fs/btrfs/relocation.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2009 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include "ctree.h"
  13 #include "disk-io.h"
  14 #include "transaction.h"
  15 #include "volumes.h"
  16 #include "locking.h"
  17 #include "btrfs_inode.h"
  18 #include "async-thread.h"
  19 #include "free-space-cache.h"
  20 #include "inode-map.h"
  21 #include "qgroup.h"
  22 #include "print-tree.h"
  23 #include "delalloc-space.h"
  24 #include "block-group.h"
  25
  26 /*
  27  * backref_node, mapping_node and tree_block start with this
  28  */
  29 struct tree_entry {
  30         struct rb_node rb_node;
  31         u64 bytenr;
  32 };
  33
  34 /*
  35  * present a tree block in the backref cache
  36  */
  37 struct backref_node {
  38         struct rb_node rb_node;
  39         u64 bytenr;
  40
  41         u64 new_bytenr;
  42         /* objectid of tree block owner, can be not uptodate */
  43         u64 owner;
  44         /* link to pending, changed or detached list */
  45         struct list_head list;
  46         /* list of upper level blocks reference this block */
  47         struct list_head upper;
  48         /* list of child blocks in the cache */
  49         struct list_head lower;
  50         /* NULL if this node is not tree root */
  51         struct btrfs_root *root;
  52         /* extent buffer got by COW the block */
  53         struct extent_buffer *eb;
  54         /* level of tree block */
  55         unsigned int level:8;
  56         /* is the block in non-reference counted tree */
  57         unsigned int cowonly:1;
  58         /* 1 if no child node in the cache */
  59         unsigned int lowest:1;
  60         /* is the extent buffer locked */
  61         unsigned int locked:1;
  62         /* has the block been processed */
  63         unsigned int processed:1;
  64         /* have backrefs of this block been checked */
  65         unsigned int checked:1;
  66         /*
  67          * 1 if corresponding block has been cowed but some upper
  68          * level block pointers may not point to the new location
  69          */
  70         unsigned int pending:1;
  71         /*
  72          * 1 if the backref node isn't connected to any other
  73          * backref node.
  74          */
  75         unsigned int detached:1;
  76 };
  77
  78 /*
  79  * present a block pointer in the backref cache
  80  */
  81 struct backref_edge {
  82         struct list_head list[2];
  83         struct backref_node *node[2];
  84 };
  85
  86 #define LOWER   0
  87 #define UPPER   1
  88 #define RELOCATION_RESERVED_NODES       256
  89
  90 struct backref_cache {
  91         /* red black tree of all backref nodes in the cache */
  92         struct rb_root rb_root;
  93         /* for passing backref nodes to btrfs_reloc_cow_block */
  94         struct backref_node *path[BTRFS_MAX_LEVEL];
  95         /*
  96          * list of blocks that have been cowed but some block
  97          * pointers in upper level blocks may not reflect the
  98          * new location
  99          */
 100         struct list_head pending[BTRFS_MAX_LEVEL];
 101         /* list of backref nodes with no child node */
 102         struct list_head leaves;
 103         /* list of blocks that have been cowed in current transaction */
 104         struct list_head changed;
 105         /* list of detached backref node. */
 106         struct list_head detached;
 107
 108         u64 last_trans;
 109
 110         int nr_nodes;
 111         int nr_edges;
 112 };
 113
 114 /*
 115  * map address of tree root to tree
 116  */
 117 struct mapping_node {
 118         struct rb_node rb_node;
 119         u64 bytenr;
 120         void *data;
 121 };
 122
 123 struct mapping_tree {
 124         struct rb_root rb_root;
 125         spinlock_t lock;
 126 };
 127
 128 /*
 129  * present a tree block to process
 130  */
 131 struct tree_block {
 132         struct rb_node rb_node;
 133         u64 bytenr;
 134         struct btrfs_key key;
 135         unsigned int level:8;
 136         unsigned int key_ready:1;
 137 };
 138
 139 #define MAX_EXTENTS 128
 140
 141 struct file_extent_cluster {
 142         u64 start;
 143         u64 end;
 144         u64 boundary[MAX_EXTENTS];
 145         unsigned int nr;
 146 };
 147
 148 struct reloc_control {
 149         /* block group to relocate */
 150         struct btrfs_block_group_cache *block_group;
 151         /* extent tree */
 152         struct btrfs_root *extent_root;
 153         /* inode for moving data */
 154         struct inode *data_inode;
 155
 156         struct btrfs_block_rsv *block_rsv;
 157
 158         struct backref_cache backref_cache;
 159
 160         struct file_extent_cluster cluster;
 161         /* tree blocks have been processed */
 162         struct extent_io_tree processed_blocks;
 163         /* map start of tree root to corresponding reloc tree */
 164         struct mapping_tree reloc_root_tree;
 165         /* list of reloc trees */
 166         struct list_head reloc_roots;
 167         /* list of subvolume trees that get relocated */
 168         struct list_head dirty_subvol_roots;
 169         /* size of metadata reservation for merging reloc trees */
 170         u64 merging_rsv_size;
 171         /* size of relocated tree nodes */
 172         u64 nodes_relocated;
 173         /* reserved size for block group relocation*/
 174         u64 reserved_bytes;
 175
 176         u64 search_start;
 177         u64 extents_found;
 178
 179         unsigned int stage:8;
 180         unsigned int create_reloc_tree:1;
 181         unsigned int merge_reloc_tree:1;
 182         unsigned int found_file_extent:1;
 183 };
 184
 185 /* stages of data relocation */
 186 #define MOVE_DATA_EXTENTS       0
 187 #define UPDATE_DATA_PTRS        1
 188
 189 static void remove_backref_node(struct backref_cache *cache,
 190                                 struct backref_node *node);
 191 static void __mark_block_processed(struct reloc_control *rc,
 192                                    struct backref_node *node);
 193
 194 static void mapping_tree_init(struct mapping_tree *tree)
 195 {
 196         tree->rb_root = RB_ROOT;
 197         spin_lock_init(&tree->lock);
 198 }
 199
 200 static void backref_cache_init(struct backref_cache *cache)
 201 {
 202         int i;
 203         cache->rb_root = RB_ROOT;
 204         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 205                 INIT_LIST_HEAD(&cache->pending[i]);
 206         INIT_LIST_HEAD(&cache->changed);
 207         INIT_LIST_HEAD(&cache->detached);
 208         INIT_LIST_HEAD(&cache->leaves);
 209 }
 210
 211 static void backref_cache_cleanup(struct backref_cache *cache)
 212 {
 213         struct backref_node *node;
 214         int i;
 215
 216         while (!list_empty(&cache->detached)) {
 217                 node = list_entry(cache->detached.next,
 218                                   struct backref_node, list);
 219                 remove_backref_node(cache, node);
 220         }
 221
 222         while (!list_empty(&cache->leaves)) {
 223                 node = list_entry(cache->leaves.next,
 224                                   struct backref_node, lower);
 225                 remove_backref_node(cache, node);
 226         }
 227
 228         cache->last_trans = 0;
 229
 230         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 231                 ASSERT(list_empty(&cache->pending[i]));
 232         ASSERT(list_empty(&cache->changed));
 233         ASSERT(list_empty(&cache->detached));
 234         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
 235         ASSERT(!cache->nr_nodes);
 236         ASSERT(!cache->nr_edges);
 237 }
 238
 239 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 240 {
 241         struct backref_node *node;
 242
 243         node = kzalloc(sizeof(*node), GFP_NOFS);
 244         if (node) {
 245                 INIT_LIST_HEAD(&node->list);
 246                 INIT_LIST_HEAD(&node->upper);
 247                 INIT_LIST_HEAD(&node->lower);
 248                 RB_CLEAR_NODE(&node->rb_node);
 249                 cache->nr_nodes++;
 250         }
 251         return node;
 252 }
 253
 254 static void free_backref_node(struct backref_cache *cache,
 255                               struct backref_node *node)
 256 {
 257         if (node) {
 258                 cache->nr_nodes--;
 259                 kfree(node);
 260         }
 261 }
 262
 263 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 264 {
 265         struct backref_edge *edge;
 266
 267         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 268         if (edge)
 269                 cache->nr_edges++;
 270         return edge;
 271 }
 272
 273 static void free_backref_edge(struct backref_cache *cache,
 274                               struct backref_edge *edge)
 275 {
 276         if (edge) {
 277                 cache->nr_edges--;
 278                 kfree(edge);
 279         }
 280 }
 281
 282 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 283                                    struct rb_node *node)
 284 {
 285         struct rb_node **p = &root->rb_node;
 286         struct rb_node *parent = NULL;
 287         struct tree_entry *entry;
 288
 289         while (*p) {
 290                 parent = *p;
 291                 entry = rb_entry(parent, struct tree_entry, rb_node);
 292
 293                 if (bytenr < entry->bytenr)
 294                         p = &(*p)->rb_left;
 295                 else if (bytenr > entry->bytenr)
 296                         p = &(*p)->rb_right;
 297                 else
 298                         return parent;
 299         }
 300
 301         rb_link_node(node, parent, p);
 302         rb_insert_color(node, root);
 303         return NULL;
 304 }
 305
 306 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 307 {
 308         struct rb_node *n = root->rb_node;
 309         struct tree_entry *entry;
 310
 311         while (n) {
 312                 entry = rb_entry(n, struct tree_entry, rb_node);
 313
 314                 if (bytenr < entry->bytenr)
 315                         n = n->rb_left;
 316                 else if (bytenr > entry->bytenr)
 317                         n = n->rb_right;
 318                 else
 319                         return n;
 320         }
 321         return NULL;
 322 }
 323
 324 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 325 {
 326
 327         struct btrfs_fs_info *fs_info = NULL;
 328         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 329                                               rb_node);
 330         if (bnode->root)
 331                 fs_info = bnode->root->fs_info;
 332         btrfs_panic(fs_info, errno,
 333                     "Inconsistency in backref cache found at offset %llu",
 334                     bytenr);
 335 }
 336
 337 /*
 338  * walk up backref nodes until reach node presents tree root
 339  */
 340 static struct backref_node *walk_up_backref(struct backref_node *node,
 341                                             struct backref_edge *edges[],
 342                                             int *index)
 343 {
 344         struct backref_edge *edge;
 345         int idx = *index;
 346
 347         while (!list_empty(&node->upper)) {
 348                 edge = list_entry(node->upper.next,
 349                                   struct backref_edge, list[LOWER]);
 350                 edges[idx++] = edge;
 351                 node = edge->node[UPPER];
 352         }
 353         BUG_ON(node->detached);
 354         *index = idx;
 355         return node;
 356 }
 357
 358 /*
 359  * walk down backref nodes to find start of next reference path
 360  */
 361 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 362                                               int *index)
 363 {
 364         struct backref_edge *edge;
 365         struct backref_node *lower;
 366         int idx = *index;
 367
 368         while (idx > 0) {
 369                 edge = edges[idx - 1];
 370                 lower = edge->node[LOWER];
 371                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 372                         idx--;
 373                         continue;
 374                 }
 375                 edge = list_entry(edge->list[LOWER].next,
 376                                   struct backref_edge, list[LOWER]);
 377                 edges[idx - 1] = edge;
 378                 *index = idx;
 379                 return edge->node[UPPER];
 380         }
 381         *index = 0;
 382         return NULL;
 383 }
 384
 385 static void unlock_node_buffer(struct backref_node *node)
 386 {
 387         if (node->locked) {
 388                 btrfs_tree_unlock(node->eb);
 389                 node->locked = 0;
 390         }
 391 }
 392
 393 static void drop_node_buffer(struct backref_node *node)
 394 {
 395         if (node->eb) {
 396                 unlock_node_buffer(node);
 397                 free_extent_buffer(node->eb);
 398                 node->eb = NULL;
 399         }
 400 }
 401
 402 static void drop_backref_node(struct backref_cache *tree,
 403                               struct backref_node *node)
 404 {
 405         BUG_ON(!list_empty(&node->upper));
 406
 407         drop_node_buffer(node);
 408         list_del(&node->list);
 409         list_del(&node->lower);
 410         if (!RB_EMPTY_NODE(&node->rb_node))
 411                 rb_erase(&node->rb_node, &tree->rb_root);
 412         free_backref_node(tree, node);
 413 }
 414
 415 /*
 416  * remove a backref node from the backref cache
 417  */
 418 static void remove_backref_node(struct backref_cache *cache,
 419                                 struct backref_node *node)
 420 {
 421         struct backref_node *upper;
 422         struct backref_edge *edge;
 423
 424         if (!node)
 425                 return;
 426
 427         BUG_ON(!node->lowest && !node->detached);
 428         while (!list_empty(&node->upper)) {
 429                 edge = list_entry(node->upper.next, struct backref_edge,
 430                                   list[LOWER]);
 431                 upper = edge->node[UPPER];
 432                 list_del(&edge->list[LOWER]);
 433                 list_del(&edge->list[UPPER]);
 434                 free_backref_edge(cache, edge);
 435
 436                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 437                         BUG_ON(!list_empty(&node->upper));
 438                         drop_backref_node(cache, node);
 439                         node = upper;
 440                         node->lowest = 1;
 441                         continue;
 442                 }
 443                 /*
 444                  * add the node to leaf node list if no other
 445                  * child block cached.
 446                  */
 447                 if (list_empty(&upper->lower)) {
 448                         list_add_tail(&upper->lower, &cache->leaves);
 449                         upper->lowest = 1;
 450                 }
 451         }
 452
 453         drop_backref_node(cache, node);
 454 }
 455
 456 static void update_backref_node(struct backref_cache *cache,
 457                                 struct backref_node *node, u64 bytenr)
 458 {
 459         struct rb_node *rb_node;
 460         rb_erase(&node->rb_node, &cache->rb_root);
 461         node->bytenr = bytenr;
 462         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 463         if (rb_node)
 464                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 465 }
 466
 467 /*
 468  * update backref cache after a transaction commit
 469  */
 470 static int update_backref_cache(struct btrfs_trans_handle *trans,
 471                                 struct backref_cache *cache)
 472 {
 473         struct backref_node *node;
 474         int level = 0;
 475
 476         if (cache->last_trans == 0) {
 477                 cache->last_trans = trans->transid;
 478                 return 0;
 479         }
 480
 481         if (cache->last_trans == trans->transid)
 482                 return 0;
 483
 484         /*
 485          * detached nodes are used to avoid unnecessary backref
 486          * lookup. transaction commit changes the extent tree.
 487          * so the detached nodes are no longer useful.
 488          */
 489         while (!list_empty(&cache->detached)) {
 490                 node = list_entry(cache->detached.next,
 491                                   struct backref_node, list);
 492                 remove_backref_node(cache, node);
 493         }
 494
 495         while (!list_empty(&cache->changed)) {
 496                 node = list_entry(cache->changed.next,
 497                                   struct backref_node, list);
 498                 list_del_init(&node->list);
 499                 BUG_ON(node->pending);
 500                 update_backref_node(cache, node, node->new_bytenr);
 501         }
 502
 503         /*
 504          * some nodes can be left in the pending list if there were
 505          * errors during processing the pending nodes.
 506          */
 507         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 508                 list_for_each_entry(node, &cache->pending[level], list) {
 509                         BUG_ON(!node->pending);
 510                         if (node->bytenr == node->new_bytenr)
 511                                 continue;
 512                         update_backref_node(cache, node, node->new_bytenr);
 513                 }
 514         }
 515
 516         cache->last_trans = 0;
 517         return 1;
 518 }
 519
 520
 521 static int should_ignore_root(struct btrfs_root *root)
 522 {
 523         struct btrfs_root *reloc_root;
 524
 525         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 526                 return 0;
 527
 528         reloc_root = root->reloc_root;
 529         if (!reloc_root)
 530                 return 0;
 531
 532         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 533             root->fs_info->running_transaction->transid - 1)
 534                 return 0;
 535         /*
 536          * if there is reloc tree and it was created in previous
 537          * transaction backref lookup can find the reloc tree,
 538          * so backref node for the fs tree root is useless for
 539          * relocation.
 540          */
 541         return 1;
 542 }
 543 /*
 544  * find reloc tree by address of tree root
 545  */
 546 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 547                                           u64 bytenr)
 548 {
 549         struct rb_node *rb_node;
 550         struct mapping_node *node;
 551         struct btrfs_root *root = NULL;
 552
 553         spin_lock(&rc->reloc_root_tree.lock);
 554         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 555         if (rb_node) {
 556                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 557                 root = (struct btrfs_root *)node->data;
 558         }
 559         spin_unlock(&rc->reloc_root_tree.lock);
 560         return root;
 561 }
 562
 563 static int is_cowonly_root(u64 root_objectid)
 564 {
 565         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 566             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 567             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 568             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 569             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 570             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 571             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 572             root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
 573             root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
 574                 return 1;
 575         return 0;
 576 }
 577
 578 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 579                                         u64 root_objectid)
 580 {
 581         struct btrfs_key key;
 582
 583         key.objectid = root_objectid;
 584         key.type = BTRFS_ROOT_ITEM_KEY;
 585         if (is_cowonly_root(root_objectid))
 586                 key.offset = 0;
 587         else
 588                 key.offset = (u64)-1;
 589
 590         return btrfs_get_fs_root(fs_info, &key, false);
 591 }
 592
 593 static noinline_for_stack
 594 int find_inline_backref(struct extent_buffer *leaf, int slot,
 595                         unsigned long *ptr, unsigned long *end)
 596 {
 597         struct btrfs_key key;
 598         struct btrfs_extent_item *ei;
 599         struct btrfs_tree_block_info *bi;
 600         u32 item_size;
 601
 602         btrfs_item_key_to_cpu(leaf, &key, slot);
 603
 604         item_size = btrfs_item_size_nr(leaf, slot);
 605         if (item_size < sizeof(*ei)) {
 606                 btrfs_print_v0_err(leaf->fs_info);
 607                 btrfs_handle_fs_error(leaf->fs_info, -EINVAL, NULL);
 608                 return 1;
 609         }
 610         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 611         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 612                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 613
 614         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 615             item_size <= sizeof(*ei) + sizeof(*bi)) {
 616                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 617                 return 1;
 618         }
 619         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 620             item_size <= sizeof(*ei)) {
 621                 WARN_ON(item_size < sizeof(*ei));
 622                 return 1;
 623         }
 624
 625         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 626                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 627                 *ptr = (unsigned long)(bi + 1);
 628         } else {
 629                 *ptr = (unsigned long)(ei + 1);
 630         }
 631         *end = (unsigned long)ei + item_size;
 632         return 0;
 633 }
 634
 635 /*
 636  * build backref tree for a given tree block. root of the backref tree
 637  * corresponds the tree block, leaves of the backref tree correspond
 638  * roots of b-trees that reference the tree block.
 639  *
 640  * the basic idea of this function is check backrefs of a given block
 641  * to find upper level blocks that reference the block, and then check
 642  * backrefs of these upper level blocks recursively. the recursion stop
 643  * when tree root is reached or backrefs for the block is cached.
 644  *
 645  * NOTE: if we find backrefs for a block are cached, we know backrefs
 646  * for all upper level blocks that directly/indirectly reference the
 647  * block are also cached.
 648  */
 649 static noinline_for_stack
 650 struct backref_node *build_backref_tree(struct reloc_control *rc,
 651                                         struct btrfs_key *node_key,
 652                                         int level, u64 bytenr)
 653 {
 654         struct backref_cache *cache = &rc->backref_cache;
 655         struct btrfs_path *path1; /* For searching extent root */
 656         struct btrfs_path *path2; /* For searching parent of TREE_BLOCK_REF */
 657         struct extent_buffer *eb;
 658         struct btrfs_root *root;
 659         struct backref_node *cur;
 660         struct backref_node *upper;
 661         struct backref_node *lower;
 662         struct backref_node *node = NULL;
 663         struct backref_node *exist = NULL;
 664         struct backref_edge *edge;
 665         struct rb_node *rb_node;
 666         struct btrfs_key key;
 667         unsigned long end;
 668         unsigned long ptr;
 669         LIST_HEAD(list); /* Pending edge list, upper node needs to be checked */
 670         LIST_HEAD(useless);
 671         int cowonly;
 672         int ret;
 673         int err = 0;
 674         bool need_check = true;
 675
 676         path1 = btrfs_alloc_path();
 677         path2 = btrfs_alloc_path();
 678         if (!path1 || !path2) {
 679                 err = -ENOMEM;
 680                 goto out;
 681         }
 682         path1->reada = READA_FORWARD;
 683         path2->reada = READA_FORWARD;
 684
 685         node = alloc_backref_node(cache);
 686         if (!node) {
 687                 err = -ENOMEM;
 688                 goto out;
 689         }
 690
 691         node->bytenr = bytenr;
 692         node->level = level;
 693         node->lowest = 1;
 694         cur = node;
 695 again:
 696         end = 0;
 697         ptr = 0;
 698         key.objectid = cur->bytenr;
 699         key.type = BTRFS_METADATA_ITEM_KEY;
 700         key.offset = (u64)-1;
 701
 702         path1->search_commit_root = 1;
 703         path1->skip_locking = 1;
 704         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 705                                 0, 0);
 706         if (ret < 0) {
 707                 err = ret;
 708                 goto out;
 709         }
 710         ASSERT(ret);
 711         ASSERT(path1->slots[0]);
 712
 713         path1->slots[0]--;
 714
 715         WARN_ON(cur->checked);
 716         if (!list_empty(&cur->upper)) {
 717                 /*
 718                  * the backref was added previously when processing
 719                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 720                  */
 721                 ASSERT(list_is_singular(&cur->upper));
 722                 edge = list_entry(cur->upper.next, struct backref_edge,
 723                                   list[LOWER]);
 724                 ASSERT(list_empty(&edge->list[UPPER]));
 725                 exist = edge->node[UPPER];
 726                 /*
 727                  * add the upper level block to pending list if we need
 728                  * check its backrefs
 729                  */
 730                 if (!exist->checked)
 731                         list_add_tail(&edge->list[UPPER], &list);
 732         } else {
 733                 exist = NULL;
 734         }
 735
 736         while (1) {
 737                 cond_resched();
 738                 eb = path1->nodes[0];
 739
 740                 if (ptr >= end) {
 741                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 742                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 743                                 if (ret < 0) {
 744                                         err = ret;
 745                                         goto out;
 746                                 }
 747                                 if (ret > 0)
 748                                         break;
 749                                 eb = path1->nodes[0];
 750                         }
 751
 752                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 753                         if (key.objectid != cur->bytenr) {
 754                                 WARN_ON(exist);
 755                                 break;
 756                         }
 757
 758                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 759                             key.type == BTRFS_METADATA_ITEM_KEY) {
 760                                 ret = find_inline_backref(eb, path1->slots[0],
 761                                                           &ptr, &end);
 762                                 if (ret)
 763                                         goto next;
 764                         }
 765                 }
 766
 767                 if (ptr < end) {
 768                         /* update key for inline back ref */
 769                         struct btrfs_extent_inline_ref *iref;
 770                         int type;
 771                         iref = (struct btrfs_extent_inline_ref *)ptr;
 772                         type = btrfs_get_extent_inline_ref_type(eb, iref,
 773                                                         BTRFS_REF_TYPE_BLOCK);
 774                         if (type == BTRFS_REF_TYPE_INVALID) {
 775                                 err = -EUCLEAN;
 776                                 goto out;
 777                         }
 778                         key.type = type;
 779                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 780
 781                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 782                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 783                 }
 784
 785                 /*
 786                  * Parent node found and matches current inline ref, no need to
 787                  * rebuild this node for this inline ref.
 788                  */
 789                 if (exist &&
 790                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 791                       exist->owner == key.offset) ||
 792                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 793                       exist->bytenr == key.offset))) {
 794                         exist = NULL;
 795                         goto next;
 796                 }
 797
 798                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
 799                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 800                         if (key.objectid == key.offset) {
 801                                 /*
 802                                  * Only root blocks of reloc trees use backref
 803                                  * pointing to itself.
 804                                  */
 805                                 root = find_reloc_root(rc, cur->bytenr);
 806                                 ASSERT(root);
 807                                 cur->root = root;
 808                                 break;
 809                         }
 810
 811                         edge = alloc_backref_edge(cache);
 812                         if (!edge) {
 813                                 err = -ENOMEM;
 814                                 goto out;
 815                         }
 816                         rb_node = tree_search(&cache->rb_root, key.offset);
 817                         if (!rb_node) {
 818                                 upper = alloc_backref_node(cache);
 819                                 if (!upper) {
 820                                         free_backref_edge(cache, edge);
 821                                         err = -ENOMEM;
 822                                         goto out;
 823                                 }
 824                                 upper->bytenr = key.offset;
 825                                 upper->level = cur->level + 1;
 826                                 /*
 827                                  *  backrefs for the upper level block isn't
 828                                  *  cached, add the block to pending list
 829                                  */
 830                                 list_add_tail(&edge->list[UPPER], &list);
 831                         } else {
 832                                 upper = rb_entry(rb_node, struct backref_node,
 833                                                  rb_node);
 834                                 ASSERT(upper->checked);
 835                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 836                         }
 837                         list_add_tail(&edge->list[LOWER], &cur->upper);
 838                         edge->node[LOWER] = cur;
 839                         edge->node[UPPER] = upper;
 840
 841                         goto next;
 842                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 843                         err = -EINVAL;
 844                         btrfs_print_v0_err(rc->extent_root->fs_info);
 845                         btrfs_handle_fs_error(rc->extent_root->fs_info, err,
 846                                               NULL);
 847                         goto out;
 848                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 849                         goto next;
 850                 }
 851
 852                 /*
 853                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
 854                  * means the root objectid. We need to search the tree to get
 855                  * its parent bytenr.
 856                  */
 857                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 858                 if (IS_ERR(root)) {
 859                         err = PTR_ERR(root);
 860                         goto out;
 861                 }
 862
 863                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 864                         cur->cowonly = 1;
 865
 866                 if (btrfs_root_level(&root->root_item) == cur->level) {
 867                         /* tree root */
 868                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
 869                                cur->bytenr);
 870                         if (should_ignore_root(root))
 871                                 list_add(&cur->list, &useless);
 872                         else
 873                                 cur->root = root;
 874                         break;
 875                 }
 876
 877                 level = cur->level + 1;
 878
 879                 /* Search the tree to find parent blocks referring the block. */
 880                 path2->search_commit_root = 1;
 881                 path2->skip_locking = 1;
 882                 path2->lowest_level = level;
 883                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 884                 path2->lowest_level = 0;
 885                 if (ret < 0) {
 886                         err = ret;
 887                         goto out;
 888                 }
 889                 if (ret > 0 && path2->slots[level] > 0)
 890                         path2->slots[level]--;
 891
 892                 eb = path2->nodes[level];
 893                 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
 894                     cur->bytenr) {
 895                         btrfs_err(root->fs_info,
 896         "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
 897                                   cur->bytenr, level - 1,
 898                                   root->root_key.objectid,
 899                                   node_key->objectid, node_key->type,
 900                                   node_key->offset);
 901                         err = -ENOENT;
 902                         goto out;
 903                 }
 904                 lower = cur;
 905                 need_check = true;
 906
 907                 /* Add all nodes and edges in the path */
 908                 for (; level < BTRFS_MAX_LEVEL; level++) {
 909                         if (!path2->nodes[level]) {
 910                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
 911                                        lower->bytenr);
 912                                 if (should_ignore_root(root))
 913                                         list_add(&lower->list, &useless);
 914                                 else
 915                                         lower->root = root;
 916                                 break;
 917                         }
 918
 919                         edge = alloc_backref_edge(cache);
 920                         if (!edge) {
 921                                 err = -ENOMEM;
 922                                 goto out;
 923                         }
 924
 925                         eb = path2->nodes[level];
 926                         rb_node = tree_search(&cache->rb_root, eb->start);
 927                         if (!rb_node) {
 928                                 upper = alloc_backref_node(cache);
 929                                 if (!upper) {
 930                                         free_backref_edge(cache, edge);
 931                                         err = -ENOMEM;
 932                                         goto out;
 933                                 }
 934                                 upper->bytenr = eb->start;
 935                                 upper->owner = btrfs_header_owner(eb);
 936                                 upper->level = lower->level + 1;
 937                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
 938                                               &root->state))
 939                                         upper->cowonly = 1;
 940
 941                                 /*
 942                                  * if we know the block isn't shared
 943                                  * we can void checking its backrefs.
 944                                  */
 945                                 if (btrfs_block_can_be_shared(root, eb))
 946                                         upper->checked = 0;
 947                                 else
 948                                         upper->checked = 1;
 949
 950                                 /*
 951                                  * add the block to pending list if we
 952                                  * need check its backrefs, we only do this once
 953                                  * while walking up a tree as we will catch
 954                                  * anything else later on.
 955                                  */
 956                                 if (!upper->checked && need_check) {
 957                                         need_check = false;
 958                                         list_add_tail(&edge->list[UPPER],
 959                                                       &list);
 960                                 } else {
 961                                         if (upper->checked)
 962                                                 need_check = true;
 963                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 964                                 }
 965                         } else {
 966                                 upper = rb_entry(rb_node, struct backref_node,
 967                                                  rb_node);
 968                                 ASSERT(upper->checked);
 969                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 970                                 if (!upper->owner)
 971                                         upper->owner = btrfs_header_owner(eb);
 972                         }
 973                         list_add_tail(&edge->list[LOWER], &lower->upper);
 974                         edge->node[LOWER] = lower;
 975                         edge->node[UPPER] = upper;
 976
 977                         if (rb_node)
 978                                 break;
 979                         lower = upper;
 980                         upper = NULL;
 981                 }
 982                 btrfs_release_path(path2);
 983 next:
 984                 if (ptr < end) {
 985                         ptr += btrfs_extent_inline_ref_size(key.type);
 986                         if (ptr >= end) {
 987                                 WARN_ON(ptr > end);
 988                                 ptr = 0;
 989                                 end = 0;
 990                         }
 991                 }
 992                 if (ptr >= end)
 993                         path1->slots[0]++;
 994         }
 995         btrfs_release_path(path1);
 996
 997         cur->checked = 1;
 998         WARN_ON(exist);
 999
1000         /* the pending list isn't empty, take the first block to process */
1001         if (!list_empty(&list)) {
1002                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1003                 list_del_init(&edge->list[UPPER]);
1004                 cur = edge->node[UPPER];
1005                 goto again;
1006         }
1007
1008         /*
1009          * everything goes well, connect backref nodes and insert backref nodes
1010          * into the cache.
1011          */
1012         ASSERT(node->checked);
1013         cowonly = node->cowonly;
1014         if (!cowonly) {
1015                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1016                                       &node->rb_node);
1017                 if (rb_node)
1018                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1019                 list_add_tail(&node->lower, &cache->leaves);
1020         }
1021
1022         list_for_each_entry(edge, &node->upper, list[LOWER])
1023                 list_add_tail(&edge->list[UPPER], &list);
1024
1025         while (!list_empty(&list)) {
1026                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1027                 list_del_init(&edge->list[UPPER]);
1028                 upper = edge->node[UPPER];
1029                 if (upper->detached) {
1030                         list_del(&edge->list[LOWER]);
1031                         lower = edge->node[LOWER];
1032                         free_backref_edge(cache, edge);
1033                         if (list_empty(&lower->upper))
1034                                 list_add(&lower->list, &useless);
1035                         continue;
1036                 }
1037
1038                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1039                         if (upper->lowest) {
1040                                 list_del_init(&upper->lower);
1041                                 upper->lowest = 0;
1042                         }
1043
1044                         list_add_tail(&edge->list[UPPER], &upper->lower);
1045                         continue;
1046                 }
1047
1048                 if (!upper->checked) {
1049                         /*
1050                          * Still want to blow up for developers since this is a
1051                          * logic bug.
1052                          */
1053                         ASSERT(0);
1054                         err = -EINVAL;
1055                         goto out;
1056                 }
1057                 if (cowonly != upper->cowonly) {
1058                         ASSERT(0);
1059                         err = -EINVAL;
1060                         goto out;
1061                 }
1062
1063                 if (!cowonly) {
1064                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1065                                               &upper->rb_node);
1066                         if (rb_node)
1067                                 backref_tree_panic(rb_node, -EEXIST,
1068                                                    upper->bytenr);
1069                 }
1070
1071                 list_add_tail(&edge->list[UPPER], &upper->lower);
1072
1073                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1074                         list_add_tail(&edge->list[UPPER], &list);
1075         }
1076         /*
1077          * process useless backref nodes. backref nodes for tree leaves
1078          * are deleted from the cache. backref nodes for upper level
1079          * tree blocks are left in the cache to avoid unnecessary backref
1080          * lookup.
1081          */
1082         while (!list_empty(&useless)) {
1083                 upper = list_entry(useless.next, struct backref_node, list);
1084                 list_del_init(&upper->list);
1085                 ASSERT(list_empty(&upper->upper));
1086                 if (upper == node)
1087                         node = NULL;
1088                 if (upper->lowest) {
1089                         list_del_init(&upper->lower);
1090                         upper->lowest = 0;
1091                 }
1092                 while (!list_empty(&upper->lower)) {
1093                         edge = list_entry(upper->lower.next,
1094                                           struct backref_edge, list[UPPER]);
1095                         list_del(&edge->list[UPPER]);
1096                         list_del(&edge->list[LOWER]);
1097                         lower = edge->node[LOWER];
1098                         free_backref_edge(cache, edge);
1099
1100                         if (list_empty(&lower->upper))
1101                                 list_add(&lower->list, &useless);
1102                 }
1103                 __mark_block_processed(rc, upper);
1104                 if (upper->level > 0) {
1105                         list_add(&upper->list, &cache->detached);
1106                         upper->detached = 1;
1107                 } else {
1108                         rb_erase(&upper->rb_node, &cache->rb_root);
1109                         free_backref_node(cache, upper);
1110                 }
1111         }
1112 out:
1113         btrfs_free_path(path1);
1114         btrfs_free_path(path2);
1115         if (err) {
1116                 while (!list_empty(&useless)) {
1117                         lower = list_entry(useless.next,
1118                                            struct backref_node, list);
1119                         list_del_init(&lower->list);
1120                 }
1121                 while (!list_empty(&list)) {
1122                         edge = list_first_entry(&list, struct backref_edge,
1123                                                 list[UPPER]);
1124                         list_del(&edge->list[UPPER]);
1125                         list_del(&edge->list[LOWER]);
1126                         lower = edge->node[LOWER];
1127                         upper = edge->node[UPPER];
1128                         free_backref_edge(cache, edge);
1129
1130                         /*
1131                          * Lower is no longer linked to any upper backref nodes
1132                          * and isn't in the cache, we can free it ourselves.
1133                          */
1134                         if (list_empty(&lower->upper) &&
1135                             RB_EMPTY_NODE(&lower->rb_node))
1136                                 list_add(&lower->list, &useless);
1137
1138                         if (!RB_EMPTY_NODE(&upper->rb_node))
1139                                 continue;
1140
1141                         /* Add this guy's upper edges to the list to process */
1142                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1143                                 list_add_tail(&edge->list[UPPER], &list);
1144                         if (list_empty(&upper->upper))
1145                                 list_add(&upper->list, &useless);
1146                 }
1147
1148                 while (!list_empty(&useless)) {
1149                         lower = list_entry(useless.next,
1150                                            struct backref_node, list);
1151                         list_del_init(&lower->list);
1152                         if (lower == node)
1153                                 node = NULL;
1154                         free_backref_node(cache, lower);
1155                 }
1156
1157                 free_backref_node(cache, node);
1158                 return ERR_PTR(err);
1159         }
1160         ASSERT(!node || !node->detached);
1161         return node;
1162 }
1163
1164 /*
1165  * helper to add backref node for the newly created snapshot.
1166  * the backref node is created by cloning backref node that
1167  * corresponds to root of source tree
1168  */
1169 static int clone_backref_node(struct btrfs_trans_handle *trans,
1170                               struct reloc_control *rc,
1171                               struct btrfs_root *src,
1172                               struct btrfs_root *dest)
1173 {
1174         struct btrfs_root *reloc_root = src->reloc_root;
1175         struct backref_cache *cache = &rc->backref_cache;
1176         struct backref_node *node = NULL;
1177         struct backref_node *new_node;
1178         struct backref_edge *edge;
1179         struct backref_edge *new_edge;
1180         struct rb_node *rb_node;
1181
1182         if (cache->last_trans > 0)
1183                 update_backref_cache(trans, cache);
1184
1185         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1186         if (rb_node) {
1187                 node = rb_entry(rb_node, struct backref_node, rb_node);
1188                 if (node->detached)
1189                         node = NULL;
1190                 else
1191                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1192         }
1193
1194         if (!node) {
1195                 rb_node = tree_search(&cache->rb_root,
1196                                       reloc_root->commit_root->start);
1197                 if (rb_node) {
1198                         node = rb_entry(rb_node, struct backref_node,
1199                                         rb_node);
1200                         BUG_ON(node->detached);
1201                 }
1202         }
1203
1204         if (!node)
1205                 return 0;
1206
1207         new_node = alloc_backref_node(cache);
1208         if (!new_node)
1209                 return -ENOMEM;
1210
1211         new_node->bytenr = dest->node->start;
1212         new_node->level = node->level;
1213         new_node->lowest = node->lowest;
1214         new_node->checked = 1;
1215         new_node->root = dest;
1216
1217         if (!node->lowest) {
1218                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1219                         new_edge = alloc_backref_edge(cache);
1220                         if (!new_edge)
1221                                 goto fail;
1222
1223                         new_edge->node[UPPER] = new_node;
1224                         new_edge->node[LOWER] = edge->node[LOWER];
1225                         list_add_tail(&new_edge->list[UPPER],
1226                                       &new_node->lower);
1227                 }
1228         } else {
1229                 list_add_tail(&new_node->lower, &cache->leaves);
1230         }
1231
1232         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1233                               &new_node->rb_node);
1234         if (rb_node)
1235                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1236
1237         if (!new_node->lowest) {
1238                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1239                         list_add_tail(&new_edge->list[LOWER],
1240                                       &new_edge->node[LOWER]->upper);
1241                 }
1242         }
1243         return 0;
1244 fail:
1245         while (!list_empty(&new_node->lower)) {
1246                 new_edge = list_entry(new_node->lower.next,
1247                                       struct backref_edge, list[UPPER]);
1248                 list_del(&new_edge->list[UPPER]);
1249                 free_backref_edge(cache, new_edge);
1250         }
1251         free_backref_node(cache, new_node);
1252         return -ENOMEM;
1253 }
1254
1255 /*
1256  * helper to add 'address of tree root -> reloc tree' mapping
1257  */
1258 static int __must_check __add_reloc_root(struct btrfs_root *root)
1259 {
1260         struct btrfs_fs_info *fs_info = root->fs_info;
1261         struct rb_node *rb_node;
1262         struct mapping_node *node;
1263         struct reloc_control *rc = fs_info->reloc_ctl;
1264
1265         node = kmalloc(sizeof(*node), GFP_NOFS);
1266         if (!node)
1267                 return -ENOMEM;
1268
1269         node->bytenr = root->node->start;
1270         node->data = root;
1271
1272         spin_lock(&rc->reloc_root_tree.lock);
1273         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1274                               node->bytenr, &node->rb_node);
1275         spin_unlock(&rc->reloc_root_tree.lock);
1276         if (rb_node) {
1277                 btrfs_panic(fs_info, -EEXIST,
1278                             "Duplicate root found for start=%llu while inserting into relocation tree",
1279                             node->bytenr);
1280         }
1281
1282         list_add_tail(&root->root_list, &rc->reloc_roots);
1283         return 0;
1284 }
1285
1286 /*
1287  * helper to delete the 'address of tree root -> reloc tree'
1288  * mapping
1289  */
1290 static void __del_reloc_root(struct btrfs_root *root)
1291 {
1292         struct btrfs_fs_info *fs_info = root->fs_info;
1293         struct rb_node *rb_node;
1294         struct mapping_node *node = NULL;
1295         struct reloc_control *rc = fs_info->reloc_ctl;
1296
1297         if (rc && root->node) {
1298                 spin_lock(&rc->reloc_root_tree.lock);
1299                 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1300                                       root->node->start);
1301                 if (rb_node) {
1302                         node = rb_entry(rb_node, struct mapping_node, rb_node);
1303                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1304                 }
1305                 spin_unlock(&rc->reloc_root_tree.lock);
1306                 if (!node)
1307                         return;
1308                 BUG_ON((struct btrfs_root *)node->data != root);
1309         }
1310
1311         spin_lock(&fs_info->trans_lock);
1312         list_del_init(&root->root_list);
1313         spin_unlock(&fs_info->trans_lock);
1314         kfree(node);
1315 }
1316
1317 /*
1318  * helper to update the 'address of tree root -> reloc tree'
1319  * mapping
1320  */
1321 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1322 {
1323         struct btrfs_fs_info *fs_info = root->fs_info;
1324         struct rb_node *rb_node;
1325         struct mapping_node *node = NULL;
1326         struct reloc_control *rc = fs_info->reloc_ctl;
1327
1328         spin_lock(&rc->reloc_root_tree.lock);
1329         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1330                               root->node->start);
1331         if (rb_node) {
1332                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1333                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1334         }
1335         spin_unlock(&rc->reloc_root_tree.lock);
1336
1337         if (!node)
1338                 return 0;
1339         BUG_ON((struct btrfs_root *)node->data != root);
1340
1341         spin_lock(&rc->reloc_root_tree.lock);
1342         node->bytenr = new_bytenr;
1343         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1344                               node->bytenr, &node->rb_node);
1345         spin_unlock(&rc->reloc_root_tree.lock);
1346         if (rb_node)
1347                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1348         return 0;
1349 }
1350
1351 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1352                                         struct btrfs_root *root, u64 objectid)
1353 {
1354         struct btrfs_fs_info *fs_info = root->fs_info;
1355         struct btrfs_root *reloc_root;
1356         struct extent_buffer *eb;
1357         struct btrfs_root_item *root_item;
1358         struct btrfs_key root_key;
1359         int ret;
1360
1361         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1362         BUG_ON(!root_item);
1363
1364         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1365         root_key.type = BTRFS_ROOT_ITEM_KEY;
1366         root_key.offset = objectid;
1367
1368         if (root->root_key.objectid == objectid) {
1369                 u64 commit_root_gen;
1370
1371                 /* called by btrfs_init_reloc_root */
1372                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1373                                       BTRFS_TREE_RELOC_OBJECTID);
1374                 BUG_ON(ret);
1375                 /*
1376                  * Set the last_snapshot field to the generation of the commit
1377                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1378                  * correctly (returns true) when the relocation root is created
1379                  * either inside the critical section of a transaction commit
1380                  * (through transaction.c:qgroup_account_snapshot()) and when
1381                  * it's created before the transaction commit is started.
1382                  */
1383                 commit_root_gen = btrfs_header_generation(root->commit_root);
1384                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1385         } else {
1386                 /*
1387                  * called by btrfs_reloc_post_snapshot_hook.
1388                  * the source tree is a reloc tree, all tree blocks
1389                  * modified after it was created have RELOC flag
1390                  * set in their headers. so it's OK to not update
1391                  * the 'last_snapshot'.
1392                  */
1393                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1394                                       BTRFS_TREE_RELOC_OBJECTID);
1395                 BUG_ON(ret);
1396         }
1397
1398         memcpy(root_item, &root->root_item, sizeof(*root_item));
1399         btrfs_set_root_bytenr(root_item, eb->start);
1400         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1401         btrfs_set_root_generation(root_item, trans->transid);
1402
1403         if (root->root_key.objectid == objectid) {
1404                 btrfs_set_root_refs(root_item, 0);
1405                 memset(&root_item->drop_progress, 0,
1406                        sizeof(struct btrfs_disk_key));
1407                 root_item->drop_level = 0;
1408         }
1409
1410         btrfs_tree_unlock(eb);
1411         free_extent_buffer(eb);
1412
1413         ret = btrfs_insert_root(trans, fs_info->tree_root,
1414                                 &root_key, root_item);
1415         BUG_ON(ret);
1416         kfree(root_item);
1417
1418         reloc_root = btrfs_read_fs_root(fs_info->tree_root, &root_key);
1419         BUG_ON(IS_ERR(reloc_root));
1420         reloc_root->last_trans = trans->transid;
1421         return reloc_root;
1422 }
1423
1424 /*
1425  * create reloc tree for a given fs tree. reloc tree is just a
1426  * snapshot of the fs tree with special root objectid.
1427  */
1428 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1429                           struct btrfs_root *root)
1430 {
1431         struct btrfs_fs_info *fs_info = root->fs_info;
1432         struct btrfs_root *reloc_root;
1433         struct reloc_control *rc = fs_info->reloc_ctl;
1434         struct btrfs_block_rsv *rsv;
1435         int clear_rsv = 0;
1436         int ret;
1437
1438         /*
1439          * The subvolume has reloc tree but the swap is finished, no need to
1440          * create/update the dead reloc tree
1441          */
1442         if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
1443                 return 0;
1444
1445         if (root->reloc_root) {
1446                 reloc_root = root->reloc_root;
1447                 reloc_root->last_trans = trans->transid;
1448                 return 0;
1449         }
1450
1451         if (!rc || !rc->create_reloc_tree ||
1452             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1453                 return 0;
1454
1455         if (!trans->reloc_reserved) {
1456                 rsv = trans->block_rsv;
1457                 trans->block_rsv = rc->block_rsv;
1458                 clear_rsv = 1;
1459         }
1460         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1461         if (clear_rsv)
1462                 trans->block_rsv = rsv;
1463
1464         ret = __add_reloc_root(reloc_root);
1465         BUG_ON(ret < 0);
1466         root->reloc_root = reloc_root;
1467         return 0;
1468 }
1469
1470 /*
1471  * update root item of reloc tree
1472  */
1473 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1474                             struct btrfs_root *root)
1475 {
1476         struct btrfs_fs_info *fs_info = root->fs_info;
1477         struct btrfs_root *reloc_root;
1478         struct btrfs_root_item *root_item;
1479         int ret;
1480
1481         if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state) ||
1482             !root->reloc_root)
1483                 goto out;
1484
1485         reloc_root = root->reloc_root;
1486         root_item = &reloc_root->root_item;
1487
1488         /* root->reloc_root will stay until current relocation finished */
1489         if (fs_info->reloc_ctl->merge_reloc_tree &&
1490             btrfs_root_refs(root_item) == 0) {
1491                 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1492                 __del_reloc_root(reloc_root);
1493         }
1494
1495         if (reloc_root->commit_root != reloc_root->node) {
1496                 btrfs_set_root_node(root_item, reloc_root->node);
1497                 free_extent_buffer(reloc_root->commit_root);
1498                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1499         }
1500
1501         ret = btrfs_update_root(trans, fs_info->tree_root,
1502                                 &reloc_root->root_key, root_item);
1503         BUG_ON(ret);
1504
1505 out:
1506         return 0;
1507 }
1508
1509 /*
1510  * helper to find first cached inode with inode number >= objectid
1511  * in a subvolume
1512  */
1513 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1514 {
1515         struct rb_node *node;
1516         struct rb_node *prev;
1517         struct btrfs_inode *entry;
1518         struct inode *inode;
1519
1520         spin_lock(&root->inode_lock);
1521 again:
1522         node = root->inode_tree.rb_node;
1523         prev = NULL;
1524         while (node) {
1525                 prev = node;
1526                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1527
1528                 if (objectid < btrfs_ino(entry))
1529                         node = node->rb_left;
1530                 else if (objectid > btrfs_ino(entry))
1531                         node = node->rb_right;
1532                 else
1533                         break;
1534         }
1535         if (!node) {
1536                 while (prev) {
1537                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1538                         if (objectid <= btrfs_ino(entry)) {
1539                                 node = prev;
1540                                 break;
1541                         }
1542                         prev = rb_next(prev);
1543                 }
1544         }
1545         while (node) {
1546                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1547                 inode = igrab(&entry->vfs_inode);
1548                 if (inode) {
1549                         spin_unlock(&root->inode_lock);
1550                         return inode;
1551                 }
1552
1553                 objectid = btrfs_ino(entry) + 1;
1554                 if (cond_resched_lock(&root->inode_lock))
1555                         goto again;
1556
1557                 node = rb_next(node);
1558         }
1559         spin_unlock(&root->inode_lock);
1560         return NULL;
1561 }
1562
1563 static int in_block_group(u64 bytenr,
1564                           struct btrfs_block_group_cache *block_group)
1565 {
1566         if (bytenr >= block_group->key.objectid &&
1567             bytenr < block_group->key.objectid + block_group->key.offset)
1568                 return 1;
1569         return 0;
1570 }
1571
1572 /*
1573  * get new location of data
1574  */
1575 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1576                             u64 bytenr, u64 num_bytes)
1577 {
1578         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1579         struct btrfs_path *path;
1580         struct btrfs_file_extent_item *fi;
1581         struct extent_buffer *leaf;
1582         int ret;
1583
1584         path = btrfs_alloc_path();
1585         if (!path)
1586                 return -ENOMEM;
1587
1588         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1589         ret = btrfs_lookup_file_extent(NULL, root, path,
1590                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1591         if (ret < 0)
1592                 goto out;
1593         if (ret > 0) {
1594                 ret = -ENOENT;
1595                 goto out;
1596         }
1597
1598         leaf = path->nodes[0];
1599         fi = btrfs_item_ptr(leaf, path->slots[0],
1600                             struct btrfs_file_extent_item);
1601
1602         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1603                btrfs_file_extent_compression(leaf, fi) ||
1604                btrfs_file_extent_encryption(leaf, fi) ||
1605                btrfs_file_extent_other_encoding(leaf, fi));
1606
1607         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1608                 ret = -EINVAL;
1609                 goto out;
1610         }
1611
1612         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1613         ret = 0;
1614 out:
1615         btrfs_free_path(path);
1616         return ret;
1617 }
1618
1619 /*
1620  * update file extent items in the tree leaf to point to
1621  * the new locations.
1622  */
1623 static noinline_for_stack
1624 int replace_file_extents(struct btrfs_trans_handle *trans,
1625                          struct reloc_control *rc,
1626                          struct btrfs_root *root,
1627                          struct extent_buffer *leaf)
1628 {
1629         struct btrfs_fs_info *fs_info = root->fs_info;
1630         struct btrfs_key key;
1631         struct btrfs_file_extent_item *fi;
1632         struct inode *inode = NULL;
1633         u64 parent;
1634         u64 bytenr;
1635         u64 new_bytenr = 0;
1636         u64 num_bytes;
1637         u64 end;
1638         u32 nritems;
1639         u32 i;
1640         int ret = 0;
1641         int first = 1;
1642         int dirty = 0;
1643
1644         if (rc->stage != UPDATE_DATA_PTRS)
1645                 return 0;
1646
1647         /* reloc trees always use full backref */
1648         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1649                 parent = leaf->start;
1650         else
1651                 parent = 0;
1652
1653         nritems = btrfs_header_nritems(leaf);
1654         for (i = 0; i < nritems; i++) {
1655                 struct btrfs_ref ref = { 0 };
1656
1657                 cond_resched();
1658                 btrfs_item_key_to_cpu(leaf, &key, i);
1659                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1660                         continue;
1661                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1662                 if (btrfs_file_extent_type(leaf, fi) ==
1663                     BTRFS_FILE_EXTENT_INLINE)
1664                         continue;
1665                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1666                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1667                 if (bytenr == 0)
1668                         continue;
1669                 if (!in_block_group(bytenr, rc->block_group))
1670                         continue;
1671
1672                 /*
1673                  * if we are modifying block in fs tree, wait for readpage
1674                  * to complete and drop the extent cache
1675                  */
1676                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1677                         if (first) {
1678                                 inode = find_next_inode(root, key.objectid);
1679                                 first = 0;
1680                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1681                                 btrfs_add_delayed_iput(inode);
1682                                 inode = find_next_inode(root, key.objectid);
1683                         }
1684                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1685                                 end = key.offset +
1686                                       btrfs_file_extent_num_bytes(leaf, fi);
1687                                 WARN_ON(!IS_ALIGNED(key.offset,
1688                                                     fs_info->sectorsize));
1689                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1690                                 end--;
1691                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1692                                                       key.offset, end);
1693                                 if (!ret)
1694                                         continue;
1695
1696                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1697                                                 key.offset,     end, 1);
1698                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1699                                               key.offset, end);
1700                         }
1701                 }
1702
1703                 ret = get_new_location(rc->data_inode, &new_bytenr,
1704                                        bytenr, num_bytes);
1705                 if (ret) {
1706                         /*
1707                          * Don't have to abort since we've not changed anything
1708                          * in the file extent yet.
1709                          */
1710                         break;
1711                 }
1712
1713                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1714                 dirty = 1;
1715
1716                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1717                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1718                                        num_bytes, parent);
1719                 ref.real_root = root->root_key.objectid;
1720                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1721                                     key.objectid, key.offset);
1722                 ret = btrfs_inc_extent_ref(trans, &ref);
1723                 if (ret) {
1724                         btrfs_abort_transaction(trans, ret);
1725                         break;
1726                 }
1727
1728                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1729                                        num_bytes, parent);
1730                 ref.real_root = root->root_key.objectid;
1731                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1732                                     key.objectid, key.offset);
1733                 ret = btrfs_free_extent(trans, &ref);
1734                 if (ret) {
1735                         btrfs_abort_transaction(trans, ret);
1736                         break;
1737                 }
1738         }
1739         if (dirty)
1740                 btrfs_mark_buffer_dirty(leaf);
1741         if (inode)
1742                 btrfs_add_delayed_iput(inode);
1743         return ret;
1744 }
1745
1746 static noinline_for_stack
1747 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1748                      struct btrfs_path *path, int level)
1749 {
1750         struct btrfs_disk_key key1;
1751         struct btrfs_disk_key key2;
1752         btrfs_node_key(eb, &key1, slot);
1753         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1754         return memcmp(&key1, &key2, sizeof(key1));
1755 }
1756
1757 /*
1758  * try to replace tree blocks in fs tree with the new blocks
1759  * in reloc tree. tree blocks haven't been modified since the
1760  * reloc tree was create can be replaced.
1761  *
1762  * if a block was replaced, level of the block + 1 is returned.
1763  * if no block got replaced, 0 is returned. if there are other
1764  * errors, a negative error number is returned.
1765  */
1766 static noinline_for_stack
1767 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1768                  struct btrfs_root *dest, struct btrfs_root *src,
1769                  struct btrfs_path *path, struct btrfs_key *next_key,
1770                  int lowest_level, int max_level)
1771 {
1772         struct btrfs_fs_info *fs_info = dest->fs_info;
1773         struct extent_buffer *eb;
1774         struct extent_buffer *parent;
1775         struct btrfs_ref ref = { 0 };
1776         struct btrfs_key key;
1777         u64 old_bytenr;
1778         u64 new_bytenr;
1779         u64 old_ptr_gen;
1780         u64 new_ptr_gen;
1781         u64 last_snapshot;
1782         u32 blocksize;
1783         int cow = 0;
1784         int level;
1785         int ret;
1786         int slot;
1787
1788         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1789         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1790
1791         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1792 again:
1793         slot = path->slots[lowest_level];
1794         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1795
1796         eb = btrfs_lock_root_node(dest);
1797         btrfs_set_lock_blocking_write(eb);
1798         level = btrfs_header_level(eb);
1799
1800         if (level < lowest_level) {
1801                 btrfs_tree_unlock(eb);
1802                 free_extent_buffer(eb);
1803                 return 0;
1804         }
1805
1806         if (cow) {
1807                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1808                 BUG_ON(ret);
1809         }
1810         btrfs_set_lock_blocking_write(eb);
1811
1812         if (next_key) {
1813                 next_key->objectid = (u64)-1;
1814                 next_key->type = (u8)-1;
1815                 next_key->offset = (u64)-1;
1816         }
1817
1818         parent = eb;
1819         while (1) {
1820                 struct btrfs_key first_key;
1821
1822                 level = btrfs_header_level(parent);
1823                 BUG_ON(level < lowest_level);
1824
1825                 ret = btrfs_bin_search(parent, &key, level, &slot);
1826                 if (ret < 0)
1827                         break;
1828                 if (ret && slot > 0)
1829                         slot--;
1830
1831                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1832                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1833
1834                 old_bytenr = btrfs_node_blockptr(parent, slot);
1835                 blocksize = fs_info->nodesize;
1836                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1837                 btrfs_node_key_to_cpu(parent, &first_key, slot);
1838
1839                 if (level <= max_level) {
1840                         eb = path->nodes[level];
1841                         new_bytenr = btrfs_node_blockptr(eb,
1842                                                         path->slots[level]);
1843                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1844                                                         path->slots[level]);
1845                 } else {
1846                         new_bytenr = 0;
1847                         new_ptr_gen = 0;
1848                 }
1849
1850                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1851                         ret = level;
1852                         break;
1853                 }
1854
1855                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1856                     memcmp_node_keys(parent, slot, path, level)) {
1857                         if (level <= lowest_level) {
1858                                 ret = 0;
1859                                 break;
1860                         }
1861
1862                         eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen,
1863                                              level - 1, &first_key);
1864                         if (IS_ERR(eb)) {
1865                                 ret = PTR_ERR(eb);
1866                                 break;
1867                         } else if (!extent_buffer_uptodate(eb)) {
1868                                 ret = -EIO;
1869                                 free_extent_buffer(eb);
1870                                 break;
1871                         }
1872                         btrfs_tree_lock(eb);
1873                         if (cow) {
1874                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1875                                                       slot, &eb);
1876                                 BUG_ON(ret);
1877                         }
1878                         btrfs_set_lock_blocking_write(eb);
1879
1880                         btrfs_tree_unlock(parent);
1881                         free_extent_buffer(parent);
1882
1883                         parent = eb;
1884                         continue;
1885                 }
1886
1887                 if (!cow) {
1888                         btrfs_tree_unlock(parent);
1889                         free_extent_buffer(parent);
1890                         cow = 1;
1891                         goto again;
1892                 }
1893
1894                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1895                                       path->slots[level]);
1896                 btrfs_release_path(path);
1897
1898                 path->lowest_level = level;
1899                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1900                 path->lowest_level = 0;
1901                 BUG_ON(ret);
1902
1903                 /*
1904                  * Info qgroup to trace both subtrees.
1905                  *
1906                  * We must trace both trees.
1907                  * 1) Tree reloc subtree
1908                  *    If not traced, we will leak data numbers
1909                  * 2) Fs subtree
1910                  *    If not traced, we will double count old data
1911                  *
1912                  * We don't scan the subtree right now, but only record
1913                  * the swapped tree blocks.
1914                  * The real subtree rescan is delayed until we have new
1915                  * CoW on the subtree root node before transaction commit.
1916                  */
1917                 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1918                                 rc->block_group, parent, slot,
1919                                 path->nodes[level], path->slots[level],
1920                                 last_snapshot);
1921                 if (ret < 0)
1922                         break;
1923                 /*
1924                  * swap blocks in fs tree and reloc tree.
1925                  */
1926                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1927                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1928                 btrfs_mark_buffer_dirty(parent);
1929
1930                 btrfs_set_node_blockptr(path->nodes[level],
1931                                         path->slots[level], old_bytenr);
1932                 btrfs_set_node_ptr_generation(path->nodes[level],
1933                                               path->slots[level], old_ptr_gen);
1934                 btrfs_mark_buffer_dirty(path->nodes[level]);
1935
1936                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1937                                        blocksize, path->nodes[level]->start);
1938                 ref.skip_qgroup = true;
1939                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1940                 ret = btrfs_inc_extent_ref(trans, &ref);
1941                 BUG_ON(ret);
1942                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1943                                        blocksize, 0);
1944                 ref.skip_qgroup = true;
1945                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1946                 ret = btrfs_inc_extent_ref(trans, &ref);
1947                 BUG_ON(ret);
1948
1949                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1950                                        blocksize, path->nodes[level]->start);
1951                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1952                 ref.skip_qgroup = true;
1953                 ret = btrfs_free_extent(trans, &ref);
1954                 BUG_ON(ret);
1955
1956                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1957                                        blocksize, 0);
1958                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1959                 ref.skip_qgroup = true;
1960                 ret = btrfs_free_extent(trans, &ref);
1961                 BUG_ON(ret);
1962
1963                 btrfs_unlock_up_safe(path, 0);
1964
1965                 ret = level;
1966                 break;
1967         }
1968         btrfs_tree_unlock(parent);
1969         free_extent_buffer(parent);
1970         return ret;
1971 }
1972
1973 /*
1974  * helper to find next relocated block in reloc tree
1975  */
1976 static noinline_for_stack
1977 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1978                        int *level)
1979 {
1980         struct extent_buffer *eb;
1981         int i;
1982         u64 last_snapshot;
1983         u32 nritems;
1984
1985         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1986
1987         for (i = 0; i < *level; i++) {
1988                 free_extent_buffer(path->nodes[i]);
1989                 path->nodes[i] = NULL;
1990         }
1991
1992         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1993                 eb = path->nodes[i];
1994                 nritems = btrfs_header_nritems(eb);
1995                 while (path->slots[i] + 1 < nritems) {
1996                         path->slots[i]++;
1997                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1998                             last_snapshot)
1999                                 continue;
2000
2001                         *level = i;
2002                         return 0;
2003                 }
2004                 free_extent_buffer(path->nodes[i]);
2005                 path->nodes[i] = NULL;
2006         }
2007         return 1;
2008 }
2009
2010 /*
2011  * walk down reloc tree to find relocated block of lowest level
2012  */
2013 static noinline_for_stack
2014 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
2015                          int *level)
2016 {
2017         struct btrfs_fs_info *fs_info = root->fs_info;
2018         struct extent_buffer *eb = NULL;
2019         int i;
2020         u64 bytenr;
2021         u64 ptr_gen = 0;
2022         u64 last_snapshot;
2023         u32 nritems;
2024
2025         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2026
2027         for (i = *level; i > 0; i--) {
2028                 struct btrfs_key first_key;
2029
2030                 eb = path->nodes[i];
2031                 nritems = btrfs_header_nritems(eb);
2032                 while (path->slots[i] < nritems) {
2033                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2034                         if (ptr_gen > last_snapshot)
2035                                 break;
2036                         path->slots[i]++;
2037                 }
2038                 if (path->slots[i] >= nritems) {
2039                         if (i == *level)
2040                                 break;
2041                         *level = i + 1;
2042                         return 0;
2043                 }
2044                 if (i == 1) {
2045                         *level = i;
2046                         return 0;
2047                 }
2048
2049                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2050                 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
2051                 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
2052                                      &first_key);
2053                 if (IS_ERR(eb)) {
2054                         return PTR_ERR(eb);
2055                 } else if (!extent_buffer_uptodate(eb)) {
2056                         free_extent_buffer(eb);
2057                         return -EIO;
2058                 }
2059                 BUG_ON(btrfs_header_level(eb) != i - 1);
2060                 path->nodes[i - 1] = eb;
2061                 path->slots[i - 1] = 0;
2062         }
2063         return 1;
2064 }
2065
2066 /*
2067  * invalidate extent cache for file extents whose key in range of
2068  * [min_key, max_key)
2069  */
2070 static int invalidate_extent_cache(struct btrfs_root *root,
2071                                    struct btrfs_key *min_key,
2072                                    struct btrfs_key *max_key)
2073 {
2074         struct btrfs_fs_info *fs_info = root->fs_info;
2075         struct inode *inode = NULL;
2076         u64 objectid;
2077         u64 start, end;
2078         u64 ino;
2079
2080         objectid = min_key->objectid;
2081         while (1) {
2082                 cond_resched();
2083                 iput(inode);
2084
2085                 if (objectid > max_key->objectid)
2086                         break;
2087
2088                 inode = find_next_inode(root, objectid);
2089                 if (!inode)
2090                         break;
2091                 ino = btrfs_ino(BTRFS_I(inode));
2092
2093                 if (ino > max_key->objectid) {
2094                         iput(inode);
2095                         break;
2096                 }
2097
2098                 objectid = ino + 1;
2099                 if (!S_ISREG(inode->i_mode))
2100                         continue;
2101
2102                 if (unlikely(min_key->objectid == ino)) {
2103                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2104                                 continue;
2105                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2106                                 start = 0;
2107                         else {
2108                                 start = min_key->offset;
2109                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
2110                         }
2111                 } else {
2112                         start = 0;
2113                 }
2114
2115                 if (unlikely(max_key->objectid == ino)) {
2116                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2117                                 continue;
2118                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2119                                 end = (u64)-1;
2120                         } else {
2121                                 if (max_key->offset == 0)
2122                                         continue;
2123                                 end = max_key->offset;
2124                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
2125                                 end--;
2126                         }
2127                 } else {
2128                         end = (u64)-1;
2129                 }
2130
2131                 /* the lock_extent waits for readpage to complete */
2132                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2133                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
2134                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2135         }
2136         return 0;
2137 }
2138
2139 static int find_next_key(struct btrfs_path *path, int level,
2140                          struct btrfs_key *key)
2141
2142 {
2143         while (level < BTRFS_MAX_LEVEL) {
2144                 if (!path->nodes[level])
2145                         break;
2146                 if (path->slots[level] + 1 <
2147                     btrfs_header_nritems(path->nodes[level])) {
2148                         btrfs_node_key_to_cpu(path->nodes[level], key,
2149                                               path->slots[level] + 1);
2150                         return 0;
2151                 }
2152                 level++;
2153         }
2154         return 1;
2155 }
2156
2157 /*
2158  * Insert current subvolume into reloc_control::dirty_subvol_roots
2159  */
2160 static void insert_dirty_subvol(struct btrfs_trans_handle *trans,
2161                                 struct reloc_control *rc,
2162                                 struct btrfs_root *root)
2163 {
2164         struct btrfs_root *reloc_root = root->reloc_root;
2165         struct btrfs_root_item *reloc_root_item;
2166
2167         /* @root must be a subvolume tree root with a valid reloc tree */
2168         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
2169         ASSERT(reloc_root);
2170
2171         reloc_root_item = &reloc_root->root_item;
2172         memset(&reloc_root_item->drop_progress, 0,
2173                 sizeof(reloc_root_item->drop_progress));
2174         reloc_root_item->drop_level = 0;
2175         btrfs_set_root_refs(reloc_root_item, 0);
2176         btrfs_update_reloc_root(trans, root);
2177
2178         if (list_empty(&root->reloc_dirty_list)) {
2179                 btrfs_grab_fs_root(root);
2180                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
2181         }
2182 }
2183
2184 static int clean_dirty_subvols(struct reloc_control *rc)
2185 {
2186         struct btrfs_root *root;
2187         struct btrfs_root *next;
2188         int ret = 0;
2189         int ret2;
2190
2191         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
2192                                  reloc_dirty_list) {
2193                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
2194                         /* Merged subvolume, cleanup its reloc root */
2195                         struct btrfs_root *reloc_root = root->reloc_root;
2196
2197                         list_del_init(&root->reloc_dirty_list);
2198                         root->reloc_root = NULL;
2199                         if (reloc_root) {
2200
2201                                 ret2 = btrfs_drop_snapshot(reloc_root, NULL, 0, 1);
2202                                 if (ret2 < 0 && !ret)
2203                                         ret = ret2;
2204                         }
2205                         clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
2206                         btrfs_put_fs_root(root);
2207                 } else {
2208                         /* Orphan reloc tree, just clean it up */
2209                         ret2 = btrfs_drop_snapshot(root, NULL, 0, 1);
2210                         if (ret2 < 0 && !ret)
2211                                 ret = ret2;
2212                 }
2213         }
2214         return ret;
2215 }
2216
2217 /*
2218  * merge the relocated tree blocks in reloc tree with corresponding
2219  * fs tree.
2220  */
2221 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2222                                                struct btrfs_root *root)
2223 {
2224         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2225         struct btrfs_key key;
2226         struct btrfs_key next_key;
2227         struct btrfs_trans_handle *trans = NULL;
2228         struct btrfs_root *reloc_root;
2229         struct btrfs_root_item *root_item;
2230         struct btrfs_path *path;
2231         struct extent_buffer *leaf;
2232         int level;
2233         int max_level;
2234         int replaced = 0;
2235         int ret;
2236         int err = 0;
2237         u32 min_reserved;
2238
2239         path = btrfs_alloc_path();
2240         if (!path)
2241                 return -ENOMEM;
2242         path->reada = READA_FORWARD;
2243
2244         reloc_root = root->reloc_root;
2245         root_item = &reloc_root->root_item;
2246
2247         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2248                 level = btrfs_root_level(root_item);
2249                 extent_buffer_get(reloc_root->node);
2250                 path->nodes[level] = reloc_root->node;
2251                 path->slots[level] = 0;
2252         } else {
2253                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2254
2255                 level = root_item->drop_level;
2256                 BUG_ON(level == 0);
2257                 path->lowest_level = level;
2258                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2259                 path->lowest_level = 0;
2260                 if (ret < 0) {
2261                         btrfs_free_path(path);
2262                         return ret;
2263                 }
2264
2265                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2266                                       path->slots[level]);
2267                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2268
2269                 btrfs_unlock_up_safe(path, 0);
2270         }
2271
2272         min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2273         memset(&next_key, 0, sizeof(next_key));
2274
2275         while (1) {
2276                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2277                                              BTRFS_RESERVE_FLUSH_ALL);
2278                 if (ret) {
2279                         err = ret;
2280                         goto out;
2281                 }
2282                 trans = btrfs_start_transaction(root, 0);
2283                 if (IS_ERR(trans)) {
2284                         err = PTR_ERR(trans);
2285                         trans = NULL;
2286                         goto out;
2287                 }
2288                 trans->block_rsv = rc->block_rsv;
2289
2290                 replaced = 0;
2291                 max_level = level;
2292
2293                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2294                 if (ret < 0) {
2295                         err = ret;
2296                         goto out;
2297                 }
2298                 if (ret > 0)
2299                         break;
2300
2301                 if (!find_next_key(path, level, &key) &&
2302                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2303                         ret = 0;
2304                 } else {
2305                         ret = replace_path(trans, rc, root, reloc_root, path,
2306                                            &next_key, level, max_level);
2307                 }
2308                 if (ret < 0) {
2309                         err = ret;
2310                         goto out;
2311                 }
2312
2313                 if (ret > 0) {
2314                         level = ret;
2315                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2316                                               path->slots[level]);
2317                         replaced = 1;
2318                 }
2319
2320                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2321                 if (ret > 0)
2322                         break;
2323
2324                 BUG_ON(level == 0);
2325                 /*
2326                  * save the merging progress in the drop_progress.
2327                  * this is OK since root refs == 1 in this case.
2328                  */
2329                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2330                                path->slots[level]);
2331                 root_item->drop_level = level;
2332
2333                 btrfs_end_transaction_throttle(trans);
2334                 trans = NULL;
2335
2336                 btrfs_btree_balance_dirty(fs_info);
2337
2338                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2339                         invalidate_extent_cache(root, &key, &next_key);
2340         }
2341
2342         /*
2343          * handle the case only one block in the fs tree need to be
2344          * relocated and the block is tree root.
2345          */
2346         leaf = btrfs_lock_root_node(root);
2347         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2348         btrfs_tree_unlock(leaf);
2349         free_extent_buffer(leaf);
2350         if (ret < 0)
2351                 err = ret;
2352 out:
2353         btrfs_free_path(path);
2354
2355         if (err == 0)
2356                 insert_dirty_subvol(trans, rc, root);
2357
2358         if (trans)
2359                 btrfs_end_transaction_throttle(trans);
2360
2361         btrfs_btree_balance_dirty(fs_info);
2362
2363         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2364                 invalidate_extent_cache(root, &key, &next_key);
2365
2366         return err;
2367 }
2368
2369 static noinline_for_stack
2370 int prepare_to_merge(struct reloc_control *rc, int err)
2371 {
2372         struct btrfs_root *root = rc->extent_root;
2373         struct btrfs_fs_info *fs_info = root->fs_info;
2374         struct btrfs_root *reloc_root;
2375         struct btrfs_trans_handle *trans;
2376         LIST_HEAD(reloc_roots);
2377         u64 num_bytes = 0;
2378         int ret;
2379
2380         mutex_lock(&fs_info->reloc_mutex);
2381         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2382         rc->merging_rsv_size += rc->nodes_relocated * 2;
2383         mutex_unlock(&fs_info->reloc_mutex);
2384
2385 again:
2386         if (!err) {
2387                 num_bytes = rc->merging_rsv_size;
2388                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2389                                           BTRFS_RESERVE_FLUSH_ALL);
2390                 if (ret)
2391                         err = ret;
2392         }
2393
2394         trans = btrfs_join_transaction(rc->extent_root);
2395         if (IS_ERR(trans)) {
2396                 if (!err)
2397                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2398                                                 num_bytes);
2399                 return PTR_ERR(trans);
2400         }
2401
2402         if (!err) {
2403                 if (num_bytes != rc->merging_rsv_size) {
2404                         btrfs_end_transaction(trans);
2405                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2406                                                 num_bytes);
2407                         goto again;
2408                 }
2409         }
2410
2411         rc->merge_reloc_tree = 1;
2412
2413         while (!list_empty(&rc->reloc_roots)) {
2414                 reloc_root = list_entry(rc->reloc_roots.next,
2415                                         struct btrfs_root, root_list);
2416                 list_del_init(&reloc_root->root_list);
2417
2418                 root = read_fs_root(fs_info, reloc_root->root_key.offset);
2419                 BUG_ON(IS_ERR(root));
2420                 BUG_ON(root->reloc_root != reloc_root);
2421
2422                 /*
2423                  * set reference count to 1, so btrfs_recover_relocation
2424                  * knows it should resumes merging
2425                  */
2426                 if (!err)
2427                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2428                 btrfs_update_reloc_root(trans, root);
2429
2430                 list_add(&reloc_root->root_list, &reloc_roots);
2431         }
2432
2433         list_splice(&reloc_roots, &rc->reloc_roots);
2434
2435         if (!err)
2436                 btrfs_commit_transaction(trans);
2437         else
2438                 btrfs_end_transaction(trans);
2439         return err;
2440 }
2441
2442 static noinline_for_stack
2443 void free_reloc_roots(struct list_head *list)
2444 {
2445         struct btrfs_root *reloc_root;
2446
2447         while (!list_empty(list)) {
2448                 reloc_root = list_entry(list->next, struct btrfs_root,
2449                                         root_list);
2450                 __del_reloc_root(reloc_root);
2451                 free_extent_buffer(reloc_root->node);
2452                 free_extent_buffer(reloc_root->commit_root);
2453                 reloc_root->node = NULL;
2454                 reloc_root->commit_root = NULL;
2455         }
2456 }
2457
2458 static noinline_for_stack
2459 void merge_reloc_roots(struct reloc_control *rc)
2460 {
2461         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2462         struct btrfs_root *root;
2463         struct btrfs_root *reloc_root;
2464         LIST_HEAD(reloc_roots);
2465         int found = 0;
2466         int ret = 0;
2467 again:
2468         root = rc->extent_root;
2469
2470         /*
2471          * this serializes us with btrfs_record_root_in_transaction,
2472          * we have to make sure nobody is in the middle of
2473          * adding their roots to the list while we are
2474          * doing this splice
2475          */
2476         mutex_lock(&fs_info->reloc_mutex);
2477         list_splice_init(&rc->reloc_roots, &reloc_roots);
2478         mutex_unlock(&fs_info->reloc_mutex);
2479
2480         while (!list_empty(&reloc_roots)) {
2481                 found = 1;
2482                 reloc_root = list_entry(reloc_roots.next,
2483                                         struct btrfs_root, root_list);
2484
2485                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2486                         root = read_fs_root(fs_info,
2487                                             reloc_root->root_key.offset);
2488                         BUG_ON(IS_ERR(root));
2489                         BUG_ON(root->reloc_root != reloc_root);
2490
2491                         ret = merge_reloc_root(rc, root);
2492                         if (ret) {
2493                                 if (list_empty(&reloc_root->root_list))
2494                                         list_add_tail(&reloc_root->root_list,
2495                                                       &reloc_roots);
2496                                 goto out;
2497                         }
2498                 } else {
2499                         list_del_init(&reloc_root->root_list);
2500                         /* Don't forget to queue this reloc root for cleanup */
2501                         list_add_tail(&reloc_root->reloc_dirty_list,
2502                                       &rc->dirty_subvol_roots);
2503                 }
2504         }
2505
2506         if (found) {
2507                 found = 0;
2508                 goto again;
2509         }
2510 out:
2511         if (ret) {
2512                 btrfs_handle_fs_error(fs_info, ret, NULL);
2513                 if (!list_empty(&reloc_roots))
2514                         free_reloc_roots(&reloc_roots);
2515
2516                 /* new reloc root may be added */
2517                 mutex_lock(&fs_info->reloc_mutex);
2518                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2519                 mutex_unlock(&fs_info->reloc_mutex);
2520                 if (!list_empty(&reloc_roots))
2521                         free_reloc_roots(&reloc_roots);
2522         }
2523
2524         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2525 }
2526
2527 static void free_block_list(struct rb_root *blocks)
2528 {
2529         struct tree_block *block;
2530         struct rb_node *rb_node;
2531         while ((rb_node = rb_first(blocks))) {
2532                 block = rb_entry(rb_node, struct tree_block, rb_node);
2533                 rb_erase(rb_node, blocks);
2534                 kfree(block);
2535         }
2536 }
2537
2538 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2539                                       struct btrfs_root *reloc_root)
2540 {
2541         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2542         struct btrfs_root *root;
2543
2544         if (reloc_root->last_trans == trans->transid)
2545                 return 0;
2546
2547         root = read_fs_root(fs_info, reloc_root->root_key.offset);
2548         BUG_ON(IS_ERR(root));
2549         BUG_ON(root->reloc_root != reloc_root);
2550
2551         return btrfs_record_root_in_trans(trans, root);
2552 }
2553
2554 static noinline_for_stack
2555 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2556                                      struct reloc_control *rc,
2557                                      struct backref_node *node,
2558                                      struct backref_edge *edges[])
2559 {
2560         struct backref_node *next;
2561         struct btrfs_root *root;
2562         int index = 0;
2563
2564         next = node;
2565         while (1) {
2566                 cond_resched();
2567                 next = walk_up_backref(next, edges, &index);
2568                 root = next->root;
2569                 BUG_ON(!root);
2570                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2571
2572                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2573                         record_reloc_root_in_trans(trans, root);
2574                         break;
2575                 }
2576
2577                 btrfs_record_root_in_trans(trans, root);
2578                 root = root->reloc_root;
2579
2580                 if (next->new_bytenr != root->node->start) {
2581                         BUG_ON(next->new_bytenr);
2582                         BUG_ON(!list_empty(&next->list));
2583                         next->new_bytenr = root->node->start;
2584                         next->root = root;
2585                         list_add_tail(&next->list,
2586                                       &rc->backref_cache.changed);
2587                         __mark_block_processed(rc, next);
2588                         break;
2589                 }
2590
2591                 WARN_ON(1);
2592                 root = NULL;
2593                 next = walk_down_backref(edges, &index);
2594                 if (!next || next->level <= node->level)
2595                         break;
2596         }
2597         if (!root)
2598                 return NULL;
2599
2600         next = node;
2601         /* setup backref node path for btrfs_reloc_cow_block */
2602         while (1) {
2603                 rc->backref_cache.path[next->level] = next;
2604                 if (--index < 0)
2605                         break;
2606                 next = edges[index]->node[UPPER];
2607         }
2608         return root;
2609 }
2610
2611 /*
2612  * select a tree root for relocation. return NULL if the block
2613  * is reference counted. we should use do_relocation() in this
2614  * case. return a tree root pointer if the block isn't reference
2615  * counted. return -ENOENT if the block is root of reloc tree.
2616  */
2617 static noinline_for_stack
2618 struct btrfs_root *select_one_root(struct backref_node *node)
2619 {
2620         struct backref_node *next;
2621         struct btrfs_root *root;
2622         struct btrfs_root *fs_root = NULL;
2623         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2624         int index = 0;
2625
2626         next = node;
2627         while (1) {
2628                 cond_resched();
2629                 next = walk_up_backref(next, edges, &index);
2630                 root = next->root;
2631                 BUG_ON(!root);
2632
2633                 /* no other choice for non-references counted tree */
2634                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2635                         return root;
2636
2637                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2638                         fs_root = root;
2639
2640                 if (next != node)
2641                         return NULL;
2642
2643                 next = walk_down_backref(edges, &index);
2644                 if (!next || next->level <= node->level)
2645                         break;
2646         }
2647
2648         if (!fs_root)
2649                 return ERR_PTR(-ENOENT);
2650         return fs_root;
2651 }
2652
2653 static noinline_for_stack
2654 u64 calcu_metadata_size(struct reloc_control *rc,
2655                         struct backref_node *node, int reserve)
2656 {
2657         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2658         struct backref_node *next = node;
2659         struct backref_edge *edge;
2660         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2661         u64 num_bytes = 0;
2662         int index = 0;
2663
2664         BUG_ON(reserve && node->processed);
2665
2666         while (next) {
2667                 cond_resched();
2668                 while (1) {
2669                         if (next->processed && (reserve || next != node))
2670                                 break;
2671
2672                         num_bytes += fs_info->nodesize;
2673
2674                         if (list_empty(&next->upper))
2675                                 break;
2676
2677                         edge = list_entry(next->upper.next,
2678                                           struct backref_edge, list[LOWER]);
2679                         edges[index++] = edge;
2680                         next = edge->node[UPPER];
2681                 }
2682                 next = walk_down_backref(edges, &index);
2683         }
2684         return num_bytes;
2685 }
2686
2687 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2688                                   struct reloc_control *rc,
2689                                   struct backref_node *node)
2690 {
2691         struct btrfs_root *root = rc->extent_root;
2692         struct btrfs_fs_info *fs_info = root->fs_info;
2693         u64 num_bytes;
2694         int ret;
2695         u64 tmp;
2696
2697         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2698
2699         trans->block_rsv = rc->block_rsv;
2700         rc->reserved_bytes += num_bytes;
2701
2702         /*
2703          * We are under a transaction here so we can only do limited flushing.
2704          * If we get an enospc just kick back -EAGAIN so we know to drop the
2705          * transaction and try to refill when we can flush all the things.
2706          */
2707         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2708                                 BTRFS_RESERVE_FLUSH_LIMIT);
2709         if (ret) {
2710                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2711                 while (tmp <= rc->reserved_bytes)
2712                         tmp <<= 1;
2713                 /*
2714                  * only one thread can access block_rsv at this point,
2715                  * so we don't need hold lock to protect block_rsv.
2716                  * we expand more reservation size here to allow enough
2717                  * space for relocation and we will return earlier in
2718                  * enospc case.
2719                  */
2720                 rc->block_rsv->size = tmp + fs_info->nodesize *
2721                                       RELOCATION_RESERVED_NODES;
2722                 return -EAGAIN;
2723         }
2724
2725         return 0;
2726 }
2727
2728 /*
2729  * relocate a block tree, and then update pointers in upper level
2730  * blocks that reference the block to point to the new location.
2731  *
2732  * if called by link_to_upper, the block has already been relocated.
2733  * in that case this function just updates pointers.
2734  */
2735 static int do_relocation(struct btrfs_trans_handle *trans,
2736                          struct reloc_control *rc,
2737                          struct backref_node *node,
2738                          struct btrfs_key *key,
2739                          struct btrfs_path *path, int lowest)
2740 {
2741         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2742         struct backref_node *upper;
2743         struct backref_edge *edge;
2744         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2745         struct btrfs_root *root;
2746         struct extent_buffer *eb;
2747         u32 blocksize;
2748         u64 bytenr;
2749         u64 generation;
2750         int slot;
2751         int ret;
2752         int err = 0;
2753
2754         BUG_ON(lowest && node->eb);
2755
2756         path->lowest_level = node->level + 1;
2757         rc->backref_cache.path[node->level] = node;
2758         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2759                 struct btrfs_key first_key;
2760                 struct btrfs_ref ref = { 0 };
2761
2762                 cond_resched();
2763
2764                 upper = edge->node[UPPER];
2765                 root = select_reloc_root(trans, rc, upper, edges);
2766                 BUG_ON(!root);
2767
2768                 if (upper->eb && !upper->locked) {
2769                         if (!lowest) {
2770                                 ret = btrfs_bin_search(upper->eb, key,
2771                                                        upper->level, &slot);
2772                                 if (ret < 0) {
2773                                         err = ret;
2774                                         goto next;
2775                                 }
2776                                 BUG_ON(ret);
2777                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2778                                 if (node->eb->start == bytenr)
2779                                         goto next;
2780                         }
2781                         drop_node_buffer(upper);
2782                 }
2783
2784                 if (!upper->eb) {
2785                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2786                         if (ret) {
2787                                 if (ret < 0)
2788                                         err = ret;
2789                                 else
2790                                         err = -ENOENT;
2791
2792                                 btrfs_release_path(path);
2793                                 break;
2794                         }
2795
2796                         if (!upper->eb) {
2797                                 upper->eb = path->nodes[upper->level];
2798                                 path->nodes[upper->level] = NULL;
2799                         } else {
2800                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2801                         }
2802
2803                         upper->locked = 1;
2804                         path->locks[upper->level] = 0;
2805
2806                         slot = path->slots[upper->level];
2807                         btrfs_release_path(path);
2808                 } else {
2809                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2810                                                &slot);
2811                         if (ret < 0) {
2812                                 err = ret;
2813                                 goto next;
2814                         }
2815                         BUG_ON(ret);
2816                 }
2817
2818                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2819                 if (lowest) {
2820                         if (bytenr != node->bytenr) {
2821                                 btrfs_err(root->fs_info,
2822                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2823                                           bytenr, node->bytenr, slot,
2824                                           upper->eb->start);
2825                                 err = -EIO;
2826                                 goto next;
2827                         }
2828                 } else {
2829                         if (node->eb->start == bytenr)
2830                                 goto next;
2831                 }
2832
2833                 blocksize = root->fs_info->nodesize;
2834                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2835                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2836                 eb = read_tree_block(fs_info, bytenr, generation,
2837                                      upper->level - 1, &first_key);
2838                 if (IS_ERR(eb)) {
2839                         err = PTR_ERR(eb);
2840                         goto next;
2841                 } else if (!extent_buffer_uptodate(eb)) {
2842                         free_extent_buffer(eb);
2843                         err = -EIO;
2844                         goto next;
2845                 }
2846                 btrfs_tree_lock(eb);
2847                 btrfs_set_lock_blocking_write(eb);
2848
2849                 if (!node->eb) {
2850                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2851                                               slot, &eb);
2852                         btrfs_tree_unlock(eb);
2853                         free_extent_buffer(eb);
2854                         if (ret < 0) {
2855                                 err = ret;
2856                                 goto next;
2857                         }
2858                         BUG_ON(node->eb != eb);
2859                 } else {
2860                         btrfs_set_node_blockptr(upper->eb, slot,
2861                                                 node->eb->start);
2862                         btrfs_set_node_ptr_generation(upper->eb, slot,
2863                                                       trans->transid);
2864                         btrfs_mark_buffer_dirty(upper->eb);
2865
2866                         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2867                                                node->eb->start, blocksize,
2868                                                upper->eb->start);
2869                         ref.real_root = root->root_key.objectid;
2870                         btrfs_init_tree_ref(&ref, node->level,
2871                                             btrfs_header_owner(upper->eb));
2872                         ret = btrfs_inc_extent_ref(trans, &ref);
2873                         BUG_ON(ret);
2874
2875                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2876                         BUG_ON(ret);
2877                 }
2878 next:
2879                 if (!upper->pending)
2880                         drop_node_buffer(upper);
2881                 else
2882                         unlock_node_buffer(upper);
2883                 if (err)
2884                         break;
2885         }
2886
2887         if (!err && node->pending) {
2888                 drop_node_buffer(node);
2889                 list_move_tail(&node->list, &rc->backref_cache.changed);
2890                 node->pending = 0;
2891         }
2892
2893         path->lowest_level = 0;
2894         BUG_ON(err == -ENOSPC);
2895         return err;
2896 }
2897
2898 static int link_to_upper(struct btrfs_trans_handle *trans,
2899                          struct reloc_control *rc,
2900                          struct backref_node *node,
2901                          struct btrfs_path *path)
2902 {
2903         struct btrfs_key key;
2904
2905         btrfs_node_key_to_cpu(node->eb, &key, 0);
2906         return do_relocation(trans, rc, node, &key, path, 0);
2907 }
2908
2909 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2910                                 struct reloc_control *rc,
2911                                 struct btrfs_path *path, int err)
2912 {
2913         LIST_HEAD(list);
2914         struct backref_cache *cache = &rc->backref_cache;
2915         struct backref_node *node;
2916         int level;
2917         int ret;
2918
2919         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2920                 while (!list_empty(&cache->pending[level])) {
2921                         node = list_entry(cache->pending[level].next,
2922                                           struct backref_node, list);
2923                         list_move_tail(&node->list, &list);
2924                         BUG_ON(!node->pending);
2925
2926                         if (!err) {
2927                                 ret = link_to_upper(trans, rc, node, path);
2928                                 if (ret < 0)
2929                                         err = ret;
2930                         }
2931                 }
2932                 list_splice_init(&list, &cache->pending[level]);
2933         }
2934         return err;
2935 }
2936
2937 static void mark_block_processed(struct reloc_control *rc,
2938                                  u64 bytenr, u32 blocksize)
2939 {
2940         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2941                         EXTENT_DIRTY);
2942 }
2943
2944 static void __mark_block_processed(struct reloc_control *rc,
2945                                    struct backref_node *node)
2946 {
2947         u32 blocksize;
2948         if (node->level == 0 ||
2949             in_block_group(node->bytenr, rc->block_group)) {
2950                 blocksize = rc->extent_root->fs_info->nodesize;
2951                 mark_block_processed(rc, node->bytenr, blocksize);
2952         }
2953         node->processed = 1;
2954 }
2955
2956 /*
2957  * mark a block and all blocks directly/indirectly reference the block
2958  * as processed.
2959  */
2960 static void update_processed_blocks(struct reloc_control *rc,
2961                                     struct backref_node *node)
2962 {
2963         struct backref_node *next = node;
2964         struct backref_edge *edge;
2965         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2966         int index = 0;
2967
2968         while (next) {
2969                 cond_resched();
2970                 while (1) {
2971                         if (next->processed)
2972                                 break;
2973
2974                         __mark_block_processed(rc, next);
2975
2976                         if (list_empty(&next->upper))
2977                                 break;
2978
2979                         edge = list_entry(next->upper.next,
2980                                           struct backref_edge, list[LOWER]);
2981                         edges[index++] = edge;
2982                         next = edge->node[UPPER];
2983                 }
2984                 next = walk_down_backref(edges, &index);
2985         }
2986 }
2987
2988 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2989 {
2990         u32 blocksize = rc->extent_root->fs_info->nodesize;
2991
2992         if (test_range_bit(&rc->processed_blocks, bytenr,
2993                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2994                 return 1;
2995         return 0;
2996 }
2997
2998 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2999                               struct tree_block *block)
3000 {
3001         struct extent_buffer *eb;
3002
3003         BUG_ON(block->key_ready);
3004         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
3005                              block->level, NULL);
3006         if (IS_ERR(eb)) {
3007                 return PTR_ERR(eb);
3008         } else if (!extent_buffer_uptodate(eb)) {
3009                 free_extent_buffer(eb);
3010                 return -EIO;
3011         }
3012         if (block->level == 0)
3013                 btrfs_item_key_to_cpu(eb, &block->key, 0);
3014         else
3015                 btrfs_node_key_to_cpu(eb, &block->key, 0);
3016         free_extent_buffer(eb);
3017         block->key_ready = 1;
3018         return 0;
3019 }
3020
3021 /*
3022  * helper function to relocate a tree block
3023  */
3024 static int relocate_tree_block(struct btrfs_trans_handle *trans,
3025                                 struct reloc_control *rc,
3026                                 struct backref_node *node,
3027                                 struct btrfs_key *key,
3028                                 struct btrfs_path *path)
3029 {
3030         struct btrfs_root *root;
3031         int ret = 0;
3032
3033         if (!node)
3034                 return 0;
3035
3036         BUG_ON(node->processed);
3037         root = select_one_root(node);
3038         if (root == ERR_PTR(-ENOENT)) {
3039                 update_processed_blocks(rc, node);
3040                 goto out;
3041         }
3042
3043         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3044                 ret = reserve_metadata_space(trans, rc, node);
3045                 if (ret)
3046                         goto out;
3047         }
3048
3049         if (root) {
3050                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3051                         BUG_ON(node->new_bytenr);
3052                         BUG_ON(!list_empty(&node->list));
3053                         btrfs_record_root_in_trans(trans, root);
3054                         root = root->reloc_root;
3055                         node->new_bytenr = root->node->start;
3056                         node->root = root;
3057                         list_add_tail(&node->list, &rc->backref_cache.changed);
3058                 } else {
3059                         path->lowest_level = node->level;
3060                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
3061                         btrfs_release_path(path);
3062                         if (ret > 0)
3063                                 ret = 0;
3064                 }
3065                 if (!ret)
3066                         update_processed_blocks(rc, node);
3067         } else {
3068                 ret = do_relocation(trans, rc, node, key, path, 1);
3069         }
3070 out:
3071         if (ret || node->level == 0 || node->cowonly)
3072                 remove_backref_node(&rc->backref_cache, node);
3073         return ret;
3074 }
3075
3076 /*
3077  * relocate a list of blocks
3078  */
3079 static noinline_for_stack
3080 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
3081                          struct reloc_control *rc, struct rb_root *blocks)
3082 {
3083         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3084         struct backref_node *node;
3085         struct btrfs_path *path;
3086         struct tree_block *block;
3087         struct tree_block *next;
3088         int ret;
3089         int err = 0;
3090
3091         path = btrfs_alloc_path();
3092         if (!path) {
3093                 err = -ENOMEM;
3094                 goto out_free_blocks;
3095         }
3096
3097         /* Kick in readahead for tree blocks with missing keys */
3098         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3099                 if (!block->key_ready)
3100                         readahead_tree_block(fs_info, block->bytenr);
3101         }
3102
3103         /* Get first keys */
3104         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3105                 if (!block->key_ready) {
3106                         err = get_tree_block_key(fs_info, block);
3107                         if (err)
3108                                 goto out_free_path;
3109                 }
3110         }
3111
3112         /* Do tree relocation */
3113         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3114                 node = build_backref_tree(rc, &block->key,
3115                                           block->level, block->bytenr);
3116                 if (IS_ERR(node)) {
3117                         err = PTR_ERR(node);
3118                         goto out;
3119                 }
3120
3121                 ret = relocate_tree_block(trans, rc, node, &block->key,
3122                                           path);
3123                 if (ret < 0) {
3124                         if (ret != -EAGAIN || &block->rb_node == rb_first(blocks))
3125                                 err = ret;
3126                         goto out;
3127                 }
3128         }
3129 out:
3130         err = finish_pending_nodes(trans, rc, path, err);
3131
3132 out_free_path:
3133         btrfs_free_path(path);
3134 out_free_blocks:
3135         free_block_list(blocks);
3136         return err;
3137 }
3138
3139 static noinline_for_stack
3140 int prealloc_file_extent_cluster(struct inode *inode,
3141                                  struct file_extent_cluster *cluster)
3142 {
3143         u64 alloc_hint = 0;
3144         u64 start;
3145         u64 end;
3146         u64 offset = BTRFS_I(inode)->index_cnt;
3147         u64 num_bytes;
3148         int nr = 0;
3149         int ret = 0;
3150         u64 prealloc_start = cluster->start - offset;
3151         u64 prealloc_end = cluster->end - offset;
3152         u64 cur_offset;
3153         struct extent_changeset *data_reserved = NULL;
3154
3155         BUG_ON(cluster->start != cluster->boundary[0]);
3156         inode_lock(inode);
3157
3158         ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
3159                                           prealloc_end + 1 - prealloc_start);
3160         if (ret)
3161                 goto out;
3162
3163         cur_offset = prealloc_start;
3164         while (nr < cluster->nr) {
3165                 start = cluster->boundary[nr] - offset;
3166                 if (nr + 1 < cluster->nr)
3167                         end = cluster->boundary[nr + 1] - 1 - offset;
3168                 else
3169                         end = cluster->end - offset;
3170
3171                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3172                 num_bytes = end + 1 - start;
3173                 if (cur_offset < start)
3174                         btrfs_free_reserved_data_space(inode, data_reserved,
3175                                         cur_offset, start - cur_offset);
3176                 ret = btrfs_prealloc_file_range(inode, 0, start,
3177                                                 num_bytes, num_bytes,
3178                                                 end + 1, &alloc_hint);
3179                 cur_offset = end + 1;
3180                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3181                 if (ret)
3182                         break;
3183                 nr++;
3184         }
3185         if (cur_offset < prealloc_end)
3186                 btrfs_free_reserved_data_space(inode, data_reserved,
3187                                 cur_offset, prealloc_end + 1 - cur_offset);
3188 out:
3189         inode_unlock(inode);
3190         extent_changeset_free(data_reserved);
3191         return ret;
3192 }
3193
3194 static noinline_for_stack
3195 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3196                          u64 block_start)
3197 {
3198         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3199         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3200         struct extent_map *em;
3201         int ret = 0;
3202
3203         em = alloc_extent_map();
3204         if (!em)
3205                 return -ENOMEM;
3206
3207         em->start = start;
3208         em->len = end + 1 - start;
3209         em->block_len = em->len;
3210         em->block_start = block_start;
3211         em->bdev = fs_info->fs_devices->latest_bdev;
3212         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3213
3214         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3215         while (1) {
3216                 write_lock(&em_tree->lock);
3217                 ret = add_extent_mapping(em_tree, em, 0);
3218                 write_unlock(&em_tree->lock);
3219                 if (ret != -EEXIST) {
3220                         free_extent_map(em);
3221                         break;
3222                 }
3223                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3224         }
3225         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3226         return ret;
3227 }
3228
3229 static int relocate_file_extent_cluster(struct inode *inode,
3230                                         struct file_extent_cluster *cluster)
3231 {
3232         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3233         u64 page_start;
3234         u64 page_end;
3235         u64 offset = BTRFS_I(inode)->index_cnt;
3236         unsigned long index;
3237         unsigned long last_index;
3238         struct page *page;
3239         struct file_ra_state *ra;
3240         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3241         int nr = 0;
3242         int ret = 0;
3243
3244         if (!cluster->nr)
3245                 return 0;
3246
3247         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3248         if (!ra)
3249                 return -ENOMEM;
3250
3251         ret = prealloc_file_extent_cluster(inode, cluster);
3252         if (ret)
3253                 goto out;
3254
3255         file_ra_state_init(ra, inode->i_mapping);
3256
3257         ret = setup_extent_mapping(inode, cluster->start - offset,
3258                                    cluster->end - offset, cluster->start);
3259         if (ret)
3260                 goto out;
3261
3262         index = (cluster->start - offset) >> PAGE_SHIFT;
3263         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3264         while (index <= last_index) {
3265                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3266                                 PAGE_SIZE);
3267                 if (ret)
3268                         goto out;
3269
3270                 page = find_lock_page(inode->i_mapping, index);
3271                 if (!page) {
3272                         page_cache_sync_readahead(inode->i_mapping,
3273                                                   ra, NULL, index,
3274                                                   last_index + 1 - index);
3275                         page = find_or_create_page(inode->i_mapping, index,
3276                                                    mask);
3277                         if (!page) {
3278                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3279                                                         PAGE_SIZE, true);
3280                                 ret = -ENOMEM;
3281                                 goto out;
3282                         }
3283                 }
3284
3285                 if (PageReadahead(page)) {
3286                         page_cache_async_readahead(inode->i_mapping,
3287                                                    ra, NULL, page, index,
3288                                                    last_index + 1 - index);
3289                 }
3290
3291                 if (!PageUptodate(page)) {
3292                         btrfs_readpage(NULL, page);
3293                         lock_page(page);
3294                         if (!PageUptodate(page)) {
3295                                 unlock_page(page);
3296                                 put_page(page);
3297                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3298                                                         PAGE_SIZE, true);
3299                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
3300                                                                PAGE_SIZE, true);
3301                                 ret = -EIO;
3302                                 goto out;
3303                         }
3304                 }
3305
3306                 page_start = page_offset(page);
3307                 page_end = page_start + PAGE_SIZE - 1;
3308
3309                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3310
3311                 set_page_extent_mapped(page);
3312
3313                 if (nr < cluster->nr &&
3314                     page_start + offset == cluster->boundary[nr]) {
3315                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3316                                         page_start, page_end,
3317                                         EXTENT_BOUNDARY);
3318                         nr++;
3319                 }
3320
3321                 ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
3322                                                 NULL);
3323                 if (ret) {
3324                         unlock_page(page);
3325                         put_page(page);
3326                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3327                                                          PAGE_SIZE, true);
3328                         btrfs_delalloc_release_extents(BTRFS_I(inode),
3329                                                        PAGE_SIZE, true);
3330
3331                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
3332                                           page_start, page_end,
3333                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
3334                         goto out;
3335
3336                 }
3337                 set_page_dirty(page);
3338
3339                 unlock_extent(&BTRFS_I(inode)->io_tree,
3340                               page_start, page_end);
3341                 unlock_page(page);
3342                 put_page(page);
3343
3344                 index++;
3345                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE,
3346                                                false);
3347                 balance_dirty_pages_ratelimited(inode->i_mapping);
3348                 btrfs_throttle(fs_info);
3349         }
3350         WARN_ON(nr != cluster->nr);
3351 out:
3352         kfree(ra);
3353         return ret;
3354 }
3355
3356 static noinline_for_stack
3357 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3358                          struct file_extent_cluster *cluster)
3359 {
3360         int ret;
3361
3362         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3363                 ret = relocate_file_extent_cluster(inode, cluster);
3364                 if (ret)
3365                         return ret;
3366                 cluster->nr = 0;
3367         }
3368
3369         if (!cluster->nr)
3370                 cluster->start = extent_key->objectid;
3371         else
3372                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3373         cluster->end = extent_key->objectid + extent_key->offset - 1;
3374         cluster->boundary[cluster->nr] = extent_key->objectid;
3375         cluster->nr++;
3376
3377         if (cluster->nr >= MAX_EXTENTS) {
3378                 ret = relocate_file_extent_cluster(inode, cluster);
3379                 if (ret)
3380                         return ret;
3381                 cluster->nr = 0;
3382         }
3383         return 0;
3384 }
3385
3386 /*
3387  * helper to add a tree block to the list.
3388  * the major work is getting the generation and level of the block
3389  */
3390 static int add_tree_block(struct reloc_control *rc,
3391                           struct btrfs_key *extent_key,
3392                           struct btrfs_path *path,
3393                           struct rb_root *blocks)
3394 {
3395         struct extent_buffer *eb;
3396         struct btrfs_extent_item *ei;
3397         struct btrfs_tree_block_info *bi;
3398         struct tree_block *block;
3399         struct rb_node *rb_node;
3400         u32 item_size;
3401         int level = -1;
3402         u64 generation;
3403
3404         eb =  path->nodes[0];
3405         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3406
3407         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3408             item_size >= sizeof(*ei) + sizeof(*bi)) {
3409                 ei = btrfs_item_ptr(eb, path->slots[0],
3410                                 struct btrfs_extent_item);
3411                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3412                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3413                         level = btrfs_tree_block_level(eb, bi);
3414                 } else {
3415                         level = (int)extent_key->offset;
3416                 }
3417                 generation = btrfs_extent_generation(eb, ei);
3418         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3419                 btrfs_print_v0_err(eb->fs_info);
3420                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3421                 return -EINVAL;
3422         } else {
3423                 BUG();
3424         }
3425
3426         btrfs_release_path(path);
3427
3428         BUG_ON(level == -1);
3429
3430         block = kmalloc(sizeof(*block), GFP_NOFS);
3431         if (!block)
3432                 return -ENOMEM;
3433
3434         block->bytenr = extent_key->objectid;
3435         block->key.objectid = rc->extent_root->fs_info->nodesize;
3436         block->key.offset = generation;
3437         block->level = level;
3438         block->key_ready = 0;
3439
3440         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3441         if (rb_node)
3442                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3443
3444         return 0;
3445 }
3446
3447 /*
3448  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3449  */
3450 static int __add_tree_block(struct reloc_control *rc,
3451                             u64 bytenr, u32 blocksize,
3452                             struct rb_root *blocks)
3453 {
3454         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3455         struct btrfs_path *path;
3456         struct btrfs_key key;
3457         int ret;
3458         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3459
3460         if (tree_block_processed(bytenr, rc))
3461                 return 0;
3462
3463         if (tree_search(blocks, bytenr))
3464                 return 0;
3465
3466         path = btrfs_alloc_path();
3467         if (!path)
3468                 return -ENOMEM;
3469 again:
3470         key.objectid = bytenr;
3471         if (skinny) {
3472                 key.type = BTRFS_METADATA_ITEM_KEY;
3473                 key.offset = (u64)-1;
3474         } else {
3475                 key.type = BTRFS_EXTENT_ITEM_KEY;
3476                 key.offset = blocksize;
3477         }
3478
3479         path->search_commit_root = 1;
3480         path->skip_locking = 1;
3481         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3482         if (ret < 0)
3483                 goto out;
3484
3485         if (ret > 0 && skinny) {
3486                 if (path->slots[0]) {
3487                         path->slots[0]--;
3488                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3489                                               path->slots[0]);
3490                         if (key.objectid == bytenr &&
3491                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3492                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3493                               key.offset == blocksize)))
3494                                 ret = 0;
3495                 }
3496
3497                 if (ret) {
3498                         skinny = false;
3499                         btrfs_release_path(path);
3500                         goto again;
3501                 }
3502         }
3503         if (ret) {
3504                 ASSERT(ret == 1);
3505                 btrfs_print_leaf(path->nodes[0]);
3506                 btrfs_err(fs_info,
3507              "tree block extent item (%llu) is not found in extent tree",
3508                      bytenr);
3509                 WARN_ON(1);
3510                 ret = -EINVAL;
3511                 goto out;
3512         }
3513
3514         ret = add_tree_block(rc, &key, path, blocks);
3515 out:
3516         btrfs_free_path(path);
3517         return ret;
3518 }
3519
3520 /*
3521  * helper to check if the block use full backrefs for pointers in it
3522  */
3523 static int block_use_full_backref(struct reloc_control *rc,
3524                                   struct extent_buffer *eb)
3525 {
3526         u64 flags;
3527         int ret;
3528
3529         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3530             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3531                 return 1;
3532
3533         ret = btrfs_lookup_extent_info(NULL, rc->extent_root->fs_info,
3534                                        eb->start, btrfs_header_level(eb), 1,
3535                                        NULL, &flags);
3536         BUG_ON(ret);
3537
3538         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3539                 ret = 1;
3540         else
3541                 ret = 0;
3542         return ret;
3543 }
3544
3545 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3546                                     struct btrfs_block_group_cache *block_group,
3547                                     struct inode *inode,
3548                                     u64 ino)
3549 {
3550         struct btrfs_key key;
3551         struct btrfs_root *root = fs_info->tree_root;
3552         struct btrfs_trans_handle *trans;
3553         int ret = 0;
3554
3555         if (inode)
3556                 goto truncate;
3557
3558         key.objectid = ino;
3559         key.type = BTRFS_INODE_ITEM_KEY;
3560         key.offset = 0;
3561
3562         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3563         if (IS_ERR(inode))
3564                 return -ENOENT;
3565
3566 truncate:
3567         ret = btrfs_check_trunc_cache_free_space(fs_info,
3568                                                  &fs_info->global_block_rsv);
3569         if (ret)
3570                 goto out;
3571
3572         trans = btrfs_join_transaction(root);
3573         if (IS_ERR(trans)) {
3574                 ret = PTR_ERR(trans);
3575                 goto out;
3576         }
3577
3578         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3579
3580         btrfs_end_transaction(trans);
3581         btrfs_btree_balance_dirty(fs_info);
3582 out:
3583         iput(inode);
3584         return ret;
3585 }
3586
3587 /*
3588  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3589  * this function scans fs tree to find blocks reference the data extent
3590  */
3591 static int find_data_references(struct reloc_control *rc,
3592                                 struct btrfs_key *extent_key,
3593                                 struct extent_buffer *leaf,
3594                                 struct btrfs_extent_data_ref *ref,
3595                                 struct rb_root *blocks)
3596 {
3597         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3598         struct btrfs_path *path;
3599         struct tree_block *block;
3600         struct btrfs_root *root;
3601         struct btrfs_file_extent_item *fi;
3602         struct rb_node *rb_node;
3603         struct btrfs_key key;
3604         u64 ref_root;
3605         u64 ref_objectid;
3606         u64 ref_offset;
3607         u32 ref_count;
3608         u32 nritems;
3609         int err = 0;
3610         int added = 0;
3611         int counted;
3612         int ret;
3613
3614         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3615         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3616         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3617         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3618
3619         /*
3620          * This is an extent belonging to the free space cache, lets just delete
3621          * it and redo the search.
3622          */
3623         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3624                 ret = delete_block_group_cache(fs_info, rc->block_group,
3625                                                NULL, ref_objectid);
3626                 if (ret != -ENOENT)
3627                         return ret;
3628                 ret = 0;
3629         }
3630
3631         path = btrfs_alloc_path();
3632         if (!path)
3633                 return -ENOMEM;
3634         path->reada = READA_FORWARD;
3635
3636         root = read_fs_root(fs_info, ref_root);
3637         if (IS_ERR(root)) {
3638                 err = PTR_ERR(root);
3639                 goto out;
3640         }
3641
3642         key.objectid = ref_objectid;
3643         key.type = BTRFS_EXTENT_DATA_KEY;
3644         if (ref_offset > ((u64)-1 << 32))
3645                 key.offset = 0;
3646         else
3647                 key.offset = ref_offset;
3648
3649         path->search_commit_root = 1;
3650         path->skip_locking = 1;
3651         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3652         if (ret < 0) {
3653                 err = ret;
3654                 goto out;
3655         }
3656
3657         leaf = path->nodes[0];
3658         nritems = btrfs_header_nritems(leaf);
3659         /*
3660          * the references in tree blocks that use full backrefs
3661          * are not counted in
3662          */
3663         if (block_use_full_backref(rc, leaf))
3664                 counted = 0;
3665         else
3666                 counted = 1;
3667         rb_node = tree_search(blocks, leaf->start);
3668         if (rb_node) {
3669                 if (counted)
3670                         added = 1;
3671                 else
3672                         path->slots[0] = nritems;
3673         }
3674
3675         while (ref_count > 0) {
3676                 while (path->slots[0] >= nritems) {
3677                         ret = btrfs_next_leaf(root, path);
3678                         if (ret < 0) {
3679                                 err = ret;
3680                                 goto out;
3681                         }
3682                         if (WARN_ON(ret > 0))
3683                                 goto out;
3684
3685                         leaf = path->nodes[0];
3686                         nritems = btrfs_header_nritems(leaf);
3687                         added = 0;
3688
3689                         if (block_use_full_backref(rc, leaf))
3690                                 counted = 0;
3691                         else
3692                                 counted = 1;
3693                         rb_node = tree_search(blocks, leaf->start);
3694                         if (rb_node) {
3695                                 if (counted)
3696                                         added = 1;
3697                                 else
3698                                         path->slots[0] = nritems;
3699                         }
3700                 }
3701
3702                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3703                 if (WARN_ON(key.objectid != ref_objectid ||
3704                     key.type != BTRFS_EXTENT_DATA_KEY))
3705                         break;
3706
3707                 fi = btrfs_item_ptr(leaf, path->slots[0],
3708                                     struct btrfs_file_extent_item);
3709
3710                 if (btrfs_file_extent_type(leaf, fi) ==
3711                     BTRFS_FILE_EXTENT_INLINE)
3712                         goto next;
3713
3714                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3715                     extent_key->objectid)
3716                         goto next;
3717
3718                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3719                 if (key.offset != ref_offset)
3720                         goto next;
3721
3722                 if (counted)
3723                         ref_count--;
3724                 if (added)
3725                         goto next;
3726
3727                 if (!tree_block_processed(leaf->start, rc)) {
3728                         block = kmalloc(sizeof(*block), GFP_NOFS);
3729                         if (!block) {
3730                                 err = -ENOMEM;
3731                                 break;
3732                         }
3733                         block->bytenr = leaf->start;
3734                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3735                         block->level = 0;
3736                         block->key_ready = 1;
3737                         rb_node = tree_insert(blocks, block->bytenr,
3738                                               &block->rb_node);
3739                         if (rb_node)
3740                                 backref_tree_panic(rb_node, -EEXIST,
3741                                                    block->bytenr);
3742                 }
3743                 if (counted)
3744                         added = 1;
3745                 else
3746                         path->slots[0] = nritems;
3747 next:
3748                 path->slots[0]++;
3749
3750         }
3751 out:
3752         btrfs_free_path(path);
3753         return err;
3754 }
3755
3756 /*
3757  * helper to find all tree blocks that reference a given data extent
3758  */
3759 static noinline_for_stack
3760 int add_data_references(struct reloc_control *rc,
3761                         struct btrfs_key *extent_key,
3762                         struct btrfs_path *path,
3763                         struct rb_root *blocks)
3764 {
3765         struct btrfs_key key;
3766         struct extent_buffer *eb;
3767         struct btrfs_extent_data_ref *dref;
3768         struct btrfs_extent_inline_ref *iref;
3769         unsigned long ptr;
3770         unsigned long end;
3771         u32 blocksize = rc->extent_root->fs_info->nodesize;
3772         int ret = 0;
3773         int err = 0;
3774
3775         eb = path->nodes[0];
3776         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3777         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3778         ptr += sizeof(struct btrfs_extent_item);
3779
3780         while (ptr < end) {
3781                 iref = (struct btrfs_extent_inline_ref *)ptr;
3782                 key.type = btrfs_get_extent_inline_ref_type(eb, iref,
3783                                                         BTRFS_REF_TYPE_DATA);
3784                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3785                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3786                         ret = __add_tree_block(rc, key.offset, blocksize,
3787                                                blocks);
3788                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3789                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3790                         ret = find_data_references(rc, extent_key,
3791                                                    eb, dref, blocks);
3792                 } else {
3793                         ret = -EUCLEAN;
3794                         btrfs_err(rc->extent_root->fs_info,
3795                      "extent %llu slot %d has an invalid inline ref type",
3796                              eb->start, path->slots[0]);
3797                 }
3798                 if (ret) {
3799                         err = ret;
3800                         goto out;
3801                 }
3802                 ptr += btrfs_extent_inline_ref_size(key.type);
3803         }
3804         WARN_ON(ptr > end);
3805
3806         while (1) {
3807                 cond_resched();
3808                 eb = path->nodes[0];
3809                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3810                         ret = btrfs_next_leaf(rc->extent_root, path);
3811                         if (ret < 0) {
3812                                 err = ret;
3813                                 break;
3814                         }
3815                         if (ret > 0)
3816                                 break;
3817                         eb = path->nodes[0];
3818                 }
3819
3820                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3821                 if (key.objectid != extent_key->objectid)
3822                         break;
3823
3824                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3825                         ret = __add_tree_block(rc, key.offset, blocksize,
3826                                                blocks);
3827                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3828                         dref = btrfs_item_ptr(eb, path->slots[0],
3829                                               struct btrfs_extent_data_ref);
3830                         ret = find_data_references(rc, extent_key,
3831                                                    eb, dref, blocks);
3832                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
3833                         btrfs_print_v0_err(eb->fs_info);
3834                         btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3835                         ret = -EINVAL;
3836                 } else {
3837                         ret = 0;
3838                 }
3839                 if (ret) {
3840                         err = ret;
3841                         break;
3842                 }
3843                 path->slots[0]++;
3844         }
3845 out:
3846         btrfs_release_path(path);
3847         if (err)
3848                 free_block_list(blocks);
3849         return err;
3850 }
3851
3852 /*
3853  * helper to find next unprocessed extent
3854  */
3855 static noinline_for_stack
3856 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3857                      struct btrfs_key *extent_key)
3858 {
3859         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3860         struct btrfs_key key;
3861         struct extent_buffer *leaf;
3862         u64 start, end, last;
3863         int ret;
3864
3865         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3866         while (1) {
3867                 cond_resched();
3868                 if (rc->search_start >= last) {
3869                         ret = 1;
3870                         break;
3871                 }
3872
3873                 key.objectid = rc->search_start;
3874                 key.type = BTRFS_EXTENT_ITEM_KEY;
3875                 key.offset = 0;
3876
3877                 path->search_commit_root = 1;
3878                 path->skip_locking = 1;
3879                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3880                                         0, 0);
3881                 if (ret < 0)
3882                         break;
3883 next:
3884                 leaf = path->nodes[0];
3885                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3886                         ret = btrfs_next_leaf(rc->extent_root, path);
3887                         if (ret != 0)
3888                                 break;
3889                         leaf = path->nodes[0];
3890                 }
3891
3892                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3893                 if (key.objectid >= last) {
3894                         ret = 1;
3895                         break;
3896                 }
3897
3898                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3899                     key.type != BTRFS_METADATA_ITEM_KEY) {
3900                         path->slots[0]++;
3901                         goto next;
3902                 }
3903
3904                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3905                     key.objectid + key.offset <= rc->search_start) {
3906                         path->slots[0]++;
3907                         goto next;
3908                 }
3909
3910                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3911                     key.objectid + fs_info->nodesize <=
3912                     rc->search_start) {
3913                         path->slots[0]++;
3914                         goto next;
3915                 }
3916
3917                 ret = find_first_extent_bit(&rc->processed_blocks,
3918                                             key.objectid, &start, &end,
3919                                             EXTENT_DIRTY, NULL);
3920
3921                 if (ret == 0 && start <= key.objectid) {
3922                         btrfs_release_path(path);
3923                         rc->search_start = end + 1;
3924                 } else {
3925                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3926                                 rc->search_start = key.objectid + key.offset;
3927                         else
3928                                 rc->search_start = key.objectid +
3929                                         fs_info->nodesize;
3930                         memcpy(extent_key, &key, sizeof(key));
3931                         return 0;
3932                 }
3933         }
3934         btrfs_release_path(path);
3935         return ret;
3936 }
3937
3938 static void set_reloc_control(struct reloc_control *rc)
3939 {
3940         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3941
3942         mutex_lock(&fs_info->reloc_mutex);
3943         fs_info->reloc_ctl = rc;
3944         mutex_unlock(&fs_info->reloc_mutex);
3945 }
3946
3947 static void unset_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 = NULL;
3953         mutex_unlock(&fs_info->reloc_mutex);
3954 }
3955
3956 static int check_extent_flags(u64 flags)
3957 {
3958         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3959             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3960                 return 1;
3961         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3962             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3963                 return 1;
3964         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3965             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3966                 return 1;
3967         return 0;
3968 }
3969
3970 static noinline_for_stack
3971 int prepare_to_relocate(struct reloc_control *rc)
3972 {
3973         struct btrfs_trans_handle *trans;
3974         int ret;
3975
3976         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3977                                               BTRFS_BLOCK_RSV_TEMP);
3978         if (!rc->block_rsv)
3979                 return -ENOMEM;
3980
3981         memset(&rc->cluster, 0, sizeof(rc->cluster));
3982         rc->search_start = rc->block_group->key.objectid;
3983         rc->extents_found = 0;
3984         rc->nodes_relocated = 0;
3985         rc->merging_rsv_size = 0;
3986         rc->reserved_bytes = 0;
3987         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3988                               RELOCATION_RESERVED_NODES;
3989         ret = btrfs_block_rsv_refill(rc->extent_root,
3990                                      rc->block_rsv, rc->block_rsv->size,
3991                                      BTRFS_RESERVE_FLUSH_ALL);
3992         if (ret)
3993                 return ret;
3994
3995         rc->create_reloc_tree = 1;
3996         set_reloc_control(rc);
3997
3998         trans = btrfs_join_transaction(rc->extent_root);
3999         if (IS_ERR(trans)) {
4000                 unset_reloc_control(rc);
4001                 /*
4002                  * extent tree is not a ref_cow tree and has no reloc_root to
4003                  * cleanup.  And callers are responsible to free the above
4004                  * block rsv.
4005                  */
4006                 return PTR_ERR(trans);
4007         }
4008         btrfs_commit_transaction(trans);
4009         return 0;
4010 }
4011
4012 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
4013 {
4014         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
4015         struct rb_root blocks = RB_ROOT;
4016         struct btrfs_key key;
4017         struct btrfs_trans_handle *trans = NULL;
4018         struct btrfs_path *path;
4019         struct btrfs_extent_item *ei;
4020         u64 flags;
4021         u32 item_size;
4022         int ret;
4023         int err = 0;
4024         int progress = 0;
4025
4026         path = btrfs_alloc_path();
4027         if (!path)
4028                 return -ENOMEM;
4029         path->reada = READA_FORWARD;
4030
4031         ret = prepare_to_relocate(rc);
4032         if (ret) {
4033                 err = ret;
4034                 goto out_free;
4035         }
4036
4037         while (1) {
4038                 rc->reserved_bytes = 0;
4039                 ret = btrfs_block_rsv_refill(rc->extent_root,
4040                                         rc->block_rsv, rc->block_rsv->size,
4041                                         BTRFS_RESERVE_FLUSH_ALL);
4042                 if (ret) {
4043                         err = ret;
4044                         break;
4045                 }
4046                 progress++;
4047                 trans = btrfs_start_transaction(rc->extent_root, 0);
4048                 if (IS_ERR(trans)) {
4049                         err = PTR_ERR(trans);
4050                         trans = NULL;
4051                         break;
4052                 }
4053 restart:
4054                 if (update_backref_cache(trans, &rc->backref_cache)) {
4055                         btrfs_end_transaction(trans);
4056                         trans = NULL;
4057                         continue;
4058                 }
4059
4060                 ret = find_next_extent(rc, path, &key);
4061                 if (ret < 0)
4062                         err = ret;
4063                 if (ret != 0)
4064                         break;
4065
4066                 rc->extents_found++;
4067
4068                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4069                                     struct btrfs_extent_item);
4070                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4071                 if (item_size >= sizeof(*ei)) {
4072                         flags = btrfs_extent_flags(path->nodes[0], ei);
4073                         ret = check_extent_flags(flags);
4074                         BUG_ON(ret);
4075                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
4076                         err = -EINVAL;
4077                         btrfs_print_v0_err(trans->fs_info);
4078                         btrfs_abort_transaction(trans, err);
4079                         break;
4080                 } else {
4081                         BUG();
4082                 }
4083
4084                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4085                         ret = add_tree_block(rc, &key, path, &blocks);
4086                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4087                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4088                         ret = add_data_references(rc, &key, path, &blocks);
4089                 } else {
4090                         btrfs_release_path(path);
4091                         ret = 0;
4092                 }
4093                 if (ret < 0) {
4094                         err = ret;
4095                         break;
4096                 }
4097
4098                 if (!RB_EMPTY_ROOT(&blocks)) {
4099                         ret = relocate_tree_blocks(trans, rc, &blocks);
4100                         if (ret < 0) {
4101                                 /*
4102                                  * if we fail to relocate tree blocks, force to update
4103                                  * backref cache when committing transaction.
4104                                  */
4105                                 rc->backref_cache.last_trans = trans->transid - 1;
4106
4107                                 if (ret != -EAGAIN) {
4108                                         err = ret;
4109                                         break;
4110                                 }
4111                                 rc->extents_found--;
4112                                 rc->search_start = key.objectid;
4113                         }
4114                 }
4115
4116                 btrfs_end_transaction_throttle(trans);
4117                 btrfs_btree_balance_dirty(fs_info);
4118                 trans = NULL;
4119
4120                 if (rc->stage == MOVE_DATA_EXTENTS &&
4121                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4122                         rc->found_file_extent = 1;
4123                         ret = relocate_data_extent(rc->data_inode,
4124                                                    &key, &rc->cluster);
4125                         if (ret < 0) {
4126                                 err = ret;
4127                                 break;
4128                         }
4129                 }
4130         }
4131         if (trans && progress && err == -ENOSPC) {
4132                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
4133                 if (ret == 1) {
4134                         err = 0;
4135                         progress = 0;
4136                         goto restart;
4137                 }
4138         }
4139
4140         btrfs_release_path(path);
4141         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4142
4143         if (trans) {
4144                 btrfs_end_transaction_throttle(trans);
4145                 btrfs_btree_balance_dirty(fs_info);
4146         }
4147
4148         if (!err) {
4149                 ret = relocate_file_extent_cluster(rc->data_inode,
4150                                                    &rc->cluster);
4151                 if (ret < 0)
4152                         err = ret;
4153         }
4154
4155         rc->create_reloc_tree = 0;
4156         set_reloc_control(rc);
4157
4158         backref_cache_cleanup(&rc->backref_cache);
4159         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4160
4161         err = prepare_to_merge(rc, err);
4162
4163         merge_reloc_roots(rc);
4164
4165         rc->merge_reloc_tree = 0;
4166         unset_reloc_control(rc);
4167         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4168
4169         /* get rid of pinned extents */
4170         trans = btrfs_join_transaction(rc->extent_root);
4171         if (IS_ERR(trans)) {
4172                 err = PTR_ERR(trans);
4173                 goto out_free;
4174         }
4175         btrfs_commit_transaction(trans);
4176         ret = clean_dirty_subvols(rc);
4177         if (ret < 0 && !err)
4178                 err = ret;
4179 out_free:
4180         btrfs_free_block_rsv(fs_info, rc->block_rsv);
4181         btrfs_free_path(path);
4182         return err;
4183 }
4184
4185 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4186                                  struct btrfs_root *root, u64 objectid)
4187 {
4188         struct btrfs_path *path;
4189         struct btrfs_inode_item *item;
4190         struct extent_buffer *leaf;
4191         int ret;
4192
4193         path = btrfs_alloc_path();
4194         if (!path)
4195                 return -ENOMEM;
4196
4197         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4198         if (ret)
4199                 goto out;
4200
4201         leaf = path->nodes[0];
4202         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4203         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
4204         btrfs_set_inode_generation(leaf, item, 1);
4205         btrfs_set_inode_size(leaf, item, 0);
4206         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4207         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4208                                           BTRFS_INODE_PREALLOC);
4209         btrfs_mark_buffer_dirty(leaf);
4210 out:
4211         btrfs_free_path(path);
4212         return ret;
4213 }
4214
4215 /*
4216  * helper to create inode for data relocation.
4217  * the inode is in data relocation tree and its link count is 0
4218  */
4219 static noinline_for_stack
4220 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4221                                  struct btrfs_block_group_cache *group)
4222 {
4223         struct inode *inode = NULL;
4224         struct btrfs_trans_handle *trans;
4225         struct btrfs_root *root;
4226         struct btrfs_key key;
4227         u64 objectid;
4228         int err = 0;
4229
4230         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4231         if (IS_ERR(root))
4232                 return ERR_CAST(root);
4233
4234         trans = btrfs_start_transaction(root, 6);
4235         if (IS_ERR(trans))
4236                 return ERR_CAST(trans);
4237
4238         err = btrfs_find_free_objectid(root, &objectid);
4239         if (err)
4240                 goto out;
4241
4242         err = __insert_orphan_inode(trans, root, objectid);
4243         BUG_ON(err);
4244
4245         key.objectid = objectid;
4246         key.type = BTRFS_INODE_ITEM_KEY;
4247         key.offset = 0;
4248         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4249         BUG_ON(IS_ERR(inode));
4250         BTRFS_I(inode)->index_cnt = group->key.objectid;
4251
4252         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4253 out:
4254         btrfs_end_transaction(trans);
4255         btrfs_btree_balance_dirty(fs_info);
4256         if (err) {
4257                 if (inode)
4258                         iput(inode);
4259                 inode = ERR_PTR(err);
4260         }
4261         return inode;
4262 }
4263
4264 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4265 {
4266         struct reloc_control *rc;
4267
4268         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4269         if (!rc)
4270                 return NULL;
4271
4272         INIT_LIST_HEAD(&rc->reloc_roots);
4273         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
4274         backref_cache_init(&rc->backref_cache);
4275         mapping_tree_init(&rc->reloc_root_tree);
4276         extent_io_tree_init(fs_info, &rc->processed_blocks,
4277                             IO_TREE_RELOC_BLOCKS, NULL);
4278         return rc;
4279 }
4280
4281 /*
4282  * Print the block group being relocated
4283  */
4284 static void describe_relocation(struct btrfs_fs_info *fs_info,
4285                                 struct btrfs_block_group_cache *block_group)
4286 {
4287         char buf[128] = {'\0'};
4288
4289         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
4290
4291         btrfs_info(fs_info,
4292                    "relocating block group %llu flags %s",
4293                    block_group->key.objectid, buf);
4294 }
4295
4296 /*
4297  * function to relocate all extents in a block group.
4298  */
4299 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4300 {
4301         struct btrfs_block_group_cache *bg;
4302         struct btrfs_root *extent_root = fs_info->extent_root;
4303         struct reloc_control *rc;
4304         struct inode *inode;
4305         struct btrfs_path *path;
4306         int ret;
4307         int rw = 0;
4308         int err = 0;
4309
4310         bg = btrfs_lookup_block_group(fs_info, group_start);
4311         if (!bg)
4312                 return -ENOENT;
4313
4314         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4315                 btrfs_put_block_group(bg);
4316                 return -ETXTBSY;
4317         }
4318
4319         rc = alloc_reloc_control(fs_info);
4320         if (!rc) {
4321                 btrfs_put_block_group(bg);
4322                 return -ENOMEM;
4323         }
4324
4325         rc->extent_root = extent_root;
4326         rc->block_group = bg;
4327
4328         ret = btrfs_inc_block_group_ro(rc->block_group);
4329         if (ret) {
4330                 err = ret;
4331                 goto out;
4332         }
4333         rw = 1;
4334
4335         path = btrfs_alloc_path();
4336         if (!path) {
4337                 err = -ENOMEM;
4338                 goto out;
4339         }
4340
4341         inode = lookup_free_space_inode(rc->block_group, path);
4342         btrfs_free_path(path);
4343
4344         if (!IS_ERR(inode))
4345                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4346         else
4347                 ret = PTR_ERR(inode);
4348
4349         if (ret && ret != -ENOENT) {
4350                 err = ret;
4351                 goto out;
4352         }
4353
4354         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4355         if (IS_ERR(rc->data_inode)) {
4356                 err = PTR_ERR(rc->data_inode);
4357                 rc->data_inode = NULL;
4358                 goto out;
4359         }
4360
4361         describe_relocation(fs_info, rc->block_group);
4362
4363         btrfs_wait_block_group_reservations(rc->block_group);
4364         btrfs_wait_nocow_writers(rc->block_group);
4365         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4366                                  rc->block_group->key.objectid,
4367                                  rc->block_group->key.offset);
4368
4369         while (1) {
4370                 mutex_lock(&fs_info->cleaner_mutex);
4371                 ret = relocate_block_group(rc);
4372                 mutex_unlock(&fs_info->cleaner_mutex);
4373                 if (ret < 0)
4374                         err = ret;
4375
4376                 /*
4377                  * We may have gotten ENOSPC after we already dirtied some
4378                  * extents.  If writeout happens while we're relocating a
4379                  * different block group we could end up hitting the
4380                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4381                  * btrfs_reloc_cow_block.  Make sure we write everything out
4382                  * properly so we don't trip over this problem, and then break
4383                  * out of the loop if we hit an error.
4384                  */
4385                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4386                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4387                                                        (u64)-1);
4388                         if (ret)
4389                                 err = ret;
4390                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4391                                                  0, -1);
4392                         rc->stage = UPDATE_DATA_PTRS;
4393                 }
4394
4395                 if (err < 0)
4396                         goto out;
4397
4398                 if (rc->extents_found == 0)
4399                         break;
4400
4401                 btrfs_info(fs_info, "found %llu extents", rc->extents_found);
4402
4403         }
4404
4405         WARN_ON(rc->block_group->pinned > 0);
4406         WARN_ON(rc->block_group->reserved > 0);
4407         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4408 out:
4409         if (err && rw)
4410                 btrfs_dec_block_group_ro(rc->block_group);
4411         iput(rc->data_inode);
4412         btrfs_put_block_group(rc->block_group);
4413         kfree(rc);
4414         return err;
4415 }
4416
4417 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4418 {
4419         struct btrfs_fs_info *fs_info = root->fs_info;
4420         struct btrfs_trans_handle *trans;
4421         int ret, err;
4422
4423         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4424         if (IS_ERR(trans))
4425                 return PTR_ERR(trans);
4426
4427         memset(&root->root_item.drop_progress, 0,
4428                 sizeof(root->root_item.drop_progress));
4429         root->root_item.drop_level = 0;
4430         btrfs_set_root_refs(&root->root_item, 0);
4431         ret = btrfs_update_root(trans, fs_info->tree_root,
4432                                 &root->root_key, &root->root_item);
4433
4434         err = btrfs_end_transaction(trans);
4435         if (err)
4436                 return err;
4437         return ret;
4438 }
4439
4440 /*
4441  * recover relocation interrupted by system crash.
4442  *
4443  * this function resumes merging reloc trees with corresponding fs trees.
4444  * this is important for keeping the sharing of tree blocks
4445  */
4446 int btrfs_recover_relocation(struct btrfs_root *root)
4447 {
4448         struct btrfs_fs_info *fs_info = root->fs_info;
4449         LIST_HEAD(reloc_roots);
4450         struct btrfs_key key;
4451         struct btrfs_root *fs_root;
4452         struct btrfs_root *reloc_root;
4453         struct btrfs_path *path;
4454         struct extent_buffer *leaf;
4455         struct reloc_control *rc = NULL;
4456         struct btrfs_trans_handle *trans;
4457         int ret;
4458         int err = 0;
4459
4460         path = btrfs_alloc_path();
4461         if (!path)
4462                 return -ENOMEM;
4463         path->reada = READA_BACK;
4464
4465         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4466         key.type = BTRFS_ROOT_ITEM_KEY;
4467         key.offset = (u64)-1;
4468
4469         while (1) {
4470                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4471                                         path, 0, 0);
4472                 if (ret < 0) {
4473                         err = ret;
4474                         goto out;
4475                 }
4476                 if (ret > 0) {
4477                         if (path->slots[0] == 0)
4478                                 break;
4479                         path->slots[0]--;
4480                 }
4481                 leaf = path->nodes[0];
4482                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4483                 btrfs_release_path(path);
4484
4485                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4486                     key.type != BTRFS_ROOT_ITEM_KEY)
4487                         break;
4488
4489                 reloc_root = btrfs_read_fs_root(root, &key);
4490                 if (IS_ERR(reloc_root)) {
4491                         err = PTR_ERR(reloc_root);
4492                         goto out;
4493                 }
4494
4495                 list_add(&reloc_root->root_list, &reloc_roots);
4496
4497                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4498                         fs_root = read_fs_root(fs_info,
4499                                                reloc_root->root_key.offset);
4500                         if (IS_ERR(fs_root)) {
4501                                 ret = PTR_ERR(fs_root);
4502                                 if (ret != -ENOENT) {
4503                                         err = ret;
4504                                         goto out;
4505                                 }
4506                                 ret = mark_garbage_root(reloc_root);
4507                                 if (ret < 0) {
4508                                         err = ret;
4509                                         goto out;
4510                                 }
4511                         }
4512                 }
4513
4514                 if (key.offset == 0)
4515                         break;
4516
4517                 key.offset--;
4518         }
4519         btrfs_release_path(path);
4520
4521         if (list_empty(&reloc_roots))
4522                 goto out;
4523
4524         rc = alloc_reloc_control(fs_info);
4525         if (!rc) {
4526                 err = -ENOMEM;
4527                 goto out;
4528         }
4529
4530         rc->extent_root = fs_info->extent_root;
4531
4532         set_reloc_control(rc);
4533
4534         trans = btrfs_join_transaction(rc->extent_root);
4535         if (IS_ERR(trans)) {
4536                 unset_reloc_control(rc);
4537                 err = PTR_ERR(trans);
4538                 goto out_free;
4539         }
4540
4541         rc->merge_reloc_tree = 1;
4542
4543         while (!list_empty(&reloc_roots)) {
4544                 reloc_root = list_entry(reloc_roots.next,
4545                                         struct btrfs_root, root_list);
4546                 list_del(&reloc_root->root_list);
4547
4548                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4549                         list_add_tail(&reloc_root->root_list,
4550                                       &rc->reloc_roots);
4551                         continue;
4552                 }
4553
4554                 fs_root = read_fs_root(fs_info, reloc_root->root_key.offset);
4555                 if (IS_ERR(fs_root)) {
4556                         err = PTR_ERR(fs_root);
4557                         goto out_free;
4558                 }
4559
4560                 err = __add_reloc_root(reloc_root);
4561                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4562                 fs_root->reloc_root = reloc_root;
4563         }
4564
4565         err = btrfs_commit_transaction(trans);
4566         if (err)
4567                 goto out_free;
4568
4569         merge_reloc_roots(rc);
4570
4571         unset_reloc_control(rc);
4572
4573         trans = btrfs_join_transaction(rc->extent_root);
4574         if (IS_ERR(trans)) {
4575                 err = PTR_ERR(trans);
4576                 goto out_free;
4577         }
4578         err = btrfs_commit_transaction(trans);
4579
4580         ret = clean_dirty_subvols(rc);
4581         if (ret < 0 && !err)
4582                 err = ret;
4583 out_free:
4584         kfree(rc);
4585 out:
4586         if (!list_empty(&reloc_roots))
4587                 free_reloc_roots(&reloc_roots);
4588
4589         btrfs_free_path(path);
4590
4591         if (err == 0) {
4592                 /* cleanup orphan inode in data relocation tree */
4593                 fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4594                 if (IS_ERR(fs_root))
4595                         err = PTR_ERR(fs_root);
4596                 else
4597                         err = btrfs_orphan_cleanup(fs_root);
4598         }
4599         return err;
4600 }
4601
4602 /*
4603  * helper to add ordered checksum for data relocation.
4604  *
4605  * cloning checksum properly handles the nodatasum extents.
4606  * it also saves CPU time to re-calculate the checksum.
4607  */
4608 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4609 {
4610         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4611         struct btrfs_ordered_sum *sums;
4612         struct btrfs_ordered_extent *ordered;
4613         int ret;
4614         u64 disk_bytenr;
4615         u64 new_bytenr;
4616         LIST_HEAD(list);
4617
4618         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4619         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4620
4621         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4622         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4623                                        disk_bytenr + len - 1, &list, 0);
4624         if (ret)
4625                 goto out;
4626
4627         while (!list_empty(&list)) {
4628                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4629                 list_del_init(&sums->list);
4630
4631                 /*
4632                  * We need to offset the new_bytenr based on where the csum is.
4633                  * We need to do this because we will read in entire prealloc
4634                  * extents but we may have written to say the middle of the
4635                  * prealloc extent, so we need to make sure the csum goes with
4636                  * the right disk offset.
4637                  *
4638                  * We can do this because the data reloc inode refers strictly
4639                  * to the on disk bytes, so we don't have to worry about
4640                  * disk_len vs real len like with real inodes since it's all
4641                  * disk length.
4642                  */
4643                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4644                 sums->bytenr = new_bytenr;
4645
4646                 btrfs_add_ordered_sum(ordered, sums);
4647         }
4648 out:
4649         btrfs_put_ordered_extent(ordered);
4650         return ret;
4651 }
4652
4653 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4654                           struct btrfs_root *root, struct extent_buffer *buf,
4655                           struct extent_buffer *cow)
4656 {
4657         struct btrfs_fs_info *fs_info = root->fs_info;
4658         struct reloc_control *rc;
4659         struct backref_node *node;
4660         int first_cow = 0;
4661         int level;
4662         int ret = 0;
4663
4664         rc = fs_info->reloc_ctl;
4665         if (!rc)
4666                 return 0;
4667
4668         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4669                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4670
4671         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4672                 if (buf == root->node)
4673                         __update_reloc_root(root, cow->start);
4674         }
4675
4676         level = btrfs_header_level(buf);
4677         if (btrfs_header_generation(buf) <=
4678             btrfs_root_last_snapshot(&root->root_item))
4679                 first_cow = 1;
4680
4681         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4682             rc->create_reloc_tree) {
4683                 WARN_ON(!first_cow && level == 0);
4684
4685                 node = rc->backref_cache.path[level];
4686                 BUG_ON(node->bytenr != buf->start &&
4687                        node->new_bytenr != buf->start);
4688
4689                 drop_node_buffer(node);
4690                 extent_buffer_get(cow);
4691                 node->eb = cow;
4692                 node->new_bytenr = cow->start;
4693
4694                 if (!node->pending) {
4695                         list_move_tail(&node->list,
4696                                        &rc->backref_cache.pending[level]);
4697                         node->pending = 1;
4698                 }
4699
4700                 if (first_cow)
4701                         __mark_block_processed(rc, node);
4702
4703                 if (first_cow && level > 0)
4704                         rc->nodes_relocated += buf->len;
4705         }
4706
4707         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4708                 ret = replace_file_extents(trans, rc, root, cow);
4709         return ret;
4710 }
4711
4712 /*
4713  * called before creating snapshot. it calculates metadata reservation
4714  * required for relocating tree blocks in the snapshot
4715  */
4716 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4717                               u64 *bytes_to_reserve)
4718 {
4719         struct btrfs_root *root = pending->root;
4720         struct reloc_control *rc = root->fs_info->reloc_ctl;
4721
4722         if (!root->reloc_root || !rc)
4723                 return;
4724
4725         if (!rc->merge_reloc_tree)
4726                 return;
4727
4728         root = root->reloc_root;
4729         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4730         /*
4731          * relocation is in the stage of merging trees. the space
4732          * used by merging a reloc tree is twice the size of
4733          * relocated tree nodes in the worst case. half for cowing
4734          * the reloc tree, half for cowing the fs tree. the space
4735          * used by cowing the reloc tree will be freed after the
4736          * tree is dropped. if we create snapshot, cowing the fs
4737          * tree may use more space than it frees. so we need
4738          * reserve extra space.
4739          */
4740         *bytes_to_reserve += rc->nodes_relocated;
4741 }
4742
4743 /*
4744  * called after snapshot is created. migrate block reservation
4745  * and create reloc root for the newly created snapshot
4746  */
4747 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4748                                struct btrfs_pending_snapshot *pending)
4749 {
4750         struct btrfs_root *root = pending->root;
4751         struct btrfs_root *reloc_root;
4752         struct btrfs_root *new_root;
4753         struct reloc_control *rc = root->fs_info->reloc_ctl;
4754         int ret;
4755
4756         if (!root->reloc_root || !rc)
4757                 return 0;
4758
4759         rc = root->fs_info->reloc_ctl;
4760         rc->merging_rsv_size += rc->nodes_relocated;
4761
4762         if (rc->merge_reloc_tree) {
4763                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4764                                               rc->block_rsv,
4765                                               rc->nodes_relocated, true);
4766                 if (ret)
4767                         return ret;
4768         }
4769
4770         new_root = pending->snap;
4771         reloc_root = create_reloc_root(trans, root->reloc_root,
4772                                        new_root->root_key.objectid);
4773         if (IS_ERR(reloc_root))
4774                 return PTR_ERR(reloc_root);
4775
4776         ret = __add_reloc_root(reloc_root);
4777         BUG_ON(ret < 0);
4778         new_root->reloc_root = reloc_root;
4779
4780         if (rc->create_reloc_tree)
4781                 ret = clone_backref_node(trans, rc, root, reloc_root);
4782         return ret;
4783 }