(cache);
 869                                 if (!upper) {
 870                                         free_backref_edge(cache, edge);
 871                                         err = -ENOMEM;
 872                                         goto out;
 873                                 }
 874                                 upper->bytenr = key.offset;
 875                                 upper->level = cur->level + 1;
 876                                 /*
 877                                  *  backrefs for the upper level block isn't
 878                                  *  cached, add the block to pending list
 879                                  */
 880                                 list_add_tail(&edge->list[UPPER], &list);
 881                         } else {
 882                                 upper = rb_entry(rb_node, struct backref_node,
 883                                                  rb_node);
 884                                 ASSERT(upper->checked);
 885                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 886                         }
 887                         list_add_tail(&edge->list[LOWER], &cur->upper);
 888                         edge->node[LOWER] = cur;
 889                         edge->node[UPPER] = upper;
 890
 891                         goto next;
 892                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 893                         goto next;
 894                 }
 895
 896                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 897                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 898                 if (IS_ERR(root)) {
 899                         err = PTR_ERR(root);
 900                         goto out;
 901                 }
 902
 903                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 904                         cur->cowonly = 1;
 905
 906                 if (btrfs_root_level(&root->root_item) == cur->level) {
 907                         /* tree root */
 908                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
 909                                cur->bytenr);
 910                         if (should_ignore_root(root))
 911                                 list_add(&cur->list, &useless);
 912                         else
 913                                 cur->root = root;
 914                         break;
 915                 }
 916
 917                 level = cur->level + 1;
 918
 919                 /*
 920                  * searching the tree to find upper level blocks
 921                  * reference the block.
 922                  */
 923                 path2->search_commit_root = 1;
 924                 path2->skip_locking = 1;
 925                 path2->lowest_level = level;
 926                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 927                 path2->lowest_level = 0;
 928                 if (ret < 0) {
 929                         err = ret;
 930                         goto out;
 931                 }
 932                 if (ret > 0 && path2->slots[level] > 0)
 933                         path2->slots[level]--;
 934
 935                 eb = path2->nodes[level];
 936                 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
 937                     cur->bytenr) {
 938                         btrfs_err(root->fs_info,
 939         "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
 940                                   cur->bytenr, level - 1, root->objectid,
 941                                   node_key->objectid, node_key->type,
 942                                   node_key->offset);
 943                         err = -ENOENT;
 944                         goto out;
 945                 }
 946                 lower = cur;
 947                 need_check = true;
 948                 for (; level < BTRFS_MAX_LEVEL; level++) {
 949                         if (!path2->nodes[level]) {
 950                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
 951                                        lower->bytenr);
 952                                 if (should_ignore_root(root))
 953                                         list_add(&lower->list, &useless);
 954                                 else
 955                                         lower->root = root;
 956                                 break;
 957                         }
 958
 959                         edge = alloc_backref_edge(cache);
 960                         if (!edge) {
 961                                 err = -ENOMEM;
 962                                 goto out;
 963                         }
 964
 965                         eb = path2->nodes[level];
 966                         rb_node = tree_search(&cache->rb_root, eb->start);
 967                         if (!rb_node) {
 968                                 upper = alloc_backref_node(cache);
 969                                 if (!upper) {
 970                                         free_backref_edge(cache, edge);
 971                                         err = -ENOMEM;
 972                                         goto out;
 973                                 }
 974                                 upper->bytenr = eb->start;
 975                                 upper->owner = btrfs_header_owner(eb);
 976                                 upper->level = lower->level + 1;
 977                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
 978                                               &root->state))
 979                                         upper->cowonly = 1;
 980
 981                                 /*
 982                                  * if we know the block isn't shared
 983                                  * we can void checking its backrefs.
 984                                  */
 985                                 if (btrfs_block_can_be_shared(root, eb))
 986                                         upper->checked = 0;
 987                                 else
 988                                         upper->checked = 1;
 989
 990                                 /*
 991                                  * add the block to pending list if we
 992                                  * need check its backrefs, we only do this once
 993                                  * while walking up a tree as we will catch
 994                                  * anything else later on.
 995                                  */
 996                                 if (!upper->checked && need_check) {
 997                                         need_check = false;
 998                                         list_add_tail(&edge->list[UPPER],
 999                                                       &list);
1000                                 } else {
1001                                         if (upper->checked)
1002                                                 need_check = true;
1003                                         INIT_LIST_HEAD(&edge->list[UPPER]);
1004                                 }
1005                         } else {
1006                                 upper = rb_entry(rb_node, struct backref_node,
1007                                                  rb_node);
1008                                 ASSERT(upper->checked);
1009                                 INIT_LIST_HEAD(&edge->list[UPPER]);
1010                                 if (!upper->owner)
1011                                         upper->owner = btrfs_header_owner(eb);
1012                         }
1013                         list_add_tail(&edge->list[LOWER], &lower->upper);
1014                         edge->node[LOWER] = lower;
1015                         edge->node[UPPER] = upper;
1016
1017                         if (rb_node)
1018                                 break;
1019                         lower = upper;
1020                         upper = NULL;
1021                 }
1022                 btrfs_release_path(path2);
1023 next:
1024                 if (ptr < end) {
1025                         ptr += btrfs_extent_inline_ref_size(key.type);
1026                         if (ptr >= end) {
1027                                 WARN_ON(ptr > end);
1028                                 ptr = 0;
1029                                 end = 0;
1030                         }
1031                 }
1032                 if (ptr >= end)
1033                         path1->slots[0]++;
1034         }
1035         btrfs_release_path(path1);
1036
1037         cur->checked = 1;
1038         WARN_ON(exist);
1039
1040         /* the pending list isn't empty, take the first block to process */
1041         if (!list_empty(&list)) {
1042                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1043                 list_del_init(&edge->list[UPPER]);
1044                 cur = edge->node[UPPER];
1045                 goto again;
1046         }
1047
1048         /*
1049          * everything goes well, connect backref nodes and insert backref nodes
1050          * into the cache.
1051          */
1052         ASSERT(node->checked);
1053         cowonly = node->cowonly;
1054         if (!cowonly) {
1055                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1056                                       &node->rb_node);
1057                 if (rb_node)
1058                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1059                 list_add_tail(&node->lower, &cache->leaves);
1060         }
1061
1062         list_for_each_entry(edge, &node->upper, list[LOWER])
1063                 list_add_tail(&edge->list[UPPER], &list);
1064
1065         while (!list_empty(&list)) {
1066                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1067                 list_del_init(&edge->list[UPPER]);
1068                 upper = edge->node[UPPER];
1069                 if (upper->detached) {
1070                         list_del(&edge->list[LOWER]);
1071                         lower = edge->node[LOWER];
1072                         free_backref_edge(cache, edge);
1073                         if (list_empty(&lower->upper))
1074                                 list_add(&lower->list, &useless);
1075                         continue;
1076                 }
1077
1078                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1079                         if (upper->lowest) {
1080                                 list_del_init(&upper->lower);
1081                                 upper->lowest = 0;
1082                         }
1083
1084                         list_add_tail(&edge->list[UPPER], &upper->lower);
1085                         continue;
1086                 }
1087
1088                 if (!upper->checked) {
1089                         /*
1090                          * Still want to blow up for developers since this is a
1091                          * logic bug.
1092                          */
1093                         ASSERT(0);
1094                         err = -EINVAL;
1095                         goto out;
1096                 }
1097                 if (cowonly != upper->cowonly) {
1098                         ASSERT(0);
1099                         err = -EINVAL;
1100                         goto out;
1101                 }
1102
1103                 if (!cowonly) {
1104                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1105                                               &upper->rb_node);
1106                         if (rb_node)
1107                                 backref_tree_panic(rb_node, -EEXIST,
1108                                                    upper->bytenr);
1109                 }
1110
1111                 list_add_tail(&edge->list[UPPER], &upper->lower);
1112
1113                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1114                         list_add_tail(&edge->list[UPPER], &list);
1115         }
1116         /*
1117          * process useless backref nodes. backref nodes for tree leaves
1118          * are deleted from the cache. backref nodes for upper level
1119          * tree blocks are left in the cache to avoid unnecessary backref
1120          * lookup.
1121          */
1122         while (!list_empty(&useless)) {
1123                 upper = list_entry(useless.next, struct backref_node, list);
1124                 list_del_init(&upper->list);
1125                 ASSERT(list_empty(&upper->upper));
1126                 if (upper == node)
1127                         node = NULL;
1128                 if (upper->lowest) {
1129                         list_del_init(&upper->lower);
1130                         upper->lowest = 0;
1131                 }
1132                 while (!list_empty(&upper->lower)) {
1133                         edge = list_entry(upper->lower.next,
1134                                           struct backref_edge, list[UPPER]);
1135                         list_del(&edge->list[UPPER]);
1136                         list_del(&edge->list[LOWER]);
1137                         lower = edge->node[LOWER];
1138                         free_backref_edge(cache, edge);
1139
1140                         if (list_empty(&lower->upper))
1141                                 list_add(&lower->list, &useless);
1142                 }
1143                 __mark_block_processed(rc, upper);
1144                 if (upper->level > 0) {
1145                         list_add(&upper->list, &cache->detached);
1146                         upper->detached = 1;
1147                 } else {
1148                         rb_erase(&upper->rb_node, &cache->rb_root);
1149                         free_backref_node(cache, upper);
1150                 }
1151         }
1152 out:
1153         btrfs_free_path(path1);
1154         btrfs_free_path(path2);
1155         if (err) {
1156                 while (!list_empty(&useless)) {
1157                         lower = list_entry(useless.next,
1158                                            struct backref_node, list);
1159                         list_del_init(&lower->list);
1160                 }
1161                 while (!list_empty(&list)) {
1162                         edge = list_first_entry(&list, struct backref_edge,
1163                                                 list[UPPER]);
1164                         list_del(&edge->list[UPPER]);
1165                         list_del(&edge->list[LOWER]);
1166                         lower = edge->node[LOWER];
1167                         upper = edge->node[UPPER];
1168                         free_backref_edge(cache, edge);
1169
1170                         /*
1171                          * Lower is no longer linked to any upper backref nodes
1172                          * and isn't in the cache, we can free it ourselves.
1173                          */
1174                         if (list_empty(&lower->upper) &&
1175                             RB_EMPTY_NODE(&lower->rb_node))
1176                                 list_add(&lower->list, &useless);
1177
1178                         if (!RB_EMPTY_NODE(&upper->rb_node))
1179                                 continue;
1180
1181                         /* Add this guy's upper edges to the list to process */
1182                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1183                                 list_add_tail(&edge->list[UPPER], &list);
1184                         if (list_empty(&upper->upper))
1185                                 list_add(&upper->list, &useless);
1186                 }
1187
1188                 while (!list_empty(&useless)) {
1189                         lower = list_entry(useless.next,
1190                                            struct backref_node, list);
1191                         list_del_init(&lower->list);
1192                         if (lower == node)
1193                                 node = NULL;
1194                         free_backref_node(cache, lower);
1195                 }
1196
1197                 free_backref_node(cache, node);
1198                 return ERR_PTR(err);
1199         }
1200         ASSERT(!node || !node->detached);
1201         return node;
1202 }
1203
1204 /*
1205  * helper to add backref node for the newly created snapshot.
1206  * the backref node is created by cloning backref node that
1207  * corresponds to root of source tree
1208  */
1209 static int clone_backref_node(struct btrfs_trans_handle *trans,
1210                               struct reloc_control *rc,
1211                               struct btrfs_root *src,
1212                               struct btrfs_root *dest)
1213 {
1214         struct btrfs_root *reloc_root = src->reloc_root;
1215         struct backref_cache *cache = &rc->backref_cache;
1216         struct backref_node *node = NULL;
1217         struct backref_node *new_node;
1218         struct backref_edge *edge;
1219         struct backref_edge *new_edge;
1220         struct rb_node *rb_node;
1221
1222         if (cache->last_trans > 0)
1223                 update_backref_cache(trans, cache);
1224
1225         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1226         if (rb_node) {
1227                 node = rb_entry(rb_node, struct backref_node, rb_node);
1228                 if (node->detached)
1229                         node = NULL;
1230                 else
1231                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1232         }
1233
1234         if (!node) {
1235                 rb_node = tree_search(&cache->rb_root,
1236                                       reloc_root->commit_root->start);
1237                 if (rb_node) {
1238                         node = rb_entry(rb_node, struct backref_node,
1239                                         rb_node);
1240                         BUG_ON(node->detached);
1241                 }
1242         }
1243
1244         if (!node)
1245                 return 0;
1246
1247         new_node = alloc_backref_node(cache);
1248         if (!new_node)
1249                 return -ENOMEM;
1250
1251         new_node->bytenr = dest->node->start;
1252         new_node->level = node->level;
1253         new_node->lowest = node->lowest;
1254         new_node->checked = 1;
1255         new_node->root = dest;
1256
1257         if (!node->lowest) {
1258                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1259                         new_edge = alloc_backref_edge(cache);
1260                         if (!new_edge)
1261                                 goto fail;
1262
1263                         new_edge->node[UPPER] = new_node;
1264                         new_edge->node[LOWER] = edge->node[LOWER];
1265                         list_add_tail(&new_edge->list[UPPER],
1266                                       &new_node->lower);
1267                 }
1268         } else {
1269                 list_add_tail(&new_node->lower, &cache->leaves);
1270         }
1271
1272         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1273                               &new_node->rb_node);
1274         if (rb_node)
1275                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1276
1277         if (!new_node->lowest) {
1278                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1279                         list_add_tail(&new_edge->list[LOWER],
1280                                       &new_edge->node[LOWER]->upper);
1281                 }
1282         }
1283         return 0;
1284 fail:
1285         while (!list_empty(&new_node->lower)) {
1286                 new_edge = list_entry(new_node->lower.next,
1287                                       struct backref_edge, list[UPPER]);
1288                 list_del(&new_edge->list[UPPER]);
1289                 free_backref_edge(cache, new_edge);
1290         }
1291         free_backref_node(cache, new_node);
1292         return -ENOMEM;
1293 }
1294
1295 /*
1296  * helper to add 'address of tree root -> reloc tree' mapping
1297  */
1298 static int __must_check __add_reloc_root(struct btrfs_root *root)
1299 {
1300         struct btrfs_fs_info *fs_info = root->fs_info;
1301         struct rb_node *rb_node;
1302         struct mapping_node *node;
1303         struct reloc_control *rc = fs_info->reloc_ctl;
1304
1305         node = kmalloc(sizeof(*node), GFP_NOFS);
1306         if (!node)
1307                 return -ENOMEM;
1308
1309         node->bytenr = root->node->start;
1310         node->data = root;
1311
1312         spin_lock(&rc->reloc_root_tree.lock);
1313         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1314                               node->bytenr, &node->rb_node);
1315         spin_unlock(&rc->reloc_root_tree.lock);
1316         if (rb_node) {
1317                 btrfs_panic(fs_info, -EEXIST,
1318                             "Duplicate root found for start=%llu while inserting into relocation tree",
1319                             node->bytenr);
1320         }
1321
1322         list_add_tail(&root->root_list, &rc->reloc_roots);
1323         return 0;
1324 }
1325
1326 /*
1327  * helper to delete the 'address of tree root -> reloc tree'
1328  * mapping
1329  */
1330 static void __del_reloc_root(struct btrfs_root *root)
1331 {
1332         struct btrfs_fs_info *fs_info = root->fs_info;
1333         struct rb_node *rb_node;
1334         struct mapping_node *node = NULL;
1335         struct reloc_control *rc = fs_info->reloc_ctl;
1336
1337         if (rc && root->node) {
1338                 spin_lock(&rc->reloc_root_tree.lock);
1339                 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1340                                       root->node->start);
1341                 if (rb_node) {
1342                         node = rb_entry(rb_node, struct mapping_node, rb_node);
1343                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1344                 }
1345                 spin_unlock(&rc->reloc_root_tree.lock);
1346                 if (!node)
1347                         return;
1348                 BUG_ON((struct btrfs_root *)node->data != root);
1349         }
1350
1351         spin_lock(&fs_info->trans_lock);
1352         list_del_init(&root->root_list);
1353         spin_unlock(&fs_info->trans_lock);
1354         kfree(node);
1355 }
1356
1357 /*
1358  * helper to update the 'address of tree root -> reloc tree'
1359  * mapping
1360  */
1361 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1362 {
1363         struct btrfs_fs_info *fs_info = root->fs_info;
1364         struct rb_node *rb_node;
1365         struct mapping_node *node = NULL;
1366         struct reloc_control *rc = fs_info->reloc_ctl;
1367
1368         spin_lock(&rc->reloc_root_tree.lock);
1369         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1370                               root->node->start);
1371         if (rb_node) {
1372                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1373                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1374         }
1375         spin_unlock(&rc->reloc_root_tree.lock);
1376
1377         if (!node)
1378                 return 0;
1379         BUG_ON((struct btrfs_root *)node->data != root);
1380
1381         spin_lock(&rc->reloc_root_tree.lock);
1382         node->bytenr = new_bytenr;
1383         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1384                               node->bytenr, &node->rb_node);
1385         spin_unlock(&rc->reloc_root_tree.lock);
1386         if (rb_node)
1387                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1388         return 0;
1389 }
1390
1391 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1392                                         struct btrfs_root *root, u64 objectid)
1393 {
1394         struct btrfs_fs_info *fs_info = root->fs_info;
1395         struct btrfs_root *reloc_root;
1396         struct extent_buffer *eb;
1397         struct btrfs_root_item *root_item;
1398         struct btrfs_key root_key;
1399         int ret;
1400
1401         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1402         BUG_ON(!root_item);
1403
1404         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1405         root_key.type = BTRFS_ROOT_ITEM_KEY;
1406         root_key.offset = objectid;
1407
1408         if (root->root_key.objectid == objectid) {
1409                 u64 commit_root_gen;
1410
1411                 /* called by btrfs_init_reloc_root */
1412                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1413                                       BTRFS_TREE_RELOC_OBJECTID);
1414                 BUG_ON(ret);
1415                 /*
1416                  * Set the last_snapshot field to the generation of the commit
1417                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1418                  * correctly (returns true) when the relocation root is created
1419                  * either inside the critical section of a transaction commit
1420                  * (through transaction.c:qgroup_account_snapshot()) and when
1421                  * it's created before the transaction commit is started.
1422                  */
1423                 commit_root_gen = btrfs_header_generation(root->commit_root);
1424                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1425         } else {
1426                 /*
1427                  * called by btrfs_reloc_post_snapshot_hook.
1428                  * the source tree is a reloc tree, all tree blocks
1429                  * modified after it was created have RELOC flag
1430                  * set in their headers. so it's OK to not update
1431                  * the 'last_snapshot'.
1432                  */
1433                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1434                                       BTRFS_TREE_RELOC_OBJECTID);
1435                 BUG_ON(ret);
1436         }
1437
1438         memcpy(root_item, &root->root_item, sizeof(*root_item));
1439         btrfs_set_root_bytenr(root_item, eb->start);
1440         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1441         btrfs_set_root_generation(root_item, trans->transid);
1442
1443         if (root->root_key.objectid == objectid) {
1444                 btrfs_set_root_refs(root_item, 0);
1445                 memset(&root_item->drop_progress, 0,
1446                        sizeof(struct btrfs_disk_key));
1447                 root_item->drop_level = 0;
1448         }
1449
1450         btrfs_tree_unlock(eb);
1451         free_extent_buffer(eb);
1452
1453         ret = btrfs_insert_root(trans, fs_info->tree_root,
1454                                 &root_key, root_item);
1455         BUG_ON(ret);
1456         kfree(root_item);
1457
1458         reloc_root = btrfs_read_fs_root(fs_info->tree_root, &root_key);
1459         BUG_ON(IS_ERR(reloc_root));
1460         reloc_root->last_trans = trans->transid;
1461         return reloc_root;
1462 }
1463
1464 /*
1465  * create reloc tree for a given fs tree. reloc tree is just a
1466  * snapshot of the fs tree with special root objectid.
1467  */
1468 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1469                           struct btrfs_root *root)
1470 {
1471         struct btrfs_fs_info *fs_info = root->fs_info;
1472         struct btrfs_root *reloc_root;
1473         struct reloc_control *rc = fs_info->reloc_ctl;
1474         struct btrfs_block_rsv *rsv;
1475         int clear_rsv = 0;
1476         int ret;
1477
1478         if (root->reloc_root) {
1479                 reloc_root = root->reloc_root;
1480                 reloc_root->last_trans = trans->transid;
1481                 return 0;
1482         }
1483
1484         if (!rc || !rc->create_reloc_tree ||
1485             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1486                 return 0;
1487
1488         if (!trans->reloc_reserved) {
1489                 rsv = trans->block_rsv;
1490                 trans->block_rsv = rc->block_rsv;
1491                 clear_rsv = 1;
1492         }
1493         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1494         if (clear_rsv)
1495                 trans->block_rsv = rsv;
1496
1497         ret = __add_reloc_root(reloc_root);
1498         BUG_ON(ret < 0);
1499         root->reloc_root = reloc_root;
1500         return 0;
1501 }
1502
1503 /*
1504  * update root item of reloc tree
1505  */
1506 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1507                             struct btrfs_root *root)
1508 {
1509         struct btrfs_fs_info *fs_info = root->fs_info;
1510         struct btrfs_root *reloc_root;
1511         struct btrfs_root_item *root_item;
1512         int ret;
1513
1514         if (!root->reloc_root)
1515                 goto out;
1516
1517         reloc_root = root->reloc_root;
1518         root_item = &reloc_root->root_item;
1519
1520         if (fs_info->reloc_ctl->merge_reloc_tree &&
1521             btrfs_root_refs(root_item) == 0) {
1522                 root->reloc_root = NULL;
1523                 __del_reloc_root(reloc_root);
1524         }
1525
1526         if (reloc_root->commit_root != reloc_root->node) {
1527                 btrfs_set_root_node(root_item, reloc_root->node);
1528                 free_extent_buffer(reloc_root->commit_root);
1529                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1530         }
1531
1532         ret = btrfs_update_root(trans, fs_info->tree_root,
1533                                 &reloc_root->root_key, root_item);
1534         BUG_ON(ret);
1535
1536 out:
1537         return 0;
1538 }
1539
1540 /*
1541  * helper to find first cached inode with inode number >= objectid
1542  * in a subvolume
1543  */
1544 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1545 {
1546         struct rb_node *node;
1547         struct rb_node *prev;
1548         struct btrfs_inode *entry;
1549         struct inode *inode;
1550
1551         spin_lock(&root->inode_lock);
1552 again:
1553         node = root->inode_tree.rb_node;
1554         prev = NULL;
1555         while (node) {
1556                 prev = node;
1557                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1558
1559                 if (objectid < btrfs_ino(entry))
1560                         node = node->rb_left;
1561                 else if (objectid > btrfs_ino(entry))
1562                         node = node->rb_right;
1563                 else
1564                         break;
1565         }
1566         if (!node) {
1567                 while (prev) {
1568                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1569                         if (objectid <= btrfs_ino(entry)) {
1570                                 node = prev;
1571                                 break;
1572                         }
1573                         prev = rb_next(prev);
1574                 }
1575         }
1576         while (node) {
1577                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1578                 inode = igrab(&entry->vfs_inode);
1579                 if (inode) {
1580                         spin_unlock(&root->inode_lock);
1581                         return inode;
1582                 }
1583
1584                 objectid = btrfs_ino(entry) + 1;
1585                 if (cond_resched_lock(&root->inode_lock))
1586                         goto again;
1587
1588                 node = rb_next(node);
1589         }
1590         spin_unlock(&root->inode_lock);
1591         return NULL;
1592 }
1593
1594 static int in_block_group(u64 bytenr,
1595                           struct btrfs_block_group_cache *block_group)
1596 {
1597         if (bytenr >= block_group->key.objectid &&
1598             bytenr < block_group->key.objectid + block_group->key.offset)
1599                 return 1;
1600         return 0;
1601 }
1602
1603 /*
1604  * get new location of data
1605  */
1606 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1607                             u64 bytenr, u64 num_bytes)
1608 {
1609         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1610         struct btrfs_path *path;
1611         struct btrfs_file_extent_item *fi;
1612         struct extent_buffer *leaf;
1613         int ret;
1614
1615         path = btrfs_alloc_path();
1616         if (!path)
1617                 return -ENOMEM;
1618
1619         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1620         ret = btrfs_lookup_file_extent(NULL, root, path,
1621                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1622         if (ret < 0)
1623                 goto out;
1624         if (ret > 0) {
1625                 ret = -ENOENT;
1626                 goto out;
1627         }
1628
1629         leaf = path->nodes[0];
1630         fi = btrfs_item_ptr(leaf, path->slots[0],
1631                             struct btrfs_file_extent_item);
1632
1633         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1634                btrfs_file_extent_compression(leaf, fi) ||
1635                btrfs_file_extent_encryption(leaf, fi) ||
1636                btrfs_file_extent_other_encoding(leaf, fi));
1637
1638         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1639                 ret = -EINVAL;
1640                 goto out;
1641         }
1642
1643         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1644         ret = 0;
1645 out:
1646         btrfs_free_path(path);
1647         return ret;
1648 }
1649
1650 /*
1651  * update file extent items in the tree leaf to point to
1652  * the new locations.
1653  */
1654 static noinline_for_stack
1655 int replace_file_extents(struct btrfs_trans_handle *trans,
1656                          struct reloc_control *rc,
1657                          struct btrfs_root *root,
1658                          struct extent_buffer *leaf)
1659 {
1660         struct btrfs_fs_info *fs_info = root->fs_info;
1661         struct btrfs_key key;
1662         struct btrfs_file_extent_item *fi;
1663         struct inode *inode = NULL;
1664         u64 parent;
1665         u64 bytenr;
1666         u64 new_bytenr = 0;
1667         u64 num_bytes;
1668         u64 end;
1669         u32 nritems;
1670         u32 i;
1671         int ret = 0;
1672         int first = 1;
1673         int dirty = 0;
1674
1675         if (rc->stage != UPDATE_DATA_PTRS)
1676                 return 0;
1677
1678         /* reloc trees always use full backref */
1679         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1680                 parent = leaf->start;
1681         else
1682                 parent = 0;
1683
1684         nritems = btrfs_header_nritems(leaf);
1685         for (i = 0; i < nritems; i++) {
1686                 cond_resched();
1687                 btrfs_item_key_to_cpu(leaf, &key, i);
1688                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1689                         continue;
1690                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1691                 if (btrfs_file_extent_type(leaf, fi) ==
1692                     BTRFS_FILE_EXTENT_INLINE)
1693                         continue;
1694                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1695                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1696                 if (bytenr == 0)
1697                         continue;
1698                 if (!in_block_group(bytenr, rc->block_group))
1699                         continue;
1700
1701                 /*
1702                  * if we are modifying block in fs tree, wait for readpage
1703                  * to complete and drop the extent cache
1704                  */
1705                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1706                         if (first) {
1707                                 inode = find_next_inode(root, key.objectid);
1708                                 first = 0;
1709                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1710                                 btrfs_add_delayed_iput(inode);
1711                                 inode = find_next_inode(root, key.objectid);
1712                         }
1713                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1714                                 end = key.offset +
1715                                       btrfs_file_extent_num_bytes(leaf, fi);
1716                                 WARN_ON(!IS_ALIGNED(key.offset,
1717                                                     fs_info->sectorsize));
1718                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1719                                 end--;
1720                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1721                                                       key.offset, end);
1722                                 if (!ret)
1723                                         continue;
1724
1725                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1726                                                 key.offset,     end, 1);
1727                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1728                                               key.offset, end);
1729                         }
1730                 }
1731
1732                 ret = get_new_location(rc->data_inode, &new_bytenr,
1733                                        bytenr, num_bytes);
1734                 if (ret) {
1735                         /*
1736                          * Don't have to abort since we've not changed anything
1737                          * in the file extent yet.
1738                          */
1739                         break;
1740                 }
1741
1742                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1743                 dirty = 1;
1744
1745                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1746                 ret = btrfs_inc_extent_ref(trans, fs_info, new_bytenr,
1747                                            num_bytes, parent,
1748                                            btrfs_header_owner(leaf),
1749                                            key.objectid, key.offset);
1750                 if (ret) {
1751                         btrfs_abort_transaction(trans, ret);
1752                         break;
1753                 }
1754
1755                 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1756                                         parent, btrfs_header_owner(leaf),
1757                                         key.objectid, key.offset);
1758                 if (ret) {
1759                         btrfs_abort_transaction(trans, ret);
1760                         break;
1761                 }
1762         }
1763         if (dirty)
1764                 btrfs_mark_buffer_dirty(leaf);
1765         if (inode)
1766                 btrfs_add_delayed_iput(inode);
1767         return ret;
1768 }
1769
1770 static noinline_for_stack
1771 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1772                      struct btrfs_path *path, int level)
1773 {
1774         struct btrfs_disk_key key1;
1775         struct btrfs_disk_key key2;
1776         btrfs_node_key(eb, &key1, slot);
1777         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1778         return memcmp(&key1, &key2, sizeof(key1));
1779 }
1780
1781 /*
1782  * try to replace tree blocks in fs tree with the new blocks
1783  * in reloc tree. tree blocks haven't been modified since the
1784  * reloc tree was create can be replaced.
1785  *
1786  * if a block was replaced, level of the block + 1 is returned.
1787  * if no block got replaced, 0 is returned. if there are other
1788  * errors, a negative error number is returned.
1789  */
1790 static noinline_for_stack
1791 int replace_path(struct btrfs_trans_handle *trans,
1792                  struct btrfs_root *dest, struct btrfs_root *src,
1793                  struct btrfs_path *path, struct btrfs_key *next_key,
1794                  int lowest_level, int max_level)
1795 {
1796         struct btrfs_fs_info *fs_info = dest->fs_info;
1797         struct extent_buffer *eb;
1798         struct extent_buffer *parent;
1799         struct btrfs_key key;
1800         u64 old_bytenr;
1801         u64 new_bytenr;
1802         u64 old_ptr_gen;
1803         u64 new_ptr_gen;
1804         u64 last_snapshot;
1805         u32 blocksize;
1806         int cow = 0;
1807         int level;
1808         int ret;
1809         int slot;
1810
1811         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1812         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1813
1814         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1815 again:
1816         slot = path->slots[lowest_level];
1817         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1818
1819         eb = btrfs_lock_root_node(dest);
1820         btrfs_set_lock_blocking(eb);
1821         level = btrfs_header_level(eb);
1822
1823         if (level < lowest_level) {
1824                 btrfs_tree_unlock(eb);
1825                 free_extent_buffer(eb);
1826                 return 0;
1827         }
1828
1829         if (cow) {
1830                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1831                 BUG_ON(ret);
1832         }
1833         btrfs_set_lock_blocking(eb);
1834
1835         if (next_key) {
1836                 next_key->objectid = (u64)-1;
1837                 next_key->type = (u8)-1;
1838                 next_key->offset = (u64)-1;
1839         }
1840
1841         parent = eb;
1842         while (1) {
1843                 level = btrfs_header_level(parent);
1844                 BUG_ON(level < lowest_level);
1845
1846                 ret = btrfs_bin_search(parent, &key, level, &slot);
1847                 if (ret && slot > 0)
1848                         slot--;
1849
1850                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1851                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1852
1853                 old_bytenr = btrfs_node_blockptr(parent, slot);
1854                 blocksize = fs_info->nodesize;
1855                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1856
1857                 if (level <= max_level) {
1858                         eb = path->nodes[level];
1859                         new_bytenr = btrfs_node_blockptr(eb,
1860                                                         path->slots[level]);
1861                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1862                                                         path->slots[level]);
1863                 } else {
1864                         new_bytenr = 0;
1865                         new_ptr_gen = 0;
1866                 }
1867
1868                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1869                         ret = level;
1870                         break;
1871                 }
1872
1873                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1874                     memcmp_node_keys(parent, slot, path, level)) {
1875                         if (level <= lowest_level) {
1876                                 ret = 0;
1877                                 break;
1878                         }
1879
1880                         eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen);
1881                         if (IS_ERR(eb)) {
1882                                 ret = PTR_ERR(eb);
1883                                 break;
1884                         } else if (!extent_buffer_uptodate(eb)) {
1885                                 ret = -EIO;
1886                                 free_extent_buffer(eb);
1887                                 break;
1888                         }
1889                         btrfs_tree_lock(eb);
1890                         if (cow) {
1891                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1892                                                       slot, &eb);
1893                                 BUG_ON(ret);
1894                         }
1895                         btrfs_set_lock_blocking(eb);
1896
1897                         btrfs_tree_unlock(parent);
1898                         free_extent_buffer(parent);
1899
1900                         parent = eb;
1901                         continue;
1902                 }
1903
1904                 if (!cow) {
1905                         btrfs_tree_unlock(parent);
1906                         free_extent_buffer(parent);
1907                         cow = 1;
1908                         goto again;
1909                 }
1910
1911                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1912                                       path->slots[level]);
1913                 btrfs_release_path(path);
1914
1915                 path->lowest_level = level;
1916                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1917                 path->lowest_level = 0;
1918                 BUG_ON(ret);
1919
1920                 /*
1921                  * Info qgroup to trace both subtrees.
1922                  *
1923                  * We must trace both trees.
1924                  * 1) Tree reloc subtree
1925                  *    If not traced, we will leak data numbers
1926                  * 2) Fs subtree
1927                  *    If not traced, we will double count old data
1928                  *    and tree block numbers, if current trans doesn't free
1929                  *    data reloc tree inode.
1930                  */
1931                 ret = btrfs_qgroup_trace_subtree(trans, src, parent,
1932                                 btrfs_header_generation(parent),
1933                                 btrfs_header_level(parent));
1934                 if (ret < 0)
1935                         break;
1936                 ret = btrfs_qgroup_trace_subtree(trans, dest,
1937                                 path->nodes[level],
1938                                 btrfs_header_generation(path->nodes[level]),
1939                                 btrfs_header_level(path->nodes[level]));
1940                 if (ret < 0)
1941                         break;
1942
1943                 /*
1944                  * swap blocks in fs tree and reloc tree.
1945                  */
1946                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1947                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1948                 btrfs_mark_buffer_dirty(parent);
1949
1950                 btrfs_set_node_blockptr(path->nodes[level],
1951                                         path->slots[level], old_bytenr);
1952                 btrfs_set_node_ptr_generation(path->nodes[level],
1953                                               path->slots[level], old_ptr_gen);
1954                 btrfs_mark_buffer_dirty(path->nodes[level]);
1955
1956                 ret = btrfs_inc_extent_ref(trans, fs_info, old_bytenr,
1957                                         blocksize, path->nodes[level]->start,
1958                                         src->root_key.objectid, level - 1, 0);
1959                 BUG_ON(ret);
1960                 ret = btrfs_inc_extent_ref(trans, fs_info, new_bytenr,
1961                                         blocksize, 0, dest->root_key.objectid,
1962                                         level - 1, 0);
1963                 BUG_ON(ret);
1964
1965                 ret = btrfs_free_extent(trans, fs_info, new_bytenr, blocksize,
1966                                         path->nodes[level]->start,
1967                                         src->root_key.objectid, level - 1, 0);
1968                 BUG_ON(ret);
1969
1970                 ret = btrfs_free_extent(trans, fs_info, old_bytenr, blocksize,
1971                                         0, dest->root_key.objectid, level - 1,
1972                                         0);
1973                 BUG_ON(ret);
1974
1975                 btrfs_unlock_up_safe(path, 0);
1976
1977                 ret = level;
1978                 break;
1979         }
1980         btrfs_tree_unlock(parent);
1981         free_extent_buffer(parent);
1982         return ret;
1983 }
1984
1985 /*
1986  * helper to find next relocated block in reloc tree
1987  */
1988 static noinline_for_stack
1989 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1990                        int *level)
1991 {
1992         struct extent_buffer *eb;
1993         int i;
1994         u64 last_snapshot;
1995         u32 nritems;
1996
1997         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1998
1999         for (i = 0; i < *level; i++) {
2000                 free_extent_buffer(path->nodes[i]);
2001                 path->nodes[i] = NULL;
2002         }
2003
2004         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
2005                 eb = path->nodes[i];
2006                 nritems = btrfs_header_nritems(eb);
2007                 while (path->slots[i] + 1 < nritems) {
2008                         path->slots[i]++;
2009                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
2010                             last_snapshot)
2011                                 continue;
2012
2013                         *level = i;
2014                         return 0;
2015                 }
2016                 free_extent_buffer(path->nodes[i]);
2017                 path->nodes[i] = NULL;
2018         }
2019         return 1;
2020 }
2021
2022 /*
2023  * walk down reloc tree to find relocated block of lowest level
2024  */
2025 static noinline_for_stack
2026 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
2027                          int *level)
2028 {
2029         struct btrfs_fs_info *fs_info = root->fs_info;
2030         struct extent_buffer *eb = NULL;
2031         int i;
2032         u64 bytenr;
2033         u64 ptr_gen = 0;
2034         u64 last_snapshot;
2035         u32 nritems;
2036
2037         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2038
2039         for (i = *level; i > 0; i--) {
2040                 eb = path->nodes[i];
2041                 nritems = btrfs_header_nritems(eb);
2042                 while (path->slots[i] < nritems) {
2043                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2044                         if (ptr_gen > last_snapshot)
2045                                 break;
2046                         path->slots[i]++;
2047                 }
2048                 if (path->slots[i] >= nritems) {
2049                         if (i == *level)
2050                                 break;
2051                         *level = i + 1;
2052                         return 0;
2053                 }
2054                 if (i == 1) {
2055                         *level = i;
2056                         return 0;
2057                 }
2058
2059                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2060                 eb = read_tree_block(fs_info, bytenr, ptr_gen);
2061                 if (IS_ERR(eb)) {
2062                         return PTR_ERR(eb);
2063                 } else if (!extent_buffer_uptodate(eb)) {
2064                         free_extent_buffer(eb);
2065                         return -EIO;
2066                 }
2067                 BUG_ON(btrfs_header_level(eb) != i - 1);
2068                 path->nodes[i - 1] = eb;
2069                 path->slots[i - 1] = 0;
2070         }
2071         return 1;
2072 }
2073
2074 /*
2075  * invalidate extent cache for file extents whose key in range of
2076  * [min_key, max_key)
2077  */
2078 static int invalidate_extent_cache(struct btrfs_root *root,
2079                                    struct btrfs_key *min_key,
2080                                    struct btrfs_key *max_key)
2081 {
2082         struct btrfs_fs_info *fs_info = root->fs_info;
2083         struct inode *inode = NULL;
2084         u64 objectid;
2085         u64 start, end;
2086         u64 ino;
2087
2088         objectid = min_key->objectid;
2089         while (1) {
2090                 cond_resched();
2091                 iput(inode);
2092
2093                 if (objectid > max_key->objectid)
2094                         break;
2095
2096                 inode = find_next_inode(root, objectid);
2097                 if (!inode)
2098                         break;
2099                 ino = btrfs_ino(BTRFS_I(inode));
2100
2101                 if (ino > max_key->objectid) {
2102                         iput(inode);
2103                         break;
2104                 }
2105
2106                 objectid = ino + 1;
2107                 if (!S_ISREG(inode->i_mode))
2108                         continue;
2109
2110                 if (unlikely(min_key->objectid == ino)) {
2111                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2112                                 continue;
2113                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2114                                 start = 0;
2115                         else {
2116                                 start = min_key->offset;
2117                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
2118                         }
2119                 } else {
2120                         start = 0;
2121                 }
2122
2123                 if (unlikely(max_key->objectid == ino)) {
2124                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2125                                 continue;
2126                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2127                                 end = (u64)-1;
2128                         } else {
2129                                 if (max_key->offset == 0)
2130                                         continue;
2131                                 end = max_key->offset;
2132                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
2133                                 end--;
2134                         }
2135                 } else {
2136                         end = (u64)-1;
2137                 }
2138
2139                 /* the lock_extent waits for readpage to complete */
2140                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2141                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
2142                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2143         }
2144         return 0;
2145 }
2146
2147 static int find_next_key(struct btrfs_path *path, int level,
2148                          struct btrfs_key *key)
2149
2150 {
2151         while (level < BTRFS_MAX_LEVEL) {
2152                 if (!path->nodes[level])
2153                         break;
2154                 if (path->slots[level] + 1 <
2155                     btrfs_header_nritems(path->nodes[level])) {
2156                         btrfs_node_key_to_cpu(path->nodes[level], key,
2157                                               path->slots[level] + 1);
2158                         return 0;
2159                 }
2160                 level++;
2161         }
2162         return 1;
2163 }
2164
2165 /*
2166  * merge the relocated tree blocks in reloc tree with corresponding
2167  * fs tree.
2168  */
2169 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2170                                                struct btrfs_root *root)
2171 {
2172         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2173         LIST_HEAD(inode_list);
2174         struct btrfs_key key;
2175         struct btrfs_key next_key;
2176         struct btrfs_trans_handle *trans = NULL;
2177         struct btrfs_root *reloc_root;
2178         struct btrfs_root_item *root_item;
2179         struct btrfs_path *path;
2180         struct extent_buffer *leaf;
2181         int level;
2182         int max_level;
2183         int replaced = 0;
2184         int ret;
2185         int err = 0;
2186         u32 min_reserved;
2187
2188         path = btrfs_alloc_path();
2189         if (!path)
2190                 return -ENOMEM;
2191         path->reada = READA_FORWARD;
2192
2193         reloc_root = root->reloc_root;
2194         root_item = &reloc_root->root_item;
2195
2196         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2197                 level = btrfs_root_level(root_item);
2198                 extent_buffer_get(reloc_root->node);
2199                 path->nodes[level] = reloc_root->node;
2200                 path->slots[level] = 0;
2201         } else {
2202                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2203
2204                 level = root_item->drop_level;
2205                 BUG_ON(level == 0);
2206                 path->lowest_level = level;
2207                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2208                 path->lowest_level = 0;
2209                 if (ret < 0) {
2210                         btrfs_free_path(path);
2211                         return ret;
2212                 }
2213
2214                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2215                                       path->slots[level]);
2216                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2217
2218                 btrfs_unlock_up_safe(path, 0);
2219         }
2220
2221         min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2222         memset(&next_key, 0, sizeof(next_key));
2223
2224         while (1) {
2225                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2226                                              BTRFS_RESERVE_FLUSH_ALL);
2227                 if (ret) {
2228                         err = ret;
2229                         goto out;
2230                 }
2231                 trans = btrfs_start_transaction(root, 0);
2232                 if (IS_ERR(trans)) {
2233                         err = PTR_ERR(trans);
2234                         trans = NULL;
2235                         goto out;
2236                 }
2237                 trans->block_rsv = rc->block_rsv;
2238
2239                 replaced = 0;
2240                 max_level = level;
2241
2242                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2243                 if (ret < 0) {
2244                         err = ret;
2245                         goto out;
2246                 }
2247                 if (ret > 0)
2248                         break;
2249
2250                 if (!find_next_key(path, level, &key) &&
2251                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2252                         ret = 0;
2253                 } else {
2254                         ret = replace_path(trans, root, reloc_root, path,
2255                                            &next_key, level, max_level);
2256                 }
2257                 if (ret < 0) {
2258                         err = ret;
2259                         goto out;
2260                 }
2261
2262                 if (ret > 0) {
2263                         level = ret;
2264                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2265                                               path->slots[level]);
2266                         replaced = 1;
2267                 }
2268
2269                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2270                 if (ret > 0)
2271                         break;
2272
2273                 BUG_ON(level == 0);
2274                 /*
2275                  * save the merging progress in the drop_progress.
2276                  * this is OK since root refs == 1 in this case.
2277                  */
2278                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2279                                path->slots[level]);
2280                 root_item->drop_level = level;
2281
2282                 btrfs_end_transaction_throttle(trans);
2283                 trans = NULL;
2284
2285                 btrfs_btree_balance_dirty(fs_info);
2286
2287                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2288                         invalidate_extent_cache(root, &key, &next_key);
2289         }
2290
2291         /*
2292          * handle the case only one block in the fs tree need to be
2293          * relocated and the block is tree root.
2294          */
2295         leaf = btrfs_lock_root_node(root);
2296         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2297         btrfs_tree_unlock(leaf);
2298         free_extent_buffer(leaf);
2299         if (ret < 0)
2300                 err = ret;
2301 out:
2302         btrfs_free_path(path);
2303
2304         if (err == 0) {
2305                 memset(&root_item->drop_progress, 0,
2306                        sizeof(root_item->drop_progress));
2307                 root_item->drop_level = 0;
2308                 btrfs_set_root_refs(root_item, 0);
2309                 btrfs_update_reloc_root(trans, root);
2310         }
2311
2312         if (trans)
2313                 btrfs_end_transaction_throttle(trans);
2314
2315         btrfs_btree_balance_dirty(fs_info);
2316
2317         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2318                 invalidate_extent_cache(root, &key, &next_key);
2319
2320         return err;
2321 }
2322
2323 static noinline_for_stack
2324 int prepare_to_merge(struct reloc_control *rc, int err)
2325 {
2326         struct btrfs_root *root = rc->extent_root;
2327         struct btrfs_fs_info *fs_info = root->fs_info;
2328         struct btrfs_root *reloc_root;
2329         struct btrfs_trans_handle *trans;
2330         LIST_HEAD(reloc_roots);
2331         u64 num_bytes = 0;
2332         int ret;
2333
2334         mutex_lock(&fs_info->reloc_mutex);
2335         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2336         rc->merging_rsv_size += rc->nodes_relocated * 2;
2337         mutex_unlock(&fs_info->reloc_mutex);
2338
2339 again:
2340         if (!err) {
2341                 num_bytes = rc->merging_rsv_size;
2342                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2343                                           BTRFS_RESERVE_FLUSH_ALL);
2344                 if (ret)
2345                         err = ret;
2346         }
2347
2348         trans = btrfs_join_transaction(rc->extent_root);
2349         if (IS_ERR(trans)) {
2350                 if (!err)
2351                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2352                                                 num_bytes);
2353                 return PTR_ERR(trans);
2354         }
2355
2356         if (!err) {
2357                 if (num_bytes != rc->merging_rsv_size) {
2358                         btrfs_end_transaction(trans);
2359                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2360                                                 num_bytes);
2361                         goto again;
2362                 }
2363         }
2364
2365         rc->merge_reloc_tree = 1;
2366
2367         while (!list_empty(&rc->reloc_roots)) {
2368                 reloc_root = list_entry(rc->reloc_roots.next,
2369                                         struct btrfs_root, root_list);
2370                 list_del_init(&reloc_root->root_list);
2371
2372                 root = read_fs_root(fs_info, reloc_root->root_key.offset);
2373                 BUG_ON(IS_ERR(root));
2374                 BUG_ON(root->reloc_root != reloc_root);
2375
2376                 /*
2377                  * set reference count to 1, so btrfs_recover_relocation
2378                  * knows it should resumes merging
2379                  */
2380                 if (!err)
2381                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2382                 btrfs_update_reloc_root(trans, root);
2383
2384                 list_add(&reloc_root->root_list, &reloc_roots);
2385         }
2386
2387         list_splice(&reloc_roots, &rc->reloc_roots);
2388
2389         if (!err)
2390                 btrfs_commit_transaction(trans);
2391         else
2392                 btrfs_end_transaction(trans);
2393         return err;
2394 }
2395
2396 static noinline_for_stack
2397 void free_reloc_roots(struct list_head *list)
2398 {
2399         struct btrfs_root *reloc_root;
2400
2401         while (!list_empty(list)) {
2402                 reloc_root = list_entry(list->next, struct btrfs_root,
2403                                         root_list);
2404                 __del_reloc_root(reloc_root);
2405                 free_extent_buffer(reloc_root->node);
2406                 free_extent_buffer(reloc_root->commit_root);
2407                 reloc_root->node = NULL;
2408                 reloc_root->commit_root = NULL;
2409         }
2410 }
2411
2412 static noinline_for_stack
2413 void merge_reloc_roots(struct reloc_control *rc)
2414 {
2415         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2416         struct btrfs_root *root;
2417         struct btrfs_root *reloc_root;
2418         LIST_HEAD(reloc_roots);
2419         int found = 0;
2420         int ret = 0;
2421 again:
2422         root = rc->extent_root;
2423
2424         /*
2425          * this serializes us with btrfs_record_root_in_transaction,
2426          * we have to make sure nobody is in the middle of
2427          * adding their roots to the list while we are
2428          * doing this splice
2429          */
2430         mutex_lock(&fs_info->reloc_mutex);
2431         list_splice_init(&rc->reloc_roots, &reloc_roots);
2432         mutex_unlock(&fs_info->reloc_mutex);
2433
2434         while (!list_empty(&reloc_roots)) {
2435                 found = 1;
2436                 reloc_root = list_entry(reloc_roots.next,
2437                                         struct btrfs_root, root_list);
2438
2439                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2440                         root = read_fs_root(fs_info,
2441                                             reloc_root->root_key.offset);
2442                         BUG_ON(IS_ERR(root));
2443                         BUG_ON(root->reloc_root != reloc_root);
2444
2445                         ret = merge_reloc_root(rc, root);
2446                         if (ret) {
2447                                 if (list_empty(&reloc_root->root_list))
2448                                         list_add_tail(&reloc_root->root_list,
2449                                                       &reloc_roots);
2450                                 goto out;
2451                         }
2452                 } else {
2453                         list_del_init(&reloc_root->root_list);
2454                 }
2455
2456                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2457                 if (ret < 0) {
2458                         if (list_empty(&reloc_root->root_list))
2459                                 list_add_tail(&reloc_root->root_list,
2460                                               &reloc_roots);
2461                         goto out;
2462                 }
2463         }
2464
2465         if (found) {
2466                 found = 0;
2467                 goto again;
2468         }
2469 out:
2470         if (ret) {
2471                 btrfs_handle_fs_error(fs_info, ret, NULL);
2472                 if (!list_empty(&reloc_roots))
2473                         free_reloc_roots(&reloc_roots);
2474
2475                 /* new reloc root may be added */
2476                 mutex_lock(&fs_info->reloc_mutex);
2477                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2478                 mutex_unlock(&fs_info->reloc_mutex);
2479                 if (!list_empty(&reloc_roots))
2480                         free_reloc_roots(&reloc_roots);
2481         }
2482
2483         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2484 }
2485
2486 static void free_block_list(struct rb_root *blocks)
2487 {
2488         struct tree_block *block;
2489         struct rb_node *rb_node;
2490         while ((rb_node = rb_first(blocks))) {
2491                 block = rb_entry(rb_node, struct tree_block, rb_node);
2492                 rb_erase(rb_node, blocks);
2493                 kfree(block);
2494         }
2495 }
2496
2497 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2498                                       struct btrfs_root *reloc_root)
2499 {
2500         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2501         struct btrfs_root *root;
2502
2503         if (reloc_root->last_trans == trans->transid)
2504                 return 0;
2505
2506         root = read_fs_root(fs_info, reloc_root->root_key.offset);
2507         BUG_ON(IS_ERR(root));
2508         BUG_ON(root->reloc_root != reloc_root);
2509
2510         return btrfs_record_root_in_trans(trans, root);
2511 }
2512
2513 static noinline_for_stack
2514 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2515                                      struct reloc_control *rc,
2516                                      struct backref_node *node,
2517                                      struct backref_edge *edges[])
2518 {
2519         struct backref_node *next;
2520         struct btrfs_root *root;
2521         int index = 0;
2522
2523         next = node;
2524         while (1) {
2525                 cond_resched();
2526                 next = walk_up_backref(next, edges, &index);
2527                 root = next->root;
2528                 BUG_ON(!root);
2529                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2530
2531                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2532                         record_reloc_root_in_trans(trans, root);
2533                         break;
2534                 }
2535
2536                 btrfs_record_root_in_trans(trans, root);
2537                 root = root->reloc_root;
2538
2539                 if (next->new_bytenr != root->node->start) {
2540                         BUG_ON(next->new_bytenr);
2541                         BUG_ON(!list_empty(&next->list));
2542                         next->new_bytenr = root->node->start;
2543                         next->root = root;
2544                         list_add_tail(&next->list,
2545                                       &rc->backref_cache.changed);
2546                         __mark_block_processed(rc, next);
2547                         break;
2548                 }
2549
2550                 WARN_ON(1);
2551                 root = NULL;
2552                 next = walk_down_backref(edges, &index);
2553                 if (!next || next->level <= node->level)
2554                         break;
2555         }
2556         if (!root)
2557                 return NULL;
2558
2559         next = node;
2560         /* setup backref node path for btrfs_reloc_cow_block */
2561         while (1) {
2562                 rc->backref_cache.path[next->level] = next;
2563                 if (--index < 0)
2564                         break;
2565                 next = edges[index]->node[UPPER];
2566         }
2567         return root;
2568 }
2569
2570 /*
2571  * select a tree root for relocation. return NULL if the block
2572  * is reference counted. we should use do_relocation() in this
2573  * case. return a tree root pointer if the block isn't reference
2574  * counted. return -ENOENT if the block is root of reloc tree.
2575  */
2576 static noinline_for_stack
2577 struct btrfs_root *select_one_root(struct backref_node *node)
2578 {
2579         struct backref_node *next;
2580         struct btrfs_root *root;
2581         struct btrfs_root *fs_root = NULL;
2582         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2583         int index = 0;
2584
2585         next = node;
2586         while (1) {
2587                 cond_resched();
2588                 next = walk_up_backref(next, edges, &index);
2589                 root = next->root;
2590                 BUG_ON(!root);
2591
2592                 /* no other choice for non-references counted tree */
2593                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2594                         return root;
2595
2596                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2597                         fs_root = root;
2598
2599                 if (next != node)
2600                         return NULL;
2601
2602                 next = walk_down_backref(edges, &index);
2603                 if (!next || next->level <= node->level)
2604                         break;
2605         }
2606
2607         if (!fs_root)
2608                 return ERR_PTR(-ENOENT);
2609         return fs_root;
2610 }
2611
2612 static noinline_for_stack
2613 u64 calcu_metadata_size(struct reloc_control *rc,
2614                         struct backref_node *node, int reserve)
2615 {
2616         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2617         struct backref_node *next = node;
2618         struct backref_edge *edge;
2619         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2620         u64 num_bytes = 0;
2621         int index = 0;
2622
2623         BUG_ON(reserve && node->processed);
2624
2625         while (next) {
2626                 cond_resched();
2627                 while (1) {
2628                         if (next->processed && (reserve || next != node))
2629                                 break;
2630
2631                         num_bytes += fs_info->nodesize;
2632
2633                         if (list_empty(&next->upper))
2634                                 break;
2635
2636                         edge = list_entry(next->upper.next,
2637                                           struct backref_edge, list[LOWER]);
2638                         edges[index++] = edge;
2639                         next = edge->node[UPPER];
2640                 }
2641                 next = walk_down_backref(edges, &index);
2642         }
2643         return num_bytes;
2644 }
2645
2646 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2647                                   struct reloc_control *rc,
2648                                   struct backref_node *node)
2649 {
2650         struct btrfs_root *root = rc->extent_root;
2651         struct btrfs_fs_info *fs_info = root->fs_info;
2652         u64 num_bytes;
2653         int ret;
2654         u64 tmp;
2655
2656         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2657
2658         trans->block_rsv = rc->block_rsv;
2659         rc->reserved_bytes += num_bytes;
2660
2661         /*
2662          * We are under a transaction here so we can only do limited flushing.
2663          * If we get an enospc just kick back -EAGAIN so we know to drop the
2664          * transaction and try to refill when we can flush all the things.
2665          */
2666         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2667                                 BTRFS_RESERVE_FLUSH_LIMIT);
2668         if (ret) {
2669                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2670                 while (tmp <= rc->reserved_bytes)
2671                         tmp <<= 1;
2672                 /*
2673                  * only one thread can access block_rsv at this point,
2674                  * so we don't need hold lock to protect block_rsv.
2675                  * we expand more reservation size here to allow enough
2676                  * space for relocation and we will return eailer in
2677                  * enospc case.
2678                  */
2679                 rc->block_rsv->size = tmp + fs_info->nodesize *
2680                                       RELOCATION_RESERVED_NODES;
2681                 return -EAGAIN;
2682         }
2683
2684         return 0;
2685 }
2686
2687 /*
2688  * relocate a block tree, and then update pointers in upper level
2689  * blocks that reference the block to point to the new location.
2690  *
2691  * if called by link_to_upper, the block has already been relocated.
2692  * in that case this function just updates pointers.
2693  */
2694 static int do_relocation(struct btrfs_trans_handle *trans,
2695                          struct reloc_control *rc,
2696                          struct backref_node *node,
2697                          struct btrfs_key *key,
2698                          struct btrfs_path *path, int lowest)
2699 {
2700         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2701         struct backref_node *upper;
2702         struct backref_edge *edge;
2703         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2704         struct btrfs_root *root;
2705         struct extent_buffer *eb;
2706         u32 blocksize;
2707         u64 bytenr;
2708         u64 generation;
2709         int slot;
2710         int ret;
2711         int err = 0;
2712
2713         BUG_ON(lowest && node->eb);
2714
2715         path->lowest_level = node->level + 1;
2716         rc->backref_cache.path[node->level] = node;
2717         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2718                 cond_resched();
2719
2720                 upper = edge->node[UPPER];
2721                 root = select_reloc_root(trans, rc, upper, edges);
2722                 BUG_ON(!root);
2723
2724                 if (upper->eb && !upper->locked) {
2725                         if (!lowest) {
2726                                 ret = btrfs_bin_search(upper->eb, key,
2727                                                        upper->level, &slot);
2728                                 BUG_ON(ret);
2729                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2730                                 if (node->eb->start == bytenr)
2731                                         goto next;
2732                         }
2733                         drop_node_buffer(upper);
2734                 }
2735
2736                 if (!upper->eb) {
2737                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2738                         if (ret) {
2739                                 if (ret < 0)
2740                                         err = ret;
2741                                 else
2742                                         err = -ENOENT;
2743
2744                                 btrfs_release_path(path);
2745                                 break;
2746                         }
2747
2748                         if (!upper->eb) {
2749                                 upper->eb = path->nodes[upper->level];
2750                                 path->nodes[upper->level] = NULL;
2751                         } else {
2752                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2753                         }
2754
2755                         upper->locked = 1;
2756                         path->locks[upper->level] = 0;
2757
2758                         slot = path->slots[upper->level];
2759                         btrfs_release_path(path);
2760                 } else {
2761                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2762                                                &slot);
2763                         BUG_ON(ret);
2764                 }
2765
2766                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2767                 if (lowest) {
2768                         if (bytenr != node->bytenr) {
2769                                 btrfs_err(root->fs_info,
2770                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2771                                           bytenr, node->bytenr, slot,
2772                                           upper->eb->start);
2773                                 err = -EIO;
2774                                 goto next;
2775                         }
2776                 } else {
2777                         if (node->eb->start == bytenr)
2778                                 goto next;
2779                 }
2780
2781                 blocksize = root->fs_info->nodesize;
2782                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2783                 eb = read_tree_block(fs_info, bytenr, generation);
2784                 if (IS_ERR(eb)) {
2785                         err = PTR_ERR(eb);
2786                         goto next;
2787                 } else if (!extent_buffer_uptodate(eb)) {
2788                         free_extent_buffer(eb);
2789                         err = -EIO;
2790                         goto next;
2791                 }
2792                 btrfs_tree_lock(eb);
2793                 btrfs_set_lock_blocking(eb);
2794
2795                 if (!node->eb) {
2796                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2797                                               slot, &eb);
2798                         btrfs_tree_unlock(eb);
2799                         free_extent_buffer(eb);
2800                         if (ret < 0) {
2801                                 err = ret;
2802                                 goto next;
2803                         }
2804                         BUG_ON(node->eb != eb);
2805                 } else {
2806                         btrfs_set_node_blockptr(upper->eb, slot,
2807                                                 node->eb->start);
2808                         btrfs_set_node_ptr_generation(upper->eb, slot,
2809                                                       trans->transid);
2810                         btrfs_mark_buffer_dirty(upper->eb);
2811
2812                         ret = btrfs_inc_extent_ref(trans, root->fs_info,
2813                                                 node->eb->start, blocksize,
2814                                                 upper->eb->start,
2815                                                 btrfs_header_owner(upper->eb),
2816                                                 node->level, 0);
2817                         BUG_ON(ret);
2818
2819                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2820                         BUG_ON(ret);
2821                 }
2822 next:
2823                 if (!upper->pending)
2824                         drop_node_buffer(upper);
2825                 else
2826                         unlock_node_buffer(upper);
2827                 if (err)
2828                         break;
2829         }
2830
2831         if (!err && node->pending) {
2832                 drop_node_buffer(node);
2833                 list_move_tail(&node->list, &rc->backref_cache.changed);
2834                 node->pending = 0;
2835         }
2836
2837         path->lowest_level = 0;
2838         BUG_ON(err == -ENOSPC);
2839         return err;
2840 }
2841
2842 static int link_to_upper(struct btrfs_trans_handle *trans,
2843                          struct reloc_control *rc,
2844                          struct backref_node *node,
2845                          struct btrfs_path *path)
2846 {
2847         struct btrfs_key key;
2848
2849         btrfs_node_key_to_cpu(node->eb, &key, 0);
2850         return do_relocation(trans, rc, node, &key, path, 0);
2851 }
2852
2853 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2854                                 struct reloc_control *rc,
2855                                 struct btrfs_path *path, int err)
2856 {
2857         LIST_HEAD(list);
2858         struct backref_cache *cache = &rc->backref_cache;
2859         struct backref_node *node;
2860         int level;
2861         int ret;
2862
2863         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2864                 while (!list_empty(&cache->pending[level])) {
2865                         node = list_entry(cache->pending[level].next,
2866                                           struct backref_node, list);
2867                         list_move_tail(&node->list, &list);
2868                         BUG_ON(!node->pending);
2869
2870                         if (!err) {
2871                                 ret = link_to_upper(trans, rc, node, path);
2872                                 if (ret < 0)
2873                                         err = ret;
2874                         }
2875                 }
2876                 list_splice_init(&list, &cache->pending[level]);
2877         }
2878         return err;
2879 }
2880
2881 static void mark_block_processed(struct reloc_control *rc,
2882                                  u64 bytenr, u32 blocksize)
2883 {
2884         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2885                         EXTENT_DIRTY);
2886 }
2887
2888 static void __mark_block_processed(struct reloc_control *rc,
2889                                    struct backref_node *node)
2890 {
2891         u32 blocksize;
2892         if (node->level == 0 ||
2893             in_block_group(node->bytenr, rc->block_group)) {
2894                 blocksize = rc->extent_root->fs_info->nodesize;
2895                 mark_block_processed(rc, node->bytenr, blocksize);
2896         }
2897         node->processed = 1;
2898 }
2899
2900 /*
2901  * mark a block and all blocks directly/indirectly reference the block
2902  * as processed.
2903  */
2904 static void update_processed_blocks(struct reloc_control *rc,
2905                                     struct backref_node *node)
2906 {
2907         struct backref_node *next = node;
2908         struct backref_edge *edge;
2909         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2910         int index = 0;
2911
2912         while (next) {
2913                 cond_resched();
2914                 while (1) {
2915                         if (next->processed)
2916                                 break;
2917
2918                         __mark_block_processed(rc, next);
2919
2920                         if (list_empty(&next->upper))
2921                                 break;
2922
2923                         edge = list_entry(next->upper.next,
2924                                           struct backref_edge, list[LOWER]);
2925                         edges[index++] = edge;
2926                         next = edge->node[UPPER];
2927                 }
2928                 next = walk_down_backref(edges, &index);
2929         }
2930 }
2931
2932 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2933 {
2934         u32 blocksize = rc->extent_root->fs_info->nodesize;
2935
2936         if (test_range_bit(&rc->processed_blocks, bytenr,
2937                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2938                 return 1;
2939         return 0;
2940 }
2941
2942 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2943                               struct tree_block *block)
2944 {
2945         struct extent_buffer *eb;
2946
2947         BUG_ON(block->key_ready);
2948         eb = read_tree_block(fs_info, block->bytenr, block->key.offset);
2949         if (IS_ERR(eb)) {
2950                 return PTR_ERR(eb);
2951         } else if (!extent_buffer_uptodate(eb)) {
2952                 free_extent_buffer(eb);
2953                 return -EIO;
2954         }
2955         WARN_ON(btrfs_header_level(eb) != block->level);
2956         if (block->level == 0)
2957                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2958         else
2959                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2960         free_extent_buffer(eb);
2961         block->key_ready = 1;
2962         return 0;
2963 }
2964
2965 /*
2966  * helper function to relocate a tree block
2967  */
2968 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2969                                 struct reloc_control *rc,
2970                                 struct backref_node *node,
2971                                 struct btrfs_key *key,
2972                                 struct btrfs_path *path)
2973 {
2974         struct btrfs_root *root;
2975         int ret = 0;
2976
2977         if (!node)
2978                 return 0;
2979
2980         BUG_ON(node->processed);
2981         root = select_one_root(node);
2982         if (root == ERR_PTR(-ENOENT)) {
2983                 update_processed_blocks(rc, node);
2984                 goto out;
2985         }
2986
2987         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2988                 ret = reserve_metadata_space(trans, rc, node);
2989                 if (ret)
2990                         goto out;
2991         }
2992
2993         if (root) {
2994                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2995                         BUG_ON(node->new_bytenr);
2996                         BUG_ON(!list_empty(&node->list));
2997                         btrfs_record_root_in_trans(trans, root);
2998                         root = root->reloc_root;
2999                         node->new_bytenr = root->node->start;
3000                         node->root = root;
3001                         list_add_tail(&node->list, &rc->backref_cache.changed);
3002                 } else {
3003                         path->lowest_level = node->level;
3004                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
3005                         btrfs_release_path(path);
3006                         if (ret > 0)
3007                                 ret = 0;
3008                 }
3009                 if (!ret)
3010                         update_processed_blocks(rc, node);
3011         } else {
3012                 ret = do_relocation(trans, rc, node, key, path, 1);
3013         }
3014 out:
3015         if (ret || node->level == 0 || node->cowonly)
3016                 remove_backref_node(&rc->backref_cache, node);
3017         return ret;
3018 }
3019
3020 /*
3021  * relocate a list of blocks
3022  */
3023 static noinline_for_stack
3024 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
3025                          struct reloc_control *rc, struct rb_root *blocks)
3026 {
3027         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3028         struct backref_node *node;
3029         struct btrfs_path *path;
3030         struct tree_block *block;
3031         struct rb_node *rb_node;
3032         int ret;
3033         int err = 0;
3034
3035         path = btrfs_alloc_path();
3036         if (!path) {
3037                 err = -ENOMEM;
3038                 goto out_free_blocks;
3039         }
3040
3041         rb_node = rb_first(blocks);
3042         while (rb_node) {
3043                 block = rb_entry(rb_node, struct tree_block, rb_node);
3044                 if (!block->key_ready)
3045                         readahead_tree_block(fs_info, block->bytenr);
3046                 rb_node = rb_next(rb_node);
3047         }
3048
3049         rb_node = rb_first(blocks);
3050         while (rb_node) {
3051                 block = rb_entry(rb_node, struct tree_block, rb_node);
3052                 if (!block->key_ready) {
3053                         err = get_tree_block_key(fs_info, block);
3054                         if (err)
3055                                 goto out_free_path;
3056                 }
3057                 rb_node = rb_next(rb_node);
3058         }
3059
3060         rb_node = rb_first(blocks);
3061         while (rb_node) {
3062                 block = rb_entry(rb_node, struct tree_block, rb_node);
3063
3064                 node = build_backref_tree(rc, &block->key,
3065                                           block->level, block->bytenr);
3066                 if (IS_ERR(node)) {
3067                         err = PTR_ERR(node);
3068                         goto out;
3069                 }
3070
3071                 ret = relocate_tree_block(trans, rc, node, &block->key,
3072                                           path);
3073                 if (ret < 0) {
3074                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3075                                 err = ret;
3076                         goto out;
3077                 }
3078                 rb_node = rb_next(rb_node);
3079         }
3080 out:
3081         err = finish_pending_nodes(trans, rc, path, err);
3082
3083 out_free_path:
3084         btrfs_free_path(path);
3085 out_free_blocks:
3086         free_block_list(blocks);
3087         return err;
3088 }
3089
3090 static noinline_for_stack
3091 int prealloc_file_extent_cluster(struct inode *inode,
3092                                  struct file_extent_cluster *cluster)
3093 {
3094         u64 alloc_hint = 0;
3095         u64 start;
3096         u64 end;
3097         u64 offset = BTRFS_I(inode)->index_cnt;
3098         u64 num_bytes;
3099         int nr = 0;
3100         int ret = 0;
3101         u64 prealloc_start = cluster->start - offset;
3102         u64 prealloc_end = cluster->end - offset;
3103         u64 cur_offset;
3104         struct extent_changeset *data_reserved = NULL;
3105
3106         BUG_ON(cluster->start != cluster->boundary[0]);
3107         inode_lock(inode);
3108
3109         ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
3110                                           prealloc_end + 1 - prealloc_start);
3111         if (ret)
3112                 goto out;
3113
3114         cur_offset = prealloc_start;
3115         while (nr < cluster->nr) {
3116                 start = cluster->boundary[nr] - offset;
3117                 if (nr + 1 < cluster->nr)
3118                         end = cluster->boundary[nr + 1] - 1 - offset;
3119                 else
3120                         end = cluster->end - offset;
3121
3122                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3123                 num_bytes = end + 1 - start;
3124                 if (cur_offset < start)
3125                         btrfs_free_reserved_data_space(inode, data_reserved,
3126                                         cur_offset, start - cur_offset);
3127                 ret = btrfs_prealloc_file_range(inode, 0, start,
3128                                                 num_bytes, num_bytes,
3129                                                 end + 1, &alloc_hint);
3130                 cur_offset = end + 1;
3131                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3132                 if (ret)
3133                         break;
3134                 nr++;
3135         }
3136         if (cur_offset < prealloc_end)
3137                 btrfs_free_reserved_data_space(inode, data_reserved,
3138                                 cur_offset, prealloc_end + 1 - cur_offset);
3139 out:
3140         inode_unlock(inode);
3141         extent_changeset_free(data_reserved);
3142         return ret;
3143 }
3144
3145 static noinline_for_stack
3146 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3147                          u64 block_start)
3148 {
3149         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3150         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3151         struct extent_map *em;
3152         int ret = 0;
3153
3154         em = alloc_extent_map();
3155         if (!em)
3156                 return -ENOMEM;
3157
3158         em->start = start;
3159         em->len = end + 1 - start;
3160         em->block_len = em->len;
3161         em->block_start = block_start;
3162         em->bdev = fs_info->fs_devices->latest_bdev;
3163         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3164
3165         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3166         while (1) {
3167                 write_lock(&em_tree->lock);
3168                 ret = add_extent_mapping(em_tree, em, 0);
3169                 write_unlock(&em_tree->lock);
3170                 if (ret != -EEXIST) {
3171                         free_extent_map(em);
3172                         break;
3173                 }
3174                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3175         }
3176         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3177         return ret;
3178 }
3179
3180 static int relocate_file_extent_cluster(struct inode *inode,
3181                                         struct file_extent_cluster *cluster)
3182 {
3183         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3184         u64 page_start;
3185         u64 page_end;
3186         u64 offset = BTRFS_I(inode)->index_cnt;
3187         unsigned long index;
3188         unsigned long last_index;
3189         struct page *page;
3190         struct file_ra_state *ra;
3191         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3192         int nr = 0;
3193         int ret = 0;
3194
3195         if (!cluster->nr)
3196                 return 0;
3197
3198         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3199         if (!ra)
3200                 return -ENOMEM;
3201
3202         ret = prealloc_file_extent_cluster(inode, cluster);
3203         if (ret)
3204                 goto out;
3205
3206         file_ra_state_init(ra, inode->i_mapping);
3207
3208         ret = setup_extent_mapping(inode, cluster->start - offset,
3209                                    cluster->end - offset, cluster->start);
3210         if (ret)
3211                 goto out;
3212
3213         index = (cluster->start - offset) >> PAGE_SHIFT;
3214         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3215         while (index <= last_index) {
3216                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3217                                 PAGE_SIZE);
3218                 if (ret)
3219                         goto out;
3220
3221                 page = find_lock_page(inode->i_mapping, index);
3222                 if (!page) {
3223                         page_cache_sync_readahead(inode->i_mapping,
3224                                                   ra, NULL, index,
3225                                                   last_index + 1 - index);
3226                         page = find_or_create_page(inode->i_mapping, index,
3227                                                    mask);
3228                         if (!page) {
3229                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3230                                                         PAGE_SIZE);
3231                                 ret = -ENOMEM;
3232                                 goto out;
3233                         }
3234                 }
3235
3236                 if (PageReadahead(page)) {
3237                         page_cache_async_readahead(inode->i_mapping,
3238                                                    ra, NULL, page, index,
3239                                                    last_index + 1 - index);
3240                 }
3241
3242                 if (!PageUptodate(page)) {
3243                         btrfs_readpage(NULL, page);
3244                         lock_page(page);
3245                         if (!PageUptodate(page)) {
3246                                 unlock_page(page);
3247                                 put_page(page);
3248                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3249                                                         PAGE_SIZE);
3250                                 ret = -EIO;
3251                                 goto out;
3252                         }
3253                 }
3254
3255                 page_start = page_offset(page);
3256                 page_end = page_start + PAGE_SIZE - 1;
3257
3258                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3259
3260                 set_page_extent_mapped(page);
3261
3262                 if (nr < cluster->nr &&
3263                     page_start + offset == cluster->boundary[nr]) {
3264                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3265                                         page_start, page_end,
3266                                         EXTENT_BOUNDARY);
3267                         nr++;
3268                 }
3269
3270                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL, 0);
3271                 set_page_dirty(page);
3272
3273                 unlock_extent(&BTRFS_I(inode)->io_tree,
3274                               page_start, page_end);
3275                 unlock_page(page);
3276                 put_page(page);
3277
3278                 index++;
3279                 balance_dirty_pages_ratelimited(inode->i_mapping);
3280                 btrfs_throttle(fs_info);
3281         }
3282         WARN_ON(nr != cluster->nr);
3283 out:
3284         kfree(ra);
3285         return ret;
3286 }
3287
3288 static noinline_for_stack
3289 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3290                          struct file_extent_cluster *cluster)
3291 {
3292         int ret;
3293
3294         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3295                 ret = relocate_file_extent_cluster(inode, cluster);
3296                 if (ret)
3297                         return ret;
3298                 cluster->nr = 0;
3299         }
3300
3301         if (!cluster->nr)
3302                 cluster->start = extent_key->objectid;
3303         else
3304                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3305         cluster->end = extent_key->objectid + extent_key->offset - 1;
3306         cluster->boundary[cluster->nr] = extent_key->objectid;
3307         cluster->nr++;
3308
3309         if (cluster->nr >= MAX_EXTENTS) {
3310                 ret = relocate_file_extent_cluster(inode, cluster);
3311                 if (ret)
3312                         return ret;
3313                 cluster->nr = 0;
3314         }
3315         return 0;
3316 }
3317
3318 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3319 static int get_ref_objectid_v0(struct reloc_control *rc,
3320                                struct btrfs_path *path,
3321                                struct btrfs_key *extent_key,
3322                                u64 *ref_objectid, int *path_change)
3323 {
3324         struct btrfs_key key;
3325         struct extent_buffer *leaf;
3326         struct btrfs_extent_ref_v0 *ref0;
3327         int ret;
3328         int slot;
3329
3330         leaf = path->nodes[0];
3331         slot = path->slots[0];
3332         while (1) {
3333                 if (slot >= btrfs_header_nritems(leaf)) {
3334                         ret = btrfs_next_leaf(rc->extent_root, path);
3335                         if (ret < 0)
3336                                 return ret;
3337                         BUG_ON(ret > 0);
3338                         leaf = path->nodes[0];
3339                         slot = path->slots[0];
3340                         if (path_change)
3341                                 *path_change = 1;
3342                 }
3343                 btrfs_item_key_to_cpu(leaf, &key, slot);
3344                 if (key.objectid != extent_key->objectid)
3345                         return -ENOENT;
3346
3347                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3348                         slot++;
3349                         continue;
3350                 }
3351                 ref0 = btrfs_item_ptr(leaf, slot,
3352                                 struct btrfs_extent_ref_v0);
3353                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3354                 break;
3355         }
3356         return 0;
3357 }
3358 #endif
3359
3360 /*
3361  * helper to add a tree block to the list.
3362  * the major work is getting the generation and level of the block
3363  */
3364 static int add_tree_block(struct reloc_control *rc,
3365                           struct btrfs_key *extent_key,
3366                           struct btrfs_path *path,
3367                           struct rb_root *blocks)
3368 {
3369         struct extent_buffer *eb;
3370         struct btrfs_extent_item *ei;
3371         struct btrfs_tree_block_info *bi;
3372         struct tree_block *block;
3373         struct rb_node *rb_node;
3374         u32 item_size;
3375         int level = -1;
3376         u64 generation;
3377
3378         eb =  path->nodes[0];
3379         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3380
3381         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3382             item_size >= sizeof(*ei) + sizeof(*bi)) {
3383                 ei = btrfs_item_ptr(eb, path->slots[0],
3384                                 struct btrfs_extent_item);
3385                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3386                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3387                         level = btrfs_tree_block_level(eb, bi);
3388                 } else {
3389                         level = (int)extent_key->offset;
3390                 }
3391                 generation = btrfs_extent_generation(eb, ei);
3392         } else {
3393 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3394                 u64 ref_owner;
3395                 int ret;
3396
3397                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3398                 ret = get_ref_objectid_v0(rc, path, extent_key,
3399                                           &ref_owner, NULL);
3400                 if (ret < 0)
3401                         return ret;
3402                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3403                 level = (int)ref_owner;
3404                 /* FIXME: get real generation */
3405                 generation = 0;
3406 #else
3407                 BUG();
3408 #endif
3409         }
3410
3411         btrfs_release_path(path);
3412
3413         BUG_ON(level == -1);
3414
3415         block = kmalloc(sizeof(*block), GFP_NOFS);
3416         if (!block)
3417                 return -ENOMEM;
3418
3419         block->bytenr = extent_key->objectid;
3420         block->key.objectid = rc->extent_root->fs_info->nodesize;
3421         block->key.offset = generation;
3422         block->level = level;
3423         block->key_ready = 0;
3424
3425         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3426         if (rb_node)
3427                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3428
3429         return 0;
3430 }
3431
3432 /*
3433  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3434  */
3435 static int __add_tree_block(struct reloc_control *rc,
3436                             u64 bytenr, u32 blocksize,
3437                             struct rb_root *blocks)
3438 {
3439         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3440         struct btrfs_path *path;
3441         struct btrfs_key key;
3442         int ret;
3443         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3444
3445         if (tree_block_processed(bytenr, rc))
3446                 return 0;
3447
3448         if (tree_search(blocks, bytenr))
3449                 return 0;
3450
3451         path = btrfs_alloc_path();
3452         if (!path)
3453                 return -ENOMEM;
3454 again:
3455         key.objectid = bytenr;
3456         if (skinny) {
3457                 key.type = BTRFS_METADATA_ITEM_KEY;
3458                 key.offset = (u64)-1;
3459         } else {
3460                 key.type = BTRFS_EXTENT_ITEM_KEY;
3461                 key.offset = blocksize;
3462         }
3463
3464         path->search_commit_root = 1;
3465         path->skip_locking = 1;
3466         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3467         if (ret < 0)
3468                 goto out;
3469
3470         if (ret > 0 && skinny) {
3471                 if (path->slots[0]) {
3472                         path->slots[0]--;
3473                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3474                                               path->slots[0]);
3475                         if (key.objectid == bytenr &&
3476                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3477                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3478                               key.offset == blocksize)))
3479                                 ret = 0;
3480                 }
3481
3482                 if (ret) {
3483                         skinny = false;
3484                         btrfs_release_path(path);
3485                         goto again;
3486                 }
3487         }
3488         if (ret) {
3489                 ASSERT(ret == 1);
3490                 btrfs_print_leaf(path->nodes[0]);
3491                 btrfs_err(fs_info,
3492              "tree block extent item (%llu) is not found in extent tree",
3493                      bytenr);
3494                 WARN_ON(1);
3495                 ret = -EINVAL;
3496                 goto out;
3497         }
3498
3499         ret = add_tree_block(rc, &key, path, blocks);
3500 out:
3501         btrfs_free_path(path);
3502         return ret;
3503 }
3504
3505 /*
3506  * helper to check if the block use full backrefs for pointers in it
3507  */
3508 static int block_use_full_backref(struct reloc_control *rc,
3509                                   struct extent_buffer *eb)
3510 {
3511         u64 flags;
3512         int ret;
3513
3514         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3515             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3516                 return 1;
3517
3518         ret = btrfs_lookup_extent_info(NULL, rc->extent_root->fs_info,
3519                                        eb->start, btrfs_header_level(eb), 1,
3520                                        NULL, &flags);
3521         BUG_ON(ret);
3522
3523         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3524                 ret = 1;
3525         else
3526                 ret = 0;
3527         return ret;
3528 }
3529
3530 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3531                                     struct btrfs_block_group_cache *block_group,
3532                                     struct inode *inode,
3533                                     u64 ino)
3534 {
3535         struct btrfs_key key;
3536         struct btrfs_root *root = fs_info->tree_root;
3537         struct btrfs_trans_handle *trans;
3538         int ret = 0;
3539
3540         if (inode)
3541                 goto truncate;
3542
3543         key.objectid = ino;
3544         key.type = BTRFS_INODE_ITEM_KEY;
3545         key.offset = 0;
3546
3547         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3548         if (IS_ERR(inode) || is_bad_inode(inode)) {
3549                 if (!IS_ERR(inode))
3550                         iput(inode);
3551                 return -ENOENT;
3552         }
3553
3554 truncate:
3555         ret = btrfs_check_trunc_cache_free_space(fs_info,
3556                                                  &fs_info->global_block_rsv);
3557         if (ret)
3558                 goto out;
3559
3560         trans = btrfs_join_transaction(root);
3561         if (IS_ERR(trans)) {
3562                 ret = PTR_ERR(trans);
3563                 goto out;
3564         }
3565
3566         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3567
3568         btrfs_end_transaction(trans);
3569         btrfs_btree_balance_dirty(fs_info);
3570 out:
3571         iput(inode);
3572         return ret;
3573 }
3574
3575 /*
3576  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3577  * this function scans fs tree to find blocks reference the data extent
3578  */
3579 static int find_data_references(struct reloc_control *rc,
3580                                 struct btrfs_key *extent_key,
3581                                 struct extent_buffer *leaf,
3582                                 struct btrfs_extent_data_ref *ref,
3583                                 struct rb_root *blocks)
3584 {
3585         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3586         struct btrfs_path *path;
3587         struct tree_block *block;
3588         struct btrfs_root *root;
3589         struct btrfs_file_extent_item *fi;
3590         struct rb_node *rb_node;
3591         struct btrfs_key key;
3592         u64 ref_root;
3593         u64 ref_objectid;
3594         u64 ref_offset;
3595         u32 ref_count;
3596         u32 nritems;
3597         int err = 0;
3598         int added = 0;
3599         int counted;
3600         int ret;
3601
3602         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3603         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3604         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3605         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3606
3607         /*
3608          * This is an extent belonging to the free space cache, lets just delete
3609          * it and redo the search.
3610          */
3611         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3612                 ret = delete_block_group_cache(fs_info, rc->block_group,
3613                                                NULL, ref_objectid);
3614                 if (ret != -ENOENT)
3615                         return ret;
3616                 ret = 0;
3617         }
3618
3619         path = btrfs_alloc_path();
3620         if (!path)
3621                 return -ENOMEM;
3622         path->reada = READA_FORWARD;
3623
3624         root = read_fs_root(fs_info, ref_root);
3625         if (IS_ERR(root)) {
3626                 err = PTR_ERR(root);
3627                 goto out;
3628         }
3629
3630         key.objectid = ref_objectid;
3631         key.type = BTRFS_EXTENT_DATA_KEY;
3632         if (ref_offset > ((u64)-1 << 32))
3633                 key.offset = 0;
3634         else
3635                 key.offset = ref_offset;
3636
3637         path->search_commit_root = 1;
3638         path->skip_locking = 1;
3639         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3640         if (ret < 0) {
3641                 err = ret;
3642                 goto out;
3643         }
3644
3645         leaf = path->nodes[0];
3646         nritems = btrfs_header_nritems(leaf);
3647         /*
3648          * the references in tree blocks that use full backrefs
3649          * are not counted in
3650          */
3651         if (block_use_full_backref(rc, leaf))
3652                 counted = 0;
3653         else
3654                 counted = 1;
3655         rb_node = tree_search(blocks, leaf->start);
3656         if (rb_node) {
3657                 if (counted)
3658                         added = 1;
3659                 else
3660                         path->slots[0] = nritems;
3661         }
3662
3663         while (ref_count > 0) {
3664                 while (path->slots[0] >= nritems) {
3665                         ret = btrfs_next_leaf(root, path);
3666                         if (ret < 0) {
3667                                 err = ret;
3668                                 goto out;
3669                         }
3670                         if (WARN_ON(ret > 0))
3671                                 goto out;
3672
3673                         leaf = path->nodes[0];
3674                         nritems = btrfs_header_nritems(leaf);
3675                         added = 0;
3676
3677                         if (block_use_full_backref(rc, leaf))
3678                                 counted = 0;
3679                         else
3680                                 counted = 1;
3681                         rb_node = tree_search(blocks, leaf->start);
3682                         if (rb_node) {
3683                                 if (counted)
3684                                         added = 1;
3685                                 else
3686                                         path->slots[0] = nritems;
3687                         }
3688                 }
3689
3690                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3691                 if (WARN_ON(key.objectid != ref_objectid ||
3692                     key.type != BTRFS_EXTENT_DATA_KEY))
3693                         break;
3694
3695                 fi = btrfs_item_ptr(leaf, path->slots[0],
3696                                     struct btrfs_file_extent_item);
3697
3698                 if (btrfs_file_extent_type(leaf, fi) ==
3699                     BTRFS_FILE_EXTENT_INLINE)
3700                         goto next;
3701
3702                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3703                     extent_key->objectid)
3704                         goto next;
3705
3706                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3707                 if (key.offset != ref_offset)
3708                         goto next;
3709
3710                 if (counted)
3711                         ref_count--;
3712                 if (added)
3713                         goto next;
3714
3715                 if (!tree_block_processed(leaf->start, rc)) {
3716                         block = kmalloc(sizeof(*block), GFP_NOFS);
3717                         if (!block) {
3718                                 err = -ENOMEM;
3719                                 break;
3720                         }
3721                         block->bytenr = leaf->start;
3722                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3723                         block->level = 0;
3724                         block->key_ready = 1;
3725                         rb_node = tree_insert(blocks, block->bytenr,
3726                                               &block->rb_node);
3727                         if (rb_node)
3728                                 backref_tree_panic(rb_node, -EEXIST,
3729                                                    block->bytenr);
3730                 }
3731                 if (counted)
3732                         added = 1;
3733                 else
3734                         path->slots[0] = nritems;
3735 next:
3736                 path->slots[0]++;
3737
3738         }
3739 out:
3740         btrfs_free_path(path);
3741         return err;
3742 }
3743
3744 /*
3745  * helper to find all tree blocks that reference a given data extent
3746  */
3747 static noinline_for_stack
3748 int add_data_references(struct reloc_control *rc,
3749                         struct btrfs_key *extent_key,
3750                         struct btrfs_path *path,
3751                         struct rb_root *blocks)
3752 {
3753         struct btrfs_key key;
3754         struct extent_buffer *eb;
3755         struct btrfs_extent_data_ref *dref;
3756         struct btrfs_extent_inline_ref *iref;
3757         unsigned long ptr;
3758         unsigned long end;
3759         u32 blocksize = rc->extent_root->fs_info->nodesize;
3760         int ret = 0;
3761         int err = 0;
3762
3763         eb = path->nodes[0];
3764         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3765         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3766 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3767         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3768                 ptr = end;
3769         else
3770 #endif
3771                 ptr += sizeof(struct btrfs_extent_item);
3772
3773         while (ptr < end) {
3774                 iref = (struct btrfs_extent_inline_ref *)ptr;
3775                 key.type = btrfs_get_extent_inline_ref_type(eb, iref,
3776                                                         BTRFS_REF_TYPE_DATA);
3777                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3778                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3779                         ret = __add_tree_block(rc, key.offset, blocksize,
3780                                                blocks);
3781                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3782                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3783                         ret = find_data_references(rc, extent_key,
3784                                                    eb, dref, blocks);
3785                 } else {
3786                         ret = -EINVAL;
3787                         btrfs_err(rc->extent_root->fs_info,
3788                      "extent %llu slot %d has an invalid inline ref type",
3789                              eb->start, path->slots[0]);
3790                 }
3791                 if (ret) {
3792                         err = ret;
3793                         goto out;
3794                 }
3795                 ptr += btrfs_extent_inline_ref_size(key.type);
3796         }
3797         WARN_ON(ptr > end);
3798
3799         while (1) {
3800                 cond_resched();
3801                 eb = path->nodes[0];
3802                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3803                         ret = btrfs_next_leaf(rc->extent_root, path);
3804                         if (ret < 0) {
3805                                 err = ret;
3806                                 break;
3807                         }
3808                         if (ret > 0)
3809                                 break;
3810                         eb = path->nodes[0];
3811                 }
3812
3813                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3814                 if (key.objectid != extent_key->objectid)
3815                         break;
3816
3817 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3818                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3819                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3820 #else
3821                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3822                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3823 #endif
3824                         ret = __add_tree_block(rc, key.offset, blocksize,
3825                                                blocks);
3826                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3827                         dref = btrfs_item_ptr(eb, path->slots[0],
3828                                               struct btrfs_extent_data_ref);
3829                         ret = find_data_references(rc, extent_key,
3830                                                    eb, dref, blocks);
3831                 } else {
3832                         ret = 0;
3833                 }
3834                 if (ret) {
3835                         err = ret;
3836                         break;
3837                 }
3838                 path->slots[0]++;
3839         }
3840 out:
3841         btrfs_release_path(path);
3842         if (err)
3843                 free_block_list(blocks);
3844         return err;
3845 }
3846
3847 /*
3848  * helper to find next unprocessed extent
3849  */
3850 static noinline_for_stack
3851 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3852                      struct btrfs_key *extent_key)
3853 {
3854         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3855         struct btrfs_key key;
3856         struct extent_buffer *leaf;
3857         u64 start, end, last;
3858         int ret;
3859
3860         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3861         while (1) {
3862                 cond_resched();
3863                 if (rc->search_start >= last) {
3864                         ret = 1;
3865                         break;
3866                 }
3867
3868                 key.objectid = rc->search_start;
3869                 key.type = BTRFS_EXTENT_ITEM_KEY;
3870                 key.offset = 0;
3871
3872                 path->search_commit_root = 1;
3873                 path->skip_locking = 1;
3874                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3875                                         0, 0);
3876                 if (ret < 0)
3877                         break;
3878 next:
3879                 leaf = path->nodes[0];
3880                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3881                         ret = btrfs_next_leaf(rc->extent_root, path);
3882                         if (ret != 0)
3883                                 break;
3884                         leaf = path->nodes[0];
3885                 }
3886
3887                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3888                 if (key.objectid >= last) {
3889                         ret = 1;
3890                         break;
3891                 }
3892
3893                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3894                     key.type != BTRFS_METADATA_ITEM_KEY) {
3895                         path->slots[0]++;
3896                         goto next;
3897                 }
3898
3899                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3900                     key.objectid + key.offset <= rc->search_start) {
3901                         path->slots[0]++;
3902                         goto next;
3903                 }
3904
3905                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3906                     key.objectid + fs_info->nodesize <=
3907                     rc->search_start) {
3908                         path->slots[0]++;
3909                         goto next;
3910                 }
3911
3912                 ret = find_first_extent_bit(&rc->processed_blocks,
3913                                             key.objectid, &start, &end,
3914                                             EXTENT_DIRTY, NULL);
3915
3916                 if (ret == 0 && start <= key.objectid) {
3917                         btrfs_release_path(path);
3918                         rc->search_start = end + 1;
3919                 } else {
3920                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3921                                 rc->search_start = key.objectid + key.offset;
3922                         else
3923                                 rc->search_start = key.objectid +
3924                                         fs_info->nodesize;
3925                         memcpy(extent_key, &key, sizeof(key));
3926                         return 0;
3927                 }
3928         }
3929         btrfs_release_path(path);
3930         return ret;
3931 }
3932
3933 static void set_reloc_control(struct reloc_control *rc)
3934 {
3935         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3936
3937         mutex_lock(&fs_info->reloc_mutex);
3938         fs_info->reloc_ctl = rc;
3939         mutex_unlock(&fs_info->reloc_mutex);
3940 }
3941
3942 static void unset_reloc_control(struct reloc_control *rc)
3943 {
3944         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3945
3946         mutex_lock(&fs_info->reloc_mutex);
3947         fs_info->reloc_ctl = NULL;
3948         mutex_unlock(&fs_info->reloc_mutex);
3949 }
3950
3951 static int check_extent_flags(u64 flags)
3952 {
3953         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3954             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3955                 return 1;
3956         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3957             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3958                 return 1;
3959         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3960             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3961                 return 1;
3962         return 0;
3963 }
3964
3965 static noinline_for_stack
3966 int prepare_to_relocate(struct reloc_control *rc)
3967 {
3968         struct btrfs_trans_handle *trans;
3969         int ret;
3970
3971         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3972                                               BTRFS_BLOCK_RSV_TEMP);
3973         if (!rc->block_rsv)
3974                 return -ENOMEM;
3975
3976         memset(&rc->cluster, 0, sizeof(rc->cluster));
3977         rc->search_start = rc->block_group->key.objectid;
3978         rc->extents_found = 0;
3979         rc->nodes_relocated = 0;
3980         rc->merging_rsv_size = 0;
3981         rc->reserved_bytes = 0;
3982         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3983                               RELOCATION_RESERVED_NODES;
3984         ret = btrfs_block_rsv_refill(rc->extent_root,
3985                                      rc->block_rsv, rc->block_rsv->size,
3986                                      BTRFS_RESERVE_FLUSH_ALL);
3987         if (ret)
3988                 return ret;
3989
3990         rc->create_reloc_tree = 1;
3991         set_reloc_control(rc);
3992
3993         trans = btrfs_join_transaction(rc->extent_root);
3994         if (IS_ERR(trans)) {
3995                 unset_reloc_control(rc);
3996                 /*
3997                  * extent tree is not a ref_cow tree and has no reloc_root to
3998                  * cleanup.  And callers are responsible to free the above
3999                  * block rsv.
4000                  */
4001                 return PTR_ERR(trans);
4002         }
4003         btrfs_commit_transaction(trans);
4004         return 0;
4005 }
4006
4007 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
4008 {
4009         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
4010         struct rb_root blocks = RB_ROOT;
4011         struct btrfs_key key;
4012         struct btrfs_trans_handle *trans = NULL;
4013         struct btrfs_path *path;
4014         struct btrfs_extent_item *ei;
4015         u64 flags;
4016         u32 item_size;
4017         int ret;
4018         int err = 0;
4019         int progress = 0;
4020
4021         path = btrfs_alloc_path();
4022         if (!path)
4023                 return -ENOMEM;
4024         path->reada = READA_FORWARD;
4025
4026         ret = prepare_to_relocate(rc);
4027         if (ret) {
4028                 err = ret;
4029                 goto out_free;
4030         }
4031
4032         while (1) {
4033                 rc->reserved_bytes = 0;
4034                 ret = btrfs_block_rsv_refill(rc->extent_root,
4035                                         rc->block_rsv, rc->block_rsv->size,
4036                                         BTRFS_RESERVE_FLUSH_ALL);
4037                 if (ret) {
4038                         err = ret;
4039                         break;
4040                 }
4041                 progress++;
4042                 trans = btrfs_start_transaction(rc->extent_root, 0);
4043                 if (IS_ERR(trans)) {
4044                         err = PTR_ERR(trans);
4045                         trans = NULL;
4046                         break;
4047                 }
4048 restart:
4049                 if (update_backref_cache(trans, &rc->backref_cache)) {
4050                         btrfs_end_transaction(trans);
4051                         trans = NULL;
4052                         continue;
4053                 }
4054
4055                 ret = find_next_extent(rc, path, &key);
4056                 if (ret < 0)
4057                         err = ret;
4058                 if (ret != 0)
4059                         break;
4060
4061                 rc->extents_found++;
4062
4063                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4064                                     struct btrfs_extent_item);
4065                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4066                 if (item_size >= sizeof(*ei)) {
4067                         flags = btrfs_extent_flags(path->nodes[0], ei);
4068                         ret = check_extent_flags(flags);
4069                         BUG_ON(ret);
4070
4071                 } else {
4072 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4073                         u64 ref_owner;
4074                         int path_change = 0;
4075
4076                         BUG_ON(item_size !=
4077                                sizeof(struct btrfs_extent_item_v0));
4078                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
4079                                                   &path_change);
4080                         if (ret < 0) {
4081                                 err = ret;
4082                                 break;
4083                         }
4084                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
4085                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
4086                         else
4087                                 flags = BTRFS_EXTENT_FLAG_DATA;
4088
4089                         if (path_change) {
4090                                 btrfs_release_path(path);
4091
4092                                 path->search_commit_root = 1;
4093                                 path->skip_locking = 1;
4094                                 ret = btrfs_search_slot(NULL, rc->extent_root,
4095                                                         &key, path, 0, 0);
4096                                 if (ret < 0) {
4097                                         err = ret;
4098                                         break;
4099                                 }
4100                                 BUG_ON(ret > 0);
4101                         }
4102 #else
4103                         BUG();
4104 #endif
4105                 }
4106
4107                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4108                         ret = add_tree_block(rc, &key, path, &blocks);
4109                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4110                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4111                         ret = add_data_references(rc, &key, path, &blocks);
4112                 } else {
4113                         btrfs_release_path(path);
4114                         ret = 0;
4115                 }
4116                 if (ret < 0) {
4117                         err = ret;
4118                         break;
4119                 }
4120
4121                 if (!RB_EMPTY_ROOT(&blocks)) {
4122                         ret = relocate_tree_blocks(trans, rc, &blocks);
4123                         if (ret < 0) {
4124                                 /*
4125                                  * if we fail to relocate tree blocks, force to update
4126                                  * backref cache when committing transaction.
4127                                  */
4128                                 rc->backref_cache.last_trans = trans->transid - 1;
4129
4130                                 if (ret != -EAGAIN) {
4131                                         err = ret;
4132                                         break;
4133                                 }
4134                                 rc->extents_found--;
4135                                 rc->search_start = key.objectid;
4136                         }
4137                 }
4138
4139                 btrfs_end_transaction_throttle(trans);
4140                 btrfs_btree_balance_dirty(fs_info);
4141                 trans = NULL;
4142
4143                 if (rc->stage == MOVE_DATA_EXTENTS &&
4144                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4145                         rc->found_file_extent = 1;
4146                         ret = relocate_data_extent(rc->data_inode,
4147                                                    &key, &rc->cluster);
4148                         if (ret < 0) {
4149                                 err = ret;
4150                                 break;
4151                         }
4152                 }
4153         }
4154         if (trans && progress && err == -ENOSPC) {
4155                 ret = btrfs_force_chunk_alloc(trans, fs_info,
4156                                               rc->block_group->flags);
4157                 if (ret == 1) {
4158                         err = 0;
4159                         progress = 0;
4160                         goto restart;
4161                 }
4162         }
4163
4164         btrfs_release_path(path);
4165         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4166
4167         if (trans) {
4168                 btrfs_end_transaction_throttle(trans);
4169                 btrfs_btree_balance_dirty(fs_info);
4170         }
4171
4172         if (!err) {
4173                 ret = relocate_file_extent_cluster(rc->data_inode,
4174                                                    &rc->cluster);
4175                 if (ret < 0)
4176                         err = ret;
4177         }
4178
4179         rc->create_reloc_tree = 0;
4180         set_reloc_control(rc);
4181
4182         backref_cache_cleanup(&rc->backref_cache);
4183         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4184
4185         err = prepare_to_merge(rc, err);
4186
4187         merge_reloc_roots(rc);
4188
4189         rc->merge_reloc_tree = 0;
4190         unset_reloc_control(rc);
4191         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4192
4193         /* get rid of pinned extents */
4194         trans = btrfs_join_transaction(rc->extent_root);
4195         if (IS_ERR(trans)) {
4196                 err = PTR_ERR(trans);
4197                 goto out_free;
4198         }
4199         btrfs_commit_transaction(trans);
4200 out_free:
4201         btrfs_free_block_rsv(fs_info, rc->block_rsv);
4202         btrfs_free_path(path);
4203         return err;
4204 }
4205
4206 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4207                                  struct btrfs_root *root, u64 objectid)
4208 {
4209         struct btrfs_path *path;
4210         struct btrfs_inode_item *item;
4211         struct extent_buffer *leaf;
4212         int ret;
4213
4214         path = btrfs_alloc_path();
4215         if (!path)
4216                 return -ENOMEM;
4217
4218         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4219         if (ret)
4220                 goto out;
4221
4222         leaf = path->nodes[0];
4223         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4224         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
4225         btrfs_set_inode_generation(leaf, item, 1);
4226         btrfs_set_inode_size(leaf, item, 0);
4227         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4228         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4229                                           BTRFS_INODE_PREALLOC);
4230         btrfs_mark_buffer_dirty(leaf);
4231 out:
4232         btrfs_free_path(path);
4233         return ret;
4234 }
4235
4236 /*
4237  * helper to create inode for data relocation.
4238  * the inode is in data relocation tree and its link count is 0
4239  */
4240 static noinline_for_stack
4241 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4242                                  struct btrfs_block_group_cache *group)
4243 {
4244         struct inode *inode = NULL;
4245         struct btrfs_trans_handle *trans;
4246         struct btrfs_root *root;
4247         struct btrfs_key key;
4248         u64 objectid;
4249         int err = 0;
4250
4251         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4252         if (IS_ERR(root))
4253                 return ERR_CAST(root);
4254
4255         trans = btrfs_start_transaction(root, 6);
4256         if (IS_ERR(trans))
4257                 return ERR_CAST(trans);
4258
4259         err = btrfs_find_free_objectid(root, &objectid);
4260         if (err)
4261                 goto out;
4262
4263         err = __insert_orphan_inode(trans, root, objectid);
4264         BUG_ON(err);
4265
4266         key.objectid = objectid;
4267         key.type = BTRFS_INODE_ITEM_KEY;
4268         key.offset = 0;
4269         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4270         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4271         BTRFS_I(inode)->index_cnt = group->key.objectid;
4272
4273         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4274 out:
4275         btrfs_end_transaction(trans);
4276         btrfs_btree_balance_dirty(fs_info);
4277         if (err) {
4278                 if (inode)
4279                         iput(inode);
4280                 inode = ERR_PTR(err);
4281         }
4282         return inode;
4283 }
4284
4285 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4286 {
4287         struct reloc_control *rc;
4288
4289         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4290         if (!rc)
4291                 return NULL;
4292
4293         INIT_LIST_HEAD(&rc->reloc_roots);
4294         backref_cache_init(&rc->backref_cache);
4295         mapping_tree_init(&rc->reloc_root_tree);
4296         extent_io_tree_init(&rc->processed_blocks, NULL);
4297         return rc;
4298 }
4299
4300 /*
4301  * Print the block group being relocated
4302  */
4303 static void describe_relocation(struct btrfs_fs_info *fs_info,
4304                                 struct btrfs_block_group_cache *block_group)
4305 {
4306         char buf[128];          /* prefixed by a '|' that'll be dropped */
4307         u64 flags = block_group->flags;
4308
4309         /* Shouldn't happen */
4310         if (!flags) {
4311                 strcpy(buf, "|NONE");
4312         } else {
4313                 char *bp = buf;
4314
4315 #define DESCRIBE_FLAG(f, d) \
4316                 if (flags & BTRFS_BLOCK_GROUP_##f) { \
4317                         bp += snprintf(bp, buf - bp + sizeof(buf), "|%s", d); \
4318                         flags &= ~BTRFS_BLOCK_GROUP_##f; \
4319                 }
4320                 DESCRIBE_FLAG(DATA,     "data");
4321                 DESCRIBE_FLAG(SYSTEM,   "system");
4322                 DESCRIBE_FLAG(METADATA, "metadata");
4323                 DESCRIBE_FLAG(RAID0,    "raid0");
4324                 DESCRIBE_FLAG(RAID1,    "raid1");
4325                 DESCRIBE_FLAG(DUP,      "dup");
4326                 DESCRIBE_FLAG(RAID10,   "raid10");
4327                 DESCRIBE_FLAG(RAID5,    "raid5");
4328                 DESCRIBE_FLAG(RAID6,    "raid6");
4329                 if (flags)
4330                         snprintf(buf, buf - bp + sizeof(buf), "|0x%llx", flags);
4331 #undef DESCRIBE_FLAG
4332         }
4333
4334         btrfs_info(fs_info,
4335                    "relocating block group %llu flags %s",
4336                    block_group->key.objectid, buf + 1);
4337 }
4338
4339 /*
4340  * function to relocate all extents in a block group.
4341  */
4342 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4343 {
4344         struct btrfs_root *extent_root = fs_info->extent_root;
4345         struct reloc_control *rc;
4346         struct inode *inode;
4347         struct btrfs_path *path;
4348         int ret;
4349         int rw = 0;
4350         int err = 0;
4351
4352         rc = alloc_reloc_control(fs_info);
4353         if (!rc)
4354                 return -ENOMEM;
4355
4356         rc->extent_root = extent_root;
4357
4358         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4359         BUG_ON(!rc->block_group);
4360
4361         ret = btrfs_inc_block_group_ro(fs_info, rc->block_group);
4362         if (ret) {
4363                 err = ret;
4364                 goto out;
4365         }
4366         rw = 1;
4367
4368         path = btrfs_alloc_path();
4369         if (!path) {
4370                 err = -ENOMEM;
4371                 goto out;
4372         }
4373
4374         inode = lookup_free_space_inode(fs_info, rc->block_group, path);
4375         btrfs_free_path(path);
4376
4377         if (!IS_ERR(inode))
4378                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4379         else
4380                 ret = PTR_ERR(inode);
4381
4382         if (ret && ret != -ENOENT) {
4383                 err = ret;
4384                 goto out;
4385         }
4386
4387         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4388         if (IS_ERR(rc->data_inode)) {
4389                 err = PTR_ERR(rc->data_inode);
4390                 rc->data_inode = NULL;
4391                 goto out;
4392         }
4393
4394         describe_relocation(fs_info, rc->block_group);
4395
4396         btrfs_wait_block_group_reservations(rc->block_group);
4397         btrfs_wait_nocow_writers(rc->block_group);
4398         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4399                                  rc->block_group->key.objectid,
4400                                  rc->block_group->key.offset);
4401
4402         while (1) {
4403                 mutex_lock(&fs_info->cleaner_mutex);
4404                 ret = relocate_block_group(rc);
4405                 mutex_unlock(&fs_info->cleaner_mutex);
4406                 if (ret < 0) {
4407                         err = ret;
4408                         goto out;
4409                 }
4410
4411                 if (rc->extents_found == 0)
4412                         break;
4413
4414                 btrfs_info(fs_info, "found %llu extents", rc->extents_found);
4415
4416                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4417                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4418                                                        (u64)-1);
4419                         if (ret) {
4420                                 err = ret;
4421                                 goto out;
4422                         }
4423                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4424                                                  0, -1);
4425                         rc->stage = UPDATE_DATA_PTRS;
4426                 }
4427         }
4428
4429         WARN_ON(rc->block_group->pinned > 0);
4430         WARN_ON(rc->block_group->reserved > 0);
4431         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4432 out:
4433         if (err && rw)
4434                 btrfs_dec_block_group_ro(rc->block_group);
4435         iput(rc->data_inode);
4436         btrfs_put_block_group(rc->block_group);
4437         kfree(rc);
4438         return err;
4439 }
4440
4441 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4442 {
4443         struct btrfs_fs_info *fs_info = root->fs_info;
4444         struct btrfs_trans_handle *trans;
4445         int ret, err;
4446
4447         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4448         if (IS_ERR(trans))
4449                 return PTR_ERR(trans);
4450
4451         memset(&root->root_item.drop_progress, 0,
4452                 sizeof(root->root_item.drop_progress));
4453         root->root_item.drop_level = 0;
4454         btrfs_set_root_refs(&root->root_item, 0);
4455         ret = btrfs_update_root(trans, fs_info->tree_root,
4456                                 &root->root_key, &root->root_item);
4457
4458         err = btrfs_end_transaction(trans);
4459         if (err)
4460                 return err;
4461         return ret;
4462 }
4463
4464 /*
4465  * recover relocation interrupted by system crash.
4466  *
4467  * this function resumes merging reloc trees with corresponding fs trees.
4468  * this is important for keeping the sharing of tree blocks
4469  */
4470 int btrfs_recover_relocation(struct btrfs_root *root)
4471 {
4472         struct btrfs_fs_info *fs_info = root->fs_info;
4473         LIST_HEAD(reloc_roots);
4474         struct btrfs_key key;
4475         struct btrfs_root *fs_root;
4476         struct btrfs_root *reloc_root;
4477         struct btrfs_path *path;
4478         struct extent_buffer *leaf;
4479         struct reloc_control *rc = NULL;
4480         struct btrfs_trans_handle *trans;
4481         int ret;
4482         int err = 0;
4483
4484         path = btrfs_alloc_path();
4485         if (!path)
4486                 return -ENOMEM;
4487         path->reada = READA_BACK;
4488
4489         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4490         key.type = BTRFS_ROOT_ITEM_KEY;
4491         key.offset = (u64)-1;
4492
4493         while (1) {
4494                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4495                                         path, 0, 0);
4496                 if (ret < 0) {
4497                         err = ret;
4498                         goto out;
4499                 }
4500                 if (ret > 0) {
4501                         if (path->slots[0] == 0)
4502                                 break;
4503                         path->slots[0]--;
4504                 }
4505                 leaf = path->nodes[0];
4506                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4507                 btrfs_release_path(path);
4508
4509                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4510                     key.type != BTRFS_ROOT_ITEM_KEY)
4511                         break;
4512
4513                 reloc_root = btrfs_read_fs_root(root, &key);
4514                 if (IS_ERR(reloc_root)) {
4515                         err = PTR_ERR(reloc_root);
4516                         goto out;
4517                 }
4518
4519                 list_add(&reloc_root->root_list, &reloc_roots);
4520
4521                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4522                         fs_root = read_fs_root(fs_info,
4523                                                reloc_root->root_key.offset);
4524                         if (IS_ERR(fs_root)) {
4525                                 ret = PTR_ERR(fs_root);
4526                                 if (ret != -ENOENT) {
4527                                         err = ret;
4528                                         goto out;
4529                                 }
4530                                 ret = mark_garbage_root(reloc_root);
4531                                 if (ret < 0) {
4532                                         err = ret;
4533                                         goto out;
4534                                 }
4535                         }
4536                 }
4537
4538                 if (key.offset == 0)
4539                         break;
4540
4541                 key.offset--;
4542         }
4543         btrfs_release_path(path);
4544
4545         if (list_empty(&reloc_roots))
4546                 goto out;
4547
4548         rc = alloc_reloc_control(fs_info);
4549         if (!rc) {
4550                 err = -ENOMEM;
4551                 goto out;
4552         }
4553
4554         rc->extent_root = fs_info->extent_root;
4555
4556         set_reloc_control(rc);
4557
4558         trans = btrfs_join_transaction(rc->extent_root);
4559         if (IS_ERR(trans)) {
4560                 unset_reloc_control(rc);
4561                 err = PTR_ERR(trans);
4562                 goto out_free;
4563         }
4564
4565         rc->merge_reloc_tree = 1;
4566
4567         while (!list_empty(&reloc_roots)) {
4568                 reloc_root = list_entry(reloc_roots.next,
4569                                         struct btrfs_root, root_list);
4570                 list_del(&reloc_root->root_list);
4571
4572                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4573                         list_add_tail(&reloc_root->root_list,
4574                                       &rc->reloc_roots);
4575                         continue;
4576                 }
4577
4578                 fs_root = read_fs_root(fs_info, reloc_root->root_key.offset);
4579                 if (IS_ERR(fs_root)) {
4580                         err = PTR_ERR(fs_root);
4581                         goto out_free;
4582                 }
4583
4584                 err = __add_reloc_root(reloc_root);
4585                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4586                 fs_root->reloc_root = reloc_root;
4587         }
4588
4589         err = btrfs_commit_transaction(trans);
4590         if (err)
4591                 goto out_free;
4592
4593         merge_reloc_roots(rc);
4594
4595         unset_reloc_control(rc);
4596
4597         trans = btrfs_join_transaction(rc->extent_root);
4598         if (IS_ERR(trans)) {
4599                 err = PTR_ERR(trans);
4600                 goto out_free;
4601         }
4602         err = btrfs_commit_transaction(trans);
4603 out_free:
4604         kfree(rc);
4605 out:
4606         if (!list_empty(&reloc_roots))
4607                 free_reloc_roots(&reloc_roots);
4608
4609         btrfs_free_path(path);
4610
4611         if (err == 0) {
4612                 /* cleanup orphan inode in data relocation tree */
4613                 fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4614                 if (IS_ERR(fs_root))
4615                         err = PTR_ERR(fs_root);
4616                 else
4617                         err = btrfs_orphan_cleanup(fs_root);
4618         }
4619         return err;
4620 }
4621
4622 /*
4623  * helper to add ordered checksum for data relocation.
4624  *
4625  * cloning checksum properly handles the nodatasum extents.
4626  * it also saves CPU time to re-calculate the checksum.
4627  */
4628 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4629 {
4630         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4631         struct btrfs_ordered_sum *sums;
4632         struct btrfs_ordered_extent *ordered;
4633         int ret;
4634         u64 disk_bytenr;
4635         u64 new_bytenr;
4636         LIST_HEAD(list);
4637
4638         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4639         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4640
4641         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4642         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4643                                        disk_bytenr + len - 1, &list, 0);
4644         if (ret)
4645                 goto out;
4646
4647         while (!list_empty(&list)) {
4648                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4649                 list_del_init(&sums->list);
4650
4651                 /*
4652                  * We need to offset the new_bytenr based on where the csum is.
4653                  * We need to do this because we will read in entire prealloc
4654                  * extents but we may have written to say the middle of the
4655                  * prealloc extent, so we need to make sure the csum goes with
4656                  * the right disk offset.
4657                  *
4658                  * We can do this because the data reloc inode refers strictly
4659                  * to the on disk bytes, so we don't have to worry about
4660                  * disk_len vs real len like with real inodes since it's all
4661                  * disk length.
4662                  */
4663                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4664                 sums->bytenr = new_bytenr;
4665
4666                 btrfs_add_ordered_sum(inode, ordered, sums);
4667         }
4668 out:
4669         btrfs_put_ordered_extent(ordered);
4670         return ret;
4671 }
4672
4673 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4674                           struct btrfs_root *root, struct extent_buffer *buf,
4675                           struct extent_buffer *cow)
4676 {
4677         struct btrfs_fs_info *fs_info = root->fs_info;
4678         struct reloc_control *rc;
4679         struct backref_node *node;
4680         int first_cow = 0;
4681         int level;
4682         int ret = 0;
4683
4684         rc = fs_info->reloc_ctl;
4685         if (!rc)
4686                 return 0;
4687
4688         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4689                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4690
4691         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4692                 if (buf == root->node)
4693                         __update_reloc_root(root, cow->start);
4694         }
4695
4696         level = btrfs_header_level(buf);
4697         if (btrfs_header_generation(buf) <=
4698             btrfs_root_last_snapshot(&root->root_item))
4699                 first_cow = 1;
4700
4701         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4702             rc->create_reloc_tree) {
4703                 WARN_ON(!first_cow && level == 0);
4704
4705                 node = rc->backref_cache.path[level];
4706                 BUG_ON(node->bytenr != buf->start &&
4707                        node->new_bytenr != buf->start);
4708
4709                 drop_node_buffer(node);
4710                 extent_buffer_get(cow);
4711                 node->eb = cow;
4712                 node->new_bytenr = cow->start;
4713
4714                 if (!node->pending) {
4715                         list_move_tail(&node->list,
4716                                        &rc->backref_cache.pending[level]);
4717                         node->pending = 1;
4718                 }
4719
4720                 if (first_cow)
4721                         __mark_block_processed(rc, node);
4722
4723                 if (first_cow && level > 0)
4724                         rc->nodes_relocated += buf->len;
4725         }
4726
4727         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4728                 ret = replace_file_extents(trans, rc, root, cow);
4729         return ret;
4730 }
4731
4732 /*
4733  * called before creating snapshot. it calculates metadata reservation
4734  * required for relocating tree blocks in the snapshot
4735  */
4736 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4737                               u64 *bytes_to_reserve)
4738 {
4739         struct btrfs_root *root;
4740         struct reloc_control *rc;
4741
4742         root = pending->root;
4743         if (!root->reloc_root)
4744                 return;
4745
4746         rc = root->fs_info->reloc_ctl;
4747         if (!rc->merge_reloc_tree)
4748                 return;
4749
4750         root = root->reloc_root;
4751         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4752         /*
4753          * relocation is in the stage of merging trees. the space
4754          * used by merging a reloc tree is twice the size of
4755          * relocated tree nodes in the worst case. half for cowing
4756          * the reloc tree, half for cowing the fs tree. the space
4757          * used by cowing the reloc tree will be freed after the
4758          * tree is dropped. if we create snapshot, cowing the fs
4759          * tree may use more space than it frees. so we need
4760          * reserve extra space.
4761          */
4762         *bytes_to_reserve += rc->nodes_relocated;
4763 }
4764
4765 /*
4766  * called after snapshot is created. migrate block reservation
4767  * and create reloc root for the newly created snapshot
4768  */
4769 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4770                                struct btrfs_pending_snapshot *pending)
4771 {
4772         struct btrfs_root *root = pending->root;
4773         struct btrfs_root *reloc_root;
4774         struct btrfs_root *new_root;
4775         struct reloc_control *rc;
4776         int ret;
4777
4778         if (!root->reloc_root)
4779                 return 0;
4780
4781         rc = root->fs_info->reloc_ctl;
4782         rc->merging_rsv_size += rc->nodes_relocated;
4783
4784         if (rc->merge_reloc_tree) {
4785                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4786                                               rc->block_rsv,
4787                                               rc->nodes_relocated, 1);
4788                 if (ret)
4789                         return ret;
4790         }
4791
4792         new_root = pending->snap;
4793         reloc_root = create_reloc_root(trans, root->reloc_root,
4794                                        new_root->root_key.objectid);
4795         if (IS_ERR(reloc_root))
4796                 return PTR_ERR(reloc_root);
4797
4798         ret = __add_reloc_root(reloc_root);
4799         BUG_ON(ret < 0);
4800         new_root->reloc_root = reloc_root;
4801
4802         if (rc->create_reloc_tree)
4803                 ret = clone_backref_node(trans, rc, root, reloc_root);
4804         return ret;
4805 }