fs/btrfs/relocation.c

   1 /*
   2  * Copyright (C) 2009 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/rbtree.h>
  24 #include <linux/slab.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "volumes.h"
  29 #include "locking.h"
  30 #include "btrfs_inode.h"
  31 #include "async-thread.h"
  32 #include "free-space-cache.h"
  33 #include "inode-map.h"
  34 #include "qgroup.h"
  35
  36 /*
  37  * backref_node, mapping_node and tree_block start with this
  38  */
  39 struct tree_entry {
  40         struct rb_node rb_node;
  41         u64 bytenr;
  42 };
  43
  44 /*
  45  * present a tree block in the backref cache
  46  */
  47 struct backref_node {
  48         struct rb_node rb_node;
  49         u64 bytenr;
  50
  51         u64 new_bytenr;
  52         /* objectid of tree block owner, can be not uptodate */
  53         u64 owner;
  54         /* link to pending, changed or detached list */
  55         struct list_head list;
  56         /* list of upper level blocks reference this block */
  57         struct list_head upper;
  58         /* list of child blocks in the cache */
  59         struct list_head lower;
  60         /* NULL if this node is not tree root */
  61         struct btrfs_root *root;
  62         /* extent buffer got by COW the block */
  63         struct extent_buffer *eb;
  64         /* level of tree block */
  65         unsigned int level:8;
  66         /* is the block in non-reference counted tree */
  67         unsigned int cowonly:1;
  68         /* 1 if no child node in the cache */
  69         unsigned int lowest:1;
  70         /* is the extent buffer locked */
  71         unsigned int locked:1;
  72         /* has the block been processed */
  73         unsigned int processed:1;
  74         /* have backrefs of this block been checked */
  75         unsigned int checked:1;
  76         /*
  77          * 1 if corresponding block has been cowed but some upper
  78          * level block pointers may not point to the new location
  79          */
  80         unsigned int pending:1;
  81         /*
  82          * 1 if the backref node isn't connected to any other
  83          * backref node.
  84          */
  85         unsigned int detached:1;
  86 };
  87
  88 /*
  89  * present a block pointer in the backref cache
  90  */
  91 struct backref_edge {
  92         struct list_head list[2];
  93         struct backref_node *node[2];
  94 };
  95
  96 #define LOWER   0
  97 #define UPPER   1
  98 #define RELOCATION_RESERVED_NODES       256
  99
 100 struct backref_cache {
 101         /* red black tree of all backref nodes in the cache */
 102         struct rb_root rb_root;
 103         /* for passing backref nodes to btrfs_reloc_cow_block */
 104         struct backref_node *path[BTRFS_MAX_LEVEL];
 105         /*
 106          * list of blocks that have been cowed but some block
 107          * pointers in upper level blocks may not reflect the
 108          * new location
 109          */
 110         struct list_head pending[BTRFS_MAX_LEVEL];
 111         /* list of backref nodes with no child node */
 112         struct list_head leaves;
 113         /* list of blocks that have been cowed in current transaction */
 114         struct list_head changed;
 115         /* list of detached backref node. */
 116         struct list_head detached;
 117
 118         u64 last_trans;
 119
 120         int nr_nodes;
 121         int nr_edges;
 122 };
 123
 124 /*
 125  * map address of tree root to tree
 126  */
 127 struct mapping_node {
 128         struct rb_node rb_node;
 129         u64 bytenr;
 130         void *data;
 131 };
 132
 133 struct mapping_tree {
 134         struct rb_root rb_root;
 135         spinlock_t lock;
 136 };
 137
 138 /*
 139  * present a tree block to process
 140  */
 141 struct tree_block {
 142         struct rb_node rb_node;
 143         u64 bytenr;
 144         struct btrfs_key key;
 145         unsigned int level:8;
 146         unsigned int key_ready:1;
 147 };
 148
 149 #define MAX_EXTENTS 128
 150
 151 struct file_extent_cluster {
 152         u64 start;
 153         u64 end;
 154         u64 boundary[MAX_EXTENTS];
 155         unsigned int nr;
 156 };
 157
 158 struct reloc_control {
 159         /* block group to relocate */
 160         struct btrfs_block_group_cache *block_group;
 161         /* extent tree */
 162         struct btrfs_root *extent_root;
 163         /* inode for moving data */
 164         struct inode *data_inode;
 165
 166         struct btrfs_block_rsv *block_rsv;
 167
 168         struct backref_cache backref_cache;
 169
 170         struct file_extent_cluster cluster;
 171         /* tree blocks have been processed */
 172         struct extent_io_tree processed_blocks;
 173         /* map start of tree root to corresponding reloc tree */
 174         struct mapping_tree reloc_root_tree;
 175         /* list of reloc trees */
 176         struct list_head reloc_roots;
 177         /* size of metadata reservation for merging reloc trees */
 178         u64 merging_rsv_size;
 179         /* size of relocated tree nodes */
 180         u64 nodes_relocated;
 181         /* reserved size for block group relocation*/
 182         u64 reserved_bytes;
 183
 184         u64 search_start;
 185         u64 extents_found;
 186
 187         unsigned int stage:8;
 188         unsigned int create_reloc_tree:1;
 189         unsigned int merge_reloc_tree:1;
 190         unsigned int found_file_extent:1;
 191 };
 192
 193 /* stages of data relocation */
 194 #define MOVE_DATA_EXTENTS       0
 195 #define UPDATE_DATA_PTRS        1
 196
 197 static void remove_backref_node(struct backref_cache *cache,
 198                                 struct backref_node *node);
 199 static void __mark_block_processed(struct reloc_control *rc,
 200                                    struct backref_node *node);
 201
 202 static void mapping_tree_init(struct mapping_tree *tree)
 203 {
 204         tree->rb_root = RB_ROOT;
 205         spin_lock_init(&tree->lock);
 206 }
 207
 208 static void backref_cache_init(struct backref_cache *cache)
 209 {
 210         int i;
 211         cache->rb_root = RB_ROOT;
 212         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 213                 INIT_LIST_HEAD(&cache->pending[i]);
 214         INIT_LIST_HEAD(&cache->changed);
 215         INIT_LIST_HEAD(&cache->detached);
 216         INIT_LIST_HEAD(&cache->leaves);
 217 }
 218
 219 static void backref_cache_cleanup(struct backref_cache *cache)
 220 {
 221         struct backref_node *node;
 222         int i;
 223
 224         while (!list_empty(&cache->detached)) {
 225                 node = list_entry(cache->detached.next,
 226                                   struct backref_node, list);
 227                 remove_backref_node(cache, node);
 228         }
 229
 230         while (!list_empty(&cache->leaves)) {
 231                 node = list_entry(cache->leaves.next,
 232                                   struct backref_node, lower);
 233                 remove_backref_node(cache, node);
 234         }
 235
 236         cache->last_trans = 0;
 237
 238         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 239                 ASSERT(list_empty(&cache->pending[i]));
 240         ASSERT(list_empty(&cache->changed));
 241         ASSERT(list_empty(&cache->detached));
 242         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
 243         ASSERT(!cache->nr_nodes);
 244         ASSERT(!cache->nr_edges);
 245 }
 246
 247 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 248 {
 249         struct backref_node *node;
 250
 251         node = kzalloc(sizeof(*node), GFP_NOFS);
 252         if (node) {
 253                 INIT_LIST_HEAD(&node->list);
 254                 INIT_LIST_HEAD(&node->upper);
 255                 INIT_LIST_HEAD(&node->lower);
 256                 RB_CLEAR_NODE(&node->rb_node);
 257                 cache->nr_nodes++;
 258         }
 259         return node;
 260 }
 261
 262 static void free_backref_node(struct backref_cache *cache,
 263                               struct backref_node *node)
 264 {
 265         if (node) {
 266                 cache->nr_nodes--;
 267                 kfree(node);
 268         }
 269 }
 270
 271 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 272 {
 273         struct backref_edge *edge;
 274
 275         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 276         if (edge)
 277                 cache->nr_edges++;
 278         return edge;
 279 }
 280
 281 static void free_backref_edge(struct backref_cache *cache,
 282                               struct backref_edge *edge)
 283 {
 284         if (edge) {
 285                 cache->nr_edges--;
 286                 kfree(edge);
 287         }
 288 }
 289
 290 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 291                                    struct rb_node *node)
 292 {
 293         struct rb_node **p = &root->rb_node;
 294         struct rb_node *parent = NULL;
 295         struct tree_entry *entry;
 296
 297         while (*p) {
 298                 parent = *p;
 299                 entry = rb_entry(parent, struct tree_entry, rb_node);
 300
 301                 if (bytenr < entry->bytenr)
 302                         p = &(*p)->rb_left;
 303                 else if (bytenr > entry->bytenr)
 304                         p = &(*p)->rb_right;
 305                 else
 306                         return parent;
 307         }
 308
 309         rb_link_node(node, parent, p);
 310         rb_insert_color(node, root);
 311         return NULL;
 312 }
 313
 314 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 315 {
 316         struct rb_node *n = root->rb_node;
 317         struct tree_entry *entry;
 318
 319         while (n) {
 320                 entry = rb_entry(n, struct tree_entry, rb_node);
 321
 322                 if (bytenr < entry->bytenr)
 323                         n = n->rb_left;
 324                 else if (bytenr > entry->bytenr)
 325                         n = n->rb_right;
 326                 else
 327                         return n;
 328         }
 329         return NULL;
 330 }
 331
 332 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 333 {
 334
 335         struct btrfs_fs_info *fs_info = NULL;
 336         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 337                                               rb_node);
 338         if (bnode->root)
 339                 fs_info = bnode->root->fs_info;
 340         btrfs_panic(fs_info, errno,
 341                     "Inconsistency in backref cache found at offset %llu",
 342                     bytenr);
 343 }
 344
 345 /*
 346  * walk up backref nodes until reach node presents tree root
 347  */
 348 static struct backref_node *walk_up_backref(struct backref_node *node,
 349                                             struct backref_edge *edges[],
 350                                             int *index)
 351 {
 352         struct backref_edge *edge;
 353         int idx = *index;
 354
 355         while (!list_empty(&node->upper)) {
 356                 edge = list_entry(node->upper.next,
 357                                   struct backref_edge, list[LOWER]);
 358                 edges[idx++] = edge;
 359                 node = edge->node[UPPER];
 360         }
 361         BUG_ON(node->detached);
 362         *index = idx;
 363         return node;
 364 }
 365
 366 /*
 367  * walk down backref nodes to find start of next reference path
 368  */
 369 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 370                                               int *index)
 371 {
 372         struct backref_edge *edge;
 373         struct backref_node *lower;
 374         int idx = *index;
 375
 376         while (idx > 0) {
 377                 edge = edges[idx - 1];
 378                 lower = edge->node[LOWER];
 379                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 380                         idx--;
 381                         continue;
 382                 }
 383                 edge = list_entry(edge->list[LOWER].next,
 384                                   struct backref_edge, list[LOWER]);
 385                 edges[idx - 1] = edge;
 386                 *index = idx;
 387                 return edge->node[UPPER];
 388         }
 389         *index = 0;
 390         return NULL;
 391 }
 392
 393 static void unlock_node_buffer(struct backref_node *node)
 394 {
 395         if (node->locked) {
 396                 btrfs_tree_unlock(node->eb);
 397                 node->locked = 0;
 398         }
 399 }
 400
 401 static void drop_node_buffer(struct backref_node *node)
 402 {
 403         if (node->eb) {
 404                 unlock_node_buffer(node);
 405                 free_extent_buffer(node->eb);
 406                 node->eb = NULL;
 407         }
 408 }
 409
 410 static void drop_backref_node(struct backref_cache *tree,
 411                               struct backref_node *node)
 412 {
 413         BUG_ON(!list_empty(&node->upper));
 414
 415         drop_node_buffer(node);
 416         list_del(&node->list);
 417         list_del(&node->lower);
 418         if (!RB_EMPTY_NODE(&node->rb_node))
 419                 rb_erase(&node->rb_node, &tree->rb_root);
 420         free_backref_node(tree, node);
 421 }
 422
 423 /*
 424  * remove a backref node from the backref cache
 425  */
 426 static void remove_backref_node(struct backref_cache *cache,
 427                                 struct backref_node *node)
 428 {
 429         struct backref_node *upper;
 430         struct backref_edge *edge;
 431
 432         if (!node)
 433                 return;
 434
 435         BUG_ON(!node->lowest && !node->detached);
 436         while (!list_empty(&node->upper)) {
 437                 edge = list_entry(node->upper.next, struct backref_edge,
 438                                   list[LOWER]);
 439                 upper = edge->node[UPPER];
 440                 list_del(&edge->list[LOWER]);
 441                 list_del(&edge->list[UPPER]);
 442                 free_backref_edge(cache, edge);
 443
 444                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 445                         BUG_ON(!list_empty(&node->upper));
 446                         drop_backref_node(cache, node);
 447                         node = upper;
 448                         node->lowest = 1;
 449                         continue;
 450                 }
 451                 /*
 452                  * add the node to leaf node list if no other
 453                  * child block cached.
 454                  */
 455                 if (list_empty(&upper->lower)) {
 456                         list_add_tail(&upper->lower, &cache->leaves);
 457                         upper->lowest = 1;
 458                 }
 459         }
 460
 461         drop_backref_node(cache, node);
 462 }
 463
 464 static void update_backref_node(struct backref_cache *cache,
 465                                 struct backref_node *node, u64 bytenr)
 466 {
 467         struct rb_node *rb_node;
 468         rb_erase(&node->rb_node, &cache->rb_root);
 469         node->bytenr = bytenr;
 470         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 471         if (rb_node)
 472                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 473 }
 474
 475 /*
 476  * update backref cache after a transaction commit
 477  */
 478 static int update_backref_cache(struct btrfs_trans_handle *trans,
 479                                 struct backref_cache *cache)
 480 {
 481         struct backref_node *node;
 482         int level = 0;
 483
 484         if (cache->last_trans == 0) {
 485                 cache->last_trans = trans->transid;
 486                 return 0;
 487         }
 488
 489         if (cache->last_trans == trans->transid)
 490                 return 0;
 491
 492         /*
 493          * detached nodes are used to avoid unnecessary backref
 494          * lookup. transaction commit changes the extent tree.
 495          * so the detached nodes are no longer useful.
 496          */
 497         while (!list_empty(&cache->detached)) {
 498                 node = list_entry(cache->detached.next,
 499                                   struct backref_node, list);
 500                 remove_backref_node(cache, node);
 501         }
 502
 503         while (!list_empty(&cache->changed)) {
 504                 node = list_entry(cache->changed.next,
 505                                   struct backref_node, list);
 506                 list_del_init(&node->list);
 507                 BUG_ON(node->pending);
 508                 update_backref_node(cache, node, node->new_bytenr);
 509         }
 510
 511         /*
 512          * some nodes can be left in the pending list if there were
 513          * errors during processing the pending nodes.
 514          */
 515         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 516                 list_for_each_entry(node, &cache->pending[level], list) {
 517                         BUG_ON(!node->pending);
 518                         if (node->bytenr == node->new_bytenr)
 519                                 continue;
 520                         update_backref_node(cache, node, node->new_bytenr);
 521                 }
 522         }
 523
 524         cache->last_trans = 0;
 525         return 1;
 526 }
 527
 528
 529 static int should_ignore_root(struct btrfs_root *root)
 530 {
 531         struct btrfs_root *reloc_root;
 532
 533         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 534                 return 0;
 535
 536         reloc_root = root->reloc_root;
 537         if (!reloc_root)
 538                 return 0;
 539
 540         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 541             root->fs_info->running_transaction->transid - 1)
 542                 return 0;
 543         /*
 544          * if there is reloc tree and it was created in previous
 545          * transaction backref lookup can find the reloc tree,
 546          * so backref node for the fs tree root is useless for
 547          * relocation.
 548          */
 549         return 1;
 550 }
 551 /*
 552  * find reloc tree by address of tree root
 553  */
 554 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 555                                           u64 bytenr)
 556 {
 557         struct rb_node *rb_node;
 558         struct mapping_node *node;
 559         struct btrfs_root *root = NULL;
 560
 561         spin_lock(&rc->reloc_root_tree.lock);
 562         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 563         if (rb_node) {
 564                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 565                 root = (struct btrfs_root *)node->data;
 566         }
 567         spin_unlock(&rc->reloc_root_tree.lock);
 568         return root;
 569 }
 570
 571 static int is_cowonly_root(u64 root_objectid)
 572 {
 573         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 574             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 575             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 576             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 577             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 578             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 579             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 580             root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
 581             root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
 582                 return 1;
 583         return 0;
 584 }
 585
 586 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 587                                         u64 root_objectid)
 588 {
 589         struct btrfs_key key;
 590
 591         key.objectid = root_objectid;
 592         key.type = BTRFS_ROOT_ITEM_KEY;
 593         if (is_cowonly_root(root_objectid))
 594                 key.offset = 0;
 595         else
 596                 key.offset = (u64)-1;
 597
 598         return btrfs_get_fs_root(fs_info, &key, false);
 599 }
 600
 601 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 602 static noinline_for_stack
 603 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 604                                   struct extent_buffer *leaf,
 605                                   struct btrfs_extent_ref_v0 *ref0)
 606 {
 607         struct btrfs_root *root;
 608         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 609         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 610
 611         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 612
 613         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 614         BUG_ON(IS_ERR(root));
 615
 616         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 617             generation != btrfs_root_generation(&root->root_item))
 618                 return NULL;
 619
 620         return root;
 621 }
 622 #endif
 623
 624 static noinline_for_stack
 625 int find_inline_backref(struct extent_buffer *leaf, int slot,
 626                         unsigned long *ptr, unsigned long *end)
 627 {
 628         struct btrfs_key key;
 629         struct btrfs_extent_item *ei;
 630         struct btrfs_tree_block_info *bi;
 631         u32 item_size;
 632
 633         btrfs_item_key_to_cpu(leaf, &key, slot);
 634
 635         item_size = btrfs_item_size_nr(leaf, slot);
 636 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 637         if (item_size < sizeof(*ei)) {
 638                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 639                 return 1;
 640         }
 641 #endif
 642         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 643         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 644                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 645
 646         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 647             item_size <= sizeof(*ei) + sizeof(*bi)) {
 648                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 649                 return 1;
 650         }
 651         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 652             item_size <= sizeof(*ei)) {
 653                 WARN_ON(item_size < sizeof(*ei));
 654                 return 1;
 655         }
 656
 657         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 658                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 659                 *ptr = (unsigned long)(bi + 1);
 660         } else {
 661                 *ptr = (unsigned long)(ei + 1);
 662         }
 663         *end = (unsigned long)ei + item_size;
 664         return 0;
 665 }
 666
 667 /*
 668  * build backref tree for a given tree block. root of the backref tree
 669  * corresponds the tree block, leaves of the backref tree correspond
 670  * roots of b-trees that reference the tree block.
 671  *
 672  * the basic idea of this function is check backrefs of a given block
 673  * to find upper level blocks that reference the block, and then check
 674  * backrefs of these upper level blocks recursively. the recursion stop
 675  * when tree root is reached or backrefs for the block is cached.
 676  *
 677  * NOTE: if we find backrefs for a block are cached, we know backrefs
 678  * for all upper level blocks that directly/indirectly reference the
 679  * block are also cached.
 680  */
 681 static noinline_for_stack
 682 struct backref_node *build_backref_tree(struct reloc_control *rc,
 683                                         struct btrfs_key *node_key,
 684                                         int level, u64 bytenr)
 685 {
 686         struct backref_cache *cache = &rc->backref_cache;
 687         struct btrfs_path *path1;
 688         struct btrfs_path *path2;
 689         struct extent_buffer *eb;
 690         struct btrfs_root *root;
 691         struct backref_node *cur;
 692         struct backref_node *upper;
 693         struct backref_node *lower;
 694         struct backref_node *node = NULL;
 695         struct backref_node *exist = NULL;
 696         struct backref_edge *edge;
 697         struct rb_node *rb_node;
 698         struct btrfs_key key;
 699         unsigned long end;
 700         unsigned long ptr;
 701         LIST_HEAD(list);
 702         LIST_HEAD(useless);
 703         int cowonly;
 704         int ret;
 705         int err = 0;
 706         bool need_check = true;
 707
 708         path1 = btrfs_alloc_path();
 709         path2 = btrfs_alloc_path();
 710         if (!path1 || !path2) {
 711                 err = -ENOMEM;
 712                 goto out;
 713         }
 714         path1->reada = READA_FORWARD;
 715         path2->reada = READA_FORWARD;
 716
 717         node = alloc_backref_node(cache);
 718         if (!node) {
 719                 err = -ENOMEM;
 720                 goto out;
 721         }
 722
 723         node->bytenr = bytenr;
 724         node->level = level;
 725         node->lowest = 1;
 726         cur = node;
 727 again:
 728         end = 0;
 729         ptr = 0;
 730         key.objectid = cur->bytenr;
 731         key.type = BTRFS_METADATA_ITEM_KEY;
 732         key.offset = (u64)-1;
 733
 734         path1->search_commit_root = 1;
 735         path1->skip_locking = 1;
 736         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 737                                 0, 0);
 738         if (ret < 0) {
 739                 err = ret;
 740                 goto out;
 741         }
 742         ASSERT(ret);
 743         ASSERT(path1->slots[0]);
 744
 745         path1->slots[0]--;
 746
 747         WARN_ON(cur->checked);
 748         if (!list_empty(&cur->upper)) {
 749                 /*
 750                  * the backref was added previously when processing
 751                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 752                  */
 753                 ASSERT(list_is_singular(&cur->upper));
 754                 edge = list_entry(cur->upper.next, struct backref_edge,
 755                                   list[LOWER]);
 756                 ASSERT(list_empty(&edge->list[UPPER]));
 757                 exist = edge->node[UPPER];
 758                 /*
 759                  * add the upper level block to pending list if we need
 760                  * check its backrefs
 761                  */
 762                 if (!exist->checked)
 763                         list_add_tail(&edge->list[UPPER], &list);
 764         } else {
 765                 exist = NULL;
 766         }
 767
 768         while (1) {
 769                 cond_resched();
 770                 eb = path1->nodes[0];
 771
 772                 if (ptr >= end) {
 773                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 774                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 775                                 if (ret < 0) {
 776                                         err = ret;
 777                                         goto out;
 778                                 }
 779                                 if (ret > 0)
 780                                         break;
 781                                 eb = path1->nodes[0];
 782                         }
 783
 784                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 785                         if (key.objectid != cur->bytenr) {
 786                                 WARN_ON(exist);
 787                                 break;
 788                         }
 789
 790                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 791                             key.type == BTRFS_METADATA_ITEM_KEY) {
 792                                 ret = find_inline_backref(eb, path1->slots[0],
 793                                                           &ptr, &end);
 794                                 if (ret)
 795                                         goto next;
 796                         }
 797                 }
 798
 799                 if (ptr < end) {
 800                         /* update key for inline back ref */
 801                         struct btrfs_extent_inline_ref *iref;
 802                         iref = (struct btrfs_extent_inline_ref *)ptr;
 803                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 804                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 805                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 806                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 807                 }
 808
 809                 if (exist &&
 810                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 811                       exist->owner == key.offset) ||
 812                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 813                       exist->bytenr == key.offset))) {
 814                         exist = NULL;
 815                         goto next;
 816                 }
 817
 818 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 819                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 820                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 821                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 822                                 struct btrfs_extent_ref_v0 *ref0;
 823                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 824                                                 struct btrfs_extent_ref_v0);
 825                                 if (key.objectid == key.offset) {
 826                                         root = find_tree_root(rc, eb, ref0);
 827                                         if (root && !should_ignore_root(root))
 828                                                 cur->root = root;
 829                                         else
 830                                                 list_add(&cur->list, &useless);
 831                                         break;
 832                                 }
 833                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 834                                                                       ref0)))
 835                                         cur->cowonly = 1;
 836                         }
 837 #else
 838                 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
 839                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 840 #endif
 841                         if (key.objectid == key.offset) {
 842                                 /*
 843                                  * only root blocks of reloc trees use
 844                                  * backref of this type.
 845                                  */
 846                                 root = find_reloc_root(rc, cur->bytenr);
 847                                 ASSERT(root);
 848                                 cur->root = root;
 849                                 break;
 850                         }
 851
 852                         edge = alloc_backref_edge(cache);
 853                         if (!edge) {
 854                                 err = -ENOMEM;
 855                                 goto out;
 856                         }
 857                         rb_node = tree_search(&cache->rb_root, key.offset);
 858                         if (!rb_node) {
 859                                 upper = alloc_backref_node(cache);
 860                                 if (!upper) {
 861                                         free_backref_edge(cache, edge);
 862                                         err = -ENOMEM;
 863                                         goto out;
 864                                 }
 865                                 upper->bytenr = key.offset;
 866                                 upper->level = cur->level + 1;
 867                                 /*
 868                                  *  backrefs for the upper level block isn't
 869                                  *  cached, add the block to pending list
 870                                  */
 871                                 list_add_tail(&edge->list[UPPER], &list);
 872                         } else {
 873                                 upper = rb_entry(rb_node, struct backref_node,
 874                                                  rb_node);
 875                                 ASSERT(upper->checked);
 876                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 877                         }
 878                         list_add_tail(&edge->list[LOWER], &cur->upper);
 879                         edge->node[LOWER] = cur;
 880                         edge->node[UPPER] = upper;
 881
 882                         goto next;
 883                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 884                         goto next;
 885                 }
 886
 887                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 888                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 889                 if (IS_ERR(root)) {
 890                         err = PTR_ERR(root);
 891                         goto out;
 892                 }
 893
 894                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 895                         cur->cowonly = 1;
 896
 897                 if (btrfs_root_level(&root->root_item) == cur->level) {
 898                         /* tree root */
 899                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
 900                                cur->bytenr);
 901                         if (should_ignore_root(root))
 902                                 list_add(&cur->list, &useless);
 903                         else
 904                                 cur->root = root;
 905                         break;
 906                 }
 907
 908                 level = cur->level + 1;
 909
 910                 /*
 911                  * searching the tree to find upper level blocks
 912                  * reference the block.
 913                  */
 914                 path2->search_commit_root = 1;
 915                 path2->skip_locking = 1;
 916                 path2->lowest_level = level;
 917                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 918                 path2->lowest_level = 0;
 919                 if (ret < 0) {
 920                         err = ret;
 921                         goto out;
 922                 }
 923                 if (ret > 0 && path2->slots[level] > 0)
 924                         path2->slots[level]--;
 925
 926                 eb = path2->nodes[level];
 927                 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
 928                     cur->bytenr) {
 929                         btrfs_err(root->fs_info,
 930         "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
 931                                   cur->bytenr, level - 1, root->objectid,
 932                                   node_key->objectid, node_key->type,
 933                                   node_key->offset);
 934                         err = -ENOENT;
 935                         goto out;
 936                 }
 937                 lower = cur;
 938                 need_check = true;
 939                 for (; level < BTRFS_MAX_LEVEL; level++) {
 940                         if (!path2->nodes[level]) {
 941                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
 942                                        lower->bytenr);
 943                                 if (should_ignore_root(root))
 944                                         list_add(&lower->list, &useless);
 945                                 else
 946                                         lower->root = root;
 947                                 break;
 948                         }
 949
 950                         edge = alloc_backref_edge(cache);
 951                         if (!edge) {
 952                                 err = -ENOMEM;
 953                                 goto out;
 954                         }
 955
 956                         eb = path2->nodes[level];
 957                         rb_node = tree_search(&cache->rb_root, eb->start);
 958                         if (!rb_node) {
 959                                 upper = alloc_backref_node(cache);
 960                                 if (!upper) {
 961                                         free_backref_edge(cache, edge);
 962                                         err = -ENOMEM;
 963                                         goto out;
 964                                 }
 965                                 upper->bytenr = eb->start;
 966                                 upper->owner = btrfs_header_owner(eb);
 967                                 upper->level = lower->level + 1;
 968                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
 969                                               &root->state))
 970                                         upper->cowonly = 1;
 971
 972                                 /*
 973                                  * if we know the block isn't shared
 974                                  * we can void checking its backrefs.
 975                                  */
 976                                 if (btrfs_block_can_be_shared(root, eb))
 977                                         upper->checked = 0;
 978                                 else
 979                                         upper->checked = 1;
 980
 981                                 /*
 982                                  * add the block to pending list if we
 983                                  * need check its backrefs, we only do this once
 984                                  * while walking up a tree as we will catch
 985                                  * anything else later on.
 986                                  */
 987                                 if (!upper->checked && need_check) {
 988                                         need_check = false;
 989                                         list_add_tail(&edge->list[UPPER],
 990                                                       &list);
 991                                 } else {
 992                                         if (upper->checked)
 993                                                 need_check = true;
 994                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 995                                 }
 996                         } else {
 997                                 upper = rb_entry(rb_node, struct backref_node,
 998                                                  rb_node);
 999                                 ASSERT(upper->checked);
1000                                 INIT_LIST_HEAD(&edge->list[UPPER]);
1001                                 if (!upper->owner)
1002                                         upper->owner = btrfs_header_owner(eb);
1003                         }
1004                         list_add_tail(&edge->list[LOWER], &lower->upper);
1005                         edge->node[LOWER] = lower;
1006                         edge->node[UPPER] = upper;
1007
1008                         if (rb_node)
1009                                 break;
1010                         lower = upper;
1011                         upper = NULL;
1012                 }
1013                 btrfs_release_path(path2);
1014 next:
1015                 if (ptr < end) {
1016                         ptr += btrfs_extent_inline_ref_size(key.type);
1017                         if (ptr >= end) {
1018                                 WARN_ON(ptr > end);
1019                                 ptr = 0;
1020                                 end = 0;
1021                         }
1022                 }
1023                 if (ptr >= end)
1024                         path1->slots[0]++;
1025         }
1026         btrfs_release_path(path1);
1027
1028         cur->checked = 1;
1029         WARN_ON(exist);
1030
1031         /* the pending list isn't empty, take the first block to process */
1032         if (!list_empty(&list)) {
1033                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1034                 list_del_init(&edge->list[UPPER]);
1035                 cur = edge->node[UPPER];
1036                 goto again;
1037         }
1038
1039         /*
1040          * everything goes well, connect backref nodes and insert backref nodes
1041          * into the cache.
1042          */
1043         ASSERT(node->checked);
1044         cowonly = node->cowonly;
1045         if (!cowonly) {
1046                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1047                                       &node->rb_node);
1048                 if (rb_node)
1049                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1050                 list_add_tail(&node->lower, &cache->leaves);
1051         }
1052
1053         list_for_each_entry(edge, &node->upper, list[LOWER])
1054                 list_add_tail(&edge->list[UPPER], &list);
1055
1056         while (!list_empty(&list)) {
1057                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1058                 list_del_init(&edge->list[UPPER]);
1059                 upper = edge->node[UPPER];
1060                 if (upper->detached) {
1061                         list_del(&edge->list[LOWER]);
1062                         lower = edge->node[LOWER];
1063                         free_backref_edge(cache, edge);
1064                         if (list_empty(&lower->upper))
1065                                 list_add(&lower->list, &useless);
1066                         continue;
1067                 }
1068
1069                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1070                         if (upper->lowest) {
1071                                 list_del_init(&upper->lower);
1072                                 upper->lowest = 0;
1073                         }
1074
1075                         list_add_tail(&edge->list[UPPER], &upper->lower);
1076                         continue;
1077                 }
1078
1079                 if (!upper->checked) {
1080                         /*
1081                          * Still want to blow up for developers since this is a
1082                          * logic bug.
1083                          */
1084                         ASSERT(0);
1085                         err = -EINVAL;
1086                         goto out;
1087                 }
1088                 if (cowonly != upper->cowonly) {
1089                         ASSERT(0);
1090                         err = -EINVAL;
1091                         goto out;
1092                 }
1093
1094                 if (!cowonly) {
1095                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1096                                               &upper->rb_node);
1097                         if (rb_node)
1098                                 backref_tree_panic(rb_node, -EEXIST,
1099                                                    upper->bytenr);
1100                 }
1101
1102                 list_add_tail(&edge->list[UPPER], &upper->lower);
1103
1104                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1105                         list_add_tail(&edge->list[UPPER], &list);
1106         }
1107         /*
1108          * process useless backref nodes. backref nodes for tree leaves
1109          * are deleted from the cache. backref nodes for upper level
1110          * tree blocks are left in the cache to avoid unnecessary backref
1111          * lookup.
1112          */
1113         while (!list_empty(&useless)) {
1114                 upper = list_entry(useless.next, struct backref_node, list);
1115                 list_del_init(&upper->list);
1116                 ASSERT(list_empty(&upper->upper));
1117                 if (upper == node)
1118                         node = NULL;
1119                 if (upper->lowest) {
1120                         list_del_init(&upper->lower);
1121                         upper->lowest = 0;
1122                 }
1123                 while (!list_empty(&upper->lower)) {
1124                         edge = list_entry(upper->lower.next,
1125                                           struct backref_edge, list[UPPER]);
1126                         list_del(&edge->list[UPPER]);
1127                         list_del(&edge->list[LOWER]);
1128                         lower = edge->node[LOWER];
1129                         free_backref_edge(cache, edge);
1130
1131                         if (list_empty(&lower->upper))
1132                                 list_add(&lower->list, &useless);
1133                 }
1134                 __mark_block_processed(rc, upper);
1135                 if (upper->level > 0) {
1136                         list_add(&upper->list, &cache->detached);
1137                         upper->detached = 1;
1138                 } else {
1139                         rb_erase(&upper->rb_node, &cache->rb_root);
1140                         free_backref_node(cache, upper);
1141                 }
1142         }
1143 out:
1144         btrfs_free_path(path1);
1145         btrfs_free_path(path2);
1146         if (err) {
1147                 while (!list_empty(&useless)) {
1148                         lower = list_entry(useless.next,
1149                                            struct backref_node, list);
1150                         list_del_init(&lower->list);
1151                 }
1152                 while (!list_empty(&list)) {
1153                         edge = list_first_entry(&list, struct backref_edge,
1154                                                 list[UPPER]);
1155                         list_del(&edge->list[UPPER]);
1156                         list_del(&edge->list[LOWER]);
1157                         lower = edge->node[LOWER];
1158                         upper = edge->node[UPPER];
1159                         free_backref_edge(cache, edge);
1160
1161                         /*
1162                          * Lower is no longer linked to any upper backref nodes
1163                          * and isn't in the cache, we can free it ourselves.
1164                          */
1165                         if (list_empty(&lower->upper) &&
1166                             RB_EMPTY_NODE(&lower->rb_node))
1167                                 list_add(&lower->list, &useless);
1168
1169                         if (!RB_EMPTY_NODE(&upper->rb_node))
1170                                 continue;
1171
1172                         /* Add this guy's upper edges to the list to process */
1173                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1174                                 list_add_tail(&edge->list[UPPER], &list);
1175                         if (list_empty(&upper->upper))
1176                                 list_add(&upper->list, &useless);
1177                 }
1178
1179                 while (!list_empty(&useless)) {
1180                         lower = list_entry(useless.next,
1181                                            struct backref_node, list);
1182                         list_del_init(&lower->list);
1183                         if (lower == node)
1184                                 node = NULL;
1185                         free_backref_node(cache, lower);
1186                 }
1187
1188                 free_backref_node(cache, node);
1189                 return ERR_PTR(err);
1190         }
1191         ASSERT(!node || !node->detached);
1192         return node;
1193 }
1194
1195 /*
1196  * helper to add backref node for the newly created snapshot.
1197  * the backref node is created by cloning backref node that
1198  * corresponds to root of source tree
1199  */
1200 static int clone_backref_node(struct btrfs_trans_handle *trans,
1201                               struct reloc_control *rc,
1202                               struct btrfs_root *src,
1203                               struct btrfs_root *dest)
1204 {
1205         struct btrfs_root *reloc_root = src->reloc_root;
1206         struct backref_cache *cache = &rc->backref_cache;
1207         struct backref_node *node = NULL;
1208         struct backref_node *new_node;
1209         struct backref_edge *edge;
1210         struct backref_edge *new_edge;
1211         struct rb_node *rb_node;
1212
1213         if (cache->last_trans > 0)
1214                 update_backref_cache(trans, cache);
1215
1216         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1217         if (rb_node) {
1218                 node = rb_entry(rb_node, struct backref_node, rb_node);
1219                 if (node->detached)
1220                         node = NULL;
1221                 else
1222                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1223         }
1224
1225         if (!node) {
1226                 rb_node = tree_search(&cache->rb_root,
1227                                       reloc_root->commit_root->start);
1228                 if (rb_node) {
1229                         node = rb_entry(rb_node, struct backref_node,
1230                                         rb_node);
1231                         BUG_ON(node->detached);
1232                 }
1233         }
1234
1235         if (!node)
1236                 return 0;
1237
1238         new_node = alloc_backref_node(cache);
1239         if (!new_node)
1240                 return -ENOMEM;
1241
1242         new_node->bytenr = dest->node->start;
1243         new_node->level = node->level;
1244         new_node->lowest = node->lowest;
1245         new_node->checked = 1;
1246         new_node->root = dest;
1247
1248         if (!node->lowest) {
1249                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1250                         new_edge = alloc_backref_edge(cache);
1251                         if (!new_edge)
1252                                 goto fail;
1253
1254                         new_edge->node[UPPER] = new_node;
1255                         new_edge->node[LOWER] = edge->node[LOWER];
1256                         list_add_tail(&new_edge->list[UPPER],
1257                                       &new_node->lower);
1258                 }
1259         } else {
1260                 list_add_tail(&new_node->lower, &cache->leaves);
1261         }
1262
1263         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1264                               &new_node->rb_node);
1265         if (rb_node)
1266                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1267
1268         if (!new_node->lowest) {
1269                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1270                         list_add_tail(&new_edge->list[LOWER],
1271                                       &new_edge->node[LOWER]->upper);
1272                 }
1273         }
1274         return 0;
1275 fail:
1276         while (!list_empty(&new_node->lower)) {
1277                 new_edge = list_entry(new_node->lower.next,
1278                                       struct backref_edge, list[UPPER]);
1279                 list_del(&new_edge->list[UPPER]);
1280                 free_backref_edge(cache, new_edge);
1281         }
1282         free_backref_node(cache, new_node);
1283         return -ENOMEM;
1284 }
1285
1286 /*
1287  * helper to add 'address of tree root -> reloc tree' mapping
1288  */
1289 static int __must_check __add_reloc_root(struct btrfs_root *root)
1290 {
1291         struct rb_node *rb_node;
1292         struct mapping_node *node;
1293         struct reloc_control *rc = root->fs_info->reloc_ctl;
1294
1295         node = kmalloc(sizeof(*node), GFP_NOFS);
1296         if (!node)
1297                 return -ENOMEM;
1298
1299         node->bytenr = root->node->start;
1300         node->data = root;
1301
1302         spin_lock(&rc->reloc_root_tree.lock);
1303         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1304                               node->bytenr, &node->rb_node);
1305         spin_unlock(&rc->reloc_root_tree.lock);
1306         if (rb_node) {
1307                 btrfs_panic(root->fs_info, -EEXIST,
1308                             "Duplicate root found for start=%llu while inserting into relocation tree",
1309                             node->bytenr);
1310                 kfree(node);
1311                 return -EEXIST;
1312         }
1313
1314         list_add_tail(&root->root_list, &rc->reloc_roots);
1315         return 0;
1316 }
1317
1318 /*
1319  * helper to delete the 'address of tree root -> reloc tree'
1320  * mapping
1321  */
1322 static void __del_reloc_root(struct btrfs_root *root)
1323 {
1324         struct rb_node *rb_node;
1325         struct mapping_node *node = NULL;
1326         struct reloc_control *rc = root->fs_info->reloc_ctl;
1327
1328         spin_lock(&rc->reloc_root_tree.lock);
1329         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1330                               root->node->start);
1331         if (rb_node) {
1332                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1333                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1334         }
1335         spin_unlock(&rc->reloc_root_tree.lock);
1336
1337         if (!node)
1338                 return;
1339         BUG_ON((struct btrfs_root *)node->data != root);
1340
1341         spin_lock(&root->fs_info->trans_lock);
1342         list_del_init(&root->root_list);
1343         spin_unlock(&root->fs_info->trans_lock);
1344         kfree(node);
1345 }
1346
1347 /*
1348  * helper to update the 'address of tree root -> reloc tree'
1349  * mapping
1350  */
1351 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1352 {
1353         struct rb_node *rb_node;
1354         struct mapping_node *node = NULL;
1355         struct reloc_control *rc = root->fs_info->reloc_ctl;
1356
1357         spin_lock(&rc->reloc_root_tree.lock);
1358         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1359                               root->node->start);
1360         if (rb_node) {
1361                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1362                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1363         }
1364         spin_unlock(&rc->reloc_root_tree.lock);
1365
1366         if (!node)
1367                 return 0;
1368         BUG_ON((struct btrfs_root *)node->data != root);
1369
1370         spin_lock(&rc->reloc_root_tree.lock);
1371         node->bytenr = new_bytenr;
1372         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1373                               node->bytenr, &node->rb_node);
1374         spin_unlock(&rc->reloc_root_tree.lock);
1375         if (rb_node)
1376                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1377         return 0;
1378 }
1379
1380 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1381                                         struct btrfs_root *root, u64 objectid)
1382 {
1383         struct btrfs_root *reloc_root;
1384         struct extent_buffer *eb;
1385         struct btrfs_root_item *root_item;
1386         struct btrfs_key root_key;
1387         int ret;
1388
1389         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1390         BUG_ON(!root_item);
1391
1392         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1393         root_key.type = BTRFS_ROOT_ITEM_KEY;
1394         root_key.offset = objectid;
1395
1396         if (root->root_key.objectid == objectid) {
1397                 u64 commit_root_gen;
1398
1399                 /* called by btrfs_init_reloc_root */
1400                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1401                                       BTRFS_TREE_RELOC_OBJECTID);
1402                 BUG_ON(ret);
1403                 /*
1404                  * Set the last_snapshot field to the generation of the commit
1405                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1406                  * correctly (returns true) when the relocation root is created
1407                  * either inside the critical section of a transaction commit
1408                  * (through transaction.c:qgroup_account_snapshot()) and when
1409                  * it's created before the transaction commit is started.
1410                  */
1411                 commit_root_gen = btrfs_header_generation(root->commit_root);
1412                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1413         } else {
1414                 /*
1415                  * called by btrfs_reloc_post_snapshot_hook.
1416                  * the source tree is a reloc tree, all tree blocks
1417                  * modified after it was created have RELOC flag
1418                  * set in their headers. so it's OK to not update
1419                  * the 'last_snapshot'.
1420                  */
1421                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1422                                       BTRFS_TREE_RELOC_OBJECTID);
1423                 BUG_ON(ret);
1424         }
1425
1426         memcpy(root_item, &root->root_item, sizeof(*root_item));
1427         btrfs_set_root_bytenr(root_item, eb->start);
1428         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1429         btrfs_set_root_generation(root_item, trans->transid);
1430
1431         if (root->root_key.objectid == objectid) {
1432                 btrfs_set_root_refs(root_item, 0);
1433                 memset(&root_item->drop_progress, 0,
1434                        sizeof(struct btrfs_disk_key));
1435                 root_item->drop_level = 0;
1436         }
1437
1438         btrfs_tree_unlock(eb);
1439         free_extent_buffer(eb);
1440
1441         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1442                                 &root_key, root_item);
1443         BUG_ON(ret);
1444         kfree(root_item);
1445
1446         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1447         BUG_ON(IS_ERR(reloc_root));
1448         reloc_root->last_trans = trans->transid;
1449         return reloc_root;
1450 }
1451
1452 /*
1453  * create reloc tree for a given fs tree. reloc tree is just a
1454  * snapshot of the fs tree with special root objectid.
1455  */
1456 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1457                           struct btrfs_root *root)
1458 {
1459         struct btrfs_root *reloc_root;
1460         struct reloc_control *rc = root->fs_info->reloc_ctl;
1461         struct btrfs_block_rsv *rsv;
1462         int clear_rsv = 0;
1463         int ret;
1464
1465         if (root->reloc_root) {
1466                 reloc_root = root->reloc_root;
1467                 reloc_root->last_trans = trans->transid;
1468                 return 0;
1469         }
1470
1471         if (!rc || !rc->create_reloc_tree ||
1472             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1473                 return 0;
1474
1475         if (!trans->reloc_reserved) {
1476                 rsv = trans->block_rsv;
1477                 trans->block_rsv = rc->block_rsv;
1478                 clear_rsv = 1;
1479         }
1480         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1481         if (clear_rsv)
1482                 trans->block_rsv = rsv;
1483
1484         ret = __add_reloc_root(reloc_root);
1485         BUG_ON(ret < 0);
1486         root->reloc_root = reloc_root;
1487         return 0;
1488 }
1489
1490 /*
1491  * update root item of reloc tree
1492  */
1493 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1494                             struct btrfs_root *root)
1495 {
1496         struct btrfs_root *reloc_root;
1497         struct btrfs_root_item *root_item;
1498         int ret;
1499
1500         if (!root->reloc_root)
1501                 goto out;
1502
1503         reloc_root = root->reloc_root;
1504         root_item = &reloc_root->root_item;
1505
1506         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1507             btrfs_root_refs(root_item) == 0) {
1508                 root->reloc_root = NULL;
1509                 __del_reloc_root(reloc_root);
1510         }
1511
1512         if (reloc_root->commit_root != reloc_root->node) {
1513                 btrfs_set_root_node(root_item, reloc_root->node);
1514                 free_extent_buffer(reloc_root->commit_root);
1515                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1516         }
1517
1518         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1519                                 &reloc_root->root_key, root_item);
1520         BUG_ON(ret);
1521
1522 out:
1523         return 0;
1524 }
1525
1526 /*
1527  * helper to find first cached inode with inode number >= objectid
1528  * in a subvolume
1529  */
1530 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1531 {
1532         struct rb_node *node;
1533         struct rb_node *prev;
1534         struct btrfs_inode *entry;
1535         struct inode *inode;
1536
1537         spin_lock(&root->inode_lock);
1538 again:
1539         node = root->inode_tree.rb_node;
1540         prev = NULL;
1541         while (node) {
1542                 prev = node;
1543                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1544
1545                 if (objectid < btrfs_ino(&entry->vfs_inode))
1546                         node = node->rb_left;
1547                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1548                         node = node->rb_right;
1549                 else
1550                         break;
1551         }
1552         if (!node) {
1553                 while (prev) {
1554                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1555                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1556                                 node = prev;
1557                                 break;
1558                         }
1559                         prev = rb_next(prev);
1560                 }
1561         }
1562         while (node) {
1563                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1564                 inode = igrab(&entry->vfs_inode);
1565                 if (inode) {
1566                         spin_unlock(&root->inode_lock);
1567                         return inode;
1568                 }
1569
1570                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1571                 if (cond_resched_lock(&root->inode_lock))
1572                         goto again;
1573
1574                 node = rb_next(node);
1575         }
1576         spin_unlock(&root->inode_lock);
1577         return NULL;
1578 }
1579
1580 static int in_block_group(u64 bytenr,
1581                           struct btrfs_block_group_cache *block_group)
1582 {
1583         if (bytenr >= block_group->key.objectid &&
1584             bytenr < block_group->key.objectid + block_group->key.offset)
1585                 return 1;
1586         return 0;
1587 }
1588
1589 /*
1590  * get new location of data
1591  */
1592 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1593                             u64 bytenr, u64 num_bytes)
1594 {
1595         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1596         struct btrfs_path *path;
1597         struct btrfs_file_extent_item *fi;
1598         struct extent_buffer *leaf;
1599         int ret;
1600
1601         path = btrfs_alloc_path();
1602         if (!path)
1603                 return -ENOMEM;
1604
1605         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1606         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1607                                        bytenr, 0);
1608         if (ret < 0)
1609                 goto out;
1610         if (ret > 0) {
1611                 ret = -ENOENT;
1612                 goto out;
1613         }
1614
1615         leaf = path->nodes[0];
1616         fi = btrfs_item_ptr(leaf, path->slots[0],
1617                             struct btrfs_file_extent_item);
1618
1619         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1620                btrfs_file_extent_compression(leaf, fi) ||
1621                btrfs_file_extent_encryption(leaf, fi) ||
1622                btrfs_file_extent_other_encoding(leaf, fi));
1623
1624         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1625                 ret = -EINVAL;
1626                 goto out;
1627         }
1628
1629         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1630         ret = 0;
1631 out:
1632         btrfs_free_path(path);
1633         return ret;
1634 }
1635
1636 /*
1637  * update file extent items in the tree leaf to point to
1638  * the new locations.
1639  */
1640 static noinline_for_stack
1641 int replace_file_extents(struct btrfs_trans_handle *trans,
1642                          struct reloc_control *rc,
1643                          struct btrfs_root *root,
1644                          struct extent_buffer *leaf)
1645 {
1646         struct btrfs_key key;
1647         struct btrfs_file_extent_item *fi;
1648         struct inode *inode = NULL;
1649         u64 parent;
1650         u64 bytenr;
1651         u64 new_bytenr = 0;
1652         u64 num_bytes;
1653         u64 end;
1654         u32 nritems;
1655         u32 i;
1656         int ret = 0;
1657         int first = 1;
1658         int dirty = 0;
1659
1660         if (rc->stage != UPDATE_DATA_PTRS)
1661                 return 0;
1662
1663         /* reloc trees always use full backref */
1664         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1665                 parent = leaf->start;
1666         else
1667                 parent = 0;
1668
1669         nritems = btrfs_header_nritems(leaf);
1670         for (i = 0; i < nritems; i++) {
1671                 cond_resched();
1672                 btrfs_item_key_to_cpu(leaf, &key, i);
1673                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1674                         continue;
1675                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1676                 if (btrfs_file_extent_type(leaf, fi) ==
1677                     BTRFS_FILE_EXTENT_INLINE)
1678                         continue;
1679                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1680                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1681                 if (bytenr == 0)
1682                         continue;
1683                 if (!in_block_group(bytenr, rc->block_group))
1684                         continue;
1685
1686                 /*
1687                  * if we are modifying block in fs tree, wait for readpage
1688                  * to complete and drop the extent cache
1689                  */
1690                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1691                         if (first) {
1692                                 inode = find_next_inode(root, key.objectid);
1693                                 first = 0;
1694                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1695                                 btrfs_add_delayed_iput(inode);
1696                                 inode = find_next_inode(root, key.objectid);
1697                         }
1698                         if (inode && btrfs_ino(inode) == key.objectid) {
1699                                 end = key.offset +
1700                                       btrfs_file_extent_num_bytes(leaf, fi);
1701                                 WARN_ON(!IS_ALIGNED(key.offset,
1702                                                     root->sectorsize));
1703                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1704                                 end--;
1705                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1706                                                       key.offset, end);
1707                                 if (!ret)
1708                                         continue;
1709
1710                                 btrfs_drop_extent_cache(inode, key.offset, end,
1711                                                         1);
1712                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1713                                               key.offset, end);
1714                         }
1715                 }
1716
1717                 ret = get_new_location(rc->data_inode, &new_bytenr,
1718                                        bytenr, num_bytes);
1719                 if (ret) {
1720                         /*
1721                          * Don't have to abort since we've not changed anything
1722                          * in the file extent yet.
1723                          */
1724                         break;
1725                 }
1726
1727                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1728                 dirty = 1;
1729
1730                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1731                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1732                                            num_bytes, parent,
1733                                            btrfs_header_owner(leaf),
1734                                            key.objectid, key.offset);
1735                 if (ret) {
1736                         btrfs_abort_transaction(trans, ret);
1737                         break;
1738                 }
1739
1740                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1741                                         parent, btrfs_header_owner(leaf),
1742                                         key.objectid, key.offset);
1743                 if (ret) {
1744                         btrfs_abort_transaction(trans, ret);
1745                         break;
1746                 }
1747         }
1748         if (dirty)
1749                 btrfs_mark_buffer_dirty(leaf);
1750         if (inode)
1751                 btrfs_add_delayed_iput(inode);
1752         return ret;
1753 }
1754
1755 static noinline_for_stack
1756 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1757                      struct btrfs_path *path, int level)
1758 {
1759         struct btrfs_disk_key key1;
1760         struct btrfs_disk_key key2;
1761         btrfs_node_key(eb, &key1, slot);
1762         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1763         return memcmp(&key1, &key2, sizeof(key1));
1764 }
1765
1766 /*
1767  * try to replace tree blocks in fs tree with the new blocks
1768  * in reloc tree. tree blocks haven't been modified since the
1769  * reloc tree was create can be replaced.
1770  *
1771  * if a block was replaced, level of the block + 1 is returned.
1772  * if no block got replaced, 0 is returned. if there are other
1773  * errors, a negative error number is returned.
1774  */
1775 static noinline_for_stack
1776 int replace_path(struct btrfs_trans_handle *trans,
1777                  struct btrfs_root *dest, struct btrfs_root *src,
1778                  struct btrfs_path *path, struct btrfs_key *next_key,
1779                  int lowest_level, int max_level)
1780 {
1781         struct extent_buffer *eb;
1782         struct extent_buffer *parent;
1783         struct btrfs_key key;
1784         u64 old_bytenr;
1785         u64 new_bytenr;
1786         u64 old_ptr_gen;
1787         u64 new_ptr_gen;
1788         u64 last_snapshot;
1789         u32 blocksize;
1790         int cow = 0;
1791         int level;
1792         int ret;
1793         int slot;
1794
1795         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1796         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1797
1798         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1799 again:
1800         slot = path->slots[lowest_level];
1801         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1802
1803         eb = btrfs_lock_root_node(dest);
1804         btrfs_set_lock_blocking(eb);
1805         level = btrfs_header_level(eb);
1806
1807         if (level < lowest_level) {
1808                 btrfs_tree_unlock(eb);
1809                 free_extent_buffer(eb);
1810                 return 0;
1811         }
1812
1813         if (cow) {
1814                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1815                 BUG_ON(ret);
1816         }
1817         btrfs_set_lock_blocking(eb);
1818
1819         if (next_key) {
1820                 next_key->objectid = (u64)-1;
1821                 next_key->type = (u8)-1;
1822                 next_key->offset = (u64)-1;
1823         }
1824
1825         parent = eb;
1826         while (1) {
1827                 level = btrfs_header_level(parent);
1828                 BUG_ON(level < lowest_level);
1829
1830                 ret = btrfs_bin_search(parent, &key, level, &slot);
1831                 if (ret && slot > 0)
1832                         slot--;
1833
1834                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1835                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1836
1837                 old_bytenr = btrfs_node_blockptr(parent, slot);
1838                 blocksize = dest->nodesize;
1839                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1840
1841                 if (level <= max_level) {
1842                         eb = path->nodes[level];
1843                         new_bytenr = btrfs_node_blockptr(eb,
1844                                                         path->slots[level]);
1845                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1846                                                         path->slots[level]);
1847                 } else {
1848                         new_bytenr = 0;
1849                         new_ptr_gen = 0;
1850                 }
1851
1852                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1853                         ret = level;
1854                         break;
1855                 }
1856
1857                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1858                     memcmp_node_keys(parent, slot, path, level)) {
1859                         if (level <= lowest_level) {
1860                                 ret = 0;
1861                                 break;
1862                         }
1863
1864                         eb = read_tree_block(dest, old_bytenr, old_ptr_gen);
1865                         if (IS_ERR(eb)) {
1866                                 ret = PTR_ERR(eb);
1867                                 break;
1868                         } else if (!extent_buffer_uptodate(eb)) {
1869                                 ret = -EIO;
1870                                 free_extent_buffer(eb);
1871                                 break;
1872                         }
1873                         btrfs_tree_lock(eb);
1874                         if (cow) {
1875                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1876                                                       slot, &eb);
1877                                 BUG_ON(ret);
1878                         }
1879                         btrfs_set_lock_blocking(eb);
1880
1881                         btrfs_tree_unlock(parent);
1882                         free_extent_buffer(parent);
1883
1884                         parent = eb;
1885                         continue;
1886                 }
1887
1888                 if (!cow) {
1889                         btrfs_tree_unlock(parent);
1890                         free_extent_buffer(parent);
1891                         cow = 1;
1892                         goto again;
1893                 }
1894
1895                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1896                                       path->slots[level]);
1897                 btrfs_release_path(path);
1898
1899                 path->lowest_level = level;
1900                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1901                 path->lowest_level = 0;
1902                 BUG_ON(ret);
1903
1904                 /*
1905                  * swap blocks in fs tree and reloc tree.
1906                  */
1907                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1908                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1909                 btrfs_mark_buffer_dirty(parent);
1910
1911                 btrfs_set_node_blockptr(path->nodes[level],
1912                                         path->slots[level], old_bytenr);
1913                 btrfs_set_node_ptr_generation(path->nodes[level],
1914                                               path->slots[level], old_ptr_gen);
1915                 btrfs_mark_buffer_dirty(path->nodes[level]);
1916
1917                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1918                                         path->nodes[level]->start,
1919                                         src->root_key.objectid, level - 1, 0);
1920                 BUG_ON(ret);
1921                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1922                                         0, dest->root_key.objectid, level - 1,
1923                                         0);
1924                 BUG_ON(ret);
1925
1926                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1927                                         path->nodes[level]->start,
1928                                         src->root_key.objectid, level - 1, 0);
1929                 BUG_ON(ret);
1930
1931                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1932                                         0, dest->root_key.objectid, level - 1,
1933                                         0);
1934                 BUG_ON(ret);
1935
1936                 btrfs_unlock_up_safe(path, 0);
1937
1938                 ret = level;
1939                 break;
1940         }
1941         btrfs_tree_unlock(parent);
1942         free_extent_buffer(parent);
1943         return ret;
1944 }
1945
1946 /*
1947  * helper to find next relocated block in reloc tree
1948  */
1949 static noinline_for_stack
1950 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1951                        int *level)
1952 {
1953         struct extent_buffer *eb;
1954         int i;
1955         u64 last_snapshot;
1956         u32 nritems;
1957
1958         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1959
1960         for (i = 0; i < *level; i++) {
1961                 free_extent_buffer(path->nodes[i]);
1962                 path->nodes[i] = NULL;
1963         }
1964
1965         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1966                 eb = path->nodes[i];
1967                 nritems = btrfs_header_nritems(eb);
1968                 while (path->slots[i] + 1 < nritems) {
1969                         path->slots[i]++;
1970                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1971                             last_snapshot)
1972                                 continue;
1973
1974                         *level = i;
1975                         return 0;
1976                 }
1977                 free_extent_buffer(path->nodes[i]);
1978                 path->nodes[i] = NULL;
1979         }
1980         return 1;
1981 }
1982
1983 /*
1984  * walk down reloc tree to find relocated block of lowest level
1985  */
1986 static noinline_for_stack
1987 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1988                          int *level)
1989 {
1990         struct extent_buffer *eb = NULL;
1991         int i;
1992         u64 bytenr;
1993         u64 ptr_gen = 0;
1994         u64 last_snapshot;
1995         u32 nritems;
1996
1997         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1998
1999         for (i = *level; i > 0; i--) {
2000                 eb = path->nodes[i];
2001                 nritems = btrfs_header_nritems(eb);
2002                 while (path->slots[i] < nritems) {
2003                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2004                         if (ptr_gen > last_snapshot)
2005                                 break;
2006                         path->slots[i]++;
2007                 }
2008                 if (path->slots[i] >= nritems) {
2009                         if (i == *level)
2010                                 break;
2011                         *level = i + 1;
2012                         return 0;
2013                 }
2014                 if (i == 1) {
2015                         *level = i;
2016                         return 0;
2017                 }
2018
2019                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2020                 eb = read_tree_block(root, bytenr, ptr_gen);
2021                 if (IS_ERR(eb)) {
2022                         return PTR_ERR(eb);
2023                 } else if (!extent_buffer_uptodate(eb)) {
2024                         free_extent_buffer(eb);
2025                         return -EIO;
2026                 }
2027                 BUG_ON(btrfs_header_level(eb) != i - 1);
2028                 path->nodes[i - 1] = eb;
2029                 path->slots[i - 1] = 0;
2030         }
2031         return 1;
2032 }
2033
2034 /*
2035  * invalidate extent cache for file extents whose key in range of
2036  * [min_key, max_key)
2037  */
2038 static int invalidate_extent_cache(struct btrfs_root *root,
2039                                    struct btrfs_key *min_key,
2040                                    struct btrfs_key *max_key)
2041 {
2042         struct inode *inode = NULL;
2043         u64 objectid;
2044         u64 start, end;
2045         u64 ino;
2046
2047         objectid = min_key->objectid;
2048         while (1) {
2049                 cond_resched();
2050                 iput(inode);
2051
2052                 if (objectid > max_key->objectid)
2053                         break;
2054
2055                 inode = find_next_inode(root, objectid);
2056                 if (!inode)
2057                         break;
2058                 ino = btrfs_ino(inode);
2059
2060                 if (ino > max_key->objectid) {
2061                         iput(inode);
2062                         break;
2063                 }
2064
2065                 objectid = ino + 1;
2066                 if (!S_ISREG(inode->i_mode))
2067                         continue;
2068
2069                 if (unlikely(min_key->objectid == ino)) {
2070                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2071                                 continue;
2072                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2073                                 start = 0;
2074                         else {
2075                                 start = min_key->offset;
2076                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2077                         }
2078                 } else {
2079                         start = 0;
2080                 }
2081
2082                 if (unlikely(max_key->objectid == ino)) {
2083                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2084                                 continue;
2085                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2086                                 end = (u64)-1;
2087                         } else {
2088                                 if (max_key->offset == 0)
2089                                         continue;
2090                                 end = max_key->offset;
2091                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2092                                 end--;
2093                         }
2094                 } else {
2095                         end = (u64)-1;
2096                 }
2097
2098                 /* the lock_extent waits for readpage to complete */
2099                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2100                 btrfs_drop_extent_cache(inode, start, end, 1);
2101                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2102         }
2103         return 0;
2104 }
2105
2106 static int find_next_key(struct btrfs_path *path, int level,
2107                          struct btrfs_key *key)
2108
2109 {
2110         while (level < BTRFS_MAX_LEVEL) {
2111                 if (!path->nodes[level])
2112                         break;
2113                 if (path->slots[level] + 1 <
2114                     btrfs_header_nritems(path->nodes[level])) {
2115                         btrfs_node_key_to_cpu(path->nodes[level], key,
2116                                               path->slots[level] + 1);
2117                         return 0;
2118                 }
2119                 level++;
2120         }
2121         return 1;
2122 }
2123
2124 /*
2125  * merge the relocated tree blocks in reloc tree with corresponding
2126  * fs tree.
2127  */
2128 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2129                                                struct btrfs_root *root)
2130 {
2131         LIST_HEAD(inode_list);
2132         struct btrfs_key key;
2133         struct btrfs_key next_key;
2134         struct btrfs_trans_handle *trans = NULL;
2135         struct btrfs_root *reloc_root;
2136         struct btrfs_root_item *root_item;
2137         struct btrfs_path *path;
2138         struct extent_buffer *leaf;
2139         int level;
2140         int max_level;
2141         int replaced = 0;
2142         int ret;
2143         int err = 0;
2144         u32 min_reserved;
2145
2146         path = btrfs_alloc_path();
2147         if (!path)
2148                 return -ENOMEM;
2149         path->reada = READA_FORWARD;
2150
2151         reloc_root = root->reloc_root;
2152         root_item = &reloc_root->root_item;
2153
2154         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2155                 level = btrfs_root_level(root_item);
2156                 extent_buffer_get(reloc_root->node);
2157                 path->nodes[level] = reloc_root->node;
2158                 path->slots[level] = 0;
2159         } else {
2160                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2161
2162                 level = root_item->drop_level;
2163                 BUG_ON(level == 0);
2164                 path->lowest_level = level;
2165                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2166                 path->lowest_level = 0;
2167                 if (ret < 0) {
2168                         btrfs_free_path(path);
2169                         return ret;
2170                 }
2171
2172                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2173                                       path->slots[level]);
2174                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2175
2176                 btrfs_unlock_up_safe(path, 0);
2177         }
2178
2179         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2180         memset(&next_key, 0, sizeof(next_key));
2181
2182         while (1) {
2183                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2184                                              BTRFS_RESERVE_FLUSH_ALL);
2185                 if (ret) {
2186                         err = ret;
2187                         goto out;
2188                 }
2189                 trans = btrfs_start_transaction(root, 0);
2190                 if (IS_ERR(trans)) {
2191                         err = PTR_ERR(trans);
2192                         trans = NULL;
2193                         goto out;
2194                 }
2195                 trans->block_rsv = rc->block_rsv;
2196
2197                 replaced = 0;
2198                 max_level = level;
2199
2200                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2201                 if (ret < 0) {
2202                         err = ret;
2203                         goto out;
2204                 }
2205                 if (ret > 0)
2206                         break;
2207
2208                 if (!find_next_key(path, level, &key) &&
2209                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2210                         ret = 0;
2211                 } else {
2212                         ret = replace_path(trans, root, reloc_root, path,
2213                                            &next_key, level, max_level);
2214                 }
2215                 if (ret < 0) {
2216                         err = ret;
2217                         goto out;
2218                 }
2219
2220                 if (ret > 0) {
2221                         level = ret;
2222                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2223                                               path->slots[level]);
2224                         replaced = 1;
2225                 }
2226
2227                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2228                 if (ret > 0)
2229                         break;
2230
2231                 BUG_ON(level == 0);
2232                 /*
2233                  * save the merging progress in the drop_progress.
2234                  * this is OK since root refs == 1 in this case.
2235                  */
2236                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2237                                path->slots[level]);
2238                 root_item->drop_level = level;
2239
2240                 btrfs_end_transaction_throttle(trans, root);
2241                 trans = NULL;
2242
2243                 btrfs_btree_balance_dirty(root);
2244
2245                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2246                         invalidate_extent_cache(root, &key, &next_key);
2247         }
2248
2249         /*
2250          * handle the case only one block in the fs tree need to be
2251          * relocated and the block is tree root.
2252          */
2253         leaf = btrfs_lock_root_node(root);
2254         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2255         btrfs_tree_unlock(leaf);
2256         free_extent_buffer(leaf);
2257         if (ret < 0)
2258                 err = ret;
2259 out:
2260         btrfs_free_path(path);
2261
2262         if (err == 0) {
2263                 memset(&root_item->drop_progress, 0,
2264                        sizeof(root_item->drop_progress));
2265                 root_item->drop_level = 0;
2266                 btrfs_set_root_refs(root_item, 0);
2267                 btrfs_update_reloc_root(trans, root);
2268         }
2269
2270         if (trans)
2271                 btrfs_end_transaction_throttle(trans, root);
2272
2273         btrfs_btree_balance_dirty(root);
2274
2275         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2276                 invalidate_extent_cache(root, &key, &next_key);
2277
2278         return err;
2279 }
2280
2281 static noinline_for_stack
2282 int prepare_to_merge(struct reloc_control *rc, int err)
2283 {
2284         struct btrfs_root *root = rc->extent_root;
2285         struct btrfs_root *reloc_root;
2286         struct btrfs_trans_handle *trans;
2287         LIST_HEAD(reloc_roots);
2288         u64 num_bytes = 0;
2289         int ret;
2290
2291         mutex_lock(&root->fs_info->reloc_mutex);
2292         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2293         rc->merging_rsv_size += rc->nodes_relocated * 2;
2294         mutex_unlock(&root->fs_info->reloc_mutex);
2295
2296 again:
2297         if (!err) {
2298                 num_bytes = rc->merging_rsv_size;
2299                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2300                                           BTRFS_RESERVE_FLUSH_ALL);
2301                 if (ret)
2302                         err = ret;
2303         }
2304
2305         trans = btrfs_join_transaction(rc->extent_root);
2306         if (IS_ERR(trans)) {
2307                 if (!err)
2308                         btrfs_block_rsv_release(rc->extent_root,
2309                                                 rc->block_rsv, num_bytes);
2310                 return PTR_ERR(trans);
2311         }
2312
2313         if (!err) {
2314                 if (num_bytes != rc->merging_rsv_size) {
2315                         btrfs_end_transaction(trans, rc->extent_root);
2316                         btrfs_block_rsv_release(rc->extent_root,
2317                                                 rc->block_rsv, num_bytes);
2318                         goto again;
2319                 }
2320         }
2321
2322         rc->merge_reloc_tree = 1;
2323
2324         while (!list_empty(&rc->reloc_roots)) {
2325                 reloc_root = list_entry(rc->reloc_roots.next,
2326                                         struct btrfs_root, root_list);
2327                 list_del_init(&reloc_root->root_list);
2328
2329                 root = read_fs_root(reloc_root->fs_info,
2330                                     reloc_root->root_key.offset);
2331                 BUG_ON(IS_ERR(root));
2332                 BUG_ON(root->reloc_root != reloc_root);
2333
2334                 /*
2335                  * set reference count to 1, so btrfs_recover_relocation
2336                  * knows it should resumes merging
2337                  */
2338                 if (!err)
2339                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2340                 btrfs_update_reloc_root(trans, root);
2341
2342                 list_add(&reloc_root->root_list, &reloc_roots);
2343         }
2344
2345         list_splice(&reloc_roots, &rc->reloc_roots);
2346
2347         if (!err)
2348                 btrfs_commit_transaction(trans, rc->extent_root);
2349         else
2350                 btrfs_end_transaction(trans, rc->extent_root);
2351         return err;
2352 }
2353
2354 static noinline_for_stack
2355 void free_reloc_roots(struct list_head *list)
2356 {
2357         struct btrfs_root *reloc_root;
2358
2359         while (!list_empty(list)) {
2360                 reloc_root = list_entry(list->next, struct btrfs_root,
2361                                         root_list);
2362                 free_extent_buffer(reloc_root->node);
2363                 free_extent_buffer(reloc_root->commit_root);
2364                 reloc_root->node = NULL;
2365                 reloc_root->commit_root = NULL;
2366                 __del_reloc_root(reloc_root);
2367         }
2368 }
2369
2370 static noinline_for_stack
2371 void merge_reloc_roots(struct reloc_control *rc)
2372 {
2373         struct btrfs_root *root;
2374         struct btrfs_root *reloc_root;
2375         LIST_HEAD(reloc_roots);
2376         int found = 0;
2377         int ret = 0;
2378 again:
2379         root = rc->extent_root;
2380
2381         /*
2382          * this serializes us with btrfs_record_root_in_transaction,
2383          * we have to make sure nobody is in the middle of
2384          * adding their roots to the list while we are
2385          * doing this splice
2386          */
2387         mutex_lock(&root->fs_info->reloc_mutex);
2388         list_splice_init(&rc->reloc_roots, &reloc_roots);
2389         mutex_unlock(&root->fs_info->reloc_mutex);
2390
2391         while (!list_empty(&reloc_roots)) {
2392                 found = 1;
2393                 reloc_root = list_entry(reloc_roots.next,
2394                                         struct btrfs_root, root_list);
2395
2396                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2397                         root = read_fs_root(reloc_root->fs_info,
2398                                             reloc_root->root_key.offset);
2399                         BUG_ON(IS_ERR(root));
2400                         BUG_ON(root->reloc_root != reloc_root);
2401
2402                         ret = merge_reloc_root(rc, root);
2403                         if (ret) {
2404                                 if (list_empty(&reloc_root->root_list))
2405                                         list_add_tail(&reloc_root->root_list,
2406                                                       &reloc_roots);
2407                                 goto out;
2408                         }
2409                 } else {
2410                         list_del_init(&reloc_root->root_list);
2411                 }
2412
2413                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2414                 if (ret < 0) {
2415                         if (list_empty(&reloc_root->root_list))
2416                                 list_add_tail(&reloc_root->root_list,
2417                                               &reloc_roots);
2418                         goto out;
2419                 }
2420         }
2421
2422         if (found) {
2423                 found = 0;
2424                 goto again;
2425         }
2426 out:
2427         if (ret) {
2428                 btrfs_handle_fs_error(root->fs_info, ret, NULL);
2429                 if (!list_empty(&reloc_roots))
2430                         free_reloc_roots(&reloc_roots);
2431
2432                 /* new reloc root may be added */
2433                 mutex_lock(&root->fs_info->reloc_mutex);
2434                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2435                 mutex_unlock(&root->fs_info->reloc_mutex);
2436                 if (!list_empty(&reloc_roots))
2437                         free_reloc_roots(&reloc_roots);
2438         }
2439
2440         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2441 }
2442
2443 static void free_block_list(struct rb_root *blocks)
2444 {
2445         struct tree_block *block;
2446         struct rb_node *rb_node;
2447         while ((rb_node = rb_first(blocks))) {
2448                 block = rb_entry(rb_node, struct tree_block, rb_node);
2449                 rb_erase(rb_node, blocks);
2450                 kfree(block);
2451         }
2452 }
2453
2454 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2455                                       struct btrfs_root *reloc_root)
2456 {
2457         struct btrfs_root *root;
2458
2459         if (reloc_root->last_trans == trans->transid)
2460                 return 0;
2461
2462         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2463         BUG_ON(IS_ERR(root));
2464         BUG_ON(root->reloc_root != reloc_root);
2465
2466         return btrfs_record_root_in_trans(trans, root);
2467 }
2468
2469 static noinline_for_stack
2470 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2471                                      struct reloc_control *rc,
2472                                      struct backref_node *node,
2473                                      struct backref_edge *edges[])
2474 {
2475         struct backref_node *next;
2476         struct btrfs_root *root;
2477         int index = 0;
2478
2479         next = node;
2480         while (1) {
2481                 cond_resched();
2482                 next = walk_up_backref(next, edges, &index);
2483                 root = next->root;
2484                 BUG_ON(!root);
2485                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2486
2487                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2488                         record_reloc_root_in_trans(trans, root);
2489                         break;
2490                 }
2491
2492                 btrfs_record_root_in_trans(trans, root);
2493                 root = root->reloc_root;
2494
2495                 if (next->new_bytenr != root->node->start) {
2496                         BUG_ON(next->new_bytenr);
2497                         BUG_ON(!list_empty(&next->list));
2498                         next->new_bytenr = root->node->start;
2499                         next->root = root;
2500                         list_add_tail(&next->list,
2501                                       &rc->backref_cache.changed);
2502                         __mark_block_processed(rc, next);
2503                         break;
2504                 }
2505
2506                 WARN_ON(1);
2507                 root = NULL;
2508                 next = walk_down_backref(edges, &index);
2509                 if (!next || next->level <= node->level)
2510                         break;
2511         }
2512         if (!root)
2513                 return NULL;
2514
2515         next = node;
2516         /* setup backref node path for btrfs_reloc_cow_block */
2517         while (1) {
2518                 rc->backref_cache.path[next->level] = next;
2519                 if (--index < 0)
2520                         break;
2521                 next = edges[index]->node[UPPER];
2522         }
2523         return root;
2524 }
2525
2526 /*
2527  * select a tree root for relocation. return NULL if the block
2528  * is reference counted. we should use do_relocation() in this
2529  * case. return a tree root pointer if the block isn't reference
2530  * counted. return -ENOENT if the block is root of reloc tree.
2531  */
2532 static noinline_for_stack
2533 struct btrfs_root *select_one_root(struct backref_node *node)
2534 {
2535         struct backref_node *next;
2536         struct btrfs_root *root;
2537         struct btrfs_root *fs_root = NULL;
2538         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2539         int index = 0;
2540
2541         next = node;
2542         while (1) {
2543                 cond_resched();
2544                 next = walk_up_backref(next, edges, &index);
2545                 root = next->root;
2546                 BUG_ON(!root);
2547
2548                 /* no other choice for non-references counted tree */
2549                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2550                         return root;
2551
2552                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2553                         fs_root = root;
2554
2555                 if (next != node)
2556                         return NULL;
2557
2558                 next = walk_down_backref(edges, &index);
2559                 if (!next || next->level <= node->level)
2560                         break;
2561         }
2562
2563         if (!fs_root)
2564                 return ERR_PTR(-ENOENT);
2565         return fs_root;
2566 }
2567
2568 static noinline_for_stack
2569 u64 calcu_metadata_size(struct reloc_control *rc,
2570                         struct backref_node *node, int reserve)
2571 {
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 += rc->extent_root->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         u64 num_bytes;
2607         int ret;
2608         u64 tmp;
2609
2610         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2611
2612         trans->block_rsv = rc->block_rsv;
2613         rc->reserved_bytes += num_bytes;
2614
2615         /*
2616          * We are under a transaction here so we can only do limited flushing.
2617          * If we get an enospc just kick back -EAGAIN so we know to drop the
2618          * transaction and try to refill when we can flush all the things.
2619          */
2620         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2621                                 BTRFS_RESERVE_FLUSH_LIMIT);
2622         if (ret) {
2623                 tmp = rc->extent_root->nodesize * RELOCATION_RESERVED_NODES;
2624                 while (tmp <= rc->reserved_bytes)
2625                         tmp <<= 1;
2626                 /*
2627                  * only one thread can access block_rsv at this point,
2628                  * so we don't need hold lock to protect block_rsv.
2629                  * we expand more reservation size here to allow enough
2630                  * space for relocation and we will return eailer in
2631                  * enospc case.
2632                  */
2633                 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2634                         RELOCATION_RESERVED_NODES;
2635                 return -EAGAIN;
2636         }
2637
2638         return 0;
2639 }
2640
2641 /*
2642  * relocate a block tree, and then update pointers in upper level
2643  * blocks that reference the block to point to the new location.
2644  *
2645  * if called by link_to_upper, the block has already been relocated.
2646  * in that case this function just updates pointers.
2647  */
2648 static int do_relocation(struct btrfs_trans_handle *trans,
2649                          struct reloc_control *rc,
2650                          struct backref_node *node,
2651                          struct btrfs_key *key,
2652                          struct btrfs_path *path, int lowest)
2653 {
2654         struct backref_node *upper;
2655         struct backref_edge *edge;
2656         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2657         struct btrfs_root *root;
2658         struct extent_buffer *eb;
2659         u32 blocksize;
2660         u64 bytenr;
2661         u64 generation;
2662         int slot;
2663         int ret;
2664         int err = 0;
2665
2666         BUG_ON(lowest && node->eb);
2667
2668         path->lowest_level = node->level + 1;
2669         rc->backref_cache.path[node->level] = node;
2670         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2671                 cond_resched();
2672
2673                 upper = edge->node[UPPER];
2674                 root = select_reloc_root(trans, rc, upper, edges);
2675                 BUG_ON(!root);
2676
2677                 if (upper->eb && !upper->locked) {
2678                         if (!lowest) {
2679                                 ret = btrfs_bin_search(upper->eb, key,
2680                                                        upper->level, &slot);
2681                                 BUG_ON(ret);
2682                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2683                                 if (node->eb->start == bytenr)
2684                                         goto next;
2685                         }
2686                         drop_node_buffer(upper);
2687                 }
2688
2689                 if (!upper->eb) {
2690                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2691                         if (ret) {
2692                                 if (ret < 0)
2693                                         err = ret;
2694                                 else
2695                                         err = -ENOENT;
2696
2697                                 btrfs_release_path(path);
2698                                 break;
2699                         }
2700
2701                         if (!upper->eb) {
2702                                 upper->eb = path->nodes[upper->level];
2703                                 path->nodes[upper->level] = NULL;
2704                         } else {
2705                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2706                         }
2707
2708                         upper->locked = 1;
2709                         path->locks[upper->level] = 0;
2710
2711                         slot = path->slots[upper->level];
2712                         btrfs_release_path(path);
2713                 } else {
2714                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2715                                                &slot);
2716                         BUG_ON(ret);
2717                 }
2718
2719                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2720                 if (lowest) {
2721                         if (bytenr != node->bytenr) {
2722                                 btrfs_err(root->fs_info,
2723                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2724                                           bytenr, node->bytenr, slot,
2725                                           upper->eb->start);
2726                                 err = -EIO;
2727                                 goto next;
2728                         }
2729                 } else {
2730                         if (node->eb->start == bytenr)
2731                                 goto next;
2732                 }
2733
2734                 blocksize = root->nodesize;
2735                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2736                 eb = read_tree_block(root, bytenr, generation);
2737                 if (IS_ERR(eb)) {
2738                         err = PTR_ERR(eb);
2739                         goto next;
2740                 } else if (!extent_buffer_uptodate(eb)) {
2741                         free_extent_buffer(eb);
2742                         err = -EIO;
2743                         goto next;
2744                 }
2745                 btrfs_tree_lock(eb);
2746                 btrfs_set_lock_blocking(eb);
2747
2748                 if (!node->eb) {
2749                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2750                                               slot, &eb);
2751                         btrfs_tree_unlock(eb);
2752                         free_extent_buffer(eb);
2753                         if (ret < 0) {
2754                                 err = ret;
2755                                 goto next;
2756                         }
2757                         BUG_ON(node->eb != eb);
2758                 } else {
2759                         btrfs_set_node_blockptr(upper->eb, slot,
2760                                                 node->eb->start);
2761                         btrfs_set_node_ptr_generation(upper->eb, slot,
2762                                                       trans->transid);
2763                         btrfs_mark_buffer_dirty(upper->eb);
2764
2765                         ret = btrfs_inc_extent_ref(trans, root,
2766                                                 node->eb->start, blocksize,
2767                                                 upper->eb->start,
2768                                                 btrfs_header_owner(upper->eb),
2769                                                 node->level, 0);
2770                         BUG_ON(ret);
2771
2772                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2773                         BUG_ON(ret);
2774                 }
2775 next:
2776                 if (!upper->pending)
2777                         drop_node_buffer(upper);
2778                 else
2779                         unlock_node_buffer(upper);
2780                 if (err)
2781                         break;
2782         }
2783
2784         if (!err && node->pending) {
2785                 drop_node_buffer(node);
2786                 list_move_tail(&node->list, &rc->backref_cache.changed);
2787                 node->pending = 0;
2788         }
2789
2790         path->lowest_level = 0;
2791         BUG_ON(err == -ENOSPC);
2792         return err;
2793 }
2794
2795 static int link_to_upper(struct btrfs_trans_handle *trans,
2796                          struct reloc_control *rc,
2797                          struct backref_node *node,
2798                          struct btrfs_path *path)
2799 {
2800         struct btrfs_key key;
2801
2802         btrfs_node_key_to_cpu(node->eb, &key, 0);
2803         return do_relocation(trans, rc, node, &key, path, 0);
2804 }
2805
2806 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2807                                 struct reloc_control *rc,
2808                                 struct btrfs_path *path, int err)
2809 {
2810         LIST_HEAD(list);
2811         struct backref_cache *cache = &rc->backref_cache;
2812         struct backref_node *node;
2813         int level;
2814         int ret;
2815
2816         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2817                 while (!list_empty(&cache->pending[level])) {
2818                         node = list_entry(cache->pending[level].next,
2819                                           struct backref_node, list);
2820                         list_move_tail(&node->list, &list);
2821                         BUG_ON(!node->pending);
2822
2823                         if (!err) {
2824                                 ret = link_to_upper(trans, rc, node, path);
2825                                 if (ret < 0)
2826                                         err = ret;
2827                         }
2828                 }
2829                 list_splice_init(&list, &cache->pending[level]);
2830         }
2831         return err;
2832 }
2833
2834 static void mark_block_processed(struct reloc_control *rc,
2835                                  u64 bytenr, u32 blocksize)
2836 {
2837         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2838                         EXTENT_DIRTY);
2839 }
2840
2841 static void __mark_block_processed(struct reloc_control *rc,
2842                                    struct backref_node *node)
2843 {
2844         u32 blocksize;
2845         if (node->level == 0 ||
2846             in_block_group(node->bytenr, rc->block_group)) {
2847                 blocksize = rc->extent_root->nodesize;
2848                 mark_block_processed(rc, node->bytenr, blocksize);
2849         }
2850         node->processed = 1;
2851 }
2852
2853 /*
2854  * mark a block and all blocks directly/indirectly reference the block
2855  * as processed.
2856  */
2857 static void update_processed_blocks(struct reloc_control *rc,
2858                                     struct backref_node *node)
2859 {
2860         struct backref_node *next = node;
2861         struct backref_edge *edge;
2862         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2863         int index = 0;
2864
2865         while (next) {
2866                 cond_resched();
2867                 while (1) {
2868                         if (next->processed)
2869                                 break;
2870
2871                         __mark_block_processed(rc, next);
2872
2873                         if (list_empty(&next->upper))
2874                                 break;
2875
2876                         edge = list_entry(next->upper.next,
2877                                           struct backref_edge, list[LOWER]);
2878                         edges[index++] = edge;
2879                         next = edge->node[UPPER];
2880                 }
2881                 next = walk_down_backref(edges, &index);
2882         }
2883 }
2884
2885 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2886 {
2887         u32 blocksize = rc->extent_root->nodesize;
2888
2889         if (test_range_bit(&rc->processed_blocks, bytenr,
2890                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2891                 return 1;
2892         return 0;
2893 }
2894
2895 static int get_tree_block_key(struct reloc_control *rc,
2896                               struct tree_block *block)
2897 {
2898         struct extent_buffer *eb;
2899
2900         BUG_ON(block->key_ready);
2901         eb = read_tree_block(rc->extent_root, block->bytenr,
2902                              block->key.offset);
2903         if (IS_ERR(eb)) {
2904                 return PTR_ERR(eb);
2905         } else if (!extent_buffer_uptodate(eb)) {
2906                 free_extent_buffer(eb);
2907                 return -EIO;
2908         }
2909         WARN_ON(btrfs_header_level(eb) != block->level);
2910         if (block->level == 0)
2911                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2912         else
2913                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2914         free_extent_buffer(eb);
2915         block->key_ready = 1;
2916         return 0;
2917 }
2918
2919 /*
2920  * helper function to relocate a tree block
2921  */
2922 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2923                                 struct reloc_control *rc,
2924                                 struct backref_node *node,
2925                                 struct btrfs_key *key,
2926                                 struct btrfs_path *path)
2927 {
2928         struct btrfs_root *root;
2929         int ret = 0;
2930
2931         if (!node)
2932                 return 0;
2933
2934         BUG_ON(node->processed);
2935         root = select_one_root(node);
2936         if (root == ERR_PTR(-ENOENT)) {
2937                 update_processed_blocks(rc, node);
2938                 goto out;
2939         }
2940
2941         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2942                 ret = reserve_metadata_space(trans, rc, node);
2943                 if (ret)
2944                         goto out;
2945         }
2946
2947         if (root) {
2948                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2949                         BUG_ON(node->new_bytenr);
2950                         BUG_ON(!list_empty(&node->list));
2951                         btrfs_record_root_in_trans(trans, root);
2952                         root = root->reloc_root;
2953                         node->new_bytenr = root->node->start;
2954                         node->root = root;
2955                         list_add_tail(&node->list, &rc->backref_cache.changed);
2956                 } else {
2957                         path->lowest_level = node->level;
2958                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2959                         btrfs_release_path(path);
2960                         if (ret > 0)
2961                                 ret = 0;
2962                 }
2963                 if (!ret)
2964                         update_processed_blocks(rc, node);
2965         } else {
2966                 ret = do_relocation(trans, rc, node, key, path, 1);
2967         }
2968 out:
2969         if (ret || node->level == 0 || node->cowonly)
2970                 remove_backref_node(&rc->backref_cache, node);
2971         return ret;
2972 }
2973
2974 /*
2975  * relocate a list of blocks
2976  */
2977 static noinline_for_stack
2978 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2979                          struct reloc_control *rc, struct rb_root *blocks)
2980 {
2981         struct backref_node *node;
2982         struct btrfs_path *path;
2983         struct tree_block *block;
2984         struct rb_node *rb_node;
2985         int ret;
2986         int err = 0;
2987
2988         path = btrfs_alloc_path();
2989         if (!path) {
2990                 err = -ENOMEM;
2991                 goto out_free_blocks;
2992         }
2993
2994         rb_node = rb_first(blocks);
2995         while (rb_node) {
2996                 block = rb_entry(rb_node, struct tree_block, rb_node);
2997                 if (!block->key_ready)
2998                         readahead_tree_block(rc->extent_root, block->bytenr);
2999                 rb_node = rb_next(rb_node);
3000         }
3001
3002         rb_node = rb_first(blocks);
3003         while (rb_node) {
3004                 block = rb_entry(rb_node, struct tree_block, rb_node);
3005                 if (!block->key_ready) {
3006                         err = get_tree_block_key(rc, block);
3007                         if (err)
3008                                 goto out_free_path;
3009                 }
3010                 rb_node = rb_next(rb_node);
3011         }
3012
3013         rb_node = rb_first(blocks);
3014         while (rb_node) {
3015                 block = rb_entry(rb_node, struct tree_block, rb_node);
3016
3017                 node = build_backref_tree(rc, &block->key,
3018                                           block->level, block->bytenr);
3019                 if (IS_ERR(node)) {
3020                         err = PTR_ERR(node);
3021                         goto out;
3022                 }
3023
3024                 ret = relocate_tree_block(trans, rc, node, &block->key,
3025                                           path);
3026                 if (ret < 0) {
3027                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3028                                 err = ret;
3029                         goto out;
3030                 }
3031                 rb_node = rb_next(rb_node);
3032         }
3033 out:
3034         err = finish_pending_nodes(trans, rc, path, err);
3035
3036 out_free_path:
3037         btrfs_free_path(path);
3038 out_free_blocks:
3039         free_block_list(blocks);
3040         return err;
3041 }
3042
3043 static noinline_for_stack
3044 int prealloc_file_extent_cluster(struct inode *inode,
3045                                  struct file_extent_cluster *cluster)
3046 {
3047         u64 alloc_hint = 0;
3048         u64 start;
3049         u64 end;
3050         u64 offset = BTRFS_I(inode)->index_cnt;
3051         u64 num_bytes;
3052         int nr = 0;
3053         int ret = 0;
3054         u64 prealloc_start = cluster->start - offset;
3055         u64 prealloc_end = cluster->end - offset;
3056         u64 cur_offset;
3057
3058         BUG_ON(cluster->start != cluster->boundary[0]);
3059         inode_lock(inode);
3060
3061         ret = btrfs_check_data_free_space(inode, prealloc_start,
3062                                           prealloc_end + 1 - prealloc_start);
3063         if (ret)
3064                 goto out;
3065
3066         cur_offset = prealloc_start;
3067         while (nr < cluster->nr) {
3068                 start = cluster->boundary[nr] - offset;
3069                 if (nr + 1 < cluster->nr)
3070                         end = cluster->boundary[nr + 1] - 1 - offset;
3071                 else
3072                         end = cluster->end - offset;
3073
3074                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3075                 num_bytes = end + 1 - start;
3076                 if (cur_offset < start)
3077                         btrfs_free_reserved_data_space(inode, cur_offset,
3078                                         start - cur_offset);
3079                 ret = btrfs_prealloc_file_range(inode, 0, start,
3080                                                 num_bytes, num_bytes,
3081                                                 end + 1, &alloc_hint);
3082                 cur_offset = end + 1;
3083                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3084                 if (ret)
3085                         break;
3086                 nr++;
3087         }
3088         if (cur_offset < prealloc_end)
3089                 btrfs_free_reserved_data_space(inode, cur_offset,
3090                                        prealloc_end + 1 - cur_offset);
3091 out:
3092         inode_unlock(inode);
3093         return ret;
3094 }
3095
3096 static noinline_for_stack
3097 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3098                          u64 block_start)
3099 {
3100         struct btrfs_root *root = BTRFS_I(inode)->root;
3101         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3102         struct extent_map *em;
3103         int ret = 0;
3104
3105         em = alloc_extent_map();
3106         if (!em)
3107                 return -ENOMEM;
3108
3109         em->start = start;
3110         em->len = end + 1 - start;
3111         em->block_len = em->len;
3112         em->block_start = block_start;
3113         em->bdev = root->fs_info->fs_devices->latest_bdev;
3114         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3115
3116         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3117         while (1) {
3118                 write_lock(&em_tree->lock);
3119                 ret = add_extent_mapping(em_tree, em, 0);
3120                 write_unlock(&em_tree->lock);
3121                 if (ret != -EEXIST) {
3122                         free_extent_map(em);
3123                         break;
3124                 }
3125                 btrfs_drop_extent_cache(inode, start, end, 0);
3126         }
3127         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3128         return ret;
3129 }
3130
3131 static int relocate_file_extent_cluster(struct inode *inode,
3132                                         struct file_extent_cluster *cluster)
3133 {
3134         u64 page_start;
3135         u64 page_end;
3136         u64 offset = BTRFS_I(inode)->index_cnt;
3137         unsigned long index;
3138         unsigned long last_index;
3139         struct page *page;
3140         struct file_ra_state *ra;
3141         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3142         int nr = 0;
3143         int ret = 0;
3144
3145         if (!cluster->nr)
3146                 return 0;
3147
3148         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3149         if (!ra)
3150                 return -ENOMEM;
3151
3152         ret = prealloc_file_extent_cluster(inode, cluster);
3153         if (ret)
3154                 goto out;
3155
3156         file_ra_state_init(ra, inode->i_mapping);
3157
3158         ret = setup_extent_mapping(inode, cluster->start - offset,
3159                                    cluster->end - offset, cluster->start);
3160         if (ret)
3161                 goto out;
3162
3163         index = (cluster->start - offset) >> PAGE_SHIFT;
3164         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3165         while (index <= last_index) {
3166                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_SIZE);
3167                 if (ret)
3168                         goto out;
3169
3170                 page = find_lock_page(inode->i_mapping, index);
3171                 if (!page) {
3172                         page_cache_sync_readahead(inode->i_mapping,
3173                                                   ra, NULL, index,
3174                                                   last_index + 1 - index);
3175                         page = find_or_create_page(inode->i_mapping, index,
3176                                                    mask);
3177                         if (!page) {
3178                                 btrfs_delalloc_release_metadata(inode,
3179                                                         PAGE_SIZE);
3180                                 ret = -ENOMEM;
3181                                 goto out;
3182                         }
3183                 }
3184
3185                 if (PageReadahead(page)) {
3186                         page_cache_async_readahead(inode->i_mapping,
3187                                                    ra, NULL, page, index,
3188                                                    last_index + 1 - index);
3189                 }
3190
3191                 if (!PageUptodate(page)) {
3192                         btrfs_readpage(NULL, page);
3193                         lock_page(page);
3194                         if (!PageUptodate(page)) {
3195                                 unlock_page(page);
3196                                 put_page(page);
3197                                 btrfs_delalloc_release_metadata(inode,
3198                                                         PAGE_SIZE);
3199                                 ret = -EIO;
3200                                 goto out;
3201                         }
3202                 }
3203
3204                 page_start = page_offset(page);
3205                 page_end = page_start + PAGE_SIZE - 1;
3206
3207                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3208
3209                 set_page_extent_mapped(page);
3210
3211                 if (nr < cluster->nr &&
3212                     page_start + offset == cluster->boundary[nr]) {
3213                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3214                                         page_start, page_end,
3215                                         EXTENT_BOUNDARY);
3216                         nr++;
3217                 }
3218
3219                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL, 0);
3220                 set_page_dirty(page);
3221
3222                 unlock_extent(&BTRFS_I(inode)->io_tree,
3223                               page_start, page_end);
3224                 unlock_page(page);
3225                 put_page(page);
3226
3227                 index++;
3228                 balance_dirty_pages_ratelimited(inode->i_mapping);
3229                 btrfs_throttle(BTRFS_I(inode)->root);
3230         }
3231         WARN_ON(nr != cluster->nr);
3232 out:
3233         kfree(ra);
3234         return ret;
3235 }
3236
3237 static noinline_for_stack
3238 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3239                          struct file_extent_cluster *cluster)
3240 {
3241         int ret;
3242
3243         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3244                 ret = relocate_file_extent_cluster(inode, cluster);
3245                 if (ret)
3246                         return ret;
3247                 cluster->nr = 0;
3248         }
3249
3250         if (!cluster->nr)
3251                 cluster->start = extent_key->objectid;
3252         else
3253                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3254         cluster->end = extent_key->objectid + extent_key->offset - 1;
3255         cluster->boundary[cluster->nr] = extent_key->objectid;
3256         cluster->nr++;
3257
3258         if (cluster->nr >= MAX_EXTENTS) {
3259                 ret = relocate_file_extent_cluster(inode, cluster);
3260                 if (ret)
3261                         return ret;
3262                 cluster->nr = 0;
3263         }
3264         return 0;
3265 }
3266
3267 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3268 static int get_ref_objectid_v0(struct reloc_control *rc,
3269                                struct btrfs_path *path,
3270                                struct btrfs_key *extent_key,
3271                                u64 *ref_objectid, int *path_change)
3272 {
3273         struct btrfs_key key;
3274         struct extent_buffer *leaf;
3275         struct btrfs_extent_ref_v0 *ref0;
3276         int ret;
3277         int slot;
3278
3279         leaf = path->nodes[0];
3280         slot = path->slots[0];
3281         while (1) {
3282                 if (slot >= btrfs_header_nritems(leaf)) {
3283                         ret = btrfs_next_leaf(rc->extent_root, path);
3284                         if (ret < 0)
3285                                 return ret;
3286                         BUG_ON(ret > 0);
3287                         leaf = path->nodes[0];
3288                         slot = path->slots[0];
3289                         if (path_change)
3290                                 *path_change = 1;
3291                 }
3292                 btrfs_item_key_to_cpu(leaf, &key, slot);
3293                 if (key.objectid != extent_key->objectid)
3294                         return -ENOENT;
3295
3296                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3297                         slot++;
3298                         continue;
3299                 }
3300                 ref0 = btrfs_item_ptr(leaf, slot,
3301                                 struct btrfs_extent_ref_v0);
3302                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3303                 break;
3304         }
3305         return 0;
3306 }
3307 #endif
3308
3309 /*
3310  * helper to add a tree block to the list.
3311  * the major work is getting the generation and level of the block
3312  */
3313 static int add_tree_block(struct reloc_control *rc,
3314                           struct btrfs_key *extent_key,
3315                           struct btrfs_path *path,
3316                           struct rb_root *blocks)
3317 {
3318         struct extent_buffer *eb;
3319         struct btrfs_extent_item *ei;
3320         struct btrfs_tree_block_info *bi;
3321         struct tree_block *block;
3322         struct rb_node *rb_node;
3323         u32 item_size;
3324         int level = -1;
3325         u64 generation;
3326
3327         eb =  path->nodes[0];
3328         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3329
3330         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3331             item_size >= sizeof(*ei) + sizeof(*bi)) {
3332                 ei = btrfs_item_ptr(eb, path->slots[0],
3333                                 struct btrfs_extent_item);
3334                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3335                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3336                         level = btrfs_tree_block_level(eb, bi);
3337                 } else {
3338                         level = (int)extent_key->offset;
3339                 }
3340                 generation = btrfs_extent_generation(eb, ei);
3341         } else {
3342 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3343                 u64 ref_owner;
3344                 int ret;
3345
3346                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3347                 ret = get_ref_objectid_v0(rc, path, extent_key,
3348                                           &ref_owner, NULL);
3349                 if (ret < 0)
3350                         return ret;
3351                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3352                 level = (int)ref_owner;
3353                 /* FIXME: get real generation */
3354                 generation = 0;
3355 #else
3356                 BUG();
3357 #endif
3358         }
3359
3360         btrfs_release_path(path);
3361
3362         BUG_ON(level == -1);
3363
3364         block = kmalloc(sizeof(*block), GFP_NOFS);
3365         if (!block)
3366                 return -ENOMEM;
3367
3368         block->bytenr = extent_key->objectid;
3369         block->key.objectid = rc->extent_root->nodesize;
3370         block->key.offset = generation;
3371         block->level = level;
3372         block->key_ready = 0;
3373
3374         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3375         if (rb_node)
3376                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3377
3378         return 0;
3379 }
3380
3381 /*
3382  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3383  */
3384 static int __add_tree_block(struct reloc_control *rc,
3385                             u64 bytenr, u32 blocksize,
3386                             struct rb_root *blocks)
3387 {
3388         struct btrfs_path *path;
3389         struct btrfs_key key;
3390         int ret;
3391         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3392                                         SKINNY_METADATA);
3393
3394         if (tree_block_processed(bytenr, rc))
3395                 return 0;
3396
3397         if (tree_search(blocks, bytenr))
3398                 return 0;
3399
3400         path = btrfs_alloc_path();
3401         if (!path)
3402                 return -ENOMEM;
3403 again:
3404         key.objectid = bytenr;
3405         if (skinny) {
3406                 key.type = BTRFS_METADATA_ITEM_KEY;
3407                 key.offset = (u64)-1;
3408         } else {
3409                 key.type = BTRFS_EXTENT_ITEM_KEY;
3410                 key.offset = blocksize;
3411         }
3412
3413         path->search_commit_root = 1;
3414         path->skip_locking = 1;
3415         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3416         if (ret < 0)
3417                 goto out;
3418
3419         if (ret > 0 && skinny) {
3420                 if (path->slots[0]) {
3421                         path->slots[0]--;
3422                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3423                                               path->slots[0]);
3424                         if (key.objectid == bytenr &&
3425                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3426                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3427                               key.offset == blocksize)))
3428                                 ret = 0;
3429                 }
3430
3431                 if (ret) {
3432                         skinny = false;
3433                         btrfs_release_path(path);
3434                         goto again;
3435                 }
3436         }
3437         BUG_ON(ret);
3438
3439         ret = add_tree_block(rc, &key, path, blocks);
3440 out:
3441         btrfs_free_path(path);
3442         return ret;
3443 }
3444
3445 /*
3446  * helper to check if the block use full backrefs for pointers in it
3447  */
3448 static int block_use_full_backref(struct reloc_control *rc,
3449                                   struct extent_buffer *eb)
3450 {
3451         u64 flags;
3452         int ret;
3453
3454         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3455             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3456                 return 1;
3457
3458         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3459                                        eb->start, btrfs_header_level(eb), 1,
3460                                        NULL, &flags);
3461         BUG_ON(ret);
3462
3463         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3464                 ret = 1;
3465         else
3466                 ret = 0;
3467         return ret;
3468 }
3469
3470 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3471                                     struct btrfs_block_group_cache *block_group,
3472                                     struct inode *inode,
3473                                     u64 ino)
3474 {
3475         struct btrfs_key key;
3476         struct btrfs_root *root = fs_info->tree_root;
3477         struct btrfs_trans_handle *trans;
3478         int ret = 0;
3479
3480         if (inode)
3481                 goto truncate;
3482
3483         key.objectid = ino;
3484         key.type = BTRFS_INODE_ITEM_KEY;
3485         key.offset = 0;
3486
3487         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3488         if (IS_ERR(inode) || is_bad_inode(inode)) {
3489                 if (!IS_ERR(inode))
3490                         iput(inode);
3491                 return -ENOENT;
3492         }
3493
3494 truncate:
3495         ret = btrfs_check_trunc_cache_free_space(root,
3496                                                  &fs_info->global_block_rsv);
3497         if (ret)
3498                 goto out;
3499
3500         trans = btrfs_join_transaction(root);
3501         if (IS_ERR(trans)) {
3502                 ret = PTR_ERR(trans);
3503                 goto out;
3504         }
3505
3506         ret = btrfs_truncate_free_space_cache(root, trans, block_group, inode);
3507
3508         btrfs_end_transaction(trans, root);
3509         btrfs_btree_balance_dirty(root);
3510 out:
3511         iput(inode);
3512         return ret;
3513 }
3514
3515 /*
3516  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3517  * this function scans fs tree to find blocks reference the data extent
3518  */
3519 static int find_data_references(struct reloc_control *rc,
3520                                 struct btrfs_key *extent_key,
3521                                 struct extent_buffer *leaf,
3522                                 struct btrfs_extent_data_ref *ref,
3523                                 struct rb_root *blocks)
3524 {
3525         struct btrfs_path *path;
3526         struct tree_block *block;
3527         struct btrfs_root *root;
3528         struct btrfs_file_extent_item *fi;
3529         struct rb_node *rb_node;
3530         struct btrfs_key key;
3531         u64 ref_root;
3532         u64 ref_objectid;
3533         u64 ref_offset;
3534         u32 ref_count;
3535         u32 nritems;
3536         int err = 0;
3537         int added = 0;
3538         int counted;
3539         int ret;
3540
3541         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3542         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3543         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3544         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3545
3546         /*
3547          * This is an extent belonging to the free space cache, lets just delete
3548          * it and redo the search.
3549          */
3550         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3551                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3552                                                rc->block_group,
3553                                                NULL, ref_objectid);
3554                 if (ret != -ENOENT)
3555                         return ret;
3556                 ret = 0;
3557         }
3558
3559         path = btrfs_alloc_path();
3560         if (!path)
3561                 return -ENOMEM;
3562         path->reada = READA_FORWARD;
3563
3564         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3565         if (IS_ERR(root)) {
3566                 err = PTR_ERR(root);
3567                 goto out;
3568         }
3569
3570         key.objectid = ref_objectid;
3571         key.type = BTRFS_EXTENT_DATA_KEY;
3572         if (ref_offset > ((u64)-1 << 32))
3573                 key.offset = 0;
3574         else
3575                 key.offset = ref_offset;
3576
3577         path->search_commit_root = 1;
3578         path->skip_locking = 1;
3579         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3580         if (ret < 0) {
3581                 err = ret;
3582                 goto out;
3583         }
3584
3585         leaf = path->nodes[0];
3586         nritems = btrfs_header_nritems(leaf);
3587         /*
3588          * the references in tree blocks that use full backrefs
3589          * are not counted in
3590          */
3591         if (block_use_full_backref(rc, leaf))
3592                 counted = 0;
3593         else
3594                 counted = 1;
3595         rb_node = tree_search(blocks, leaf->start);
3596         if (rb_node) {
3597                 if (counted)
3598                         added = 1;
3599                 else
3600                         path->slots[0] = nritems;
3601         }
3602
3603         while (ref_count > 0) {
3604                 while (path->slots[0] >= nritems) {
3605                         ret = btrfs_next_leaf(root, path);
3606                         if (ret < 0) {
3607                                 err = ret;
3608                                 goto out;
3609                         }
3610                         if (WARN_ON(ret > 0))
3611                                 goto out;
3612
3613                         leaf = path->nodes[0];
3614                         nritems = btrfs_header_nritems(leaf);
3615                         added = 0;
3616
3617                         if (block_use_full_backref(rc, leaf))
3618                                 counted = 0;
3619                         else
3620                                 counted = 1;
3621                         rb_node = tree_search(blocks, leaf->start);
3622                         if (rb_node) {
3623                                 if (counted)
3624                                         added = 1;
3625                                 else
3626                                         path->slots[0] = nritems;
3627                         }
3628                 }
3629
3630                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3631                 if (WARN_ON(key.objectid != ref_objectid ||
3632                     key.type != BTRFS_EXTENT_DATA_KEY))
3633                         break;
3634
3635                 fi = btrfs_item_ptr(leaf, path->slots[0],
3636                                     struct btrfs_file_extent_item);
3637
3638                 if (btrfs_file_extent_type(leaf, fi) ==
3639                     BTRFS_FILE_EXTENT_INLINE)
3640                         goto next;
3641
3642                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3643                     extent_key->objectid)
3644                         goto next;
3645
3646                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3647                 if (key.offset != ref_offset)
3648                         goto next;
3649
3650                 if (counted)
3651                         ref_count--;
3652                 if (added)
3653                         goto next;
3654
3655                 if (!tree_block_processed(leaf->start, rc)) {
3656                         block = kmalloc(sizeof(*block), GFP_NOFS);
3657                         if (!block) {
3658                                 err = -ENOMEM;
3659                                 break;
3660                         }
3661                         block->bytenr = leaf->start;
3662                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3663                         block->level = 0;
3664                         block->key_ready = 1;
3665                         rb_node = tree_insert(blocks, block->bytenr,
3666                                               &block->rb_node);
3667                         if (rb_node)
3668                                 backref_tree_panic(rb_node, -EEXIST,
3669                                                    block->bytenr);
3670                 }
3671                 if (counted)
3672                         added = 1;
3673                 else
3674                         path->slots[0] = nritems;
3675 next:
3676                 path->slots[0]++;
3677
3678         }
3679 out:
3680         btrfs_free_path(path);
3681         return err;
3682 }
3683
3684 /*
3685  * helper to find all tree blocks that reference a given data extent
3686  */
3687 static noinline_for_stack
3688 int add_data_references(struct reloc_control *rc,
3689                         struct btrfs_key *extent_key,
3690                         struct btrfs_path *path,
3691                         struct rb_root *blocks)
3692 {
3693         struct btrfs_key key;
3694         struct extent_buffer *eb;
3695         struct btrfs_extent_data_ref *dref;
3696         struct btrfs_extent_inline_ref *iref;
3697         unsigned long ptr;
3698         unsigned long end;
3699         u32 blocksize = rc->extent_root->nodesize;
3700         int ret = 0;
3701         int err = 0;
3702
3703         eb = path->nodes[0];
3704         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3705         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3706 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3707         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3708                 ptr = end;
3709         else
3710 #endif
3711                 ptr += sizeof(struct btrfs_extent_item);
3712
3713         while (ptr < end) {
3714                 iref = (struct btrfs_extent_inline_ref *)ptr;
3715                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3716                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3717                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3718                         ret = __add_tree_block(rc, key.offset, blocksize,
3719                                                blocks);
3720                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3721                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3722                         ret = find_data_references(rc, extent_key,
3723                                                    eb, dref, blocks);
3724                 } else {
3725                         BUG();
3726                 }
3727                 if (ret) {
3728                         err = ret;
3729                         goto out;
3730                 }
3731                 ptr += btrfs_extent_inline_ref_size(key.type);
3732         }
3733         WARN_ON(ptr > end);
3734
3735         while (1) {
3736                 cond_resched();
3737                 eb = path->nodes[0];
3738                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3739                         ret = btrfs_next_leaf(rc->extent_root, path);
3740                         if (ret < 0) {
3741                                 err = ret;
3742                                 break;
3743                         }
3744                         if (ret > 0)
3745                                 break;
3746                         eb = path->nodes[0];
3747                 }
3748
3749                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3750                 if (key.objectid != extent_key->objectid)
3751                         break;
3752
3753 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3754                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3755                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3756 #else
3757                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3758                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3759 #endif
3760                         ret = __add_tree_block(rc, key.offset, blocksize,
3761                                                blocks);
3762                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3763                         dref = btrfs_item_ptr(eb, path->slots[0],
3764                                               struct btrfs_extent_data_ref);
3765                         ret = find_data_references(rc, extent_key,
3766                                                    eb, dref, blocks);
3767                 } else {
3768                         ret = 0;
3769                 }
3770                 if (ret) {
3771                         err = ret;
3772                         break;
3773                 }
3774                 path->slots[0]++;
3775         }
3776 out:
3777         btrfs_release_path(path);
3778         if (err)
3779                 free_block_list(blocks);
3780         return err;
3781 }
3782
3783 /*
3784  * helper to find next unprocessed extent
3785  */
3786 static noinline_for_stack
3787 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3788                      struct btrfs_key *extent_key)
3789 {
3790         struct btrfs_key key;
3791         struct extent_buffer *leaf;
3792         u64 start, end, last;
3793         int ret;
3794
3795         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3796         while (1) {
3797                 cond_resched();
3798                 if (rc->search_start >= last) {
3799                         ret = 1;
3800                         break;
3801                 }
3802
3803                 key.objectid = rc->search_start;
3804                 key.type = BTRFS_EXTENT_ITEM_KEY;
3805                 key.offset = 0;
3806
3807                 path->search_commit_root = 1;
3808                 path->skip_locking = 1;
3809                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3810                                         0, 0);
3811                 if (ret < 0)
3812                         break;
3813 next:
3814                 leaf = path->nodes[0];
3815                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3816                         ret = btrfs_next_leaf(rc->extent_root, path);
3817                         if (ret != 0)
3818                                 break;
3819                         leaf = path->nodes[0];
3820                 }
3821
3822                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3823                 if (key.objectid >= last) {
3824                         ret = 1;
3825                         break;
3826                 }
3827
3828                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3829                     key.type != BTRFS_METADATA_ITEM_KEY) {
3830                         path->slots[0]++;
3831                         goto next;
3832                 }
3833
3834                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3835                     key.objectid + key.offset <= rc->search_start) {
3836                         path->slots[0]++;
3837                         goto next;
3838                 }
3839
3840                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3841                     key.objectid + rc->extent_root->nodesize <=
3842                     rc->search_start) {
3843                         path->slots[0]++;
3844                         goto next;
3845                 }
3846
3847                 ret = find_first_extent_bit(&rc->processed_blocks,
3848                                             key.objectid, &start, &end,
3849                                             EXTENT_DIRTY, NULL);
3850
3851                 if (ret == 0 && start <= key.objectid) {
3852                         btrfs_release_path(path);
3853                         rc->search_start = end + 1;
3854                 } else {
3855                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3856                                 rc->search_start = key.objectid + key.offset;
3857                         else
3858                                 rc->search_start = key.objectid +
3859                                         rc->extent_root->nodesize;
3860                         memcpy(extent_key, &key, sizeof(key));
3861                         return 0;
3862                 }
3863         }
3864         btrfs_release_path(path);
3865         return ret;
3866 }
3867
3868 static void set_reloc_control(struct reloc_control *rc)
3869 {
3870         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3871
3872         mutex_lock(&fs_info->reloc_mutex);
3873         fs_info->reloc_ctl = rc;
3874         mutex_unlock(&fs_info->reloc_mutex);
3875 }
3876
3877 static void unset_reloc_control(struct reloc_control *rc)
3878 {
3879         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3880
3881         mutex_lock(&fs_info->reloc_mutex);
3882         fs_info->reloc_ctl = NULL;
3883         mutex_unlock(&fs_info->reloc_mutex);
3884 }
3885
3886 static int check_extent_flags(u64 flags)
3887 {
3888         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3889             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3890                 return 1;
3891         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3892             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3893                 return 1;
3894         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3895             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3896                 return 1;
3897         return 0;
3898 }
3899
3900 static noinline_for_stack
3901 int prepare_to_relocate(struct reloc_control *rc)
3902 {
3903         struct btrfs_trans_handle *trans;
3904         int ret;
3905
3906         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3907                                               BTRFS_BLOCK_RSV_TEMP);
3908         if (!rc->block_rsv)
3909                 return -ENOMEM;
3910
3911         memset(&rc->cluster, 0, sizeof(rc->cluster));
3912         rc->search_start = rc->block_group->key.objectid;
3913         rc->extents_found = 0;
3914         rc->nodes_relocated = 0;
3915         rc->merging_rsv_size = 0;
3916         rc->reserved_bytes = 0;
3917         rc->block_rsv->size = rc->extent_root->nodesize *
3918                               RELOCATION_RESERVED_NODES;
3919         ret = btrfs_block_rsv_refill(rc->extent_root,
3920                                      rc->block_rsv, rc->block_rsv->size,
3921                                      BTRFS_RESERVE_FLUSH_ALL);
3922         if (ret)
3923                 return ret;
3924
3925         rc->create_reloc_tree = 1;
3926         set_reloc_control(rc);
3927
3928         trans = btrfs_join_transaction(rc->extent_root);
3929         if (IS_ERR(trans)) {
3930                 unset_reloc_control(rc);
3931                 /*
3932                  * extent tree is not a ref_cow tree and has no reloc_root to
3933                  * cleanup.  And callers are responsible to free the above
3934                  * block rsv.
3935                  */
3936                 return PTR_ERR(trans);
3937         }
3938         btrfs_commit_transaction(trans, rc->extent_root);
3939         return 0;
3940 }
3941
3942 /*
3943  * Qgroup fixer for data chunk relocation.
3944  * The data relocation is done in the following steps
3945  * 1) Copy data extents into data reloc tree
3946  * 2) Create tree reloc tree(special snapshot) for related subvolumes
3947  * 3) Modify file extents in tree reloc tree
3948  * 4) Merge tree reloc tree with original fs tree, by swapping tree blocks
3949  *
3950  * The problem is, data and tree reloc tree are not accounted to qgroup,
3951  * and 4) will only info qgroup to track tree blocks change, not file extents
3952  * in the tree blocks.
3953  *
3954  * The good news is, related data extents are all in data reloc tree, so we
3955  * only need to info qgroup to track all file extents in data reloc tree
3956  * before commit trans.
3957  */
3958 static int qgroup_fix_relocated_data_extents(struct btrfs_trans_handle *trans,
3959                                              struct reloc_control *rc)
3960 {
3961         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3962         struct inode *inode = rc->data_inode;
3963         struct btrfs_root *data_reloc_root = BTRFS_I(inode)->root;
3964         struct btrfs_path *path;
3965         struct btrfs_key key;
3966         int ret = 0;
3967
3968         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
3969                 return 0;
3970
3971         /*
3972          * Only for stage where we update data pointers the qgroup fix is
3973          * valid.
3974          * For MOVING_DATA stage, we will miss the timing of swapping tree
3975          * blocks, and won't fix it.
3976          */
3977         if (!(rc->stage == UPDATE_DATA_PTRS && rc->extents_found))
3978                 return 0;
3979
3980         path = btrfs_alloc_path();
3981         if (!path)
3982                 return -ENOMEM;
3983         key.objectid = btrfs_ino(inode);
3984         key.type = BTRFS_EXTENT_DATA_KEY;
3985         key.offset = 0;
3986
3987         ret = btrfs_search_slot(NULL, data_reloc_root, &key, path, 0, 0);
3988         if (ret < 0)
3989                 goto out;
3990
3991         lock_extent(&BTRFS_I(inode)->io_tree, 0, (u64)-1);
3992         while (1) {
3993                 struct btrfs_file_extent_item *fi;
3994
3995                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3996                 if (key.objectid > btrfs_ino(inode))
3997                         break;
3998                 if (key.type != BTRFS_EXTENT_DATA_KEY)
3999                         goto next;
4000                 fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
4001                                     struct btrfs_file_extent_item);
4002                 if (btrfs_file_extent_type(path->nodes[0], fi) !=
4003                                 BTRFS_FILE_EXTENT_REG)
4004                         goto next;
4005                 ret = btrfs_qgroup_insert_dirty_extent(trans, fs_info,
4006                         btrfs_file_extent_disk_bytenr(path->nodes[0], fi),
4007                         btrfs_file_extent_disk_num_bytes(path->nodes[0], fi),
4008                         GFP_NOFS);
4009                 if (ret < 0)
4010                         break;
4011 next:
4012                 ret = btrfs_next_item(data_reloc_root, path);
4013                 if (ret < 0)
4014                         break;
4015                 if (ret > 0) {
4016                         ret = 0;
4017                         break;
4018                 }
4019         }
4020         unlock_extent(&BTRFS_I(inode)->io_tree, 0 , (u64)-1);
4021 out:
4022         btrfs_free_path(path);
4023         return ret;
4024 }
4025
4026 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
4027 {
4028         struct rb_root blocks = RB_ROOT;
4029         struct btrfs_key key;
4030         struct btrfs_trans_handle *trans = NULL;
4031         struct btrfs_path *path;
4032         struct btrfs_extent_item *ei;
4033         u64 flags;
4034         u32 item_size;
4035         int ret;
4036         int err = 0;
4037         int progress = 0;
4038
4039         path = btrfs_alloc_path();
4040         if (!path)
4041                 return -ENOMEM;
4042         path->reada = READA_FORWARD;
4043
4044         ret = prepare_to_relocate(rc);
4045         if (ret) {
4046                 err = ret;
4047                 goto out_free;
4048         }
4049
4050         while (1) {
4051                 rc->reserved_bytes = 0;
4052                 ret = btrfs_block_rsv_refill(rc->extent_root,
4053                                         rc->block_rsv, rc->block_rsv->size,
4054                                         BTRFS_RESERVE_FLUSH_ALL);
4055                 if (ret) {
4056                         err = ret;
4057                         break;
4058                 }
4059                 progress++;
4060                 trans = btrfs_start_transaction(rc->extent_root, 0);
4061                 if (IS_ERR(trans)) {
4062                         err = PTR_ERR(trans);
4063                         trans = NULL;
4064                         break;
4065                 }
4066 restart:
4067                 if (update_backref_cache(trans, &rc->backref_cache)) {
4068                         btrfs_end_transaction(trans, rc->extent_root);
4069                         continue;
4070                 }
4071
4072                 ret = find_next_extent(rc, path, &key);
4073                 if (ret < 0)
4074                         err = ret;
4075                 if (ret != 0)
4076                         break;
4077
4078                 rc->extents_found++;
4079
4080                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4081                                     struct btrfs_extent_item);
4082                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4083                 if (item_size >= sizeof(*ei)) {
4084                         flags = btrfs_extent_flags(path->nodes[0], ei);
4085                         ret = check_extent_flags(flags);
4086                         BUG_ON(ret);
4087
4088                 } else {
4089 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4090                         u64 ref_owner;
4091                         int path_change = 0;
4092
4093                         BUG_ON(item_size !=
4094                                sizeof(struct btrfs_extent_item_v0));
4095                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
4096                                                   &path_change);
4097                         if (ret < 0) {
4098                                 err = ret;
4099                                 break;
4100                         }
4101                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
4102                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
4103                         else
4104                                 flags = BTRFS_EXTENT_FLAG_DATA;
4105
4106                         if (path_change) {
4107                                 btrfs_release_path(path);
4108
4109                                 path->search_commit_root = 1;
4110                                 path->skip_locking = 1;
4111                                 ret = btrfs_search_slot(NULL, rc->extent_root,
4112                                                         &key, path, 0, 0);
4113                                 if (ret < 0) {
4114                                         err = ret;
4115                                         break;
4116                                 }
4117                                 BUG_ON(ret > 0);
4118                         }
4119 #else
4120                         BUG();
4121 #endif
4122                 }
4123
4124                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4125                         ret = add_tree_block(rc, &key, path, &blocks);
4126                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4127                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4128                         ret = add_data_references(rc, &key, path, &blocks);
4129                 } else {
4130                         btrfs_release_path(path);
4131                         ret = 0;
4132                 }
4133                 if (ret < 0) {
4134                         err = ret;
4135                         break;
4136                 }
4137
4138                 if (!RB_EMPTY_ROOT(&blocks)) {
4139                         ret = relocate_tree_blocks(trans, rc, &blocks);
4140                         if (ret < 0) {
4141                                 /*
4142                                  * if we fail to relocate tree blocks, force to update
4143                                  * backref cache when committing transaction.
4144                                  */
4145                                 rc->backref_cache.last_trans = trans->transid - 1;
4146
4147                                 if (ret != -EAGAIN) {
4148                                         err = ret;
4149                                         break;
4150                                 }
4151                                 rc->extents_found--;
4152                                 rc->search_start = key.objectid;
4153                         }
4154                 }
4155
4156                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4157                 btrfs_btree_balance_dirty(rc->extent_root);
4158                 trans = NULL;
4159
4160                 if (rc->stage == MOVE_DATA_EXTENTS &&
4161                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4162                         rc->found_file_extent = 1;
4163                         ret = relocate_data_extent(rc->data_inode,
4164                                                    &key, &rc->cluster);
4165                         if (ret < 0) {
4166                                 err = ret;
4167                                 break;
4168                         }
4169                 }
4170         }
4171         if (trans && progress && err == -ENOSPC) {
4172                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4173                                               rc->block_group->flags);
4174                 if (ret == 1) {
4175                         err = 0;
4176                         progress = 0;
4177                         goto restart;
4178                 }
4179         }
4180
4181         btrfs_release_path(path);
4182         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4183
4184         if (trans) {
4185                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4186                 btrfs_btree_balance_dirty(rc->extent_root);
4187         }
4188
4189         if (!err) {
4190                 ret = relocate_file_extent_cluster(rc->data_inode,
4191                                                    &rc->cluster);
4192                 if (ret < 0)
4193                         err = ret;
4194         }
4195
4196         rc->create_reloc_tree = 0;
4197         set_reloc_control(rc);
4198
4199         backref_cache_cleanup(&rc->backref_cache);
4200         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4201
4202         err = prepare_to_merge(rc, err);
4203
4204         merge_reloc_roots(rc);
4205
4206         rc->merge_reloc_tree = 0;
4207         unset_reloc_control(rc);
4208         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4209
4210         /* get rid of pinned extents */
4211         trans = btrfs_join_transaction(rc->extent_root);
4212         if (IS_ERR(trans)) {
4213                 err = PTR_ERR(trans);
4214                 goto out_free;
4215         }
4216         ret = qgroup_fix_relocated_data_extents(trans, rc);
4217         if (ret < 0) {
4218                 btrfs_abort_transaction(trans, ret);
4219                 if (!err)
4220                         err = ret;
4221                 goto out_free;
4222         }
4223         btrfs_commit_transaction(trans, rc->extent_root);
4224 out_free:
4225         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4226         btrfs_free_path(path);
4227         return err;
4228 }
4229
4230 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4231                                  struct btrfs_root *root, u64 objectid)
4232 {
4233         struct btrfs_path *path;
4234         struct btrfs_inode_item *item;
4235         struct extent_buffer *leaf;
4236         int ret;
4237
4238         path = btrfs_alloc_path();
4239         if (!path)
4240                 return -ENOMEM;
4241
4242         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4243         if (ret)
4244                 goto out;
4245
4246         leaf = path->nodes[0];
4247         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4248         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4249         btrfs_set_inode_generation(leaf, item, 1);
4250         btrfs_set_inode_size(leaf, item, 0);
4251         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4252         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4253                                           BTRFS_INODE_PREALLOC);
4254         btrfs_mark_buffer_dirty(leaf);
4255 out:
4256         btrfs_free_path(path);
4257         return ret;
4258 }
4259
4260 /*
4261  * helper to create inode for data relocation.
4262  * the inode is in data relocation tree and its link count is 0
4263  */
4264 static noinline_for_stack
4265 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4266                                  struct btrfs_block_group_cache *group)
4267 {
4268         struct inode *inode = NULL;
4269         struct btrfs_trans_handle *trans;
4270         struct btrfs_root *root;
4271         struct btrfs_key key;
4272         u64 objectid;
4273         int err = 0;
4274
4275         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4276         if (IS_ERR(root))
4277                 return ERR_CAST(root);
4278
4279         trans = btrfs_start_transaction(root, 6);
4280         if (IS_ERR(trans))
4281                 return ERR_CAST(trans);
4282
4283         err = btrfs_find_free_objectid(root, &objectid);
4284         if (err)
4285                 goto out;
4286
4287         err = __insert_orphan_inode(trans, root, objectid);
4288         BUG_ON(err);
4289
4290         key.objectid = objectid;
4291         key.type = BTRFS_INODE_ITEM_KEY;
4292         key.offset = 0;
4293         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4294         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4295         BTRFS_I(inode)->index_cnt = group->key.objectid;
4296
4297         err = btrfs_orphan_add(trans, inode);
4298 out:
4299         btrfs_end_transaction(trans, root);
4300         btrfs_btree_balance_dirty(root);
4301         if (err) {
4302                 if (inode)
4303                         iput(inode);
4304                 inode = ERR_PTR(err);
4305         }
4306         return inode;
4307 }
4308
4309 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4310 {
4311         struct reloc_control *rc;
4312
4313         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4314         if (!rc)
4315                 return NULL;
4316
4317         INIT_LIST_HEAD(&rc->reloc_roots);
4318         backref_cache_init(&rc->backref_cache);
4319         mapping_tree_init(&rc->reloc_root_tree);
4320         extent_io_tree_init(&rc->processed_blocks,
4321                             fs_info->btree_inode->i_mapping);
4322         return rc;
4323 }
4324
4325 /*
4326  * function to relocate all extents in a block group.
4327  */
4328 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4329 {
4330         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4331         struct reloc_control *rc;
4332         struct inode *inode;
4333         struct btrfs_path *path;
4334         int ret;
4335         int rw = 0;
4336         int err = 0;
4337
4338         rc = alloc_reloc_control(fs_info);
4339         if (!rc)
4340                 return -ENOMEM;
4341
4342         rc->extent_root = extent_root;
4343
4344         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4345         BUG_ON(!rc->block_group);
4346
4347         ret = btrfs_inc_block_group_ro(extent_root, rc->block_group);
4348         if (ret) {
4349                 err = ret;
4350                 goto out;
4351         }
4352         rw = 1;
4353
4354         path = btrfs_alloc_path();
4355         if (!path) {
4356                 err = -ENOMEM;
4357                 goto out;
4358         }
4359
4360         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4361                                         path);
4362         btrfs_free_path(path);
4363
4364         if (!IS_ERR(inode))
4365                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4366         else
4367                 ret = PTR_ERR(inode);
4368
4369         if (ret && ret != -ENOENT) {
4370                 err = ret;
4371                 goto out;
4372         }
4373
4374         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4375         if (IS_ERR(rc->data_inode)) {
4376                 err = PTR_ERR(rc->data_inode);
4377                 rc->data_inode = NULL;
4378                 goto out;
4379         }
4380
4381         btrfs_info(extent_root->fs_info,
4382                    "relocating block group %llu flags %llu",
4383                    rc->block_group->key.objectid, rc->block_group->flags);
4384
4385         btrfs_wait_block_group_reservations(rc->block_group);
4386         btrfs_wait_nocow_writers(rc->block_group);
4387         btrfs_wait_ordered_roots(fs_info, -1,
4388                                  rc->block_group->key.objectid,
4389                                  rc->block_group->key.offset);
4390
4391         while (1) {
4392                 mutex_lock(&fs_info->cleaner_mutex);
4393                 ret = relocate_block_group(rc);
4394                 mutex_unlock(&fs_info->cleaner_mutex);
4395                 if (ret < 0) {
4396                         err = ret;
4397                         goto out;
4398                 }
4399
4400                 if (rc->extents_found == 0)
4401                         break;
4402
4403                 btrfs_info(extent_root->fs_info, "found %llu extents",
4404                         rc->extents_found);
4405
4406                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4407                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4408                                                        (u64)-1);
4409                         if (ret) {
4410                                 err = ret;
4411                                 goto out;
4412                         }
4413                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4414                                                  0, -1);
4415                         rc->stage = UPDATE_DATA_PTRS;
4416                 }
4417         }
4418
4419         WARN_ON(rc->block_group->pinned > 0);
4420         WARN_ON(rc->block_group->reserved > 0);
4421         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4422 out:
4423         if (err && rw)
4424                 btrfs_dec_block_group_ro(extent_root, rc->block_group);
4425         iput(rc->data_inode);
4426         btrfs_put_block_group(rc->block_group);
4427         kfree(rc);
4428         return err;
4429 }
4430
4431 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4432 {
4433         struct btrfs_trans_handle *trans;
4434         int ret, err;
4435
4436         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4437         if (IS_ERR(trans))
4438                 return PTR_ERR(trans);
4439
4440         memset(&root->root_item.drop_progress, 0,
4441                 sizeof(root->root_item.drop_progress));
4442         root->root_item.drop_level = 0;
4443         btrfs_set_root_refs(&root->root_item, 0);
4444         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4445                                 &root->root_key, &root->root_item);
4446
4447         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4448         if (err)
4449                 return err;
4450         return ret;
4451 }
4452
4453 /*
4454  * recover relocation interrupted by system crash.
4455  *
4456  * this function resumes merging reloc trees with corresponding fs trees.
4457  * this is important for keeping the sharing of tree blocks
4458  */
4459 int btrfs_recover_relocation(struct btrfs_root *root)
4460 {
4461         LIST_HEAD(reloc_roots);
4462         struct btrfs_key key;
4463         struct btrfs_root *fs_root;
4464         struct btrfs_root *reloc_root;
4465         struct btrfs_path *path;
4466         struct extent_buffer *leaf;
4467         struct reloc_control *rc = NULL;
4468         struct btrfs_trans_handle *trans;
4469         int ret;
4470         int err = 0;
4471
4472         path = btrfs_alloc_path();
4473         if (!path)
4474                 return -ENOMEM;
4475         path->reada = READA_BACK;
4476
4477         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4478         key.type = BTRFS_ROOT_ITEM_KEY;
4479         key.offset = (u64)-1;
4480
4481         while (1) {
4482                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4483                                         path, 0, 0);
4484                 if (ret < 0) {
4485                         err = ret;
4486                         goto out;
4487                 }
4488                 if (ret > 0) {
4489                         if (path->slots[0] == 0)
4490                                 break;
4491                         path->slots[0]--;
4492                 }
4493                 leaf = path->nodes[0];
4494                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4495                 btrfs_release_path(path);
4496
4497                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4498                     key.type != BTRFS_ROOT_ITEM_KEY)
4499                         break;
4500
4501                 reloc_root = btrfs_read_fs_root(root, &key);
4502                 if (IS_ERR(reloc_root)) {
4503                         err = PTR_ERR(reloc_root);
4504                         goto out;
4505                 }
4506
4507                 list_add(&reloc_root->root_list, &reloc_roots);
4508
4509                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4510                         fs_root = read_fs_root(root->fs_info,
4511                                                reloc_root->root_key.offset);
4512                         if (IS_ERR(fs_root)) {
4513                                 ret = PTR_ERR(fs_root);
4514                                 if (ret != -ENOENT) {
4515                                         err = ret;
4516                                         goto out;
4517                                 }
4518                                 ret = mark_garbage_root(reloc_root);
4519                                 if (ret < 0) {
4520                                         err = ret;
4521                                         goto out;
4522                                 }
4523                         }
4524                 }
4525
4526                 if (key.offset == 0)
4527                         break;
4528
4529                 key.offset--;
4530         }
4531         btrfs_release_path(path);
4532
4533         if (list_empty(&reloc_roots))
4534                 goto out;
4535
4536         rc = alloc_reloc_control(root->fs_info);
4537         if (!rc) {
4538                 err = -ENOMEM;
4539                 goto out;
4540         }
4541
4542         rc->extent_root = root->fs_info->extent_root;
4543
4544         set_reloc_control(rc);
4545
4546         trans = btrfs_join_transaction(rc->extent_root);
4547         if (IS_ERR(trans)) {
4548                 unset_reloc_control(rc);
4549                 err = PTR_ERR(trans);
4550                 goto out_free;
4551         }
4552
4553         rc->merge_reloc_tree = 1;
4554
4555         while (!list_empty(&reloc_roots)) {
4556                 reloc_root = list_entry(reloc_roots.next,
4557                                         struct btrfs_root, root_list);
4558                 list_del(&reloc_root->root_list);
4559
4560                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4561                         list_add_tail(&reloc_root->root_list,
4562                                       &rc->reloc_roots);
4563                         continue;
4564                 }
4565
4566                 fs_root = read_fs_root(root->fs_info,
4567                                        reloc_root->root_key.offset);
4568                 if (IS_ERR(fs_root)) {
4569                         err = PTR_ERR(fs_root);
4570                         goto out_free;
4571                 }
4572
4573                 err = __add_reloc_root(reloc_root);
4574                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4575                 fs_root->reloc_root = reloc_root;
4576         }
4577
4578         err = btrfs_commit_transaction(trans, rc->extent_root);
4579         if (err)
4580                 goto out_free;
4581
4582         merge_reloc_roots(rc);
4583
4584         unset_reloc_control(rc);
4585
4586         trans = btrfs_join_transaction(rc->extent_root);
4587         if (IS_ERR(trans)) {
4588                 err = PTR_ERR(trans);
4589                 goto out_free;
4590         }
4591         err = qgroup_fix_relocated_data_extents(trans, rc);
4592         if (err < 0) {
4593                 btrfs_abort_transaction(trans, err);
4594                 goto out_free;
4595         }
4596         err = btrfs_commit_transaction(trans, rc->extent_root);
4597 out_free:
4598         kfree(rc);
4599 out:
4600         if (!list_empty(&reloc_roots))
4601                 free_reloc_roots(&reloc_roots);
4602
4603         btrfs_free_path(path);
4604
4605         if (err == 0) {
4606                 /* cleanup orphan inode in data relocation tree */
4607                 fs_root = read_fs_root(root->fs_info,
4608                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4609                 if (IS_ERR(fs_root))
4610                         err = PTR_ERR(fs_root);
4611                 else
4612                         err = btrfs_orphan_cleanup(fs_root);
4613         }
4614         return err;
4615 }
4616
4617 /*
4618  * helper to add ordered checksum for data relocation.
4619  *
4620  * cloning checksum properly handles the nodatasum extents.
4621  * it also saves CPU time to re-calculate the checksum.
4622  */
4623 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4624 {
4625         struct btrfs_ordered_sum *sums;
4626         struct btrfs_ordered_extent *ordered;
4627         struct btrfs_root *root = BTRFS_I(inode)->root;
4628         int ret;
4629         u64 disk_bytenr;
4630         u64 new_bytenr;
4631         LIST_HEAD(list);
4632
4633         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4634         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4635
4636         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4637         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4638                                        disk_bytenr + len - 1, &list, 0);
4639         if (ret)
4640                 goto out;
4641
4642         while (!list_empty(&list)) {
4643                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4644                 list_del_init(&sums->list);
4645
4646                 /*
4647                  * We need to offset the new_bytenr based on where the csum is.
4648                  * We need to do this because we will read in entire prealloc
4649                  * extents but we may have written to say the middle of the
4650                  * prealloc extent, so we need to make sure the csum goes with
4651                  * the right disk offset.
4652                  *
4653                  * We can do this because the data reloc inode refers strictly
4654                  * to the on disk bytes, so we don't have to worry about
4655                  * disk_len vs real len like with real inodes since it's all
4656                  * disk length.
4657                  */
4658                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4659                 sums->bytenr = new_bytenr;
4660
4661                 btrfs_add_ordered_sum(inode, ordered, sums);
4662         }
4663 out:
4664         btrfs_put_ordered_extent(ordered);
4665         return ret;
4666 }
4667
4668 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4669                           struct btrfs_root *root, struct extent_buffer *buf,
4670                           struct extent_buffer *cow)
4671 {
4672         struct reloc_control *rc;
4673         struct backref_node *node;
4674         int first_cow = 0;
4675         int level;
4676         int ret = 0;
4677
4678         rc = root->fs_info->reloc_ctl;
4679         if (!rc)
4680                 return 0;
4681
4682         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4683                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4684
4685         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4686                 if (buf == root->node)
4687                         __update_reloc_root(root, cow->start);
4688         }
4689
4690         level = btrfs_header_level(buf);
4691         if (btrfs_header_generation(buf) <=
4692             btrfs_root_last_snapshot(&root->root_item))
4693                 first_cow = 1;
4694
4695         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4696             rc->create_reloc_tree) {
4697                 WARN_ON(!first_cow && level == 0);
4698
4699                 node = rc->backref_cache.path[level];
4700                 BUG_ON(node->bytenr != buf->start &&
4701                        node->new_bytenr != buf->start);
4702
4703                 drop_node_buffer(node);
4704                 extent_buffer_get(cow);
4705                 node->eb = cow;
4706                 node->new_bytenr = cow->start;
4707
4708                 if (!node->pending) {
4709                         list_move_tail(&node->list,
4710                                        &rc->backref_cache.pending[level]);
4711                         node->pending = 1;
4712                 }
4713
4714                 if (first_cow)
4715                         __mark_block_processed(rc, node);
4716
4717                 if (first_cow && level > 0)
4718                         rc->nodes_relocated += buf->len;
4719         }
4720
4721         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4722                 ret = replace_file_extents(trans, rc, root, cow);
4723         return ret;
4724 }
4725
4726 /*
4727  * called before creating snapshot. it calculates metadata reservation
4728  * required for relocating tree blocks in the snapshot
4729  */
4730 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4731                               u64 *bytes_to_reserve)
4732 {
4733         struct btrfs_root *root;
4734         struct reloc_control *rc;
4735
4736         root = pending->root;
4737         if (!root->reloc_root)
4738                 return;
4739
4740         rc = root->fs_info->reloc_ctl;
4741         if (!rc->merge_reloc_tree)
4742                 return;
4743
4744         root = root->reloc_root;
4745         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4746         /*
4747          * relocation is in the stage of merging trees. the space
4748          * used by merging a reloc tree is twice the size of
4749          * relocated tree nodes in the worst case. half for cowing
4750          * the reloc tree, half for cowing the fs tree. the space
4751          * used by cowing the reloc tree will be freed after the
4752          * tree is dropped. if we create snapshot, cowing the fs
4753          * tree may use more space than it frees. so we need
4754          * reserve extra space.
4755          */
4756         *bytes_to_reserve += rc->nodes_relocated;
4757 }
4758
4759 /*
4760  * called after snapshot is created. migrate block reservation
4761  * and create reloc root for the newly created snapshot
4762  */
4763 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4764                                struct btrfs_pending_snapshot *pending)
4765 {
4766         struct btrfs_root *root = pending->root;
4767         struct btrfs_root *reloc_root;
4768         struct btrfs_root *new_root;
4769         struct reloc_control *rc;
4770         int ret;
4771
4772         if (!root->reloc_root)
4773                 return 0;
4774
4775         rc = root->fs_info->reloc_ctl;
4776         rc->merging_rsv_size += rc->nodes_relocated;
4777
4778         if (rc->merge_reloc_tree) {
4779                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4780                                               rc->block_rsv,
4781                                               rc->nodes_relocated, 1);
4782                 if (ret)
4783                         return ret;
4784         }
4785
4786         new_root = pending->snap;
4787         reloc_root = create_reloc_root(trans, root->reloc_root,
4788                                        new_root->root_key.objectid);
4789         if (IS_ERR(reloc_root))
4790                 return PTR_ERR(reloc_root);
4791
4792         ret = __add_reloc_root(reloc_root);
4793         BUG_ON(ret < 0);
4794         new_root->reloc_root = reloc_root;
4795
4796         if (rc->create_reloc_tree)
4797                 ret = clone_backref_node(trans, rc, root, reloc_root);
4798         return ret;
4799 }