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
  24 /*
  25  * backref_node, mapping_node and tree_block start with this
  26  */
  27 struct tree_entry {
  28         struct rb_node rb_node;
  29         u64 bytenr;
  30 };
  31
  32 /*
  33  * present a tree block in the backref cache
  34  */
  35 struct backref_node {
  36         struct rb_node rb_node;
  37         u64 bytenr;
  38
  39         u64 new_bytenr;
  40         /* objectid of tree block owner, can be not uptodate */
  41         u64 owner;
  42         /* link to pending, changed or detached list */
  43         struct list_head list;
  44         /* list of upper level blocks reference this block */
  45         struct list_head upper;
  46         /* list of child blocks in the cache */
  47         struct list_head lower;
  48         /* NULL if this node is not tree root */
  49         struct btrfs_root *root;
  50         /* extent buffer got by COW the block */
  51         struct extent_buffer *eb;
  52         /* level of tree block */
  53         unsigned int level:8;
  54         /* is the block in non-reference counted tree */
  55         unsigned int cowonly:1;
  56         /* 1 if no child node in the cache */
  57         unsigned int lowest:1;
  58         /* is the extent buffer locked */
  59         unsigned int locked:1;
  60         /* has the block been processed */
  61         unsigned int processed:1;
  62         /* have backrefs of this block been checked */
  63         unsigned int checked:1;
  64         /*
  65          * 1 if corresponding block has been cowed but some upper
  66          * level block pointers may not point to the new location
  67          */
  68         unsigned int pending:1;
  69         /*
  70          * 1 if the backref node isn't connected to any other
  71          * backref node.
  72          */
  73         unsigned int detached:1;
  74 };
  75
  76 /*
  77  * present a block pointer in the backref cache
  78  */
  79 struct backref_edge {
  80         struct list_head list[2];
  81         struct backref_node *node[2];
  82 };
  83
  84 #define LOWER   0
  85 #define UPPER   1
  86 #define RELOCATION_RESERVED_NODES       256
  87
  88 struct backref_cache {
  89         /* red black tree of all backref nodes in the cache */
  90         struct rb_root rb_root;
  91         /* for passing backref nodes to btrfs_reloc_cow_block */
  92         struct backref_node *path[BTRFS_MAX_LEVEL];
  93         /*
  94          * list of blocks that have been cowed but some block
  95          * pointers in upper level blocks may not reflect the
  96          * new location
  97          */
  98         struct list_head pending[BTRFS_MAX_LEVEL];
  99         /* list of backref nodes with no child node */
 100         struct list_head leaves;
 101         /* list of blocks that have been cowed in current transaction */
 102         struct list_head changed;
 103         /* list of detached backref node. */
 104         struct list_head detached;
 105
 106         u64 last_trans;
 107
 108         int nr_nodes;
 109         int nr_edges;
 110 };
 111
 112 /*
 113  * map address of tree root to tree
 114  */
 115 struct mapping_node {
 116         struct rb_node rb_node;
 117         u64 bytenr;
 118         void *data;
 119 };
 120
 121 struct mapping_tree {
 122         struct rb_root rb_root;
 123         spinlock_t lock;
 124 };
 125
 126 /*
 127  * present a tree block to process
 128  */
 129 struct tree_block {
 130         struct rb_node rb_node;
 131         u64 bytenr;
 132         struct btrfs_key key;
 133         unsigned int level:8;
 134         unsigned int key_ready:1;
 135 };
 136
 137 #define MAX_EXTENTS 128
 138
 139 struct file_extent_cluster {
 140         u64 start;
 141         u64 end;
 142         u64 boundary[MAX_EXTENTS];
 143         unsigned int nr;
 144 };
 145
 146 struct reloc_control {
 147         /* block group to relocate */
 148         struct btrfs_block_group_cache *block_group;
 149         /* extent tree */
 150         struct btrfs_root *extent_root;
 151         /* inode for moving data */
 152         struct inode *data_inode;
 153
 154         struct btrfs_block_rsv *block_rsv;
 155
 156         struct backref_cache backref_cache;
 157
 158         struct file_extent_cluster cluster;
 159         /* tree blocks have been processed */
 160         struct extent_io_tree processed_blocks;
 161         /* map start of tree root to corresponding reloc tree */
 162         struct mapping_tree reloc_root_tree;
 163         /* list of reloc trees */
 164         struct list_head reloc_roots;
 165         /* size of metadata reservation for merging reloc trees */
 166         u64 merging_rsv_size;
 167         /* size of relocated tree nodes */
 168         u64 nodes_relocated;
 169         /* reserved size for block group relocation*/
 170         u64 reserved_bytes;
 171
 172         u64 search_start;
 173         u64 extents_found;
 174
 175         unsigned int stage:8;
 176         unsigned int create_reloc_tree:1;
 177         unsigned int merge_reloc_tree:1;
 178         unsigned int found_file_extent:1;
 179 };
 180
 181 /* stages of data relocation */
 182 #define MOVE_DATA_EXTENTS       0
 183 #define UPDATE_DATA_PTRS        1
 184
 185 static void remove_backref_node(struct backref_cache *cache,
 186                                 struct backref_node *node);
 187 static void __mark_block_processed(struct reloc_control *rc,
 188                                    struct backref_node *node);
 189
 190 static void mapping_tree_init(struct mapping_tree *tree)
 191 {
 192         tree->rb_root = RB_ROOT;
 193         spin_lock_init(&tree->lock);
 194 }
 195
 196 static void backref_cache_init(struct backref_cache *cache)
 197 {
 198         int i;
 199         cache->rb_root = RB_ROOT;
 200         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 201                 INIT_LIST_HEAD(&cache->pending[i]);
 202         INIT_LIST_HEAD(&cache->changed);
 203         INIT_LIST_HEAD(&cache->detached);
 204         INIT_LIST_HEAD(&cache->leaves);
 205 }
 206
 207 static void backref_cache_cleanup(struct backref_cache *cache)
 208 {
 209         struct backref_node *node;
 210         int i;
 211
 212         while (!list_empty(&cache->detached)) {
 213                 node = list_entry(cache->detached.next,
 214                                   struct backref_node, list);
 215                 remove_backref_node(cache, node);
 216         }
 217
 218         while (!list_empty(&cache->leaves)) {
 219                 node = list_entry(cache->leaves.next,
 220                                   struct backref_node, lower);
 221                 remove_backref_node(cache, node);
 222         }
 223
 224         cache->last_trans = 0;
 225
 226         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 227                 ASSERT(list_empty(&cache->pending[i]));
 228         ASSERT(list_empty(&cache->changed));
 229         ASSERT(list_empty(&cache->detached));
 230         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
 231         ASSERT(!cache->nr_nodes);
 232         ASSERT(!cache->nr_edges);
 233 }
 234
 235 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 236 {
 237         struct backref_node *node;
 238
 239         node = kzalloc(sizeof(*node), GFP_NOFS);
 240         if (node) {
 241                 INIT_LIST_HEAD(&node->list);
 242                 INIT_LIST_HEAD(&node->upper);
 243                 INIT_LIST_HEAD(&node->lower);
 244                 RB_CLEAR_NODE(&node->rb_node);
 245                 cache->nr_nodes++;
 246         }
 247         return node;
 248 }
 249
 250 static void free_backref_node(struct backref_cache *cache,
 251                               struct backref_node *node)
 252 {
 253         if (node) {
 254                 cache->nr_nodes--;
 255                 kfree(node);
 256         }
 257 }
 258
 259 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 260 {
 261         struct backref_edge *edge;
 262
 263         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 264         if (edge)
 265                 cache->nr_edges++;
 266         return edge;
 267 }
 268
 269 static void free_backref_edge(struct backref_cache *cache,
 270                               struct backref_edge *edge)
 271 {
 272         if (edge) {
 273                 cache->nr_edges--;
 274                 kfree(edge);
 275         }
 276 }
 277
 278 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 279                                    struct rb_node *node)
 280 {
 281         struct rb_node **p = &root->rb_node;
 282         struct rb_node *parent = NULL;
 283         struct tree_entry *entry;
 284
 285         while (*p) {
 286                 parent = *p;
 287                 entry = rb_entry(parent, struct tree_entry, rb_node);
 288
 289                 if (bytenr < entry->bytenr)
 290                         p = &(*p)->rb_left;
 291                 else if (bytenr > entry->bytenr)
 292                         p = &(*p)->rb_right;
 293                 else
 294                         return parent;
 295         }
 296
 297         rb_link_node(node, parent, p);
 298         rb_insert_color(node, root);
 299         return NULL;
 300 }
 301
 302 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 303 {
 304         struct rb_node *n = root->rb_node;
 305         struct tree_entry *entry;
 306
 307         while (n) {
 308                 entry = rb_entry(n, struct tree_entry, rb_node);
 309
 310                 if (bytenr < entry->bytenr)
 311                         n = n->rb_left;
 312                 else if (bytenr > entry->bytenr)
 313                         n = n->rb_right;
 314                 else
 315                         return n;
 316         }
 317         return NULL;
 318 }
 319
 320 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 321 {
 322
 323         struct btrfs_fs_info *fs_info = NULL;
 324         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 325                                               rb_node);
 326         if (bnode->root)
 327                 fs_info = bnode->root->fs_info;
 328         btrfs_panic(fs_info, errno,
 329                     "Inconsistency in backref cache found at offset %llu",
 330                     bytenr);
 331 }
 332
 333 /*
 334  * walk up backref nodes until reach node presents tree root
 335  */
 336 static struct backref_node *walk_up_backref(struct backref_node *node,
 337                                             struct backref_edge *edges[],
 338                                             int *index)
 339 {
 340         struct backref_edge *edge;
 341         int idx = *index;
 342
 343         while (!list_empty(&node->upper)) {
 344                 edge = list_entry(node->upper.next,
 345                                   struct backref_edge, list[LOWER]);
 346                 edges[idx++] = edge;
 347                 node = edge->node[UPPER];
 348         }
 349         BUG_ON(node->detached);
 350         *index = idx;
 351         return node;
 352 }
 353
 354 /*
 355  * walk down backref nodes to find start of next reference path
 356  */
 357 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 358                                               int *index)
 359 {
 360         struct backref_edge *edge;
 361         struct backref_node *lower;
 362         int idx = *index;
 363
 364         while (idx > 0) {
 365                 edge = edges[idx - 1];
 366                 lower = edge->node[LOWER];
 367                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 368                         idx--;
 369                         continue;
 370                 }
 371                 edge = list_entry(edge->list[LOWER].next,
 372                                   struct backref_edge, list[LOWER]);
 373                 edges[idx - 1] = edge;
 374                 *index = idx;
 375                 return edge->node[UPPER];
 376         }
 377         *index = 0;
 378         return NULL;
 379 }
 380
 381 static void unlock_node_buffer(struct backref_node *node)
 382 {
 383         if (node->locked) {
 384                 btrfs_tree_unlock(node->eb);
 385                 node->locked = 0;
 386         }
 387 }
 388
 389 static void drop_node_buffer(struct backref_node *node)
 390 {
 391         if (node->eb) {
 392                 unlock_node_buffer(node);
 393                 free_extent_buffer(node->eb);
 394                 node->eb = NULL;
 395         }
 396 }
 397
 398 static void drop_backref_node(struct backref_cache *tree,
 399                               struct backref_node *node)
 400 {
 401         BUG_ON(!list_empty(&node->upper));
 402
 403         drop_node_buffer(node);
 404         list_del(&node->list);
 405         list_del(&node->lower);
 406         if (!RB_EMPTY_NODE(&node->rb_node))
 407                 rb_erase(&node->rb_node, &tree->rb_root);
 408         free_backref_node(tree, node);
 409 }
 410
 411 /*
 412  * remove a backref node from the backref cache
 413  */
 414 static void remove_backref_node(struct backref_cache *cache,
 415                                 struct backref_node *node)
 416 {
 417         struct backref_node *upper;
 418         struct backref_edge *edge;
 419
 420         if (!node)
 421                 return;
 422
 423         BUG_ON(!node->lowest && !node->detached);
 424         while (!list_empty(&node->upper)) {
 425                 edge = list_entry(node->upper.next, struct backref_edge,
 426                                   list[LOWER]);
 427                 upper = edge->node[UPPER];
 428                 list_del(&edge->list[LOWER]);
 429                 list_del(&edge->list[UPPER]);
 430                 free_backref_edge(cache, edge);
 431
 432                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 433                         BUG_ON(!list_empty(&node->upper));
 434                         drop_backref_node(cache, node);
 435                         node = upper;
 436                         node->lowest = 1;
 437                         continue;
 438                 }
 439                 /*
 440                  * add the node to leaf node list if no other
 441                  * child block cached.
 442                  */
 443                 if (list_empty(&upper->lower)) {
 444                         list_add_tail(&upper->lower, &cache->leaves);
 445                         upper->lowest = 1;
 446                 }
 447         }
 448
 449         drop_backref_node(cache, node);
 450 }
 451
 452 static void update_backref_node(struct backref_cache *cache,
 453                                 struct backref_node *node, u64 bytenr)
 454 {
 455         struct rb_node *rb_node;
 456         rb_erase(&node->rb_node, &cache->rb_root);
 457         node->bytenr = bytenr;
 458         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 459         if (rb_node)
 460                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 461 }
 462
 463 /*
 464  * update backref cache after a transaction commit
 465  */
 466 static int update_backref_cache(struct btrfs_trans_handle *trans,
 467                                 struct backref_cache *cache)
 468 {
 469         struct backref_node *node;
 470         int level = 0;
 471
 472         if (cache->last_trans == 0) {
 473                 cache->last_trans = trans->transid;
 474                 return 0;
 475         }
 476
 477         if (cache->last_trans == trans->transid)
 478                 return 0;
 479
 480         /*
 481          * detached nodes are used to avoid unnecessary backref
 482          * lookup. transaction commit changes the extent tree.
 483          * so the detached nodes are no longer useful.
 484          */
 485         while (!list_empty(&cache->detached)) {
 486                 node = list_entry(cache->detached.next,
 487                                   struct backref_node, list);
 488                 remove_backref_node(cache, node);
 489         }
 490
 491         while (!list_empty(&cache->changed)) {
 492                 node = list_entry(cache->changed.next,
 493                                   struct backref_node, list);
 494                 list_del_init(&node->list);
 495                 BUG_ON(node->pending);
 496                 update_backref_node(cache, node, node->new_bytenr);
 497         }
 498
 499         /*
 500          * some nodes can be left in the pending list if there were
 501          * errors during processing the pending nodes.
 502          */
 503         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 504                 list_for_each_entry(node, &cache->pending[level], list) {
 505                         BUG_ON(!node->pending);
 506                         if (node->bytenr == node->new_bytenr)
 507                                 continue;
 508                         update_backref_node(cache, node, node->new_bytenr);
 509                 }
 510         }
 511
 512         cache->last_trans = 0;
 513         return 1;
 514 }
 515
 516
 517 static int should_ignore_root(struct btrfs_root *root)
 518 {
 519         struct btrfs_root *reloc_root;
 520
 521         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 522                 return 0;
 523
 524         reloc_root = root->reloc_root;
 525         if (!reloc_root)
 526                 return 0;
 527
 528         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 529             root->fs_info->running_transaction->transid - 1)
 530                 return 0;
 531         /*
 532          * if there is reloc tree and it was created in previous
 533          * transaction backref lookup can find the reloc tree,
 534          * so backref node for the fs tree root is useless for
 535          * relocation.
 536          */
 537         return 1;
 538 }
 539 /*
 540  * find reloc tree by address of tree root
 541  */
 542 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 543                                           u64 bytenr)
 544 {
 545         struct rb_node *rb_node;
 546         struct mapping_node *node;
 547         struct btrfs_root *root = NULL;
 548
 549         spin_lock(&rc->reloc_root_tree.lock);
 550         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 551         if (rb_node) {
 552                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 553                 root = (struct btrfs_root *)node->data;
 554         }
 555         spin_unlock(&rc->reloc_root_tree.lock);
 556         return root;
 557 }
 558
 559 static int is_cowonly_root(u64 root_objectid)
 560 {
 561         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 562             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 563             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 564             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 565             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 566             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 567             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 568             root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
 569             root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
 570                 return 1;
 571         return 0;
 572 }
 573
 574 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 575                                         u64 root_objectid)
 576 {
 577         struct btrfs_key key;
 578
 579         key.objectid = root_objectid;
 580         key.type = BTRFS_ROOT_ITEM_KEY;
 581         if (is_cowonly_root(root_objectid))
 582                 key.offset = 0;
 583         else
 584                 key.offset = (u64)-1;
 585
 586         return btrfs_get_fs_root(fs_info, &key, false);
 587 }
 588
 589 static noinline_for_stack
 590 int find_inline_backref(struct extent_buffer *leaf, int slot,
 591                         unsigned long *ptr, unsigned long *end)
 592 {
 593         struct btrfs_key key;
 594         struct btrfs_extent_item *ei;
 595         struct btrfs_tree_block_info *bi;
 596         u32 item_size;
 597
 598         btrfs_item_key_to_cpu(leaf, &key, slot);
 599
 600         item_size = btrfs_item_size_nr(leaf, slot);
 601         if (item_size < sizeof(*ei)) {
 602                 btrfs_print_v0_err(leaf->fs_info);
 603                 btrfs_handle_fs_error(leaf->fs_info, -EINVAL, NULL);
 604                 return 1;
 605         }
 606         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 607         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 608                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 609
 610         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 611             item_size <= sizeof(*ei) + sizeof(*bi)) {
 612                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 613                 return 1;
 614         }
 615         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 616             item_size <= sizeof(*ei)) {
 617                 WARN_ON(item_size < sizeof(*ei));
 618                 return 1;
 619         }
 620
 621         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 622                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 623                 *ptr = (unsigned long)(bi + 1);
 624         } else {
 625                 *ptr = (unsigned long)(ei + 1);
 626         }
 627         *end = (unsigned long)ei + item_size;
 628         return 0;
 629 }
 630
 631 /*
 632  * build backref tree for a given tree block. root of the backref tree
 633  * corresponds the tree block, leaves of the backref tree correspond
 634  * roots of b-trees that reference the tree block.
 635  *
 636  * the basic idea of this function is check backrefs of a given block
 637  * to find upper level blocks that reference the block, and then check
 638  * backrefs of these upper level blocks recursively. the recursion stop
 639  * when tree root is reached or backrefs for the block is cached.
 640  *
 641  * NOTE: if we find backrefs for a block are cached, we know backrefs
 642  * for all upper level blocks that directly/indirectly reference the
 643  * block are also cached.
 644  */
 645 static noinline_for_stack
 646 struct backref_node *build_backref_tree(struct reloc_control *rc,
 647                                         struct btrfs_key *node_key,
 648                                         int level, u64 bytenr)
 649 {
 650         struct backref_cache *cache = &rc->backref_cache;
 651         struct btrfs_path *path1;
 652         struct btrfs_path *path2;
 653         struct extent_buffer *eb;
 654         struct btrfs_root *root;
 655         struct backref_node *cur;
 656         struct backref_node *upper;
 657         struct backref_node *lower;
 658         struct backref_node *node = NULL;
 659         struct backref_node *exist = NULL;
 660         struct backref_edge *edge;
 661         struct rb_node *rb_node;
 662         struct btrfs_key key;
 663         unsigned long end;
 664         unsigned long ptr;
 665         LIST_HEAD(list);
 666         LIST_HEAD(useless);
 667         int cowonly;
 668         int ret;
 669         int err = 0;
 670         bool need_check = true;
 671
 672         path1 = btrfs_alloc_path();
 673         path2 = btrfs_alloc_path();
 674         if (!path1 || !path2) {
 675                 err = -ENOMEM;
 676                 goto out;
 677         }
 678         path1->reada = READA_FORWARD;
 679         path2->reada = READA_FORWARD;
 680
 681         node = alloc_backref_node(cache);
 682         if (!node) {
 683                 err = -ENOMEM;
 684                 goto out;
 685         }
 686
 687         node->bytenr = bytenr;
 688         node->level = level;
 689         node->lowest = 1;
 690         cur = node;
 691 again:
 692         end = 0;
 693         ptr = 0;
 694         key.objectid = cur->bytenr;
 695         key.type = BTRFS_METADATA_ITEM_KEY;
 696         key.offset = (u64)-1;
 697
 698         path1->search_commit_root = 1;
 699         path1->skip_locking = 1;
 700         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 701                                 0, 0);
 702         if (ret < 0) {
 703                 err = ret;
 704                 goto out;
 705         }
 706         ASSERT(ret);
 707         ASSERT(path1->slots[0]);
 708
 709         path1->slots[0]--;
 710
 711         WARN_ON(cur->checked);
 712         if (!list_empty(&cur->upper)) {
 713                 /*
 714                  * the backref was added previously when processing
 715                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 716                  */
 717                 ASSERT(list_is_singular(&cur->upper));
 718                 edge = list_entry(cur->upper.next, struct backref_edge,
 719                                   list[LOWER]);
 720                 ASSERT(list_empty(&edge->list[UPPER]));
 721                 exist = edge->node[UPPER];
 722                 /*
 723                  * add the upper level block to pending list if we need
 724                  * check its backrefs
 725                  */
 726                 if (!exist->checked)
 727                         list_add_tail(&edge->list[UPPER], &list);
 728         } else {
 729                 exist = NULL;
 730         }
 731
 732         while (1) {
 733                 cond_resched();
 734                 eb = path1->nodes[0];
 735
 736                 if (ptr >= end) {
 737                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 738                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 739                                 if (ret < 0) {
 740                                         err = ret;
 741                                         goto out;
 742                                 }
 743                                 if (ret > 0)
 744                                         break;
 745                                 eb = path1->nodes[0];
 746                         }
 747
 748                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 749                         if (key.objectid != cur->bytenr) {
 750                                 WARN_ON(exist);
 751                                 break;
 752                         }
 753
 754                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 755                             key.type == BTRFS_METADATA_ITEM_KEY) {
 756                                 ret = find_inline_backref(eb, path1->slots[0],
 757                                                           &ptr, &end);
 758                                 if (ret)
 759                                         goto next;
 760                         }
 761                 }
 762
 763                 if (ptr < end) {
 764                         /* update key for inline back ref */
 765                         struct btrfs_extent_inline_ref *iref;
 766                         int type;
 767                         iref = (struct btrfs_extent_inline_ref *)ptr;
 768                         type = btrfs_get_extent_inline_ref_type(eb, iref,
 769                                                         BTRFS_REF_TYPE_BLOCK);
 770                         if (type == BTRFS_REF_TYPE_INVALID) {
 771                                 err = -EUCLEAN;
 772                                 goto out;
 773                         }
 774                         key.type = type;
 775                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 776
 777                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 778                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 779                 }
 780
 781                 if (exist &&
 782                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 783                       exist->owner == key.offset) ||
 784                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 785                       exist->bytenr == key.offset))) {
 786                         exist = NULL;
 787                         goto next;
 788                 }
 789
 790                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 791                         if (key.objectid == key.offset) {
 792                                 /*
 793                                  * only root blocks of reloc trees use
 794                                  * backref of this type.
 795                                  */
 796                                 root = find_reloc_root(rc, cur->bytenr);
 797                                 ASSERT(root);
 798                                 cur->root = root;
 799                                 break;
 800                         }
 801
 802                         edge = alloc_backref_edge(cache);
 803                         if (!edge) {
 804                                 err = -ENOMEM;
 805                                 goto out;
 806                         }
 807                         rb_node = tree_search(&cache->rb_root, key.offset);
 808                         if (!rb_node) {
 809                                 upper = alloc_backref_node(cache);
 810                                 if (!upper) {
 811                                         free_backref_edge(cache, edge);
 812                                         err = -ENOMEM;
 813                                         goto out;
 814                                 }
 815                                 upper->bytenr = key.offset;
 816                                 upper->level = cur->level + 1;
 817                                 /*
 818                                  *  backrefs for the upper level block isn't
 819                                  *  cached, add the block to pending list
 820                                  */
 821                                 list_add_tail(&edge->list[UPPER], &list);
 822                         } else {
 823                                 upper = rb_entry(rb_node, struct backref_node,
 824                                                  rb_node);
 825                                 ASSERT(upper->checked);
 826                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 827                         }
 828                         list_add_tail(&edge->list[LOWER], &cur->upper);
 829                         edge->node[LOWER] = cur;
 830                         edge->node[UPPER] = upper;
 831
 832                         goto next;
 833                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 834                         err = -EINVAL;
 835                         btrfs_print_v0_err(rc->extent_root->fs_info);
 836                         btrfs_handle_fs_error(rc->extent_root->fs_info, err,
 837                                               NULL);
 838                         goto out;
 839                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 840                         goto next;
 841                 }
 842
 843                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 844                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 845                 if (IS_ERR(root)) {
 846                         err = PTR_ERR(root);
 847                         goto out;
 848                 }
 849
 850                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 851                         cur->cowonly = 1;
 852
 853                 if (btrfs_root_level(&root->root_item) == cur->level) {
 854                         /* tree root */
 855                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
 856                                cur->bytenr);
 857                         if (should_ignore_root(root))
 858                                 list_add(&cur->list, &useless);
 859                         else
 860                                 cur->root = root;
 861                         break;
 862                 }
 863
 864                 level = cur->level + 1;
 865
 866                 /*
 867                  * searching the tree to find upper level blocks
 868                  * reference the block.
 869                  */
 870                 path2->search_commit_root = 1;
 871                 path2->skip_locking = 1;
 872                 path2->lowest_level = level;
 873                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 874                 path2->lowest_level = 0;
 875                 if (ret < 0) {
 876                         err = ret;
 877                         goto out;
 878                 }
 879                 if (ret > 0 && path2->slots[level] > 0)
 880                         path2->slots[level]--;
 881
 882                 eb = path2->nodes[level];
 883                 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
 884                     cur->bytenr) {
 885                         btrfs_err(root->fs_info,
 886         "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
 887                                   cur->bytenr, level - 1,
 888                                   root->root_key.objectid,
 889                                   node_key->objectid, node_key->type,
 890                                   node_key->offset);
 891                         err = -ENOENT;
 892                         goto out;
 893                 }
 894                 lower = cur;
 895                 need_check = true;
 896                 for (; level < BTRFS_MAX_LEVEL; level++) {
 897                         if (!path2->nodes[level]) {
 898                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
 899                                        lower->bytenr);
 900                                 if (should_ignore_root(root))
 901                                         list_add(&lower->list, &useless);
 902                                 else
 903                                         lower->root = root;
 904                                 break;
 905                         }
 906
 907                         edge = alloc_backref_edge(cache);
 908                         if (!edge) {
 909                                 err = -ENOMEM;
 910                                 goto out;
 911                         }
 912
 913                         eb = path2->nodes[level];
 914                         rb_node = tree_search(&cache->rb_root, eb->start);
 915                         if (!rb_node) {
 916                                 upper = alloc_backref_node(cache);
 917                                 if (!upper) {
 918                                         free_backref_edge(cache, edge);
 919                                         err = -ENOMEM;
 920                                         goto out;
 921                                 }
 922                                 upper->bytenr = eb->start;
 923                                 upper->owner = btrfs_header_owner(eb);
 924                                 upper->level = lower->level + 1;
 925                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
 926                                               &root->state))
 927                                         upper->cowonly = 1;
 928
 929                                 /*
 930                                  * if we know the block isn't shared
 931                                  * we can void checking its backrefs.
 932                                  */
 933                                 if (btrfs_block_can_be_shared(root, eb))
 934                                         upper->checked = 0;
 935                                 else
 936                                         upper->checked = 1;
 937
 938                                 /*
 939                                  * add the block to pending list if we
 940                                  * need check its backrefs, we only do this once
 941                                  * while walking up a tree as we will catch
 942                                  * anything else later on.
 943                                  */
 944                                 if (!upper->checked && need_check) {
 945                                         need_check = false;
 946                                         list_add_tail(&edge->list[UPPER],
 947                                                       &list);
 948                                 } else {
 949                                         if (upper->checked)
 950                                                 need_check = true;
 951                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 952                                 }
 953                         } else {
 954                                 upper = rb_entry(rb_node, struct backref_node,
 955                                                  rb_node);
 956                                 ASSERT(upper->checked);
 957                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 958                                 if (!upper->owner)
 959                                         upper->owner = btrfs_header_owner(eb);
 960                         }
 961                         list_add_tail(&edge->list[LOWER], &lower->upper);
 962                         edge->node[LOWER] = lower;
 963                         edge->node[UPPER] = upper;
 964
 965                         if (rb_node)
 966                                 break;
 967                         lower = upper;
 968                         upper = NULL;
 969                 }
 970                 btrfs_release_path(path2);
 971 next:
 972                 if (ptr < end) {
 973                         ptr += btrfs_extent_inline_ref_size(key.type);
 974                         if (ptr >= end) {
 975                                 WARN_ON(ptr > end);
 976                                 ptr = 0;
 977                                 end = 0;
 978                         }
 979                 }
 980                 if (ptr >= end)
 981                         path1->slots[0]++;
 982         }
 983         btrfs_release_path(path1);
 984
 985         cur->checked = 1;
 986         WARN_ON(exist);
 987
 988         /* the pending list isn't empty, take the first block to process */
 989         if (!list_empty(&list)) {
 990                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
 991                 list_del_init(&edge->list[UPPER]);
 992                 cur = edge->node[UPPER];
 993                 goto again;
 994         }
 995
 996         /*
 997          * everything goes well, connect backref nodes and insert backref nodes
 998          * into the cache.
 999          */
1000         ASSERT(node->checked);
1001         cowonly = node->cowonly;
1002         if (!cowonly) {
1003                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1004                                       &node->rb_node);
1005                 if (rb_node)
1006                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1007                 list_add_tail(&node->lower, &cache->leaves);
1008         }
1009
1010         list_for_each_entry(edge, &node->upper, list[LOWER])
1011                 list_add_tail(&edge->list[UPPER], &list);
1012
1013         while (!list_empty(&list)) {
1014                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1015                 list_del_init(&edge->list[UPPER]);
1016                 upper = edge->node[UPPER];
1017                 if (upper->detached) {
1018                         list_del(&edge->list[LOWER]);
1019                         lower = edge->node[LOWER];
1020                         free_backref_edge(cache, edge);
1021                         if (list_empty(&lower->upper))
1022                                 list_add(&lower->list, &useless);
1023                         continue;
1024                 }
1025
1026                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1027                         if (upper->lowest) {
1028                                 list_del_init(&upper->lower);
1029                                 upper->lowest = 0;
1030                         }
1031
1032                         list_add_tail(&edge->list[UPPER], &upper->lower);
1033                         continue;
1034                 }
1035
1036                 if (!upper->checked) {
1037                         /*
1038                          * Still want to blow up for developers since this is a
1039                          * logic bug.
1040                          */
1041                         ASSERT(0);
1042                         err = -EINVAL;
1043                         goto out;
1044                 }
1045                 if (cowonly != upper->cowonly) {
1046                         ASSERT(0);
1047                         err = -EINVAL;
1048                         goto out;
1049                 }
1050
1051                 if (!cowonly) {
1052                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1053                                               &upper->rb_node);
1054                         if (rb_node)
1055                                 backref_tree_panic(rb_node, -EEXIST,
1056                                                    upper->bytenr);
1057                 }
1058
1059                 list_add_tail(&edge->list[UPPER], &upper->lower);
1060
1061                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1062                         list_add_tail(&edge->list[UPPER], &list);
1063         }
1064         /*
1065          * process useless backref nodes. backref nodes for tree leaves
1066          * are deleted from the cache. backref nodes for upper level
1067          * tree blocks are left in the cache to avoid unnecessary backref
1068          * lookup.
1069          */
1070         while (!list_empty(&useless)) {
1071                 upper = list_entry(useless.next, struct backref_node, list);
1072                 list_del_init(&upper->list);
1073                 ASSERT(list_empty(&upper->upper));
1074                 if (upper == node)
1075                         node = NULL;
1076                 if (upper->lowest) {
1077                         list_del_init(&upper->lower);
1078                         upper->lowest = 0;
1079                 }
1080                 while (!list_empty(&upper->lower)) {
1081                         edge = list_entry(upper->lower.next,
1082                                           struct backref_edge, list[UPPER]);
1083                         list_del(&edge->list[UPPER]);
1084                         list_del(&edge->list[LOWER]);
1085                         lower = edge->node[LOWER];
1086                         free_backref_edge(cache, edge);
1087
1088                         if (list_empty(&lower->upper))
1089                                 list_add(&lower->list, &useless);
1090                 }
1091                 __mark_block_processed(rc, upper);
1092                 if (upper->level > 0) {
1093                         list_add(&upper->list, &cache->detached);
1094                         upper->detached = 1;
1095                 } else {
1096                         rb_erase(&upper->rb_node, &cache->rb_root);
1097                         free_backref_node(cache, upper);
1098                 }
1099         }
1100 out:
1101         btrfs_free_path(path1);
1102         btrfs_free_path(path2);
1103         if (err) {
1104                 while (!list_empty(&useless)) {
1105                         lower = list_entry(useless.next,
1106                                            struct backref_node, list);
1107                         list_del_init(&lower->list);
1108                 }
1109                 while (!list_empty(&list)) {
1110                         edge = list_first_entry(&list, struct backref_edge,
1111                                                 list[UPPER]);
1112                         list_del(&edge->list[UPPER]);
1113                         list_del(&edge->list[LOWER]);
1114                         lower = edge->node[LOWER];
1115                         upper = edge->node[UPPER];
1116                         free_backref_edge(cache, edge);
1117
1118                         /*
1119                          * Lower is no longer linked to any upper backref nodes
1120                          * and isn't in the cache, we can free it ourselves.
1121                          */
1122                         if (list_empty(&lower->upper) &&
1123                             RB_EMPTY_NODE(&lower->rb_node))
1124                                 list_add(&lower->list, &useless);
1125
1126                         if (!RB_EMPTY_NODE(&upper->rb_node))
1127                                 continue;
1128
1129                         /* Add this guy's upper edges to the list to process */
1130                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1131                                 list_add_tail(&edge->list[UPPER], &list);
1132                         if (list_empty(&upper->upper))
1133                                 list_add(&upper->list, &useless);
1134                 }
1135
1136                 while (!list_empty(&useless)) {
1137                         lower = list_entry(useless.next,
1138                                            struct backref_node, list);
1139                         list_del_init(&lower->list);
1140                         if (lower == node)
1141                                 node = NULL;
1142                         free_backref_node(cache, lower);
1143                 }
1144
1145                 free_backref_node(cache, node);
1146                 return ERR_PTR(err);
1147         }
1148         ASSERT(!node || !node->detached);
1149         return node;
1150 }
1151
1152 /*
1153  * helper to add backref node for the newly created snapshot.
1154  * the backref node is created by cloning backref node that
1155  * corresponds to root of source tree
1156  */
1157 static int clone_backref_node(struct btrfs_trans_handle *trans,
1158                               struct reloc_control *rc,
1159                               struct btrfs_root *src,
1160                               struct btrfs_root *dest)
1161 {
1162         struct btrfs_root *reloc_root = src->reloc_root;
1163         struct backref_cache *cache = &rc->backref_cache;
1164         struct backref_node *node = NULL;
1165         struct backref_node *new_node;
1166         struct backref_edge *edge;
1167         struct backref_edge *new_edge;
1168         struct rb_node *rb_node;
1169
1170         if (cache->last_trans > 0)
1171                 update_backref_cache(trans, cache);
1172
1173         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1174         if (rb_node) {
1175                 node = rb_entry(rb_node, struct backref_node, rb_node);
1176                 if (node->detached)
1177                         node = NULL;
1178                 else
1179                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1180         }
1181
1182         if (!node) {
1183                 rb_node = tree_search(&cache->rb_root,
1184                                       reloc_root->commit_root->start);
1185                 if (rb_node) {
1186                         node = rb_entry(rb_node, struct backref_node,
1187                                         rb_node);
1188                         BUG_ON(node->detached);
1189                 }
1190         }
1191
1192         if (!node)
1193                 return 0;
1194
1195         new_node = alloc_backref_node(cache);
1196         if (!new_node)
1197                 return -ENOMEM;
1198
1199         new_node->bytenr = dest->node->start;
1200         new_node->level = node->level;
1201         new_node->lowest = node->lowest;
1202         new_node->checked = 1;
1203         new_node->root = dest;
1204
1205         if (!node->lowest) {
1206                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1207                         new_edge = alloc_backref_edge(cache);
1208                         if (!new_edge)
1209                                 goto fail;
1210
1211                         new_edge->node[UPPER] = new_node;
1212                         new_edge->node[LOWER] = edge->node[LOWER];
1213                         list_add_tail(&new_edge->list[UPPER],
1214                                       &new_node->lower);
1215                 }
1216         } else {
1217                 list_add_tail(&new_node->lower, &cache->leaves);
1218         }
1219
1220         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1221                               &new_node->rb_node);
1222         if (rb_node)
1223                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1224
1225         if (!new_node->lowest) {
1226                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1227                         list_add_tail(&new_edge->list[LOWER],
1228                                       &new_edge->node[LOWER]->upper);
1229                 }
1230         }
1231         return 0;
1232 fail:
1233         while (!list_empty(&new_node->lower)) {
1234                 new_edge = list_entry(new_node->lower.next,
1235                                       struct backref_edge, list[UPPER]);
1236                 list_del(&new_edge->list[UPPER]);
1237                 free_backref_edge(cache, new_edge);
1238         }
1239         free_backref_node(cache, new_node);
1240         return -ENOMEM;
1241 }
1242
1243 /*
1244  * helper to add 'address of tree root -> reloc tree' mapping
1245  */
1246 static int __must_check __add_reloc_root(struct btrfs_root *root)
1247 {
1248         struct btrfs_fs_info *fs_info = root->fs_info;
1249         struct rb_node *rb_node;
1250         struct mapping_node *node;
1251         struct reloc_control *rc = fs_info->reloc_ctl;
1252
1253         node = kmalloc(sizeof(*node), GFP_NOFS);
1254         if (!node)
1255                 return -ENOMEM;
1256
1257         node->bytenr = root->node->start;
1258         node->data = root;
1259
1260         spin_lock(&rc->reloc_root_tree.lock);
1261         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1262                               node->bytenr, &node->rb_node);
1263         spin_unlock(&rc->reloc_root_tree.lock);
1264         if (rb_node) {
1265                 btrfs_panic(fs_info, -EEXIST,
1266                             "Duplicate root found for start=%llu while inserting into relocation tree",
1267                             node->bytenr);
1268         }
1269
1270         list_add_tail(&root->root_list, &rc->reloc_roots);
1271         return 0;
1272 }
1273
1274 /*
1275  * helper to delete the 'address of tree root -> reloc tree'
1276  * mapping
1277  */
1278 static void __del_reloc_root(struct btrfs_root *root)
1279 {
1280         struct btrfs_fs_info *fs_info = root->fs_info;
1281         struct rb_node *rb_node;
1282         struct mapping_node *node = NULL;
1283         struct reloc_control *rc = fs_info->reloc_ctl;
1284
1285         if (rc && root->node) {
1286                 spin_lock(&rc->reloc_root_tree.lock);
1287                 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1288                                       root->node->start);
1289                 if (rb_node) {
1290                         node = rb_entry(rb_node, struct mapping_node, rb_node);
1291                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1292                 }
1293                 spin_unlock(&rc->reloc_root_tree.lock);
1294                 if (!node)
1295                         return;
1296                 BUG_ON((struct btrfs_root *)node->data != root);
1297         }
1298
1299         spin_lock(&fs_info->trans_lock);
1300         list_del_init(&root->root_list);
1301         spin_unlock(&fs_info->trans_lock);
1302         kfree(node);
1303 }
1304
1305 /*
1306  * helper to update the 'address of tree root -> reloc tree'
1307  * mapping
1308  */
1309 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1310 {
1311         struct btrfs_fs_info *fs_info = root->fs_info;
1312         struct rb_node *rb_node;
1313         struct mapping_node *node = NULL;
1314         struct reloc_control *rc = fs_info->reloc_ctl;
1315
1316         spin_lock(&rc->reloc_root_tree.lock);
1317         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1318                               root->node->start);
1319         if (rb_node) {
1320                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1321                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1322         }
1323         spin_unlock(&rc->reloc_root_tree.lock);
1324
1325         if (!node)
1326                 return 0;
1327         BUG_ON((struct btrfs_root *)node->data != root);
1328
1329         spin_lock(&rc->reloc_root_tree.lock);
1330         node->bytenr = new_bytenr;
1331         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1332                               node->bytenr, &node->rb_node);
1333         spin_unlock(&rc->reloc_root_tree.lock);
1334         if (rb_node)
1335                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1336         return 0;
1337 }
1338
1339 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1340                                         struct btrfs_root *root, u64 objectid)
1341 {
1342         struct btrfs_fs_info *fs_info = root->fs_info;
1343         struct btrfs_root *reloc_root;
1344         struct extent_buffer *eb;
1345         struct btrfs_root_item *root_item;
1346         struct btrfs_key root_key;
1347         int ret;
1348
1349         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1350         BUG_ON(!root_item);
1351
1352         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1353         root_key.type = BTRFS_ROOT_ITEM_KEY;
1354         root_key.offset = objectid;
1355
1356         if (root->root_key.objectid == objectid) {
1357                 u64 commit_root_gen;
1358
1359                 /* called by btrfs_init_reloc_root */
1360                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1361                                       BTRFS_TREE_RELOC_OBJECTID);
1362                 BUG_ON(ret);
1363                 /*
1364                  * Set the last_snapshot field to the generation of the commit
1365                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1366                  * correctly (returns true) when the relocation root is created
1367                  * either inside the critical section of a transaction commit
1368                  * (through transaction.c:qgroup_account_snapshot()) and when
1369                  * it's created before the transaction commit is started.
1370                  */
1371                 commit_root_gen = btrfs_header_generation(root->commit_root);
1372                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1373         } else {
1374                 /*
1375                  * called by btrfs_reloc_post_snapshot_hook.
1376                  * the source tree is a reloc tree, all tree blocks
1377                  * modified after it was created have RELOC flag
1378                  * set in their headers. so it's OK to not update
1379                  * the 'last_snapshot'.
1380                  */
1381                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1382                                       BTRFS_TREE_RELOC_OBJECTID);
1383                 BUG_ON(ret);
1384         }
1385
1386         memcpy(root_item, &root->root_item, sizeof(*root_item));
1387         btrfs_set_root_bytenr(root_item, eb->start);
1388         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1389         btrfs_set_root_generation(root_item, trans->transid);
1390
1391         if (root->root_key.objectid == objectid) {
1392                 btrfs_set_root_refs(root_item, 0);
1393                 memset(&root_item->drop_progress, 0,
1394                        sizeof(struct btrfs_disk_key));
1395                 root_item->drop_level = 0;
1396         }
1397
1398         btrfs_tree_unlock(eb);
1399         free_extent_buffer(eb);
1400
1401         ret = btrfs_insert_root(trans, fs_info->tree_root,
1402                                 &root_key, root_item);
1403         BUG_ON(ret);
1404         kfree(root_item);
1405
1406         reloc_root = btrfs_read_fs_root(fs_info->tree_root, &root_key);
1407         BUG_ON(IS_ERR(reloc_root));
1408         reloc_root->last_trans = trans->transid;
1409         return reloc_root;
1410 }
1411
1412 /*
1413  * create reloc tree for a given fs tree. reloc tree is just a
1414  * snapshot of the fs tree with special root objectid.
1415  */
1416 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1417                           struct btrfs_root *root)
1418 {
1419         struct btrfs_fs_info *fs_info = root->fs_info;
1420         struct btrfs_root *reloc_root;
1421         struct reloc_control *rc = fs_info->reloc_ctl;
1422         struct btrfs_block_rsv *rsv;
1423         int clear_rsv = 0;
1424         int ret;
1425
1426         if (root->reloc_root) {
1427                 reloc_root = root->reloc_root;
1428                 reloc_root->last_trans = trans->transid;
1429                 return 0;
1430         }
1431
1432         if (!rc || !rc->create_reloc_tree ||
1433             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1434                 return 0;
1435
1436         if (!trans->reloc_reserved) {
1437                 rsv = trans->block_rsv;
1438                 trans->block_rsv = rc->block_rsv;
1439                 clear_rsv = 1;
1440         }
1441         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1442         if (clear_rsv)
1443                 trans->block_rsv = rsv;
1444
1445         ret = __add_reloc_root(reloc_root);
1446         BUG_ON(ret < 0);
1447         root->reloc_root = reloc_root;
1448         return 0;
1449 }
1450
1451 /*
1452  * update root item of reloc tree
1453  */
1454 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1455                             struct btrfs_root *root)
1456 {
1457         struct btrfs_fs_info *fs_info = root->fs_info;
1458         struct btrfs_root *reloc_root;
1459         struct btrfs_root_item *root_item;
1460         int ret;
1461
1462         if (!root->reloc_root)
1463                 goto out;
1464
1465         reloc_root = root->reloc_root;
1466         root_item = &reloc_root->root_item;
1467
1468         if (fs_info->reloc_ctl->merge_reloc_tree &&
1469             btrfs_root_refs(root_item) == 0) {
1470                 root->reloc_root = NULL;
1471                 __del_reloc_root(reloc_root);
1472         }
1473
1474         if (reloc_root->commit_root != reloc_root->node) {
1475                 btrfs_set_root_node(root_item, reloc_root->node);
1476                 free_extent_buffer(reloc_root->commit_root);
1477                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1478         }
1479
1480         ret = btrfs_update_root(trans, fs_info->tree_root,
1481                                 &reloc_root->root_key, root_item);
1482         BUG_ON(ret);
1483
1484 out:
1485         return 0;
1486 }
1487
1488 /*
1489  * helper to find first cached inode with inode number >= objectid
1490  * in a subvolume
1491  */
1492 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1493 {
1494         struct rb_node *node;
1495         struct rb_node *prev;
1496         struct btrfs_inode *entry;
1497         struct inode *inode;
1498
1499         spin_lock(&root->inode_lock);
1500 again:
1501         node = root->inode_tree.rb_node;
1502         prev = NULL;
1503         while (node) {
1504                 prev = node;
1505                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1506
1507                 if (objectid < btrfs_ino(entry))
1508                         node = node->rb_left;
1509                 else if (objectid > btrfs_ino(entry))
1510                         node = node->rb_right;
1511                 else
1512                         break;
1513         }
1514         if (!node) {
1515                 while (prev) {
1516                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1517                         if (objectid <= btrfs_ino(entry)) {
1518                                 node = prev;
1519                                 break;
1520                         }
1521                         prev = rb_next(prev);
1522                 }
1523         }
1524         while (node) {
1525                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1526                 inode = igrab(&entry->vfs_inode);
1527                 if (inode) {
1528                         spin_unlock(&root->inode_lock);
1529                         return inode;
1530                 }
1531
1532                 objectid = btrfs_ino(entry) + 1;
1533                 if (cond_resched_lock(&root->inode_lock))
1534                         goto again;
1535
1536                 node = rb_next(node);
1537         }
1538         spin_unlock(&root->inode_lock);
1539         return NULL;
1540 }
1541
1542 static int in_block_group(u64 bytenr,
1543                           struct btrfs_block_group_cache *block_group)
1544 {
1545         if (bytenr >= block_group->key.objectid &&
1546             bytenr < block_group->key.objectid + block_group->key.offset)
1547                 return 1;
1548         return 0;
1549 }
1550
1551 /*
1552  * get new location of data
1553  */
1554 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1555                             u64 bytenr, u64 num_bytes)
1556 {
1557         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1558         struct btrfs_path *path;
1559         struct btrfs_file_extent_item *fi;
1560         struct extent_buffer *leaf;
1561         int ret;
1562
1563         path = btrfs_alloc_path();
1564         if (!path)
1565                 return -ENOMEM;
1566
1567         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1568         ret = btrfs_lookup_file_extent(NULL, root, path,
1569                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1570         if (ret < 0)
1571                 goto out;
1572         if (ret > 0) {
1573                 ret = -ENOENT;
1574                 goto out;
1575         }
1576
1577         leaf = path->nodes[0];
1578         fi = btrfs_item_ptr(leaf, path->slots[0],
1579                             struct btrfs_file_extent_item);
1580
1581         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1582                btrfs_file_extent_compression(leaf, fi) ||
1583                btrfs_file_extent_encryption(leaf, fi) ||
1584                btrfs_file_extent_other_encoding(leaf, fi));
1585
1586         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1587                 ret = -EINVAL;
1588                 goto out;
1589         }
1590
1591         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1592         ret = 0;
1593 out:
1594         btrfs_free_path(path);
1595         return ret;
1596 }
1597
1598 /*
1599  * update file extent items in the tree leaf to point to
1600  * the new locations.
1601  */
1602 static noinline_for_stack
1603 int replace_file_extents(struct btrfs_trans_handle *trans,
1604                          struct reloc_control *rc,
1605                          struct btrfs_root *root,
1606                          struct extent_buffer *leaf)
1607 {
1608         struct btrfs_fs_info *fs_info = root->fs_info;
1609         struct btrfs_key key;
1610         struct btrfs_file_extent_item *fi;
1611         struct inode *inode = NULL;
1612         u64 parent;
1613         u64 bytenr;
1614         u64 new_bytenr = 0;
1615         u64 num_bytes;
1616         u64 end;
1617         u32 nritems;
1618         u32 i;
1619         int ret = 0;
1620         int first = 1;
1621         int dirty = 0;
1622
1623         if (rc->stage != UPDATE_DATA_PTRS)
1624                 return 0;
1625
1626         /* reloc trees always use full backref */
1627         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1628                 parent = leaf->start;
1629         else
1630                 parent = 0;
1631
1632         nritems = btrfs_header_nritems(leaf);
1633         for (i = 0; i < nritems; i++) {
1634                 cond_resched();
1635                 btrfs_item_key_to_cpu(leaf, &key, i);
1636                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1637                         continue;
1638                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1639                 if (btrfs_file_extent_type(leaf, fi) ==
1640                     BTRFS_FILE_EXTENT_INLINE)
1641                         continue;
1642                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1643                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1644                 if (bytenr == 0)
1645                         continue;
1646                 if (!in_block_group(bytenr, rc->block_group))
1647                         continue;
1648
1649                 /*
1650                  * if we are modifying block in fs tree, wait for readpage
1651                  * to complete and drop the extent cache
1652                  */
1653                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1654                         if (first) {
1655                                 inode = find_next_inode(root, key.objectid);
1656                                 first = 0;
1657                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1658                                 btrfs_add_delayed_iput(inode);
1659                                 inode = find_next_inode(root, key.objectid);
1660                         }
1661                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1662                                 end = key.offset +
1663                                       btrfs_file_extent_num_bytes(leaf, fi);
1664                                 WARN_ON(!IS_ALIGNED(key.offset,
1665                                                     fs_info->sectorsize));
1666                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1667                                 end--;
1668                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1669                                                       key.offset, end);
1670                                 if (!ret)
1671                                         continue;
1672
1673                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1674                                                 key.offset,     end, 1);
1675                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1676                                               key.offset, end);
1677                         }
1678                 }
1679
1680                 ret = get_new_location(rc->data_inode, &new_bytenr,
1681                                        bytenr, num_bytes);
1682                 if (ret) {
1683                         /*
1684                          * Don't have to abort since we've not changed anything
1685                          * in the file extent yet.
1686                          */
1687                         break;
1688                 }
1689
1690                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1691                 dirty = 1;
1692
1693                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1694                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1695                                            num_bytes, parent,
1696                                            btrfs_header_owner(leaf),
1697                                            key.objectid, key.offset);
1698                 if (ret) {
1699                         btrfs_abort_transaction(trans, ret);
1700                         break;
1701                 }
1702
1703                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1704                                         parent, btrfs_header_owner(leaf),
1705                                         key.objectid, key.offset);
1706                 if (ret) {
1707                         btrfs_abort_transaction(trans, ret);
1708                         break;
1709                 }
1710         }
1711         if (dirty)
1712                 btrfs_mark_buffer_dirty(leaf);
1713         if (inode)
1714                 btrfs_add_delayed_iput(inode);
1715         return ret;
1716 }
1717
1718 static noinline_for_stack
1719 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1720                      struct btrfs_path *path, int level)
1721 {
1722         struct btrfs_disk_key key1;
1723         struct btrfs_disk_key key2;
1724         btrfs_node_key(eb, &key1, slot);
1725         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1726         return memcmp(&key1, &key2, sizeof(key1));
1727 }
1728
1729 /*
1730  * try to replace tree blocks in fs tree with the new blocks
1731  * in reloc tree. tree blocks haven't been modified since the
1732  * reloc tree was create can be replaced.
1733  *
1734  * if a block was replaced, level of the block + 1 is returned.
1735  * if no block got replaced, 0 is returned. if there are other
1736  * errors, a negative error number is returned.
1737  */
1738 static noinline_for_stack
1739 int replace_path(struct btrfs_trans_handle *trans,
1740                  struct btrfs_root *dest, struct btrfs_root *src,
1741                  struct btrfs_path *path, struct btrfs_key *next_key,
1742                  int lowest_level, int max_level)
1743 {
1744         struct btrfs_fs_info *fs_info = dest->fs_info;
1745         struct extent_buffer *eb;
1746         struct extent_buffer *parent;
1747         struct btrfs_key key;
1748         u64 old_bytenr;
1749         u64 new_bytenr;
1750         u64 old_ptr_gen;
1751         u64 new_ptr_gen;
1752         u64 last_snapshot;
1753         u32 blocksize;
1754         int cow = 0;
1755         int level;
1756         int ret;
1757         int slot;
1758
1759         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1760         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1761
1762         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1763 again:
1764         slot = path->slots[lowest_level];
1765         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1766
1767         eb = btrfs_lock_root_node(dest);
1768         btrfs_set_lock_blocking(eb);
1769         level = btrfs_header_level(eb);
1770
1771         if (level < lowest_level) {
1772                 btrfs_tree_unlock(eb);
1773                 free_extent_buffer(eb);
1774                 return 0;
1775         }
1776
1777         if (cow) {
1778                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1779                 BUG_ON(ret);
1780         }
1781         btrfs_set_lock_blocking(eb);
1782
1783         if (next_key) {
1784                 next_key->objectid = (u64)-1;
1785                 next_key->type = (u8)-1;
1786                 next_key->offset = (u64)-1;
1787         }
1788
1789         parent = eb;
1790         while (1) {
1791                 struct btrfs_key first_key;
1792
1793                 level = btrfs_header_level(parent);
1794                 BUG_ON(level < lowest_level);
1795
1796                 ret = btrfs_bin_search(parent, &key, level, &slot);
1797                 if (ret && slot > 0)
1798                         slot--;
1799
1800                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1801                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1802
1803                 old_bytenr = btrfs_node_blockptr(parent, slot);
1804                 blocksize = fs_info->nodesize;
1805                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1806                 btrfs_node_key_to_cpu(parent, &first_key, slot);
1807
1808                 if (level <= max_level) {
1809                         eb = path->nodes[level];
1810                         new_bytenr = btrfs_node_blockptr(eb,
1811                                                         path->slots[level]);
1812                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1813                                                         path->slots[level]);
1814                 } else {
1815                         new_bytenr = 0;
1816                         new_ptr_gen = 0;
1817                 }
1818
1819                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1820                         ret = level;
1821                         break;
1822                 }
1823
1824                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1825                     memcmp_node_keys(parent, slot, path, level)) {
1826                         if (level <= lowest_level) {
1827                                 ret = 0;
1828                                 break;
1829                         }
1830
1831                         eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen,
1832                                              level - 1, &first_key);
1833                         if (IS_ERR(eb)) {
1834                                 ret = PTR_ERR(eb);
1835                                 break;
1836                         } else if (!extent_buffer_uptodate(eb)) {
1837                                 ret = -EIO;
1838                                 free_extent_buffer(eb);
1839                                 break;
1840                         }
1841                         btrfs_tree_lock(eb);
1842                         if (cow) {
1843                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1844                                                       slot, &eb);
1845                                 BUG_ON(ret);
1846                         }
1847                         btrfs_set_lock_blocking(eb);
1848
1849                         btrfs_tree_unlock(parent);
1850                         free_extent_buffer(parent);
1851
1852                         parent = eb;
1853                         continue;
1854                 }
1855
1856                 if (!cow) {
1857                         btrfs_tree_unlock(parent);
1858                         free_extent_buffer(parent);
1859                         cow = 1;
1860                         goto again;
1861                 }
1862
1863                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1864                                       path->slots[level]);
1865                 btrfs_release_path(path);
1866
1867                 path->lowest_level = level;
1868                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1869                 path->lowest_level = 0;
1870                 BUG_ON(ret);
1871
1872                 /*
1873                  * Info qgroup to trace both subtrees.
1874                  *
1875                  * We must trace both trees.
1876                  * 1) Tree reloc subtree
1877                  *    If not traced, we will leak data numbers
1878                  * 2) Fs subtree
1879                  *    If not traced, we will double count old data
1880                  *    and tree block numbers, if current trans doesn't free
1881                  *    data reloc tree inode.
1882                  */
1883                 ret = btrfs_qgroup_trace_subtree(trans, parent,
1884                                 btrfs_header_generation(parent),
1885                                 btrfs_header_level(parent));
1886                 if (ret < 0)
1887                         break;
1888                 ret = btrfs_qgroup_trace_subtree(trans, path->nodes[level],
1889                                 btrfs_header_generation(path->nodes[level]),
1890                                 btrfs_header_level(path->nodes[level]));
1891                 if (ret < 0)
1892                         break;
1893
1894                 /*
1895                  * swap blocks in fs tree and reloc tree.
1896                  */
1897                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1898                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1899                 btrfs_mark_buffer_dirty(parent);
1900
1901                 btrfs_set_node_blockptr(path->nodes[level],
1902                                         path->slots[level], old_bytenr);
1903                 btrfs_set_node_ptr_generation(path->nodes[level],
1904                                               path->slots[level], old_ptr_gen);
1905                 btrfs_mark_buffer_dirty(path->nodes[level]);
1906
1907                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr,
1908                                         blocksize, path->nodes[level]->start,
1909                                         src->root_key.objectid, level - 1, 0);
1910                 BUG_ON(ret);
1911                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr,
1912                                         blocksize, 0, dest->root_key.objectid,
1913                                         level - 1, 0);
1914                 BUG_ON(ret);
1915
1916                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1917                                         path->nodes[level]->start,
1918                                         src->root_key.objectid, level - 1, 0);
1919                 BUG_ON(ret);
1920
1921                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1922                                         0, dest->root_key.objectid, level - 1,
1923                                         0);
1924                 BUG_ON(ret);
1925
1926                 btrfs_unlock_up_safe(path, 0);
1927
1928                 ret = level;
1929                 break;
1930         }
1931         btrfs_tree_unlock(parent);
1932         free_extent_buffer(parent);
1933         return ret;
1934 }
1935
1936 /*
1937  * helper to find next relocated block in reloc tree
1938  */
1939 static noinline_for_stack
1940 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1941                        int *level)
1942 {
1943         struct extent_buffer *eb;
1944         int i;
1945         u64 last_snapshot;
1946         u32 nritems;
1947
1948         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1949
1950         for (i = 0; i < *level; i++) {
1951                 free_extent_buffer(path->nodes[i]);
1952                 path->nodes[i] = NULL;
1953         }
1954
1955         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1956                 eb = path->nodes[i];
1957                 nritems = btrfs_header_nritems(eb);
1958                 while (path->slots[i] + 1 < nritems) {
1959                         path->slots[i]++;
1960                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1961                             last_snapshot)
1962                                 continue;
1963
1964                         *level = i;
1965                         return 0;
1966                 }
1967                 free_extent_buffer(path->nodes[i]);
1968                 path->nodes[i] = NULL;
1969         }
1970         return 1;
1971 }
1972
1973 /*
1974  * walk down reloc tree to find relocated block of lowest level
1975  */
1976 static noinline_for_stack
1977 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1978                          int *level)
1979 {
1980         struct btrfs_fs_info *fs_info = root->fs_info;
1981         struct extent_buffer *eb = NULL;
1982         int i;
1983         u64 bytenr;
1984         u64 ptr_gen = 0;
1985         u64 last_snapshot;
1986         u32 nritems;
1987
1988         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1989
1990         for (i = *level; i > 0; i--) {
1991                 struct btrfs_key first_key;
1992
1993                 eb = path->nodes[i];
1994                 nritems = btrfs_header_nritems(eb);
1995                 while (path->slots[i] < nritems) {
1996                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1997                         if (ptr_gen > last_snapshot)
1998                                 break;
1999                         path->slots[i]++;
2000                 }
2001                 if (path->slots[i] >= nritems) {
2002                         if (i == *level)
2003                                 break;
2004                         *level = i + 1;
2005                         return 0;
2006                 }
2007                 if (i == 1) {
2008                         *level = i;
2009                         return 0;
2010                 }
2011
2012                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2013                 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
2014                 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
2015                                      &first_key);
2016                 if (IS_ERR(eb)) {
2017                         return PTR_ERR(eb);
2018                 } else if (!extent_buffer_uptodate(eb)) {
2019                         free_extent_buffer(eb);
2020                         return -EIO;
2021                 }
2022                 BUG_ON(btrfs_header_level(eb) != i - 1);
2023                 path->nodes[i - 1] = eb;
2024                 path->slots[i - 1] = 0;
2025         }
2026         return 1;
2027 }
2028
2029 /*
2030  * invalidate extent cache for file extents whose key in range of
2031  * [min_key, max_key)
2032  */
2033 static int invalidate_extent_cache(struct btrfs_root *root,
2034                                    struct btrfs_key *min_key,
2035                                    struct btrfs_key *max_key)
2036 {
2037         struct btrfs_fs_info *fs_info = root->fs_info;
2038         struct inode *inode = NULL;
2039         u64 objectid;
2040         u64 start, end;
2041         u64 ino;
2042
2043         objectid = min_key->objectid;
2044         while (1) {
2045                 cond_resched();
2046                 iput(inode);
2047
2048                 if (objectid > max_key->objectid)
2049                         break;
2050
2051                 inode = find_next_inode(root, objectid);
2052                 if (!inode)
2053                         break;
2054                 ino = btrfs_ino(BTRFS_I(inode));
2055
2056                 if (ino > max_key->objectid) {
2057                         iput(inode);
2058                         break;
2059                 }
2060
2061                 objectid = ino + 1;
2062                 if (!S_ISREG(inode->i_mode))
2063                         continue;
2064
2065                 if (unlikely(min_key->objectid == ino)) {
2066                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2067                                 continue;
2068                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2069                                 start = 0;
2070                         else {
2071                                 start = min_key->offset;
2072                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
2073                         }
2074                 } else {
2075                         start = 0;
2076                 }
2077
2078                 if (unlikely(max_key->objectid == ino)) {
2079                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2080                                 continue;
2081                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2082                                 end = (u64)-1;
2083                         } else {
2084                                 if (max_key->offset == 0)
2085                                         continue;
2086                                 end = max_key->offset;
2087                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
2088                                 end--;
2089                         }
2090                 } else {
2091                         end = (u64)-1;
2092                 }
2093
2094                 /* the lock_extent waits for readpage to complete */
2095                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2096                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
2097                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2098         }
2099         return 0;
2100 }
2101
2102 static int find_next_key(struct btrfs_path *path, int level,
2103                          struct btrfs_key *key)
2104
2105 {
2106         while (level < BTRFS_MAX_LEVEL) {
2107                 if (!path->nodes[level])
2108                         break;
2109                 if (path->slots[level] + 1 <
2110                     btrfs_header_nritems(path->nodes[level])) {
2111                         btrfs_node_key_to_cpu(path->nodes[level], key,
2112                                               path->slots[level] + 1);
2113                         return 0;
2114                 }
2115                 level++;
2116         }
2117         return 1;
2118 }
2119
2120 /*
2121  * merge the relocated tree blocks in reloc tree with corresponding
2122  * fs tree.
2123  */
2124 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2125                                                struct btrfs_root *root)
2126 {
2127         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2128         LIST_HEAD(inode_list);
2129         struct btrfs_key key;
2130         struct btrfs_key next_key;
2131         struct btrfs_trans_handle *trans = NULL;
2132         struct btrfs_root *reloc_root;
2133         struct btrfs_root_item *root_item;
2134         struct btrfs_path *path;
2135         struct extent_buffer *leaf;
2136         int level;
2137         int max_level;
2138         int replaced = 0;
2139         int ret;
2140         int err = 0;
2141         u32 min_reserved;
2142
2143         path = btrfs_alloc_path();
2144         if (!path)
2145                 return -ENOMEM;
2146         path->reada = READA_FORWARD;
2147
2148         reloc_root = root->reloc_root;
2149         root_item = &reloc_root->root_item;
2150
2151         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2152                 level = btrfs_root_level(root_item);
2153                 extent_buffer_get(reloc_root->node);
2154                 path->nodes[level] = reloc_root->node;
2155                 path->slots[level] = 0;
2156         } else {
2157                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2158
2159                 level = root_item->drop_level;
2160                 BUG_ON(level == 0);
2161                 path->lowest_level = level;
2162                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2163                 path->lowest_level = 0;
2164                 if (ret < 0) {
2165                         btrfs_free_path(path);
2166                         return ret;
2167                 }
2168
2169                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2170                                       path->slots[level]);
2171                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2172
2173                 btrfs_unlock_up_safe(path, 0);
2174         }
2175
2176         min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2177         memset(&next_key, 0, sizeof(next_key));
2178
2179         while (1) {
2180                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2181                                              BTRFS_RESERVE_FLUSH_ALL);
2182                 if (ret) {
2183                         err = ret;
2184                         goto out;
2185                 }
2186                 trans = btrfs_start_transaction(root, 0);
2187                 if (IS_ERR(trans)) {
2188                         err = PTR_ERR(trans);
2189                         trans = NULL;
2190                         goto out;
2191                 }
2192                 trans->block_rsv = rc->block_rsv;
2193
2194                 replaced = 0;
2195                 max_level = level;
2196
2197                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2198                 if (ret < 0) {
2199                         err = ret;
2200                         goto out;
2201                 }
2202                 if (ret > 0)
2203                         break;
2204
2205                 if (!find_next_key(path, level, &key) &&
2206                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2207                         ret = 0;
2208                 } else {
2209                         ret = replace_path(trans, root, reloc_root, path,
2210                                            &next_key, level, max_level);
2211                 }
2212                 if (ret < 0) {
2213                         err = ret;
2214                         goto out;
2215                 }
2216
2217                 if (ret > 0) {
2218                         level = ret;
2219                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2220                                               path->slots[level]);
2221                         replaced = 1;
2222                 }
2223
2224                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2225                 if (ret > 0)
2226                         break;
2227
2228                 BUG_ON(level == 0);
2229                 /*
2230                  * save the merging progress in the drop_progress.
2231                  * this is OK since root refs == 1 in this case.
2232                  */
2233                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2234                                path->slots[level]);
2235                 root_item->drop_level = level;
2236
2237                 btrfs_end_transaction_throttle(trans);
2238                 trans = NULL;
2239
2240                 btrfs_btree_balance_dirty(fs_info);
2241
2242                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2243                         invalidate_extent_cache(root, &key, &next_key);
2244         }
2245
2246         /*
2247          * handle the case only one block in the fs tree need to be
2248          * relocated and the block is tree root.
2249          */
2250         leaf = btrfs_lock_root_node(root);
2251         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2252         btrfs_tree_unlock(leaf);
2253         free_extent_buffer(leaf);
2254         if (ret < 0)
2255                 err = ret;
2256 out:
2257         btrfs_free_path(path);
2258
2259         if (err == 0) {
2260                 memset(&root_item->drop_progress, 0,
2261                        sizeof(root_item->drop_progress));
2262                 root_item->drop_level = 0;
2263                 btrfs_set_root_refs(root_item, 0);
2264                 btrfs_update_reloc_root(trans, root);
2265         }
2266
2267         if (trans)
2268                 btrfs_end_transaction_throttle(trans);
2269
2270         btrfs_btree_balance_dirty(fs_info);
2271
2272         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2273                 invalidate_extent_cache(root, &key, &next_key);
2274
2275         return err;
2276 }
2277
2278 static noinline_for_stack
2279 int prepare_to_merge(struct reloc_control *rc, int err)
2280 {
2281         struct btrfs_root *root = rc->extent_root;
2282         struct btrfs_fs_info *fs_info = root->fs_info;
2283         struct btrfs_root *reloc_root;
2284         struct btrfs_trans_handle *trans;
2285         LIST_HEAD(reloc_roots);
2286         u64 num_bytes = 0;
2287         int ret;
2288
2289         mutex_lock(&fs_info->reloc_mutex);
2290         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2291         rc->merging_rsv_size += rc->nodes_relocated * 2;
2292         mutex_unlock(&fs_info->reloc_mutex);
2293
2294 again:
2295         if (!err) {
2296                 num_bytes = rc->merging_rsv_size;
2297                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2298                                           BTRFS_RESERVE_FLUSH_ALL);
2299                 if (ret)
2300                         err = ret;
2301         }
2302
2303         trans = btrfs_join_transaction(rc->extent_root);
2304         if (IS_ERR(trans)) {
2305                 if (!err)
2306                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2307                                                 num_bytes);
2308                 return PTR_ERR(trans);
2309         }
2310
2311         if (!err) {
2312                 if (num_bytes != rc->merging_rsv_size) {
2313                         btrfs_end_transaction(trans);
2314                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2315                                                 num_bytes);
2316                         goto again;
2317                 }
2318         }
2319
2320         rc->merge_reloc_tree = 1;
2321
2322         while (!list_empty(&rc->reloc_roots)) {
2323                 reloc_root = list_entry(rc->reloc_roots.next,
2324                                         struct btrfs_root, root_list);
2325                 list_del_init(&reloc_root->root_list);
2326
2327                 root = read_fs_root(fs_info, reloc_root->root_key.offset);
2328                 BUG_ON(IS_ERR(root));
2329                 BUG_ON(root->reloc_root != reloc_root);
2330
2331                 /*
2332                  * set reference count to 1, so btrfs_recover_relocation
2333                  * knows it should resumes merging
2334                  */
2335                 if (!err)
2336                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2337                 btrfs_update_reloc_root(trans, root);
2338
2339                 list_add(&reloc_root->root_list, &reloc_roots);
2340         }
2341
2342         list_splice(&reloc_roots, &rc->reloc_roots);
2343
2344         if (!err)
2345                 btrfs_commit_transaction(trans);
2346         else
2347                 btrfs_end_transaction(trans);
2348         return err;
2349 }
2350
2351 static noinline_for_stack
2352 void free_reloc_roots(struct list_head *list)
2353 {
2354         struct btrfs_root *reloc_root;
2355
2356         while (!list_empty(list)) {
2357                 reloc_root = list_entry(list->next, struct btrfs_root,
2358                                         root_list);
2359                 __del_reloc_root(reloc_root);
2360                 free_extent_buffer(reloc_root->node);
2361                 free_extent_buffer(reloc_root->commit_root);
2362                 reloc_root->node = NULL;
2363                 reloc_root->commit_root = NULL;
2364         }
2365 }
2366
2367 static noinline_for_stack
2368 void merge_reloc_roots(struct reloc_control *rc)
2369 {
2370         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2371         struct btrfs_root *root;
2372         struct btrfs_root *reloc_root;
2373         LIST_HEAD(reloc_roots);
2374         int found = 0;
2375         int ret = 0;
2376 again:
2377         root = rc->extent_root;
2378
2379         /*
2380          * this serializes us with btrfs_record_root_in_transaction,
2381          * we have to make sure nobody is in the middle of
2382          * adding their roots to the list while we are
2383          * doing this splice
2384          */
2385         mutex_lock(&fs_info->reloc_mutex);
2386         list_splice_init(&rc->reloc_roots, &reloc_roots);
2387         mutex_unlock(&fs_info->reloc_mutex);
2388
2389         while (!list_empty(&reloc_roots)) {
2390                 found = 1;
2391                 reloc_root = list_entry(reloc_roots.next,
2392                                         struct btrfs_root, root_list);
2393
2394                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2395                         root = read_fs_root(fs_info,
2396                                             reloc_root->root_key.offset);
2397                         BUG_ON(IS_ERR(root));
2398                         BUG_ON(root->reloc_root != reloc_root);
2399
2400                         ret = merge_reloc_root(rc, root);
2401                         if (ret) {
2402                                 if (list_empty(&reloc_root->root_list))
2403                                         list_add_tail(&reloc_root->root_list,
2404                                                       &reloc_roots);
2405                                 goto out;
2406                         }
2407                 } else {
2408                         list_del_init(&reloc_root->root_list);
2409                 }
2410
2411                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2412                 if (ret < 0) {
2413                         if (list_empty(&reloc_root->root_list))
2414                                 list_add_tail(&reloc_root->root_list,
2415                                               &reloc_roots);
2416                         goto out;
2417                 }
2418         }
2419
2420         if (found) {
2421                 found = 0;
2422                 goto again;
2423         }
2424 out:
2425         if (ret) {
2426                 btrfs_handle_fs_error(fs_info, ret, NULL);
2427                 if (!list_empty(&reloc_roots))
2428                         free_reloc_roots(&reloc_roots);
2429
2430                 /* new reloc root may be added */
2431                 mutex_lock(&fs_info->reloc_mutex);
2432                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2433                 mutex_unlock(&fs_info->reloc_mutex);
2434                 if (!list_empty(&reloc_roots))
2435                         free_reloc_roots(&reloc_roots);
2436         }
2437
2438         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2439 }
2440
2441 static void free_block_list(struct rb_root *blocks)
2442 {
2443         struct tree_block *block;
2444         struct rb_node *rb_node;
2445         while ((rb_node = rb_first(blocks))) {
2446                 block = rb_entry(rb_node, struct tree_block, rb_node);
2447                 rb_erase(rb_node, blocks);
2448                 kfree(block);
2449         }
2450 }
2451
2452 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2453                                       struct btrfs_root *reloc_root)
2454 {
2455         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2456         struct btrfs_root *root;
2457
2458         if (reloc_root->last_trans == trans->transid)
2459                 return 0;
2460
2461         root = read_fs_root(fs_info, reloc_root->root_key.offset);
2462         BUG_ON(IS_ERR(root));
2463         BUG_ON(root->reloc_root != reloc_root);
2464
2465         return btrfs_record_root_in_trans(trans, root);
2466 }
2467
2468 static noinline_for_stack
2469 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2470                                      struct reloc_control *rc,
2471                                      struct backref_node *node,
2472                                      struct backref_edge *edges[])
2473 {
2474         struct backref_node *next;
2475         struct btrfs_root *root;
2476         int index = 0;
2477
2478         next = node;
2479         while (1) {
2480                 cond_resched();
2481                 next = walk_up_backref(next, edges, &index);
2482                 root = next->root;
2483                 BUG_ON(!root);
2484                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2485
2486                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2487                         record_reloc_root_in_trans(trans, root);
2488                         break;
2489                 }
2490
2491                 btrfs_record_root_in_trans(trans, root);
2492                 root = root->reloc_root;
2493
2494                 if (next->new_bytenr != root->node->start) {
2495                         BUG_ON(next->new_bytenr);
2496                         BUG_ON(!list_empty(&next->list));
2497                         next->new_bytenr = root->node->start;
2498                         next->root = root;
2499                         list_add_tail(&next->list,
2500                                       &rc->backref_cache.changed);
2501                         __mark_block_processed(rc, next);
2502                         break;
2503                 }
2504
2505                 WARN_ON(1);
2506                 root = NULL;
2507                 next = walk_down_backref(edges, &index);
2508                 if (!next || next->level <= node->level)
2509                         break;
2510         }
2511         if (!root)
2512                 return NULL;
2513
2514         next = node;
2515         /* setup backref node path for btrfs_reloc_cow_block */
2516         while (1) {
2517                 rc->backref_cache.path[next->level] = next;
2518                 if (--index < 0)
2519                         break;
2520                 next = edges[index]->node[UPPER];
2521         }
2522         return root;
2523 }
2524
2525 /*
2526  * select a tree root for relocation. return NULL if the block
2527  * is reference counted. we should use do_relocation() in this
2528  * case. return a tree root pointer if the block isn't reference
2529  * counted. return -ENOENT if the block is root of reloc tree.
2530  */
2531 static noinline_for_stack
2532 struct btrfs_root *select_one_root(struct backref_node *node)
2533 {
2534         struct backref_node *next;
2535         struct btrfs_root *root;
2536         struct btrfs_root *fs_root = NULL;
2537         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2538         int index = 0;
2539
2540         next = node;
2541         while (1) {
2542                 cond_resched();
2543                 next = walk_up_backref(next, edges, &index);
2544                 root = next->root;
2545                 BUG_ON(!root);
2546
2547                 /* no other choice for non-references counted tree */
2548                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2549                         return root;
2550
2551                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2552                         fs_root = root;
2553
2554                 if (next != node)
2555                         return NULL;
2556
2557                 next = walk_down_backref(edges, &index);
2558                 if (!next || next->level <= node->level)
2559                         break;
2560         }
2561
2562         if (!fs_root)
2563                 return ERR_PTR(-ENOENT);
2564         return fs_root;
2565 }
2566
2567 static noinline_for_stack
2568 u64 calcu_metadata_size(struct reloc_control *rc,
2569                         struct backref_node *node, int reserve)
2570 {
2571         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2572         struct backref_node *next = node;
2573         struct backref_edge *edge;
2574         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2575         u64 num_bytes = 0;
2576         int index = 0;
2577
2578         BUG_ON(reserve && node->processed);
2579
2580         while (next) {
2581                 cond_resched();
2582                 while (1) {
2583                         if (next->processed && (reserve || next != node))
2584                                 break;
2585
2586                         num_bytes += fs_info->nodesize;
2587
2588                         if (list_empty(&next->upper))
2589                                 break;
2590
2591                         edge = list_entry(next->upper.next,
2592                                           struct backref_edge, list[LOWER]);
2593                         edges[index++] = edge;
2594                         next = edge->node[UPPER];
2595                 }
2596                 next = walk_down_backref(edges, &index);
2597         }
2598         return num_bytes;
2599 }
2600
2601 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2602                                   struct reloc_control *rc,
2603                                   struct backref_node *node)
2604 {
2605         struct btrfs_root *root = rc->extent_root;
2606         struct btrfs_fs_info *fs_info = root->fs_info;
2607         u64 num_bytes;
2608         int ret;
2609         u64 tmp;
2610
2611         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2612
2613         trans->block_rsv = rc->block_rsv;
2614         rc->reserved_bytes += num_bytes;
2615
2616         /*
2617          * We are under a transaction here so we can only do limited flushing.
2618          * If we get an enospc just kick back -EAGAIN so we know to drop the
2619          * transaction and try to refill when we can flush all the things.
2620          */
2621         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2622                                 BTRFS_RESERVE_FLUSH_LIMIT);
2623         if (ret) {
2624                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2625                 while (tmp <= rc->reserved_bytes)
2626                         tmp <<= 1;
2627                 /*
2628                  * only one thread can access block_rsv at this point,
2629                  * so we don't need hold lock to protect block_rsv.
2630                  * we expand more reservation size here to allow enough
2631                  * space for relocation and we will return eailer in
2632                  * enospc case.
2633                  */
2634                 rc->block_rsv->size = tmp + fs_info->nodesize *
2635                                       RELOCATION_RESERVED_NODES;
2636                 return -EAGAIN;
2637         }
2638
2639         return 0;
2640 }
2641
2642 /*
2643  * relocate a block tree, and then update pointers in upper level
2644  * blocks that reference the block to point to the new location.
2645  *
2646  * if called by link_to_upper, the block has already been relocated.
2647  * in that case this function just updates pointers.
2648  */
2649 static int do_relocation(struct btrfs_trans_handle *trans,
2650                          struct reloc_control *rc,
2651                          struct backref_node *node,
2652                          struct btrfs_key *key,
2653                          struct btrfs_path *path, int lowest)
2654 {
2655         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2656         struct backref_node *upper;
2657         struct backref_edge *edge;
2658         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2659         struct btrfs_root *root;
2660         struct extent_buffer *eb;
2661         u32 blocksize;
2662         u64 bytenr;
2663         u64 generation;
2664         int slot;
2665         int ret;
2666         int err = 0;
2667
2668         BUG_ON(lowest && node->eb);
2669
2670         path->lowest_level = node->level + 1;
2671         rc->backref_cache.path[node->level] = node;
2672         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2673                 struct btrfs_key first_key;
2674
2675                 cond_resched();
2676
2677                 upper = edge->node[UPPER];
2678                 root = select_reloc_root(trans, rc, upper, edges);
2679                 BUG_ON(!root);
2680
2681                 if (upper->eb && !upper->locked) {
2682                         if (!lowest) {
2683                                 ret = btrfs_bin_search(upper->eb, key,
2684                                                        upper->level, &slot);
2685                                 BUG_ON(ret);
2686                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2687                                 if (node->eb->start == bytenr)
2688                                         goto next;
2689                         }
2690                         drop_node_buffer(upper);
2691                 }
2692
2693                 if (!upper->eb) {
2694                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2695                         if (ret) {
2696                                 if (ret < 0)
2697                                         err = ret;
2698                                 else
2699                                         err = -ENOENT;
2700
2701                                 btrfs_release_path(path);
2702                                 break;
2703                         }
2704
2705                         if (!upper->eb) {
2706                                 upper->eb = path->nodes[upper->level];
2707                                 path->nodes[upper->level] = NULL;
2708                         } else {
2709                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2710                         }
2711
2712                         upper->locked = 1;
2713                         path->locks[upper->level] = 0;
2714
2715                         slot = path->slots[upper->level];
2716                         btrfs_release_path(path);
2717                 } else {
2718                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2719                                                &slot);
2720                         BUG_ON(ret);
2721                 }
2722
2723                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2724                 if (lowest) {
2725                         if (bytenr != node->bytenr) {
2726                                 btrfs_err(root->fs_info,
2727                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2728                                           bytenr, node->bytenr, slot,
2729                                           upper->eb->start);
2730                                 err = -EIO;
2731                                 goto next;
2732                         }
2733                 } else {
2734                         if (node->eb->start == bytenr)
2735                                 goto next;
2736                 }
2737
2738                 blocksize = root->fs_info->nodesize;
2739                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2740                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2741                 eb = read_tree_block(fs_info, bytenr, generation,
2742                                      upper->level - 1, &first_key);
2743                 if (IS_ERR(eb)) {
2744                         err = PTR_ERR(eb);
2745                         goto next;
2746                 } else if (!extent_buffer_uptodate(eb)) {
2747                         free_extent_buffer(eb);
2748                         err = -EIO;
2749                         goto next;
2750                 }
2751                 btrfs_tree_lock(eb);
2752                 btrfs_set_lock_blocking(eb);
2753
2754                 if (!node->eb) {
2755                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2756                                               slot, &eb);
2757                         btrfs_tree_unlock(eb);
2758                         free_extent_buffer(eb);
2759                         if (ret < 0) {
2760                                 err = ret;
2761                                 goto next;
2762                         }
2763                         BUG_ON(node->eb != eb);
2764                 } else {
2765                         btrfs_set_node_blockptr(upper->eb, slot,
2766                                                 node->eb->start);
2767                         btrfs_set_node_ptr_generation(upper->eb, slot,
2768                                                       trans->transid);
2769                         btrfs_mark_buffer_dirty(upper->eb);
2770
2771                         ret = btrfs_inc_extent_ref(trans, root,
2772                                                 node->eb->start, blocksize,
2773                                                 upper->eb->start,
2774                                                 btrfs_header_owner(upper->eb),
2775                                                 node->level, 0);
2776                         BUG_ON(ret);
2777
2778                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2779                         BUG_ON(ret);
2780                 }
2781 next:
2782                 if (!upper->pending)
2783                         drop_node_buffer(upper);
2784                 else
2785                         unlock_node_buffer(upper);
2786                 if (err)
2787                         break;
2788         }
2789
2790         if (!err && node->pending) {
2791                 drop_node_buffer(node);
2792                 list_move_tail(&node->list, &rc->backref_cache.changed);
2793                 node->pending = 0;
2794         }
2795
2796         path->lowest_level = 0;
2797         BUG_ON(err == -ENOSPC);
2798         return err;
2799 }
2800
2801 static int link_to_upper(struct btrfs_trans_handle *trans,
2802                          struct reloc_control *rc,
2803                          struct backref_node *node,
2804                          struct btrfs_path *path)
2805 {
2806         struct btrfs_key key;
2807
2808         btrfs_node_key_to_cpu(node->eb, &key, 0);
2809         return do_relocation(trans, rc, node, &key, path, 0);
2810 }
2811
2812 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2813                                 struct reloc_control *rc,
2814                                 struct btrfs_path *path, int err)
2815 {
2816         LIST_HEAD(list);
2817         struct backref_cache *cache = &rc->backref_cache;
2818         struct backref_node *node;
2819         int level;
2820         int ret;
2821
2822         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2823                 while (!list_empty(&cache->pending[level])) {
2824                         node = list_entry(cache->pending[level].next,
2825                                           struct backref_node, list);
2826                         list_move_tail(&node->list, &list);
2827                         BUG_ON(!node->pending);
2828
2829                         if (!err) {
2830                                 ret = link_to_upper(trans, rc, node, path);
2831                                 if (ret < 0)
2832                                         err = ret;
2833                         }
2834                 }
2835                 list_splice_init(&list, &cache->pending[level]);
2836         }
2837         return err;
2838 }
2839
2840 static void mark_block_processed(struct reloc_control *rc,
2841                                  u64 bytenr, u32 blocksize)
2842 {
2843         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2844                         EXTENT_DIRTY);
2845 }
2846
2847 static void __mark_block_processed(struct reloc_control *rc,
2848                                    struct backref_node *node)
2849 {
2850         u32 blocksize;
2851         if (node->level == 0 ||
2852             in_block_group(node->bytenr, rc->block_group)) {
2853                 blocksize = rc->extent_root->fs_info->nodesize;
2854                 mark_block_processed(rc, node->bytenr, blocksize);
2855         }
2856         node->processed = 1;
2857 }
2858
2859 /*
2860  * mark a block and all blocks directly/indirectly reference the block
2861  * as processed.
2862  */
2863 static void update_processed_blocks(struct reloc_control *rc,
2864                                     struct backref_node *node)
2865 {
2866         struct backref_node *next = node;
2867         struct backref_edge *edge;
2868         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2869         int index = 0;
2870
2871         while (next) {
2872                 cond_resched();
2873                 while (1) {
2874                         if (next->processed)
2875                                 break;
2876
2877                         __mark_block_processed(rc, next);
2878
2879                         if (list_empty(&next->upper))
2880                                 break;
2881
2882                         edge = list_entry(next->upper.next,
2883                                           struct backref_edge, list[LOWER]);
2884                         edges[index++] = edge;
2885                         next = edge->node[UPPER];
2886                 }
2887                 next = walk_down_backref(edges, &index);
2888         }
2889 }
2890
2891 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2892 {
2893         u32 blocksize = rc->extent_root->fs_info->nodesize;
2894
2895         if (test_range_bit(&rc->processed_blocks, bytenr,
2896                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2897                 return 1;
2898         return 0;
2899 }
2900
2901 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2902                               struct tree_block *block)
2903 {
2904         struct extent_buffer *eb;
2905
2906         BUG_ON(block->key_ready);
2907         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2908                              block->level, NULL);
2909         if (IS_ERR(eb)) {
2910                 return PTR_ERR(eb);
2911         } else if (!extent_buffer_uptodate(eb)) {
2912                 free_extent_buffer(eb);
2913                 return -EIO;
2914         }
2915         WARN_ON(btrfs_header_level(eb) != block->level);
2916         if (block->level == 0)
2917                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2918         else
2919                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2920         free_extent_buffer(eb);
2921         block->key_ready = 1;
2922         return 0;
2923 }
2924
2925 /*
2926  * helper function to relocate a tree block
2927  */
2928 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2929                                 struct reloc_control *rc,
2930                                 struct backref_node *node,
2931                                 struct btrfs_key *key,
2932                                 struct btrfs_path *path)
2933 {
2934         struct btrfs_root *root;
2935         int ret = 0;
2936
2937         if (!node)
2938                 return 0;
2939
2940         BUG_ON(node->processed);
2941         root = select_one_root(node);
2942         if (root == ERR_PTR(-ENOENT)) {
2943                 update_processed_blocks(rc, node);
2944                 goto out;
2945         }
2946
2947         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2948                 ret = reserve_metadata_space(trans, rc, node);
2949                 if (ret)
2950                         goto out;
2951         }
2952
2953         if (root) {
2954                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2955                         BUG_ON(node->new_bytenr);
2956                         BUG_ON(!list_empty(&node->list));
2957                         btrfs_record_root_in_trans(trans, root);
2958                         root = root->reloc_root;
2959                         node->new_bytenr = root->node->start;
2960                         node->root = root;
2961                         list_add_tail(&node->list, &rc->backref_cache.changed);
2962                 } else {
2963                         path->lowest_level = node->level;
2964                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2965                         btrfs_release_path(path);
2966                         if (ret > 0)
2967                                 ret = 0;
2968                 }
2969                 if (!ret)
2970                         update_processed_blocks(rc, node);
2971         } else {
2972                 ret = do_relocation(trans, rc, node, key, path, 1);
2973         }
2974 out:
2975         if (ret || node->level == 0 || node->cowonly)
2976                 remove_backref_node(&rc->backref_cache, node);
2977         return ret;
2978 }
2979
2980 /*
2981  * relocate a list of blocks
2982  */
2983 static noinline_for_stack
2984 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2985                          struct reloc_control *rc, struct rb_root *blocks)
2986 {
2987         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2988         struct backref_node *node;
2989         struct btrfs_path *path;
2990         struct tree_block *block;
2991         struct rb_node *rb_node;
2992         int ret;
2993         int err = 0;
2994
2995         path = btrfs_alloc_path();
2996         if (!path) {
2997                 err = -ENOMEM;
2998                 goto out_free_blocks;
2999         }
3000
3001         rb_node = rb_first(blocks);
3002         while (rb_node) {
3003                 block = rb_entry(rb_node, struct tree_block, rb_node);
3004                 if (!block->key_ready)
3005                         readahead_tree_block(fs_info, block->bytenr);
3006                 rb_node = rb_next(rb_node);
3007         }
3008
3009         rb_node = rb_first(blocks);
3010         while (rb_node) {
3011                 block = rb_entry(rb_node, struct tree_block, rb_node);
3012                 if (!block->key_ready) {
3013                         err = get_tree_block_key(fs_info, block);
3014                         if (err)
3015                                 goto out_free_path;
3016                 }
3017                 rb_node = rb_next(rb_node);
3018         }
3019
3020         rb_node = rb_first(blocks);
3021         while (rb_node) {
3022                 block = rb_entry(rb_node, struct tree_block, rb_node);
3023
3024                 node = build_backref_tree(rc, &block->key,
3025                                           block->level, block->bytenr);
3026                 if (IS_ERR(node)) {
3027                         err = PTR_ERR(node);
3028                         goto out;
3029                 }
3030
3031                 ret = relocate_tree_block(trans, rc, node, &block->key,
3032                                           path);
3033                 if (ret < 0) {
3034                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3035                                 err = ret;
3036                         goto out;
3037                 }
3038                 rb_node = rb_next(rb_node);
3039         }
3040 out:
3041         err = finish_pending_nodes(trans, rc, path, err);
3042
3043 out_free_path:
3044         btrfs_free_path(path);
3045 out_free_blocks:
3046         free_block_list(blocks);
3047         return err;
3048 }
3049
3050 static noinline_for_stack
3051 int prealloc_file_extent_cluster(struct inode *inode,
3052                                  struct file_extent_cluster *cluster)
3053 {
3054         u64 alloc_hint = 0;
3055         u64 start;
3056         u64 end;
3057         u64 offset = BTRFS_I(inode)->index_cnt;
3058         u64 num_bytes;
3059         int nr = 0;
3060         int ret = 0;
3061         u64 prealloc_start = cluster->start - offset;
3062         u64 prealloc_end = cluster->end - offset;
3063         u64 cur_offset;
3064         struct extent_changeset *data_reserved = NULL;
3065
3066         BUG_ON(cluster->start != cluster->boundary[0]);
3067         inode_lock(inode);
3068
3069         ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
3070                                           prealloc_end + 1 - prealloc_start);
3071         if (ret)
3072                 goto out;
3073
3074         cur_offset = prealloc_start;
3075         while (nr < cluster->nr) {
3076                 start = cluster->boundary[nr] - offset;
3077                 if (nr + 1 < cluster->nr)
3078                         end = cluster->boundary[nr + 1] - 1 - offset;
3079                 else
3080                         end = cluster->end - offset;
3081
3082                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3083                 num_bytes = end + 1 - start;
3084                 if (cur_offset < start)
3085                         btrfs_free_reserved_data_space(inode, data_reserved,
3086                                         cur_offset, start - cur_offset);
3087                 ret = btrfs_prealloc_file_range(inode, 0, start,
3088                                                 num_bytes, num_bytes,
3089                                                 end + 1, &alloc_hint);
3090                 cur_offset = end + 1;
3091                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3092                 if (ret)
3093                         break;
3094                 nr++;
3095         }
3096         if (cur_offset < prealloc_end)
3097                 btrfs_free_reserved_data_space(inode, data_reserved,
3098                                 cur_offset, prealloc_end + 1 - cur_offset);
3099 out:
3100         inode_unlock(inode);
3101         extent_changeset_free(data_reserved);
3102         return ret;
3103 }
3104
3105 static noinline_for_stack
3106 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3107                          u64 block_start)
3108 {
3109         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3110         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3111         struct extent_map *em;
3112         int ret = 0;
3113
3114         em = alloc_extent_map();
3115         if (!em)
3116                 return -ENOMEM;
3117
3118         em->start = start;
3119         em->len = end + 1 - start;
3120         em->block_len = em->len;
3121         em->block_start = block_start;
3122         em->bdev = fs_info->fs_devices->latest_bdev;
3123         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3124
3125         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3126         while (1) {
3127                 write_lock(&em_tree->lock);
3128                 ret = add_extent_mapping(em_tree, em, 0);
3129                 write_unlock(&em_tree->lock);
3130                 if (ret != -EEXIST) {
3131                         free_extent_map(em);
3132                         break;
3133                 }
3134                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3135         }
3136         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3137         return ret;
3138 }
3139
3140 static int relocate_file_extent_cluster(struct inode *inode,
3141                                         struct file_extent_cluster *cluster)
3142 {
3143         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3144         u64 page_start;
3145         u64 page_end;
3146         u64 offset = BTRFS_I(inode)->index_cnt;
3147         unsigned long index;
3148         unsigned long last_index;
3149         struct page *page;
3150         struct file_ra_state *ra;
3151         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3152         int nr = 0;
3153         int ret = 0;
3154
3155         if (!cluster->nr)
3156                 return 0;
3157
3158         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3159         if (!ra)
3160                 return -ENOMEM;
3161
3162         ret = prealloc_file_extent_cluster(inode, cluster);
3163         if (ret)
3164                 goto out;
3165
3166         file_ra_state_init(ra, inode->i_mapping);
3167
3168         ret = setup_extent_mapping(inode, cluster->start - offset,
3169                                    cluster->end - offset, cluster->start);
3170         if (ret)
3171                 goto out;
3172
3173         index = (cluster->start - offset) >> PAGE_SHIFT;
3174         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3175         while (index <= last_index) {
3176                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3177                                 PAGE_SIZE);
3178                 if (ret)
3179                         goto out;
3180
3181                 page = find_lock_page(inode->i_mapping, index);
3182                 if (!page) {
3183                         page_cache_sync_readahead(inode->i_mapping,
3184                                                   ra, NULL, index,
3185                                                   last_index + 1 - index);
3186                         page = find_or_create_page(inode->i_mapping, index,
3187                                                    mask);
3188                         if (!page) {
3189                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3190                                                         PAGE_SIZE, true);
3191                                 ret = -ENOMEM;
3192                                 goto out;
3193                         }
3194                 }
3195
3196                 if (PageReadahead(page)) {
3197                         page_cache_async_readahead(inode->i_mapping,
3198                                                    ra, NULL, page, index,
3199                                                    last_index + 1 - index);
3200                 }
3201
3202                 if (!PageUptodate(page)) {
3203                         btrfs_readpage(NULL, page);
3204                         lock_page(page);
3205                         if (!PageUptodate(page)) {
3206                                 unlock_page(page);
3207                                 put_page(page);
3208                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3209                                                         PAGE_SIZE, true);
3210                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
3211                                                                PAGE_SIZE, true);
3212                                 ret = -EIO;
3213                                 goto out;
3214                         }
3215                 }
3216
3217                 page_start = page_offset(page);
3218                 page_end = page_start + PAGE_SIZE - 1;
3219
3220                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3221
3222                 set_page_extent_mapped(page);
3223
3224                 if (nr < cluster->nr &&
3225                     page_start + offset == cluster->boundary[nr]) {
3226                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3227                                         page_start, page_end,
3228                                         EXTENT_BOUNDARY);
3229                         nr++;
3230                 }
3231
3232                 ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
3233                                                 NULL, 0);
3234                 if (ret) {
3235                         unlock_page(page);
3236                         put_page(page);
3237                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3238                                                          PAGE_SIZE, true);
3239                         btrfs_delalloc_release_extents(BTRFS_I(inode),
3240                                                        PAGE_SIZE, true);
3241
3242                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
3243                                           page_start, page_end,
3244                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
3245                         goto out;
3246
3247                 }
3248                 set_page_dirty(page);
3249
3250                 unlock_extent(&BTRFS_I(inode)->io_tree,
3251                               page_start, page_end);
3252                 unlock_page(page);
3253                 put_page(page);
3254
3255                 index++;
3256                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE,
3257                                                false);
3258                 balance_dirty_pages_ratelimited(inode->i_mapping);
3259                 btrfs_throttle(fs_info);
3260         }
3261         WARN_ON(nr != cluster->nr);
3262 out:
3263         kfree(ra);
3264         return ret;
3265 }
3266
3267 static noinline_for_stack
3268 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3269                          struct file_extent_cluster *cluster)
3270 {
3271         int ret;
3272
3273         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3274                 ret = relocate_file_extent_cluster(inode, cluster);
3275                 if (ret)
3276                         return ret;
3277                 cluster->nr = 0;
3278         }
3279
3280         if (!cluster->nr)
3281                 cluster->start = extent_key->objectid;
3282         else
3283                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3284         cluster->end = extent_key->objectid + extent_key->offset - 1;
3285         cluster->boundary[cluster->nr] = extent_key->objectid;
3286         cluster->nr++;
3287
3288         if (cluster->nr >= MAX_EXTENTS) {
3289                 ret = relocate_file_extent_cluster(inode, cluster);
3290                 if (ret)
3291                         return ret;
3292                 cluster->nr = 0;
3293         }
3294         return 0;
3295 }
3296
3297 /*
3298  * helper to add a tree block to the list.
3299  * the major work is getting the generation and level of the block
3300  */
3301 static int add_tree_block(struct reloc_control *rc,
3302                           struct btrfs_key *extent_key,
3303                           struct btrfs_path *path,
3304                           struct rb_root *blocks)
3305 {
3306         struct extent_buffer *eb;
3307         struct btrfs_extent_item *ei;
3308         struct btrfs_tree_block_info *bi;
3309         struct tree_block *block;
3310         struct rb_node *rb_node;
3311         u32 item_size;
3312         int level = -1;
3313         u64 generation;
3314
3315         eb =  path->nodes[0];
3316         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3317
3318         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3319             item_size >= sizeof(*ei) + sizeof(*bi)) {
3320                 ei = btrfs_item_ptr(eb, path->slots[0],
3321                                 struct btrfs_extent_item);
3322                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3323                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3324                         level = btrfs_tree_block_level(eb, bi);
3325                 } else {
3326                         level = (int)extent_key->offset;
3327                 }
3328                 generation = btrfs_extent_generation(eb, ei);
3329         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3330                 btrfs_print_v0_err(eb->fs_info);
3331                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3332                 return -EINVAL;
3333         } else {
3334                 BUG();
3335         }
3336
3337         btrfs_release_path(path);
3338
3339         BUG_ON(level == -1);
3340
3341         block = kmalloc(sizeof(*block), GFP_NOFS);
3342         if (!block)
3343                 return -ENOMEM;
3344
3345         block->bytenr = extent_key->objectid;
3346         block->key.objectid = rc->extent_root->fs_info->nodesize;
3347         block->key.offset = generation;
3348         block->level = level;
3349         block->key_ready = 0;
3350
3351         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3352         if (rb_node)
3353                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3354
3355         return 0;
3356 }
3357
3358 /*
3359  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3360  */
3361 static int __add_tree_block(struct reloc_control *rc,
3362                             u64 bytenr, u32 blocksize,
3363                             struct rb_root *blocks)
3364 {
3365         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3366         struct btrfs_path *path;
3367         struct btrfs_key key;
3368         int ret;
3369         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3370
3371         if (tree_block_processed(bytenr, rc))
3372                 return 0;
3373
3374         if (tree_search(blocks, bytenr))
3375                 return 0;
3376
3377         path = btrfs_alloc_path();
3378         if (!path)
3379                 return -ENOMEM;
3380 again:
3381         key.objectid = bytenr;
3382         if (skinny) {
3383                 key.type = BTRFS_METADATA_ITEM_KEY;
3384                 key.offset = (u64)-1;
3385         } else {
3386                 key.type = BTRFS_EXTENT_ITEM_KEY;
3387                 key.offset = blocksize;
3388         }
3389
3390         path->search_commit_root = 1;
3391         path->skip_locking = 1;
3392         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3393         if (ret < 0)
3394                 goto out;
3395
3396         if (ret > 0 && skinny) {
3397                 if (path->slots[0]) {
3398                         path->slots[0]--;
3399                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3400                                               path->slots[0]);
3401                         if (key.objectid == bytenr &&
3402                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3403                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3404                               key.offset == blocksize)))
3405                                 ret = 0;
3406                 }
3407
3408                 if (ret) {
3409                         skinny = false;
3410                         btrfs_release_path(path);
3411                         goto again;
3412                 }
3413         }
3414         if (ret) {
3415                 ASSERT(ret == 1);
3416                 btrfs_print_leaf(path->nodes[0]);
3417                 btrfs_err(fs_info,
3418              "tree block extent item (%llu) is not found in extent tree",
3419                      bytenr);
3420                 WARN_ON(1);
3421                 ret = -EINVAL;
3422                 goto out;
3423         }
3424
3425         ret = add_tree_block(rc, &key, path, blocks);
3426 out:
3427         btrfs_free_path(path);
3428         return ret;
3429 }
3430
3431 /*
3432  * helper to check if the block use full backrefs for pointers in it
3433  */
3434 static int block_use_full_backref(struct reloc_control *rc,
3435                                   struct extent_buffer *eb)
3436 {
3437         u64 flags;
3438         int ret;
3439
3440         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3441             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3442                 return 1;
3443
3444         ret = btrfs_lookup_extent_info(NULL, rc->extent_root->fs_info,
3445                                        eb->start, btrfs_header_level(eb), 1,
3446                                        NULL, &flags);
3447         BUG_ON(ret);
3448
3449         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3450                 ret = 1;
3451         else
3452                 ret = 0;
3453         return ret;
3454 }
3455
3456 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3457                                     struct btrfs_block_group_cache *block_group,
3458                                     struct inode *inode,
3459                                     u64 ino)
3460 {
3461         struct btrfs_key key;
3462         struct btrfs_root *root = fs_info->tree_root;
3463         struct btrfs_trans_handle *trans;
3464         int ret = 0;
3465
3466         if (inode)
3467                 goto truncate;
3468
3469         key.objectid = ino;
3470         key.type = BTRFS_INODE_ITEM_KEY;
3471         key.offset = 0;
3472
3473         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3474         if (IS_ERR(inode))
3475                 return -ENOENT;
3476
3477 truncate:
3478         ret = btrfs_check_trunc_cache_free_space(fs_info,
3479                                                  &fs_info->global_block_rsv);
3480         if (ret)
3481                 goto out;
3482
3483         trans = btrfs_join_transaction(root);
3484         if (IS_ERR(trans)) {
3485                 ret = PTR_ERR(trans);
3486                 goto out;
3487         }
3488
3489         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3490
3491         btrfs_end_transaction(trans);
3492         btrfs_btree_balance_dirty(fs_info);
3493 out:
3494         iput(inode);
3495         return ret;
3496 }
3497
3498 /*
3499  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3500  * this function scans fs tree to find blocks reference the data extent
3501  */
3502 static int find_data_references(struct reloc_control *rc,
3503                                 struct btrfs_key *extent_key,
3504                                 struct extent_buffer *leaf,
3505                                 struct btrfs_extent_data_ref *ref,
3506                                 struct rb_root *blocks)
3507 {
3508         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3509         struct btrfs_path *path;
3510         struct tree_block *block;
3511         struct btrfs_root *root;
3512         struct btrfs_file_extent_item *fi;
3513         struct rb_node *rb_node;
3514         struct btrfs_key key;
3515         u64 ref_root;
3516         u64 ref_objectid;
3517         u64 ref_offset;
3518         u32 ref_count;
3519         u32 nritems;
3520         int err = 0;
3521         int added = 0;
3522         int counted;
3523         int ret;
3524
3525         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3526         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3527         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3528         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3529
3530         /*
3531          * This is an extent belonging to the free space cache, lets just delete
3532          * it and redo the search.
3533          */
3534         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3535                 ret = delete_block_group_cache(fs_info, rc->block_group,
3536                                                NULL, ref_objectid);
3537                 if (ret != -ENOENT)
3538                         return ret;
3539                 ret = 0;
3540         }
3541
3542         path = btrfs_alloc_path();
3543         if (!path)
3544                 return -ENOMEM;
3545         path->reada = READA_FORWARD;
3546
3547         root = read_fs_root(fs_info, ref_root);
3548         if (IS_ERR(root)) {
3549                 err = PTR_ERR(root);
3550                 goto out;
3551         }
3552
3553         key.objectid = ref_objectid;
3554         key.type = BTRFS_EXTENT_DATA_KEY;
3555         if (ref_offset > ((u64)-1 << 32))
3556                 key.offset = 0;
3557         else
3558                 key.offset = ref_offset;
3559
3560         path->search_commit_root = 1;
3561         path->skip_locking = 1;
3562         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3563         if (ret < 0) {
3564                 err = ret;
3565                 goto out;
3566         }
3567
3568         leaf = path->nodes[0];
3569         nritems = btrfs_header_nritems(leaf);
3570         /*
3571          * the references in tree blocks that use full backrefs
3572          * are not counted in
3573          */
3574         if (block_use_full_backref(rc, leaf))
3575                 counted = 0;
3576         else
3577                 counted = 1;
3578         rb_node = tree_search(blocks, leaf->start);
3579         if (rb_node) {
3580                 if (counted)
3581                         added = 1;
3582                 else
3583                         path->slots[0] = nritems;
3584         }
3585
3586         while (ref_count > 0) {
3587                 while (path->slots[0] >= nritems) {
3588                         ret = btrfs_next_leaf(root, path);
3589                         if (ret < 0) {
3590                                 err = ret;
3591                                 goto out;
3592                         }
3593                         if (WARN_ON(ret > 0))
3594                                 goto out;
3595
3596                         leaf = path->nodes[0];
3597                         nritems = btrfs_header_nritems(leaf);
3598                         added = 0;
3599
3600                         if (block_use_full_backref(rc, leaf))
3601                                 counted = 0;
3602                         else
3603                                 counted = 1;
3604                         rb_node = tree_search(blocks, leaf->start);
3605                         if (rb_node) {
3606                                 if (counted)
3607                                         added = 1;
3608                                 else
3609                                         path->slots[0] = nritems;
3610                         }
3611                 }
3612
3613                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3614                 if (WARN_ON(key.objectid != ref_objectid ||
3615                     key.type != BTRFS_EXTENT_DATA_KEY))
3616                         break;
3617
3618                 fi = btrfs_item_ptr(leaf, path->slots[0],
3619                                     struct btrfs_file_extent_item);
3620
3621                 if (btrfs_file_extent_type(leaf, fi) ==
3622                     BTRFS_FILE_EXTENT_INLINE)
3623                         goto next;
3624
3625                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3626                     extent_key->objectid)
3627                         goto next;
3628
3629                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3630                 if (key.offset != ref_offset)
3631                         goto next;
3632
3633                 if (counted)
3634                         ref_count--;
3635                 if (added)
3636                         goto next;
3637
3638                 if (!tree_block_processed(leaf->start, rc)) {
3639                         block = kmalloc(sizeof(*block), GFP_NOFS);
3640                         if (!block) {
3641                                 err = -ENOMEM;
3642                                 break;
3643                         }
3644                         block->bytenr = leaf->start;
3645                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3646                         block->level = 0;
3647                         block->key_ready = 1;
3648                         rb_node = tree_insert(blocks, block->bytenr,
3649                                               &block->rb_node);
3650                         if (rb_node)
3651                                 backref_tree_panic(rb_node, -EEXIST,
3652                                                    block->bytenr);
3653                 }
3654                 if (counted)
3655                         added = 1;
3656                 else
3657                         path->slots[0] = nritems;
3658 next:
3659                 path->slots[0]++;
3660
3661         }
3662 out:
3663         btrfs_free_path(path);
3664         return err;
3665 }
3666
3667 /*
3668  * helper to find all tree blocks that reference a given data extent
3669  */
3670 static noinline_for_stack
3671 int add_data_references(struct reloc_control *rc,
3672                         struct btrfs_key *extent_key,
3673                         struct btrfs_path *path,
3674                         struct rb_root *blocks)
3675 {
3676         struct btrfs_key key;
3677         struct extent_buffer *eb;
3678         struct btrfs_extent_data_ref *dref;
3679         struct btrfs_extent_inline_ref *iref;
3680         unsigned long ptr;
3681         unsigned long end;
3682         u32 blocksize = rc->extent_root->fs_info->nodesize;
3683         int ret = 0;
3684         int err = 0;
3685
3686         eb = path->nodes[0];
3687         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3688         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3689         ptr += sizeof(struct btrfs_extent_item);
3690
3691         while (ptr < end) {
3692                 iref = (struct btrfs_extent_inline_ref *)ptr;
3693                 key.type = btrfs_get_extent_inline_ref_type(eb, iref,
3694                                                         BTRFS_REF_TYPE_DATA);
3695                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3696                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3697                         ret = __add_tree_block(rc, key.offset, blocksize,
3698                                                blocks);
3699                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3700                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3701                         ret = find_data_references(rc, extent_key,
3702                                                    eb, dref, blocks);
3703                 } else {
3704                         ret = -EUCLEAN;
3705                         btrfs_err(rc->extent_root->fs_info,
3706                      "extent %llu slot %d has an invalid inline ref type",
3707                              eb->start, path->slots[0]);
3708                 }
3709                 if (ret) {
3710                         err = ret;
3711                         goto out;
3712                 }
3713                 ptr += btrfs_extent_inline_ref_size(key.type);
3714         }
3715         WARN_ON(ptr > end);
3716
3717         while (1) {
3718                 cond_resched();
3719                 eb = path->nodes[0];
3720                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3721                         ret = btrfs_next_leaf(rc->extent_root, path);
3722                         if (ret < 0) {
3723                                 err = ret;
3724                                 break;
3725                         }
3726                         if (ret > 0)
3727                                 break;
3728                         eb = path->nodes[0];
3729                 }
3730
3731                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3732                 if (key.objectid != extent_key->objectid)
3733                         break;
3734
3735                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3736                         ret = __add_tree_block(rc, key.offset, blocksize,
3737                                                blocks);
3738                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3739                         dref = btrfs_item_ptr(eb, path->slots[0],
3740                                               struct btrfs_extent_data_ref);
3741                         ret = find_data_references(rc, extent_key,
3742                                                    eb, dref, blocks);
3743                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
3744                         btrfs_print_v0_err(eb->fs_info);
3745                         btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3746                         ret = -EINVAL;
3747                 } else {
3748                         ret = 0;
3749                 }
3750                 if (ret) {
3751                         err = ret;
3752                         break;
3753                 }
3754                 path->slots[0]++;
3755         }
3756 out:
3757         btrfs_release_path(path);
3758         if (err)
3759                 free_block_list(blocks);
3760         return err;
3761 }
3762
3763 /*
3764  * helper to find next unprocessed extent
3765  */
3766 static noinline_for_stack
3767 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3768                      struct btrfs_key *extent_key)
3769 {
3770         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3771         struct btrfs_key key;
3772         struct extent_buffer *leaf;
3773         u64 start, end, last;
3774         int ret;
3775
3776         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3777         while (1) {
3778                 cond_resched();
3779                 if (rc->search_start >= last) {
3780                         ret = 1;
3781                         break;
3782                 }
3783
3784                 key.objectid = rc->search_start;
3785                 key.type = BTRFS_EXTENT_ITEM_KEY;
3786                 key.offset = 0;
3787
3788                 path->search_commit_root = 1;
3789                 path->skip_locking = 1;
3790                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3791                                         0, 0);
3792                 if (ret < 0)
3793                         break;
3794 next:
3795                 leaf = path->nodes[0];
3796                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3797                         ret = btrfs_next_leaf(rc->extent_root, path);
3798                         if (ret != 0)
3799                                 break;
3800                         leaf = path->nodes[0];
3801                 }
3802
3803                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3804                 if (key.objectid >= last) {
3805                         ret = 1;
3806                         break;
3807                 }
3808
3809                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3810                     key.type != BTRFS_METADATA_ITEM_KEY) {
3811                         path->slots[0]++;
3812                         goto next;
3813                 }
3814
3815                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3816                     key.objectid + key.offset <= rc->search_start) {
3817                         path->slots[0]++;
3818                         goto next;
3819                 }
3820
3821                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3822                     key.objectid + fs_info->nodesize <=
3823                     rc->search_start) {
3824                         path->slots[0]++;
3825                         goto next;
3826                 }
3827
3828                 ret = find_first_extent_bit(&rc->processed_blocks,
3829                                             key.objectid, &start, &end,
3830                                             EXTENT_DIRTY, NULL);
3831
3832                 if (ret == 0 && start <= key.objectid) {
3833                         btrfs_release_path(path);
3834                         rc->search_start = end + 1;
3835                 } else {
3836                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3837                                 rc->search_start = key.objectid + key.offset;
3838                         else
3839                                 rc->search_start = key.objectid +
3840                                         fs_info->nodesize;
3841                         memcpy(extent_key, &key, sizeof(key));
3842                         return 0;
3843                 }
3844         }
3845         btrfs_release_path(path);
3846         return ret;
3847 }
3848
3849 static void set_reloc_control(struct reloc_control *rc)
3850 {
3851         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3852
3853         mutex_lock(&fs_info->reloc_mutex);
3854         fs_info->reloc_ctl = rc;
3855         mutex_unlock(&fs_info->reloc_mutex);
3856 }
3857
3858 static void unset_reloc_control(struct reloc_control *rc)
3859 {
3860         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3861
3862         mutex_lock(&fs_info->reloc_mutex);
3863         fs_info->reloc_ctl = NULL;
3864         mutex_unlock(&fs_info->reloc_mutex);
3865 }
3866
3867 static int check_extent_flags(u64 flags)
3868 {
3869         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3870             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3871                 return 1;
3872         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3873             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3874                 return 1;
3875         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3876             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3877                 return 1;
3878         return 0;
3879 }
3880
3881 static noinline_for_stack
3882 int prepare_to_relocate(struct reloc_control *rc)
3883 {
3884         struct btrfs_trans_handle *trans;
3885         int ret;
3886
3887         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3888                                               BTRFS_BLOCK_RSV_TEMP);
3889         if (!rc->block_rsv)
3890                 return -ENOMEM;
3891
3892         memset(&rc->cluster, 0, sizeof(rc->cluster));
3893         rc->search_start = rc->block_group->key.objectid;
3894         rc->extents_found = 0;
3895         rc->nodes_relocated = 0;
3896         rc->merging_rsv_size = 0;
3897         rc->reserved_bytes = 0;
3898         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3899                               RELOCATION_RESERVED_NODES;
3900         ret = btrfs_block_rsv_refill(rc->extent_root,
3901                                      rc->block_rsv, rc->block_rsv->size,
3902                                      BTRFS_RESERVE_FLUSH_ALL);
3903         if (ret)
3904                 return ret;
3905
3906         rc->create_reloc_tree = 1;
3907         set_reloc_control(rc);
3908
3909         trans = btrfs_join_transaction(rc->extent_root);
3910         if (IS_ERR(trans)) {
3911                 unset_reloc_control(rc);
3912                 /*
3913                  * extent tree is not a ref_cow tree and has no reloc_root to
3914                  * cleanup.  And callers are responsible to free the above
3915                  * block rsv.
3916                  */
3917                 return PTR_ERR(trans);
3918         }
3919         btrfs_commit_transaction(trans);
3920         return 0;
3921 }
3922
3923 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3924 {
3925         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3926         struct rb_root blocks = RB_ROOT;
3927         struct btrfs_key key;
3928         struct btrfs_trans_handle *trans = NULL;
3929         struct btrfs_path *path;
3930         struct btrfs_extent_item *ei;
3931         u64 flags;
3932         u32 item_size;
3933         int ret;
3934         int err = 0;
3935         int progress = 0;
3936
3937         path = btrfs_alloc_path();
3938         if (!path)
3939                 return -ENOMEM;
3940         path->reada = READA_FORWARD;
3941
3942         ret = prepare_to_relocate(rc);
3943         if (ret) {
3944                 err = ret;
3945                 goto out_free;
3946         }
3947
3948         while (1) {
3949                 rc->reserved_bytes = 0;
3950                 ret = btrfs_block_rsv_refill(rc->extent_root,
3951                                         rc->block_rsv, rc->block_rsv->size,
3952                                         BTRFS_RESERVE_FLUSH_ALL);
3953                 if (ret) {
3954                         err = ret;
3955                         break;
3956                 }
3957                 progress++;
3958                 trans = btrfs_start_transaction(rc->extent_root, 0);
3959                 if (IS_ERR(trans)) {
3960                         err = PTR_ERR(trans);
3961                         trans = NULL;
3962                         break;
3963                 }
3964 restart:
3965                 if (update_backref_cache(trans, &rc->backref_cache)) {
3966                         btrfs_end_transaction(trans);
3967                         continue;
3968                 }
3969
3970                 ret = find_next_extent(rc, path, &key);
3971                 if (ret < 0)
3972                         err = ret;
3973                 if (ret != 0)
3974                         break;
3975
3976                 rc->extents_found++;
3977
3978                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3979                                     struct btrfs_extent_item);
3980                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3981                 if (item_size >= sizeof(*ei)) {
3982                         flags = btrfs_extent_flags(path->nodes[0], ei);
3983                         ret = check_extent_flags(flags);
3984                         BUG_ON(ret);
3985                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3986                         err = -EINVAL;
3987                         btrfs_print_v0_err(trans->fs_info);
3988                         btrfs_abort_transaction(trans, err);
3989                         break;
3990                 } else {
3991                         BUG();
3992                 }
3993
3994                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3995                         ret = add_tree_block(rc, &key, path, &blocks);
3996                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3997                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3998                         ret = add_data_references(rc, &key, path, &blocks);
3999                 } else {
4000                         btrfs_release_path(path);
4001                         ret = 0;
4002                 }
4003                 if (ret < 0) {
4004                         err = ret;
4005                         break;
4006                 }
4007
4008                 if (!RB_EMPTY_ROOT(&blocks)) {
4009                         ret = relocate_tree_blocks(trans, rc, &blocks);
4010                         if (ret < 0) {
4011                                 /*
4012                                  * if we fail to relocate tree blocks, force to update
4013                                  * backref cache when committing transaction.
4014                                  */
4015                                 rc->backref_cache.last_trans = trans->transid - 1;
4016
4017                                 if (ret != -EAGAIN) {
4018                                         err = ret;
4019                                         break;
4020                                 }
4021                                 rc->extents_found--;
4022                                 rc->search_start = key.objectid;
4023                         }
4024                 }
4025
4026                 btrfs_end_transaction_throttle(trans);
4027                 btrfs_btree_balance_dirty(fs_info);
4028                 trans = NULL;
4029
4030                 if (rc->stage == MOVE_DATA_EXTENTS &&
4031                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4032                         rc->found_file_extent = 1;
4033                         ret = relocate_data_extent(rc->data_inode,
4034                                                    &key, &rc->cluster);
4035                         if (ret < 0) {
4036                                 err = ret;
4037                                 break;
4038                         }
4039                 }
4040         }
4041         if (trans && progress && err == -ENOSPC) {
4042                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
4043                 if (ret == 1) {
4044                         err = 0;
4045                         progress = 0;
4046                         goto restart;
4047                 }
4048         }
4049
4050         btrfs_release_path(path);
4051         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4052
4053         if (trans) {
4054                 btrfs_end_transaction_throttle(trans);
4055                 btrfs_btree_balance_dirty(fs_info);
4056         }
4057
4058         if (!err) {
4059                 ret = relocate_file_extent_cluster(rc->data_inode,
4060                                                    &rc->cluster);
4061                 if (ret < 0)
4062                         err = ret;
4063         }
4064
4065         rc->create_reloc_tree = 0;
4066         set_reloc_control(rc);
4067
4068         backref_cache_cleanup(&rc->backref_cache);
4069         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4070
4071         err = prepare_to_merge(rc, err);
4072
4073         merge_reloc_roots(rc);
4074
4075         rc->merge_reloc_tree = 0;
4076         unset_reloc_control(rc);
4077         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4078
4079         /* get rid of pinned extents */
4080         trans = btrfs_join_transaction(rc->extent_root);
4081         if (IS_ERR(trans)) {
4082                 err = PTR_ERR(trans);
4083                 goto out_free;
4084         }
4085         btrfs_commit_transaction(trans);
4086 out_free:
4087         btrfs_free_block_rsv(fs_info, rc->block_rsv);
4088         btrfs_free_path(path);
4089         return err;
4090 }
4091
4092 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4093                                  struct btrfs_root *root, u64 objectid)
4094 {
4095         struct btrfs_path *path;
4096         struct btrfs_inode_item *item;
4097         struct extent_buffer *leaf;
4098         int ret;
4099
4100         path = btrfs_alloc_path();
4101         if (!path)
4102                 return -ENOMEM;
4103
4104         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4105         if (ret)
4106                 goto out;
4107
4108         leaf = path->nodes[0];
4109         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4110         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
4111         btrfs_set_inode_generation(leaf, item, 1);
4112         btrfs_set_inode_size(leaf, item, 0);
4113         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4114         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4115                                           BTRFS_INODE_PREALLOC);
4116         btrfs_mark_buffer_dirty(leaf);
4117 out:
4118         btrfs_free_path(path);
4119         return ret;
4120 }
4121
4122 /*
4123  * helper to create inode for data relocation.
4124  * the inode is in data relocation tree and its link count is 0
4125  */
4126 static noinline_for_stack
4127 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4128                                  struct btrfs_block_group_cache *group)
4129 {
4130         struct inode *inode = NULL;
4131         struct btrfs_trans_handle *trans;
4132         struct btrfs_root *root;
4133         struct btrfs_key key;
4134         u64 objectid;
4135         int err = 0;
4136
4137         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4138         if (IS_ERR(root))
4139                 return ERR_CAST(root);
4140
4141         trans = btrfs_start_transaction(root, 6);
4142         if (IS_ERR(trans))
4143                 return ERR_CAST(trans);
4144
4145         err = btrfs_find_free_objectid(root, &objectid);
4146         if (err)
4147                 goto out;
4148
4149         err = __insert_orphan_inode(trans, root, objectid);
4150         BUG_ON(err);
4151
4152         key.objectid = objectid;
4153         key.type = BTRFS_INODE_ITEM_KEY;
4154         key.offset = 0;
4155         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4156         BUG_ON(IS_ERR(inode));
4157         BTRFS_I(inode)->index_cnt = group->key.objectid;
4158
4159         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4160 out:
4161         btrfs_end_transaction(trans);
4162         btrfs_btree_balance_dirty(fs_info);
4163         if (err) {
4164                 if (inode)
4165                         iput(inode);
4166                 inode = ERR_PTR(err);
4167         }
4168         return inode;
4169 }
4170
4171 static struct reloc_control *alloc_reloc_control(void)
4172 {
4173         struct reloc_control *rc;
4174
4175         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4176         if (!rc)
4177                 return NULL;
4178
4179         INIT_LIST_HEAD(&rc->reloc_roots);
4180         backref_cache_init(&rc->backref_cache);
4181         mapping_tree_init(&rc->reloc_root_tree);
4182         extent_io_tree_init(&rc->processed_blocks, NULL);
4183         return rc;
4184 }
4185
4186 /*
4187  * Print the block group being relocated
4188  */
4189 static void describe_relocation(struct btrfs_fs_info *fs_info,
4190                                 struct btrfs_block_group_cache *block_group)
4191 {
4192         char buf[128];          /* prefixed by a '|' that'll be dropped */
4193         u64 flags = block_group->flags;
4194
4195         /* Shouldn't happen */
4196         if (!flags) {
4197                 strcpy(buf, "|NONE");
4198         } else {
4199                 char *bp = buf;
4200
4201 #define DESCRIBE_FLAG(f, d) \
4202                 if (flags & BTRFS_BLOCK_GROUP_##f) { \
4203                         bp += snprintf(bp, buf - bp + sizeof(buf), "|%s", d); \
4204                         flags &= ~BTRFS_BLOCK_GROUP_##f; \
4205                 }
4206                 DESCRIBE_FLAG(DATA,     "data");
4207                 DESCRIBE_FLAG(SYSTEM,   "system");
4208                 DESCRIBE_FLAG(METADATA, "metadata");
4209                 DESCRIBE_FLAG(RAID0,    "raid0");
4210                 DESCRIBE_FLAG(RAID1,    "raid1");
4211                 DESCRIBE_FLAG(DUP,      "dup");
4212                 DESCRIBE_FLAG(RAID10,   "raid10");
4213                 DESCRIBE_FLAG(RAID5,    "raid5");
4214                 DESCRIBE_FLAG(RAID6,    "raid6");
4215                 if (flags)
4216                         snprintf(bp, buf - bp + sizeof(buf), "|0x%llx", flags);
4217 #undef DESCRIBE_FLAG
4218         }
4219
4220         btrfs_info(fs_info,
4221                    "relocating block group %llu flags %s",
4222                    block_group->key.objectid, buf + 1);
4223 }
4224
4225 /*
4226  * function to relocate all extents in a block group.
4227  */
4228 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4229 {
4230         struct btrfs_root *extent_root = fs_info->extent_root;
4231         struct reloc_control *rc;
4232         struct inode *inode;
4233         struct btrfs_path *path;
4234         int ret;
4235         int rw = 0;
4236         int err = 0;
4237
4238         rc = alloc_reloc_control();
4239         if (!rc)
4240                 return -ENOMEM;
4241
4242         rc->extent_root = extent_root;
4243
4244         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4245         BUG_ON(!rc->block_group);
4246
4247         ret = btrfs_inc_block_group_ro(rc->block_group);
4248         if (ret) {
4249                 err = ret;
4250                 goto out;
4251         }
4252         rw = 1;
4253
4254         path = btrfs_alloc_path();
4255         if (!path) {
4256                 err = -ENOMEM;
4257                 goto out;
4258         }
4259
4260         inode = lookup_free_space_inode(fs_info, rc->block_group, path);
4261         btrfs_free_path(path);
4262
4263         if (!IS_ERR(inode))
4264                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4265         else
4266                 ret = PTR_ERR(inode);
4267
4268         if (ret && ret != -ENOENT) {
4269                 err = ret;
4270                 goto out;
4271         }
4272
4273         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4274         if (IS_ERR(rc->data_inode)) {
4275                 err = PTR_ERR(rc->data_inode);
4276                 rc->data_inode = NULL;
4277                 goto out;
4278         }
4279
4280         describe_relocation(fs_info, rc->block_group);
4281
4282         btrfs_wait_block_group_reservations(rc->block_group);
4283         btrfs_wait_nocow_writers(rc->block_group);
4284         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4285                                  rc->block_group->key.objectid,
4286                                  rc->block_group->key.offset);
4287
4288         while (1) {
4289                 mutex_lock(&fs_info->cleaner_mutex);
4290                 ret = relocate_block_group(rc);
4291                 mutex_unlock(&fs_info->cleaner_mutex);
4292                 if (ret < 0) {
4293                         err = ret;
4294                         goto out;
4295                 }
4296
4297                 if (rc->extents_found == 0)
4298                         break;
4299
4300                 btrfs_info(fs_info, "found %llu extents", rc->extents_found);
4301
4302                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4303                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4304                                                        (u64)-1);
4305                         if (ret) {
4306                                 err = ret;
4307                                 goto out;
4308                         }
4309                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4310                                                  0, -1);
4311                         rc->stage = UPDATE_DATA_PTRS;
4312                 }
4313         }
4314
4315         WARN_ON(rc->block_group->pinned > 0);
4316         WARN_ON(rc->block_group->reserved > 0);
4317         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4318 out:
4319         if (err && rw)
4320                 btrfs_dec_block_group_ro(rc->block_group);
4321         iput(rc->data_inode);
4322         btrfs_put_block_group(rc->block_group);
4323         kfree(rc);
4324         return err;
4325 }
4326
4327 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4328 {
4329         struct btrfs_fs_info *fs_info = root->fs_info;
4330         struct btrfs_trans_handle *trans;
4331         int ret, err;
4332
4333         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4334         if (IS_ERR(trans))
4335                 return PTR_ERR(trans);
4336
4337         memset(&root->root_item.drop_progress, 0,
4338                 sizeof(root->root_item.drop_progress));
4339         root->root_item.drop_level = 0;
4340         btrfs_set_root_refs(&root->root_item, 0);
4341         ret = btrfs_update_root(trans, fs_info->tree_root,
4342                                 &root->root_key, &root->root_item);
4343
4344         err = btrfs_end_transaction(trans);
4345         if (err)
4346                 return err;
4347         return ret;
4348 }
4349
4350 /*
4351  * recover relocation interrupted by system crash.
4352  *
4353  * this function resumes merging reloc trees with corresponding fs trees.
4354  * this is important for keeping the sharing of tree blocks
4355  */
4356 int btrfs_recover_relocation(struct btrfs_root *root)
4357 {
4358         struct btrfs_fs_info *fs_info = root->fs_info;
4359         LIST_HEAD(reloc_roots);
4360         struct btrfs_key key;
4361         struct btrfs_root *fs_root;
4362         struct btrfs_root *reloc_root;
4363         struct btrfs_path *path;
4364         struct extent_buffer *leaf;
4365         struct reloc_control *rc = NULL;
4366         struct btrfs_trans_handle *trans;
4367         int ret;
4368         int err = 0;
4369
4370         path = btrfs_alloc_path();
4371         if (!path)
4372                 return -ENOMEM;
4373         path->reada = READA_BACK;
4374
4375         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4376         key.type = BTRFS_ROOT_ITEM_KEY;
4377         key.offset = (u64)-1;
4378
4379         while (1) {
4380                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4381                                         path, 0, 0);
4382                 if (ret < 0) {
4383                         err = ret;
4384                         goto out;
4385                 }
4386                 if (ret > 0) {
4387                         if (path->slots[0] == 0)
4388                                 break;
4389                         path->slots[0]--;
4390                 }
4391                 leaf = path->nodes[0];
4392                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4393                 btrfs_release_path(path);
4394
4395                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4396                     key.type != BTRFS_ROOT_ITEM_KEY)
4397                         break;
4398
4399                 reloc_root = btrfs_read_fs_root(root, &key);
4400                 if (IS_ERR(reloc_root)) {
4401                         err = PTR_ERR(reloc_root);
4402                         goto out;
4403                 }
4404
4405                 list_add(&reloc_root->root_list, &reloc_roots);
4406
4407                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4408                         fs_root = read_fs_root(fs_info,
4409                                                reloc_root->root_key.offset);
4410                         if (IS_ERR(fs_root)) {
4411                                 ret = PTR_ERR(fs_root);
4412                                 if (ret != -ENOENT) {
4413                                         err = ret;
4414                                         goto out;
4415                                 }
4416                                 ret = mark_garbage_root(reloc_root);
4417                                 if (ret < 0) {
4418                                         err = ret;
4419                                         goto out;
4420                                 }
4421                         }
4422                 }
4423
4424                 if (key.offset == 0)
4425                         break;
4426
4427                 key.offset--;
4428         }
4429         btrfs_release_path(path);
4430
4431         if (list_empty(&reloc_roots))
4432                 goto out;
4433
4434         rc = alloc_reloc_control();
4435         if (!rc) {
4436                 err = -ENOMEM;
4437                 goto out;
4438         }
4439
4440         rc->extent_root = fs_info->extent_root;
4441
4442         set_reloc_control(rc);
4443
4444         trans = btrfs_join_transaction(rc->extent_root);
4445         if (IS_ERR(trans)) {
4446                 unset_reloc_control(rc);
4447                 err = PTR_ERR(trans);
4448                 goto out_free;
4449         }
4450
4451         rc->merge_reloc_tree = 1;
4452
4453         while (!list_empty(&reloc_roots)) {
4454                 reloc_root = list_entry(reloc_roots.next,
4455                                         struct btrfs_root, root_list);
4456                 list_del(&reloc_root->root_list);
4457
4458                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4459                         list_add_tail(&reloc_root->root_list,
4460                                       &rc->reloc_roots);
4461                         continue;
4462                 }
4463
4464                 fs_root = read_fs_root(fs_info, reloc_root->root_key.offset);
4465                 if (IS_ERR(fs_root)) {
4466                         err = PTR_ERR(fs_root);
4467                         goto out_free;
4468                 }
4469
4470                 err = __add_reloc_root(reloc_root);
4471                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4472                 fs_root->reloc_root = reloc_root;
4473         }
4474
4475         err = btrfs_commit_transaction(trans);
4476         if (err)
4477                 goto out_free;
4478
4479         merge_reloc_roots(rc);
4480
4481         unset_reloc_control(rc);
4482
4483         trans = btrfs_join_transaction(rc->extent_root);
4484         if (IS_ERR(trans)) {
4485                 err = PTR_ERR(trans);
4486                 goto out_free;
4487         }
4488         err = btrfs_commit_transaction(trans);
4489 out_free:
4490         kfree(rc);
4491 out:
4492         if (!list_empty(&reloc_roots))
4493                 free_reloc_roots(&reloc_roots);
4494
4495         btrfs_free_path(path);
4496
4497         if (err == 0) {
4498                 /* cleanup orphan inode in data relocation tree */
4499                 fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4500                 if (IS_ERR(fs_root))
4501                         err = PTR_ERR(fs_root);
4502                 else
4503                         err = btrfs_orphan_cleanup(fs_root);
4504         }
4505         return err;
4506 }
4507
4508 /*
4509  * helper to add ordered checksum for data relocation.
4510  *
4511  * cloning checksum properly handles the nodatasum extents.
4512  * it also saves CPU time to re-calculate the checksum.
4513  */
4514 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4515 {
4516         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4517         struct btrfs_ordered_sum *sums;
4518         struct btrfs_ordered_extent *ordered;
4519         int ret;
4520         u64 disk_bytenr;
4521         u64 new_bytenr;
4522         LIST_HEAD(list);
4523
4524         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4525         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4526
4527         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4528         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4529                                        disk_bytenr + len - 1, &list, 0);
4530         if (ret)
4531                 goto out;
4532
4533         while (!list_empty(&list)) {
4534                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4535                 list_del_init(&sums->list);
4536
4537                 /*
4538                  * We need to offset the new_bytenr based on where the csum is.
4539                  * We need to do this because we will read in entire prealloc
4540                  * extents but we may have written to say the middle of the
4541                  * prealloc extent, so we need to make sure the csum goes with
4542                  * the right disk offset.
4543                  *
4544                  * We can do this because the data reloc inode refers strictly
4545                  * to the on disk bytes, so we don't have to worry about
4546                  * disk_len vs real len like with real inodes since it's all
4547                  * disk length.
4548                  */
4549                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4550                 sums->bytenr = new_bytenr;
4551
4552                 btrfs_add_ordered_sum(inode, ordered, sums);
4553         }
4554 out:
4555         btrfs_put_ordered_extent(ordered);
4556         return ret;
4557 }
4558
4559 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4560                           struct btrfs_root *root, struct extent_buffer *buf,
4561                           struct extent_buffer *cow)
4562 {
4563         struct btrfs_fs_info *fs_info = root->fs_info;
4564         struct reloc_control *rc;
4565         struct backref_node *node;
4566         int first_cow = 0;
4567         int level;
4568         int ret = 0;
4569
4570         rc = fs_info->reloc_ctl;
4571         if (!rc)
4572                 return 0;
4573
4574         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4575                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4576
4577         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4578                 if (buf == root->node)
4579                         __update_reloc_root(root, cow->start);
4580         }
4581
4582         level = btrfs_header_level(buf);
4583         if (btrfs_header_generation(buf) <=
4584             btrfs_root_last_snapshot(&root->root_item))
4585                 first_cow = 1;
4586
4587         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4588             rc->create_reloc_tree) {
4589                 WARN_ON(!first_cow && level == 0);
4590
4591                 node = rc->backref_cache.path[level];
4592                 BUG_ON(node->bytenr != buf->start &&
4593                        node->new_bytenr != buf->start);
4594
4595                 drop_node_buffer(node);
4596                 extent_buffer_get(cow);
4597                 node->eb = cow;
4598                 node->new_bytenr = cow->start;
4599
4600                 if (!node->pending) {
4601                         list_move_tail(&node->list,
4602                                        &rc->backref_cache.pending[level]);
4603                         node->pending = 1;
4604                 }
4605
4606                 if (first_cow)
4607                         __mark_block_processed(rc, node);
4608
4609                 if (first_cow && level > 0)
4610                         rc->nodes_relocated += buf->len;
4611         }
4612
4613         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4614                 ret = replace_file_extents(trans, rc, root, cow);
4615         return ret;
4616 }
4617
4618 /*
4619  * called before creating snapshot. it calculates metadata reservation
4620  * required for relocating tree blocks in the snapshot
4621  */
4622 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4623                               u64 *bytes_to_reserve)
4624 {
4625         struct btrfs_root *root;
4626         struct reloc_control *rc;
4627
4628         root = pending->root;
4629         if (!root->reloc_root)
4630                 return;
4631
4632         rc = root->fs_info->reloc_ctl;
4633         if (!rc->merge_reloc_tree)
4634                 return;
4635
4636         root = root->reloc_root;
4637         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4638         /*
4639          * relocation is in the stage of merging trees. the space
4640          * used by merging a reloc tree is twice the size of
4641          * relocated tree nodes in the worst case. half for cowing
4642          * the reloc tree, half for cowing the fs tree. the space
4643          * used by cowing the reloc tree will be freed after the
4644          * tree is dropped. if we create snapshot, cowing the fs
4645          * tree may use more space than it frees. so we need
4646          * reserve extra space.
4647          */
4648         *bytes_to_reserve += rc->nodes_relocated;
4649 }
4650
4651 /*
4652  * called after snapshot is created. migrate block reservation
4653  * and create reloc root for the newly created snapshot
4654  */
4655 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4656                                struct btrfs_pending_snapshot *pending)
4657 {
4658         struct btrfs_root *root = pending->root;
4659         struct btrfs_root *reloc_root;
4660         struct btrfs_root *new_root;
4661         struct reloc_control *rc;
4662         int ret;
4663
4664         if (!root->reloc_root)
4665                 return 0;
4666
4667         rc = root->fs_info->reloc_ctl;
4668         rc->merging_rsv_size += rc->nodes_relocated;
4669
4670         if (rc->merge_reloc_tree) {
4671                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4672                                               rc->block_rsv,
4673                                               rc->nodes_relocated, true);
4674                 if (ret)
4675                         return ret;
4676         }
4677
4678         new_root = pending->snap;
4679         reloc_root = create_reloc_root(trans, root->reloc_root,
4680                                        new_root->root_key.objectid);
4681         if (IS_ERR(reloc_root))
4682                 return PTR_ERR(reloc_root);
4683
4684         ret = __add_reloc_root(reloc_root);
4685         BUG_ON(ret < 0);
4686         new_root->reloc_root = reloc_root;
4687
4688         if (rc->create_reloc_tree)
4689                 ret = clone_backref_node(trans, rc, root, reloc_root);
4690         return ret;
4691 }