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 <linux/error-injection.h>
  13 #include "ctree.h"
  14 #include "disk-io.h"
  15 #include "transaction.h"
  16 #include "volumes.h"
  17 #include "locking.h"
  18 #include "btrfs_inode.h"
  19 #include "async-thread.h"
  20 #include "free-space-cache.h"
  21 #include "inode-map.h"
  22 #include "qgroup.h"
  23 #include "print-tree.h"
  24 #include "delalloc-space.h"
  25 #include "block-group.h"
  26 #include "backref.h"
  27 #include "misc.h"
  28
  29 /*
  30  * Relocation overview
  31  *
  32  * [What does relocation do]
  33  *
  34  * The objective of relocation is to relocate all extents of the target block
  35  * group to other block groups.
  36  * This is utilized by resize (shrink only), profile converting, compacting
  37  * space, or balance routine to spread chunks over devices.
  38  *
  39  *              Before          |               After
  40  * ------------------------------------------------------------------
  41  *  BG A: 10 data extents       | BG A: deleted
  42  *  BG B:  2 data extents       | BG B: 10 data extents (2 old + 8 relocated)
  43  *  BG C:  1 extents            | BG C:  3 data extents (1 old + 2 relocated)
  44  *
  45  * [How does relocation work]
  46  *
  47  * 1.   Mark the target block group read-only
  48  *      New extents won't be allocated from the target block group.
  49  *
  50  * 2.1  Record each extent in the target block group
  51  *      To build a proper map of extents to be relocated.
  52  *
  53  * 2.2  Build data reloc tree and reloc trees
  54  *      Data reloc tree will contain an inode, recording all newly relocated
  55  *      data extents.
  56  *      There will be only one data reloc tree for one data block group.
  57  *
  58  *      Reloc tree will be a special snapshot of its source tree, containing
  59  *      relocated tree blocks.
  60  *      Each tree referring to a tree block in target block group will get its
  61  *      reloc tree built.
  62  *
  63  * 2.3  Swap source tree with its corresponding reloc tree
  64  *      Each involved tree only refers to new extents after swap.
  65  *
  66  * 3.   Cleanup reloc trees and data reloc tree.
  67  *      As old extents in the target block group are still referenced by reloc
  68  *      trees, we need to clean them up before really freeing the target block
  69  *      group.
  70  *
  71  * The main complexity is in steps 2.2 and 2.3.
  72  *
  73  * The entry point of relocation is relocate_block_group() function.
  74  */
  75
  76 #define RELOCATION_RESERVED_NODES       256
  77 /*
  78  * map address of tree root to tree
  79  */
  80 struct mapping_node {
  81         struct {
  82                 struct rb_node rb_node;
  83                 u64 bytenr;
  84         }; /* Use rb_simle_node for search/insert */
  85         void *data;
  86 };
  87
  88 struct mapping_tree {
  89         struct rb_root rb_root;
  90         spinlock_t lock;
  91 };
  92
  93 /*
  94  * present a tree block to process
  95  */
  96 struct tree_block {
  97         struct {
  98                 struct rb_node rb_node;
  99                 u64 bytenr;
 100         }; /* Use rb_simple_node for search/insert */
 101         struct btrfs_key key;
 102         unsigned int level:8;
 103         unsigned int key_ready:1;
 104 };
 105
 106 #define MAX_EXTENTS 128
 107
 108 struct file_extent_cluster {
 109         u64 start;
 110         u64 end;
 111         u64 boundary[MAX_EXTENTS];
 112         unsigned int nr;
 113 };
 114
 115 struct reloc_control {
 116         /* block group to relocate */
 117         struct btrfs_block_group *block_group;
 118         /* extent tree */
 119         struct btrfs_root *extent_root;
 120         /* inode for moving data */
 121         struct inode *data_inode;
 122
 123         struct btrfs_block_rsv *block_rsv;
 124
 125         struct btrfs_backref_cache backref_cache;
 126
 127         struct file_extent_cluster cluster;
 128         /* tree blocks have been processed */
 129         struct extent_io_tree processed_blocks;
 130         /* map start of tree root to corresponding reloc tree */
 131         struct mapping_tree reloc_root_tree;
 132         /* list of reloc trees */
 133         struct list_head reloc_roots;
 134         /* list of subvolume trees that get relocated */
 135         struct list_head dirty_subvol_roots;
 136         /* size of metadata reservation for merging reloc trees */
 137         u64 merging_rsv_size;
 138         /* size of relocated tree nodes */
 139         u64 nodes_relocated;
 140         /* reserved size for block group relocation*/
 141         u64 reserved_bytes;
 142
 143         u64 search_start;
 144         u64 extents_found;
 145
 146         unsigned int stage:8;
 147         unsigned int create_reloc_tree:1;
 148         unsigned int merge_reloc_tree:1;
 149         unsigned int found_file_extent:1;
 150 };
 151
 152 /* stages of data relocation */
 153 #define MOVE_DATA_EXTENTS       0
 154 #define UPDATE_DATA_PTRS        1
 155
 156 static void mark_block_processed(struct reloc_control *rc,
 157                                  struct btrfs_backref_node *node)
 158 {
 159         u32 blocksize;
 160
 161         if (node->level == 0 ||
 162             in_range(node->bytenr, rc->block_group->start,
 163                      rc->block_group->length)) {
 164                 blocksize = rc->extent_root->fs_info->nodesize;
 165                 set_extent_bits(&rc->processed_blocks, node->bytenr,
 166                                 node->bytenr + blocksize - 1, EXTENT_DIRTY);
 167         }
 168         node->processed = 1;
 169 }
 170
 171
 172 static void mapping_tree_init(struct mapping_tree *tree)
 173 {
 174         tree->rb_root = RB_ROOT;
 175         spin_lock_init(&tree->lock);
 176 }
 177
 178 /*
 179  * walk up backref nodes until reach node presents tree root
 180  */
 181 static struct btrfs_backref_node *walk_up_backref(
 182                 struct btrfs_backref_node *node,
 183                 struct btrfs_backref_edge *edges[], int *index)
 184 {
 185         struct btrfs_backref_edge *edge;
 186         int idx = *index;
 187
 188         while (!list_empty(&node->upper)) {
 189                 edge = list_entry(node->upper.next,
 190                                   struct btrfs_backref_edge, list[LOWER]);
 191                 edges[idx++] = edge;
 192                 node = edge->node[UPPER];
 193         }
 194         BUG_ON(node->detached);
 195         *index = idx;
 196         return node;
 197 }
 198
 199 /*
 200  * walk down backref nodes to find start of next reference path
 201  */
 202 static struct btrfs_backref_node *walk_down_backref(
 203                 struct btrfs_backref_edge *edges[], int *index)
 204 {
 205         struct btrfs_backref_edge *edge;
 206         struct btrfs_backref_node *lower;
 207         int idx = *index;
 208
 209         while (idx > 0) {
 210                 edge = edges[idx - 1];
 211                 lower = edge->node[LOWER];
 212                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 213                         idx--;
 214                         continue;
 215                 }
 216                 edge = list_entry(edge->list[LOWER].next,
 217                                   struct btrfs_backref_edge, list[LOWER]);
 218                 edges[idx - 1] = edge;
 219                 *index = idx;
 220                 return edge->node[UPPER];
 221         }
 222         *index = 0;
 223         return NULL;
 224 }
 225
 226 static void update_backref_node(struct btrfs_backref_cache *cache,
 227                                 struct btrfs_backref_node *node, u64 bytenr)
 228 {
 229         struct rb_node *rb_node;
 230         rb_erase(&node->rb_node, &cache->rb_root);
 231         node->bytenr = bytenr;
 232         rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 233         if (rb_node)
 234                 btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
 235 }
 236
 237 /*
 238  * update backref cache after a transaction commit
 239  */
 240 static int update_backref_cache(struct btrfs_trans_handle *trans,
 241                                 struct btrfs_backref_cache *cache)
 242 {
 243         struct btrfs_backref_node *node;
 244         int level = 0;
 245
 246         if (cache->last_trans == 0) {
 247                 cache->last_trans = trans->transid;
 248                 return 0;
 249         }
 250
 251         if (cache->last_trans == trans->transid)
 252                 return 0;
 253
 254         /*
 255          * detached nodes are used to avoid unnecessary backref
 256          * lookup. transaction commit changes the extent tree.
 257          * so the detached nodes are no longer useful.
 258          */
 259         while (!list_empty(&cache->detached)) {
 260                 node = list_entry(cache->detached.next,
 261                                   struct btrfs_backref_node, list);
 262                 btrfs_backref_cleanup_node(cache, node);
 263         }
 264
 265         while (!list_empty(&cache->changed)) {
 266                 node = list_entry(cache->changed.next,
 267                                   struct btrfs_backref_node, list);
 268                 list_del_init(&node->list);
 269                 BUG_ON(node->pending);
 270                 update_backref_node(cache, node, node->new_bytenr);
 271         }
 272
 273         /*
 274          * some nodes can be left in the pending list if there were
 275          * errors during processing the pending nodes.
 276          */
 277         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 278                 list_for_each_entry(node, &cache->pending[level], list) {
 279                         BUG_ON(!node->pending);
 280                         if (node->bytenr == node->new_bytenr)
 281                                 continue;
 282                         update_backref_node(cache, node, node->new_bytenr);
 283                 }
 284         }
 285
 286         cache->last_trans = 0;
 287         return 1;
 288 }
 289
 290 static bool reloc_root_is_dead(struct btrfs_root *root)
 291 {
 292         /*
 293          * Pair with set_bit/clear_bit in clean_dirty_subvols and
 294          * btrfs_update_reloc_root. We need to see the updated bit before
 295          * trying to access reloc_root
 296          */
 297         smp_rmb();
 298         if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
 299                 return true;
 300         return false;
 301 }
 302
 303 /*
 304  * Check if this subvolume tree has valid reloc tree.
 305  *
 306  * Reloc tree after swap is considered dead, thus not considered as valid.
 307  * This is enough for most callers, as they don't distinguish dead reloc root
 308  * from no reloc root.  But btrfs_should_ignore_reloc_root() below is a
 309  * special case.
 310  */
 311 static bool have_reloc_root(struct btrfs_root *root)
 312 {
 313         if (reloc_root_is_dead(root))
 314                 return false;
 315         if (!root->reloc_root)
 316                 return false;
 317         return true;
 318 }
 319
 320 int btrfs_should_ignore_reloc_root(struct btrfs_root *root)
 321 {
 322         struct btrfs_root *reloc_root;
 323
 324         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
 325                 return 0;
 326
 327         /* This root has been merged with its reloc tree, we can ignore it */
 328         if (reloc_root_is_dead(root))
 329                 return 1;
 330
 331         reloc_root = root->reloc_root;
 332         if (!reloc_root)
 333                 return 0;
 334
 335         if (btrfs_header_generation(reloc_root->commit_root) ==
 336             root->fs_info->running_transaction->transid)
 337                 return 0;
 338         /*
 339          * if there is reloc tree and it was created in previous
 340          * transaction backref lookup can find the reloc tree,
 341          * so backref node for the fs tree root is useless for
 342          * relocation.
 343          */
 344         return 1;
 345 }
 346
 347 /*
 348  * find reloc tree by address of tree root
 349  */
 350 struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
 351 {
 352         struct reloc_control *rc = fs_info->reloc_ctl;
 353         struct rb_node *rb_node;
 354         struct mapping_node *node;
 355         struct btrfs_root *root = NULL;
 356
 357         ASSERT(rc);
 358         spin_lock(&rc->reloc_root_tree.lock);
 359         rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
 360         if (rb_node) {
 361                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 362                 root = (struct btrfs_root *)node->data;
 363         }
 364         spin_unlock(&rc->reloc_root_tree.lock);
 365         return btrfs_grab_root(root);
 366 }
 367
 368 /*
 369  * For useless nodes, do two major clean ups:
 370  *
 371  * - Cleanup the children edges and nodes
 372  *   If child node is also orphan (no parent) during cleanup, then the child
 373  *   node will also be cleaned up.
 374  *
 375  * - Freeing up leaves (level 0), keeps nodes detached
 376  *   For nodes, the node is still cached as "detached"
 377  *
 378  * Return false if @node is not in the @useless_nodes list.
 379  * Return true if @node is in the @useless_nodes list.
 380  */
 381 static bool handle_useless_nodes(struct reloc_control *rc,
 382                                  struct btrfs_backref_node *node)
 383 {
 384         struct btrfs_backref_cache *cache = &rc->backref_cache;
 385         struct list_head *useless_node = &cache->useless_node;
 386         bool ret = false;
 387
 388         while (!list_empty(useless_node)) {
 389                 struct btrfs_backref_node *cur;
 390
 391                 cur = list_first_entry(useless_node, struct btrfs_backref_node,
 392                                  list);
 393                 list_del_init(&cur->list);
 394
 395                 /* Only tree root nodes can be added to @useless_nodes */
 396                 ASSERT(list_empty(&cur->upper));
 397
 398                 if (cur == node)
 399                         ret = true;
 400
 401                 /* The node is the lowest node */
 402                 if (cur->lowest) {
 403                         list_del_init(&cur->lower);
 404                         cur->lowest = 0;
 405                 }
 406
 407                 /* Cleanup the lower edges */
 408                 while (!list_empty(&cur->lower)) {
 409                         struct btrfs_backref_edge *edge;
 410                         struct btrfs_backref_node *lower;
 411
 412                         edge = list_entry(cur->lower.next,
 413                                         struct btrfs_backref_edge, list[UPPER]);
 414                         list_del(&edge->list[UPPER]);
 415                         list_del(&edge->list[LOWER]);
 416                         lower = edge->node[LOWER];
 417                         btrfs_backref_free_edge(cache, edge);
 418
 419                         /* Child node is also orphan, queue for cleanup */
 420                         if (list_empty(&lower->upper))
 421                                 list_add(&lower->list, useless_node);
 422                 }
 423                 /* Mark this block processed for relocation */
 424                 mark_block_processed(rc, cur);
 425
 426                 /*
 427                  * Backref nodes for tree leaves are deleted from the cache.
 428                  * Backref nodes for upper level tree blocks are left in the
 429                  * cache to avoid unnecessary backref lookup.
 430                  */
 431                 if (cur->level > 0) {
 432                         list_add(&cur->list, &cache->detached);
 433                         cur->detached = 1;
 434                 } else {
 435                         rb_erase(&cur->rb_node, &cache->rb_root);
 436                         btrfs_backref_free_node(cache, cur);
 437                 }
 438         }
 439         return ret;
 440 }
 441
 442 /*
 443  * Build backref tree for a given tree block. Root of the backref tree
 444  * corresponds the tree block, leaves of the backref tree correspond roots of
 445  * b-trees that reference the tree block.
 446  *
 447  * The basic idea of this function is check backrefs of a given block to find
 448  * upper level blocks that reference the block, and then check backrefs of
 449  * these upper level blocks recursively. The recursion stops when tree root is
 450  * reached or backrefs for the block is cached.
 451  *
 452  * NOTE: if we find that backrefs for a block are cached, we know backrefs for
 453  * all upper level blocks that directly/indirectly reference the block are also
 454  * cached.
 455  */
 456 static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
 457                         struct reloc_control *rc, struct btrfs_key *node_key,
 458                         int level, u64 bytenr)
 459 {
 460         struct btrfs_backref_iter *iter;
 461         struct btrfs_backref_cache *cache = &rc->backref_cache;
 462         /* For searching parent of TREE_BLOCK_REF */
 463         struct btrfs_path *path;
 464         struct btrfs_backref_node *cur;
 465         struct btrfs_backref_node *node = NULL;
 466         struct btrfs_backref_edge *edge;
 467         int ret;
 468         int err = 0;
 469
 470         iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS);
 471         if (!iter)
 472                 return ERR_PTR(-ENOMEM);
 473         path = btrfs_alloc_path();
 474         if (!path) {
 475                 err = -ENOMEM;
 476                 goto out;
 477         }
 478
 479         node = btrfs_backref_alloc_node(cache, bytenr, level);
 480         if (!node) {
 481                 err = -ENOMEM;
 482                 goto out;
 483         }
 484
 485         node->lowest = 1;
 486         cur = node;
 487
 488         /* Breadth-first search to build backref cache */
 489         do {
 490                 ret = btrfs_backref_add_tree_node(cache, path, iter, node_key,
 491                                                   cur);
 492                 if (ret < 0) {
 493                         err = ret;
 494                         goto out;
 495                 }
 496                 edge = list_first_entry_or_null(&cache->pending_edge,
 497                                 struct btrfs_backref_edge, list[UPPER]);
 498                 /*
 499                  * The pending list isn't empty, take the first block to
 500                  * process
 501                  */
 502                 if (edge) {
 503                         list_del_init(&edge->list[UPPER]);
 504                         cur = edge->node[UPPER];
 505                 }
 506         } while (edge);
 507
 508         /* Finish the upper linkage of newly added edges/nodes */
 509         ret = btrfs_backref_finish_upper_links(cache, node);
 510         if (ret < 0) {
 511                 err = ret;
 512                 goto out;
 513         }
 514
 515         if (handle_useless_nodes(rc, node))
 516                 node = NULL;
 517 out:
 518         btrfs_backref_iter_free(iter);
 519         btrfs_free_path(path);
 520         if (err) {
 521                 btrfs_backref_error_cleanup(cache, node);
 522                 return ERR_PTR(err);
 523         }
 524         ASSERT(!node || !node->detached);
 525         ASSERT(list_empty(&cache->useless_node) &&
 526                list_empty(&cache->pending_edge));
 527         return node;
 528 }
 529
 530 /*
 531  * helper to add backref node for the newly created snapshot.
 532  * the backref node is created by cloning backref node that
 533  * corresponds to root of source tree
 534  */
 535 static int clone_backref_node(struct btrfs_trans_handle *trans,
 536                               struct reloc_control *rc,
 537                               struct btrfs_root *src,
 538                               struct btrfs_root *dest)
 539 {
 540         struct btrfs_root *reloc_root = src->reloc_root;
 541         struct btrfs_backref_cache *cache = &rc->backref_cache;
 542         struct btrfs_backref_node *node = NULL;
 543         struct btrfs_backref_node *new_node;
 544         struct btrfs_backref_edge *edge;
 545         struct btrfs_backref_edge *new_edge;
 546         struct rb_node *rb_node;
 547
 548         if (cache->last_trans > 0)
 549                 update_backref_cache(trans, cache);
 550
 551         rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
 552         if (rb_node) {
 553                 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
 554                 if (node->detached)
 555                         node = NULL;
 556                 else
 557                         BUG_ON(node->new_bytenr != reloc_root->node->start);
 558         }
 559
 560         if (!node) {
 561                 rb_node = rb_simple_search(&cache->rb_root,
 562                                            reloc_root->commit_root->start);
 563                 if (rb_node) {
 564                         node = rb_entry(rb_node, struct btrfs_backref_node,
 565                                         rb_node);
 566                         BUG_ON(node->detached);
 567                 }
 568         }
 569
 570         if (!node)
 571                 return 0;
 572
 573         new_node = btrfs_backref_alloc_node(cache, dest->node->start,
 574                                             node->level);
 575         if (!new_node)
 576                 return -ENOMEM;
 577
 578         new_node->lowest = node->lowest;
 579         new_node->checked = 1;
 580         new_node->root = btrfs_grab_root(dest);
 581         ASSERT(new_node->root);
 582
 583         if (!node->lowest) {
 584                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
 585                         new_edge = btrfs_backref_alloc_edge(cache);
 586                         if (!new_edge)
 587                                 goto fail;
 588
 589                         btrfs_backref_link_edge(new_edge, edge->node[LOWER],
 590                                                 new_node, LINK_UPPER);
 591                 }
 592         } else {
 593                 list_add_tail(&new_node->lower, &cache->leaves);
 594         }
 595
 596         rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
 597                                    &new_node->rb_node);
 598         if (rb_node)
 599                 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
 600
 601         if (!new_node->lowest) {
 602                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
 603                         list_add_tail(&new_edge->list[LOWER],
 604                                       &new_edge->node[LOWER]->upper);
 605                 }
 606         }
 607         return 0;
 608 fail:
 609         while (!list_empty(&new_node->lower)) {
 610                 new_edge = list_entry(new_node->lower.next,
 611                                       struct btrfs_backref_edge, list[UPPER]);
 612                 list_del(&new_edge->list[UPPER]);
 613                 btrfs_backref_free_edge(cache, new_edge);
 614         }
 615         btrfs_backref_free_node(cache, new_node);
 616         return -ENOMEM;
 617 }
 618
 619 /*
 620  * helper to add 'address of tree root -> reloc tree' mapping
 621  */
 622 static int __must_check __add_reloc_root(struct btrfs_root *root)
 623 {
 624         struct btrfs_fs_info *fs_info = root->fs_info;
 625         struct rb_node *rb_node;
 626         struct mapping_node *node;
 627         struct reloc_control *rc = fs_info->reloc_ctl;
 628
 629         node = kmalloc(sizeof(*node), GFP_NOFS);
 630         if (!node)
 631                 return -ENOMEM;
 632
 633         node->bytenr = root->commit_root->start;
 634         node->data = root;
 635
 636         spin_lock(&rc->reloc_root_tree.lock);
 637         rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
 638                                    node->bytenr, &node->rb_node);
 639         spin_unlock(&rc->reloc_root_tree.lock);
 640         if (rb_node) {
 641                 btrfs_panic(fs_info, -EEXIST,
 642                             "Duplicate root found for start=%llu while inserting into relocation tree",
 643                             node->bytenr);
 644         }
 645
 646         list_add_tail(&root->root_list, &rc->reloc_roots);
 647         return 0;
 648 }
 649
 650 /*
 651  * helper to delete the 'address of tree root -> reloc tree'
 652  * mapping
 653  */
 654 static void __del_reloc_root(struct btrfs_root *root)
 655 {
 656         struct btrfs_fs_info *fs_info = root->fs_info;
 657         struct rb_node *rb_node;
 658         struct mapping_node *node = NULL;
 659         struct reloc_control *rc = fs_info->reloc_ctl;
 660         bool put_ref = false;
 661
 662         if (rc && root->node) {
 663                 spin_lock(&rc->reloc_root_tree.lock);
 664                 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
 665                                            root->commit_root->start);
 666                 if (rb_node) {
 667                         node = rb_entry(rb_node, struct mapping_node, rb_node);
 668                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
 669                         RB_CLEAR_NODE(&node->rb_node);
 670                 }
 671                 spin_unlock(&rc->reloc_root_tree.lock);
 672                 if (!node)
 673                         return;
 674                 BUG_ON((struct btrfs_root *)node->data != root);
 675         }
 676
 677         /*
 678          * We only put the reloc root here if it's on the list.  There's a lot
 679          * of places where the pattern is to splice the rc->reloc_roots, process
 680          * the reloc roots, and then add the reloc root back onto
 681          * rc->reloc_roots.  If we call __del_reloc_root while it's off of the
 682          * list we don't want the reference being dropped, because the guy
 683          * messing with the list is in charge of the reference.
 684          */
 685         spin_lock(&fs_info->trans_lock);
 686         if (!list_empty(&root->root_list)) {
 687                 put_ref = true;
 688                 list_del_init(&root->root_list);
 689         }
 690         spin_unlock(&fs_info->trans_lock);
 691         if (put_ref)
 692                 btrfs_put_root(root);
 693         kfree(node);
 694 }
 695
 696 /*
 697  * helper to update the 'address of tree root -> reloc tree'
 698  * mapping
 699  */
 700 static int __update_reloc_root(struct btrfs_root *root)
 701 {
 702         struct btrfs_fs_info *fs_info = root->fs_info;
 703         struct rb_node *rb_node;
 704         struct mapping_node *node = NULL;
 705         struct reloc_control *rc = fs_info->reloc_ctl;
 706
 707         spin_lock(&rc->reloc_root_tree.lock);
 708         rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
 709                                    root->commit_root->start);
 710         if (rb_node) {
 711                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 712                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
 713         }
 714         spin_unlock(&rc->reloc_root_tree.lock);
 715
 716         if (!node)
 717                 return 0;
 718         BUG_ON((struct btrfs_root *)node->data != root);
 719
 720         spin_lock(&rc->reloc_root_tree.lock);
 721         node->bytenr = root->node->start;
 722         rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
 723                                    node->bytenr, &node->rb_node);
 724         spin_unlock(&rc->reloc_root_tree.lock);
 725         if (rb_node)
 726                 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
 727         return 0;
 728 }
 729
 730 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
 731                                         struct btrfs_root *root, u64 objectid)
 732 {
 733         struct btrfs_fs_info *fs_info = root->fs_info;
 734         struct btrfs_root *reloc_root;
 735         struct extent_buffer *eb;
 736         struct btrfs_root_item *root_item;
 737         struct btrfs_key root_key;
 738         int ret;
 739
 740         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
 741         BUG_ON(!root_item);
 742
 743         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
 744         root_key.type = BTRFS_ROOT_ITEM_KEY;
 745         root_key.offset = objectid;
 746
 747         if (root->root_key.objectid == objectid) {
 748                 u64 commit_root_gen;
 749
 750                 /* called by btrfs_init_reloc_root */
 751                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
 752                                       BTRFS_TREE_RELOC_OBJECTID);
 753                 BUG_ON(ret);
 754                 /*
 755                  * Set the last_snapshot field to the generation of the commit
 756                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
 757                  * correctly (returns true) when the relocation root is created
 758                  * either inside the critical section of a transaction commit
 759                  * (through transaction.c:qgroup_account_snapshot()) and when
 760                  * it's created before the transaction commit is started.
 761                  */
 762                 commit_root_gen = btrfs_header_generation(root->commit_root);
 763                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
 764         } else {
 765                 /*
 766                  * called by btrfs_reloc_post_snapshot_hook.
 767                  * the source tree is a reloc tree, all tree blocks
 768                  * modified after it was created have RELOC flag
 769                  * set in their headers. so it's OK to not update
 770                  * the 'last_snapshot'.
 771                  */
 772                 ret = btrfs_copy_root(trans, root, root->node, &eb,
 773                                       BTRFS_TREE_RELOC_OBJECTID);
 774                 BUG_ON(ret);
 775         }
 776
 777         memcpy(root_item, &root->root_item, sizeof(*root_item));
 778         btrfs_set_root_bytenr(root_item, eb->start);
 779         btrfs_set_root_level(root_item, btrfs_header_level(eb));
 780         btrfs_set_root_generation(root_item, trans->transid);
 781
 782         if (root->root_key.objectid == objectid) {
 783                 btrfs_set_root_refs(root_item, 0);
 784                 memset(&root_item->drop_progress, 0,
 785                        sizeof(struct btrfs_disk_key));
 786                 root_item->drop_level = 0;
 787         }
 788
 789         btrfs_tree_unlock(eb);
 790         free_extent_buffer(eb);
 791
 792         ret = btrfs_insert_root(trans, fs_info->tree_root,
 793                                 &root_key, root_item);
 794         BUG_ON(ret);
 795         kfree(root_item);
 796
 797         reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
 798         BUG_ON(IS_ERR(reloc_root));
 799         set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
 800         reloc_root->last_trans = trans->transid;
 801         return reloc_root;
 802 }
 803
 804 /*
 805  * create reloc tree for a given fs tree. reloc tree is just a
 806  * snapshot of the fs tree with special root objectid.
 807  *
 808  * The reloc_root comes out of here with two references, one for
 809  * root->reloc_root, and another for being on the rc->reloc_roots list.
 810  */
 811 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
 812                           struct btrfs_root *root)
 813 {
 814         struct btrfs_fs_info *fs_info = root->fs_info;
 815         struct btrfs_root *reloc_root;
 816         struct reloc_control *rc = fs_info->reloc_ctl;
 817         struct btrfs_block_rsv *rsv;
 818         int clear_rsv = 0;
 819         int ret;
 820
 821         if (!rc)
 822                 return 0;
 823
 824         /*
 825          * The subvolume has reloc tree but the swap is finished, no need to
 826          * create/update the dead reloc tree
 827          */
 828         if (reloc_root_is_dead(root))
 829                 return 0;
 830
 831         /*
 832          * This is subtle but important.  We do not do
 833          * record_root_in_transaction for reloc roots, instead we record their
 834          * corresponding fs root, and then here we update the last trans for the
 835          * reloc root.  This means that we have to do this for the entire life
 836          * of the reloc root, regardless of which stage of the relocation we are
 837          * in.
 838          */
 839         if (root->reloc_root) {
 840                 reloc_root = root->reloc_root;
 841                 reloc_root->last_trans = trans->transid;
 842                 return 0;
 843         }
 844
 845         /*
 846          * We are merging reloc roots, we do not need new reloc trees.  Also
 847          * reloc trees never need their own reloc tree.
 848          */
 849         if (!rc->create_reloc_tree ||
 850             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 851                 return 0;
 852
 853         if (!trans->reloc_reserved) {
 854                 rsv = trans->block_rsv;
 855                 trans->block_rsv = rc->block_rsv;
 856                 clear_rsv = 1;
 857         }
 858         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
 859         if (clear_rsv)
 860                 trans->block_rsv = rsv;
 861
 862         ret = __add_reloc_root(reloc_root);
 863         BUG_ON(ret < 0);
 864         root->reloc_root = btrfs_grab_root(reloc_root);
 865         return 0;
 866 }
 867
 868 /*
 869  * update root item of reloc tree
 870  */
 871 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
 872                             struct btrfs_root *root)
 873 {
 874         struct btrfs_fs_info *fs_info = root->fs_info;
 875         struct btrfs_root *reloc_root;
 876         struct btrfs_root_item *root_item;
 877         int ret;
 878
 879         if (!have_reloc_root(root))
 880                 goto out;
 881
 882         reloc_root = root->reloc_root;
 883         root_item = &reloc_root->root_item;
 884
 885         /*
 886          * We are probably ok here, but __del_reloc_root() will drop its ref of
 887          * the root.  We have the ref for root->reloc_root, but just in case
 888          * hold it while we update the reloc root.
 889          */
 890         btrfs_grab_root(reloc_root);
 891
 892         /* root->reloc_root will stay until current relocation finished */
 893         if (fs_info->reloc_ctl->merge_reloc_tree &&
 894             btrfs_root_refs(root_item) == 0) {
 895                 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
 896                 /*
 897                  * Mark the tree as dead before we change reloc_root so
 898                  * have_reloc_root will not touch it from now on.
 899                  */
 900                 smp_wmb();
 901                 __del_reloc_root(reloc_root);
 902         }
 903
 904         if (reloc_root->commit_root != reloc_root->node) {
 905                 __update_reloc_root(reloc_root);
 906                 btrfs_set_root_node(root_item, reloc_root->node);
 907                 free_extent_buffer(reloc_root->commit_root);
 908                 reloc_root->commit_root = btrfs_root_node(reloc_root);
 909         }
 910
 911         ret = btrfs_update_root(trans, fs_info->tree_root,
 912                                 &reloc_root->root_key, root_item);
 913         BUG_ON(ret);
 914         btrfs_put_root(reloc_root);
 915 out:
 916         return 0;
 917 }
 918
 919 /*
 920  * helper to find first cached inode with inode number >= objectid
 921  * in a subvolume
 922  */
 923 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
 924 {
 925         struct rb_node *node;
 926         struct rb_node *prev;
 927         struct btrfs_inode *entry;
 928         struct inode *inode;
 929
 930         spin_lock(&root->inode_lock);
 931 again:
 932         node = root->inode_tree.rb_node;
 933         prev = NULL;
 934         while (node) {
 935                 prev = node;
 936                 entry = rb_entry(node, struct btrfs_inode, rb_node);
 937
 938                 if (objectid < btrfs_ino(entry))
 939                         node = node->rb_left;
 940                 else if (objectid > btrfs_ino(entry))
 941                         node = node->rb_right;
 942                 else
 943                         break;
 944         }
 945         if (!node) {
 946                 while (prev) {
 947                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
 948                         if (objectid <= btrfs_ino(entry)) {
 949                                 node = prev;
 950                                 break;
 951                         }
 952                         prev = rb_next(prev);
 953                 }
 954         }
 955         while (node) {
 956                 entry = rb_entry(node, struct btrfs_inode, rb_node);
 957                 inode = igrab(&entry->vfs_inode);
 958                 if (inode) {
 959                         spin_unlock(&root->inode_lock);
 960                         return inode;
 961                 }
 962
 963                 objectid = btrfs_ino(entry) + 1;
 964                 if (cond_resched_lock(&root->inode_lock))
 965                         goto again;
 966
 967                 node = rb_next(node);
 968         }
 969         spin_unlock(&root->inode_lock);
 970         return NULL;
 971 }
 972
 973 /*
 974  * get new location of data
 975  */
 976 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
 977                             u64 bytenr, u64 num_bytes)
 978 {
 979         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
 980         struct btrfs_path *path;
 981         struct btrfs_file_extent_item *fi;
 982         struct extent_buffer *leaf;
 983         int ret;
 984
 985         path = btrfs_alloc_path();
 986         if (!path)
 987                 return -ENOMEM;
 988
 989         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
 990         ret = btrfs_lookup_file_extent(NULL, root, path,
 991                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
 992         if (ret < 0)
 993                 goto out;
 994         if (ret > 0) {
 995                 ret = -ENOENT;
 996                 goto out;
 997         }
 998
 999         leaf = path->nodes[0];
1000         fi = btrfs_item_ptr(leaf, path->slots[0],
1001                             struct btrfs_file_extent_item);
1002
1003         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1004                btrfs_file_extent_compression(leaf, fi) ||
1005                btrfs_file_extent_encryption(leaf, fi) ||
1006                btrfs_file_extent_other_encoding(leaf, fi));
1007
1008         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1009                 ret = -EINVAL;
1010                 goto out;
1011         }
1012
1013         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1014         ret = 0;
1015 out:
1016         btrfs_free_path(path);
1017         return ret;
1018 }
1019
1020 /*
1021  * update file extent items in the tree leaf to point to
1022  * the new locations.
1023  */
1024 static noinline_for_stack
1025 int replace_file_extents(struct btrfs_trans_handle *trans,
1026                          struct reloc_control *rc,
1027                          struct btrfs_root *root,
1028                          struct extent_buffer *leaf)
1029 {
1030         struct btrfs_fs_info *fs_info = root->fs_info;
1031         struct btrfs_key key;
1032         struct btrfs_file_extent_item *fi;
1033         struct inode *inode = NULL;
1034         u64 parent;
1035         u64 bytenr;
1036         u64 new_bytenr = 0;
1037         u64 num_bytes;
1038         u64 end;
1039         u32 nritems;
1040         u32 i;
1041         int ret = 0;
1042         int first = 1;
1043         int dirty = 0;
1044
1045         if (rc->stage != UPDATE_DATA_PTRS)
1046                 return 0;
1047
1048         /* reloc trees always use full backref */
1049         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1050                 parent = leaf->start;
1051         else
1052                 parent = 0;
1053
1054         nritems = btrfs_header_nritems(leaf);
1055         for (i = 0; i < nritems; i++) {
1056                 struct btrfs_ref ref = { 0 };
1057
1058                 cond_resched();
1059                 btrfs_item_key_to_cpu(leaf, &key, i);
1060                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1061                         continue;
1062                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1063                 if (btrfs_file_extent_type(leaf, fi) ==
1064                     BTRFS_FILE_EXTENT_INLINE)
1065                         continue;
1066                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1067                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1068                 if (bytenr == 0)
1069                         continue;
1070                 if (!in_range(bytenr, rc->block_group->start,
1071                               rc->block_group->length))
1072                         continue;
1073
1074                 /*
1075                  * if we are modifying block in fs tree, wait for readpage
1076                  * to complete and drop the extent cache
1077                  */
1078                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1079                         if (first) {
1080                                 inode = find_next_inode(root, key.objectid);
1081                                 first = 0;
1082                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1083                                 btrfs_add_delayed_iput(inode);
1084                                 inode = find_next_inode(root, key.objectid);
1085                         }
1086                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1087                                 end = key.offset +
1088                                       btrfs_file_extent_num_bytes(leaf, fi);
1089                                 WARN_ON(!IS_ALIGNED(key.offset,
1090                                                     fs_info->sectorsize));
1091                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1092                                 end--;
1093                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1094                                                       key.offset, end);
1095                                 if (!ret)
1096                                         continue;
1097
1098                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1099                                                 key.offset,     end, 1);
1100                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1101                                               key.offset, end);
1102                         }
1103                 }
1104
1105                 ret = get_new_location(rc->data_inode, &new_bytenr,
1106                                        bytenr, num_bytes);
1107                 if (ret) {
1108                         /*
1109                          * Don't have to abort since we've not changed anything
1110                          * in the file extent yet.
1111                          */
1112                         break;
1113                 }
1114
1115                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1116                 dirty = 1;
1117
1118                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1119                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1120                                        num_bytes, parent);
1121                 ref.real_root = root->root_key.objectid;
1122                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1123                                     key.objectid, key.offset);
1124                 ret = btrfs_inc_extent_ref(trans, &ref);
1125                 if (ret) {
1126                         btrfs_abort_transaction(trans, ret);
1127                         break;
1128                 }
1129
1130                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1131                                        num_bytes, parent);
1132                 ref.real_root = root->root_key.objectid;
1133                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1134                                     key.objectid, key.offset);
1135                 ret = btrfs_free_extent(trans, &ref);
1136                 if (ret) {
1137                         btrfs_abort_transaction(trans, ret);
1138                         break;
1139                 }
1140         }
1141         if (dirty)
1142                 btrfs_mark_buffer_dirty(leaf);
1143         if (inode)
1144                 btrfs_add_delayed_iput(inode);
1145         return ret;
1146 }
1147
1148 static noinline_for_stack
1149 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1150                      struct btrfs_path *path, int level)
1151 {
1152         struct btrfs_disk_key key1;
1153         struct btrfs_disk_key key2;
1154         btrfs_node_key(eb, &key1, slot);
1155         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1156         return memcmp(&key1, &key2, sizeof(key1));
1157 }
1158
1159 /*
1160  * try to replace tree blocks in fs tree with the new blocks
1161  * in reloc tree. tree blocks haven't been modified since the
1162  * reloc tree was create can be replaced.
1163  *
1164  * if a block was replaced, level of the block + 1 is returned.
1165  * if no block got replaced, 0 is returned. if there are other
1166  * errors, a negative error number is returned.
1167  */
1168 static noinline_for_stack
1169 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1170                  struct btrfs_root *dest, struct btrfs_root *src,
1171                  struct btrfs_path *path, struct btrfs_key *next_key,
1172                  int lowest_level, int max_level)
1173 {
1174         struct btrfs_fs_info *fs_info = dest->fs_info;
1175         struct extent_buffer *eb;
1176         struct extent_buffer *parent;
1177         struct btrfs_ref ref = { 0 };
1178         struct btrfs_key key;
1179         u64 old_bytenr;
1180         u64 new_bytenr;
1181         u64 old_ptr_gen;
1182         u64 new_ptr_gen;
1183         u64 last_snapshot;
1184         u32 blocksize;
1185         int cow = 0;
1186         int level;
1187         int ret;
1188         int slot;
1189
1190         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1191         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1192
1193         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1194 again:
1195         slot = path->slots[lowest_level];
1196         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1197
1198         eb = btrfs_lock_root_node(dest);
1199         btrfs_set_lock_blocking_write(eb);
1200         level = btrfs_header_level(eb);
1201
1202         if (level < lowest_level) {
1203                 btrfs_tree_unlock(eb);
1204                 free_extent_buffer(eb);
1205                 return 0;
1206         }
1207
1208         if (cow) {
1209                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1210                 BUG_ON(ret);
1211         }
1212         btrfs_set_lock_blocking_write(eb);
1213
1214         if (next_key) {
1215                 next_key->objectid = (u64)-1;
1216                 next_key->type = (u8)-1;
1217                 next_key->offset = (u64)-1;
1218         }
1219
1220         parent = eb;
1221         while (1) {
1222                 struct btrfs_key first_key;
1223
1224                 level = btrfs_header_level(parent);
1225                 BUG_ON(level < lowest_level);
1226
1227                 ret = btrfs_bin_search(parent, &key, &slot);
1228                 if (ret < 0)
1229                         break;
1230                 if (ret && slot > 0)
1231                         slot--;
1232
1233                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1234                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1235
1236                 old_bytenr = btrfs_node_blockptr(parent, slot);
1237                 blocksize = fs_info->nodesize;
1238                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1239                 btrfs_node_key_to_cpu(parent, &first_key, slot);
1240
1241                 if (level <= max_level) {
1242                         eb = path->nodes[level];
1243                         new_bytenr = btrfs_node_blockptr(eb,
1244                                                         path->slots[level]);
1245                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1246                                                         path->slots[level]);
1247                 } else {
1248                         new_bytenr = 0;
1249                         new_ptr_gen = 0;
1250                 }
1251
1252                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1253                         ret = level;
1254                         break;
1255                 }
1256
1257                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1258                     memcmp_node_keys(parent, slot, path, level)) {
1259                         if (level <= lowest_level) {
1260                                 ret = 0;
1261                                 break;
1262                         }
1263
1264                         eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen,
1265                                              level - 1, &first_key);
1266                         if (IS_ERR(eb)) {
1267                                 ret = PTR_ERR(eb);
1268                                 break;
1269                         } else if (!extent_buffer_uptodate(eb)) {
1270                                 ret = -EIO;
1271                                 free_extent_buffer(eb);
1272                                 break;
1273                         }
1274                         btrfs_tree_lock(eb);
1275                         if (cow) {
1276                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1277                                                       slot, &eb);
1278                                 BUG_ON(ret);
1279                         }
1280                         btrfs_set_lock_blocking_write(eb);
1281
1282                         btrfs_tree_unlock(parent);
1283                         free_extent_buffer(parent);
1284
1285                         parent = eb;
1286                         continue;
1287                 }
1288
1289                 if (!cow) {
1290                         btrfs_tree_unlock(parent);
1291                         free_extent_buffer(parent);
1292                         cow = 1;
1293                         goto again;
1294                 }
1295
1296                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1297                                       path->slots[level]);
1298                 btrfs_release_path(path);
1299
1300                 path->lowest_level = level;
1301                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1302                 path->lowest_level = 0;
1303                 BUG_ON(ret);
1304
1305                 /*
1306                  * Info qgroup to trace both subtrees.
1307                  *
1308                  * We must trace both trees.
1309                  * 1) Tree reloc subtree
1310                  *    If not traced, we will leak data numbers
1311                  * 2) Fs subtree
1312                  *    If not traced, we will double count old data
1313                  *
1314                  * We don't scan the subtree right now, but only record
1315                  * the swapped tree blocks.
1316                  * The real subtree rescan is delayed until we have new
1317                  * CoW on the subtree root node before transaction commit.
1318                  */
1319                 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1320                                 rc->block_group, parent, slot,
1321                                 path->nodes[level], path->slots[level],
1322                                 last_snapshot);
1323                 if (ret < 0)
1324                         break;
1325                 /*
1326                  * swap blocks in fs tree and reloc tree.
1327                  */
1328                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1329                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1330                 btrfs_mark_buffer_dirty(parent);
1331
1332                 btrfs_set_node_blockptr(path->nodes[level],
1333                                         path->slots[level], old_bytenr);
1334                 btrfs_set_node_ptr_generation(path->nodes[level],
1335                                               path->slots[level], old_ptr_gen);
1336                 btrfs_mark_buffer_dirty(path->nodes[level]);
1337
1338                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1339                                        blocksize, path->nodes[level]->start);
1340                 ref.skip_qgroup = true;
1341                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1342                 ret = btrfs_inc_extent_ref(trans, &ref);
1343                 BUG_ON(ret);
1344                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1345                                        blocksize, 0);
1346                 ref.skip_qgroup = true;
1347                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1348                 ret = btrfs_inc_extent_ref(trans, &ref);
1349                 BUG_ON(ret);
1350
1351                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1352                                        blocksize, path->nodes[level]->start);
1353                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1354                 ref.skip_qgroup = true;
1355                 ret = btrfs_free_extent(trans, &ref);
1356                 BUG_ON(ret);
1357
1358                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1359                                        blocksize, 0);
1360                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1361                 ref.skip_qgroup = true;
1362                 ret = btrfs_free_extent(trans, &ref);
1363                 BUG_ON(ret);
1364
1365                 btrfs_unlock_up_safe(path, 0);
1366
1367                 ret = level;
1368                 break;
1369         }
1370         btrfs_tree_unlock(parent);
1371         free_extent_buffer(parent);
1372         return ret;
1373 }
1374
1375 /*
1376  * helper to find next relocated block in reloc tree
1377  */
1378 static noinline_for_stack
1379 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1380                        int *level)
1381 {
1382         struct extent_buffer *eb;
1383         int i;
1384         u64 last_snapshot;
1385         u32 nritems;
1386
1387         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1388
1389         for (i = 0; i < *level; i++) {
1390                 free_extent_buffer(path->nodes[i]);
1391                 path->nodes[i] = NULL;
1392         }
1393
1394         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1395                 eb = path->nodes[i];
1396                 nritems = btrfs_header_nritems(eb);
1397                 while (path->slots[i] + 1 < nritems) {
1398                         path->slots[i]++;
1399                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1400                             last_snapshot)
1401                                 continue;
1402
1403                         *level = i;
1404                         return 0;
1405                 }
1406                 free_extent_buffer(path->nodes[i]);
1407                 path->nodes[i] = NULL;
1408         }
1409         return 1;
1410 }
1411
1412 /*
1413  * walk down reloc tree to find relocated block of lowest level
1414  */
1415 static noinline_for_stack
1416 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1417                          int *level)
1418 {
1419         struct btrfs_fs_info *fs_info = root->fs_info;
1420         struct extent_buffer *eb = NULL;
1421         int i;
1422         u64 bytenr;
1423         u64 ptr_gen = 0;
1424         u64 last_snapshot;
1425         u32 nritems;
1426
1427         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1428
1429         for (i = *level; i > 0; i--) {
1430                 struct btrfs_key first_key;
1431
1432                 eb = path->nodes[i];
1433                 nritems = btrfs_header_nritems(eb);
1434                 while (path->slots[i] < nritems) {
1435                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1436                         if (ptr_gen > last_snapshot)
1437                                 break;
1438                         path->slots[i]++;
1439                 }
1440                 if (path->slots[i] >= nritems) {
1441                         if (i == *level)
1442                                 break;
1443                         *level = i + 1;
1444                         return 0;
1445                 }
1446                 if (i == 1) {
1447                         *level = i;
1448                         return 0;
1449                 }
1450
1451                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1452                 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
1453                 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
1454                                      &first_key);
1455                 if (IS_ERR(eb)) {
1456                         return PTR_ERR(eb);
1457                 } else if (!extent_buffer_uptodate(eb)) {
1458                         free_extent_buffer(eb);
1459                         return -EIO;
1460                 }
1461                 BUG_ON(btrfs_header_level(eb) != i - 1);
1462                 path->nodes[i - 1] = eb;
1463                 path->slots[i - 1] = 0;
1464         }
1465         return 1;
1466 }
1467
1468 /*
1469  * invalidate extent cache for file extents whose key in range of
1470  * [min_key, max_key)
1471  */
1472 static int invalidate_extent_cache(struct btrfs_root *root,
1473                                    struct btrfs_key *min_key,
1474                                    struct btrfs_key *max_key)
1475 {
1476         struct btrfs_fs_info *fs_info = root->fs_info;
1477         struct inode *inode = NULL;
1478         u64 objectid;
1479         u64 start, end;
1480         u64 ino;
1481
1482         objectid = min_key->objectid;
1483         while (1) {
1484                 cond_resched();
1485                 iput(inode);
1486
1487                 if (objectid > max_key->objectid)
1488                         break;
1489
1490                 inode = find_next_inode(root, objectid);
1491                 if (!inode)
1492                         break;
1493                 ino = btrfs_ino(BTRFS_I(inode));
1494
1495                 if (ino > max_key->objectid) {
1496                         iput(inode);
1497                         break;
1498                 }
1499
1500                 objectid = ino + 1;
1501                 if (!S_ISREG(inode->i_mode))
1502                         continue;
1503
1504                 if (unlikely(min_key->objectid == ino)) {
1505                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1506                                 continue;
1507                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1508                                 start = 0;
1509                         else {
1510                                 start = min_key->offset;
1511                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1512                         }
1513                 } else {
1514                         start = 0;
1515                 }
1516
1517                 if (unlikely(max_key->objectid == ino)) {
1518                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1519                                 continue;
1520                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1521                                 end = (u64)-1;
1522                         } else {
1523                                 if (max_key->offset == 0)
1524                                         continue;
1525                                 end = max_key->offset;
1526                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1527                                 end--;
1528                         }
1529                 } else {
1530                         end = (u64)-1;
1531                 }
1532
1533                 /* the lock_extent waits for readpage to complete */
1534                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
1535                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
1536                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1537         }
1538         return 0;
1539 }
1540
1541 static int find_next_key(struct btrfs_path *path, int level,
1542                          struct btrfs_key *key)
1543
1544 {
1545         while (level < BTRFS_MAX_LEVEL) {
1546                 if (!path->nodes[level])
1547                         break;
1548                 if (path->slots[level] + 1 <
1549                     btrfs_header_nritems(path->nodes[level])) {
1550                         btrfs_node_key_to_cpu(path->nodes[level], key,
1551                                               path->slots[level] + 1);
1552                         return 0;
1553                 }
1554                 level++;
1555         }
1556         return 1;
1557 }
1558
1559 /*
1560  * Insert current subvolume into reloc_control::dirty_subvol_roots
1561  */
1562 static void insert_dirty_subvol(struct btrfs_trans_handle *trans,
1563                                 struct reloc_control *rc,
1564                                 struct btrfs_root *root)
1565 {
1566         struct btrfs_root *reloc_root = root->reloc_root;
1567         struct btrfs_root_item *reloc_root_item;
1568
1569         /* @root must be a subvolume tree root with a valid reloc tree */
1570         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1571         ASSERT(reloc_root);
1572
1573         reloc_root_item = &reloc_root->root_item;
1574         memset(&reloc_root_item->drop_progress, 0,
1575                 sizeof(reloc_root_item->drop_progress));
1576         reloc_root_item->drop_level = 0;
1577         btrfs_set_root_refs(reloc_root_item, 0);
1578         btrfs_update_reloc_root(trans, root);
1579
1580         if (list_empty(&root->reloc_dirty_list)) {
1581                 btrfs_grab_root(root);
1582                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1583         }
1584 }
1585
1586 static int clean_dirty_subvols(struct reloc_control *rc)
1587 {
1588         struct btrfs_root *root;
1589         struct btrfs_root *next;
1590         int ret = 0;
1591         int ret2;
1592
1593         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1594                                  reloc_dirty_list) {
1595                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1596                         /* Merged subvolume, cleanup its reloc root */
1597                         struct btrfs_root *reloc_root = root->reloc_root;
1598
1599                         list_del_init(&root->reloc_dirty_list);
1600                         root->reloc_root = NULL;
1601                         /*
1602                          * Need barrier to ensure clear_bit() only happens after
1603                          * root->reloc_root = NULL. Pairs with have_reloc_root.
1604                          */
1605                         smp_wmb();
1606                         clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1607                         if (reloc_root) {
1608                                 /*
1609                                  * btrfs_drop_snapshot drops our ref we hold for
1610                                  * ->reloc_root.  If it fails however we must
1611                                  * drop the ref ourselves.
1612                                  */
1613                                 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1614                                 if (ret2 < 0) {
1615                                         btrfs_put_root(reloc_root);
1616                                         if (!ret)
1617                                                 ret = ret2;
1618                                 }
1619                         }
1620                         btrfs_put_root(root);
1621                 } else {
1622                         /* Orphan reloc tree, just clean it up */
1623                         ret2 = btrfs_drop_snapshot(root, 0, 1);
1624                         if (ret2 < 0) {
1625                                 btrfs_put_root(root);
1626                                 if (!ret)
1627                                         ret = ret2;
1628                         }
1629                 }
1630         }
1631         return ret;
1632 }
1633
1634 /*
1635  * merge the relocated tree blocks in reloc tree with corresponding
1636  * fs tree.
1637  */
1638 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1639                                                struct btrfs_root *root)
1640 {
1641         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1642         struct btrfs_key key;
1643         struct btrfs_key next_key;
1644         struct btrfs_trans_handle *trans = NULL;
1645         struct btrfs_root *reloc_root;
1646         struct btrfs_root_item *root_item;
1647         struct btrfs_path *path;
1648         struct extent_buffer *leaf;
1649         int level;
1650         int max_level;
1651         int replaced = 0;
1652         int ret;
1653         int err = 0;
1654         u32 min_reserved;
1655
1656         path = btrfs_alloc_path();
1657         if (!path)
1658                 return -ENOMEM;
1659         path->reada = READA_FORWARD;
1660
1661         reloc_root = root->reloc_root;
1662         root_item = &reloc_root->root_item;
1663
1664         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1665                 level = btrfs_root_level(root_item);
1666                 atomic_inc(&reloc_root->node->refs);
1667                 path->nodes[level] = reloc_root->node;
1668                 path->slots[level] = 0;
1669         } else {
1670                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1671
1672                 level = root_item->drop_level;
1673                 BUG_ON(level == 0);
1674                 path->lowest_level = level;
1675                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1676                 path->lowest_level = 0;
1677                 if (ret < 0) {
1678                         btrfs_free_path(path);
1679                         return ret;
1680                 }
1681
1682                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1683                                       path->slots[level]);
1684                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1685
1686                 btrfs_unlock_up_safe(path, 0);
1687         }
1688
1689         min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1690         memset(&next_key, 0, sizeof(next_key));
1691
1692         while (1) {
1693                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
1694                                              BTRFS_RESERVE_FLUSH_ALL);
1695                 if (ret) {
1696                         err = ret;
1697                         goto out;
1698                 }
1699                 trans = btrfs_start_transaction(root, 0);
1700                 if (IS_ERR(trans)) {
1701                         err = PTR_ERR(trans);
1702                         trans = NULL;
1703                         goto out;
1704                 }
1705
1706                 /*
1707                  * At this point we no longer have a reloc_control, so we can't
1708                  * depend on btrfs_init_reloc_root to update our last_trans.
1709                  *
1710                  * But that's ok, we started the trans handle on our
1711                  * corresponding fs_root, which means it's been added to the
1712                  * dirty list.  At commit time we'll still call
1713                  * btrfs_update_reloc_root() and update our root item
1714                  * appropriately.
1715                  */
1716                 reloc_root->last_trans = trans->transid;
1717                 trans->block_rsv = rc->block_rsv;
1718
1719                 replaced = 0;
1720                 max_level = level;
1721
1722                 ret = walk_down_reloc_tree(reloc_root, path, &level);
1723                 if (ret < 0) {
1724                         err = ret;
1725                         goto out;
1726                 }
1727                 if (ret > 0)
1728                         break;
1729
1730                 if (!find_next_key(path, level, &key) &&
1731                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1732                         ret = 0;
1733                 } else {
1734                         ret = replace_path(trans, rc, root, reloc_root, path,
1735                                            &next_key, level, max_level);
1736                 }
1737                 if (ret < 0) {
1738                         err = ret;
1739                         goto out;
1740                 }
1741
1742                 if (ret > 0) {
1743                         level = ret;
1744                         btrfs_node_key_to_cpu(path->nodes[level], &key,
1745                                               path->slots[level]);
1746                         replaced = 1;
1747                 }
1748
1749                 ret = walk_up_reloc_tree(reloc_root, path, &level);
1750                 if (ret > 0)
1751                         break;
1752
1753                 BUG_ON(level == 0);
1754                 /*
1755                  * save the merging progress in the drop_progress.
1756                  * this is OK since root refs == 1 in this case.
1757                  */
1758                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1759                                path->slots[level]);
1760                 root_item->drop_level = level;
1761
1762                 btrfs_end_transaction_throttle(trans);
1763                 trans = NULL;
1764
1765                 btrfs_btree_balance_dirty(fs_info);
1766
1767                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1768                         invalidate_extent_cache(root, &key, &next_key);
1769         }
1770
1771         /*
1772          * handle the case only one block in the fs tree need to be
1773          * relocated and the block is tree root.
1774          */
1775         leaf = btrfs_lock_root_node(root);
1776         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
1777         btrfs_tree_unlock(leaf);
1778         free_extent_buffer(leaf);
1779         if (ret < 0)
1780                 err = ret;
1781 out:
1782         btrfs_free_path(path);
1783
1784         if (err == 0)
1785                 insert_dirty_subvol(trans, rc, root);
1786
1787         if (trans)
1788                 btrfs_end_transaction_throttle(trans);
1789
1790         btrfs_btree_balance_dirty(fs_info);
1791
1792         if (replaced && rc->stage == UPDATE_DATA_PTRS)
1793                 invalidate_extent_cache(root, &key, &next_key);
1794
1795         return err;
1796 }
1797
1798 static noinline_for_stack
1799 int prepare_to_merge(struct reloc_control *rc, int err)
1800 {
1801         struct btrfs_root *root = rc->extent_root;
1802         struct btrfs_fs_info *fs_info = root->fs_info;
1803         struct btrfs_root *reloc_root;
1804         struct btrfs_trans_handle *trans;
1805         LIST_HEAD(reloc_roots);
1806         u64 num_bytes = 0;
1807         int ret;
1808
1809         mutex_lock(&fs_info->reloc_mutex);
1810         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1811         rc->merging_rsv_size += rc->nodes_relocated * 2;
1812         mutex_unlock(&fs_info->reloc_mutex);
1813
1814 again:
1815         if (!err) {
1816                 num_bytes = rc->merging_rsv_size;
1817                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
1818                                           BTRFS_RESERVE_FLUSH_ALL);
1819                 if (ret)
1820                         err = ret;
1821         }
1822
1823         trans = btrfs_join_transaction(rc->extent_root);
1824         if (IS_ERR(trans)) {
1825                 if (!err)
1826                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1827                                                 num_bytes, NULL);
1828                 return PTR_ERR(trans);
1829         }
1830
1831         if (!err) {
1832                 if (num_bytes != rc->merging_rsv_size) {
1833                         btrfs_end_transaction(trans);
1834                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1835                                                 num_bytes, NULL);
1836                         goto again;
1837                 }
1838         }
1839
1840         rc->merge_reloc_tree = 1;
1841
1842         while (!list_empty(&rc->reloc_roots)) {
1843                 reloc_root = list_entry(rc->reloc_roots.next,
1844                                         struct btrfs_root, root_list);
1845                 list_del_init(&reloc_root->root_list);
1846
1847                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1848                                 false);
1849                 BUG_ON(IS_ERR(root));
1850                 BUG_ON(root->reloc_root != reloc_root);
1851
1852                 /*
1853                  * set reference count to 1, so btrfs_recover_relocation
1854                  * knows it should resumes merging
1855                  */
1856                 if (!err)
1857                         btrfs_set_root_refs(&reloc_root->root_item, 1);
1858                 btrfs_update_reloc_root(trans, root);
1859
1860                 list_add(&reloc_root->root_list, &reloc_roots);
1861                 btrfs_put_root(root);
1862         }
1863
1864         list_splice(&reloc_roots, &rc->reloc_roots);
1865
1866         if (!err)
1867                 btrfs_commit_transaction(trans);
1868         else
1869                 btrfs_end_transaction(trans);
1870         return err;
1871 }
1872
1873 static noinline_for_stack
1874 void free_reloc_roots(struct list_head *list)
1875 {
1876         struct btrfs_root *reloc_root, *tmp;
1877
1878         list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1879                 __del_reloc_root(reloc_root);
1880 }
1881
1882 static noinline_for_stack
1883 void merge_reloc_roots(struct reloc_control *rc)
1884 {
1885         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1886         struct btrfs_root *root;
1887         struct btrfs_root *reloc_root;
1888         LIST_HEAD(reloc_roots);
1889         int found = 0;
1890         int ret = 0;
1891 again:
1892         root = rc->extent_root;
1893
1894         /*
1895          * this serializes us with btrfs_record_root_in_transaction,
1896          * we have to make sure nobody is in the middle of
1897          * adding their roots to the list while we are
1898          * doing this splice
1899          */
1900         mutex_lock(&fs_info->reloc_mutex);
1901         list_splice_init(&rc->reloc_roots, &reloc_roots);
1902         mutex_unlock(&fs_info->reloc_mutex);
1903
1904         while (!list_empty(&reloc_roots)) {
1905                 found = 1;
1906                 reloc_root = list_entry(reloc_roots.next,
1907                                         struct btrfs_root, root_list);
1908
1909                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1910                                          false);
1911                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1912                         BUG_ON(IS_ERR(root));
1913                         BUG_ON(root->reloc_root != reloc_root);
1914                         ret = merge_reloc_root(rc, root);
1915                         btrfs_put_root(root);
1916                         if (ret) {
1917                                 if (list_empty(&reloc_root->root_list))
1918                                         list_add_tail(&reloc_root->root_list,
1919                                                       &reloc_roots);
1920                                 goto out;
1921                         }
1922                 } else {
1923                         if (!IS_ERR(root)) {
1924                                 if (root->reloc_root == reloc_root) {
1925                                         root->reloc_root = NULL;
1926                                         btrfs_put_root(reloc_root);
1927                                 }
1928                                 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
1929                                           &root->state);
1930                                 btrfs_put_root(root);
1931                         }
1932
1933                         list_del_init(&reloc_root->root_list);
1934                         /* Don't forget to queue this reloc root for cleanup */
1935                         list_add_tail(&reloc_root->reloc_dirty_list,
1936                                       &rc->dirty_subvol_roots);
1937                 }
1938         }
1939
1940         if (found) {
1941                 found = 0;
1942                 goto again;
1943         }
1944 out:
1945         if (ret) {
1946                 btrfs_handle_fs_error(fs_info, ret, NULL);
1947                 free_reloc_roots(&reloc_roots);
1948
1949                 /* new reloc root may be added */
1950                 mutex_lock(&fs_info->reloc_mutex);
1951                 list_splice_init(&rc->reloc_roots, &reloc_roots);
1952                 mutex_unlock(&fs_info->reloc_mutex);
1953                 free_reloc_roots(&reloc_roots);
1954         }
1955
1956         /*
1957          * We used to have
1958          *
1959          * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
1960          *
1961          * here, but it's wrong.  If we fail to start the transaction in
1962          * prepare_to_merge() we will have only 0 ref reloc roots, none of which
1963          * have actually been removed from the reloc_root_tree rb tree.  This is
1964          * fine because we're bailing here, and we hold a reference on the root
1965          * for the list that holds it, so these roots will be cleaned up when we
1966          * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
1967          * will be cleaned up on unmount.
1968          *
1969          * The remaining nodes will be cleaned up by free_reloc_control.
1970          */
1971 }
1972
1973 static void free_block_list(struct rb_root *blocks)
1974 {
1975         struct tree_block *block;
1976         struct rb_node *rb_node;
1977         while ((rb_node = rb_first(blocks))) {
1978                 block = rb_entry(rb_node, struct tree_block, rb_node);
1979                 rb_erase(rb_node, blocks);
1980                 kfree(block);
1981         }
1982 }
1983
1984 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
1985                                       struct btrfs_root *reloc_root)
1986 {
1987         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
1988         struct btrfs_root *root;
1989         int ret;
1990
1991         if (reloc_root->last_trans == trans->transid)
1992                 return 0;
1993
1994         root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
1995         BUG_ON(IS_ERR(root));
1996         BUG_ON(root->reloc_root != reloc_root);
1997         ret = btrfs_record_root_in_trans(trans, root);
1998         btrfs_put_root(root);
1999
2000         return ret;
2001 }
2002
2003 static noinline_for_stack
2004 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2005                                      struct reloc_control *rc,
2006                                      struct btrfs_backref_node *node,
2007                                      struct btrfs_backref_edge *edges[])
2008 {
2009         struct btrfs_backref_node *next;
2010         struct btrfs_root *root;
2011         int index = 0;
2012
2013         next = node;
2014         while (1) {
2015                 cond_resched();
2016                 next = walk_up_backref(next, edges, &index);
2017                 root = next->root;
2018                 BUG_ON(!root);
2019                 BUG_ON(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state));
2020
2021                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2022                         record_reloc_root_in_trans(trans, root);
2023                         break;
2024                 }
2025
2026                 btrfs_record_root_in_trans(trans, root);
2027                 root = root->reloc_root;
2028
2029                 if (next->new_bytenr != root->node->start) {
2030                         BUG_ON(next->new_bytenr);
2031                         BUG_ON(!list_empty(&next->list));
2032                         next->new_bytenr = root->node->start;
2033                         btrfs_put_root(next->root);
2034                         next->root = btrfs_grab_root(root);
2035                         ASSERT(next->root);
2036                         list_add_tail(&next->list,
2037                                       &rc->backref_cache.changed);
2038                         mark_block_processed(rc, next);
2039                         break;
2040                 }
2041
2042                 WARN_ON(1);
2043                 root = NULL;
2044                 next = walk_down_backref(edges, &index);
2045                 if (!next || next->level <= node->level)
2046                         break;
2047         }
2048         if (!root)
2049                 return NULL;
2050
2051         next = node;
2052         /* setup backref node path for btrfs_reloc_cow_block */
2053         while (1) {
2054                 rc->backref_cache.path[next->level] = next;
2055                 if (--index < 0)
2056                         break;
2057                 next = edges[index]->node[UPPER];
2058         }
2059         return root;
2060 }
2061
2062 /*
2063  * Select a tree root for relocation.
2064  *
2065  * Return NULL if the block is not shareable. We should use do_relocation() in
2066  * this case.
2067  *
2068  * Return a tree root pointer if the block is shareable.
2069  * Return -ENOENT if the block is root of reloc tree.
2070  */
2071 static noinline_for_stack
2072 struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2073 {
2074         struct btrfs_backref_node *next;
2075         struct btrfs_root *root;
2076         struct btrfs_root *fs_root = NULL;
2077         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2078         int index = 0;
2079
2080         next = node;
2081         while (1) {
2082                 cond_resched();
2083                 next = walk_up_backref(next, edges, &index);
2084                 root = next->root;
2085                 BUG_ON(!root);
2086
2087                 /* No other choice for non-shareable tree */
2088                 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2089                         return root;
2090
2091                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2092                         fs_root = root;
2093
2094                 if (next != node)
2095                         return NULL;
2096
2097                 next = walk_down_backref(edges, &index);
2098                 if (!next || next->level <= node->level)
2099                         break;
2100         }
2101
2102         if (!fs_root)
2103                 return ERR_PTR(-ENOENT);
2104         return fs_root;
2105 }
2106
2107 static noinline_for_stack
2108 u64 calcu_metadata_size(struct reloc_control *rc,
2109                         struct btrfs_backref_node *node, int reserve)
2110 {
2111         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2112         struct btrfs_backref_node *next = node;
2113         struct btrfs_backref_edge *edge;
2114         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2115         u64 num_bytes = 0;
2116         int index = 0;
2117
2118         BUG_ON(reserve && node->processed);
2119
2120         while (next) {
2121                 cond_resched();
2122                 while (1) {
2123                         if (next->processed && (reserve || next != node))
2124                                 break;
2125
2126                         num_bytes += fs_info->nodesize;
2127
2128                         if (list_empty(&next->upper))
2129                                 break;
2130
2131                         edge = list_entry(next->upper.next,
2132                                         struct btrfs_backref_edge, list[LOWER]);
2133                         edges[index++] = edge;
2134                         next = edge->node[UPPER];
2135                 }
2136                 next = walk_down_backref(edges, &index);
2137         }
2138         return num_bytes;
2139 }
2140
2141 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2142                                   struct reloc_control *rc,
2143                                   struct btrfs_backref_node *node)
2144 {
2145         struct btrfs_root *root = rc->extent_root;
2146         struct btrfs_fs_info *fs_info = root->fs_info;
2147         u64 num_bytes;
2148         int ret;
2149         u64 tmp;
2150
2151         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2152
2153         trans->block_rsv = rc->block_rsv;
2154         rc->reserved_bytes += num_bytes;
2155
2156         /*
2157          * We are under a transaction here so we can only do limited flushing.
2158          * If we get an enospc just kick back -EAGAIN so we know to drop the
2159          * transaction and try to refill when we can flush all the things.
2160          */
2161         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2162                                 BTRFS_RESERVE_FLUSH_LIMIT);
2163         if (ret) {
2164                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2165                 while (tmp <= rc->reserved_bytes)
2166                         tmp <<= 1;
2167                 /*
2168                  * only one thread can access block_rsv at this point,
2169                  * so we don't need hold lock to protect block_rsv.
2170                  * we expand more reservation size here to allow enough
2171                  * space for relocation and we will return earlier in
2172                  * enospc case.
2173                  */
2174                 rc->block_rsv->size = tmp + fs_info->nodesize *
2175                                       RELOCATION_RESERVED_NODES;
2176                 return -EAGAIN;
2177         }
2178
2179         return 0;
2180 }
2181
2182 /*
2183  * relocate a block tree, and then update pointers in upper level
2184  * blocks that reference the block to point to the new location.
2185  *
2186  * if called by link_to_upper, the block has already been relocated.
2187  * in that case this function just updates pointers.
2188  */
2189 static int do_relocation(struct btrfs_trans_handle *trans,
2190                          struct reloc_control *rc,
2191                          struct btrfs_backref_node *node,
2192                          struct btrfs_key *key,
2193                          struct btrfs_path *path, int lowest)
2194 {
2195         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2196         struct btrfs_backref_node *upper;
2197         struct btrfs_backref_edge *edge;
2198         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2199         struct btrfs_root *root;
2200         struct extent_buffer *eb;
2201         u32 blocksize;
2202         u64 bytenr;
2203         u64 generation;
2204         int slot;
2205         int ret;
2206         int err = 0;
2207
2208         BUG_ON(lowest && node->eb);
2209
2210         path->lowest_level = node->level + 1;
2211         rc->backref_cache.path[node->level] = node;
2212         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2213                 struct btrfs_key first_key;
2214                 struct btrfs_ref ref = { 0 };
2215
2216                 cond_resched();
2217
2218                 upper = edge->node[UPPER];
2219                 root = select_reloc_root(trans, rc, upper, edges);
2220                 BUG_ON(!root);
2221
2222                 if (upper->eb && !upper->locked) {
2223                         if (!lowest) {
2224                                 ret = btrfs_bin_search(upper->eb, key, &slot);
2225                                 if (ret < 0) {
2226                                         err = ret;
2227                                         goto next;
2228                                 }
2229                                 BUG_ON(ret);
2230                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2231                                 if (node->eb->start == bytenr)
2232                                         goto next;
2233                         }
2234                         btrfs_backref_drop_node_buffer(upper);
2235                 }
2236
2237                 if (!upper->eb) {
2238                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2239                         if (ret) {
2240                                 if (ret < 0)
2241                                         err = ret;
2242                                 else
2243                                         err = -ENOENT;
2244
2245                                 btrfs_release_path(path);
2246                                 break;
2247                         }
2248
2249                         if (!upper->eb) {
2250                                 upper->eb = path->nodes[upper->level];
2251                                 path->nodes[upper->level] = NULL;
2252                         } else {
2253                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2254                         }
2255
2256                         upper->locked = 1;
2257                         path->locks[upper->level] = 0;
2258
2259                         slot = path->slots[upper->level];
2260                         btrfs_release_path(path);
2261                 } else {
2262                         ret = btrfs_bin_search(upper->eb, key, &slot);
2263                         if (ret < 0) {
2264                                 err = ret;
2265                                 goto next;
2266                         }
2267                         BUG_ON(ret);
2268                 }
2269
2270                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2271                 if (lowest) {
2272                         if (bytenr != node->bytenr) {
2273                                 btrfs_err(root->fs_info,
2274                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2275                                           bytenr, node->bytenr, slot,
2276                                           upper->eb->start);
2277                                 err = -EIO;
2278                                 goto next;
2279                         }
2280                 } else {
2281                         if (node->eb->start == bytenr)
2282                                 goto next;
2283                 }
2284
2285                 blocksize = root->fs_info->nodesize;
2286                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2287                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2288                 eb = read_tree_block(fs_info, bytenr, generation,
2289                                      upper->level - 1, &first_key);
2290                 if (IS_ERR(eb)) {
2291                         err = PTR_ERR(eb);
2292                         goto next;
2293                 } else if (!extent_buffer_uptodate(eb)) {
2294                         free_extent_buffer(eb);
2295                         err = -EIO;
2296                         goto next;
2297                 }
2298                 btrfs_tree_lock(eb);
2299                 btrfs_set_lock_blocking_write(eb);
2300
2301                 if (!node->eb) {
2302                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2303                                               slot, &eb);
2304                         btrfs_tree_unlock(eb);
2305                         free_extent_buffer(eb);
2306                         if (ret < 0) {
2307                                 err = ret;
2308                                 goto next;
2309                         }
2310                         BUG_ON(node->eb != eb);
2311                 } else {
2312                         btrfs_set_node_blockptr(upper->eb, slot,
2313                                                 node->eb->start);
2314                         btrfs_set_node_ptr_generation(upper->eb, slot,
2315                                                       trans->transid);
2316                         btrfs_mark_buffer_dirty(upper->eb);
2317
2318                         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2319                                                node->eb->start, blocksize,
2320                                                upper->eb->start);
2321                         ref.real_root = root->root_key.objectid;
2322                         btrfs_init_tree_ref(&ref, node->level,
2323                                             btrfs_header_owner(upper->eb));
2324                         ret = btrfs_inc_extent_ref(trans, &ref);
2325                         BUG_ON(ret);
2326
2327                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2328                         BUG_ON(ret);
2329                 }
2330 next:
2331                 if (!upper->pending)
2332                         btrfs_backref_drop_node_buffer(upper);
2333                 else
2334                         btrfs_backref_unlock_node_buffer(upper);
2335                 if (err)
2336                         break;
2337         }
2338
2339         if (!err && node->pending) {
2340                 btrfs_backref_drop_node_buffer(node);
2341                 list_move_tail(&node->list, &rc->backref_cache.changed);
2342                 node->pending = 0;
2343         }
2344
2345         path->lowest_level = 0;
2346         BUG_ON(err == -ENOSPC);
2347         return err;
2348 }
2349
2350 static int link_to_upper(struct btrfs_trans_handle *trans,
2351                          struct reloc_control *rc,
2352                          struct btrfs_backref_node *node,
2353                          struct btrfs_path *path)
2354 {
2355         struct btrfs_key key;
2356
2357         btrfs_node_key_to_cpu(node->eb, &key, 0);
2358         return do_relocation(trans, rc, node, &key, path, 0);
2359 }
2360
2361 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2362                                 struct reloc_control *rc,
2363                                 struct btrfs_path *path, int err)
2364 {
2365         LIST_HEAD(list);
2366         struct btrfs_backref_cache *cache = &rc->backref_cache;
2367         struct btrfs_backref_node *node;
2368         int level;
2369         int ret;
2370
2371         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2372                 while (!list_empty(&cache->pending[level])) {
2373                         node = list_entry(cache->pending[level].next,
2374                                           struct btrfs_backref_node, list);
2375                         list_move_tail(&node->list, &list);
2376                         BUG_ON(!node->pending);
2377
2378                         if (!err) {
2379                                 ret = link_to_upper(trans, rc, node, path);
2380                                 if (ret < 0)
2381                                         err = ret;
2382                         }
2383                 }
2384                 list_splice_init(&list, &cache->pending[level]);
2385         }
2386         return err;
2387 }
2388
2389 /*
2390  * mark a block and all blocks directly/indirectly reference the block
2391  * as processed.
2392  */
2393 static void update_processed_blocks(struct reloc_control *rc,
2394                                     struct btrfs_backref_node *node)
2395 {
2396         struct btrfs_backref_node *next = node;
2397         struct btrfs_backref_edge *edge;
2398         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2399         int index = 0;
2400
2401         while (next) {
2402                 cond_resched();
2403                 while (1) {
2404                         if (next->processed)
2405                                 break;
2406
2407                         mark_block_processed(rc, next);
2408
2409                         if (list_empty(&next->upper))
2410                                 break;
2411
2412                         edge = list_entry(next->upper.next,
2413                                         struct btrfs_backref_edge, list[LOWER]);
2414                         edges[index++] = edge;
2415                         next = edge->node[UPPER];
2416                 }
2417                 next = walk_down_backref(edges, &index);
2418         }
2419 }
2420
2421 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2422 {
2423         u32 blocksize = rc->extent_root->fs_info->nodesize;
2424
2425         if (test_range_bit(&rc->processed_blocks, bytenr,
2426                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2427                 return 1;
2428         return 0;
2429 }
2430
2431 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2432                               struct tree_block *block)
2433 {
2434         struct extent_buffer *eb;
2435
2436         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2437                              block->level, NULL);
2438         if (IS_ERR(eb)) {
2439                 return PTR_ERR(eb);
2440         } else if (!extent_buffer_uptodate(eb)) {
2441                 free_extent_buffer(eb);
2442                 return -EIO;
2443         }
2444         if (block->level == 0)
2445                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2446         else
2447                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2448         free_extent_buffer(eb);
2449         block->key_ready = 1;
2450         return 0;
2451 }
2452
2453 /*
2454  * helper function to relocate a tree block
2455  */
2456 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2457                                 struct reloc_control *rc,
2458                                 struct btrfs_backref_node *node,
2459                                 struct btrfs_key *key,
2460                                 struct btrfs_path *path)
2461 {
2462         struct btrfs_root *root;
2463         int ret = 0;
2464
2465         if (!node)
2466                 return 0;
2467
2468         /*
2469          * If we fail here we want to drop our backref_node because we are going
2470          * to start over and regenerate the tree for it.
2471          */
2472         ret = reserve_metadata_space(trans, rc, node);
2473         if (ret)
2474                 goto out;
2475
2476         BUG_ON(node->processed);
2477         root = select_one_root(node);
2478         if (root == ERR_PTR(-ENOENT)) {
2479                 update_processed_blocks(rc, node);
2480                 goto out;
2481         }
2482
2483         if (root) {
2484                 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2485                         BUG_ON(node->new_bytenr);
2486                         BUG_ON(!list_empty(&node->list));
2487                         btrfs_record_root_in_trans(trans, root);
2488                         root = root->reloc_root;
2489                         node->new_bytenr = root->node->start;
2490                         btrfs_put_root(node->root);
2491                         node->root = btrfs_grab_root(root);
2492                         ASSERT(node->root);
2493                         list_add_tail(&node->list, &rc->backref_cache.changed);
2494                 } else {
2495                         path->lowest_level = node->level;
2496                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2497                         btrfs_release_path(path);
2498                         if (ret > 0)
2499                                 ret = 0;
2500                 }
2501                 if (!ret)
2502                         update_processed_blocks(rc, node);
2503         } else {
2504                 ret = do_relocation(trans, rc, node, key, path, 1);
2505         }
2506 out:
2507         if (ret || node->level == 0 || node->cowonly)
2508                 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2509         return ret;
2510 }
2511
2512 /*
2513  * relocate a list of blocks
2514  */
2515 static noinline_for_stack
2516 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2517                          struct reloc_control *rc, struct rb_root *blocks)
2518 {
2519         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2520         struct btrfs_backref_node *node;
2521         struct btrfs_path *path;
2522         struct tree_block *block;
2523         struct tree_block *next;
2524         int ret;
2525         int err = 0;
2526
2527         path = btrfs_alloc_path();
2528         if (!path) {
2529                 err = -ENOMEM;
2530                 goto out_free_blocks;
2531         }
2532
2533         /* Kick in readahead for tree blocks with missing keys */
2534         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2535                 if (!block->key_ready)
2536                         readahead_tree_block(fs_info, block->bytenr);
2537         }
2538
2539         /* Get first keys */
2540         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2541                 if (!block->key_ready) {
2542                         err = get_tree_block_key(fs_info, block);
2543                         if (err)
2544                                 goto out_free_path;
2545                 }
2546         }
2547
2548         /* Do tree relocation */
2549         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2550                 node = build_backref_tree(rc, &block->key,
2551                                           block->level, block->bytenr);
2552                 if (IS_ERR(node)) {
2553                         err = PTR_ERR(node);
2554                         goto out;
2555                 }
2556
2557                 ret = relocate_tree_block(trans, rc, node, &block->key,
2558                                           path);
2559                 if (ret < 0) {
2560                         err = ret;
2561                         break;
2562                 }
2563         }
2564 out:
2565         err = finish_pending_nodes(trans, rc, path, err);
2566
2567 out_free_path:
2568         btrfs_free_path(path);
2569 out_free_blocks:
2570         free_block_list(blocks);
2571         return err;
2572 }
2573
2574 static noinline_for_stack
2575 int prealloc_file_extent_cluster(struct inode *inode,
2576                                  struct file_extent_cluster *cluster)
2577 {
2578         u64 alloc_hint = 0;
2579         u64 start;
2580         u64 end;
2581         u64 offset = BTRFS_I(inode)->index_cnt;
2582         u64 num_bytes;
2583         int nr = 0;
2584         int ret = 0;
2585         u64 prealloc_start = cluster->start - offset;
2586         u64 prealloc_end = cluster->end - offset;
2587         u64 cur_offset;
2588         struct extent_changeset *data_reserved = NULL;
2589
2590         BUG_ON(cluster->start != cluster->boundary[0]);
2591         inode_lock(inode);
2592
2593         ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
2594                                           prealloc_end + 1 - prealloc_start);
2595         if (ret)
2596                 goto out;
2597
2598         cur_offset = prealloc_start;
2599         while (nr < cluster->nr) {
2600                 start = cluster->boundary[nr] - offset;
2601                 if (nr + 1 < cluster->nr)
2602                         end = cluster->boundary[nr + 1] - 1 - offset;
2603                 else
2604                         end = cluster->end - offset;
2605
2606                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2607                 num_bytes = end + 1 - start;
2608                 if (cur_offset < start)
2609                         btrfs_free_reserved_data_space(inode, data_reserved,
2610                                         cur_offset, start - cur_offset);
2611                 ret = btrfs_prealloc_file_range(inode, 0, start,
2612                                                 num_bytes, num_bytes,
2613                                                 end + 1, &alloc_hint);
2614                 cur_offset = end + 1;
2615                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2616                 if (ret)
2617                         break;
2618                 nr++;
2619         }
2620         if (cur_offset < prealloc_end)
2621                 btrfs_free_reserved_data_space(inode, data_reserved,
2622                                 cur_offset, prealloc_end + 1 - cur_offset);
2623 out:
2624         inode_unlock(inode);
2625         extent_changeset_free(data_reserved);
2626         return ret;
2627 }
2628
2629 static noinline_for_stack
2630 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2631                          u64 block_start)
2632 {
2633         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2634         struct extent_map *em;
2635         int ret = 0;
2636
2637         em = alloc_extent_map();
2638         if (!em)
2639                 return -ENOMEM;
2640
2641         em->start = start;
2642         em->len = end + 1 - start;
2643         em->block_len = em->len;
2644         em->block_start = block_start;
2645         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2646
2647         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2648         while (1) {
2649                 write_lock(&em_tree->lock);
2650                 ret = add_extent_mapping(em_tree, em, 0);
2651                 write_unlock(&em_tree->lock);
2652                 if (ret != -EEXIST) {
2653                         free_extent_map(em);
2654                         break;
2655                 }
2656                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
2657         }
2658         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2659         return ret;
2660 }
2661
2662 /*
2663  * Allow error injection to test balance cancellation
2664  */
2665 int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2666 {
2667         return atomic_read(&fs_info->balance_cancel_req);
2668 }
2669 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2670
2671 static int relocate_file_extent_cluster(struct inode *inode,
2672                                         struct file_extent_cluster *cluster)
2673 {
2674         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2675         u64 page_start;
2676         u64 page_end;
2677         u64 offset = BTRFS_I(inode)->index_cnt;
2678         unsigned long index;
2679         unsigned long last_index;
2680         struct page *page;
2681         struct file_ra_state *ra;
2682         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2683         int nr = 0;
2684         int ret = 0;
2685
2686         if (!cluster->nr)
2687                 return 0;
2688
2689         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2690         if (!ra)
2691                 return -ENOMEM;
2692
2693         ret = prealloc_file_extent_cluster(inode, cluster);
2694         if (ret)
2695                 goto out;
2696
2697         file_ra_state_init(ra, inode->i_mapping);
2698
2699         ret = setup_extent_mapping(inode, cluster->start - offset,
2700                                    cluster->end - offset, cluster->start);
2701         if (ret)
2702                 goto out;
2703
2704         index = (cluster->start - offset) >> PAGE_SHIFT;
2705         last_index = (cluster->end - offset) >> PAGE_SHIFT;
2706         while (index <= last_index) {
2707                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
2708                                 PAGE_SIZE);
2709                 if (ret)
2710                         goto out;
2711
2712                 page = find_lock_page(inode->i_mapping, index);
2713                 if (!page) {
2714                         page_cache_sync_readahead(inode->i_mapping,
2715                                                   ra, NULL, index,
2716                                                   last_index + 1 - index);
2717                         page = find_or_create_page(inode->i_mapping, index,
2718                                                    mask);
2719                         if (!page) {
2720                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2721                                                         PAGE_SIZE, true);
2722                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
2723                                                         PAGE_SIZE);
2724                                 ret = -ENOMEM;
2725                                 goto out;
2726                         }
2727                 }
2728
2729                 if (PageReadahead(page)) {
2730                         page_cache_async_readahead(inode->i_mapping,
2731                                                    ra, NULL, page, index,
2732                                                    last_index + 1 - index);
2733                 }
2734
2735                 if (!PageUptodate(page)) {
2736                         btrfs_readpage(NULL, page);
2737                         lock_page(page);
2738                         if (!PageUptodate(page)) {
2739                                 unlock_page(page);
2740                                 put_page(page);
2741                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2742                                                         PAGE_SIZE, true);
2743                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
2744                                                                PAGE_SIZE);
2745                                 ret = -EIO;
2746                                 goto out;
2747                         }
2748                 }
2749
2750                 page_start = page_offset(page);
2751                 page_end = page_start + PAGE_SIZE - 1;
2752
2753                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
2754
2755                 set_page_extent_mapped(page);
2756
2757                 if (nr < cluster->nr &&
2758                     page_start + offset == cluster->boundary[nr]) {
2759                         set_extent_bits(&BTRFS_I(inode)->io_tree,
2760                                         page_start, page_end,
2761                                         EXTENT_BOUNDARY);
2762                         nr++;
2763                 }
2764
2765                 ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2766                                                 NULL);
2767                 if (ret) {
2768                         unlock_page(page);
2769                         put_page(page);
2770                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
2771                                                          PAGE_SIZE, true);
2772                         btrfs_delalloc_release_extents(BTRFS_I(inode),
2773                                                        PAGE_SIZE);
2774
2775                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
2776                                           page_start, page_end,
2777                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
2778                         goto out;
2779
2780                 }
2781                 set_page_dirty(page);
2782
2783                 unlock_extent(&BTRFS_I(inode)->io_tree,
2784                               page_start, page_end);
2785                 unlock_page(page);
2786                 put_page(page);
2787
2788                 index++;
2789                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2790                 balance_dirty_pages_ratelimited(inode->i_mapping);
2791                 btrfs_throttle(fs_info);
2792                 if (btrfs_should_cancel_balance(fs_info)) {
2793                         ret = -ECANCELED;
2794                         goto out;
2795                 }
2796         }
2797         WARN_ON(nr != cluster->nr);
2798 out:
2799         kfree(ra);
2800         return ret;
2801 }
2802
2803 static noinline_for_stack
2804 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
2805                          struct file_extent_cluster *cluster)
2806 {
2807         int ret;
2808
2809         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
2810                 ret = relocate_file_extent_cluster(inode, cluster);
2811                 if (ret)
2812                         return ret;
2813                 cluster->nr = 0;
2814         }
2815
2816         if (!cluster->nr)
2817                 cluster->start = extent_key->objectid;
2818         else
2819                 BUG_ON(cluster->nr >= MAX_EXTENTS);
2820         cluster->end = extent_key->objectid + extent_key->offset - 1;
2821         cluster->boundary[cluster->nr] = extent_key->objectid;
2822         cluster->nr++;
2823
2824         if (cluster->nr >= MAX_EXTENTS) {
2825                 ret = relocate_file_extent_cluster(inode, cluster);
2826                 if (ret)
2827                         return ret;
2828                 cluster->nr = 0;
2829         }
2830         return 0;
2831 }
2832
2833 /*
2834  * helper to add a tree block to the list.
2835  * the major work is getting the generation and level of the block
2836  */
2837 static int add_tree_block(struct reloc_control *rc,
2838                           struct btrfs_key *extent_key,
2839                           struct btrfs_path *path,
2840                           struct rb_root *blocks)
2841 {
2842         struct extent_buffer *eb;
2843         struct btrfs_extent_item *ei;
2844         struct btrfs_tree_block_info *bi;
2845         struct tree_block *block;
2846         struct rb_node *rb_node;
2847         u32 item_size;
2848         int level = -1;
2849         u64 generation;
2850
2851         eb =  path->nodes[0];
2852         item_size = btrfs_item_size_nr(eb, path->slots[0]);
2853
2854         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
2855             item_size >= sizeof(*ei) + sizeof(*bi)) {
2856                 ei = btrfs_item_ptr(eb, path->slots[0],
2857                                 struct btrfs_extent_item);
2858                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
2859                         bi = (struct btrfs_tree_block_info *)(ei + 1);
2860                         level = btrfs_tree_block_level(eb, bi);
2861                 } else {
2862                         level = (int)extent_key->offset;
2863                 }
2864                 generation = btrfs_extent_generation(eb, ei);
2865         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
2866                 btrfs_print_v0_err(eb->fs_info);
2867                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
2868                 return -EINVAL;
2869         } else {
2870                 BUG();
2871         }
2872
2873         btrfs_release_path(path);
2874
2875         BUG_ON(level == -1);
2876
2877         block = kmalloc(sizeof(*block), GFP_NOFS);
2878         if (!block)
2879                 return -ENOMEM;
2880
2881         block->bytenr = extent_key->objectid;
2882         block->key.objectid = rc->extent_root->fs_info->nodesize;
2883         block->key.offset = generation;
2884         block->level = level;
2885         block->key_ready = 0;
2886
2887         rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
2888         if (rb_node)
2889                 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
2890                                     -EEXIST);
2891
2892         return 0;
2893 }
2894
2895 /*
2896  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
2897  */
2898 static int __add_tree_block(struct reloc_control *rc,
2899                             u64 bytenr, u32 blocksize,
2900                             struct rb_root *blocks)
2901 {
2902         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2903         struct btrfs_path *path;
2904         struct btrfs_key key;
2905         int ret;
2906         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2907
2908         if (tree_block_processed(bytenr, rc))
2909                 return 0;
2910
2911         if (rb_simple_search(blocks, bytenr))
2912                 return 0;
2913
2914         path = btrfs_alloc_path();
2915         if (!path)
2916                 return -ENOMEM;
2917 again:
2918         key.objectid = bytenr;
2919         if (skinny) {
2920                 key.type = BTRFS_METADATA_ITEM_KEY;
2921                 key.offset = (u64)-1;
2922         } else {
2923                 key.type = BTRFS_EXTENT_ITEM_KEY;
2924                 key.offset = blocksize;
2925         }
2926
2927         path->search_commit_root = 1;
2928         path->skip_locking = 1;
2929         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
2930         if (ret < 0)
2931                 goto out;
2932
2933         if (ret > 0 && skinny) {
2934                 if (path->slots[0]) {
2935                         path->slots[0]--;
2936                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2937                                               path->slots[0]);
2938                         if (key.objectid == bytenr &&
2939                             (key.type == BTRFS_METADATA_ITEM_KEY ||
2940                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
2941                               key.offset == blocksize)))
2942                                 ret = 0;
2943                 }
2944
2945                 if (ret) {
2946                         skinny = false;
2947                         btrfs_release_path(path);
2948                         goto again;
2949                 }
2950         }
2951         if (ret) {
2952                 ASSERT(ret == 1);
2953                 btrfs_print_leaf(path->nodes[0]);
2954                 btrfs_err(fs_info,
2955              "tree block extent item (%llu) is not found in extent tree",
2956                      bytenr);
2957                 WARN_ON(1);
2958                 ret = -EINVAL;
2959                 goto out;
2960         }
2961
2962         ret = add_tree_block(rc, &key, path, blocks);
2963 out:
2964         btrfs_free_path(path);
2965         return ret;
2966 }
2967
2968 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
2969                                     struct btrfs_block_group *block_group,
2970                                     struct inode *inode,
2971                                     u64 ino)
2972 {
2973         struct btrfs_root *root = fs_info->tree_root;
2974         struct btrfs_trans_handle *trans;
2975         int ret = 0;
2976
2977         if (inode)
2978                 goto truncate;
2979
2980         inode = btrfs_iget(fs_info->sb, ino, root);
2981         if (IS_ERR(inode))
2982                 return -ENOENT;
2983
2984 truncate:
2985         ret = btrfs_check_trunc_cache_free_space(fs_info,
2986                                                  &fs_info->global_block_rsv);
2987         if (ret)
2988                 goto out;
2989
2990         trans = btrfs_join_transaction(root);
2991         if (IS_ERR(trans)) {
2992                 ret = PTR_ERR(trans);
2993                 goto out;
2994         }
2995
2996         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
2997
2998         btrfs_end_transaction(trans);
2999         btrfs_btree_balance_dirty(fs_info);
3000 out:
3001         iput(inode);
3002         return ret;
3003 }
3004
3005 /*
3006  * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3007  * cache inode, to avoid free space cache data extent blocking data relocation.
3008  */
3009 static int delete_v1_space_cache(struct extent_buffer *leaf,
3010                                  struct btrfs_block_group *block_group,
3011                                  u64 data_bytenr)
3012 {
3013         u64 space_cache_ino;
3014         struct btrfs_file_extent_item *ei;
3015         struct btrfs_key key;
3016         bool found = false;
3017         int i;
3018         int ret;
3019
3020         if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3021                 return 0;
3022
3023         for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3024                 btrfs_item_key_to_cpu(leaf, &key, i);
3025                 if (key.type != BTRFS_EXTENT_DATA_KEY)
3026                         continue;
3027                 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3028                 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG &&
3029                     btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3030                         found = true;
3031                         space_cache_ino = key.objectid;
3032                         break;
3033                 }
3034         }
3035         if (!found)
3036                 return -ENOENT;
3037         ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3038                                         space_cache_ino);
3039         return ret;
3040 }
3041
3042 /*
3043  * helper to find all tree blocks that reference a given data extent
3044  */
3045 static noinline_for_stack
3046 int add_data_references(struct reloc_control *rc,
3047                         struct btrfs_key *extent_key,
3048                         struct btrfs_path *path,
3049                         struct rb_root *blocks)
3050 {
3051         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3052         struct ulist *leaves = NULL;
3053         struct ulist_iterator leaf_uiter;
3054         struct ulist_node *ref_node = NULL;
3055         const u32 blocksize = fs_info->nodesize;
3056         int ret = 0;
3057
3058         btrfs_release_path(path);
3059         ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid,
3060                                    0, &leaves, NULL, true);
3061         if (ret < 0)
3062                 return ret;
3063
3064         ULIST_ITER_INIT(&leaf_uiter);
3065         while ((ref_node = ulist_next(leaves, &leaf_uiter))) {
3066                 struct extent_buffer *eb;
3067
3068                 eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL);
3069                 if (IS_ERR(eb)) {
3070                         ret = PTR_ERR(eb);
3071                         break;
3072                 }
3073                 ret = delete_v1_space_cache(eb, rc->block_group,
3074                                             extent_key->objectid);
3075                 free_extent_buffer(eb);
3076                 if (ret < 0)
3077                         break;
3078                 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3079                 if (ret < 0)
3080                         break;
3081         }
3082         if (ret < 0)
3083                 free_block_list(blocks);
3084         ulist_free(leaves);
3085         return ret;
3086 }
3087
3088 /*
3089  * helper to find next unprocessed extent
3090  */
3091 static noinline_for_stack
3092 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3093                      struct btrfs_key *extent_key)
3094 {
3095         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3096         struct btrfs_key key;
3097         struct extent_buffer *leaf;
3098         u64 start, end, last;
3099         int ret;
3100
3101         last = rc->block_group->start + rc->block_group->length;
3102         while (1) {
3103                 cond_resched();
3104                 if (rc->search_start >= last) {
3105                         ret = 1;
3106                         break;
3107                 }
3108
3109                 key.objectid = rc->search_start;
3110                 key.type = BTRFS_EXTENT_ITEM_KEY;
3111                 key.offset = 0;
3112
3113                 path->search_commit_root = 1;
3114                 path->skip_locking = 1;
3115                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3116                                         0, 0);
3117                 if (ret < 0)
3118                         break;
3119 next:
3120                 leaf = path->nodes[0];
3121                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3122                         ret = btrfs_next_leaf(rc->extent_root, path);
3123                         if (ret != 0)
3124                                 break;
3125                         leaf = path->nodes[0];
3126                 }
3127
3128                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3129                 if (key.objectid >= last) {
3130                         ret = 1;
3131                         break;
3132                 }
3133
3134                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3135                     key.type != BTRFS_METADATA_ITEM_KEY) {
3136                         path->slots[0]++;
3137                         goto next;
3138                 }
3139
3140                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3141                     key.objectid + key.offset <= rc->search_start) {
3142                         path->slots[0]++;
3143                         goto next;
3144                 }
3145
3146                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3147                     key.objectid + fs_info->nodesize <=
3148                     rc->search_start) {
3149                         path->slots[0]++;
3150                         goto next;
3151                 }
3152
3153                 ret = find_first_extent_bit(&rc->processed_blocks,
3154                                             key.objectid, &start, &end,
3155                                             EXTENT_DIRTY, NULL);
3156
3157                 if (ret == 0 && start <= key.objectid) {
3158                         btrfs_release_path(path);
3159                         rc->search_start = end + 1;
3160                 } else {
3161                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3162                                 rc->search_start = key.objectid + key.offset;
3163                         else
3164                                 rc->search_start = key.objectid +
3165                                         fs_info->nodesize;
3166                         memcpy(extent_key, &key, sizeof(key));
3167                         return 0;
3168                 }
3169         }
3170         btrfs_release_path(path);
3171         return ret;
3172 }
3173
3174 static void set_reloc_control(struct reloc_control *rc)
3175 {
3176         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3177
3178         mutex_lock(&fs_info->reloc_mutex);
3179         fs_info->reloc_ctl = rc;
3180         mutex_unlock(&fs_info->reloc_mutex);
3181 }
3182
3183 static void unset_reloc_control(struct reloc_control *rc)
3184 {
3185         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3186
3187         mutex_lock(&fs_info->reloc_mutex);
3188         fs_info->reloc_ctl = NULL;
3189         mutex_unlock(&fs_info->reloc_mutex);
3190 }
3191
3192 static int check_extent_flags(u64 flags)
3193 {
3194         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3195             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3196                 return 1;
3197         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3198             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3199                 return 1;
3200         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3201             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3202                 return 1;
3203         return 0;
3204 }
3205
3206 static noinline_for_stack
3207 int prepare_to_relocate(struct reloc_control *rc)
3208 {
3209         struct btrfs_trans_handle *trans;
3210         int ret;
3211
3212         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3213                                               BTRFS_BLOCK_RSV_TEMP);
3214         if (!rc->block_rsv)
3215                 return -ENOMEM;
3216
3217         memset(&rc->cluster, 0, sizeof(rc->cluster));
3218         rc->search_start = rc->block_group->start;
3219         rc->extents_found = 0;
3220         rc->nodes_relocated = 0;
3221         rc->merging_rsv_size = 0;
3222         rc->reserved_bytes = 0;
3223         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3224                               RELOCATION_RESERVED_NODES;
3225         ret = btrfs_block_rsv_refill(rc->extent_root,
3226                                      rc->block_rsv, rc->block_rsv->size,
3227                                      BTRFS_RESERVE_FLUSH_ALL);
3228         if (ret)
3229                 return ret;
3230
3231         rc->create_reloc_tree = 1;
3232         set_reloc_control(rc);
3233
3234         trans = btrfs_join_transaction(rc->extent_root);
3235         if (IS_ERR(trans)) {
3236                 unset_reloc_control(rc);
3237                 /*
3238                  * extent tree is not a ref_cow tree and has no reloc_root to
3239                  * cleanup.  And callers are responsible to free the above
3240                  * block rsv.
3241                  */
3242                 return PTR_ERR(trans);
3243         }
3244         btrfs_commit_transaction(trans);
3245         return 0;
3246 }
3247
3248 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3249 {
3250         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3251         struct rb_root blocks = RB_ROOT;
3252         struct btrfs_key key;
3253         struct btrfs_trans_handle *trans = NULL;
3254         struct btrfs_path *path;
3255         struct btrfs_extent_item *ei;
3256         u64 flags;
3257         u32 item_size;
3258         int ret;
3259         int err = 0;
3260         int progress = 0;
3261
3262         path = btrfs_alloc_path();
3263         if (!path)
3264                 return -ENOMEM;
3265         path->reada = READA_FORWARD;
3266
3267         ret = prepare_to_relocate(rc);
3268         if (ret) {
3269                 err = ret;
3270                 goto out_free;
3271         }
3272
3273         while (1) {
3274                 rc->reserved_bytes = 0;
3275                 ret = btrfs_block_rsv_refill(rc->extent_root,
3276                                         rc->block_rsv, rc->block_rsv->size,
3277                                         BTRFS_RESERVE_FLUSH_ALL);
3278                 if (ret) {
3279                         err = ret;
3280                         break;
3281                 }
3282                 progress++;
3283                 trans = btrfs_start_transaction(rc->extent_root, 0);
3284                 if (IS_ERR(trans)) {
3285                         err = PTR_ERR(trans);
3286                         trans = NULL;
3287                         break;
3288                 }
3289 restart:
3290                 if (update_backref_cache(trans, &rc->backref_cache)) {
3291                         btrfs_end_transaction(trans);
3292                         trans = NULL;
3293                         continue;
3294                 }
3295
3296                 ret = find_next_extent(rc, path, &key);
3297                 if (ret < 0)
3298                         err = ret;
3299                 if (ret != 0)
3300                         break;
3301
3302                 rc->extents_found++;
3303
3304                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3305                                     struct btrfs_extent_item);
3306                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3307                 if (item_size >= sizeof(*ei)) {
3308                         flags = btrfs_extent_flags(path->nodes[0], ei);
3309                         ret = check_extent_flags(flags);
3310                         BUG_ON(ret);
3311                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3312                         err = -EINVAL;
3313                         btrfs_print_v0_err(trans->fs_info);
3314                         btrfs_abort_transaction(trans, err);
3315                         break;
3316                 } else {
3317                         BUG();
3318                 }
3319
3320                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3321                         ret = add_tree_block(rc, &key, path, &blocks);
3322                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3323                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3324                         ret = add_data_references(rc, &key, path, &blocks);
3325                 } else {
3326                         btrfs_release_path(path);
3327                         ret = 0;
3328                 }
3329                 if (ret < 0) {
3330                         err = ret;
3331                         break;
3332                 }
3333
3334                 if (!RB_EMPTY_ROOT(&blocks)) {
3335                         ret = relocate_tree_blocks(trans, rc, &blocks);
3336                         if (ret < 0) {
3337                                 if (ret != -EAGAIN) {
3338                                         err = ret;
3339                                         break;
3340                                 }
3341                                 rc->extents_found--;
3342                                 rc->search_start = key.objectid;
3343                         }
3344                 }
3345
3346                 btrfs_end_transaction_throttle(trans);
3347                 btrfs_btree_balance_dirty(fs_info);
3348                 trans = NULL;
3349
3350                 if (rc->stage == MOVE_DATA_EXTENTS &&
3351                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3352                         rc->found_file_extent = 1;
3353                         ret = relocate_data_extent(rc->data_inode,
3354                                                    &key, &rc->cluster);
3355                         if (ret < 0) {
3356                                 err = ret;
3357                                 break;
3358                         }
3359                 }
3360                 if (btrfs_should_cancel_balance(fs_info)) {
3361                         err = -ECANCELED;
3362                         break;
3363                 }
3364         }
3365         if (trans && progress && err == -ENOSPC) {
3366                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3367                 if (ret == 1) {
3368                         err = 0;
3369                         progress = 0;
3370                         goto restart;
3371                 }
3372         }
3373
3374         btrfs_release_path(path);
3375         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3376
3377         if (trans) {
3378                 btrfs_end_transaction_throttle(trans);
3379                 btrfs_btree_balance_dirty(fs_info);
3380         }
3381
3382         if (!err) {
3383                 ret = relocate_file_extent_cluster(rc->data_inode,
3384                                                    &rc->cluster);
3385                 if (ret < 0)
3386                         err = ret;
3387         }
3388
3389         rc->create_reloc_tree = 0;
3390         set_reloc_control(rc);
3391
3392         btrfs_backref_release_cache(&rc->backref_cache);
3393         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3394
3395         /*
3396          * Even in the case when the relocation is cancelled, we should all go
3397          * through prepare_to_merge() and merge_reloc_roots().
3398          *
3399          * For error (including cancelled balance), prepare_to_merge() will
3400          * mark all reloc trees orphan, then queue them for cleanup in
3401          * merge_reloc_roots()
3402          */
3403         err = prepare_to_merge(rc, err);
3404
3405         merge_reloc_roots(rc);
3406
3407         rc->merge_reloc_tree = 0;
3408         unset_reloc_control(rc);
3409         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3410
3411         /* get rid of pinned extents */
3412         trans = btrfs_join_transaction(rc->extent_root);
3413         if (IS_ERR(trans)) {
3414                 err = PTR_ERR(trans);
3415                 goto out_free;
3416         }
3417         btrfs_commit_transaction(trans);
3418 out_free:
3419         ret = clean_dirty_subvols(rc);
3420         if (ret < 0 && !err)
3421                 err = ret;
3422         btrfs_free_block_rsv(fs_info, rc->block_rsv);
3423         btrfs_free_path(path);
3424         return err;
3425 }
3426
3427 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3428                                  struct btrfs_root *root, u64 objectid)
3429 {
3430         struct btrfs_path *path;
3431         struct btrfs_inode_item *item;
3432         struct extent_buffer *leaf;
3433         int ret;
3434
3435         path = btrfs_alloc_path();
3436         if (!path)
3437                 return -ENOMEM;
3438
3439         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3440         if (ret)
3441                 goto out;
3442
3443         leaf = path->nodes[0];
3444         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3445         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3446         btrfs_set_inode_generation(leaf, item, 1);
3447         btrfs_set_inode_size(leaf, item, 0);
3448         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3449         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3450                                           BTRFS_INODE_PREALLOC);
3451         btrfs_mark_buffer_dirty(leaf);
3452 out:
3453         btrfs_free_path(path);
3454         return ret;
3455 }
3456
3457 /*
3458  * helper to create inode for data relocation.
3459  * the inode is in data relocation tree and its link count is 0
3460  */
3461 static noinline_for_stack
3462 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3463                                  struct btrfs_block_group *group)
3464 {
3465         struct inode *inode = NULL;
3466         struct btrfs_trans_handle *trans;
3467         struct btrfs_root *root;
3468         u64 objectid;
3469         int err = 0;
3470
3471         root = btrfs_grab_root(fs_info->data_reloc_root);
3472         trans = btrfs_start_transaction(root, 6);
3473         if (IS_ERR(trans)) {
3474                 btrfs_put_root(root);
3475                 return ERR_CAST(trans);
3476         }
3477
3478         err = btrfs_find_free_objectid(root, &objectid);
3479         if (err)
3480                 goto out;
3481
3482         err = __insert_orphan_inode(trans, root, objectid);
3483         BUG_ON(err);
3484
3485         inode = btrfs_iget(fs_info->sb, objectid, root);
3486         BUG_ON(IS_ERR(inode));
3487         BTRFS_I(inode)->index_cnt = group->start;
3488
3489         err = btrfs_orphan_add(trans, BTRFS_I(inode));
3490 out:
3491         btrfs_put_root(root);
3492         btrfs_end_transaction(trans);
3493         btrfs_btree_balance_dirty(fs_info);
3494         if (err) {
3495                 if (inode)
3496                         iput(inode);
3497                 inode = ERR_PTR(err);
3498         }
3499         return inode;
3500 }
3501
3502 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3503 {
3504         struct reloc_control *rc;
3505
3506         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3507         if (!rc)
3508                 return NULL;
3509
3510         INIT_LIST_HEAD(&rc->reloc_roots);
3511         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3512         btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3513         mapping_tree_init(&rc->reloc_root_tree);
3514         extent_io_tree_init(fs_info, &rc->processed_blocks,
3515                             IO_TREE_RELOC_BLOCKS, NULL);
3516         return rc;
3517 }
3518
3519 static void free_reloc_control(struct reloc_control *rc)
3520 {
3521         struct mapping_node *node, *tmp;
3522
3523         free_reloc_roots(&rc->reloc_roots);
3524         rbtree_postorder_for_each_entry_safe(node, tmp,
3525                         &rc->reloc_root_tree.rb_root, rb_node)
3526                 kfree(node);
3527
3528         kfree(rc);
3529 }
3530
3531 /*
3532  * Print the block group being relocated
3533  */
3534 static void describe_relocation(struct btrfs_fs_info *fs_info,
3535                                 struct btrfs_block_group *block_group)
3536 {
3537         char buf[128] = {'\0'};
3538
3539         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3540
3541         btrfs_info(fs_info,
3542                    "relocating block group %llu flags %s",
3543                    block_group->start, buf);
3544 }
3545
3546 static const char *stage_to_string(int stage)
3547 {
3548         if (stage == MOVE_DATA_EXTENTS)
3549                 return "move data extents";
3550         if (stage == UPDATE_DATA_PTRS)
3551                 return "update data pointers";
3552         return "unknown";
3553 }
3554
3555 /*
3556  * function to relocate all extents in a block group.
3557  */
3558 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3559 {
3560         struct btrfs_block_group *bg;
3561         struct btrfs_root *extent_root = fs_info->extent_root;
3562         struct reloc_control *rc;
3563         struct inode *inode;
3564         struct btrfs_path *path;
3565         int ret;
3566         int rw = 0;
3567         int err = 0;
3568
3569         bg = btrfs_lookup_block_group(fs_info, group_start);
3570         if (!bg)
3571                 return -ENOENT;
3572
3573         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
3574                 btrfs_put_block_group(bg);
3575                 return -ETXTBSY;
3576         }
3577
3578         rc = alloc_reloc_control(fs_info);
3579         if (!rc) {
3580                 btrfs_put_block_group(bg);
3581                 return -ENOMEM;
3582         }
3583
3584         rc->extent_root = extent_root;
3585         rc->block_group = bg;
3586
3587         ret = btrfs_inc_block_group_ro(rc->block_group, true);
3588         if (ret) {
3589                 err = ret;
3590                 goto out;
3591         }
3592         rw = 1;
3593
3594         path = btrfs_alloc_path();
3595         if (!path) {
3596                 err = -ENOMEM;
3597                 goto out;
3598         }
3599
3600         inode = lookup_free_space_inode(rc->block_group, path);
3601         btrfs_free_path(path);
3602
3603         if (!IS_ERR(inode))
3604                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
3605         else
3606                 ret = PTR_ERR(inode);
3607
3608         if (ret && ret != -ENOENT) {
3609                 err = ret;
3610                 goto out;
3611         }
3612
3613         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3614         if (IS_ERR(rc->data_inode)) {
3615                 err = PTR_ERR(rc->data_inode);
3616                 rc->data_inode = NULL;
3617                 goto out;
3618         }
3619
3620         describe_relocation(fs_info, rc->block_group);
3621
3622         btrfs_wait_block_group_reservations(rc->block_group);
3623         btrfs_wait_nocow_writers(rc->block_group);
3624         btrfs_wait_ordered_roots(fs_info, U64_MAX,
3625                                  rc->block_group->start,
3626                                  rc->block_group->length);
3627
3628         while (1) {
3629                 int finishes_stage;
3630
3631                 mutex_lock(&fs_info->cleaner_mutex);
3632                 ret = relocate_block_group(rc);
3633                 mutex_unlock(&fs_info->cleaner_mutex);
3634                 if (ret < 0)
3635                         err = ret;
3636
3637                 finishes_stage = rc->stage;
3638                 /*
3639                  * We may have gotten ENOSPC after we already dirtied some
3640                  * extents.  If writeout happens while we're relocating a
3641                  * different block group we could end up hitting the
3642                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
3643                  * btrfs_reloc_cow_block.  Make sure we write everything out
3644                  * properly so we don't trip over this problem, and then break
3645                  * out of the loop if we hit an error.
3646                  */
3647                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
3648                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
3649                                                        (u64)-1);
3650                         if (ret)
3651                                 err = ret;
3652                         invalidate_mapping_pages(rc->data_inode->i_mapping,
3653                                                  0, -1);
3654                         rc->stage = UPDATE_DATA_PTRS;
3655                 }
3656
3657                 if (err < 0)
3658                         goto out;
3659
3660                 if (rc->extents_found == 0)
3661                         break;
3662
3663                 btrfs_info(fs_info, "found %llu extents, stage: %s",
3664                            rc->extents_found, stage_to_string(finishes_stage));
3665         }
3666
3667         WARN_ON(rc->block_group->pinned > 0);
3668         WARN_ON(rc->block_group->reserved > 0);
3669         WARN_ON(rc->block_group->used > 0);
3670 out:
3671         if (err && rw)
3672                 btrfs_dec_block_group_ro(rc->block_group);
3673         iput(rc->data_inode);
3674         btrfs_put_block_group(rc->block_group);
3675         free_reloc_control(rc);
3676         return err;
3677 }
3678
3679 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
3680 {
3681         struct btrfs_fs_info *fs_info = root->fs_info;
3682         struct btrfs_trans_handle *trans;
3683         int ret, err;
3684
3685         trans = btrfs_start_transaction(fs_info->tree_root, 0);
3686         if (IS_ERR(trans))
3687                 return PTR_ERR(trans);
3688
3689         memset(&root->root_item.drop_progress, 0,
3690                 sizeof(root->root_item.drop_progress));
3691         root->root_item.drop_level = 0;
3692         btrfs_set_root_refs(&root->root_item, 0);
3693         ret = btrfs_update_root(trans, fs_info->tree_root,
3694                                 &root->root_key, &root->root_item);
3695
3696         err = btrfs_end_transaction(trans);
3697         if (err)
3698                 return err;
3699         return ret;
3700 }
3701
3702 /*
3703  * recover relocation interrupted by system crash.
3704  *
3705  * this function resumes merging reloc trees with corresponding fs trees.
3706  * this is important for keeping the sharing of tree blocks
3707  */
3708 int btrfs_recover_relocation(struct btrfs_root *root)
3709 {
3710         struct btrfs_fs_info *fs_info = root->fs_info;
3711         LIST_HEAD(reloc_roots);
3712         struct btrfs_key key;
3713         struct btrfs_root *fs_root;
3714         struct btrfs_root *reloc_root;
3715         struct btrfs_path *path;
3716         struct extent_buffer *leaf;
3717         struct reloc_control *rc = NULL;
3718         struct btrfs_trans_handle *trans;
3719         int ret;
3720         int err = 0;
3721
3722         path = btrfs_alloc_path();
3723         if (!path)
3724                 return -ENOMEM;
3725         path->reada = READA_BACK;
3726
3727         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
3728         key.type = BTRFS_ROOT_ITEM_KEY;
3729         key.offset = (u64)-1;
3730
3731         while (1) {
3732                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
3733                                         path, 0, 0);
3734                 if (ret < 0) {
3735                         err = ret;
3736                         goto out;
3737                 }
3738                 if (ret > 0) {
3739                         if (path->slots[0] == 0)
3740                                 break;
3741                         path->slots[0]--;
3742                 }
3743                 leaf = path->nodes[0];
3744                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3745                 btrfs_release_path(path);
3746
3747                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
3748                     key.type != BTRFS_ROOT_ITEM_KEY)
3749                         break;
3750
3751                 reloc_root = btrfs_read_tree_root(root, &key);
3752                 if (IS_ERR(reloc_root)) {
3753                         err = PTR_ERR(reloc_root);
3754                         goto out;
3755                 }
3756
3757                 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
3758                 list_add(&reloc_root->root_list, &reloc_roots);
3759
3760                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
3761                         fs_root = btrfs_get_fs_root(fs_info,
3762                                         reloc_root->root_key.offset, false);
3763                         if (IS_ERR(fs_root)) {
3764                                 ret = PTR_ERR(fs_root);
3765                                 if (ret != -ENOENT) {
3766                                         err = ret;
3767                                         goto out;
3768                                 }
3769                                 ret = mark_garbage_root(reloc_root);
3770                                 if (ret < 0) {
3771                                         err = ret;
3772                                         goto out;
3773                                 }
3774                         } else {
3775                                 btrfs_put_root(fs_root);
3776                         }
3777                 }
3778
3779                 if (key.offset == 0)
3780                         break;
3781
3782                 key.offset--;
3783         }
3784         btrfs_release_path(path);
3785
3786         if (list_empty(&reloc_roots))
3787                 goto out;
3788
3789         rc = alloc_reloc_control(fs_info);
3790         if (!rc) {
3791                 err = -ENOMEM;
3792                 goto out;
3793         }
3794
3795         rc->extent_root = fs_info->extent_root;
3796
3797         set_reloc_control(rc);
3798
3799         trans = btrfs_join_transaction(rc->extent_root);
3800         if (IS_ERR(trans)) {
3801                 err = PTR_ERR(trans);
3802                 goto out_unset;
3803         }
3804
3805         rc->merge_reloc_tree = 1;
3806
3807         while (!list_empty(&reloc_roots)) {
3808                 reloc_root = list_entry(reloc_roots.next,
3809                                         struct btrfs_root, root_list);
3810                 list_del(&reloc_root->root_list);
3811
3812                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
3813                         list_add_tail(&reloc_root->root_list,
3814                                       &rc->reloc_roots);
3815                         continue;
3816                 }
3817
3818                 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
3819                                             false);
3820                 if (IS_ERR(fs_root)) {
3821                         err = PTR_ERR(fs_root);
3822                         list_add_tail(&reloc_root->root_list, &reloc_roots);
3823                         btrfs_end_transaction(trans);
3824                         goto out_unset;
3825                 }
3826
3827                 err = __add_reloc_root(reloc_root);
3828                 BUG_ON(err < 0); /* -ENOMEM or logic error */
3829                 fs_root->reloc_root = btrfs_grab_root(reloc_root);
3830                 btrfs_put_root(fs_root);
3831         }
3832
3833         err = btrfs_commit_transaction(trans);
3834         if (err)
3835                 goto out_unset;
3836
3837         merge_reloc_roots(rc);
3838
3839         unset_reloc_control(rc);
3840
3841         trans = btrfs_join_transaction(rc->extent_root);
3842         if (IS_ERR(trans)) {
3843                 err = PTR_ERR(trans);
3844                 goto out_clean;
3845         }
3846         err = btrfs_commit_transaction(trans);
3847 out_clean:
3848         ret = clean_dirty_subvols(rc);
3849         if (ret < 0 && !err)
3850                 err = ret;
3851 out_unset:
3852         unset_reloc_control(rc);
3853         free_reloc_control(rc);
3854 out:
3855         free_reloc_roots(&reloc_roots);
3856
3857         btrfs_free_path(path);
3858
3859         if (err == 0) {
3860                 /* cleanup orphan inode in data relocation tree */
3861                 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
3862                 ASSERT(fs_root);
3863                 err = btrfs_orphan_cleanup(fs_root);
3864                 btrfs_put_root(fs_root);
3865         }
3866         return err;
3867 }
3868
3869 /*
3870  * helper to add ordered checksum for data relocation.
3871  *
3872  * cloning checksum properly handles the nodatasum extents.
3873  * it also saves CPU time to re-calculate the checksum.
3874  */
3875 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
3876 {
3877         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3878         struct btrfs_ordered_sum *sums;
3879         struct btrfs_ordered_extent *ordered;
3880         int ret;
3881         u64 disk_bytenr;
3882         u64 new_bytenr;
3883         LIST_HEAD(list);
3884
3885         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
3886         BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
3887
3888         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
3889         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
3890                                        disk_bytenr + len - 1, &list, 0);
3891         if (ret)
3892                 goto out;
3893
3894         while (!list_empty(&list)) {
3895                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
3896                 list_del_init(&sums->list);
3897
3898                 /*
3899                  * We need to offset the new_bytenr based on where the csum is.
3900                  * We need to do this because we will read in entire prealloc
3901                  * extents but we may have written to say the middle of the
3902                  * prealloc extent, so we need to make sure the csum goes with
3903                  * the right disk offset.
3904                  *
3905                  * We can do this because the data reloc inode refers strictly
3906                  * to the on disk bytes, so we don't have to worry about
3907                  * disk_len vs real len like with real inodes since it's all
3908                  * disk length.
3909                  */
3910                 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
3911                 sums->bytenr = new_bytenr;
3912
3913                 btrfs_add_ordered_sum(ordered, sums);
3914         }
3915 out:
3916         btrfs_put_ordered_extent(ordered);
3917         return ret;
3918 }
3919
3920 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
3921                           struct btrfs_root *root, struct extent_buffer *buf,
3922                           struct extent_buffer *cow)
3923 {
3924         struct btrfs_fs_info *fs_info = root->fs_info;
3925         struct reloc_control *rc;
3926         struct btrfs_backref_node *node;
3927         int first_cow = 0;
3928         int level;
3929         int ret = 0;
3930
3931         rc = fs_info->reloc_ctl;
3932         if (!rc)
3933                 return 0;
3934
3935         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
3936                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
3937
3938         level = btrfs_header_level(buf);
3939         if (btrfs_header_generation(buf) <=
3940             btrfs_root_last_snapshot(&root->root_item))
3941                 first_cow = 1;
3942
3943         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
3944             rc->create_reloc_tree) {
3945                 WARN_ON(!first_cow && level == 0);
3946
3947                 node = rc->backref_cache.path[level];
3948                 BUG_ON(node->bytenr != buf->start &&
3949                        node->new_bytenr != buf->start);
3950
3951                 btrfs_backref_drop_node_buffer(node);
3952                 atomic_inc(&cow->refs);
3953                 node->eb = cow;
3954                 node->new_bytenr = cow->start;
3955
3956                 if (!node->pending) {
3957                         list_move_tail(&node->list,
3958                                        &rc->backref_cache.pending[level]);
3959                         node->pending = 1;
3960                 }
3961
3962                 if (first_cow)
3963                         mark_block_processed(rc, node);
3964
3965                 if (first_cow && level > 0)
3966                         rc->nodes_relocated += buf->len;
3967         }
3968
3969         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
3970                 ret = replace_file_extents(trans, rc, root, cow);
3971         return ret;
3972 }
3973
3974 /*
3975  * called before creating snapshot. it calculates metadata reservation
3976  * required for relocating tree blocks in the snapshot
3977  */
3978 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
3979                               u64 *bytes_to_reserve)
3980 {
3981         struct btrfs_root *root = pending->root;
3982         struct reloc_control *rc = root->fs_info->reloc_ctl;
3983
3984         if (!rc || !have_reloc_root(root))
3985                 return;
3986
3987         if (!rc->merge_reloc_tree)
3988                 return;
3989
3990         root = root->reloc_root;
3991         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
3992         /*
3993          * relocation is in the stage of merging trees. the space
3994          * used by merging a reloc tree is twice the size of
3995          * relocated tree nodes in the worst case. half for cowing
3996          * the reloc tree, half for cowing the fs tree. the space
3997          * used by cowing the reloc tree will be freed after the
3998          * tree is dropped. if we create snapshot, cowing the fs
3999          * tree may use more space than it frees. so we need
4000          * reserve extra space.
4001          */
4002         *bytes_to_reserve += rc->nodes_relocated;
4003 }
4004
4005 /*
4006  * called after snapshot is created. migrate block reservation
4007  * and create reloc root for the newly created snapshot
4008  *
4009  * This is similar to btrfs_init_reloc_root(), we come out of here with two
4010  * references held on the reloc_root, one for root->reloc_root and one for
4011  * rc->reloc_roots.
4012  */
4013 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4014                                struct btrfs_pending_snapshot *pending)
4015 {
4016         struct btrfs_root *root = pending->root;
4017         struct btrfs_root *reloc_root;
4018         struct btrfs_root *new_root;
4019         struct reloc_control *rc = root->fs_info->reloc_ctl;
4020         int ret;
4021
4022         if (!rc || !have_reloc_root(root))
4023                 return 0;
4024
4025         rc = root->fs_info->reloc_ctl;
4026         rc->merging_rsv_size += rc->nodes_relocated;
4027
4028         if (rc->merge_reloc_tree) {
4029                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4030                                               rc->block_rsv,
4031                                               rc->nodes_relocated, true);
4032                 if (ret)
4033                         return ret;
4034         }
4035
4036         new_root = pending->snap;
4037         reloc_root = create_reloc_root(trans, root->reloc_root,
4038                                        new_root->root_key.objectid);
4039         if (IS_ERR(reloc_root))
4040                 return PTR_ERR(reloc_root);
4041
4042         ret = __add_reloc_root(reloc_root);
4043         BUG_ON(ret < 0);
4044         new_root->reloc_root = btrfs_grab_root(reloc_root);
4045
4046         if (rc->create_reloc_tree)
4047                 ret = clone_backref_node(trans, rc, root, reloc_root);
4048         return ret;
4049 }