fs/btrfs/relocation.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2009 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include "ctree.h"
  13 #include "disk-io.h"
  14 #include "transaction.h"
  15 #include "volumes.h"
  16 #include "locking.h"
  17 #include "btrfs_inode.h"
  18 #include "async-thread.h"
  19 #include "free-space-cache.h"
  20 #include "inode-map.h"
  21 #include "qgroup.h"
  22 #include "print-tree.h"
  23 #include "delalloc-space.h"
  24 #include "block-group.h"
  25
  26 /*
  27  * backref_node, mapping_node and tree_block start with this
  28  */
  29 struct tree_entry {
  30         struct rb_node rb_node;
  31         u64 bytenr;
  32 };
  33
  34 /*
  35  * present a tree block in the backref cache
  36  */
  37 struct backref_node {
  38         struct rb_node rb_node;
  39         u64 bytenr;
  40
  41         u64 new_bytenr;
  42         /* objectid of tree block owner, can be not uptodate */
  43         u64 owner;
  44         /* link to pending, changed or detached list */
  45         struct list_head list;
  46         /* list of upper level blocks reference this block */
  47         struct list_head upper;
  48         /* list of child blocks in the cache */
  49         struct list_head lower;
  50         /* NULL if this node is not tree root */
  51         struct btrfs_root *root;
  52         /* extent buffer got by COW the block */
  53         struct extent_buffer *eb;
  54         /* level of tree block */
  55         unsigned int level:8;
  56         /* is the block in non-reference counted tree */
  57         unsigned int cowonly:1;
  58         /* 1 if no child node in the cache */
  59         unsigned int lowest:1;
  60         /* is the extent buffer locked */
  61         unsigned int locked:1;
  62         /* has the block been processed */
  63         unsigned int processed:1;
  64         /* have backrefs of this block been checked */
  65         unsigned int checked:1;
  66         /*
  67          * 1 if corresponding block has been cowed but some upper
  68          * level block pointers may not point to the new location
  69          */
  70         unsigned int pending:1;
  71         /*
  72          * 1 if the backref node isn't connected to any other
  73          * backref node.
  74          */
  75         unsigned int detached:1;
  76 };
  77
  78 /*
  79  * present a block pointer in the backref cache
  80  */
  81 struct backref_edge {
  82         struct list_head list[2];
  83         struct backref_node *node[2];
  84 };
  85
  86 #define LOWER   0
  87 #define UPPER   1
  88 #define RELOCATION_RESERVED_NODES       256
  89
  90 struct backref_cache {
  91         /* red black tree of all backref nodes in the cache */
  92         struct rb_root rb_root;
  93         /* for passing backref nodes to btrfs_reloc_cow_block */
  94         struct backref_node *path[BTRFS_MAX_LEVEL];
  95         /*
  96          * list of blocks that have been cowed but some block
  97          * pointers in upper level blocks may not reflect the
  98          * new location
  99          */
 100         struct list_head pending[BTRFS_MAX_LEVEL];
 101         /* list of backref nodes with no child node */
 102         struct list_head leaves;
 103         /* list of blocks that have been cowed in current transaction */
 104         struct list_head changed;
 105         /* list of detached backref node. */
 106         struct list_head detached;
 107
 108         u64 last_trans;
 109
 110         int nr_nodes;
 111         int nr_edges;
 112 };
 113
 114 /*
 115  * map address of tree root to tree
 116  */
 117 struct mapping_node {
 118         struct rb_node rb_node;
 119         u64 bytenr;
 120         void *data;
 121 };
 122
 123 struct mapping_tree {
 124         struct rb_root rb_root;
 125         spinlock_t lock;
 126 };
 127
 128 /*
 129  * present a tree block to process
 130  */
 131 struct tree_block {
 132         struct rb_node rb_node;
 133         u64 bytenr;
 134         struct btrfs_key key;
 135         unsigned int level:8;
 136         unsigned int key_ready:1;
 137 };
 138
 139 #define MAX_EXTENTS 128
 140
 141 struct file_extent_cluster {
 142         u64 start;
 143         u64 end;
 144         u64 boundary[MAX_EXTENTS];
 145         unsigned int nr;
 146 };
 147
 148 struct reloc_control {
 149         /* block group to relocate */
 150         struct btrfs_block_group_cache *block_group;
 151         /* extent tree */
 152         struct btrfs_root *extent_root;
 153         /* inode for moving data */
 154         struct inode *data_inode;
 155
 156         struct btrfs_block_rsv *block_rsv;
 157
 158         struct backref_cache backref_cache;
 159
 160         struct file_extent_cluster cluster;
 161         /* tree blocks have been processed */
 162         struct extent_io_tree processed_blocks;
 163         /* map start of tree root to corresponding reloc tree */
 164         struct mapping_tree reloc_root_tree;
 165         /* list of reloc trees */
 166         struct list_head reloc_roots;
 167         /* list of subvolume trees that get relocated */
 168         struct list_head dirty_subvol_roots;
 169         /* size of metadata reservation for merging reloc trees */
 170         u64 merging_rsv_size;
 171         /* size of relocated tree nodes */
 172         u64 nodes_relocated;
 173         /* reserved size for block group relocation*/
 174         u64 reserved_bytes;
 175
 176         u64 search_start;
 177         u64 extents_found;
 178
 179         unsigned int stage:8;
 180         unsigned int create_reloc_tree:1;
 181         unsigned int merge_reloc_tree:1;
 182         unsigned int found_file_extent:1;
 183 };
 184
 185 /* stages of data relocation */
 186 #define MOVE_DATA_EXTENTS       0
 187 #define UPDATE_DATA_PTRS        1
 188
 189 static void remove_backref_node(struct backref_cache *cache,
 190                                 struct backref_node *node);
 191 static void __mark_block_processed(struct reloc_control *rc,
 192                                    struct backref_node *node);
 193
 194 static void mapping_tree_init(struct mapping_tree *tree)
 195 {
 196         tree->rb_root = RB_ROOT;
 197         spin_lock_init(&tree->lock);
 198 }
 199
 200 static void backref_cache_init(struct backref_cache *cache)
 201 {
 202         int i;
 203         cache->rb_root = RB_ROOT;
 204         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 205                 INIT_LIST_HEAD(&cache->pending[i]);
 206         INIT_LIST_HEAD(&cache->changed);
 207         INIT_LIST_HEAD(&cache->detached);
 208         INIT_LIST_HEAD(&cache->leaves);
 209 }
 210
 211 static void backref_cache_cleanup(struct backref_cache *cache)
 212 {
 213         struct backref_node *node;
 214         int i;
 215
 216         while (!list_empty(&cache->detached)) {
 217                 node = list_entry(cache->detached.next,
 218                                   struct backref_node, list);
 219                 remove_backref_node(cache, node);
 220         }
 221
 222         while (!list_empty(&cache->leaves)) {
 223                 node = list_entry(cache->leaves.next,
 224                                   struct backref_node, lower);
 225                 remove_backref_node(cache, node);
 226         }
 227
 228         cache->last_trans = 0;
 229
 230         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 231                 ASSERT(list_empty(&cache->pending[i]));
 232         ASSERT(list_empty(&cache->changed));
 233         ASSERT(list_empty(&cache->detached));
 234         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
 235         ASSERT(!cache->nr_nodes);
 236         ASSERT(!cache->nr_edges);
 237 }
 238
 239 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 240 {
 241         struct backref_node *node;
 242
 243         node = kzalloc(sizeof(*node), GFP_NOFS);
 244         if (node) {
 245                 INIT_LIST_HEAD(&node->list);
 246                 INIT_LIST_HEAD(&node->upper);
 247                 INIT_LIST_HEAD(&node->lower);
 248                 RB_CLEAR_NODE(&node->rb_node);
 249                 cache->nr_nodes++;
 250         }
 251         return node;
 252 }
 253
 254 static void free_backref_node(struct backref_cache *cache,
 255                               struct backref_node *node)
 256 {
 257         if (node) {
 258                 cache->nr_nodes--;
 259                 kfree(node);
 260         }
 261 }
 262
 263 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 264 {
 265         struct backref_edge *edge;
 266
 267         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 268         if (edge)
 269                 cache->nr_edges++;
 270         return edge;
 271 }
 272
 273 static void free_backref_edge(struct backref_cache *cache,
 274                               struct backref_edge *edge)
 275 {
 276         if (edge) {
 277                 cache->nr_edges--;
 278                 kfree(edge);
 279         }
 280 }
 281
 282 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 283                                    struct rb_node *node)
 284 {
 285         struct rb_node **p = &root->rb_node;
 286         struct rb_node *parent = NULL;
 287         struct tree_entry *entry;
 288
 289         while (*p) {
 290                 parent = *p;
 291                 entry = rb_entry(parent, struct tree_entry, rb_node);
 292
 293                 if (bytenr < entry->bytenr)
 294                         p = &(*p)->rb_left;
 295                 else if (bytenr > entry->bytenr)
 296                         p = &(*p)->rb_right;
 297                 else
 298                         return parent;
 299         }
 300
 301         rb_link_node(node, parent, p);
 302         rb_insert_color(node, root);
 303         return NULL;
 304 }
 305
 306 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 307 {
 308         struct rb_node *n = root->rb_node;
 309         struct tree_entry *entry;
 310
 311         while (n) {
 312                 entry = rb_entry(n, struct tree_entry, rb_node);
 313
 314                 if (bytenr < entry->bytenr)
 315                         n = n->rb_left;
 316                 else if (bytenr > entry->bytenr)
 317                         n = n->rb_right;
 318                 else
 319                         return n;
 320         }
 321         return NULL;
 322 }
 323
 324 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 325 {
 326
 327         struct btrfs_fs_info *fs_info = NULL;
 328         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 329                                               rb_node);
 330         if (bnode->root)
 331                 fs_info = bnode->root->fs_info;
 332         btrfs_panic(fs_info, errno,
 333                     "Inconsistency in backref cache found at offset %llu",
 334                     bytenr);
 335 }
 336
 337 /*
 338  * walk up backref nodes until reach node presents tree root
 339  */
 340 static struct backref_node *walk_up_backref(struct backref_node *node,
 341                                             struct backref_edge *edges[],
 342                                             int *index)
 343 {
 344         struct backref_edge *edge;
 345         int idx = *index;
 346
 347         while (!list_empty(&node->upper)) {
 348                 edge = list_entry(node->upper.next,
 349                                   struct backref_edge, list[LOWER]);
 350                 edges[idx++] = edge;
 351                 node = edge->node[UPPER];
 352         }
 353         BUG_ON(node->detached);
 354         *index = idx;
 355         return node;
 356 }
 357
 358 /*
 359  * walk down backref nodes to find start of next reference path
 360  */
 361 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 362                                               int *index)
 363 {
 364         struct backref_edge *edge;
 365         struct backref_node *lower;
 366         int idx = *index;
 367
 368         while (idx > 0) {
 369                 edge = edges[idx - 1];
 370                 lower = edge->node[LOWER];
 371                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 372                         idx--;
 373                         continue;
 374                 }
 375                 edge = list_entry(edge->list[LOWER].next,
 376                                   struct backref_edge, list[LOWER]);
 377                 edges[idx - 1] = edge;
 378                 *index = idx;
 379                 return edge->node[UPPER];
 380         }
 381         *index = 0;
 382         return NULL;
 383 }
 384
 385 static void unlock_node_buffer(struct backref_node *node)
 386 {
 387         if (node->locked) {
 388                 btrfs_tree_unlock(node->eb);
 389                 node->locked = 0;
 390         }
 391 }
 392
 393 static void drop_node_buffer(struct backref_node *node)
 394 {
 395         if (node->eb) {
 396                 unlock_node_buffer(node);
 397                 free_extent_buffer(node->eb);
 398                 node->eb = NULL;
 399         }
 400 }
 401
 402 static void drop_backref_node(struct backref_cache *tree,
 403                               struct backref_node *node)
 404 {
 405         BUG_ON(!list_empty(&node->upper));
 406
 407         drop_node_buffer(node);
 408         list_del(&node->list);
 409         list_del(&node->lower);
 410         if (!RB_EMPTY_NODE(&node->rb_node))
 411                 rb_erase(&node->rb_node, &tree->rb_root);
 412         free_backref_node(tree, node);
 413 }
 414
 415 /*
 416  * remove a backref node from the backref cache
 417  */
 418 static void remove_backref_node(struct backref_cache *cache,
 419                                 struct backref_node *node)
 420 {
 421         struct backref_node *upper;
 422         struct backref_edge *edge;
 423
 424         if (!node)
 425                 return;
 426
 427         BUG_ON(!node->lowest && !node->detached);
 428         while (!list_empty(&node->upper)) {
 429                 edge = list_entry(node->upper.next, struct backref_edge,
 430                                   list[LOWER]);
 431                 upper = edge->node[UPPER];
 432                 list_del(&edge->list[LOWER]);
 433                 list_del(&edge->list[UPPER]);
 434                 free_backref_edge(cache, edge);
 435
 436                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 437                         BUG_ON(!list_empty(&node->upper));
 438                         drop_backref_node(cache, node);
 439                         node = upper;
 440                         node->lowest = 1;
 441                         continue;
 442                 }
 443                 /*
 444                  * add the node to leaf node list if no other
 445                  * child block cached.
 446                  */
 447                 if (list_empty(&upper->lower)) {
 448                         list_add_tail(&upper->lower, &cache->leaves);
 449                         upper->lowest = 1;
 450                 }
 451         }
 452
 453         drop_backref_node(cache, node);
 454 }
 455
 456 static void update_backref_node(struct backref_cache *cache,
 457                                 struct backref_node *node, u64 bytenr)
 458 {
 459         struct rb_node *rb_node;
 460         rb_erase(&node->rb_node, &cache->rb_root);
 461         node->bytenr = bytenr;
 462         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 463         if (rb_node)
 464                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 465 }
 466
 467 /*
 468  * update backref cache after a transaction commit
 469  */
 470 static int update_backref_cache(struct btrfs_trans_handle *trans,
 471                                 struct backref_cache *cache)
 472 {
 473         struct backref_node *node;
 474         int level = 0;
 475
 476         if (cache->last_trans == 0) {
 477                 cache->last_trans = trans->transid;
 478                 return 0;
 479         }
 480
 481         if (cache->last_trans == trans->transid)
 482                 return 0;
 483
 484         /*
 485          * detached nodes are used to avoid unnecessary backref
 486          * lookup. transaction commit changes the extent tree.
 487          * so the detached nodes are no longer useful.
 488          */
 489         while (!list_empty(&cache->detached)) {
 490                 node = list_entry(cache->detached.next,
 491                                   struct backref_node, list);
 492                 remove_backref_node(cache, node);
 493         }
 494
 495         while (!list_empty(&cache->changed)) {
 496                 node = list_entry(cache->changed.next,
 497                                   struct backref_node, list);
 498                 list_del_init(&node->list);
 499                 BUG_ON(node->pending);
 500                 update_backref_node(cache, node, node->new_bytenr);
 501         }
 502
 503         /*
 504          * some nodes can be left in the pending list if there were
 505          * errors during processing the pending nodes.
 506          */
 507         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 508                 list_for_each_entry(node, &cache->pending[level], list) {
 509                         BUG_ON(!node->pending);
 510                         if (node->bytenr == node->new_bytenr)
 511                                 continue;
 512                         update_backref_node(cache, node, node->new_bytenr);
 513                 }
 514         }
 515
 516         cache->last_trans = 0;
 517         return 1;
 518 }
 519
 520
 521 static int should_ignore_root(struct btrfs_root *root)
 522 {
 523         struct btrfs_root *reloc_root;
 524
 525         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 526                 return 0;
 527
 528         reloc_root = root->reloc_root;
 529         if (!reloc_root)
 530                 return 0;
 531
 532         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 533             root->fs_info->running_transaction->transid - 1)
 534                 return 0;
 535         /*
 536          * if there is reloc tree and it was created in previous
 537          * transaction backref lookup can find the reloc tree,
 538          * so backref node for the fs tree root is useless for
 539          * relocation.
 540          */
 541         return 1;
 542 }
 543 /*
 544  * find reloc tree by address of tree root
 545  */
 546 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 547                                           u64 bytenr)
 548 {
 549         struct rb_node *rb_node;
 550         struct mapping_node *node;
 551         struct btrfs_root *root = NULL;
 552
 553         spin_lock(&rc->reloc_root_tree.lock);
 554         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 555         if (rb_node) {
 556                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 557                 root = (struct btrfs_root *)node->data;
 558         }
 559         spin_unlock(&rc->reloc_root_tree.lock);
 560         return root;
 561 }
 562
 563 static int is_cowonly_root(u64 root_objectid)
 564 {
 565         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 566             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 567             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 568             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 569             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 570             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 571             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 572             root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
 573             root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
 574                 return 1;
 575         return 0;
 576 }
 577
 578 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 579                                         u64 root_objectid)
 580 {
 581         struct btrfs_key key;
 582
 583         key.objectid = root_objectid;
 584         key.type = BTRFS_ROOT_ITEM_KEY;
 585         if (is_cowonly_root(root_objectid))
 586                 key.offset = 0;
 587         else
 588                 key.offset = (u64)-1;
 589
 590         return btrfs_get_fs_root(fs_info, &key, false);
 591 }
 592
 593 static noinline_for_stack
 594 int find_inline_backref(struct extent_buffer *leaf, int slot,
 595                         unsigned long *ptr, unsigned long *end)
 596 {
 597         struct btrfs_key key;
 598         struct btrfs_extent_item *ei;
 599         struct btrfs_tree_block_info *bi;
 600         u32 item_size;
 601
 602         btrfs_item_key_to_cpu(leaf, &key, slot);
 603
 604         item_size = btrfs_item_size_nr(leaf, slot);
 605         if (item_size < sizeof(*ei)) {
 606                 btrfs_print_v0_err(leaf->fs_info);
 607                 btrfs_handle_fs_error(leaf->fs_info, -EINVAL, NULL);
 608                 return 1;
 609         }
 610         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 611         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 612                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 613
 614         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 615             item_size <= sizeof(*ei) + sizeof(*bi)) {
 616                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 617                 return 1;
 618         }
 619         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 620             item_size <= sizeof(*ei)) {
 621                 WARN_ON(item_size < sizeof(*ei));
 622                 return 1;
 623         }
 624
 625         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 626                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 627                 *ptr = (unsigned long)(bi + 1);
 628         } else {
 629                 *ptr = (unsigned long)(ei + 1);
 630         }
 631         *end = (unsigned long)ei + item_size;
 632         return 0;
 633 }
 634
 635 /*
 636  * build backref tree for a given tree block. root of the backref tree
 637  * corresponds the tree block, leaves of the backref tree correspond
 638  * roots of b-trees that reference the tree block.
 639  *
 640  * the basic idea of this function is check backrefs of a given block
 641  * to find upper level blocks that reference the block, and then check
 642  * backrefs of these upper level blocks recursively. the recursion stop
 643  * when tree root is reached or backrefs for the block is cached.
 644  *
 645  * NOTE: if we find backrefs for a block are cached, we know backrefs
 646  * for all upper level blocks that directly/indirectly reference the
 647  * block are also cached.
 648  */
 649 static noinline_for_stack
 650 struct backref_node *build_backref_tree(struct reloc_control *rc,
 651                                         struct btrfs_key *node_key,
 652                                         int level, u64 bytenr)
 653 {
 654         struct backref_cache *cache = &rc->backref_cache;
 655         struct btrfs_path *path1; /* For searching extent root */
 656         struct btrfs_path *path2; /* For searching parent of TREE_BLOCK_REF */
 657         struct extent_buffer *eb;
 658         struct btrfs_root *root;
 659         struct backref_node *cur;
 660         struct backref_node *upper;
 661         struct backref_node *lower;
 662         struct backref_node *node = NULL;
 663         struct backref_node *exist = NULL;
 664         struct backref_edge *edge;
 665         struct rb_node *rb_node;
 666         struct btrfs_key key;
 667         unsigned long end;
 668         unsigned long ptr;
 669         LIST_HEAD(list); /* Pending edge list, upper node needs to be checked */
 670         LIST_HEAD(useless);
 671         int cowonly;
 672         int ret;
 673         int err = 0;
 674         bool need_check = true;
 675
 676         path1 = btrfs_alloc_path();
 677         path2 = btrfs_alloc_path();
 678         if (!path1 || !path2) {
 679                 err = -ENOMEM;
 680                 goto out;
 681         }
 682         path1->reada = READA_FORWARD;
 683         path2->reada = READA_FORWARD;
 684
 685         node = alloc_backref_node(cache);
 686         if (!node) {
 687                 err = -ENOMEM;
 688                 goto out;
 689         }
 690
 691         node->bytenr = bytenr;
 692         node->level = level;
 693         node->lowest = 1;
 694         cur = node;
 695 again:
 696         end = 0;
 697         ptr = 0;
 698         key.objectid = cur->bytenr;
 699         key.type = BTRFS_METADATA_ITEM_KEY;
 700         key.offset = (u64)-1;
 701
 702         path1->search_commit_root = 1;
 703         path1->skip_locking = 1;
 704         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 705                                 0, 0);
 706         if (ret < 0) {
 707                 err = ret;
 708                 goto out;
 709         }
 710         ASSERT(ret);
 711         ASSERT(path1->slots[0]);
 712
 713         path1->slots[0]--;
 714
 715         WARN_ON(cur->checked);
 716         if (!list_empty(&cur->upper)) {
 717                 /*
 718                  * the backref was added previously when processing
 719                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 720                  */
 721                 ASSERT(list_is_singular(&cur->upper));
 722                 edge = list_entry(cur->upper.next, struct backref_edge,
 723                                   list[LOWER]);
 724                 ASSERT(list_empty(&edge->list[UPPER]));
 725                 exist = edge->node[UPPER];
 726                 /*
 727                  * add the upper level block to pending list if we need
 728                  * check its backrefs
 729                  */
 730                 if (!exist->checked)
 731                         list_add_tail(&edge->list[UPPER], &list);
 732         } else {
 733                 exist = NULL;
 734         }
 735
 736         while (1) {
 737                 cond_resched();
 738                 eb = path1->nodes[0];
 739
 740                 if (ptr >= end) {
 741                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 742                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 743                                 if (ret < 0) {
 744                                         err = ret;
 745                                         goto out;
 746                                 }
 747                                 if (ret > 0)
 748                                         break;
 749                                 eb = path1->nodes[0];
 750                         }
 751
 752                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 753                         if (key.objectid != cur->bytenr) {
 754                                 WARN_ON(exist);
 755                                 break;
 756                         }
 757
 758                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 759                             key.type == BTRFS_METADATA_ITEM_KEY) {
 760                                 ret = find_inline_backref(eb, path1->slots[0],
 761                                                           &ptr, &end);
 762                                 if (ret)
 763                                         goto next;
 764                         }
 765                 }
 766
 767                 if (ptr < end) {
 768                         /* update key for inline back ref */
 769                         struct btrfs_extent_inline_ref *iref;
 770                         int type;
 771                         iref = (struct btrfs_extent_inline_ref *)ptr;
 772                         type = btrfs_get_extent_inline_ref_type(eb, iref,
 773                                                         BTRFS_REF_TYPE_BLOCK);
 774                         if (type == BTRFS_REF_TYPE_INVALID) {
 775                                 err = -EUCLEAN;
 776                                 goto out;
 777                         }
 778                         key.type = type;
 779                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 780
 781                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 782                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 783                 }
 784
 785                 /*
 786                  * Parent node found and matches current inline ref, no need to
 787                  * rebuild this node for this inline ref.
 788                  */
 789                 if (exist &&
 790                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 791                       exist->owner == key.offset) ||
 792                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 793                       exist->bytenr == key.offset))) {
 794                         exist = NULL;
 795                         goto next;
 796                 }
 797
 798                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
 799                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 800                         if (key.objectid == key.offset) {
 801                                 /*
 802                                  * Only root blocks of reloc trees use backref
 803                                  * pointing to itself.
 804                                  */
 805                                 root = find_reloc_root(rc, cur->bytenr);
 806                                 ASSERT(root);
 807                                 cur->root = root;
 808                                 break;
 809                         }
 810
 811                         edge = alloc_backref_edge(cache);
 812                         if (!edge) {
 813                                 err = -ENOMEM;
 814                                 goto out;
 815                         }
 816                         rb_node = tree_search(&cache->rb_root, key.offset);
 817                         if (!rb_node) {
 818                                 upper = alloc_backref_node(cache);
 819                                 if (!upper) {
 820                                         free_backref_edge(cache, edge);
 821                                         err = -ENOMEM;
 822                                         goto out;
 823                                 }
 824                                 upper->bytenr = key.offset;
 825                                 upper->level = cur->level + 1;
 826                                 /*
 827                                  *  backrefs for the upper level block isn't
 828                                  *  cached, add the block to pending list
 829                                  */
 830                                 list_add_tail(&edge->list[UPPER], &list);
 831                         } else {
 832                                 upper = rb_entry(rb_node, struct backref_node,
 833                                                  rb_node);
 834                                 ASSERT(upper->checked);
 835                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 836                         }
 837                         list_add_tail(&edge->list[LOWER], &cur->upper);
 838                         edge->node[LOWER] = cur;
 839                         edge->node[UPPER] = upper;
 840
 841                         goto next;
 842                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 843                         err = -EINVAL;
 844                         btrfs_print_v0_err(rc->extent_root->fs_info);
 845                         btrfs_handle_fs_error(rc->extent_root->fs_info, err,
 846                                               NULL);
 847                         goto out;
 848                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 849                         goto next;
 850                 }
 851
 852                 /*
 853                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
 854                  * means the root objectid. We need to search the tree to get
 855                  * its parent bytenr.
 856                  */
 857                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 858                 if (IS_ERR(root)) {
 859                         err = PTR_ERR(root);
 860                         goto out;
 861                 }
 862
 863                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 864                         cur->cowonly = 1;
 865
 866                 if (btrfs_root_level(&root->root_item) == cur->level) {
 867                         /* tree root */
 868                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
 869                                cur->bytenr);
 870                         if (should_ignore_root(root))
 871                                 list_add(&cur->list, &useless);
 872                         else
 873                                 cur->root = root;
 874                         break;
 875                 }
 876
 877                 level = cur->level + 1;
 878
 879                 /* Search the tree to find parent blocks referring the block. */
 880                 path2->search_commit_root = 1;
 881                 path2->skip_locking = 1;
 882                 path2->lowest_level = level;
 883                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 884                 path2->lowest_level = 0;
 885                 if (ret < 0) {
 886                         err = ret;
 887                         goto out;
 888                 }
 889                 if (ret > 0 && path2->slots[level] > 0)
 890                         path2->slots[level]--;
 891
 892                 eb = path2->nodes[level];
 893                 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
 894                     cur->bytenr) {
 895                         btrfs_err(root->fs_info,
 896         "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
 897                                   cur->bytenr, level - 1,
 898                                   root->root_key.objectid,
 899                                   node_key->objectid, node_key->type,
 900                                   node_key->offset);
 901                         err = -ENOENT;
 902                         goto out;
 903                 }
 904                 lower = cur;
 905                 need_check = true;
 906
 907                 /* Add all nodes and edges in the path */
 908                 for (; level < BTRFS_MAX_LEVEL; level++) {
 909                         if (!path2->nodes[level]) {
 910                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
 911                                        lower->bytenr);
 912                                 if (should_ignore_root(root))
 913                                         list_add(&lower->list, &useless);
 914                                 else
 915                                         lower->root = root;
 916                                 break;
 917                         }
 918
 919                         edge = alloc_backref_edge(cache);
 920                         if (!edge) {
 921                                 err = -ENOMEM;
 922                                 goto out;
 923                         }
 924
 925                         eb = path2->nodes[level];
 926                         rb_node = tree_search(&cache->rb_root, eb->start);
 927                         if (!rb_node) {
 928                                 upper = alloc_backref_node(cache);
 929                                 if (!upper) {
 930                                         free_backref_edge(cache, edge);
 931                                         err = -ENOMEM;
 932                                         goto out;
 933                                 }
 934                                 upper->bytenr = eb->start;
 935                                 upper->owner = btrfs_header_owner(eb);
 936                                 upper->level = lower->level + 1;
 937                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
 938                                               &root->state))
 939                                         upper->cowonly = 1;
 940
 941                                 /*
 942                                  * if we know the block isn't shared
 943                                  * we can void checking its backrefs.
 944                                  */
 945                                 if (btrfs_block_can_be_shared(root, eb))
 946                                         upper->checked = 0;
 947                                 else
 948                                         upper->checked = 1;
 949
 950                                 /*
 951                                  * add the block to pending list if we
 952                                  * need check its backrefs, we only do this once
 953                                  * while walking up a tree as we will catch
 954                                  * anything else later on.
 955                                  */
 956                                 if (!upper->checked && need_check) {
 957                                         need_check = false;
 958                                         list_add_tail(&edge->list[UPPER],
 959                                                       &list);
 960                                 } else {
 961                                         if (upper->checked)
 962                                                 need_check = true;
 963                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 964                                 }
 965                         } else {
 966                                 upper = rb_entry(rb_node, struct backref_node,
 967                                                  rb_node);
 968                                 ASSERT(upper->checked);
 969                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 970                                 if (!upper->owner)
 971                                         upper->owner = btrfs_header_owner(eb);
 972                         }
 973                         list_add_tail(&edge->list[LOWER], &lower->upper);
 974                         edge->node[LOWER] = lower;
 975                         edge->node[UPPER] = upper;
 976
 977                         if (rb_node)
 978                                 break;
 979                         lower = upper;
 980                         upper = NULL;
 981                 }
 982                 btrfs_release_path(path2);
 983 next:
 984                 if (ptr < end) {
 985                         ptr += btrfs_extent_inline_ref_size(key.type);
 986                         if (ptr >= end) {
 987                                 WARN_ON(ptr > end);
 988                                 ptr = 0;
 989                                 end = 0;
 990                         }
 991                 }
 992                 if (ptr >= end)
 993                         path1->slots[0]++;
 994         }
 995         btrfs_release_path(path1);
 996
 997         cur->checked = 1;
 998         WARN_ON(exist);
 999
1000         /* the pending list isn't empty, take the first block to process */
1001         if (!list_empty(&list)) {
1002                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1003                 list_del_init(&edge->list[UPPER]);
1004                 cur = edge->node[UPPER];
1005                 goto again;
1006         }
1007
1008         /*
1009          * everything goes well, connect backref nodes and insert backref nodes
1010          * into the cache.
1011          */
1012         ASSERT(node->checked);
1013         cowonly = node->cowonly;
1014         if (!cowonly) {
1015                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1016                                       &node->rb_node);
1017                 if (rb_node)
1018                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1019                 list_add_tail(&node->lower, &cache->leaves);
1020         }
1021
1022         list_for_each_entry(edge, &node->upper, list[LOWER])
1023                 list_add_tail(&edge->list[UPPER], &list);
1024
1025         while (!list_empty(&list)) {
1026                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1027                 list_del_init(&edge->list[UPPER]);
1028                 upper = edge->node[UPPER];
1029                 if (upper->detached) {
1030                         list_del(&edge->list[LOWER]);
1031                         lower = edge->node[LOWER];
1032                         free_backref_edge(cache, edge);
1033                         if (list_empty(&lower->upper))
1034                                 list_add(&lower->list, &useless);
1035                         continue;
1036                 }
1037
1038                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1039                         if (upper->lowest) {
1040                                 list_del_init(&upper->lower);
1041                                 upper->lowest = 0;
1042                         }
1043
1044                         list_add_tail(&edge->list[UPPER], &upper->lower);
1045                         continue;
1046                 }
1047
1048                 if (!upper->checked) {
1049                         /*
1050                          * Still want to blow up for developers since this is a
1051                          * logic bug.
1052                          */
1053                         ASSERT(0);
1054                         err = -EINVAL;
1055                         goto out;
1056                 }
1057                 if (cowonly != upper->cowonly) {
1058                         ASSERT(0);
1059                         err = -EINVAL;
1060                         goto out;
1061                 }
1062
1063                 if (!cowonly) {
1064                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1065                                               &upper->rb_node);
1066                         if (rb_node)
1067                                 backref_tree_panic(rb_node, -EEXIST,
1068                                                    upper->bytenr);
1069                 }
1070
1071                 list_add_tail(&edge->list[UPPER], &upper->lower);
1072
1073                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1074                         list_add_tail(&edge->list[UPPER], &list);
1075         }
1076         /*
1077          * process useless backref nodes. backref nodes for tree leaves
1078          * are deleted from the cache. backref nodes for upper level
1079          * tree blocks are left in the cache to avoid unnecessary backref
1080          * lookup.
1081          */
1082         while (!list_empty(&useless)) {
1083                 upper = list_entry(useless.next, struct backref_node, list);
1084                 list_del_init(&upper->list);
1085                 ASSERT(list_empty(&upper->upper));
1086                 if (upper == node)
1087                         node = NULL;
1088                 if (upper->lowest) {
1089                         list_del_init(&upper->lower);
1090                         upper->lowest = 0;
1091                 }
1092                 while (!list_empty(&upper->lower)) {
1093                         edge = list_entry(upper->lower.next,
1094                                           struct backref_edge, list[UPPER]);
1095                         list_del(&edge->list[UPPER]);
1096                         list_del(&edge->list[LOWER]);
1097                         lower = edge->node[LOWER];
1098                         free_backref_edge(cache, edge);
1099
1100                         if (list_empty(&lower->upper))
1101                                 list_add(&lower->list, &useless);
1102                 }
1103                 __mark_block_processed(rc, upper);
1104                 if (upper->level > 0) {
1105                         list_add(&upper->list, &cache->detached);
1106                         upper->detached = 1;
1107                 } else {
1108                         rb_erase(&upper->rb_node, &cache->rb_root);
1109                         free_backref_node(cache, upper);
1110                 }
1111         }
1112 out:
1113         btrfs_free_path(path1);
1114         btrfs_free_path(path2);
1115         if (err) {
1116                 while (!list_empty(&useless)) {
1117                         lower = list_entry(useless.next,
1118                                            struct backref_node, list);
1119                         list_del_init(&lower->list);
1120                 }
1121                 while (!list_empty(&list)) {
1122                         edge = list_first_entry(&list, struct backref_edge,
1123                                                 list[UPPER]);
1124                         list_del(&edge->list[UPPER]);
1125                         list_del(&edge->list[LOWER]);
1126                         lower = edge->node[LOWER];
1127                         upper = edge->node[UPPER];
1128                         free_backref_edge(cache, edge);
1129
1130                         /*
1131                          * Lower is no longer linked to any upper backref nodes
1132                          * and isn't in the cache, we can free it ourselves.
1133                          */
1134                         if (list_empty(&lower->upper) &&
1135                             RB_EMPTY_NODE(&lower->rb_node))
1136                                 list_add(&lower->list, &useless);
1137
1138                         if (!RB_EMPTY_NODE(&upper->rb_node))
1139                                 continue;
1140
1141                         /* Add this guy's upper edges to the list to process */
1142                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1143                                 list_add_tail(&edge->list[UPPER], &list);
1144                         if (list_empty(&upper->upper))
1145                                 list_add(&upper->list, &useless);
1146                 }
1147
1148                 while (!list_empty(&useless)) {
1149                         lower = list_entry(useless.next,
1150                                            struct backref_node, list);
1151                         list_del_init(&lower->list);
1152                         if (lower == node)
1153                                 node = NULL;
1154                         free_backref_node(cache, lower);
1155                 }
1156
1157                 free_backref_node(cache, node);
1158                 return ERR_PTR(err);
1159         }
1160         ASSERT(!node || !node->detached);
1161         return node;
1162 }
1163
1164 /*
1165  * helper to add backref node for the newly created snapshot.
1166  * the backref node is created by cloning backref node that
1167  * corresponds to root of source tree
1168  */
1169 static int clone_backref_node(struct btrfs_trans_handle *trans,
1170                               struct reloc_control *rc,
1171                               struct btrfs_root *src,
1172                               struct btrfs_root *dest)
1173 {
1174         struct btrfs_root *reloc_root = src->reloc_root;
1175         struct backref_cache *cache = &rc->backref_cache;
1176         struct backref_node *node = NULL;
1177         struct backref_node *new_node;
1178         struct backref_edge *edge;
1179         struct backref_edge *new_edge;
1180         struct rb_node *rb_node;
1181
1182         if (cache->last_trans > 0)
1183                 update_backref_cache(trans, cache);
1184
1185         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1186         if (rb_node) {
1187                 node = rb_entry(rb_node, struct backref_node, rb_node);
1188                 if (node->detached)
1189                         node = NULL;
1190                 else
1191                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1192         }
1193
1194         if (!node) {
1195                 rb_node = tree_search(&cache->rb_root,
1196                                       reloc_root->commit_root->start);
1197                 if (rb_node) {
1198                         node = rb_entry(rb_node, struct backref_node,
1199                                         rb_node);
1200                         BUG_ON(node->detached);
1201                 }
1202         }
1203
1204         if (!node)
1205                 return 0;
1206
1207         new_node = alloc_backref_node(cache);
1208         if (!new_node)
1209                 return -ENOMEM;
1210
1211         new_node->bytenr = dest->node->start;
1212         new_node->level = node->level;
1213         new_node->lowest = node->lowest;
1214         new_node->checked = 1;
1215         new_node->root = dest;
1216
1217         if (!node->lowest) {
1218                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1219                         new_edge = alloc_backref_edge(cache);
1220                         if (!new_edge)
1221                                 goto fail;
1222
1223                         new_edge->node[UPPER] = new_node;
1224                         new_edge->node[LOWER] = edge->node[LOWER];
1225                         list_add_tail(&new_edge->list[UPPER],
1226                                       &new_node->lower);
1227                 }
1228         } else {
1229                 list_add_tail(&new_node->lower, &cache->leaves);
1230         }
1231
1232         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1233                               &new_node->rb_node);
1234         if (rb_node)
1235                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1236
1237         if (!new_node->lowest) {
1238                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1239                         list_add_tail(&new_edge->list[LOWER],
1240                                       &new_edge->node[LOWER]->upper);
1241                 }
1242         }
1243         return 0;
1244 fail:
1245         while (!list_empty(&new_node->lower)) {
1246                 new_edge = list_entry(new_node->lower.next,
1247                                       struct backref_edge, list[UPPER]);
1248                 list_del(&new_edge->list[UPPER]);
1249                 free_backref_edge(cache, new_edge);
1250         }
1251         free_backref_node(cache, new_node);
1252         return -ENOMEM;
1253 }
1254
1255 /*
1256  * helper to add 'address of tree root -> reloc tree' mapping
1257  */
1258 static int __must_check __add_reloc_root(struct btrfs_root *root)
1259 {
1260         struct btrfs_fs_info *fs_info = root->fs_info;
1261         struct rb_node *rb_node;
1262         struct mapping_node *node;
1263         struct reloc_control *rc = fs_info->reloc_ctl;
1264
1265         node = kmalloc(sizeof(*node), GFP_NOFS);
1266         if (!node)
1267                 return -ENOMEM;
1268
1269         node->bytenr = root->node->start;
1270         node->data = root;
1271
1272         spin_lock(&rc->reloc_root_tree.lock);
1273         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1274                               node->bytenr, &node->rb_node);
1275         spin_unlock(&rc->reloc_root_tree.lock);
1276         if (rb_node) {
1277                 btrfs_panic(fs_info, -EEXIST,
1278                             "Duplicate root found for start=%llu while inserting into relocation tree",
1279                             node->bytenr);
1280         }
1281
1282         list_add_tail(&root->root_list, &rc->reloc_roots);
1283         return 0;
1284 }
1285
1286 /*
1287  * helper to delete the 'address of tree root -> reloc tree'
1288  * mapping
1289  */
1290 static void __del_reloc_root(struct btrfs_root *root)
1291 {
1292         struct btrfs_fs_info *fs_info = root->fs_info;
1293         struct rb_node *rb_node;
1294         struct mapping_node *node = NULL;
1295         struct reloc_control *rc = fs_info->reloc_ctl;
1296
1297         if (rc && root->node) {
1298                 spin_lock(&rc->reloc_root_tree.lock);
1299                 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1300                                       root->node->start);
1301                 if (rb_node) {
1302                         node = rb_entry(rb_node, struct mapping_node, rb_node);
1303                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1304                 }
1305                 spin_unlock(&rc->reloc_root_tree.lock);
1306                 if (!node)
1307                         return;
1308                 BUG_ON((struct btrfs_root *)node->data != root);
1309         }
1310
1311         spin_lock(&fs_info->trans_lock);
1312         list_del_init(&root->root_list);
1313         spin_unlock(&fs_info->trans_lock);
1314         kfree(node);
1315 }
1316
1317 /*
1318  * helper to update the 'address of tree root -> reloc tree'
1319  * mapping
1320  */
1321 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1322 {
1323         struct btrfs_fs_info *fs_info = root->fs_info;
1324         struct rb_node *rb_node;
1325         struct mapping_node *node = NULL;
1326         struct reloc_control *rc = fs_info->reloc_ctl;
1327
1328         spin_lock(&rc->reloc_root_tree.lock);
1329         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1330                               root->node->start);
1331         if (rb_node) {
1332                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1333                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1334         }
1335         spin_unlock(&rc->reloc_root_tree.lock);
1336
1337         if (!node)
1338                 return 0;
1339         BUG_ON((struct btrfs_root *)node->data != root);
1340
1341         spin_lock(&rc->reloc_root_tree.lock);
1342         node->bytenr = new_bytenr;
1343         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1344                               node->bytenr, &node->rb_node);
1345         spin_unlock(&rc->reloc_root_tree.lock);
1346         if (rb_node)
1347                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1348         return 0;
1349 }
1350
1351 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1352                                         struct btrfs_root *root, u64 objectid)
1353 {
1354         struct btrfs_fs_info *fs_info = root->fs_info;
1355         struct btrfs_root *reloc_root;
1356         struct extent_buffer *eb;
1357         struct btrfs_root_item *root_item;
1358         struct btrfs_key root_key;
1359         int ret;
1360
1361         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1362         BUG_ON(!root_item);
1363
1364         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1365         root_key.type = BTRFS_ROOT_ITEM_KEY;
1366         root_key.offset = objectid;
1367
1368         if (root->root_key.objectid == objectid) {
1369                 u64 commit_root_gen;
1370
1371                 /* called by btrfs_init_reloc_root */
1372                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1373                                       BTRFS_TREE_RELOC_OBJECTID);
1374                 BUG_ON(ret);
1375                 /*
1376                  * Set the last_snapshot field to the generation of the commit
1377                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1378                  * correctly (returns true) when the relocation root is created
1379                  * either inside the critical section of a transaction commit
1380                  * (through transaction.c:qgroup_account_snapshot()) and when
1381                  * it's created before the transaction commit is started.
1382                  */
1383                 commit_root_gen = btrfs_header_generation(root->commit_root);
1384                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1385         } else {
1386                 /*
1387                  * called by btrfs_reloc_post_snapshot_hook.
1388                  * the source tree is a reloc tree, all tree blocks
1389                  * modified after it was created have RELOC flag
1390                  * set in their headers. so it's OK to not update
1391                  * the 'last_snapshot'.
1392                  */
1393                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1394                                       BTRFS_TREE_RELOC_OBJECTID);
1395                 BUG_ON(ret);
1396         }
1397
1398         memcpy(root_item, &root->root_item, sizeof(*root_item));
1399         btrfs_set_root_bytenr(root_item, eb->start);
1400         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1401         btrfs_set_root_generation(root_item, trans->transid);
1402
1403         if (root->root_key.objectid == objectid) {
1404                 btrfs_set_root_refs(root_item, 0);
1405                 memset(&root_item->drop_progress, 0,
1406                        sizeof(struct btrfs_disk_key));
1407                 root_item->drop_level = 0;
1408         }
1409
1410         btrfs_tree_unlock(eb);
1411         free_extent_buffer(eb);
1412
1413         ret = btrfs_insert_root(trans, fs_info->tree_root,
1414                                 &root_key, root_item);
1415         BUG_ON(ret);
1416         kfree(root_item);
1417
1418         reloc_root = btrfs_read_fs_root(fs_info->tree_root, &root_key);
1419         BUG_ON(IS_ERR(reloc_root));
1420         reloc_root->last_trans = trans->transid;
1421         return reloc_root;
1422 }
1423
1424 /*
1425  * create reloc tree for a given fs tree. reloc tree is just a
1426  * snapshot of the fs tree with special root objectid.
1427  */
1428 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1429                           struct btrfs_root *root)
1430 {
1431         struct btrfs_fs_info *fs_info = root->fs_info;
1432         struct btrfs_root *reloc_root;
1433         struct reloc_control *rc = fs_info->reloc_ctl;
1434         struct btrfs_block_rsv *rsv;
1435         int clear_rsv = 0;
1436         int ret;
1437
1438         if (root->reloc_root) {
1439                 reloc_root = root->reloc_root;
1440                 reloc_root->last_trans = trans->transid;
1441                 return 0;
1442         }
1443
1444         if (!rc || !rc->create_reloc_tree ||
1445             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1446                 return 0;
1447
1448         if (!trans->reloc_reserved) {
1449                 rsv = trans->block_rsv;
1450                 trans->block_rsv = rc->block_rsv;
1451                 clear_rsv = 1;
1452         }
1453         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1454         if (clear_rsv)
1455                 trans->block_rsv = rsv;
1456
1457         ret = __add_reloc_root(reloc_root);
1458         BUG_ON(ret < 0);
1459         root->reloc_root = reloc_root;
1460         return 0;
1461 }
1462
1463 /*
1464  * update root item of reloc tree
1465  */
1466 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1467                             struct btrfs_root *root)
1468 {
1469         struct btrfs_fs_info *fs_info = root->fs_info;
1470         struct btrfs_root *reloc_root;
1471         struct btrfs_root_item *root_item;
1472         int ret;
1473
1474         if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state) ||
1475             !root->reloc_root)
1476                 goto out;
1477
1478         reloc_root = root->reloc_root;
1479         root_item = &reloc_root->root_item;
1480
1481         /* root->reloc_root will stay until current relocation finished */
1482         if (fs_info->reloc_ctl->merge_reloc_tree &&
1483             btrfs_root_refs(root_item) == 0) {
1484                 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1485                 __del_reloc_root(reloc_root);
1486         }
1487
1488         if (reloc_root->commit_root != reloc_root->node) {
1489                 btrfs_set_root_node(root_item, reloc_root->node);
1490                 free_extent_buffer(reloc_root->commit_root);
1491                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1492         }
1493
1494         ret = btrfs_update_root(trans, fs_info->tree_root,
1495                                 &reloc_root->root_key, root_item);
1496         BUG_ON(ret);
1497
1498 out:
1499         return 0;
1500 }
1501
1502 /*
1503  * helper to find first cached inode with inode number >= objectid
1504  * in a subvolume
1505  */
1506 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1507 {
1508         struct rb_node *node;
1509         struct rb_node *prev;
1510         struct btrfs_inode *entry;
1511         struct inode *inode;
1512
1513         spin_lock(&root->inode_lock);
1514 again:
1515         node = root->inode_tree.rb_node;
1516         prev = NULL;
1517         while (node) {
1518                 prev = node;
1519                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1520
1521                 if (objectid < btrfs_ino(entry))
1522                         node = node->rb_left;
1523                 else if (objectid > btrfs_ino(entry))
1524                         node = node->rb_right;
1525                 else
1526                         break;
1527         }
1528         if (!node) {
1529                 while (prev) {
1530                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1531                         if (objectid <= btrfs_ino(entry)) {
1532                                 node = prev;
1533                                 break;
1534                         }
1535                         prev = rb_next(prev);
1536                 }
1537         }
1538         while (node) {
1539                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1540                 inode = igrab(&entry->vfs_inode);
1541                 if (inode) {
1542                         spin_unlock(&root->inode_lock);
1543                         return inode;
1544                 }
1545
1546                 objectid = btrfs_ino(entry) + 1;
1547                 if (cond_resched_lock(&root->inode_lock))
1548                         goto again;
1549
1550                 node = rb_next(node);
1551         }
1552         spin_unlock(&root->inode_lock);
1553         return NULL;
1554 }
1555
1556 static int in_block_group(u64 bytenr,
1557                           struct btrfs_block_group_cache *block_group)
1558 {
1559         if (bytenr >= block_group->key.objectid &&
1560             bytenr < block_group->key.objectid + block_group->key.offset)
1561                 return 1;
1562         return 0;
1563 }
1564
1565 /*
1566  * get new location of data
1567  */
1568 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1569                             u64 bytenr, u64 num_bytes)
1570 {
1571         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1572         struct btrfs_path *path;
1573         struct btrfs_file_extent_item *fi;
1574         struct extent_buffer *leaf;
1575         int ret;
1576
1577         path = btrfs_alloc_path();
1578         if (!path)
1579                 return -ENOMEM;
1580
1581         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1582         ret = btrfs_lookup_file_extent(NULL, root, path,
1583                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1584         if (ret < 0)
1585                 goto out;
1586         if (ret > 0) {
1587                 ret = -ENOENT;
1588                 goto out;
1589         }
1590
1591         leaf = path->nodes[0];
1592         fi = btrfs_item_ptr(leaf, path->slots[0],
1593                             struct btrfs_file_extent_item);
1594
1595         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1596                btrfs_file_extent_compression(leaf, fi) ||
1597                btrfs_file_extent_encryption(leaf, fi) ||
1598                btrfs_file_extent_other_encoding(leaf, fi));
1599
1600         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1601                 ret = -EINVAL;
1602                 goto out;
1603         }
1604
1605         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1606         ret = 0;
1607 out:
1608         btrfs_free_path(path);
1609         return ret;
1610 }
1611
1612 /*
1613  * update file extent items in the tree leaf to point to
1614  * the new locations.
1615  */
1616 static noinline_for_stack
1617 int replace_file_extents(struct btrfs_trans_handle *trans,
1618                          struct reloc_control *rc,
1619                          struct btrfs_root *root,
1620                          struct extent_buffer *leaf)
1621 {
1622         struct btrfs_fs_info *fs_info = root->fs_info;
1623         struct btrfs_key key;
1624         struct btrfs_file_extent_item *fi;
1625         struct inode *inode = NULL;
1626         u64 parent;
1627         u64 bytenr;
1628         u64 new_bytenr = 0;
1629         u64 num_bytes;
1630         u64 end;
1631         u32 nritems;
1632         u32 i;
1633         int ret = 0;
1634         int first = 1;
1635         int dirty = 0;
1636
1637         if (rc->stage != UPDATE_DATA_PTRS)
1638                 return 0;
1639
1640         /* reloc trees always use full backref */
1641         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1642                 parent = leaf->start;
1643         else
1644                 parent = 0;
1645
1646         nritems = btrfs_header_nritems(leaf);
1647         for (i = 0; i < nritems; i++) {
1648                 struct btrfs_ref ref = { 0 };
1649
1650                 cond_resched();
1651                 btrfs_item_key_to_cpu(leaf, &key, i);
1652                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1653                         continue;
1654                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1655                 if (btrfs_file_extent_type(leaf, fi) ==
1656                     BTRFS_FILE_EXTENT_INLINE)
1657                         continue;
1658                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1659                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1660                 if (bytenr == 0)
1661                         continue;
1662                 if (!in_block_group(bytenr, rc->block_group))
1663                         continue;
1664
1665                 /*
1666                  * if we are modifying block in fs tree, wait for readpage
1667                  * to complete and drop the extent cache
1668                  */
1669                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1670                         if (first) {
1671                                 inode = find_next_inode(root, key.objectid);
1672                                 first = 0;
1673                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1674                                 btrfs_add_delayed_iput(inode);
1675                                 inode = find_next_inode(root, key.objectid);
1676                         }
1677                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1678                                 end = key.offset +
1679                                       btrfs_file_extent_num_bytes(leaf, fi);
1680                                 WARN_ON(!IS_ALIGNED(key.offset,
1681                                                     fs_info->sectorsize));
1682                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1683                                 end--;
1684                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1685                                                       key.offset, end);
1686                                 if (!ret)
1687                                         continue;
1688
1689                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1690                                                 key.offset,     end, 1);
1691                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1692                                               key.offset, end);
1693                         }
1694                 }
1695
1696                 ret = get_new_location(rc->data_inode, &new_bytenr,
1697                                        bytenr, num_bytes);
1698                 if (ret) {
1699                         /*
1700                          * Don't have to abort since we've not changed anything
1701                          * in the file extent yet.
1702                          */
1703                         break;
1704                 }
1705
1706                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1707                 dirty = 1;
1708
1709                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1710                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1711                                        num_bytes, parent);
1712                 ref.real_root = root->root_key.objectid;
1713                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1714                                     key.objectid, key.offset);
1715                 ret = btrfs_inc_extent_ref(trans, &ref);
1716                 if (ret) {
1717                         btrfs_abort_transaction(trans, ret);
1718                         break;
1719                 }
1720
1721                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1722                                        num_bytes, parent);
1723                 ref.real_root = root->root_key.objectid;
1724                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1725                                     key.objectid, key.offset);
1726                 ret = btrfs_free_extent(trans, &ref);
1727                 if (ret) {
1728                         btrfs_abort_transaction(trans, ret);
1729                         break;
1730                 }
1731         }
1732         if (dirty)
1733                 btrfs_mark_buffer_dirty(leaf);
1734         if (inode)
1735                 btrfs_add_delayed_iput(inode);
1736         return ret;
1737 }
1738
1739 static noinline_for_stack
1740 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1741                      struct btrfs_path *path, int level)
1742 {
1743         struct btrfs_disk_key key1;
1744         struct btrfs_disk_key key2;
1745         btrfs_node_key(eb, &key1, slot);
1746         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1747         return memcmp(&key1, &key2, sizeof(key1));
1748 }
1749
1750 /*
1751  * try to replace tree blocks in fs tree with the new blocks
1752  * in reloc tree. tree blocks haven't been modified since the
1753  * reloc tree was create can be replaced.
1754  *
1755  * if a block was replaced, level of the block + 1 is returned.
1756  * if no block got replaced, 0 is returned. if there are other
1757  * errors, a negative error number is returned.
1758  */
1759 static noinline_for_stack
1760 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1761                  struct btrfs_root *dest, struct btrfs_root *src,
1762                  struct btrfs_path *path, struct btrfs_key *next_key,
1763                  int lowest_level, int max_level)
1764 {
1765         struct btrfs_fs_info *fs_info = dest->fs_info;
1766         struct extent_buffer *eb;
1767         struct extent_buffer *parent;
1768         struct btrfs_ref ref = { 0 };
1769         struct btrfs_key key;
1770         u64 old_bytenr;
1771         u64 new_bytenr;
1772         u64 old_ptr_gen;
1773         u64 new_ptr_gen;
1774         u64 last_snapshot;
1775         u32 blocksize;
1776         int cow = 0;
1777         int level;
1778         int ret;
1779         int slot;
1780
1781         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1782         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1783
1784         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1785 again:
1786         slot = path->slots[lowest_level];
1787         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1788
1789         eb = btrfs_lock_root_node(dest);
1790         btrfs_set_lock_blocking_write(eb);
1791         level = btrfs_header_level(eb);
1792
1793         if (level < lowest_level) {
1794                 btrfs_tree_unlock(eb);
1795                 free_extent_buffer(eb);
1796                 return 0;
1797         }
1798
1799         if (cow) {
1800                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1801                 BUG_ON(ret);
1802         }
1803         btrfs_set_lock_blocking_write(eb);
1804
1805         if (next_key) {
1806                 next_key->objectid = (u64)-1;
1807                 next_key->type = (u8)-1;
1808                 next_key->offset = (u64)-1;
1809         }
1810
1811         parent = eb;
1812         while (1) {
1813                 struct btrfs_key first_key;
1814
1815                 level = btrfs_header_level(parent);
1816                 BUG_ON(level < lowest_level);
1817
1818                 ret = btrfs_bin_search(parent, &key, level, &slot);
1819                 if (ret < 0)
1820                         break;
1821                 if (ret && slot > 0)
1822                         slot--;
1823
1824                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1825                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1826
1827                 old_bytenr = btrfs_node_blockptr(parent, slot);
1828                 blocksize = fs_info->nodesize;
1829                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1830                 btrfs_node_key_to_cpu(parent, &first_key, slot);
1831
1832                 if (level <= max_level) {
1833                         eb = path->nodes[level];
1834                         new_bytenr = btrfs_node_blockptr(eb,
1835                                                         path->slots[level]);
1836                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1837                                                         path->slots[level]);
1838                 } else {
1839                         new_bytenr = 0;
1840                         new_ptr_gen = 0;
1841                 }
1842
1843                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1844                         ret = level;
1845                         break;
1846                 }
1847
1848                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1849                     memcmp_node_keys(parent, slot, path, level)) {
1850                         if (level <= lowest_level) {
1851                                 ret = 0;
1852                                 break;
1853                         }
1854
1855                         eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen,
1856                                              level - 1, &first_key);
1857                         if (IS_ERR(eb)) {
1858                                 ret = PTR_ERR(eb);
1859                                 break;
1860                         } else if (!extent_buffer_uptodate(eb)) {
1861                                 ret = -EIO;
1862                                 free_extent_buffer(eb);
1863                                 break;
1864                         }
1865                         btrfs_tree_lock(eb);
1866                         if (cow) {
1867                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1868                                                       slot, &eb);
1869                                 BUG_ON(ret);
1870                         }
1871                         btrfs_set_lock_blocking_write(eb);
1872
1873                         btrfs_tree_unlock(parent);
1874                         free_extent_buffer(parent);
1875
1876                         parent = eb;
1877                         continue;
1878                 }
1879
1880                 if (!cow) {
1881                         btrfs_tree_unlock(parent);
1882                         free_extent_buffer(parent);
1883                         cow = 1;
1884                         goto again;
1885                 }
1886
1887                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1888                                       path->slots[level]);
1889                 btrfs_release_path(path);
1890
1891                 path->lowest_level = level;
1892                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1893                 path->lowest_level = 0;
1894                 BUG_ON(ret);
1895
1896                 /*
1897                  * Info qgroup to trace both subtrees.
1898                  *
1899                  * We must trace both trees.
1900                  * 1) Tree reloc subtree
1901                  *    If not traced, we will leak data numbers
1902                  * 2) Fs subtree
1903                  *    If not traced, we will double count old data
1904                  *
1905                  * We don't scan the subtree right now, but only record
1906                  * the swapped tree blocks.
1907                  * The real subtree rescan is delayed until we have new
1908                  * CoW on the subtree root node before transaction commit.
1909                  */
1910                 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1911                                 rc->block_group, parent, slot,
1912                                 path->nodes[level], path->slots[level],
1913                                 last_snapshot);
1914                 if (ret < 0)
1915                         break;
1916                 /*
1917                  * swap blocks in fs tree and reloc tree.
1918                  */
1919                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1920                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1921                 btrfs_mark_buffer_dirty(parent);
1922
1923                 btrfs_set_node_blockptr(path->nodes[level],
1924                                         path->slots[level], old_bytenr);
1925                 btrfs_set_node_ptr_generation(path->nodes[level],
1926                                               path->slots[level], old_ptr_gen);
1927                 btrfs_mark_buffer_dirty(path->nodes[level]);
1928
1929                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1930                                        blocksize, path->nodes[level]->start);
1931                 ref.skip_qgroup = true;
1932                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1933                 ret = btrfs_inc_extent_ref(trans, &ref);
1934                 BUG_ON(ret);
1935                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1936                                        blocksize, 0);
1937                 ref.skip_qgroup = true;
1938                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1939                 ret = btrfs_inc_extent_ref(trans, &ref);
1940                 BUG_ON(ret);
1941
1942                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1943                                        blocksize, path->nodes[level]->start);
1944                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1945                 ref.skip_qgroup = true;
1946                 ret = btrfs_free_extent(trans, &ref);
1947                 BUG_ON(ret);
1948
1949                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1950                                        blocksize, 0);
1951                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1952                 ref.skip_qgroup = true;
1953                 ret = btrfs_free_extent(trans, &ref);
1954                 BUG_ON(ret);
1955
1956                 btrfs_unlock_up_safe(path, 0);
1957
1958                 ret = level;
1959                 break;
1960         }
1961         btrfs_tree_unlock(parent);
1962         free_extent_buffer(parent);
1963         return ret;
1964 }
1965
1966 /*
1967  * helper to find next relocated block in reloc tree
1968  */
1969 static noinline_for_stack
1970 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1971                        int *level)
1972 {
1973         struct extent_buffer *eb;
1974         int i;
1975         u64 last_snapshot;
1976         u32 nritems;
1977
1978         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1979
1980         for (i = 0; i < *level; i++) {
1981                 free_extent_buffer(path->nodes[i]);
1982                 path->nodes[i] = NULL;
1983         }
1984
1985         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1986                 eb = path->nodes[i];
1987                 nritems = btrfs_header_nritems(eb);
1988                 while (path->slots[i] + 1 < nritems) {
1989                         path->slots[i]++;
1990                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1991                             last_snapshot)
1992                                 continue;
1993
1994                         *level = i;
1995                         return 0;
1996                 }
1997                 free_extent_buffer(path->nodes[i]);
1998                 path->nodes[i] = NULL;
1999         }
2000         return 1;
2001 }
2002
2003 /*
2004  * walk down reloc tree to find relocated block of lowest level
2005  */
2006 static noinline_for_stack
2007 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
2008                          int *level)
2009 {
2010         struct btrfs_fs_info *fs_info = root->fs_info;
2011         struct extent_buffer *eb = NULL;
2012         int i;
2013         u64 bytenr;
2014         u64 ptr_gen = 0;
2015         u64 last_snapshot;
2016         u32 nritems;
2017
2018         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2019
2020         for (i = *level; i > 0; i--) {
2021                 struct btrfs_key first_key;
2022
2023                 eb = path->nodes[i];
2024                 nritems = btrfs_header_nritems(eb);
2025                 while (path->slots[i] < nritems) {
2026                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2027                         if (ptr_gen > last_snapshot)
2028                                 break;
2029                         path->slots[i]++;
2030                 }
2031                 if (path->slots[i] >= nritems) {
2032                         if (i == *level)
2033                                 break;
2034                         *level = i + 1;
2035                         return 0;
2036                 }
2037                 if (i == 1) {
2038                         *level = i;
2039                         return 0;
2040                 }
2041
2042                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2043                 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
2044                 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
2045                                      &first_key);
2046                 if (IS_ERR(eb)) {
2047                         return PTR_ERR(eb);
2048                 } else if (!extent_buffer_uptodate(eb)) {
2049                         free_extent_buffer(eb);
2050                         return -EIO;
2051                 }
2052                 BUG_ON(btrfs_header_level(eb) != i - 1);
2053                 path->nodes[i - 1] = eb;
2054                 path->slots[i - 1] = 0;
2055         }
2056         return 1;
2057 }
2058
2059 /*
2060  * invalidate extent cache for file extents whose key in range of
2061  * [min_key, max_key)
2062  */
2063 static int invalidate_extent_cache(struct btrfs_root *root,
2064                                    struct btrfs_key *min_key,
2065                                    struct btrfs_key *max_key)
2066 {
2067         struct btrfs_fs_info *fs_info = root->fs_info;
2068         struct inode *inode = NULL;
2069         u64 objectid;
2070         u64 start, end;
2071         u64 ino;
2072
2073         objectid = min_key->objectid;
2074         while (1) {
2075                 cond_resched();
2076                 iput(inode);
2077
2078                 if (objectid > max_key->objectid)
2079                         break;
2080
2081                 inode = find_next_inode(root, objectid);
2082                 if (!inode)
2083                         break;
2084                 ino = btrfs_ino(BTRFS_I(inode));
2085
2086                 if (ino > max_key->objectid) {
2087                         iput(inode);
2088                         break;
2089                 }
2090
2091                 objectid = ino + 1;
2092                 if (!S_ISREG(inode->i_mode))
2093                         continue;
2094
2095                 if (unlikely(min_key->objectid == ino)) {
2096                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2097                                 continue;
2098                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2099                                 start = 0;
2100                         else {
2101                                 start = min_key->offset;
2102                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
2103                         }
2104                 } else {
2105                         start = 0;
2106                 }
2107
2108                 if (unlikely(max_key->objectid == ino)) {
2109                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2110                                 continue;
2111                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2112                                 end = (u64)-1;
2113                         } else {
2114                                 if (max_key->offset == 0)
2115                                         continue;
2116                                 end = max_key->offset;
2117                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
2118                                 end--;
2119                         }
2120                 } else {
2121                         end = (u64)-1;
2122                 }
2123
2124                 /* the lock_extent waits for readpage to complete */
2125                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2126                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
2127                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2128         }
2129         return 0;
2130 }
2131
2132 static int find_next_key(struct btrfs_path *path, int level,
2133                          struct btrfs_key *key)
2134
2135 {
2136         while (level < BTRFS_MAX_LEVEL) {
2137                 if (!path->nodes[level])
2138                         break;
2139                 if (path->slots[level] + 1 <
2140                     btrfs_header_nritems(path->nodes[level])) {
2141                         btrfs_node_key_to_cpu(path->nodes[level], key,
2142                                               path->slots[level] + 1);
2143                         return 0;
2144                 }
2145                 level++;
2146         }
2147         return 1;
2148 }
2149
2150 /*
2151  * Insert current subvolume into reloc_control::dirty_subvol_roots
2152  */
2153 static void insert_dirty_subvol(struct btrfs_trans_handle *trans,
2154                                 struct reloc_control *rc,
2155                                 struct btrfs_root *root)
2156 {
2157         struct btrfs_root *reloc_root = root->reloc_root;
2158         struct btrfs_root_item *reloc_root_item;
2159
2160         /* @root must be a subvolume tree root with a valid reloc tree */
2161         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
2162         ASSERT(reloc_root);
2163
2164         reloc_root_item = &reloc_root->root_item;
2165         memset(&reloc_root_item->drop_progress, 0,
2166                 sizeof(reloc_root_item->drop_progress));
2167         reloc_root_item->drop_level = 0;
2168         btrfs_set_root_refs(reloc_root_item, 0);
2169         btrfs_update_reloc_root(trans, root);
2170
2171         if (list_empty(&root->reloc_dirty_list)) {
2172                 btrfs_grab_fs_root(root);
2173                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
2174         }
2175 }
2176
2177 static int clean_dirty_subvols(struct reloc_control *rc)
2178 {
2179         struct btrfs_root *root;
2180         struct btrfs_root *next;
2181         int ret = 0;
2182         int ret2;
2183
2184         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
2185                                  reloc_dirty_list) {
2186                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
2187                         /* Merged subvolume, cleanup its reloc root */
2188                         struct btrfs_root *reloc_root = root->reloc_root;
2189
2190                         clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
2191                         list_del_init(&root->reloc_dirty_list);
2192                         root->reloc_root = NULL;
2193                         if (reloc_root) {
2194
2195                                 ret2 = btrfs_drop_snapshot(reloc_root, NULL, 0, 1);
2196                                 if (ret2 < 0 && !ret)
2197                                         ret = ret2;
2198                         }
2199                         btrfs_put_fs_root(root);
2200                 } else {
2201                         /* Orphan reloc tree, just clean it up */
2202                         ret2 = btrfs_drop_snapshot(root, NULL, 0, 1);
2203                         if (ret2 < 0 && !ret)
2204                                 ret = ret2;
2205                 }
2206         }
2207         return ret;
2208 }
2209
2210 /*
2211  * merge the relocated tree blocks in reloc tree with corresponding
2212  * fs tree.
2213  */
2214 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2215                                                struct btrfs_root *root)
2216 {
2217         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2218         struct btrfs_key key;
2219         struct btrfs_key next_key;
2220         struct btrfs_trans_handle *trans = NULL;
2221         struct btrfs_root *reloc_root;
2222         struct btrfs_root_item *root_item;
2223         struct btrfs_path *path;
2224         struct extent_buffer *leaf;
2225         int level;
2226         int max_level;
2227         int replaced = 0;
2228         int ret;
2229         int err = 0;
2230         u32 min_reserved;
2231
2232         path = btrfs_alloc_path();
2233         if (!path)
2234                 return -ENOMEM;
2235         path->reada = READA_FORWARD;
2236
2237         reloc_root = root->reloc_root;
2238         root_item = &reloc_root->root_item;
2239
2240         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2241                 level = btrfs_root_level(root_item);
2242                 extent_buffer_get(reloc_root->node);
2243                 path->nodes[level] = reloc_root->node;
2244                 path->slots[level] = 0;
2245         } else {
2246                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2247
2248                 level = root_item->drop_level;
2249                 BUG_ON(level == 0);
2250                 path->lowest_level = level;
2251                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2252                 path->lowest_level = 0;
2253                 if (ret < 0) {
2254                         btrfs_free_path(path);
2255                         return ret;
2256                 }
2257
2258                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2259                                       path->slots[level]);
2260                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2261
2262                 btrfs_unlock_up_safe(path, 0);
2263         }
2264
2265         min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2266         memset(&next_key, 0, sizeof(next_key));
2267
2268         while (1) {
2269                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2270                                              BTRFS_RESERVE_FLUSH_ALL);
2271                 if (ret) {
2272                         err = ret;
2273                         goto out;
2274                 }
2275                 trans = btrfs_start_transaction(root, 0);
2276                 if (IS_ERR(trans)) {
2277                         err = PTR_ERR(trans);
2278                         trans = NULL;
2279                         goto out;
2280                 }
2281                 trans->block_rsv = rc->block_rsv;
2282
2283                 replaced = 0;
2284                 max_level = level;
2285
2286                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2287                 if (ret < 0) {
2288                         err = ret;
2289                         goto out;
2290                 }
2291                 if (ret > 0)
2292                         break;
2293
2294                 if (!find_next_key(path, level, &key) &&
2295                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2296                         ret = 0;
2297                 } else {
2298                         ret = replace_path(trans, rc, root, reloc_root, path,
2299                                            &next_key, level, max_level);
2300                 }
2301                 if (ret < 0) {
2302                         err = ret;
2303                         goto out;
2304                 }
2305
2306                 if (ret > 0) {
2307                         level = ret;
2308                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2309                                               path->slots[level]);
2310                         replaced = 1;
2311                 }
2312
2313                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2314                 if (ret > 0)
2315                         break;
2316
2317                 BUG_ON(level == 0);
2318                 /*
2319                  * save the merging progress in the drop_progress.
2320                  * this is OK since root refs == 1 in this case.
2321                  */
2322                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2323                                path->slots[level]);
2324                 root_item->drop_level = level;
2325
2326                 btrfs_end_transaction_throttle(trans);
2327                 trans = NULL;
2328
2329                 btrfs_btree_balance_dirty(fs_info);
2330
2331                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2332                         invalidate_extent_cache(root, &key, &next_key);
2333         }
2334
2335         /*
2336          * handle the case only one block in the fs tree need to be
2337          * relocated and the block is tree root.
2338          */
2339         leaf = btrfs_lock_root_node(root);
2340         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2341         btrfs_tree_unlock(leaf);
2342         free_extent_buffer(leaf);
2343         if (ret < 0)
2344                 err = ret;
2345 out:
2346         btrfs_free_path(path);
2347
2348         if (err == 0)
2349                 insert_dirty_subvol(trans, rc, root);
2350
2351         if (trans)
2352                 btrfs_end_transaction_throttle(trans);
2353
2354         btrfs_btree_balance_dirty(fs_info);
2355
2356         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2357                 invalidate_extent_cache(root, &key, &next_key);
2358
2359         return err;
2360 }
2361
2362 static noinline_for_stack
2363 int prepare_to_merge(struct reloc_control *rc, int err)
2364 {
2365         struct btrfs_root *root = rc->extent_root;
2366         struct btrfs_fs_info *fs_info = root->fs_info;
2367         struct btrfs_root *reloc_root;
2368         struct btrfs_trans_handle *trans;
2369         LIST_HEAD(reloc_roots);
2370         u64 num_bytes = 0;
2371         int ret;
2372
2373         mutex_lock(&fs_info->reloc_mutex);
2374         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2375         rc->merging_rsv_size += rc->nodes_relocated * 2;
2376         mutex_unlock(&fs_info->reloc_mutex);
2377
2378 again:
2379         if (!err) {
2380                 num_bytes = rc->merging_rsv_size;
2381                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2382                                           BTRFS_RESERVE_FLUSH_ALL);
2383                 if (ret)
2384                         err = ret;
2385         }
2386
2387         trans = btrfs_join_transaction(rc->extent_root);
2388         if (IS_ERR(trans)) {
2389                 if (!err)
2390                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2391                                                 num_bytes);
2392                 return PTR_ERR(trans);
2393         }
2394
2395         if (!err) {
2396                 if (num_bytes != rc->merging_rsv_size) {
2397                         btrfs_end_transaction(trans);
2398                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2399                                                 num_bytes);
2400                         goto again;
2401                 }
2402         }
2403
2404         rc->merge_reloc_tree = 1;
2405
2406         while (!list_empty(&rc->reloc_roots)) {
2407                 reloc_root = list_entry(rc->reloc_roots.next,
2408                                         struct btrfs_root, root_list);
2409                 list_del_init(&reloc_root->root_list);
2410
2411                 root = read_fs_root(fs_info, reloc_root->root_key.offset);
2412                 BUG_ON(IS_ERR(root));
2413                 BUG_ON(root->reloc_root != reloc_root);
2414
2415                 /*
2416                  * set reference count to 1, so btrfs_recover_relocation
2417                  * knows it should resumes merging
2418                  */
2419                 if (!err)
2420                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2421                 btrfs_update_reloc_root(trans, root);
2422
2423                 list_add(&reloc_root->root_list, &reloc_roots);
2424         }
2425
2426         list_splice(&reloc_roots, &rc->reloc_roots);
2427
2428         if (!err)
2429                 btrfs_commit_transaction(trans);
2430         else
2431                 btrfs_end_transaction(trans);
2432         return err;
2433 }
2434
2435 static noinline_for_stack
2436 void free_reloc_roots(struct list_head *list)
2437 {
2438         struct btrfs_root *reloc_root;
2439
2440         while (!list_empty(list)) {
2441                 reloc_root = list_entry(list->next, struct btrfs_root,
2442                                         root_list);
2443                 __del_reloc_root(reloc_root);
2444                 free_extent_buffer(reloc_root->node);
2445                 free_extent_buffer(reloc_root->commit_root);
2446                 reloc_root->node = NULL;
2447                 reloc_root->commit_root = NULL;
2448         }
2449 }
2450
2451 static noinline_for_stack
2452 void merge_reloc_roots(struct reloc_control *rc)
2453 {
2454         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2455         struct btrfs_root *root;
2456         struct btrfs_root *reloc_root;
2457         LIST_HEAD(reloc_roots);
2458         int found = 0;
2459         int ret = 0;
2460 again:
2461         root = rc->extent_root;
2462
2463         /*
2464          * this serializes us with btrfs_record_root_in_transaction,
2465          * we have to make sure nobody is in the middle of
2466          * adding their roots to the list while we are
2467          * doing this splice
2468          */
2469         mutex_lock(&fs_info->reloc_mutex);
2470         list_splice_init(&rc->reloc_roots, &reloc_roots);
2471         mutex_unlock(&fs_info->reloc_mutex);
2472
2473         while (!list_empty(&reloc_roots)) {
2474                 found = 1;
2475                 reloc_root = list_entry(reloc_roots.next,
2476                                         struct btrfs_root, root_list);
2477
2478                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2479                         root = read_fs_root(fs_info,
2480                                             reloc_root->root_key.offset);
2481                         BUG_ON(IS_ERR(root));
2482                         BUG_ON(root->reloc_root != reloc_root);
2483
2484                         ret = merge_reloc_root(rc, root);
2485                         if (ret) {
2486                                 if (list_empty(&reloc_root->root_list))
2487                                         list_add_tail(&reloc_root->root_list,
2488                                                       &reloc_roots);
2489                                 goto out;
2490                         }
2491                 } else {
2492                         list_del_init(&reloc_root->root_list);
2493                         /* Don't forget to queue this reloc root for cleanup */
2494                         list_add_tail(&reloc_root->reloc_dirty_list,
2495                                       &rc->dirty_subvol_roots);
2496                 }
2497         }
2498
2499         if (found) {
2500                 found = 0;
2501                 goto again;
2502         }
2503 out:
2504         if (ret) {
2505                 btrfs_handle_fs_error(fs_info, ret, NULL);
2506                 if (!list_empty(&reloc_roots))
2507                         free_reloc_roots(&reloc_roots);
2508
2509                 /* new reloc root may be added */
2510                 mutex_lock(&fs_info->reloc_mutex);
2511                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2512                 mutex_unlock(&fs_info->reloc_mutex);
2513                 if (!list_empty(&reloc_roots))
2514                         free_reloc_roots(&reloc_roots);
2515         }
2516
2517         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2518 }
2519
2520 static void free_block_list(struct rb_root *blocks)
2521 {
2522         struct tree_block *block;
2523         struct rb_node *rb_node;
2524         while ((rb_node = rb_first(blocks))) {
2525                 block = rb_entry(rb_node, struct tree_block, rb_node);
2526                 rb_erase(rb_node, blocks);
2527                 kfree(block);
2528         }
2529 }
2530
2531 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2532                                       struct btrfs_root *reloc_root)
2533 {
2534         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2535         struct btrfs_root *root;
2536
2537         if (reloc_root->last_trans == trans->transid)
2538                 return 0;
2539
2540         root = read_fs_root(fs_info, reloc_root->root_key.offset);
2541         BUG_ON(IS_ERR(root));
2542         BUG_ON(root->reloc_root != reloc_root);
2543
2544         return btrfs_record_root_in_trans(trans, root);
2545 }
2546
2547 static noinline_for_stack
2548 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2549                                      struct reloc_control *rc,
2550                                      struct backref_node *node,
2551                                      struct backref_edge *edges[])
2552 {
2553         struct backref_node *next;
2554         struct btrfs_root *root;
2555         int index = 0;
2556
2557         next = node;
2558         while (1) {
2559                 cond_resched();
2560                 next = walk_up_backref(next, edges, &index);
2561                 root = next->root;
2562                 BUG_ON(!root);
2563                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2564
2565                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2566                         record_reloc_root_in_trans(trans, root);
2567                         break;
2568                 }
2569
2570                 btrfs_record_root_in_trans(trans, root);
2571                 root = root->reloc_root;
2572
2573                 if (next->new_bytenr != root->node->start) {
2574                         BUG_ON(next->new_bytenr);
2575                         BUG_ON(!list_empty(&next->list));
2576                         next->new_bytenr = root->node->start;
2577                         next->root = root;
2578                         list_add_tail(&next->list,
2579                                       &rc->backref_cache.changed);
2580                         __mark_block_processed(rc, next);
2581                         break;
2582                 }
2583
2584                 WARN_ON(1);
2585                 root = NULL;
2586                 next = walk_down_backref(edges, &index);
2587                 if (!next || next->level <= node->level)
2588                         break;
2589         }
2590         if (!root)
2591                 return NULL;
2592
2593         next = node;
2594         /* setup backref node path for btrfs_reloc_cow_block */
2595         while (1) {
2596                 rc->backref_cache.path[next->level] = next;
2597                 if (--index < 0)
2598                         break;
2599                 next = edges[index]->node[UPPER];
2600         }
2601         return root;
2602 }
2603
2604 /*
2605  * select a tree root for relocation. return NULL if the block
2606  * is reference counted. we should use do_relocation() in this
2607  * case. return a tree root pointer if the block isn't reference
2608  * counted. return -ENOENT if the block is root of reloc tree.
2609  */
2610 static noinline_for_stack
2611 struct btrfs_root *select_one_root(struct backref_node *node)
2612 {
2613         struct backref_node *next;
2614         struct btrfs_root *root;
2615         struct btrfs_root *fs_root = NULL;
2616         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2617         int index = 0;
2618
2619         next = node;
2620         while (1) {
2621                 cond_resched();
2622                 next = walk_up_backref(next, edges, &index);
2623                 root = next->root;
2624                 BUG_ON(!root);
2625
2626                 /* no other choice for non-references counted tree */
2627                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2628                         return root;
2629
2630                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2631                         fs_root = root;
2632
2633                 if (next != node)
2634                         return NULL;
2635
2636                 next = walk_down_backref(edges, &index);
2637                 if (!next || next->level <= node->level)
2638                         break;
2639         }
2640
2641         if (!fs_root)
2642                 return ERR_PTR(-ENOENT);
2643         return fs_root;
2644 }
2645
2646 static noinline_for_stack
2647 u64 calcu_metadata_size(struct reloc_control *rc,
2648                         struct backref_node *node, int reserve)
2649 {
2650         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2651         struct backref_node *next = node;
2652         struct backref_edge *edge;
2653         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2654         u64 num_bytes = 0;
2655         int index = 0;
2656
2657         BUG_ON(reserve && node->processed);
2658
2659         while (next) {
2660                 cond_resched();
2661                 while (1) {
2662                         if (next->processed && (reserve || next != node))
2663                                 break;
2664
2665                         num_bytes += fs_info->nodesize;
2666
2667                         if (list_empty(&next->upper))
2668                                 break;
2669
2670                         edge = list_entry(next->upper.next,
2671                                           struct backref_edge, list[LOWER]);
2672                         edges[index++] = edge;
2673                         next = edge->node[UPPER];
2674                 }
2675                 next = walk_down_backref(edges, &index);
2676         }
2677         return num_bytes;
2678 }
2679
2680 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2681                                   struct reloc_control *rc,
2682                                   struct backref_node *node)
2683 {
2684         struct btrfs_root *root = rc->extent_root;
2685         struct btrfs_fs_info *fs_info = root->fs_info;
2686         u64 num_bytes;
2687         int ret;
2688         u64 tmp;
2689
2690         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2691
2692         trans->block_rsv = rc->block_rsv;
2693         rc->reserved_bytes += num_bytes;
2694
2695         /*
2696          * We are under a transaction here so we can only do limited flushing.
2697          * If we get an enospc just kick back -EAGAIN so we know to drop the
2698          * transaction and try to refill when we can flush all the things.
2699          */
2700         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2701                                 BTRFS_RESERVE_FLUSH_LIMIT);
2702         if (ret) {
2703                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2704                 while (tmp <= rc->reserved_bytes)
2705                         tmp <<= 1;
2706                 /*
2707                  * only one thread can access block_rsv at this point,
2708                  * so we don't need hold lock to protect block_rsv.
2709                  * we expand more reservation size here to allow enough
2710                  * space for relocation and we will return earlier in
2711                  * enospc case.
2712                  */
2713                 rc->block_rsv->size = tmp + fs_info->nodesize *
2714                                       RELOCATION_RESERVED_NODES;
2715                 return -EAGAIN;
2716         }
2717
2718         return 0;
2719 }
2720
2721 /*
2722  * relocate a block tree, and then update pointers in upper level
2723  * blocks that reference the block to point to the new location.
2724  *
2725  * if called by link_to_upper, the block has already been relocated.
2726  * in that case this function just updates pointers.
2727  */
2728 static int do_relocation(struct btrfs_trans_handle *trans,
2729                          struct reloc_control *rc,
2730                          struct backref_node *node,
2731                          struct btrfs_key *key,
2732                          struct btrfs_path *path, int lowest)
2733 {
2734         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2735         struct backref_node *upper;
2736         struct backref_edge *edge;
2737         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2738         struct btrfs_root *root;
2739         struct extent_buffer *eb;
2740         u32 blocksize;
2741         u64 bytenr;
2742         u64 generation;
2743         int slot;
2744         int ret;
2745         int err = 0;
2746
2747         BUG_ON(lowest && node->eb);
2748
2749         path->lowest_level = node->level + 1;
2750         rc->backref_cache.path[node->level] = node;
2751         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2752                 struct btrfs_key first_key;
2753                 struct btrfs_ref ref = { 0 };
2754
2755                 cond_resched();
2756
2757                 upper = edge->node[UPPER];
2758                 root = select_reloc_root(trans, rc, upper, edges);
2759                 BUG_ON(!root);
2760
2761                 if (upper->eb && !upper->locked) {
2762                         if (!lowest) {
2763                                 ret = btrfs_bin_search(upper->eb, key,
2764                                                        upper->level, &slot);
2765                                 if (ret < 0) {
2766                                         err = ret;
2767                                         goto next;
2768                                 }
2769                                 BUG_ON(ret);
2770                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2771                                 if (node->eb->start == bytenr)
2772                                         goto next;
2773                         }
2774                         drop_node_buffer(upper);
2775                 }
2776
2777                 if (!upper->eb) {
2778                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2779                         if (ret) {
2780                                 if (ret < 0)
2781                                         err = ret;
2782                                 else
2783                                         err = -ENOENT;
2784
2785                                 btrfs_release_path(path);
2786                                 break;
2787                         }
2788
2789                         if (!upper->eb) {
2790                                 upper->eb = path->nodes[upper->level];
2791                                 path->nodes[upper->level] = NULL;
2792                         } else {
2793                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2794                         }
2795
2796                         upper->locked = 1;
2797                         path->locks[upper->level] = 0;
2798
2799                         slot = path->slots[upper->level];
2800                         btrfs_release_path(path);
2801                 } else {
2802                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2803                                                &slot);
2804                         if (ret < 0) {
2805                                 err = ret;
2806                                 goto next;
2807                         }
2808                         BUG_ON(ret);
2809                 }
2810
2811                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2812                 if (lowest) {
2813                         if (bytenr != node->bytenr) {
2814                                 btrfs_err(root->fs_info,
2815                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2816                                           bytenr, node->bytenr, slot,
2817                                           upper->eb->start);
2818                                 err = -EIO;
2819                                 goto next;
2820                         }
2821                 } else {
2822                         if (node->eb->start == bytenr)
2823                                 goto next;
2824                 }
2825
2826                 blocksize = root->fs_info->nodesize;
2827                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2828                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2829                 eb = read_tree_block(fs_info, bytenr, generation,
2830                                      upper->level - 1, &first_key);
2831                 if (IS_ERR(eb)) {
2832                         err = PTR_ERR(eb);
2833                         goto next;
2834                 } else if (!extent_buffer_uptodate(eb)) {
2835                         free_extent_buffer(eb);
2836                         err = -EIO;
2837                         goto next;
2838                 }
2839                 btrfs_tree_lock(eb);
2840                 btrfs_set_lock_blocking_write(eb);
2841
2842                 if (!node->eb) {
2843                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2844                                               slot, &eb);
2845                         btrfs_tree_unlock(eb);
2846                         free_extent_buffer(eb);
2847                         if (ret < 0) {
2848                                 err = ret;
2849                                 goto next;
2850                         }
2851                         BUG_ON(node->eb != eb);
2852                 } else {
2853                         btrfs_set_node_blockptr(upper->eb, slot,
2854                                                 node->eb->start);
2855                         btrfs_set_node_ptr_generation(upper->eb, slot,
2856                                                       trans->transid);
2857                         btrfs_mark_buffer_dirty(upper->eb);
2858
2859                         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2860                                                node->eb->start, blocksize,
2861                                                upper->eb->start);
2862                         ref.real_root = root->root_key.objectid;
2863                         btrfs_init_tree_ref(&ref, node->level,
2864                                             btrfs_header_owner(upper->eb));
2865                         ret = btrfs_inc_extent_ref(trans, &ref);
2866                         BUG_ON(ret);
2867
2868                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2869                         BUG_ON(ret);
2870                 }
2871 next:
2872                 if (!upper->pending)
2873                         drop_node_buffer(upper);
2874                 else
2875                         unlock_node_buffer(upper);
2876                 if (err)
2877                         break;
2878         }
2879
2880         if (!err && node->pending) {
2881                 drop_node_buffer(node);
2882                 list_move_tail(&node->list, &rc->backref_cache.changed);
2883                 node->pending = 0;
2884         }
2885
2886         path->lowest_level = 0;
2887         BUG_ON(err == -ENOSPC);
2888         return err;
2889 }
2890
2891 static int link_to_upper(struct btrfs_trans_handle *trans,
2892                          struct reloc_control *rc,
2893                          struct backref_node *node,
2894                          struct btrfs_path *path)
2895 {
2896         struct btrfs_key key;
2897
2898         btrfs_node_key_to_cpu(node->eb, &key, 0);
2899         return do_relocation(trans, rc, node, &key, path, 0);
2900 }
2901
2902 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2903                                 struct reloc_control *rc,
2904                                 struct btrfs_path *path, int err)
2905 {
2906         LIST_HEAD(list);
2907         struct backref_cache *cache = &rc->backref_cache;
2908         struct backref_node *node;
2909         int level;
2910         int ret;
2911
2912         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2913                 while (!list_empty(&cache->pending[level])) {
2914                         node = list_entry(cache->pending[level].next,
2915                                           struct backref_node, list);
2916                         list_move_tail(&node->list, &list);
2917                         BUG_ON(!node->pending);
2918
2919                         if (!err) {
2920                                 ret = link_to_upper(trans, rc, node, path);
2921                                 if (ret < 0)
2922                                         err = ret;
2923                         }
2924                 }
2925                 list_splice_init(&list, &cache->pending[level]);
2926         }
2927         return err;
2928 }
2929
2930 static void mark_block_processed(struct reloc_control *rc,
2931                                  u64 bytenr, u32 blocksize)
2932 {
2933         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2934                         EXTENT_DIRTY);
2935 }
2936
2937 static void __mark_block_processed(struct reloc_control *rc,
2938                                    struct backref_node *node)
2939 {
2940         u32 blocksize;
2941         if (node->level == 0 ||
2942             in_block_group(node->bytenr, rc->block_group)) {
2943                 blocksize = rc->extent_root->fs_info->nodesize;
2944                 mark_block_processed(rc, node->bytenr, blocksize);
2945         }
2946         node->processed = 1;
2947 }
2948
2949 /*
2950  * mark a block and all blocks directly/indirectly reference the block
2951  * as processed.
2952  */
2953 static void update_processed_blocks(struct reloc_control *rc,
2954                                     struct backref_node *node)
2955 {
2956         struct backref_node *next = node;
2957         struct backref_edge *edge;
2958         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2959         int index = 0;
2960
2961         while (next) {
2962                 cond_resched();
2963                 while (1) {
2964                         if (next->processed)
2965                                 break;
2966
2967                         __mark_block_processed(rc, next);
2968
2969                         if (list_empty(&next->upper))
2970                                 break;
2971
2972                         edge = list_entry(next->upper.next,
2973                                           struct backref_edge, list[LOWER]);
2974                         edges[index++] = edge;
2975                         next = edge->node[UPPER];
2976                 }
2977                 next = walk_down_backref(edges, &index);
2978         }
2979 }
2980
2981 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2982 {
2983         u32 blocksize = rc->extent_root->fs_info->nodesize;
2984
2985         if (test_range_bit(&rc->processed_blocks, bytenr,
2986                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2987                 return 1;
2988         return 0;
2989 }
2990
2991 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2992                               struct tree_block *block)
2993 {
2994         struct extent_buffer *eb;
2995
2996         BUG_ON(block->key_ready);
2997         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2998                              block->level, NULL);
2999         if (IS_ERR(eb)) {
3000                 return PTR_ERR(eb);
3001         } else if (!extent_buffer_uptodate(eb)) {
3002                 free_extent_buffer(eb);
3003                 return -EIO;
3004         }
3005         if (block->level == 0)
3006                 btrfs_item_key_to_cpu(eb, &block->key, 0);
3007         else
3008                 btrfs_node_key_to_cpu(eb, &block->key, 0);
3009         free_extent_buffer(eb);
3010         block->key_ready = 1;
3011         return 0;
3012 }
3013
3014 /*
3015  * helper function to relocate a tree block
3016  */
3017 static int relocate_tree_block(struct btrfs_trans_handle *trans,
3018                                 struct reloc_control *rc,
3019                                 struct backref_node *node,
3020                                 struct btrfs_key *key,
3021                                 struct btrfs_path *path)
3022 {
3023         struct btrfs_root *root;
3024         int ret = 0;
3025
3026         if (!node)
3027                 return 0;
3028
3029         BUG_ON(node->processed);
3030         root = select_one_root(node);
3031         if (root == ERR_PTR(-ENOENT)) {
3032                 update_processed_blocks(rc, node);
3033                 goto out;
3034         }
3035
3036         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3037                 ret = reserve_metadata_space(trans, rc, node);
3038                 if (ret)
3039                         goto out;
3040         }
3041
3042         if (root) {
3043                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3044                         BUG_ON(node->new_bytenr);
3045                         BUG_ON(!list_empty(&node->list));
3046                         btrfs_record_root_in_trans(trans, root);
3047                         root = root->reloc_root;
3048                         node->new_bytenr = root->node->start;
3049                         node->root = root;
3050                         list_add_tail(&node->list, &rc->backref_cache.changed);
3051                 } else {
3052                         path->lowest_level = node->level;
3053                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
3054                         btrfs_release_path(path);
3055                         if (ret > 0)
3056                                 ret = 0;
3057                 }
3058                 if (!ret)
3059                         update_processed_blocks(rc, node);
3060         } else {
3061                 ret = do_relocation(trans, rc, node, key, path, 1);
3062         }
3063 out:
3064         if (ret || node->level == 0 || node->cowonly)
3065                 remove_backref_node(&rc->backref_cache, node);
3066         return ret;
3067 }
3068
3069 /*
3070  * relocate a list of blocks
3071  */
3072 static noinline_for_stack
3073 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
3074                          struct reloc_control *rc, struct rb_root *blocks)
3075 {
3076         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3077         struct backref_node *node;
3078         struct btrfs_path *path;
3079         struct tree_block *block;
3080         struct tree_block *next;
3081         int ret;
3082         int err = 0;
3083
3084         path = btrfs_alloc_path();
3085         if (!path) {
3086                 err = -ENOMEM;
3087                 goto out_free_blocks;
3088         }
3089
3090         /* Kick in readahead for tree blocks with missing keys */
3091         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3092                 if (!block->key_ready)
3093                         readahead_tree_block(fs_info, block->bytenr);
3094         }
3095
3096         /* Get first keys */
3097         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3098                 if (!block->key_ready) {
3099                         err = get_tree_block_key(fs_info, block);
3100                         if (err)
3101                                 goto out_free_path;
3102                 }
3103         }
3104
3105         /* Do tree relocation */
3106         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3107                 node = build_backref_tree(rc, &block->key,
3108                                           block->level, block->bytenr);
3109                 if (IS_ERR(node)) {
3110                         err = PTR_ERR(node);
3111                         goto out;
3112                 }
3113
3114                 ret = relocate_tree_block(trans, rc, node, &block->key,
3115                                           path);
3116                 if (ret < 0) {
3117                         if (ret != -EAGAIN || &block->rb_node == rb_first(blocks))
3118                                 err = ret;
3119                         goto out;
3120                 }
3121         }
3122 out:
3123         err = finish_pending_nodes(trans, rc, path, err);
3124
3125 out_free_path:
3126         btrfs_free_path(path);
3127 out_free_blocks:
3128         free_block_list(blocks);
3129         return err;
3130 }
3131
3132 static noinline_for_stack
3133 int prealloc_file_extent_cluster(struct inode *inode,
3134                                  struct file_extent_cluster *cluster)
3135 {
3136         u64 alloc_hint = 0;
3137         u64 start;
3138         u64 end;
3139         u64 offset = BTRFS_I(inode)->index_cnt;
3140         u64 num_bytes;
3141         int nr = 0;
3142         int ret = 0;
3143         u64 prealloc_start = cluster->start - offset;
3144         u64 prealloc_end = cluster->end - offset;
3145         u64 cur_offset;
3146         struct extent_changeset *data_reserved = NULL;
3147
3148         BUG_ON(cluster->start != cluster->boundary[0]);
3149         inode_lock(inode);
3150
3151         ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
3152                                           prealloc_end + 1 - prealloc_start);
3153         if (ret)
3154                 goto out;
3155
3156         cur_offset = prealloc_start;
3157         while (nr < cluster->nr) {
3158                 start = cluster->boundary[nr] - offset;
3159                 if (nr + 1 < cluster->nr)
3160                         end = cluster->boundary[nr + 1] - 1 - offset;
3161                 else
3162                         end = cluster->end - offset;
3163
3164                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3165                 num_bytes = end + 1 - start;
3166                 if (cur_offset < start)
3167                         btrfs_free_reserved_data_space(inode, data_reserved,
3168                                         cur_offset, start - cur_offset);
3169                 ret = btrfs_prealloc_file_range(inode, 0, start,
3170                                                 num_bytes, num_bytes,
3171                                                 end + 1, &alloc_hint);
3172                 cur_offset = end + 1;
3173                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3174                 if (ret)
3175                         break;
3176                 nr++;
3177         }
3178         if (cur_offset < prealloc_end)
3179                 btrfs_free_reserved_data_space(inode, data_reserved,
3180                                 cur_offset, prealloc_end + 1 - cur_offset);
3181 out:
3182         inode_unlock(inode);
3183         extent_changeset_free(data_reserved);
3184         return ret;
3185 }
3186
3187 static noinline_for_stack
3188 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3189                          u64 block_start)
3190 {
3191         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3192         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3193         struct extent_map *em;
3194         int ret = 0;
3195
3196         em = alloc_extent_map();
3197         if (!em)
3198                 return -ENOMEM;
3199
3200         em->start = start;
3201         em->len = end + 1 - start;
3202         em->block_len = em->len;
3203         em->block_start = block_start;
3204         em->bdev = fs_info->fs_devices->latest_bdev;
3205         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3206
3207         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3208         while (1) {
3209                 write_lock(&em_tree->lock);
3210                 ret = add_extent_mapping(em_tree, em, 0);
3211                 write_unlock(&em_tree->lock);
3212                 if (ret != -EEXIST) {
3213                         free_extent_map(em);
3214                         break;
3215                 }
3216                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3217         }
3218         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3219         return ret;
3220 }
3221
3222 static int relocate_file_extent_cluster(struct inode *inode,
3223                                         struct file_extent_cluster *cluster)
3224 {
3225         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3226         u64 page_start;
3227         u64 page_end;
3228         u64 offset = BTRFS_I(inode)->index_cnt;
3229         unsigned long index;
3230         unsigned long last_index;
3231         struct page *page;
3232         struct file_ra_state *ra;
3233         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3234         int nr = 0;
3235         int ret = 0;
3236
3237         if (!cluster->nr)
3238                 return 0;
3239
3240         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3241         if (!ra)
3242                 return -ENOMEM;
3243
3244         ret = prealloc_file_extent_cluster(inode, cluster);
3245         if (ret)
3246                 goto out;
3247
3248         file_ra_state_init(ra, inode->i_mapping);
3249
3250         ret = setup_extent_mapping(inode, cluster->start - offset,
3251                                    cluster->end - offset, cluster->start);
3252         if (ret)
3253                 goto out;
3254
3255         index = (cluster->start - offset) >> PAGE_SHIFT;
3256         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3257         while (index <= last_index) {
3258                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3259                                 PAGE_SIZE);
3260                 if (ret)
3261                         goto out;
3262
3263                 page = find_lock_page(inode->i_mapping, index);
3264                 if (!page) {
3265                         page_cache_sync_readahead(inode->i_mapping,
3266                                                   ra, NULL, index,
3267                                                   last_index + 1 - index);
3268                         page = find_or_create_page(inode->i_mapping, index,
3269                                                    mask);
3270                         if (!page) {
3271                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3272                                                         PAGE_SIZE, true);
3273                                 ret = -ENOMEM;
3274                                 goto out;
3275                         }
3276                 }
3277
3278                 if (PageReadahead(page)) {
3279                         page_cache_async_readahead(inode->i_mapping,
3280                                                    ra, NULL, page, index,
3281                                                    last_index + 1 - index);
3282                 }
3283
3284                 if (!PageUptodate(page)) {
3285                         btrfs_readpage(NULL, page);
3286                         lock_page(page);
3287                         if (!PageUptodate(page)) {
3288                                 unlock_page(page);
3289                                 put_page(page);
3290                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3291                                                         PAGE_SIZE, true);
3292                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
3293                                                                PAGE_SIZE, true);
3294                                 ret = -EIO;
3295                                 goto out;
3296                         }
3297                 }
3298
3299                 page_start = page_offset(page);
3300                 page_end = page_start + PAGE_SIZE - 1;
3301
3302                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3303
3304                 set_page_extent_mapped(page);
3305
3306                 if (nr < cluster->nr &&
3307                     page_start + offset == cluster->boundary[nr]) {
3308                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3309                                         page_start, page_end,
3310                                         EXTENT_BOUNDARY);
3311                         nr++;
3312                 }
3313
3314                 ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
3315                                                 NULL);
3316                 if (ret) {
3317                         unlock_page(page);
3318                         put_page(page);
3319                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3320                                                          PAGE_SIZE, true);
3321                         btrfs_delalloc_release_extents(BTRFS_I(inode),
3322                                                        PAGE_SIZE, true);
3323
3324                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
3325                                           page_start, page_end,
3326                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
3327                         goto out;
3328
3329                 }
3330                 set_page_dirty(page);
3331
3332                 unlock_extent(&BTRFS_I(inode)->io_tree,
3333                               page_start, page_end);
3334                 unlock_page(page);
3335                 put_page(page);
3336
3337                 index++;
3338                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE,
3339                                                false);
3340                 balance_dirty_pages_ratelimited(inode->i_mapping);
3341                 btrfs_throttle(fs_info);
3342         }
3343         WARN_ON(nr != cluster->nr);
3344 out:
3345         kfree(ra);
3346         return ret;
3347 }
3348
3349 static noinline_for_stack
3350 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3351                          struct file_extent_cluster *cluster)
3352 {
3353         int ret;
3354
3355         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3356                 ret = relocate_file_extent_cluster(inode, cluster);
3357                 if (ret)
3358                         return ret;
3359                 cluster->nr = 0;
3360         }
3361
3362         if (!cluster->nr)
3363                 cluster->start = extent_key->objectid;
3364         else
3365                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3366         cluster->end = extent_key->objectid + extent_key->offset - 1;
3367         cluster->boundary[cluster->nr] = extent_key->objectid;
3368         cluster->nr++;
3369
3370         if (cluster->nr >= MAX_EXTENTS) {
3371                 ret = relocate_file_extent_cluster(inode, cluster);
3372                 if (ret)
3373                         return ret;
3374                 cluster->nr = 0;
3375         }
3376         return 0;
3377 }
3378
3379 /*
3380  * helper to add a tree block to the list.
3381  * the major work is getting the generation and level of the block
3382  */
3383 static int add_tree_block(struct reloc_control *rc,
3384                           struct btrfs_key *extent_key,
3385                           struct btrfs_path *path,
3386                           struct rb_root *blocks)
3387 {
3388         struct extent_buffer *eb;
3389         struct btrfs_extent_item *ei;
3390         struct btrfs_tree_block_info *bi;
3391         struct tree_block *block;
3392         struct rb_node *rb_node;
3393         u32 item_size;
3394         int level = -1;
3395         u64 generation;
3396
3397         eb =  path->nodes[0];
3398         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3399
3400         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3401             item_size >= sizeof(*ei) + sizeof(*bi)) {
3402                 ei = btrfs_item_ptr(eb, path->slots[0],
3403                                 struct btrfs_extent_item);
3404                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3405                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3406                         level = btrfs_tree_block_level(eb, bi);
3407                 } else {
3408                         level = (int)extent_key->offset;
3409                 }
3410                 generation = btrfs_extent_generation(eb, ei);
3411         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3412                 btrfs_print_v0_err(eb->fs_info);
3413                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3414                 return -EINVAL;
3415         } else {
3416                 BUG();
3417         }
3418
3419         btrfs_release_path(path);
3420
3421         BUG_ON(level == -1);
3422
3423         block = kmalloc(sizeof(*block), GFP_NOFS);
3424         if (!block)
3425                 return -ENOMEM;
3426
3427         block->bytenr = extent_key->objectid;
3428         block->key.objectid = rc->extent_root->fs_info->nodesize;
3429         block->key.offset = generation;
3430         block->level = level;
3431         block->key_ready = 0;
3432
3433         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3434         if (rb_node)
3435                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3436
3437         return 0;
3438 }
3439
3440 /*
3441  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3442  */
3443 static int __add_tree_block(struct reloc_control *rc,
3444                             u64 bytenr, u32 blocksize,
3445                             struct rb_root *blocks)
3446 {
3447         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3448         struct btrfs_path *path;
3449         struct btrfs_key key;
3450         int ret;
3451         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3452
3453         if (tree_block_processed(bytenr, rc))
3454                 return 0;
3455
3456         if (tree_search(blocks, bytenr))
3457                 return 0;
3458
3459         path = btrfs_alloc_path();
3460         if (!path)
3461                 return -ENOMEM;
3462 again:
3463         key.objectid = bytenr;
3464         if (skinny) {
3465                 key.type = BTRFS_METADATA_ITEM_KEY;
3466                 key.offset = (u64)-1;
3467         } else {
3468                 key.type = BTRFS_EXTENT_ITEM_KEY;
3469                 key.offset = blocksize;
3470         }
3471
3472         path->search_commit_root = 1;
3473         path->skip_locking = 1;
3474         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3475         if (ret < 0)
3476                 goto out;
3477
3478         if (ret > 0 && skinny) {
3479                 if (path->slots[0]) {
3480                         path->slots[0]--;
3481                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3482                                               path->slots[0]);
3483                         if (key.objectid == bytenr &&
3484                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3485                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3486                               key.offset == blocksize)))
3487                                 ret = 0;
3488                 }
3489
3490                 if (ret) {
3491                         skinny = false;
3492                         btrfs_release_path(path);
3493                         goto again;
3494                 }
3495         }
3496         if (ret) {
3497                 ASSERT(ret == 1);
3498                 btrfs_print_leaf(path->nodes[0]);
3499                 btrfs_err(fs_info,
3500              "tree block extent item (%llu) is not found in extent tree",
3501                      bytenr);
3502                 WARN_ON(1);
3503                 ret = -EINVAL;
3504                 goto out;
3505         }
3506
3507         ret = add_tree_block(rc, &key, path, blocks);
3508 out:
3509         btrfs_free_path(path);
3510         return ret;
3511 }
3512
3513 /*
3514  * helper to check if the block use full backrefs for pointers in it
3515  */
3516 static int block_use_full_backref(struct reloc_control *rc,
3517                                   struct extent_buffer *eb)
3518 {
3519         u64 flags;
3520         int ret;
3521
3522         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3523             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3524                 return 1;
3525
3526         ret = btrfs_lookup_extent_info(NULL, rc->extent_root->fs_info,
3527                                        eb->start, btrfs_header_level(eb), 1,
3528                                        NULL, &flags);
3529         BUG_ON(ret);
3530
3531         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3532                 ret = 1;
3533         else
3534                 ret = 0;
3535         return ret;
3536 }
3537
3538 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3539                                     struct btrfs_block_group_cache *block_group,
3540                                     struct inode *inode,
3541                                     u64 ino)
3542 {
3543         struct btrfs_key key;
3544         struct btrfs_root *root = fs_info->tree_root;
3545         struct btrfs_trans_handle *trans;
3546         int ret = 0;
3547
3548         if (inode)
3549                 goto truncate;
3550
3551         key.objectid = ino;
3552         key.type = BTRFS_INODE_ITEM_KEY;
3553         key.offset = 0;
3554
3555         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3556         if (IS_ERR(inode))
3557                 return -ENOENT;
3558
3559 truncate:
3560         ret = btrfs_check_trunc_cache_free_space(fs_info,
3561                                                  &fs_info->global_block_rsv);
3562         if (ret)
3563                 goto out;
3564
3565         trans = btrfs_join_transaction(root);
3566         if (IS_ERR(trans)) {
3567                 ret = PTR_ERR(trans);
3568                 goto out;
3569         }
3570
3571         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3572
3573         btrfs_end_transaction(trans);
3574         btrfs_btree_balance_dirty(fs_info);
3575 out:
3576         iput(inode);
3577         return ret;
3578 }
3579
3580 /*
3581  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3582  * this function scans fs tree to find blocks reference the data extent
3583  */
3584 static int find_data_references(struct reloc_control *rc,
3585                                 struct btrfs_key *extent_key,
3586                                 struct extent_buffer *leaf,
3587                                 struct btrfs_extent_data_ref *ref,
3588                                 struct rb_root *blocks)
3589 {
3590         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3591         struct btrfs_path *path;
3592         struct tree_block *block;
3593         struct btrfs_root *root;
3594         struct btrfs_file_extent_item *fi;
3595         struct rb_node *rb_node;
3596         struct btrfs_key key;
3597         u64 ref_root;
3598         u64 ref_objectid;
3599         u64 ref_offset;
3600         u32 ref_count;
3601         u32 nritems;
3602         int err = 0;
3603         int added = 0;
3604         int counted;
3605         int ret;
3606
3607         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3608         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3609         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3610         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3611
3612         /*
3613          * This is an extent belonging to the free space cache, lets just delete
3614          * it and redo the search.
3615          */
3616         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3617                 ret = delete_block_group_cache(fs_info, rc->block_group,
3618                                                NULL, ref_objectid);
3619                 if (ret != -ENOENT)
3620                         return ret;
3621                 ret = 0;
3622         }
3623
3624         path = btrfs_alloc_path();
3625         if (!path)
3626                 return -ENOMEM;
3627         path->reada = READA_FORWARD;
3628
3629         root = read_fs_root(fs_info, ref_root);
3630         if (IS_ERR(root)) {
3631                 err = PTR_ERR(root);
3632                 goto out;
3633         }
3634
3635         key.objectid = ref_objectid;
3636         key.type = BTRFS_EXTENT_DATA_KEY;
3637         if (ref_offset > ((u64)-1 << 32))
3638                 key.offset = 0;
3639         else
3640                 key.offset = ref_offset;
3641
3642         path->search_commit_root = 1;
3643         path->skip_locking = 1;
3644         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3645         if (ret < 0) {
3646                 err = ret;
3647                 goto out;
3648         }
3649
3650         leaf = path->nodes[0];
3651         nritems = btrfs_header_nritems(leaf);
3652         /*
3653          * the references in tree blocks that use full backrefs
3654          * are not counted in
3655          */
3656         if (block_use_full_backref(rc, leaf))
3657                 counted = 0;
3658         else
3659                 counted = 1;
3660         rb_node = tree_search(blocks, leaf->start);
3661         if (rb_node) {
3662                 if (counted)
3663                         added = 1;
3664                 else
3665                         path->slots[0] = nritems;
3666         }
3667
3668         while (ref_count > 0) {
3669                 while (path->slots[0] >= nritems) {
3670                         ret = btrfs_next_leaf(root, path);
3671                         if (ret < 0) {
3672                                 err = ret;
3673                                 goto out;
3674                         }
3675                         if (WARN_ON(ret > 0))
3676                                 goto out;
3677
3678                         leaf = path->nodes[0];
3679                         nritems = btrfs_header_nritems(leaf);
3680                         added = 0;
3681
3682                         if (block_use_full_backref(rc, leaf))
3683                                 counted = 0;
3684                         else
3685                                 counted = 1;
3686                         rb_node = tree_search(blocks, leaf->start);
3687                         if (rb_node) {
3688                                 if (counted)
3689                                         added = 1;
3690                                 else
3691                                         path->slots[0] = nritems;
3692                         }
3693                 }
3694
3695                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3696                 if (WARN_ON(key.objectid != ref_objectid ||
3697                     key.type != BTRFS_EXTENT_DATA_KEY))
3698                         break;
3699
3700                 fi = btrfs_item_ptr(leaf, path->slots[0],
3701                                     struct btrfs_file_extent_item);
3702
3703                 if (btrfs_file_extent_type(leaf, fi) ==
3704                     BTRFS_FILE_EXTENT_INLINE)
3705                         goto next;
3706
3707                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3708                     extent_key->objectid)
3709                         goto next;
3710
3711                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3712                 if (key.offset != ref_offset)
3713                         goto next;
3714
3715                 if (counted)
3716                         ref_count--;
3717                 if (added)
3718                         goto next;
3719
3720                 if (!tree_block_processed(leaf->start, rc)) {
3721                         block = kmalloc(sizeof(*block), GFP_NOFS);
3722                         if (!block) {
3723                                 err = -ENOMEM;
3724                                 break;
3725                         }
3726                         block->bytenr = leaf->start;
3727                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3728                         block->level = 0;
3729                         block->key_ready = 1;
3730                         rb_node = tree_insert(blocks, block->bytenr,
3731                                               &block->rb_node);
3732                         if (rb_node)
3733                                 backref_tree_panic(rb_node, -EEXIST,
3734                                                    block->bytenr);
3735                 }
3736                 if (counted)
3737                         added = 1;
3738                 else
3739                         path->slots[0] = nritems;
3740 next:
3741                 path->slots[0]++;
3742
3743         }
3744 out:
3745         btrfs_free_path(path);
3746         return err;
3747 }
3748
3749 /*
3750  * helper to find all tree blocks that reference a given data extent
3751  */
3752 static noinline_for_stack
3753 int add_data_references(struct reloc_control *rc,
3754                         struct btrfs_key *extent_key,
3755                         struct btrfs_path *path,
3756                         struct rb_root *blocks)
3757 {
3758         struct btrfs_key key;
3759         struct extent_buffer *eb;
3760         struct btrfs_extent_data_ref *dref;
3761         struct btrfs_extent_inline_ref *iref;
3762         unsigned long ptr;
3763         unsigned long end;
3764         u32 blocksize = rc->extent_root->fs_info->nodesize;
3765         int ret = 0;
3766         int err = 0;
3767
3768         eb = path->nodes[0];
3769         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3770         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3771         ptr += sizeof(struct btrfs_extent_item);
3772
3773         while (ptr < end) {
3774                 iref = (struct btrfs_extent_inline_ref *)ptr;
3775                 key.type = btrfs_get_extent_inline_ref_type(eb, iref,
3776                                                         BTRFS_REF_TYPE_DATA);
3777                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3778                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3779                         ret = __add_tree_block(rc, key.offset, blocksize,
3780                                                blocks);
3781                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3782                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3783                         ret = find_data_references(rc, extent_key,
3784                                                    eb, dref, blocks);
3785                 } else {
3786                         ret = -EUCLEAN;
3787                         btrfs_err(rc->extent_root->fs_info,
3788                      "extent %llu slot %d has an invalid inline ref type",
3789                              eb->start, path->slots[0]);
3790                 }
3791                 if (ret) {
3792                         err = ret;
3793                         goto out;
3794                 }
3795                 ptr += btrfs_extent_inline_ref_size(key.type);
3796         }
3797         WARN_ON(ptr > end);
3798
3799         while (1) {
3800                 cond_resched();
3801                 eb = path->nodes[0];
3802                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3803                         ret = btrfs_next_leaf(rc->extent_root, path);
3804                         if (ret < 0) {
3805                                 err = ret;
3806                                 break;
3807                         }
3808                         if (ret > 0)
3809                                 break;
3810                         eb = path->nodes[0];
3811                 }
3812
3813                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3814                 if (key.objectid != extent_key->objectid)
3815                         break;
3816
3817                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3818                         ret = __add_tree_block(rc, key.offset, blocksize,
3819                                                blocks);
3820                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3821                         dref = btrfs_item_ptr(eb, path->slots[0],
3822                                               struct btrfs_extent_data_ref);
3823                         ret = find_data_references(rc, extent_key,
3824                                                    eb, dref, blocks);
3825                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
3826                         btrfs_print_v0_err(eb->fs_info);
3827                         btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3828                         ret = -EINVAL;
3829                 } else {
3830                         ret = 0;
3831                 }
3832                 if (ret) {
3833                         err = ret;
3834                         break;
3835                 }
3836                 path->slots[0]++;
3837         }
3838 out:
3839         btrfs_release_path(path);
3840         if (err)
3841                 free_block_list(blocks);
3842         return err;
3843 }
3844
3845 /*
3846  * helper to find next unprocessed extent
3847  */
3848 static noinline_for_stack
3849 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3850                      struct btrfs_key *extent_key)
3851 {
3852         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3853         struct btrfs_key key;
3854         struct extent_buffer *leaf;
3855         u64 start, end, last;
3856         int ret;
3857
3858         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3859         while (1) {
3860                 cond_resched();
3861                 if (rc->search_start >= last) {
3862                         ret = 1;
3863                         break;
3864                 }
3865
3866                 key.objectid = rc->search_start;
3867                 key.type = BTRFS_EXTENT_ITEM_KEY;
3868                 key.offset = 0;
3869
3870                 path->search_commit_root = 1;
3871                 path->skip_locking = 1;
3872                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3873                                         0, 0);
3874                 if (ret < 0)
3875                         break;
3876 next:
3877                 leaf = path->nodes[0];
3878                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3879                         ret = btrfs_next_leaf(rc->extent_root, path);
3880                         if (ret != 0)
3881                                 break;
3882                         leaf = path->nodes[0];
3883                 }
3884
3885                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3886                 if (key.objectid >= last) {
3887                         ret = 1;
3888                         break;
3889                 }
3890
3891                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3892                     key.type != BTRFS_METADATA_ITEM_KEY) {
3893                         path->slots[0]++;
3894                         goto next;
3895                 }
3896
3897                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3898                     key.objectid + key.offset <= rc->search_start) {
3899                         path->slots[0]++;
3900                         goto next;
3901                 }
3902
3903                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3904                     key.objectid + fs_info->nodesize <=
3905                     rc->search_start) {
3906                         path->slots[0]++;
3907                         goto next;
3908                 }
3909
3910                 ret = find_first_extent_bit(&rc->processed_blocks,
3911                                             key.objectid, &start, &end,
3912                                             EXTENT_DIRTY, NULL);
3913
3914                 if (ret == 0 && start <= key.objectid) {
3915                         btrfs_release_path(path);
3916                         rc->search_start = end + 1;
3917                 } else {
3918                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3919                                 rc->search_start = key.objectid + key.offset;
3920                         else
3921                                 rc->search_start = key.objectid +
3922                                         fs_info->nodesize;
3923                         memcpy(extent_key, &key, sizeof(key));
3924                         return 0;
3925                 }
3926         }
3927         btrfs_release_path(path);
3928         return ret;
3929 }
3930
3931 static void set_reloc_control(struct reloc_control *rc)
3932 {
3933         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3934
3935         mutex_lock(&fs_info->reloc_mutex);
3936         fs_info->reloc_ctl = rc;
3937         mutex_unlock(&fs_info->reloc_mutex);
3938 }
3939
3940 static void unset_reloc_control(struct reloc_control *rc)
3941 {
3942         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3943
3944         mutex_lock(&fs_info->reloc_mutex);
3945         fs_info->reloc_ctl = NULL;
3946         mutex_unlock(&fs_info->reloc_mutex);
3947 }
3948
3949 static int check_extent_flags(u64 flags)
3950 {
3951         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3952             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3953                 return 1;
3954         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3955             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3956                 return 1;
3957         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3958             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3959                 return 1;
3960         return 0;
3961 }
3962
3963 static noinline_for_stack
3964 int prepare_to_relocate(struct reloc_control *rc)
3965 {
3966         struct btrfs_trans_handle *trans;
3967         int ret;
3968
3969         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3970                                               BTRFS_BLOCK_RSV_TEMP);
3971         if (!rc->block_rsv)
3972                 return -ENOMEM;
3973
3974         memset(&rc->cluster, 0, sizeof(rc->cluster));
3975         rc->search_start = rc->block_group->key.objectid;
3976         rc->extents_found = 0;
3977         rc->nodes_relocated = 0;
3978         rc->merging_rsv_size = 0;
3979         rc->reserved_bytes = 0;
3980         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3981                               RELOCATION_RESERVED_NODES;
3982         ret = btrfs_block_rsv_refill(rc->extent_root,
3983                                      rc->block_rsv, rc->block_rsv->size,
3984                                      BTRFS_RESERVE_FLUSH_ALL);
3985         if (ret)
3986                 return ret;
3987
3988         rc->create_reloc_tree = 1;
3989         set_reloc_control(rc);
3990
3991         trans = btrfs_join_transaction(rc->extent_root);
3992         if (IS_ERR(trans)) {
3993                 unset_reloc_control(rc);
3994                 /*
3995                  * extent tree is not a ref_cow tree and has no reloc_root to
3996                  * cleanup.  And callers are responsible to free the above
3997                  * block rsv.
3998                  */
3999                 return PTR_ERR(trans);
4000         }
4001         btrfs_commit_transaction(trans);
4002         return 0;
4003 }
4004
4005 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
4006 {
4007         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
4008         struct rb_root blocks = RB_ROOT;
4009         struct btrfs_key key;
4010         struct btrfs_trans_handle *trans = NULL;
4011         struct btrfs_path *path;
4012         struct btrfs_extent_item *ei;
4013         u64 flags;
4014         u32 item_size;
4015         int ret;
4016         int err = 0;
4017         int progress = 0;
4018
4019         path = btrfs_alloc_path();
4020         if (!path)
4021                 return -ENOMEM;
4022         path->reada = READA_FORWARD;
4023
4024         ret = prepare_to_relocate(rc);
4025         if (ret) {
4026                 err = ret;
4027                 goto out_free;
4028         }
4029
4030         while (1) {
4031                 rc->reserved_bytes = 0;
4032                 ret = btrfs_block_rsv_refill(rc->extent_root,
4033                                         rc->block_rsv, rc->block_rsv->size,
4034                                         BTRFS_RESERVE_FLUSH_ALL);
4035                 if (ret) {
4036                         err = ret;
4037                         break;
4038                 }
4039                 progress++;
4040                 trans = btrfs_start_transaction(rc->extent_root, 0);
4041                 if (IS_ERR(trans)) {
4042                         err = PTR_ERR(trans);
4043                         trans = NULL;
4044                         break;
4045                 }
4046 restart:
4047                 if (update_backref_cache(trans, &rc->backref_cache)) {
4048                         btrfs_end_transaction(trans);
4049                         trans = NULL;
4050                         continue;
4051                 }
4052
4053                 ret = find_next_extent(rc, path, &key);
4054                 if (ret < 0)
4055                         err = ret;
4056                 if (ret != 0)
4057                         break;
4058
4059                 rc->extents_found++;
4060
4061                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4062                                     struct btrfs_extent_item);
4063                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4064                 if (item_size >= sizeof(*ei)) {
4065                         flags = btrfs_extent_flags(path->nodes[0], ei);
4066                         ret = check_extent_flags(flags);
4067                         BUG_ON(ret);
4068                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
4069                         err = -EINVAL;
4070                         btrfs_print_v0_err(trans->fs_info);
4071                         btrfs_abort_transaction(trans, err);
4072                         break;
4073                 } else {
4074                         BUG();
4075                 }
4076
4077                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4078                         ret = add_tree_block(rc, &key, path, &blocks);
4079                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4080                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4081                         ret = add_data_references(rc, &key, path, &blocks);
4082                 } else {
4083                         btrfs_release_path(path);
4084                         ret = 0;
4085                 }
4086                 if (ret < 0) {
4087                         err = ret;
4088                         break;
4089                 }
4090
4091                 if (!RB_EMPTY_ROOT(&blocks)) {
4092                         ret = relocate_tree_blocks(trans, rc, &blocks);
4093                         if (ret < 0) {
4094                                 /*
4095                                  * if we fail to relocate tree blocks, force to update
4096                                  * backref cache when committing transaction.
4097                                  */
4098                                 rc->backref_cache.last_trans = trans->transid - 1;
4099
4100                                 if (ret != -EAGAIN) {
4101                                         err = ret;
4102                                         break;
4103                                 }
4104                                 rc->extents_found--;
4105                                 rc->search_start = key.objectid;
4106                         }
4107                 }
4108
4109                 btrfs_end_transaction_throttle(trans);
4110                 btrfs_btree_balance_dirty(fs_info);
4111                 trans = NULL;
4112
4113                 if (rc->stage == MOVE_DATA_EXTENTS &&
4114                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4115                         rc->found_file_extent = 1;
4116                         ret = relocate_data_extent(rc->data_inode,
4117                                                    &key, &rc->cluster);
4118                         if (ret < 0) {
4119                                 err = ret;
4120                                 break;
4121                         }
4122                 }
4123         }
4124         if (trans && progress && err == -ENOSPC) {
4125                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
4126                 if (ret == 1) {
4127                         err = 0;
4128                         progress = 0;
4129                         goto restart;
4130                 }
4131         }
4132
4133         btrfs_release_path(path);
4134         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4135
4136         if (trans) {
4137                 btrfs_end_transaction_throttle(trans);
4138                 btrfs_btree_balance_dirty(fs_info);
4139         }
4140
4141         if (!err) {
4142                 ret = relocate_file_extent_cluster(rc->data_inode,
4143                                                    &rc->cluster);
4144                 if (ret < 0)
4145                         err = ret;
4146         }
4147
4148         rc->create_reloc_tree = 0;
4149         set_reloc_control(rc);
4150
4151         backref_cache_cleanup(&rc->backref_cache);
4152         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4153
4154         err = prepare_to_merge(rc, err);
4155
4156         merge_reloc_roots(rc);
4157
4158         rc->merge_reloc_tree = 0;
4159         unset_reloc_control(rc);
4160         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4161
4162         /* get rid of pinned extents */
4163         trans = btrfs_join_transaction(rc->extent_root);
4164         if (IS_ERR(trans)) {
4165                 err = PTR_ERR(trans);
4166                 goto out_free;
4167         }
4168         btrfs_commit_transaction(trans);
4169         ret = clean_dirty_subvols(rc);
4170         if (ret < 0 && !err)
4171                 err = ret;
4172 out_free:
4173         btrfs_free_block_rsv(fs_info, rc->block_rsv);
4174         btrfs_free_path(path);
4175         return err;
4176 }
4177
4178 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4179                                  struct btrfs_root *root, u64 objectid)
4180 {
4181         struct btrfs_path *path;
4182         struct btrfs_inode_item *item;
4183         struct extent_buffer *leaf;
4184         int ret;
4185
4186         path = btrfs_alloc_path();
4187         if (!path)
4188                 return -ENOMEM;
4189
4190         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4191         if (ret)
4192                 goto out;
4193
4194         leaf = path->nodes[0];
4195         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4196         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
4197         btrfs_set_inode_generation(leaf, item, 1);
4198         btrfs_set_inode_size(leaf, item, 0);
4199         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4200         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4201                                           BTRFS_INODE_PREALLOC);
4202         btrfs_mark_buffer_dirty(leaf);
4203 out:
4204         btrfs_free_path(path);
4205         return ret;
4206 }
4207
4208 /*
4209  * helper to create inode for data relocation.
4210  * the inode is in data relocation tree and its link count is 0
4211  */
4212 static noinline_for_stack
4213 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4214                                  struct btrfs_block_group_cache *group)
4215 {
4216         struct inode *inode = NULL;
4217         struct btrfs_trans_handle *trans;
4218         struct btrfs_root *root;
4219         struct btrfs_key key;
4220         u64 objectid;
4221         int err = 0;
4222
4223         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4224         if (IS_ERR(root))
4225                 return ERR_CAST(root);
4226
4227         trans = btrfs_start_transaction(root, 6);
4228         if (IS_ERR(trans))
4229                 return ERR_CAST(trans);
4230
4231         err = btrfs_find_free_objectid(root, &objectid);
4232         if (err)
4233                 goto out;
4234
4235         err = __insert_orphan_inode(trans, root, objectid);
4236         BUG_ON(err);
4237
4238         key.objectid = objectid;
4239         key.type = BTRFS_INODE_ITEM_KEY;
4240         key.offset = 0;
4241         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4242         BUG_ON(IS_ERR(inode));
4243         BTRFS_I(inode)->index_cnt = group->key.objectid;
4244
4245         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4246 out:
4247         btrfs_end_transaction(trans);
4248         btrfs_btree_balance_dirty(fs_info);
4249         if (err) {
4250                 if (inode)
4251                         iput(inode);
4252                 inode = ERR_PTR(err);
4253         }
4254         return inode;
4255 }
4256
4257 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4258 {
4259         struct reloc_control *rc;
4260
4261         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4262         if (!rc)
4263                 return NULL;
4264
4265         INIT_LIST_HEAD(&rc->reloc_roots);
4266         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
4267         backref_cache_init(&rc->backref_cache);
4268         mapping_tree_init(&rc->reloc_root_tree);
4269         extent_io_tree_init(fs_info, &rc->processed_blocks,
4270                             IO_TREE_RELOC_BLOCKS, NULL);
4271         return rc;
4272 }
4273
4274 /*
4275  * Print the block group being relocated
4276  */
4277 static void describe_relocation(struct btrfs_fs_info *fs_info,
4278                                 struct btrfs_block_group_cache *block_group)
4279 {
4280         char buf[128] = {'\0'};
4281
4282         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
4283
4284         btrfs_info(fs_info,
4285                    "relocating block group %llu flags %s",
4286                    block_group->key.objectid, buf);
4287 }
4288
4289 /*
4290  * function to relocate all extents in a block group.
4291  */
4292 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4293 {
4294         struct btrfs_block_group_cache *bg;
4295         struct btrfs_root *extent_root = fs_info->extent_root;
4296         struct reloc_control *rc;
4297         struct inode *inode;
4298         struct btrfs_path *path;
4299         int ret;
4300         int rw = 0;
4301         int err = 0;
4302
4303         bg = btrfs_lookup_block_group(fs_info, group_start);
4304         if (!bg)
4305                 return -ENOENT;
4306
4307         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4308                 btrfs_put_block_group(bg);
4309                 return -ETXTBSY;
4310         }
4311
4312         rc = alloc_reloc_control(fs_info);
4313         if (!rc) {
4314                 btrfs_put_block_group(bg);
4315                 return -ENOMEM;
4316         }
4317
4318         rc->extent_root = extent_root;
4319         rc->block_group = bg;
4320
4321         ret = btrfs_inc_block_group_ro(rc->block_group);
4322         if (ret) {
4323                 err = ret;
4324                 goto out;
4325         }
4326         rw = 1;
4327
4328         path = btrfs_alloc_path();
4329         if (!path) {
4330                 err = -ENOMEM;
4331                 goto out;
4332         }
4333
4334         inode = lookup_free_space_inode(rc->block_group, path);
4335         btrfs_free_path(path);
4336
4337         if (!IS_ERR(inode))
4338                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4339         else
4340                 ret = PTR_ERR(inode);
4341
4342         if (ret && ret != -ENOENT) {
4343                 err = ret;
4344                 goto out;
4345         }
4346
4347         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4348         if (IS_ERR(rc->data_inode)) {
4349                 err = PTR_ERR(rc->data_inode);
4350                 rc->data_inode = NULL;
4351                 goto out;
4352         }
4353
4354         describe_relocation(fs_info, rc->block_group);
4355
4356         btrfs_wait_block_group_reservations(rc->block_group);
4357         btrfs_wait_nocow_writers(rc->block_group);
4358         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4359                                  rc->block_group->key.objectid,
4360                                  rc->block_group->key.offset);
4361
4362         while (1) {
4363                 mutex_lock(&fs_info->cleaner_mutex);
4364                 ret = relocate_block_group(rc);
4365                 mutex_unlock(&fs_info->cleaner_mutex);
4366                 if (ret < 0)
4367                         err = ret;
4368
4369                 /*
4370                  * We may have gotten ENOSPC after we already dirtied some
4371                  * extents.  If writeout happens while we're relocating a
4372                  * different block group we could end up hitting the
4373                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4374                  * btrfs_reloc_cow_block.  Make sure we write everything out
4375                  * properly so we don't trip over this problem, and then break
4376                  * out of the loop if we hit an error.
4377                  */
4378                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4379                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4380                                                        (u64)-1);
4381                         if (ret)
4382                                 err = ret;
4383                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4384                                                  0, -1);
4385                         rc->stage = UPDATE_DATA_PTRS;
4386                 }
4387
4388                 if (err < 0)
4389                         goto out;
4390
4391                 if (rc->extents_found == 0)
4392                         break;
4393
4394                 btrfs_info(fs_info, "found %llu extents", rc->extents_found);
4395
4396         }
4397
4398         WARN_ON(rc->block_group->pinned > 0);
4399         WARN_ON(rc->block_group->reserved > 0);
4400         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4401 out:
4402         if (err && rw)
4403                 btrfs_dec_block_group_ro(rc->block_group);
4404         iput(rc->data_inode);
4405         btrfs_put_block_group(rc->block_group);
4406         kfree(rc);
4407         return err;
4408 }
4409
4410 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4411 {
4412         struct btrfs_fs_info *fs_info = root->fs_info;
4413         struct btrfs_trans_handle *trans;
4414         int ret, err;
4415
4416         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4417         if (IS_ERR(trans))
4418                 return PTR_ERR(trans);
4419
4420         memset(&root->root_item.drop_progress, 0,
4421                 sizeof(root->root_item.drop_progress));
4422         root->root_item.drop_level = 0;
4423         btrfs_set_root_refs(&root->root_item, 0);
4424         ret = btrfs_update_root(trans, fs_info->tree_root,
4425                                 &root->root_key, &root->root_item);
4426
4427         err = btrfs_end_transaction(trans);
4428         if (err)
4429                 return err;
4430         return ret;
4431 }
4432
4433 /*
4434  * recover relocation interrupted by system crash.
4435  *
4436  * this function resumes merging reloc trees with corresponding fs trees.
4437  * this is important for keeping the sharing of tree blocks
4438  */
4439 int btrfs_recover_relocation(struct btrfs_root *root)
4440 {
4441         struct btrfs_fs_info *fs_info = root->fs_info;
4442         LIST_HEAD(reloc_roots);
4443         struct btrfs_key key;
4444         struct btrfs_root *fs_root;
4445         struct btrfs_root *reloc_root;
4446         struct btrfs_path *path;
4447         struct extent_buffer *leaf;
4448         struct reloc_control *rc = NULL;
4449         struct btrfs_trans_handle *trans;
4450         int ret;
4451         int err = 0;
4452
4453         path = btrfs_alloc_path();
4454         if (!path)
4455                 return -ENOMEM;
4456         path->reada = READA_BACK;
4457
4458         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4459         key.type = BTRFS_ROOT_ITEM_KEY;
4460         key.offset = (u64)-1;
4461
4462         while (1) {
4463                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4464                                         path, 0, 0);
4465                 if (ret < 0) {
4466                         err = ret;
4467                         goto out;
4468                 }
4469                 if (ret > 0) {
4470                         if (path->slots[0] == 0)
4471                                 break;
4472                         path->slots[0]--;
4473                 }
4474                 leaf = path->nodes[0];
4475                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4476                 btrfs_release_path(path);
4477
4478                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4479                     key.type != BTRFS_ROOT_ITEM_KEY)
4480                         break;
4481
4482                 reloc_root = btrfs_read_fs_root(root, &key);
4483                 if (IS_ERR(reloc_root)) {
4484                         err = PTR_ERR(reloc_root);
4485                         goto out;
4486                 }
4487
4488                 list_add(&reloc_root->root_list, &reloc_roots);
4489
4490                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4491                         fs_root = read_fs_root(fs_info,
4492                                                reloc_root->root_key.offset);
4493                         if (IS_ERR(fs_root)) {
4494                                 ret = PTR_ERR(fs_root);
4495                                 if (ret != -ENOENT) {
4496                                         err = ret;
4497                                         goto out;
4498                                 }
4499                                 ret = mark_garbage_root(reloc_root);
4500                                 if (ret < 0) {
4501                                         err = ret;
4502                                         goto out;
4503                                 }
4504                         }
4505                 }
4506
4507                 if (key.offset == 0)
4508                         break;
4509
4510                 key.offset--;
4511         }
4512         btrfs_release_path(path);
4513
4514         if (list_empty(&reloc_roots))
4515                 goto out;
4516
4517         rc = alloc_reloc_control(fs_info);
4518         if (!rc) {
4519                 err = -ENOMEM;
4520                 goto out;
4521         }
4522
4523         rc->extent_root = fs_info->extent_root;
4524
4525         set_reloc_control(rc);
4526
4527         trans = btrfs_join_transaction(rc->extent_root);
4528         if (IS_ERR(trans)) {
4529                 unset_reloc_control(rc);
4530                 err = PTR_ERR(trans);
4531                 goto out_free;
4532         }
4533
4534         rc->merge_reloc_tree = 1;
4535
4536         while (!list_empty(&reloc_roots)) {
4537                 reloc_root = list_entry(reloc_roots.next,
4538                                         struct btrfs_root, root_list);
4539                 list_del(&reloc_root->root_list);
4540
4541                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4542                         list_add_tail(&reloc_root->root_list,
4543                                       &rc->reloc_roots);
4544                         continue;
4545                 }
4546
4547                 fs_root = read_fs_root(fs_info, reloc_root->root_key.offset);
4548                 if (IS_ERR(fs_root)) {
4549                         err = PTR_ERR(fs_root);
4550                         goto out_free;
4551                 }
4552
4553                 err = __add_reloc_root(reloc_root);
4554                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4555                 fs_root->reloc_root = reloc_root;
4556         }
4557
4558         err = btrfs_commit_transaction(trans);
4559         if (err)
4560                 goto out_free;
4561
4562         merge_reloc_roots(rc);
4563
4564         unset_reloc_control(rc);
4565
4566         trans = btrfs_join_transaction(rc->extent_root);
4567         if (IS_ERR(trans)) {
4568                 err = PTR_ERR(trans);
4569                 goto out_free;
4570         }
4571         err = btrfs_commit_transaction(trans);
4572
4573         ret = clean_dirty_subvols(rc);
4574         if (ret < 0 && !err)
4575                 err = ret;
4576 out_free:
4577         kfree(rc);
4578 out:
4579         if (!list_empty(&reloc_roots))
4580                 free_reloc_roots(&reloc_roots);
4581
4582         btrfs_free_path(path);
4583
4584         if (err == 0) {
4585                 /* cleanup orphan inode in data relocation tree */
4586                 fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4587                 if (IS_ERR(fs_root))
4588                         err = PTR_ERR(fs_root);
4589                 else
4590                         err = btrfs_orphan_cleanup(fs_root);
4591         }
4592         return err;
4593 }
4594
4595 /*
4596  * helper to add ordered checksum for data relocation.
4597  *
4598  * cloning checksum properly handles the nodatasum extents.
4599  * it also saves CPU time to re-calculate the checksum.
4600  */
4601 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4602 {
4603         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4604         struct btrfs_ordered_sum *sums;
4605         struct btrfs_ordered_extent *ordered;
4606         int ret;
4607         u64 disk_bytenr;
4608         u64 new_bytenr;
4609         LIST_HEAD(list);
4610
4611         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4612         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4613
4614         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4615         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4616                                        disk_bytenr + len - 1, &list, 0);
4617         if (ret)
4618                 goto out;
4619
4620         while (!list_empty(&list)) {
4621                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4622                 list_del_init(&sums->list);
4623
4624                 /*
4625                  * We need to offset the new_bytenr based on where the csum is.
4626                  * We need to do this because we will read in entire prealloc
4627                  * extents but we may have written to say the middle of the
4628                  * prealloc extent, so we need to make sure the csum goes with
4629                  * the right disk offset.
4630                  *
4631                  * We can do this because the data reloc inode refers strictly
4632                  * to the on disk bytes, so we don't have to worry about
4633                  * disk_len vs real len like with real inodes since it's all
4634                  * disk length.
4635                  */
4636                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4637                 sums->bytenr = new_bytenr;
4638
4639                 btrfs_add_ordered_sum(ordered, sums);
4640         }
4641 out:
4642         btrfs_put_ordered_extent(ordered);
4643         return ret;
4644 }
4645
4646 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4647                           struct btrfs_root *root, struct extent_buffer *buf,
4648                           struct extent_buffer *cow)
4649 {
4650         struct btrfs_fs_info *fs_info = root->fs_info;
4651         struct reloc_control *rc;
4652         struct backref_node *node;
4653         int first_cow = 0;
4654         int level;
4655         int ret = 0;
4656
4657         rc = fs_info->reloc_ctl;
4658         if (!rc)
4659                 return 0;
4660
4661         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4662                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4663
4664         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4665                 if (buf == root->node)
4666                         __update_reloc_root(root, cow->start);
4667         }
4668
4669         level = btrfs_header_level(buf);
4670         if (btrfs_header_generation(buf) <=
4671             btrfs_root_last_snapshot(&root->root_item))
4672                 first_cow = 1;
4673
4674         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4675             rc->create_reloc_tree) {
4676                 WARN_ON(!first_cow && level == 0);
4677
4678                 node = rc->backref_cache.path[level];
4679                 BUG_ON(node->bytenr != buf->start &&
4680                        node->new_bytenr != buf->start);
4681
4682                 drop_node_buffer(node);
4683                 extent_buffer_get(cow);
4684                 node->eb = cow;
4685                 node->new_bytenr = cow->start;
4686
4687                 if (!node->pending) {
4688                         list_move_tail(&node->list,
4689                                        &rc->backref_cache.pending[level]);
4690                         node->pending = 1;
4691                 }
4692
4693                 if (first_cow)
4694                         __mark_block_processed(rc, node);
4695
4696                 if (first_cow && level > 0)
4697                         rc->nodes_relocated += buf->len;
4698         }
4699
4700         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4701                 ret = replace_file_extents(trans, rc, root, cow);
4702         return ret;
4703 }
4704
4705 /*
4706  * called before creating snapshot. it calculates metadata reservation
4707  * required for relocating tree blocks in the snapshot
4708  */
4709 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4710                               u64 *bytes_to_reserve)
4711 {
4712         struct btrfs_root *root = pending->root;
4713         struct reloc_control *rc = root->fs_info->reloc_ctl;
4714
4715         if (!root->reloc_root || !rc)
4716                 return;
4717
4718         if (!rc->merge_reloc_tree)
4719                 return;
4720
4721         root = root->reloc_root;
4722         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4723         /*
4724          * relocation is in the stage of merging trees. the space
4725          * used by merging a reloc tree is twice the size of
4726          * relocated tree nodes in the worst case. half for cowing
4727          * the reloc tree, half for cowing the fs tree. the space
4728          * used by cowing the reloc tree will be freed after the
4729          * tree is dropped. if we create snapshot, cowing the fs
4730          * tree may use more space than it frees. so we need
4731          * reserve extra space.
4732          */
4733         *bytes_to_reserve += rc->nodes_relocated;
4734 }
4735
4736 /*
4737  * called after snapshot is created. migrate block reservation
4738  * and create reloc root for the newly created snapshot
4739  */
4740 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4741                                struct btrfs_pending_snapshot *pending)
4742 {
4743         struct btrfs_root *root = pending->root;
4744         struct btrfs_root *reloc_root;
4745         struct btrfs_root *new_root;
4746         struct reloc_control *rc = root->fs_info->reloc_ctl;
4747         int ret;
4748
4749         if (!root->reloc_root || !rc)
4750                 return 0;
4751
4752         rc = root->fs_info->reloc_ctl;
4753         rc->merging_rsv_size += rc->nodes_relocated;
4754
4755         if (rc->merge_reloc_tree) {
4756                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4757                                               rc->block_rsv,
4758                                               rc->nodes_relocated, true);
4759                 if (ret)
4760                         return ret;
4761         }
4762
4763         new_root = pending->snap;
4764         reloc_root = create_reloc_root(trans, root->reloc_root,
4765                                        new_root->root_key.objectid);
4766         if (IS_ERR(reloc_root))
4767                 return PTR_ERR(reloc_root);
4768
4769         ret = __add_reloc_root(reloc_root);
4770         BUG_ON(ret < 0);
4771         new_root->reloc_root = reloc_root;
4772
4773         if (rc->create_reloc_tree)
4774                 ret = clone_backref_node(trans, rc, root, reloc_root);
4775         return ret;
4776 }