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_header_generation(reloc_root->commit_root) ==
 542             root->fs_info->running_transaction->transid)
 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                 remove_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->commit_root->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->commit_root->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                         RB_CLEAR_NODE(&node->rb_node);
1345                 }
1346                 spin_unlock(&rc->reloc_root_tree.lock);
1347                 ASSERT(!node || (struct btrfs_root *)node->data == root);
1348         }
1349
1350         spin_lock(&fs_info->trans_lock);
1351         list_del_init(&root->root_list);
1352         spin_unlock(&fs_info->trans_lock);
1353         kfree(node);
1354 }
1355
1356 /*
1357  * helper to update the 'address of tree root -> reloc tree'
1358  * mapping
1359  */
1360 static int __update_reloc_root(struct btrfs_root *root)
1361 {
1362         struct btrfs_fs_info *fs_info = root->fs_info;
1363         struct rb_node *rb_node;
1364         struct mapping_node *node = NULL;
1365         struct reloc_control *rc = fs_info->reloc_ctl;
1366
1367         spin_lock(&rc->reloc_root_tree.lock);
1368         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1369                               root->commit_root->start);
1370         if (rb_node) {
1371                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1372                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1373         }
1374         spin_unlock(&rc->reloc_root_tree.lock);
1375
1376         if (!node)
1377                 return 0;
1378         BUG_ON((struct btrfs_root *)node->data != root);
1379
1380         spin_lock(&rc->reloc_root_tree.lock);
1381         node->bytenr = root->node->start;
1382         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1383                               node->bytenr, &node->rb_node);
1384         spin_unlock(&rc->reloc_root_tree.lock);
1385         if (rb_node)
1386                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1387         return 0;
1388 }
1389
1390 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1391                                         struct btrfs_root *root, u64 objectid)
1392 {
1393         struct btrfs_fs_info *fs_info = root->fs_info;
1394         struct btrfs_root *reloc_root;
1395         struct extent_buffer *eb;
1396         struct btrfs_root_item *root_item;
1397         struct btrfs_key root_key;
1398         int ret;
1399
1400         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1401         BUG_ON(!root_item);
1402
1403         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1404         root_key.type = BTRFS_ROOT_ITEM_KEY;
1405         root_key.offset = objectid;
1406
1407         if (root->root_key.objectid == objectid) {
1408                 u64 commit_root_gen;
1409
1410                 /* called by btrfs_init_reloc_root */
1411                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1412                                       BTRFS_TREE_RELOC_OBJECTID);
1413                 BUG_ON(ret);
1414                 /*
1415                  * Set the last_snapshot field to the generation of the commit
1416                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1417                  * correctly (returns true) when the relocation root is created
1418                  * either inside the critical section of a transaction commit
1419                  * (through transaction.c:qgroup_account_snapshot()) and when
1420                  * it's created before the transaction commit is started.
1421                  */
1422                 commit_root_gen = btrfs_header_generation(root->commit_root);
1423                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1424         } else {
1425                 /*
1426                  * called by btrfs_reloc_post_snapshot_hook.
1427                  * the source tree is a reloc tree, all tree blocks
1428                  * modified after it was created have RELOC flag
1429                  * set in their headers. so it's OK to not update
1430                  * the 'last_snapshot'.
1431                  */
1432                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1433                                       BTRFS_TREE_RELOC_OBJECTID);
1434                 BUG_ON(ret);
1435         }
1436
1437         memcpy(root_item, &root->root_item, sizeof(*root_item));
1438         btrfs_set_root_bytenr(root_item, eb->start);
1439         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1440         btrfs_set_root_generation(root_item, trans->transid);
1441
1442         if (root->root_key.objectid == objectid) {
1443                 btrfs_set_root_refs(root_item, 0);
1444                 memset(&root_item->drop_progress, 0,
1445                        sizeof(struct btrfs_disk_key));
1446                 root_item->drop_level = 0;
1447         }
1448
1449         btrfs_tree_unlock(eb);
1450         free_extent_buffer(eb);
1451
1452         ret = btrfs_insert_root(trans, fs_info->tree_root,
1453                                 &root_key, root_item);
1454         BUG_ON(ret);
1455         kfree(root_item);
1456
1457         reloc_root = btrfs_read_fs_root(fs_info->tree_root, &root_key);
1458         BUG_ON(IS_ERR(reloc_root));
1459         reloc_root->last_trans = trans->transid;
1460         return reloc_root;
1461 }
1462
1463 /*
1464  * create reloc tree for a given fs tree. reloc tree is just a
1465  * snapshot of the fs tree with special root objectid.
1466  */
1467 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1468                           struct btrfs_root *root)
1469 {
1470         struct btrfs_fs_info *fs_info = root->fs_info;
1471         struct btrfs_root *reloc_root;
1472         struct reloc_control *rc = fs_info->reloc_ctl;
1473         struct btrfs_block_rsv *rsv;
1474         int clear_rsv = 0;
1475         int ret;
1476
1477         if (root->reloc_root) {
1478                 reloc_root = root->reloc_root;
1479                 reloc_root->last_trans = trans->transid;
1480                 return 0;
1481         }
1482
1483         if (!rc || !rc->create_reloc_tree ||
1484             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1485                 return 0;
1486
1487         if (!trans->reloc_reserved) {
1488                 rsv = trans->block_rsv;
1489                 trans->block_rsv = rc->block_rsv;
1490                 clear_rsv = 1;
1491         }
1492         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1493         if (clear_rsv)
1494                 trans->block_rsv = rsv;
1495
1496         ret = __add_reloc_root(reloc_root);
1497         BUG_ON(ret < 0);
1498         root->reloc_root = reloc_root;
1499         return 0;
1500 }
1501
1502 /*
1503  * update root item of reloc tree
1504  */
1505 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1506                             struct btrfs_root *root)
1507 {
1508         struct btrfs_fs_info *fs_info = root->fs_info;
1509         struct btrfs_root *reloc_root;
1510         struct btrfs_root_item *root_item;
1511         int ret;
1512
1513         if (!root->reloc_root)
1514                 goto out;
1515
1516         reloc_root = root->reloc_root;
1517         root_item = &reloc_root->root_item;
1518
1519         if (fs_info->reloc_ctl->merge_reloc_tree &&
1520             btrfs_root_refs(root_item) == 0) {
1521                 root->reloc_root = NULL;
1522                 __del_reloc_root(reloc_root);
1523         }
1524
1525         if (reloc_root->commit_root != reloc_root->node) {
1526                 __update_reloc_root(reloc_root);
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         ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1812         ASSERT(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                 ASSERT(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         /*
2484          * We used to have
2485          *
2486          * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2487          *
2488          * here, but it's wrong.  If we fail to start the transaction in
2489          * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2490          * have actually been removed from the reloc_root_tree rb tree.  This is
2491          * fine because we're bailing here, and we hold a reference on the root
2492          * for the list that holds it, so these roots will be cleaned up when we
2493          * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
2494          * will be cleaned up on unmount.
2495          *
2496          * The remaining nodes will be cleaned up by free_reloc_control.
2497          */
2498 }
2499
2500 static void free_block_list(struct rb_root *blocks)
2501 {
2502         struct tree_block *block;
2503         struct rb_node *rb_node;
2504         while ((rb_node = rb_first(blocks))) {
2505                 block = rb_entry(rb_node, struct tree_block, rb_node);
2506                 rb_erase(rb_node, blocks);
2507                 kfree(block);
2508         }
2509 }
2510
2511 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2512                                       struct btrfs_root *reloc_root)
2513 {
2514         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2515         struct btrfs_root *root;
2516
2517         if (reloc_root->last_trans == trans->transid)
2518                 return 0;
2519
2520         root = read_fs_root(fs_info, reloc_root->root_key.offset);
2521         BUG_ON(IS_ERR(root));
2522         BUG_ON(root->reloc_root != reloc_root);
2523
2524         return btrfs_record_root_in_trans(trans, root);
2525 }
2526
2527 static noinline_for_stack
2528 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2529                                      struct reloc_control *rc,
2530                                      struct backref_node *node,
2531                                      struct backref_edge *edges[])
2532 {
2533         struct backref_node *next;
2534         struct btrfs_root *root;
2535         int index = 0;
2536
2537         next = node;
2538         while (1) {
2539                 cond_resched();
2540                 next = walk_up_backref(next, edges, &index);
2541                 root = next->root;
2542                 BUG_ON(!root);
2543                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2544
2545                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2546                         record_reloc_root_in_trans(trans, root);
2547                         break;
2548                 }
2549
2550                 btrfs_record_root_in_trans(trans, root);
2551                 root = root->reloc_root;
2552
2553                 if (next->new_bytenr != root->node->start) {
2554                         BUG_ON(next->new_bytenr);
2555                         BUG_ON(!list_empty(&next->list));
2556                         next->new_bytenr = root->node->start;
2557                         next->root = root;
2558                         list_add_tail(&next->list,
2559                                       &rc->backref_cache.changed);
2560                         __mark_block_processed(rc, next);
2561                         break;
2562                 }
2563
2564                 WARN_ON(1);
2565                 root = NULL;
2566                 next = walk_down_backref(edges, &index);
2567                 if (!next || next->level <= node->level)
2568                         break;
2569         }
2570         if (!root)
2571                 return NULL;
2572
2573         next = node;
2574         /* setup backref node path for btrfs_reloc_cow_block */
2575         while (1) {
2576                 rc->backref_cache.path[next->level] = next;
2577                 if (--index < 0)
2578                         break;
2579                 next = edges[index]->node[UPPER];
2580         }
2581         return root;
2582 }
2583
2584 /*
2585  * select a tree root for relocation. return NULL if the block
2586  * is reference counted. we should use do_relocation() in this
2587  * case. return a tree root pointer if the block isn't reference
2588  * counted. return -ENOENT if the block is root of reloc tree.
2589  */
2590 static noinline_for_stack
2591 struct btrfs_root *select_one_root(struct backref_node *node)
2592 {
2593         struct backref_node *next;
2594         struct btrfs_root *root;
2595         struct btrfs_root *fs_root = NULL;
2596         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2597         int index = 0;
2598
2599         next = node;
2600         while (1) {
2601                 cond_resched();
2602                 next = walk_up_backref(next, edges, &index);
2603                 root = next->root;
2604                 BUG_ON(!root);
2605
2606                 /* no other choice for non-references counted tree */
2607                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2608                         return root;
2609
2610                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2611                         fs_root = root;
2612
2613                 if (next != node)
2614                         return NULL;
2615
2616                 next = walk_down_backref(edges, &index);
2617                 if (!next || next->level <= node->level)
2618                         break;
2619         }
2620
2621         if (!fs_root)
2622                 return ERR_PTR(-ENOENT);
2623         return fs_root;
2624 }
2625
2626 static noinline_for_stack
2627 u64 calcu_metadata_size(struct reloc_control *rc,
2628                         struct backref_node *node, int reserve)
2629 {
2630         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2631         struct backref_node *next = node;
2632         struct backref_edge *edge;
2633         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2634         u64 num_bytes = 0;
2635         int index = 0;
2636
2637         BUG_ON(reserve && node->processed);
2638
2639         while (next) {
2640                 cond_resched();
2641                 while (1) {
2642                         if (next->processed && (reserve || next != node))
2643                                 break;
2644
2645                         num_bytes += fs_info->nodesize;
2646
2647                         if (list_empty(&next->upper))
2648                                 break;
2649
2650                         edge = list_entry(next->upper.next,
2651                                           struct backref_edge, list[LOWER]);
2652                         edges[index++] = edge;
2653                         next = edge->node[UPPER];
2654                 }
2655                 next = walk_down_backref(edges, &index);
2656         }
2657         return num_bytes;
2658 }
2659
2660 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2661                                   struct reloc_control *rc,
2662                                   struct backref_node *node)
2663 {
2664         struct btrfs_root *root = rc->extent_root;
2665         struct btrfs_fs_info *fs_info = root->fs_info;
2666         u64 num_bytes;
2667         int ret;
2668         u64 tmp;
2669
2670         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2671
2672         trans->block_rsv = rc->block_rsv;
2673         rc->reserved_bytes += num_bytes;
2674
2675         /*
2676          * We are under a transaction here so we can only do limited flushing.
2677          * If we get an enospc just kick back -EAGAIN so we know to drop the
2678          * transaction and try to refill when we can flush all the things.
2679          */
2680         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2681                                 BTRFS_RESERVE_FLUSH_LIMIT);
2682         if (ret) {
2683                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2684                 while (tmp <= rc->reserved_bytes)
2685                         tmp <<= 1;
2686                 /*
2687                  * only one thread can access block_rsv at this point,
2688                  * so we don't need hold lock to protect block_rsv.
2689                  * we expand more reservation size here to allow enough
2690                  * space for relocation and we will return eailer in
2691                  * enospc case.
2692                  */
2693                 rc->block_rsv->size = tmp + fs_info->nodesize *
2694                                       RELOCATION_RESERVED_NODES;
2695                 return -EAGAIN;
2696         }
2697
2698         return 0;
2699 }
2700
2701 /*
2702  * relocate a block tree, and then update pointers in upper level
2703  * blocks that reference the block to point to the new location.
2704  *
2705  * if called by link_to_upper, the block has already been relocated.
2706  * in that case this function just updates pointers.
2707  */
2708 static int do_relocation(struct btrfs_trans_handle *trans,
2709                          struct reloc_control *rc,
2710                          struct backref_node *node,
2711                          struct btrfs_key *key,
2712                          struct btrfs_path *path, int lowest)
2713 {
2714         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2715         struct backref_node *upper;
2716         struct backref_edge *edge;
2717         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2718         struct btrfs_root *root;
2719         struct extent_buffer *eb;
2720         u32 blocksize;
2721         u64 bytenr;
2722         u64 generation;
2723         int slot;
2724         int ret;
2725         int err = 0;
2726
2727         BUG_ON(lowest && node->eb);
2728
2729         path->lowest_level = node->level + 1;
2730         rc->backref_cache.path[node->level] = node;
2731         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2732                 cond_resched();
2733
2734                 upper = edge->node[UPPER];
2735                 root = select_reloc_root(trans, rc, upper, edges);
2736                 BUG_ON(!root);
2737
2738                 if (upper->eb && !upper->locked) {
2739                         if (!lowest) {
2740                                 ret = btrfs_bin_search(upper->eb, key,
2741                                                        upper->level, &slot);
2742                                 BUG_ON(ret);
2743                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2744                                 if (node->eb->start == bytenr)
2745                                         goto next;
2746                         }
2747                         drop_node_buffer(upper);
2748                 }
2749
2750                 if (!upper->eb) {
2751                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2752                         if (ret) {
2753                                 if (ret < 0)
2754                                         err = ret;
2755                                 else
2756                                         err = -ENOENT;
2757
2758                                 btrfs_release_path(path);
2759                                 break;
2760                         }
2761
2762                         if (!upper->eb) {
2763                                 upper->eb = path->nodes[upper->level];
2764                                 path->nodes[upper->level] = NULL;
2765                         } else {
2766                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2767                         }
2768
2769                         upper->locked = 1;
2770                         path->locks[upper->level] = 0;
2771
2772                         slot = path->slots[upper->level];
2773                         btrfs_release_path(path);
2774                 } else {
2775                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2776                                                &slot);
2777                         BUG_ON(ret);
2778                 }
2779
2780                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2781                 if (lowest) {
2782                         if (bytenr != node->bytenr) {
2783                                 btrfs_err(root->fs_info,
2784                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2785                                           bytenr, node->bytenr, slot,
2786                                           upper->eb->start);
2787                                 err = -EIO;
2788                                 goto next;
2789                         }
2790                 } else {
2791                         if (node->eb->start == bytenr)
2792                                 goto next;
2793                 }
2794
2795                 blocksize = root->fs_info->nodesize;
2796                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2797                 eb = read_tree_block(fs_info, bytenr, generation);
2798                 if (IS_ERR(eb)) {
2799                         err = PTR_ERR(eb);
2800                         goto next;
2801                 } else if (!extent_buffer_uptodate(eb)) {
2802                         free_extent_buffer(eb);
2803                         err = -EIO;
2804                         goto next;
2805                 }
2806                 btrfs_tree_lock(eb);
2807                 btrfs_set_lock_blocking(eb);
2808
2809                 if (!node->eb) {
2810                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2811                                               slot, &eb);
2812                         btrfs_tree_unlock(eb);
2813                         free_extent_buffer(eb);
2814                         if (ret < 0) {
2815                                 err = ret;
2816                                 goto next;
2817                         }
2818                         BUG_ON(node->eb != eb);
2819                 } else {
2820                         btrfs_set_node_blockptr(upper->eb, slot,
2821                                                 node->eb->start);
2822                         btrfs_set_node_ptr_generation(upper->eb, slot,
2823                                                       trans->transid);
2824                         btrfs_mark_buffer_dirty(upper->eb);
2825
2826                         ret = btrfs_inc_extent_ref(trans, root->fs_info,
2827                                                 node->eb->start, blocksize,
2828                                                 upper->eb->start,
2829                                                 btrfs_header_owner(upper->eb),
2830                                                 node->level, 0);
2831                         BUG_ON(ret);
2832
2833                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2834                         BUG_ON(ret);
2835                 }
2836 next:
2837                 if (!upper->pending)
2838                         drop_node_buffer(upper);
2839                 else
2840                         unlock_node_buffer(upper);
2841                 if (err)
2842                         break;
2843         }
2844
2845         if (!err && node->pending) {
2846                 drop_node_buffer(node);
2847                 list_move_tail(&node->list, &rc->backref_cache.changed);
2848                 node->pending = 0;
2849         }
2850
2851         path->lowest_level = 0;
2852         BUG_ON(err == -ENOSPC);
2853         return err;
2854 }
2855
2856 static int link_to_upper(struct btrfs_trans_handle *trans,
2857                          struct reloc_control *rc,
2858                          struct backref_node *node,
2859                          struct btrfs_path *path)
2860 {
2861         struct btrfs_key key;
2862
2863         btrfs_node_key_to_cpu(node->eb, &key, 0);
2864         return do_relocation(trans, rc, node, &key, path, 0);
2865 }
2866
2867 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2868                                 struct reloc_control *rc,
2869                                 struct btrfs_path *path, int err)
2870 {
2871         LIST_HEAD(list);
2872         struct backref_cache *cache = &rc->backref_cache;
2873         struct backref_node *node;
2874         int level;
2875         int ret;
2876
2877         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2878                 while (!list_empty(&cache->pending[level])) {
2879                         node = list_entry(cache->pending[level].next,
2880                                           struct backref_node, list);
2881                         list_move_tail(&node->list, &list);
2882                         BUG_ON(!node->pending);
2883
2884                         if (!err) {
2885                                 ret = link_to_upper(trans, rc, node, path);
2886                                 if (ret < 0)
2887                                         err = ret;
2888                         }
2889                 }
2890                 list_splice_init(&list, &cache->pending[level]);
2891         }
2892         return err;
2893 }
2894
2895 static void mark_block_processed(struct reloc_control *rc,
2896                                  u64 bytenr, u32 blocksize)
2897 {
2898         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2899                         EXTENT_DIRTY);
2900 }
2901
2902 static void __mark_block_processed(struct reloc_control *rc,
2903                                    struct backref_node *node)
2904 {
2905         u32 blocksize;
2906         if (node->level == 0 ||
2907             in_block_group(node->bytenr, rc->block_group)) {
2908                 blocksize = rc->extent_root->fs_info->nodesize;
2909                 mark_block_processed(rc, node->bytenr, blocksize);
2910         }
2911         node->processed = 1;
2912 }
2913
2914 /*
2915  * mark a block and all blocks directly/indirectly reference the block
2916  * as processed.
2917  */
2918 static void update_processed_blocks(struct reloc_control *rc,
2919                                     struct backref_node *node)
2920 {
2921         struct backref_node *next = node;
2922         struct backref_edge *edge;
2923         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2924         int index = 0;
2925
2926         while (next) {
2927                 cond_resched();
2928                 while (1) {
2929                         if (next->processed)
2930                                 break;
2931
2932                         __mark_block_processed(rc, next);
2933
2934                         if (list_empty(&next->upper))
2935                                 break;
2936
2937                         edge = list_entry(next->upper.next,
2938                                           struct backref_edge, list[LOWER]);
2939                         edges[index++] = edge;
2940                         next = edge->node[UPPER];
2941                 }
2942                 next = walk_down_backref(edges, &index);
2943         }
2944 }
2945
2946 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2947 {
2948         u32 blocksize = rc->extent_root->fs_info->nodesize;
2949
2950         if (test_range_bit(&rc->processed_blocks, bytenr,
2951                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2952                 return 1;
2953         return 0;
2954 }
2955
2956 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2957                               struct tree_block *block)
2958 {
2959         struct extent_buffer *eb;
2960
2961         BUG_ON(block->key_ready);
2962         eb = read_tree_block(fs_info, block->bytenr, block->key.offset);
2963         if (IS_ERR(eb)) {
2964                 return PTR_ERR(eb);
2965         } else if (!extent_buffer_uptodate(eb)) {
2966                 free_extent_buffer(eb);
2967                 return -EIO;
2968         }
2969         WARN_ON(btrfs_header_level(eb) != block->level);
2970         if (block->level == 0)
2971                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2972         else
2973                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2974         free_extent_buffer(eb);
2975         block->key_ready = 1;
2976         return 0;
2977 }
2978
2979 /*
2980  * helper function to relocate a tree block
2981  */
2982 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2983                                 struct reloc_control *rc,
2984                                 struct backref_node *node,
2985                                 struct btrfs_key *key,
2986                                 struct btrfs_path *path)
2987 {
2988         struct btrfs_root *root;
2989         int ret = 0;
2990
2991         if (!node)
2992                 return 0;
2993
2994         BUG_ON(node->processed);
2995         root = select_one_root(node);
2996         if (root == ERR_PTR(-ENOENT)) {
2997                 update_processed_blocks(rc, node);
2998                 goto out;
2999         }
3000
3001         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3002                 ret = reserve_metadata_space(trans, rc, node);
3003                 if (ret)
3004                         goto out;
3005         }
3006
3007         if (root) {
3008                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3009                         BUG_ON(node->new_bytenr);
3010                         BUG_ON(!list_empty(&node->list));
3011                         btrfs_record_root_in_trans(trans, root);
3012                         root = root->reloc_root;
3013                         node->new_bytenr = root->node->start;
3014                         node->root = root;
3015                         list_add_tail(&node->list, &rc->backref_cache.changed);
3016                 } else {
3017                         path->lowest_level = node->level;
3018                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
3019                         btrfs_release_path(path);
3020                         if (ret > 0)
3021                                 ret = 0;
3022                 }
3023                 if (!ret)
3024                         update_processed_blocks(rc, node);
3025         } else {
3026                 ret = do_relocation(trans, rc, node, key, path, 1);
3027         }
3028 out:
3029         if (ret || node->level == 0 || node->cowonly)
3030                 remove_backref_node(&rc->backref_cache, node);
3031         return ret;
3032 }
3033
3034 /*
3035  * relocate a list of blocks
3036  */
3037 static noinline_for_stack
3038 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
3039                          struct reloc_control *rc, struct rb_root *blocks)
3040 {
3041         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3042         struct backref_node *node;
3043         struct btrfs_path *path;
3044         struct tree_block *block;
3045         struct rb_node *rb_node;
3046         int ret;
3047         int err = 0;
3048
3049         path = btrfs_alloc_path();
3050         if (!path) {
3051                 err = -ENOMEM;
3052                 goto out_free_blocks;
3053         }
3054
3055         rb_node = rb_first(blocks);
3056         while (rb_node) {
3057                 block = rb_entry(rb_node, struct tree_block, rb_node);
3058                 if (!block->key_ready)
3059                         readahead_tree_block(fs_info, block->bytenr);
3060                 rb_node = rb_next(rb_node);
3061         }
3062
3063         rb_node = rb_first(blocks);
3064         while (rb_node) {
3065                 block = rb_entry(rb_node, struct tree_block, rb_node);
3066                 if (!block->key_ready) {
3067                         err = get_tree_block_key(fs_info, block);
3068                         if (err)
3069                                 goto out_free_path;
3070                 }
3071                 rb_node = rb_next(rb_node);
3072         }
3073
3074         rb_node = rb_first(blocks);
3075         while (rb_node) {
3076                 block = rb_entry(rb_node, struct tree_block, rb_node);
3077
3078                 node = build_backref_tree(rc, &block->key,
3079                                           block->level, block->bytenr);
3080                 if (IS_ERR(node)) {
3081                         err = PTR_ERR(node);
3082                         goto out;
3083                 }
3084
3085                 ret = relocate_tree_block(trans, rc, node, &block->key,
3086                                           path);
3087                 if (ret < 0) {
3088                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3089                                 err = ret;
3090                         goto out;
3091                 }
3092                 rb_node = rb_next(rb_node);
3093         }
3094 out:
3095         err = finish_pending_nodes(trans, rc, path, err);
3096
3097 out_free_path:
3098         btrfs_free_path(path);
3099 out_free_blocks:
3100         free_block_list(blocks);
3101         return err;
3102 }
3103
3104 static noinline_for_stack
3105 int prealloc_file_extent_cluster(struct inode *inode,
3106                                  struct file_extent_cluster *cluster)
3107 {
3108         u64 alloc_hint = 0;
3109         u64 start;
3110         u64 end;
3111         u64 offset = BTRFS_I(inode)->index_cnt;
3112         u64 num_bytes;
3113         int nr = 0;
3114         int ret = 0;
3115         u64 prealloc_start = cluster->start - offset;
3116         u64 prealloc_end = cluster->end - offset;
3117         u64 cur_offset;
3118         struct extent_changeset *data_reserved = NULL;
3119
3120         BUG_ON(cluster->start != cluster->boundary[0]);
3121         inode_lock(inode);
3122
3123         ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
3124                                           prealloc_end + 1 - prealloc_start);
3125         if (ret)
3126                 goto out;
3127
3128         cur_offset = prealloc_start;
3129         while (nr < cluster->nr) {
3130                 start = cluster->boundary[nr] - offset;
3131                 if (nr + 1 < cluster->nr)
3132                         end = cluster->boundary[nr + 1] - 1 - offset;
3133                 else
3134                         end = cluster->end - offset;
3135
3136                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3137                 num_bytes = end + 1 - start;
3138                 if (cur_offset < start)
3139                         btrfs_free_reserved_data_space(inode, data_reserved,
3140                                         cur_offset, start - cur_offset);
3141                 ret = btrfs_prealloc_file_range(inode, 0, start,
3142                                                 num_bytes, num_bytes,
3143                                                 end + 1, &alloc_hint);
3144                 cur_offset = end + 1;
3145                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3146                 if (ret)
3147                         break;
3148                 nr++;
3149         }
3150         if (cur_offset < prealloc_end)
3151                 btrfs_free_reserved_data_space(inode, data_reserved,
3152                                 cur_offset, prealloc_end + 1 - cur_offset);
3153 out:
3154         inode_unlock(inode);
3155         extent_changeset_free(data_reserved);
3156         return ret;
3157 }
3158
3159 static noinline_for_stack
3160 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3161                          u64 block_start)
3162 {
3163         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3164         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3165         struct extent_map *em;
3166         int ret = 0;
3167
3168         em = alloc_extent_map();
3169         if (!em)
3170                 return -ENOMEM;
3171
3172         em->start = start;
3173         em->len = end + 1 - start;
3174         em->block_len = em->len;
3175         em->block_start = block_start;
3176         em->bdev = fs_info->fs_devices->latest_bdev;
3177         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3178
3179         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3180         while (1) {
3181                 write_lock(&em_tree->lock);
3182                 ret = add_extent_mapping(em_tree, em, 0);
3183                 write_unlock(&em_tree->lock);
3184                 if (ret != -EEXIST) {
3185                         free_extent_map(em);
3186                         break;
3187                 }
3188                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3189         }
3190         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3191         return ret;
3192 }
3193
3194 static int relocate_file_extent_cluster(struct inode *inode,
3195                                         struct file_extent_cluster *cluster)
3196 {
3197         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3198         u64 page_start;
3199         u64 page_end;
3200         u64 offset = BTRFS_I(inode)->index_cnt;
3201         unsigned long index;
3202         unsigned long last_index;
3203         struct page *page;
3204         struct file_ra_state *ra;
3205         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3206         int nr = 0;
3207         int ret = 0;
3208
3209         if (!cluster->nr)
3210                 return 0;
3211
3212         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3213         if (!ra)
3214                 return -ENOMEM;
3215
3216         ret = prealloc_file_extent_cluster(inode, cluster);
3217         if (ret)
3218                 goto out;
3219
3220         file_ra_state_init(ra, inode->i_mapping);
3221
3222         ret = setup_extent_mapping(inode, cluster->start - offset,
3223                                    cluster->end - offset, cluster->start);
3224         if (ret)
3225                 goto out;
3226
3227         index = (cluster->start - offset) >> PAGE_SHIFT;
3228         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3229         while (index <= last_index) {
3230                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3231                                 PAGE_SIZE);
3232                 if (ret)
3233                         goto out;
3234
3235                 page = find_lock_page(inode->i_mapping, index);
3236                 if (!page) {
3237                         page_cache_sync_readahead(inode->i_mapping,
3238                                                   ra, NULL, index,
3239                                                   last_index + 1 - index);
3240                         page = find_or_create_page(inode->i_mapping, index,
3241                                                    mask);
3242                         if (!page) {
3243                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3244                                                         PAGE_SIZE);
3245                                 ret = -ENOMEM;
3246                                 goto out;
3247                         }
3248                 }
3249
3250                 if (PageReadahead(page)) {
3251                         page_cache_async_readahead(inode->i_mapping,
3252                                                    ra, NULL, page, index,
3253                                                    last_index + 1 - index);
3254                 }
3255
3256                 if (!PageUptodate(page)) {
3257                         btrfs_readpage(NULL, page);
3258                         lock_page(page);
3259                         if (!PageUptodate(page)) {
3260                                 unlock_page(page);
3261                                 put_page(page);
3262                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3263                                                         PAGE_SIZE);
3264                                 ret = -EIO;
3265                                 goto out;
3266                         }
3267                 }
3268
3269                 page_start = page_offset(page);
3270                 page_end = page_start + PAGE_SIZE - 1;
3271
3272                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3273
3274                 set_page_extent_mapped(page);
3275
3276                 if (nr < cluster->nr &&
3277                     page_start + offset == cluster->boundary[nr]) {
3278                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3279                                         page_start, page_end,
3280                                         EXTENT_BOUNDARY);
3281                         nr++;
3282                 }
3283
3284                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL, 0);
3285                 set_page_dirty(page);
3286
3287                 unlock_extent(&BTRFS_I(inode)->io_tree,
3288                               page_start, page_end);
3289                 unlock_page(page);
3290                 put_page(page);
3291
3292                 index++;
3293                 balance_dirty_pages_ratelimited(inode->i_mapping);
3294                 btrfs_throttle(fs_info);
3295         }
3296         WARN_ON(nr != cluster->nr);
3297 out:
3298         kfree(ra);
3299         return ret;
3300 }
3301
3302 static noinline_for_stack
3303 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3304                          struct file_extent_cluster *cluster)
3305 {
3306         int ret;
3307
3308         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3309                 ret = relocate_file_extent_cluster(inode, cluster);
3310                 if (ret)
3311                         return ret;
3312                 cluster->nr = 0;
3313         }
3314
3315         if (!cluster->nr)
3316                 cluster->start = extent_key->objectid;
3317         else
3318                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3319         cluster->end = extent_key->objectid + extent_key->offset - 1;
3320         cluster->boundary[cluster->nr] = extent_key->objectid;
3321         cluster->nr++;
3322
3323         if (cluster->nr >= MAX_EXTENTS) {
3324                 ret = relocate_file_extent_cluster(inode, cluster);
3325                 if (ret)
3326                         return ret;
3327                 cluster->nr = 0;
3328         }
3329         return 0;
3330 }
3331
3332 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3333 static int get_ref_objectid_v0(struct reloc_control *rc,
3334                                struct btrfs_path *path,
3335                                struct btrfs_key *extent_key,
3336                                u64 *ref_objectid, int *path_change)
3337 {
3338         struct btrfs_key key;
3339         struct extent_buffer *leaf;
3340         struct btrfs_extent_ref_v0 *ref0;
3341         int ret;
3342         int slot;
3343
3344         leaf = path->nodes[0];
3345         slot = path->slots[0];
3346         while (1) {
3347                 if (slot >= btrfs_header_nritems(leaf)) {
3348                         ret = btrfs_next_leaf(rc->extent_root, path);
3349                         if (ret < 0)
3350                                 return ret;
3351                         BUG_ON(ret > 0);
3352                         leaf = path->nodes[0];
3353                         slot = path->slots[0];
3354                         if (path_change)
3355                                 *path_change = 1;
3356                 }
3357                 btrfs_item_key_to_cpu(leaf, &key, slot);
3358                 if (key.objectid != extent_key->objectid)
3359                         return -ENOENT;
3360
3361                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3362                         slot++;
3363                         continue;
3364                 }
3365                 ref0 = btrfs_item_ptr(leaf, slot,
3366                                 struct btrfs_extent_ref_v0);
3367                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3368                 break;
3369         }
3370         return 0;
3371 }
3372 #endif
3373
3374 /*
3375  * helper to add a tree block to the list.
3376  * the major work is getting the generation and level of the block
3377  */
3378 static int add_tree_block(struct reloc_control *rc,
3379                           struct btrfs_key *extent_key,
3380                           struct btrfs_path *path,
3381                           struct rb_root *blocks)
3382 {
3383         struct extent_buffer *eb;
3384         struct btrfs_extent_item *ei;
3385         struct btrfs_tree_block_info *bi;
3386         struct tree_block *block;
3387         struct rb_node *rb_node;
3388         u32 item_size;
3389         int level = -1;
3390         u64 generation;
3391
3392         eb =  path->nodes[0];
3393         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3394
3395         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3396             item_size >= sizeof(*ei) + sizeof(*bi)) {
3397                 ei = btrfs_item_ptr(eb, path->slots[0],
3398                                 struct btrfs_extent_item);
3399                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3400                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3401                         level = btrfs_tree_block_level(eb, bi);
3402                 } else {
3403                         level = (int)extent_key->offset;
3404                 }
3405                 generation = btrfs_extent_generation(eb, ei);
3406         } else {
3407 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3408                 u64 ref_owner;
3409                 int ret;
3410
3411                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3412                 ret = get_ref_objectid_v0(rc, path, extent_key,
3413                                           &ref_owner, NULL);
3414                 if (ret < 0)
3415                         return ret;
3416                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3417                 level = (int)ref_owner;
3418                 /* FIXME: get real generation */
3419                 generation = 0;
3420 #else
3421                 BUG();
3422 #endif
3423         }
3424
3425         btrfs_release_path(path);
3426
3427         BUG_ON(level == -1);
3428
3429         block = kmalloc(sizeof(*block), GFP_NOFS);
3430         if (!block)
3431                 return -ENOMEM;
3432
3433         block->bytenr = extent_key->objectid;
3434         block->key.objectid = rc->extent_root->fs_info->nodesize;
3435         block->key.offset = generation;
3436         block->level = level;
3437         block->key_ready = 0;
3438
3439         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3440         if (rb_node)
3441                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3442
3443         return 0;
3444 }
3445
3446 /*
3447  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3448  */
3449 static int __add_tree_block(struct reloc_control *rc,
3450                             u64 bytenr, u32 blocksize,
3451                             struct rb_root *blocks)
3452 {
3453         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3454         struct btrfs_path *path;
3455         struct btrfs_key key;
3456         int ret;
3457         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3458
3459         if (tree_block_processed(bytenr, rc))
3460                 return 0;
3461
3462         if (tree_search(blocks, bytenr))
3463                 return 0;
3464
3465         path = btrfs_alloc_path();
3466         if (!path)
3467                 return -ENOMEM;
3468 again:
3469         key.objectid = bytenr;
3470         if (skinny) {
3471                 key.type = BTRFS_METADATA_ITEM_KEY;
3472                 key.offset = (u64)-1;
3473         } else {
3474                 key.type = BTRFS_EXTENT_ITEM_KEY;
3475                 key.offset = blocksize;
3476         }
3477
3478         path->search_commit_root = 1;
3479         path->skip_locking = 1;
3480         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3481         if (ret < 0)
3482                 goto out;
3483
3484         if (ret > 0 && skinny) {
3485                 if (path->slots[0]) {
3486                         path->slots[0]--;
3487                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3488                                               path->slots[0]);
3489                         if (key.objectid == bytenr &&
3490                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3491                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3492                               key.offset == blocksize)))
3493                                 ret = 0;
3494                 }
3495
3496                 if (ret) {
3497                         skinny = false;
3498                         btrfs_release_path(path);
3499                         goto again;
3500                 }
3501         }
3502         if (ret) {
3503                 ASSERT(ret == 1);
3504                 btrfs_print_leaf(path->nodes[0]);
3505                 btrfs_err(fs_info,
3506              "tree block extent item (%llu) is not found in extent tree",
3507                      bytenr);
3508                 WARN_ON(1);
3509                 ret = -EINVAL;
3510                 goto out;
3511         }
3512
3513         ret = add_tree_block(rc, &key, path, blocks);
3514 out:
3515         btrfs_free_path(path);
3516         return ret;
3517 }
3518
3519 /*
3520  * helper to check if the block use full backrefs for pointers in it
3521  */
3522 static int block_use_full_backref(struct reloc_control *rc,
3523                                   struct extent_buffer *eb)
3524 {
3525         u64 flags;
3526         int ret;
3527
3528         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3529             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3530                 return 1;
3531
3532         ret = btrfs_lookup_extent_info(NULL, rc->extent_root->fs_info,
3533                                        eb->start, btrfs_header_level(eb), 1,
3534                                        NULL, &flags);
3535         BUG_ON(ret);
3536
3537         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3538                 ret = 1;
3539         else
3540                 ret = 0;
3541         return ret;
3542 }
3543
3544 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3545                                     struct btrfs_block_group_cache *block_group,
3546                                     struct inode *inode,
3547                                     u64 ino)
3548 {
3549         struct btrfs_key key;
3550         struct btrfs_root *root = fs_info->tree_root;
3551         struct btrfs_trans_handle *trans;
3552         int ret = 0;
3553
3554         if (inode)
3555                 goto truncate;
3556
3557         key.objectid = ino;
3558         key.type = BTRFS_INODE_ITEM_KEY;
3559         key.offset = 0;
3560
3561         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3562         if (IS_ERR(inode) || is_bad_inode(inode)) {
3563                 if (!IS_ERR(inode))
3564                         iput(inode);
3565                 return -ENOENT;
3566         }
3567
3568 truncate:
3569         ret = btrfs_check_trunc_cache_free_space(fs_info,
3570                                                  &fs_info->global_block_rsv);
3571         if (ret)
3572                 goto out;
3573
3574         trans = btrfs_join_transaction(root);
3575         if (IS_ERR(trans)) {
3576                 ret = PTR_ERR(trans);
3577                 goto out;
3578         }
3579
3580         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3581
3582         btrfs_end_transaction(trans);
3583         btrfs_btree_balance_dirty(fs_info);
3584 out:
3585         iput(inode);
3586         return ret;
3587 }
3588
3589 /*
3590  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3591  * this function scans fs tree to find blocks reference the data extent
3592  */
3593 static int find_data_references(struct reloc_control *rc,
3594                                 struct btrfs_key *extent_key,
3595                                 struct extent_buffer *leaf,
3596                                 struct btrfs_extent_data_ref *ref,
3597                                 struct rb_root *blocks)
3598 {
3599         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3600         struct btrfs_path *path;
3601         struct tree_block *block;
3602         struct btrfs_root *root;
3603         struct btrfs_file_extent_item *fi;
3604         struct rb_node *rb_node;
3605         struct btrfs_key key;
3606         u64 ref_root;
3607         u64 ref_objectid;
3608         u64 ref_offset;
3609         u32 ref_count;
3610         u32 nritems;
3611         int err = 0;
3612         int added = 0;
3613         int counted;
3614         int ret;
3615
3616         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3617         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3618         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3619         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3620
3621         /*
3622          * This is an extent belonging to the free space cache, lets just delete
3623          * it and redo the search.
3624          */
3625         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3626                 ret = delete_block_group_cache(fs_info, rc->block_group,
3627                                                NULL, ref_objectid);
3628                 if (ret != -ENOENT)
3629                         return ret;
3630                 ret = 0;
3631         }
3632
3633         path = btrfs_alloc_path();
3634         if (!path)
3635                 return -ENOMEM;
3636         path->reada = READA_FORWARD;
3637
3638         root = read_fs_root(fs_info, ref_root);
3639         if (IS_ERR(root)) {
3640                 err = PTR_ERR(root);
3641                 goto out;
3642         }
3643
3644         key.objectid = ref_objectid;
3645         key.type = BTRFS_EXTENT_DATA_KEY;
3646         if (ref_offset > ((u64)-1 << 32))
3647                 key.offset = 0;
3648         else
3649                 key.offset = ref_offset;
3650
3651         path->search_commit_root = 1;
3652         path->skip_locking = 1;
3653         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3654         if (ret < 0) {
3655                 err = ret;
3656                 goto out;
3657         }
3658
3659         leaf = path->nodes[0];
3660         nritems = btrfs_header_nritems(leaf);
3661         /*
3662          * the references in tree blocks that use full backrefs
3663          * are not counted in
3664          */
3665         if (block_use_full_backref(rc, leaf))
3666                 counted = 0;
3667         else
3668                 counted = 1;
3669         rb_node = tree_search(blocks, leaf->start);
3670         if (rb_node) {
3671                 if (counted)
3672                         added = 1;
3673                 else
3674                         path->slots[0] = nritems;
3675         }
3676
3677         while (ref_count > 0) {
3678                 while (path->slots[0] >= nritems) {
3679                         ret = btrfs_next_leaf(root, path);
3680                         if (ret < 0) {
3681                                 err = ret;
3682                                 goto out;
3683                         }
3684                         if (WARN_ON(ret > 0))
3685                                 goto out;
3686
3687                         leaf = path->nodes[0];
3688                         nritems = btrfs_header_nritems(leaf);
3689                         added = 0;
3690
3691                         if (block_use_full_backref(rc, leaf))
3692                                 counted = 0;
3693                         else
3694                                 counted = 1;
3695                         rb_node = tree_search(blocks, leaf->start);
3696                         if (rb_node) {
3697                                 if (counted)
3698                                         added = 1;
3699                                 else
3700                                         path->slots[0] = nritems;
3701                         }
3702                 }
3703
3704                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3705                 if (WARN_ON(key.objectid != ref_objectid ||
3706                     key.type != BTRFS_EXTENT_DATA_KEY))
3707                         break;
3708
3709                 fi = btrfs_item_ptr(leaf, path->slots[0],
3710                                     struct btrfs_file_extent_item);
3711
3712                 if (btrfs_file_extent_type(leaf, fi) ==
3713                     BTRFS_FILE_EXTENT_INLINE)
3714                         goto next;
3715
3716                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3717                     extent_key->objectid)
3718                         goto next;
3719
3720                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3721                 if (key.offset != ref_offset)
3722                         goto next;
3723
3724                 if (counted)
3725                         ref_count--;
3726                 if (added)
3727                         goto next;
3728
3729                 if (!tree_block_processed(leaf->start, rc)) {
3730                         block = kmalloc(sizeof(*block), GFP_NOFS);
3731                         if (!block) {
3732                                 err = -ENOMEM;
3733                                 break;
3734                         }
3735                         block->bytenr = leaf->start;
3736                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3737                         block->level = 0;
3738                         block->key_ready = 1;
3739                         rb_node = tree_insert(blocks, block->bytenr,
3740                                               &block->rb_node);
3741                         if (rb_node)
3742                                 backref_tree_panic(rb_node, -EEXIST,
3743                                                    block->bytenr);
3744                 }
3745                 if (counted)
3746                         added = 1;
3747                 else
3748                         path->slots[0] = nritems;
3749 next:
3750                 path->slots[0]++;
3751
3752         }
3753 out:
3754         btrfs_free_path(path);
3755         return err;
3756 }
3757
3758 /*
3759  * helper to find all tree blocks that reference a given data extent
3760  */
3761 static noinline_for_stack
3762 int add_data_references(struct reloc_control *rc,
3763                         struct btrfs_key *extent_key,
3764                         struct btrfs_path *path,
3765                         struct rb_root *blocks)
3766 {
3767         struct btrfs_key key;
3768         struct extent_buffer *eb;
3769         struct btrfs_extent_data_ref *dref;
3770         struct btrfs_extent_inline_ref *iref;
3771         unsigned long ptr;
3772         unsigned long end;
3773         u32 blocksize = rc->extent_root->fs_info->nodesize;
3774         int ret = 0;
3775         int err = 0;
3776
3777         eb = path->nodes[0];
3778         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3779         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3780 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3781         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3782                 ptr = end;
3783         else
3784 #endif
3785                 ptr += sizeof(struct btrfs_extent_item);
3786
3787         while (ptr < end) {
3788                 iref = (struct btrfs_extent_inline_ref *)ptr;
3789                 key.type = btrfs_get_extent_inline_ref_type(eb, iref,
3790                                                         BTRFS_REF_TYPE_DATA);
3791                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3792                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3793                         ret = __add_tree_block(rc, key.offset, blocksize,
3794                                                blocks);
3795                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3796                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3797                         ret = find_data_references(rc, extent_key,
3798                                                    eb, dref, blocks);
3799                 } else {
3800                         ret = -EINVAL;
3801                         btrfs_err(rc->extent_root->fs_info,
3802                      "extent %llu slot %d has an invalid inline ref type",
3803                              eb->start, path->slots[0]);
3804                 }
3805                 if (ret) {
3806                         err = ret;
3807                         goto out;
3808                 }
3809                 ptr += btrfs_extent_inline_ref_size(key.type);
3810         }
3811         WARN_ON(ptr > end);
3812
3813         while (1) {
3814                 cond_resched();
3815                 eb = path->nodes[0];
3816                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3817                         ret = btrfs_next_leaf(rc->extent_root, path);
3818                         if (ret < 0) {
3819                                 err = ret;
3820                                 break;
3821                         }
3822                         if (ret > 0)
3823                                 break;
3824                         eb = path->nodes[0];
3825                 }
3826
3827                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3828                 if (key.objectid != extent_key->objectid)
3829                         break;
3830
3831 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3832                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3833                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3834 #else
3835                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3836                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3837 #endif
3838                         ret = __add_tree_block(rc, key.offset, blocksize,
3839                                                blocks);
3840                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3841                         dref = btrfs_item_ptr(eb, path->slots[0],
3842                                               struct btrfs_extent_data_ref);
3843                         ret = find_data_references(rc, extent_key,
3844                                                    eb, dref, blocks);
3845                 } else {
3846                         ret = 0;
3847                 }
3848                 if (ret) {
3849                         err = ret;
3850                         break;
3851                 }
3852                 path->slots[0]++;
3853         }
3854 out:
3855         btrfs_release_path(path);
3856         if (err)
3857                 free_block_list(blocks);
3858         return err;
3859 }
3860
3861 /*
3862  * helper to find next unprocessed extent
3863  */
3864 static noinline_for_stack
3865 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3866                      struct btrfs_key *extent_key)
3867 {
3868         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3869         struct btrfs_key key;
3870         struct extent_buffer *leaf;
3871         u64 start, end, last;
3872         int ret;
3873
3874         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3875         while (1) {
3876                 cond_resched();
3877                 if (rc->search_start >= last) {
3878                         ret = 1;
3879                         break;
3880                 }
3881
3882                 key.objectid = rc->search_start;
3883                 key.type = BTRFS_EXTENT_ITEM_KEY;
3884                 key.offset = 0;
3885
3886                 path->search_commit_root = 1;
3887                 path->skip_locking = 1;
3888                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3889                                         0, 0);
3890                 if (ret < 0)
3891                         break;
3892 next:
3893                 leaf = path->nodes[0];
3894                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3895                         ret = btrfs_next_leaf(rc->extent_root, path);
3896                         if (ret != 0)
3897                                 break;
3898                         leaf = path->nodes[0];
3899                 }
3900
3901                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3902                 if (key.objectid >= last) {
3903                         ret = 1;
3904                         break;
3905                 }
3906
3907                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3908                     key.type != BTRFS_METADATA_ITEM_KEY) {
3909                         path->slots[0]++;
3910                         goto next;
3911                 }
3912
3913                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3914                     key.objectid + key.offset <= rc->search_start) {
3915                         path->slots[0]++;
3916                         goto next;
3917                 }
3918
3919                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3920                     key.objectid + fs_info->nodesize <=
3921                     rc->search_start) {
3922                         path->slots[0]++;
3923                         goto next;
3924                 }
3925
3926                 ret = find_first_extent_bit(&rc->processed_blocks,
3927                                             key.objectid, &start, &end,
3928                                             EXTENT_DIRTY, NULL);
3929
3930                 if (ret == 0 && start <= key.objectid) {
3931                         btrfs_release_path(path);
3932                         rc->search_start = end + 1;
3933                 } else {
3934                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3935                                 rc->search_start = key.objectid + key.offset;
3936                         else
3937                                 rc->search_start = key.objectid +
3938                                         fs_info->nodesize;
3939                         memcpy(extent_key, &key, sizeof(key));
3940                         return 0;
3941                 }
3942         }
3943         btrfs_release_path(path);
3944         return ret;
3945 }
3946
3947 static void set_reloc_control(struct reloc_control *rc)
3948 {
3949         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3950
3951         mutex_lock(&fs_info->reloc_mutex);
3952         fs_info->reloc_ctl = rc;
3953         mutex_unlock(&fs_info->reloc_mutex);
3954 }
3955
3956 static void unset_reloc_control(struct reloc_control *rc)
3957 {
3958         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3959
3960         mutex_lock(&fs_info->reloc_mutex);
3961         fs_info->reloc_ctl = NULL;
3962         mutex_unlock(&fs_info->reloc_mutex);
3963 }
3964
3965 static int check_extent_flags(u64 flags)
3966 {
3967         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3968             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3969                 return 1;
3970         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3971             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3972                 return 1;
3973         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3974             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3975                 return 1;
3976         return 0;
3977 }
3978
3979 static noinline_for_stack
3980 int prepare_to_relocate(struct reloc_control *rc)
3981 {
3982         struct btrfs_trans_handle *trans;
3983         int ret;
3984
3985         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3986                                               BTRFS_BLOCK_RSV_TEMP);
3987         if (!rc->block_rsv)
3988                 return -ENOMEM;
3989
3990         memset(&rc->cluster, 0, sizeof(rc->cluster));
3991         rc->search_start = rc->block_group->key.objectid;
3992         rc->extents_found = 0;
3993         rc->nodes_relocated = 0;
3994         rc->merging_rsv_size = 0;
3995         rc->reserved_bytes = 0;
3996         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3997                               RELOCATION_RESERVED_NODES;
3998         ret = btrfs_block_rsv_refill(rc->extent_root,
3999                                      rc->block_rsv, rc->block_rsv->size,
4000                                      BTRFS_RESERVE_FLUSH_ALL);
4001         if (ret)
4002                 return ret;
4003
4004         rc->create_reloc_tree = 1;
4005         set_reloc_control(rc);
4006
4007         trans = btrfs_join_transaction(rc->extent_root);
4008         if (IS_ERR(trans)) {
4009                 unset_reloc_control(rc);
4010                 /*
4011                  * extent tree is not a ref_cow tree and has no reloc_root to
4012                  * cleanup.  And callers are responsible to free the above
4013                  * block rsv.
4014                  */
4015                 return PTR_ERR(trans);
4016         }
4017         btrfs_commit_transaction(trans);
4018         return 0;
4019 }
4020
4021 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
4022 {
4023         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
4024         struct rb_root blocks = RB_ROOT;
4025         struct btrfs_key key;
4026         struct btrfs_trans_handle *trans = NULL;
4027         struct btrfs_path *path;
4028         struct btrfs_extent_item *ei;
4029         u64 flags;
4030         u32 item_size;
4031         int ret;
4032         int err = 0;
4033         int progress = 0;
4034
4035         path = btrfs_alloc_path();
4036         if (!path)
4037                 return -ENOMEM;
4038         path->reada = READA_FORWARD;
4039
4040         ret = prepare_to_relocate(rc);
4041         if (ret) {
4042                 err = ret;
4043                 goto out_free;
4044         }
4045
4046         while (1) {
4047                 rc->reserved_bytes = 0;
4048                 ret = btrfs_block_rsv_refill(rc->extent_root,
4049                                         rc->block_rsv, rc->block_rsv->size,
4050                                         BTRFS_RESERVE_FLUSH_ALL);
4051                 if (ret) {
4052                         err = ret;
4053                         break;
4054                 }
4055                 progress++;
4056                 trans = btrfs_start_transaction(rc->extent_root, 0);
4057                 if (IS_ERR(trans)) {
4058                         err = PTR_ERR(trans);
4059                         trans = NULL;
4060                         break;
4061                 }
4062 restart:
4063                 if (update_backref_cache(trans, &rc->backref_cache)) {
4064                         btrfs_end_transaction(trans);
4065                         trans = NULL;
4066                         continue;
4067                 }
4068
4069                 ret = find_next_extent(rc, path, &key);
4070                 if (ret < 0)
4071                         err = ret;
4072                 if (ret != 0)
4073                         break;
4074
4075                 rc->extents_found++;
4076
4077                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4078                                     struct btrfs_extent_item);
4079                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4080                 if (item_size >= sizeof(*ei)) {
4081                         flags = btrfs_extent_flags(path->nodes[0], ei);
4082                         ret = check_extent_flags(flags);
4083                         BUG_ON(ret);
4084
4085                 } else {
4086 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4087                         u64 ref_owner;
4088                         int path_change = 0;
4089
4090                         BUG_ON(item_size !=
4091                                sizeof(struct btrfs_extent_item_v0));
4092                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
4093                                                   &path_change);
4094                         if (ret < 0) {
4095                                 err = ret;
4096                                 break;
4097                         }
4098                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
4099                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
4100                         else
4101                                 flags = BTRFS_EXTENT_FLAG_DATA;
4102
4103                         if (path_change) {
4104                                 btrfs_release_path(path);
4105
4106                                 path->search_commit_root = 1;
4107                                 path->skip_locking = 1;
4108                                 ret = btrfs_search_slot(NULL, rc->extent_root,
4109                                                         &key, path, 0, 0);
4110                                 if (ret < 0) {
4111                                         err = ret;
4112                                         break;
4113                                 }
4114                                 BUG_ON(ret > 0);
4115                         }
4116 #else
4117                         BUG();
4118 #endif
4119                 }
4120
4121                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4122                         ret = add_tree_block(rc, &key, path, &blocks);
4123                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4124                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4125                         ret = add_data_references(rc, &key, path, &blocks);
4126                 } else {
4127                         btrfs_release_path(path);
4128                         ret = 0;
4129                 }
4130                 if (ret < 0) {
4131                         err = ret;
4132                         break;
4133                 }
4134
4135                 if (!RB_EMPTY_ROOT(&blocks)) {
4136                         ret = relocate_tree_blocks(trans, rc, &blocks);
4137                         if (ret < 0) {
4138                                 /*
4139                                  * if we fail to relocate tree blocks, force to update
4140                                  * backref cache when committing transaction.
4141                                  */
4142                                 rc->backref_cache.last_trans = trans->transid - 1;
4143
4144                                 if (ret != -EAGAIN) {
4145                                         err = ret;
4146                                         break;
4147                                 }
4148                                 rc->extents_found--;
4149                                 rc->search_start = key.objectid;
4150                         }
4151                 }
4152
4153                 btrfs_end_transaction_throttle(trans);
4154                 btrfs_btree_balance_dirty(fs_info);
4155                 trans = NULL;
4156
4157                 if (rc->stage == MOVE_DATA_EXTENTS &&
4158                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4159                         rc->found_file_extent = 1;
4160                         ret = relocate_data_extent(rc->data_inode,
4161                                                    &key, &rc->cluster);
4162                         if (ret < 0) {
4163                                 err = ret;
4164                                 break;
4165                         }
4166                 }
4167         }
4168         if (trans && progress && err == -ENOSPC) {
4169                 ret = btrfs_force_chunk_alloc(trans, fs_info,
4170                                               rc->block_group->flags);
4171                 if (ret == 1) {
4172                         err = 0;
4173                         progress = 0;
4174                         goto restart;
4175                 }
4176         }
4177
4178         btrfs_release_path(path);
4179         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4180
4181         if (trans) {
4182                 btrfs_end_transaction_throttle(trans);
4183                 btrfs_btree_balance_dirty(fs_info);
4184         }
4185
4186         if (!err) {
4187                 ret = relocate_file_extent_cluster(rc->data_inode,
4188                                                    &rc->cluster);
4189                 if (ret < 0)
4190                         err = ret;
4191         }
4192
4193         rc->create_reloc_tree = 0;
4194         set_reloc_control(rc);
4195
4196         backref_cache_cleanup(&rc->backref_cache);
4197         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4198
4199         err = prepare_to_merge(rc, err);
4200
4201         merge_reloc_roots(rc);
4202
4203         rc->merge_reloc_tree = 0;
4204         unset_reloc_control(rc);
4205         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4206
4207         /* get rid of pinned extents */
4208         trans = btrfs_join_transaction(rc->extent_root);
4209         if (IS_ERR(trans)) {
4210                 err = PTR_ERR(trans);
4211                 goto out_free;
4212         }
4213         btrfs_commit_transaction(trans);
4214 out_free:
4215         btrfs_free_block_rsv(fs_info, rc->block_rsv);
4216         btrfs_free_path(path);
4217         return err;
4218 }
4219
4220 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4221                                  struct btrfs_root *root, u64 objectid)
4222 {
4223         struct btrfs_path *path;
4224         struct btrfs_inode_item *item;
4225         struct extent_buffer *leaf;
4226         int ret;
4227
4228         path = btrfs_alloc_path();
4229         if (!path)
4230                 return -ENOMEM;
4231
4232         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4233         if (ret)
4234                 goto out;
4235
4236         leaf = path->nodes[0];
4237         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4238         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
4239         btrfs_set_inode_generation(leaf, item, 1);
4240         btrfs_set_inode_size(leaf, item, 0);
4241         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4242         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4243                                           BTRFS_INODE_PREALLOC);
4244         btrfs_mark_buffer_dirty(leaf);
4245 out:
4246         btrfs_free_path(path);
4247         return ret;
4248 }
4249
4250 /*
4251  * helper to create inode for data relocation.
4252  * the inode is in data relocation tree and its link count is 0
4253  */
4254 static noinline_for_stack
4255 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4256                                  struct btrfs_block_group_cache *group)
4257 {
4258         struct inode *inode = NULL;
4259         struct btrfs_trans_handle *trans;
4260         struct btrfs_root *root;
4261         struct btrfs_key key;
4262         u64 objectid;
4263         int err = 0;
4264
4265         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4266         if (IS_ERR(root))
4267                 return ERR_CAST(root);
4268
4269         trans = btrfs_start_transaction(root, 6);
4270         if (IS_ERR(trans))
4271                 return ERR_CAST(trans);
4272
4273         err = btrfs_find_free_objectid(root, &objectid);
4274         if (err)
4275                 goto out;
4276
4277         err = __insert_orphan_inode(trans, root, objectid);
4278         BUG_ON(err);
4279
4280         key.objectid = objectid;
4281         key.type = BTRFS_INODE_ITEM_KEY;
4282         key.offset = 0;
4283         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4284         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4285         BTRFS_I(inode)->index_cnt = group->key.objectid;
4286
4287         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4288 out:
4289         btrfs_end_transaction(trans);
4290         btrfs_btree_balance_dirty(fs_info);
4291         if (err) {
4292                 if (inode)
4293                         iput(inode);
4294                 inode = ERR_PTR(err);
4295         }
4296         return inode;
4297 }
4298
4299 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4300 {
4301         struct reloc_control *rc;
4302
4303         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4304         if (!rc)
4305                 return NULL;
4306
4307         INIT_LIST_HEAD(&rc->reloc_roots);
4308         backref_cache_init(&rc->backref_cache);
4309         mapping_tree_init(&rc->reloc_root_tree);
4310         extent_io_tree_init(&rc->processed_blocks, NULL);
4311         return rc;
4312 }
4313
4314 /*
4315  * Print the block group being relocated
4316  */
4317 static void describe_relocation(struct btrfs_fs_info *fs_info,
4318                                 struct btrfs_block_group_cache *block_group)
4319 {
4320         char buf[128];          /* prefixed by a '|' that'll be dropped */
4321         u64 flags = block_group->flags;
4322
4323         /* Shouldn't happen */
4324         if (!flags) {
4325                 strcpy(buf, "|NONE");
4326         } else {
4327                 char *bp = buf;
4328
4329 #define DESCRIBE_FLAG(f, d) \
4330                 if (flags & BTRFS_BLOCK_GROUP_##f) { \
4331                         bp += snprintf(bp, buf - bp + sizeof(buf), "|%s", d); \
4332                         flags &= ~BTRFS_BLOCK_GROUP_##f; \
4333                 }
4334                 DESCRIBE_FLAG(DATA,     "data");
4335                 DESCRIBE_FLAG(SYSTEM,   "system");
4336                 DESCRIBE_FLAG(METADATA, "metadata");
4337                 DESCRIBE_FLAG(RAID0,    "raid0");
4338                 DESCRIBE_FLAG(RAID1,    "raid1");
4339                 DESCRIBE_FLAG(DUP,      "dup");
4340                 DESCRIBE_FLAG(RAID10,   "raid10");
4341                 DESCRIBE_FLAG(RAID5,    "raid5");
4342                 DESCRIBE_FLAG(RAID6,    "raid6");
4343                 if (flags)
4344                         snprintf(buf, buf - bp + sizeof(buf), "|0x%llx", flags);
4345 #undef DESCRIBE_FLAG
4346         }
4347
4348         btrfs_info(fs_info,
4349                    "relocating block group %llu flags %s",
4350                    block_group->key.objectid, buf + 1);
4351 }
4352
4353 /*
4354  * function to relocate all extents in a block group.
4355  */
4356 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4357 {
4358         struct btrfs_root *extent_root = fs_info->extent_root;
4359         struct reloc_control *rc;
4360         struct inode *inode;
4361         struct btrfs_path *path;
4362         int ret;
4363         int rw = 0;
4364         int err = 0;
4365
4366         rc = alloc_reloc_control(fs_info);
4367         if (!rc)
4368                 return -ENOMEM;
4369
4370         rc->extent_root = extent_root;
4371
4372         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4373         BUG_ON(!rc->block_group);
4374
4375         ret = btrfs_inc_block_group_ro(fs_info, rc->block_group);
4376         if (ret) {
4377                 err = ret;
4378                 goto out;
4379         }
4380         rw = 1;
4381
4382         path = btrfs_alloc_path();
4383         if (!path) {
4384                 err = -ENOMEM;
4385                 goto out;
4386         }
4387
4388         inode = lookup_free_space_inode(fs_info, rc->block_group, path);
4389         btrfs_free_path(path);
4390
4391         if (!IS_ERR(inode))
4392                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4393         else
4394                 ret = PTR_ERR(inode);
4395
4396         if (ret && ret != -ENOENT) {
4397                 err = ret;
4398                 goto out;
4399         }
4400
4401         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4402         if (IS_ERR(rc->data_inode)) {
4403                 err = PTR_ERR(rc->data_inode);
4404                 rc->data_inode = NULL;
4405                 goto out;
4406         }
4407
4408         describe_relocation(fs_info, rc->block_group);
4409
4410         btrfs_wait_block_group_reservations(rc->block_group);
4411         btrfs_wait_nocow_writers(rc->block_group);
4412         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4413                                  rc->block_group->key.objectid,
4414                                  rc->block_group->key.offset);
4415
4416         while (1) {
4417                 mutex_lock(&fs_info->cleaner_mutex);
4418                 ret = relocate_block_group(rc);
4419                 mutex_unlock(&fs_info->cleaner_mutex);
4420                 if (ret < 0)
4421                         err = ret;
4422
4423                 /*
4424                  * We may have gotten ENOSPC after we already dirtied some
4425                  * extents.  If writeout happens while we're relocating a
4426                  * different block group we could end up hitting the
4427                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4428                  * btrfs_reloc_cow_block.  Make sure we write everything out
4429                  * properly so we don't trip over this problem, and then break
4430                  * out of the loop if we hit an error.
4431                  */
4432                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4433                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4434                                                        (u64)-1);
4435                         if (ret)
4436                                 err = ret;
4437                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4438                                                  0, -1);
4439                         rc->stage = UPDATE_DATA_PTRS;
4440                 }
4441
4442                 if (err < 0)
4443                         goto out;
4444
4445                 if (rc->extents_found == 0)
4446                         break;
4447
4448                 btrfs_info(fs_info, "found %llu extents", rc->extents_found);
4449
4450         }
4451
4452         WARN_ON(rc->block_group->pinned > 0);
4453         WARN_ON(rc->block_group->reserved > 0);
4454         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4455 out:
4456         if (err && rw)
4457                 btrfs_dec_block_group_ro(rc->block_group);
4458         iput(rc->data_inode);
4459         btrfs_put_block_group(rc->block_group);
4460         kfree(rc);
4461         return err;
4462 }
4463
4464 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4465 {
4466         struct btrfs_fs_info *fs_info = root->fs_info;
4467         struct btrfs_trans_handle *trans;
4468         int ret, err;
4469
4470         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4471         if (IS_ERR(trans))
4472                 return PTR_ERR(trans);
4473
4474         memset(&root->root_item.drop_progress, 0,
4475                 sizeof(root->root_item.drop_progress));
4476         root->root_item.drop_level = 0;
4477         btrfs_set_root_refs(&root->root_item, 0);
4478         ret = btrfs_update_root(trans, fs_info->tree_root,
4479                                 &root->root_key, &root->root_item);
4480
4481         err = btrfs_end_transaction(trans);
4482         if (err)
4483                 return err;
4484         return ret;
4485 }
4486
4487 /*
4488  * recover relocation interrupted by system crash.
4489  *
4490  * this function resumes merging reloc trees with corresponding fs trees.
4491  * this is important for keeping the sharing of tree blocks
4492  */
4493 int btrfs_recover_relocation(struct btrfs_root *root)
4494 {
4495         struct btrfs_fs_info *fs_info = root->fs_info;
4496         LIST_HEAD(reloc_roots);
4497         struct btrfs_key key;
4498         struct btrfs_root *fs_root;
4499         struct btrfs_root *reloc_root;
4500         struct btrfs_path *path;
4501         struct extent_buffer *leaf;
4502         struct reloc_control *rc = NULL;
4503         struct btrfs_trans_handle *trans;
4504         int ret;
4505         int err = 0;
4506
4507         path = btrfs_alloc_path();
4508         if (!path)
4509                 return -ENOMEM;
4510         path->reada = READA_BACK;
4511
4512         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4513         key.type = BTRFS_ROOT_ITEM_KEY;
4514         key.offset = (u64)-1;
4515
4516         while (1) {
4517                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4518                                         path, 0, 0);
4519                 if (ret < 0) {
4520                         err = ret;
4521                         goto out;
4522                 }
4523                 if (ret > 0) {
4524                         if (path->slots[0] == 0)
4525                                 break;
4526                         path->slots[0]--;
4527                 }
4528                 leaf = path->nodes[0];
4529                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4530                 btrfs_release_path(path);
4531
4532                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4533                     key.type != BTRFS_ROOT_ITEM_KEY)
4534                         break;
4535
4536                 reloc_root = btrfs_read_fs_root(root, &key);
4537                 if (IS_ERR(reloc_root)) {
4538                         err = PTR_ERR(reloc_root);
4539                         goto out;
4540                 }
4541
4542                 list_add(&reloc_root->root_list, &reloc_roots);
4543
4544                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4545                         fs_root = read_fs_root(fs_info,
4546                                                reloc_root->root_key.offset);
4547                         if (IS_ERR(fs_root)) {
4548                                 ret = PTR_ERR(fs_root);
4549                                 if (ret != -ENOENT) {
4550                                         err = ret;
4551                                         goto out;
4552                                 }
4553                                 ret = mark_garbage_root(reloc_root);
4554                                 if (ret < 0) {
4555                                         err = ret;
4556                                         goto out;
4557                                 }
4558                         }
4559                 }
4560
4561                 if (key.offset == 0)
4562                         break;
4563
4564                 key.offset--;
4565         }
4566         btrfs_release_path(path);
4567
4568         if (list_empty(&reloc_roots))
4569                 goto out;
4570
4571         rc = alloc_reloc_control(fs_info);
4572         if (!rc) {
4573                 err = -ENOMEM;
4574                 goto out;
4575         }
4576
4577         rc->extent_root = fs_info->extent_root;
4578
4579         set_reloc_control(rc);
4580
4581         trans = btrfs_join_transaction(rc->extent_root);
4582         if (IS_ERR(trans)) {
4583                 unset_reloc_control(rc);
4584                 err = PTR_ERR(trans);
4585                 goto out_free;
4586         }
4587
4588         rc->merge_reloc_tree = 1;
4589
4590         while (!list_empty(&reloc_roots)) {
4591                 reloc_root = list_entry(reloc_roots.next,
4592                                         struct btrfs_root, root_list);
4593                 list_del(&reloc_root->root_list);
4594
4595                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4596                         list_add_tail(&reloc_root->root_list,
4597                                       &rc->reloc_roots);
4598                         continue;
4599                 }
4600
4601                 fs_root = read_fs_root(fs_info, reloc_root->root_key.offset);
4602                 if (IS_ERR(fs_root)) {
4603                         err = PTR_ERR(fs_root);
4604                         list_add_tail(&reloc_root->root_list, &reloc_roots);
4605                         goto out_free;
4606                 }
4607
4608                 err = __add_reloc_root(reloc_root);
4609                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4610                 fs_root->reloc_root = reloc_root;
4611         }
4612
4613         err = btrfs_commit_transaction(trans);
4614         if (err)
4615                 goto out_free;
4616
4617         merge_reloc_roots(rc);
4618
4619         unset_reloc_control(rc);
4620
4621         trans = btrfs_join_transaction(rc->extent_root);
4622         if (IS_ERR(trans)) {
4623                 err = PTR_ERR(trans);
4624                 goto out_free;
4625         }
4626         err = btrfs_commit_transaction(trans);
4627 out_free:
4628         kfree(rc);
4629 out:
4630         if (!list_empty(&reloc_roots))
4631                 free_reloc_roots(&reloc_roots);
4632
4633         btrfs_free_path(path);
4634
4635         if (err == 0) {
4636                 /* cleanup orphan inode in data relocation tree */
4637                 fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4638                 if (IS_ERR(fs_root))
4639                         err = PTR_ERR(fs_root);
4640                 else
4641                         err = btrfs_orphan_cleanup(fs_root);
4642         }
4643         return err;
4644 }
4645
4646 /*
4647  * helper to add ordered checksum for data relocation.
4648  *
4649  * cloning checksum properly handles the nodatasum extents.
4650  * it also saves CPU time to re-calculate the checksum.
4651  */
4652 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4653 {
4654         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4655         struct btrfs_ordered_sum *sums;
4656         struct btrfs_ordered_extent *ordered;
4657         int ret;
4658         u64 disk_bytenr;
4659         u64 new_bytenr;
4660         LIST_HEAD(list);
4661
4662         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4663         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4664
4665         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4666         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4667                                        disk_bytenr + len - 1, &list, 0);
4668         if (ret)
4669                 goto out;
4670
4671         while (!list_empty(&list)) {
4672                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4673                 list_del_init(&sums->list);
4674
4675                 /*
4676                  * We need to offset the new_bytenr based on where the csum is.
4677                  * We need to do this because we will read in entire prealloc
4678                  * extents but we may have written to say the middle of the
4679                  * prealloc extent, so we need to make sure the csum goes with
4680                  * the right disk offset.
4681                  *
4682                  * We can do this because the data reloc inode refers strictly
4683                  * to the on disk bytes, so we don't have to worry about
4684                  * disk_len vs real len like with real inodes since it's all
4685                  * disk length.
4686                  */
4687                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4688                 sums->bytenr = new_bytenr;
4689
4690                 btrfs_add_ordered_sum(inode, ordered, sums);
4691         }
4692 out:
4693         btrfs_put_ordered_extent(ordered);
4694         return ret;
4695 }
4696
4697 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4698                           struct btrfs_root *root, struct extent_buffer *buf,
4699                           struct extent_buffer *cow)
4700 {
4701         struct btrfs_fs_info *fs_info = root->fs_info;
4702         struct reloc_control *rc;
4703         struct backref_node *node;
4704         int first_cow = 0;
4705         int level;
4706         int ret = 0;
4707
4708         rc = fs_info->reloc_ctl;
4709         if (!rc)
4710                 return 0;
4711
4712         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4713                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4714
4715         level = btrfs_header_level(buf);
4716         if (btrfs_header_generation(buf) <=
4717             btrfs_root_last_snapshot(&root->root_item))
4718                 first_cow = 1;
4719
4720         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4721             rc->create_reloc_tree) {
4722                 WARN_ON(!first_cow && level == 0);
4723
4724                 node = rc->backref_cache.path[level];
4725                 BUG_ON(node->bytenr != buf->start &&
4726                        node->new_bytenr != buf->start);
4727
4728                 drop_node_buffer(node);
4729                 extent_buffer_get(cow);
4730                 node->eb = cow;
4731                 node->new_bytenr = cow->start;
4732
4733                 if (!node->pending) {
4734                         list_move_tail(&node->list,
4735                                        &rc->backref_cache.pending[level]);
4736                         node->pending = 1;
4737                 }
4738
4739                 if (first_cow)
4740                         __mark_block_processed(rc, node);
4741
4742                 if (first_cow && level > 0)
4743                         rc->nodes_relocated += buf->len;
4744         }
4745
4746         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4747                 ret = replace_file_extents(trans, rc, root, cow);
4748         return ret;
4749 }
4750
4751 /*
4752  * called before creating snapshot. it calculates metadata reservation
4753  * required for relocating tree blocks in the snapshot
4754  */
4755 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4756                               u64 *bytes_to_reserve)
4757 {
4758         struct btrfs_root *root;
4759         struct reloc_control *rc;
4760
4761         root = pending->root;
4762         if (!root->reloc_root)
4763                 return;
4764
4765         rc = root->fs_info->reloc_ctl;
4766         if (!rc->merge_reloc_tree)
4767                 return;
4768
4769         root = root->reloc_root;
4770         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4771         /*
4772          * relocation is in the stage of merging trees. the space
4773          * used by merging a reloc tree is twice the size of
4774          * relocated tree nodes in the worst case. half for cowing
4775          * the reloc tree, half for cowing the fs tree. the space
4776          * used by cowing the reloc tree will be freed after the
4777          * tree is dropped. if we create snapshot, cowing the fs
4778          * tree may use more space than it frees. so we need
4779          * reserve extra space.
4780          */
4781         *bytes_to_reserve += rc->nodes_relocated;
4782 }
4783
4784 /*
4785  * called after snapshot is created. migrate block reservation
4786  * and create reloc root for the newly created snapshot
4787  */
4788 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4789                                struct btrfs_pending_snapshot *pending)
4790 {
4791         struct btrfs_root *root = pending->root;
4792         struct btrfs_root *reloc_root;
4793         struct btrfs_root *new_root;
4794         struct reloc_control *rc;
4795         int ret;
4796
4797         if (!root->reloc_root)
4798                 return 0;
4799
4800         rc = root->fs_info->reloc_ctl;
4801         rc->merging_rsv_size += rc->nodes_relocated;
4802
4803         if (rc->merge_reloc_tree) {
4804                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4805                                               rc->block_rsv,
4806                                               rc->nodes_relocated, 1);
4807                 if (ret)
4808                         return ret;
4809         }
4810
4811         new_root = pending->snap;
4812         reloc_root = create_reloc_root(trans, root->reloc_root,
4813                                        new_root->root_key.objectid);
4814         if (IS_ERR(reloc_root))
4815                 return PTR_ERR(reloc_root);
4816
4817         ret = __add_reloc_root(reloc_root);
4818         BUG_ON(ret < 0);
4819         new_root->reloc_root = reloc_root;
4820
4821         if (rc->create_reloc_tree)
4822                 ret = clone_backref_node(trans, rc, root, reloc_root);
4823         return ret;
4824 }