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
  35 /*
  36  * backref_node, mapping_node and tree_block start with this
  37  */
  38 struct tree_entry {
  39         struct rb_node rb_node;
  40         u64 bytenr;
  41 };
  42
  43 /*
  44  * present a tree block in the backref cache
  45  */
  46 struct backref_node {
  47         struct rb_node rb_node;
  48         u64 bytenr;
  49
  50         u64 new_bytenr;
  51         /* objectid of tree block owner, can be not uptodate */
  52         u64 owner;
  53         /* link to pending, changed or detached list */
  54         struct list_head list;
  55         /* list of upper level blocks reference this block */
  56         struct list_head upper;
  57         /* list of child blocks in the cache */
  58         struct list_head lower;
  59         /* NULL if this node is not tree root */
  60         struct btrfs_root *root;
  61         /* extent buffer got by COW the block */
  62         struct extent_buffer *eb;
  63         /* level of tree block */
  64         unsigned int level:8;
  65         /* is the block in non-reference counted tree */
  66         unsigned int cowonly:1;
  67         /* 1 if no child node in the cache */
  68         unsigned int lowest:1;
  69         /* is the extent buffer locked */
  70         unsigned int locked:1;
  71         /* has the block been processed */
  72         unsigned int processed:1;
  73         /* have backrefs of this block been checked */
  74         unsigned int checked:1;
  75         /*
  76          * 1 if corresponding block has been cowed but some upper
  77          * level block pointers may not point to the new location
  78          */
  79         unsigned int pending:1;
  80         /*
  81          * 1 if the backref node isn't connected to any other
  82          * backref node.
  83          */
  84         unsigned int detached:1;
  85 };
  86
  87 /*
  88  * present a block pointer in the backref cache
  89  */
  90 struct backref_edge {
  91         struct list_head list[2];
  92         struct backref_node *node[2];
  93 };
  94
  95 #define LOWER   0
  96 #define UPPER   1
  97
  98 struct backref_cache {
  99         /* red black tree of all backref nodes in the cache */
 100         struct rb_root rb_root;
 101         /* for passing backref nodes to btrfs_reloc_cow_block */
 102         struct backref_node *path[BTRFS_MAX_LEVEL];
 103         /*
 104          * list of blocks that have been cowed but some block
 105          * pointers in upper level blocks may not reflect the
 106          * new location
 107          */
 108         struct list_head pending[BTRFS_MAX_LEVEL];
 109         /* list of backref nodes with no child node */
 110         struct list_head leaves;
 111         /* list of blocks that have been cowed in current transaction */
 112         struct list_head changed;
 113         /* list of detached backref node. */
 114         struct list_head detached;
 115
 116         u64 last_trans;
 117
 118         int nr_nodes;
 119         int nr_edges;
 120 };
 121
 122 /*
 123  * map address of tree root to tree
 124  */
 125 struct mapping_node {
 126         struct rb_node rb_node;
 127         u64 bytenr;
 128         void *data;
 129 };
 130
 131 struct mapping_tree {
 132         struct rb_root rb_root;
 133         spinlock_t lock;
 134 };
 135
 136 /*
 137  * present a tree block to process
 138  */
 139 struct tree_block {
 140         struct rb_node rb_node;
 141         u64 bytenr;
 142         struct btrfs_key key;
 143         unsigned int level:8;
 144         unsigned int key_ready:1;
 145 };
 146
 147 #define MAX_EXTENTS 128
 148
 149 struct file_extent_cluster {
 150         u64 start;
 151         u64 end;
 152         u64 boundary[MAX_EXTENTS];
 153         unsigned int nr;
 154 };
 155
 156 struct reloc_control {
 157         /* block group to relocate */
 158         struct btrfs_block_group_cache *block_group;
 159         /* extent tree */
 160         struct btrfs_root *extent_root;
 161         /* inode for moving data */
 162         struct inode *data_inode;
 163
 164         struct btrfs_block_rsv *block_rsv;
 165
 166         struct backref_cache backref_cache;
 167
 168         struct file_extent_cluster cluster;
 169         /* tree blocks have been processed */
 170         struct extent_io_tree processed_blocks;
 171         /* map start of tree root to corresponding reloc tree */
 172         struct mapping_tree reloc_root_tree;
 173         /* list of reloc trees */
 174         struct list_head reloc_roots;
 175         /* size of metadata reservation for merging reloc trees */
 176         u64 merging_rsv_size;
 177         /* size of relocated tree nodes */
 178         u64 nodes_relocated;
 179
 180         u64 search_start;
 181         u64 extents_found;
 182
 183         unsigned int stage:8;
 184         unsigned int create_reloc_tree:1;
 185         unsigned int merge_reloc_tree:1;
 186         unsigned int found_file_extent:1;
 187         unsigned int commit_transaction:1;
 188 };
 189
 190 /* stages of data relocation */
 191 #define MOVE_DATA_EXTENTS       0
 192 #define UPDATE_DATA_PTRS        1
 193
 194 static void remove_backref_node(struct backref_cache *cache,
 195                                 struct backref_node *node);
 196 static void __mark_block_processed(struct reloc_control *rc,
 197                                    struct backref_node *node);
 198
 199 static void mapping_tree_init(struct mapping_tree *tree)
 200 {
 201         tree->rb_root = RB_ROOT;
 202         spin_lock_init(&tree->lock);
 203 }
 204
 205 static void backref_cache_init(struct backref_cache *cache)
 206 {
 207         int i;
 208         cache->rb_root = RB_ROOT;
 209         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 210                 INIT_LIST_HEAD(&cache->pending[i]);
 211         INIT_LIST_HEAD(&cache->changed);
 212         INIT_LIST_HEAD(&cache->detached);
 213         INIT_LIST_HEAD(&cache->leaves);
 214 }
 215
 216 static void backref_cache_cleanup(struct backref_cache *cache)
 217 {
 218         struct backref_node *node;
 219         int i;
 220
 221         while (!list_empty(&cache->detached)) {
 222                 node = list_entry(cache->detached.next,
 223                                   struct backref_node, list);
 224                 remove_backref_node(cache, node);
 225         }
 226
 227         while (!list_empty(&cache->leaves)) {
 228                 node = list_entry(cache->leaves.next,
 229                                   struct backref_node, lower);
 230                 remove_backref_node(cache, node);
 231         }
 232
 233         cache->last_trans = 0;
 234
 235         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 236                 BUG_ON(!list_empty(&cache->pending[i]));
 237         BUG_ON(!list_empty(&cache->changed));
 238         BUG_ON(!list_empty(&cache->detached));
 239         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
 240         BUG_ON(cache->nr_nodes);
 241         BUG_ON(cache->nr_edges);
 242 }
 243
 244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 245 {
 246         struct backref_node *node;
 247
 248         node = kzalloc(sizeof(*node), GFP_NOFS);
 249         if (node) {
 250                 INIT_LIST_HEAD(&node->list);
 251                 INIT_LIST_HEAD(&node->upper);
 252                 INIT_LIST_HEAD(&node->lower);
 253                 RB_CLEAR_NODE(&node->rb_node);
 254                 cache->nr_nodes++;
 255         }
 256         return node;
 257 }
 258
 259 static void free_backref_node(struct backref_cache *cache,
 260                               struct backref_node *node)
 261 {
 262         if (node) {
 263                 cache->nr_nodes--;
 264                 kfree(node);
 265         }
 266 }
 267
 268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 269 {
 270         struct backref_edge *edge;
 271
 272         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 273         if (edge)
 274                 cache->nr_edges++;
 275         return edge;
 276 }
 277
 278 static void free_backref_edge(struct backref_cache *cache,
 279                               struct backref_edge *edge)
 280 {
 281         if (edge) {
 282                 cache->nr_edges--;
 283                 kfree(edge);
 284         }
 285 }
 286
 287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 288                                    struct rb_node *node)
 289 {
 290         struct rb_node **p = &root->rb_node;
 291         struct rb_node *parent = NULL;
 292         struct tree_entry *entry;
 293
 294         while (*p) {
 295                 parent = *p;
 296                 entry = rb_entry(parent, struct tree_entry, rb_node);
 297
 298                 if (bytenr < entry->bytenr)
 299                         p = &(*p)->rb_left;
 300                 else if (bytenr > entry->bytenr)
 301                         p = &(*p)->rb_right;
 302                 else
 303                         return parent;
 304         }
 305
 306         rb_link_node(node, parent, p);
 307         rb_insert_color(node, root);
 308         return NULL;
 309 }
 310
 311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 312 {
 313         struct rb_node *n = root->rb_node;
 314         struct tree_entry *entry;
 315
 316         while (n) {
 317                 entry = rb_entry(n, struct tree_entry, rb_node);
 318
 319                 if (bytenr < entry->bytenr)
 320                         n = n->rb_left;
 321                 else if (bytenr > entry->bytenr)
 322                         n = n->rb_right;
 323                 else
 324                         return n;
 325         }
 326         return NULL;
 327 }
 328
 329 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 330 {
 331
 332         struct btrfs_fs_info *fs_info = NULL;
 333         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 334                                               rb_node);
 335         if (bnode->root)
 336                 fs_info = bnode->root->fs_info;
 337         btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
 338                     "found at offset %llu\n", bytenr);
 339 }
 340
 341 /*
 342  * walk up backref nodes until reach node presents tree root
 343  */
 344 static struct backref_node *walk_up_backref(struct backref_node *node,
 345                                             struct backref_edge *edges[],
 346                                             int *index)
 347 {
 348         struct backref_edge *edge;
 349         int idx = *index;
 350
 351         while (!list_empty(&node->upper)) {
 352                 edge = list_entry(node->upper.next,
 353                                   struct backref_edge, list[LOWER]);
 354                 edges[idx++] = edge;
 355                 node = edge->node[UPPER];
 356         }
 357         BUG_ON(node->detached);
 358         *index = idx;
 359         return node;
 360 }
 361
 362 /*
 363  * walk down backref nodes to find start of next reference path
 364  */
 365 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 366                                               int *index)
 367 {
 368         struct backref_edge *edge;
 369         struct backref_node *lower;
 370         int idx = *index;
 371
 372         while (idx > 0) {
 373                 edge = edges[idx - 1];
 374                 lower = edge->node[LOWER];
 375                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 376                         idx--;
 377                         continue;
 378                 }
 379                 edge = list_entry(edge->list[LOWER].next,
 380                                   struct backref_edge, list[LOWER]);
 381                 edges[idx - 1] = edge;
 382                 *index = idx;
 383                 return edge->node[UPPER];
 384         }
 385         *index = 0;
 386         return NULL;
 387 }
 388
 389 static void unlock_node_buffer(struct backref_node *node)
 390 {
 391         if (node->locked) {
 392                 btrfs_tree_unlock(node->eb);
 393                 node->locked = 0;
 394         }
 395 }
 396
 397 static void drop_node_buffer(struct backref_node *node)
 398 {
 399         if (node->eb) {
 400                 unlock_node_buffer(node);
 401                 free_extent_buffer(node->eb);
 402                 node->eb = NULL;
 403         }
 404 }
 405
 406 static void drop_backref_node(struct backref_cache *tree,
 407                               struct backref_node *node)
 408 {
 409         BUG_ON(!list_empty(&node->upper));
 410
 411         drop_node_buffer(node);
 412         list_del(&node->list);
 413         list_del(&node->lower);
 414         if (!RB_EMPTY_NODE(&node->rb_node))
 415                 rb_erase(&node->rb_node, &tree->rb_root);
 416         free_backref_node(tree, node);
 417 }
 418
 419 /*
 420  * remove a backref node from the backref cache
 421  */
 422 static void remove_backref_node(struct backref_cache *cache,
 423                                 struct backref_node *node)
 424 {
 425         struct backref_node *upper;
 426         struct backref_edge *edge;
 427
 428         if (!node)
 429                 return;
 430
 431         BUG_ON(!node->lowest && !node->detached);
 432         while (!list_empty(&node->upper)) {
 433                 edge = list_entry(node->upper.next, struct backref_edge,
 434                                   list[LOWER]);
 435                 upper = edge->node[UPPER];
 436                 list_del(&edge->list[LOWER]);
 437                 list_del(&edge->list[UPPER]);
 438                 free_backref_edge(cache, edge);
 439
 440                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 441                         BUG_ON(!list_empty(&node->upper));
 442                         drop_backref_node(cache, node);
 443                         node = upper;
 444                         node->lowest = 1;
 445                         continue;
 446                 }
 447                 /*
 448                  * add the node to leaf node list if no other
 449                  * child block cached.
 450                  */
 451                 if (list_empty(&upper->lower)) {
 452                         list_add_tail(&upper->lower, &cache->leaves);
 453                         upper->lowest = 1;
 454                 }
 455         }
 456
 457         drop_backref_node(cache, node);
 458 }
 459
 460 static void update_backref_node(struct backref_cache *cache,
 461                                 struct backref_node *node, u64 bytenr)
 462 {
 463         struct rb_node *rb_node;
 464         rb_erase(&node->rb_node, &cache->rb_root);
 465         node->bytenr = bytenr;
 466         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 467         if (rb_node)
 468                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 469 }
 470
 471 /*
 472  * update backref cache after a transaction commit
 473  */
 474 static int update_backref_cache(struct btrfs_trans_handle *trans,
 475                                 struct backref_cache *cache)
 476 {
 477         struct backref_node *node;
 478         int level = 0;
 479
 480         if (cache->last_trans == 0) {
 481                 cache->last_trans = trans->transid;
 482                 return 0;
 483         }
 484
 485         if (cache->last_trans == trans->transid)
 486                 return 0;
 487
 488         /*
 489          * detached nodes are used to avoid unnecessary backref
 490          * lookup. transaction commit changes the extent tree.
 491          * so the detached nodes are no longer useful.
 492          */
 493         while (!list_empty(&cache->detached)) {
 494                 node = list_entry(cache->detached.next,
 495                                   struct backref_node, list);
 496                 remove_backref_node(cache, node);
 497         }
 498
 499         while (!list_empty(&cache->changed)) {
 500                 node = list_entry(cache->changed.next,
 501                                   struct backref_node, list);
 502                 list_del_init(&node->list);
 503                 BUG_ON(node->pending);
 504                 update_backref_node(cache, node, node->new_bytenr);
 505         }
 506
 507         /*
 508          * some nodes can be left in the pending list if there were
 509          * errors during processing the pending nodes.
 510          */
 511         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 512                 list_for_each_entry(node, &cache->pending[level], list) {
 513                         BUG_ON(!node->pending);
 514                         if (node->bytenr == node->new_bytenr)
 515                                 continue;
 516                         update_backref_node(cache, node, node->new_bytenr);
 517                 }
 518         }
 519
 520         cache->last_trans = 0;
 521         return 1;
 522 }
 523
 524
 525 static int should_ignore_root(struct btrfs_root *root)
 526 {
 527         struct btrfs_root *reloc_root;
 528
 529         if (!root->ref_cows)
 530                 return 0;
 531
 532         reloc_root = root->reloc_root;
 533         if (!reloc_root)
 534                 return 0;
 535
 536         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 537             root->fs_info->running_transaction->transid - 1)
 538                 return 0;
 539         /*
 540          * if there is reloc tree and it was created in previous
 541          * transaction backref lookup can find the reloc tree,
 542          * so backref node for the fs tree root is useless for
 543          * relocation.
 544          */
 545         return 1;
 546 }
 547 /*
 548  * find reloc tree by address of tree root
 549  */
 550 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 551                                           u64 bytenr)
 552 {
 553         struct rb_node *rb_node;
 554         struct mapping_node *node;
 555         struct btrfs_root *root = NULL;
 556
 557         spin_lock(&rc->reloc_root_tree.lock);
 558         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 559         if (rb_node) {
 560                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 561                 root = (struct btrfs_root *)node->data;
 562         }
 563         spin_unlock(&rc->reloc_root_tree.lock);
 564         return root;
 565 }
 566
 567 static int is_cowonly_root(u64 root_objectid)
 568 {
 569         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 570             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 571             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 572             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 573             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 574             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
 575                 return 1;
 576         return 0;
 577 }
 578
 579 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 580                                         u64 root_objectid)
 581 {
 582         struct btrfs_key key;
 583
 584         key.objectid = root_objectid;
 585         key.type = BTRFS_ROOT_ITEM_KEY;
 586         if (is_cowonly_root(root_objectid))
 587                 key.offset = 0;
 588         else
 589                 key.offset = (u64)-1;
 590
 591         return btrfs_read_fs_root_no_name(fs_info, &key);
 592 }
 593
 594 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 595 static noinline_for_stack
 596 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 597                                   struct extent_buffer *leaf,
 598                                   struct btrfs_extent_ref_v0 *ref0)
 599 {
 600         struct btrfs_root *root;
 601         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 602         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 603
 604         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 605
 606         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 607         BUG_ON(IS_ERR(root));
 608
 609         if (root->ref_cows &&
 610             generation != btrfs_root_generation(&root->root_item))
 611                 return NULL;
 612
 613         return root;
 614 }
 615 #endif
 616
 617 static noinline_for_stack
 618 int find_inline_backref(struct extent_buffer *leaf, int slot,
 619                         unsigned long *ptr, unsigned long *end)
 620 {
 621         struct btrfs_key key;
 622         struct btrfs_extent_item *ei;
 623         struct btrfs_tree_block_info *bi;
 624         u32 item_size;
 625
 626         btrfs_item_key_to_cpu(leaf, &key, slot);
 627
 628         item_size = btrfs_item_size_nr(leaf, slot);
 629 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 630         if (item_size < sizeof(*ei)) {
 631                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 632                 return 1;
 633         }
 634 #endif
 635         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 636         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 637                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 638
 639         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 640             item_size <= sizeof(*ei) + sizeof(*bi)) {
 641                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 642                 return 1;
 643         }
 644         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 645             item_size <= sizeof(*ei)) {
 646                 WARN_ON(item_size < sizeof(*ei));
 647                 return 1;
 648         }
 649
 650         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 651                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 652                 *ptr = (unsigned long)(bi + 1);
 653         } else {
 654                 *ptr = (unsigned long)(ei + 1);
 655         }
 656         *end = (unsigned long)ei + item_size;
 657         return 0;
 658 }
 659
 660 /*
 661  * build backref tree for a given tree block. root of the backref tree
 662  * corresponds the tree block, leaves of the backref tree correspond
 663  * roots of b-trees that reference the tree block.
 664  *
 665  * the basic idea of this function is check backrefs of a given block
 666  * to find upper level blocks that refernece the block, and then check
 667  * bakcrefs of these upper level blocks recursively. the recursion stop
 668  * when tree root is reached or backrefs for the block is cached.
 669  *
 670  * NOTE: if we find backrefs for a block are cached, we know backrefs
 671  * for all upper level blocks that directly/indirectly reference the
 672  * block are also cached.
 673  */
 674 static noinline_for_stack
 675 struct backref_node *build_backref_tree(struct reloc_control *rc,
 676                                         struct btrfs_key *node_key,
 677                                         int level, u64 bytenr)
 678 {
 679         struct backref_cache *cache = &rc->backref_cache;
 680         struct btrfs_path *path1;
 681         struct btrfs_path *path2;
 682         struct extent_buffer *eb;
 683         struct btrfs_root *root;
 684         struct backref_node *cur;
 685         struct backref_node *upper;
 686         struct backref_node *lower;
 687         struct backref_node *node = NULL;
 688         struct backref_node *exist = NULL;
 689         struct backref_edge *edge;
 690         struct rb_node *rb_node;
 691         struct btrfs_key key;
 692         unsigned long end;
 693         unsigned long ptr;
 694         LIST_HEAD(list);
 695         LIST_HEAD(useless);
 696         int cowonly;
 697         int ret;
 698         int err = 0;
 699         bool need_check = true;
 700
 701         path1 = btrfs_alloc_path();
 702         path2 = btrfs_alloc_path();
 703         if (!path1 || !path2) {
 704                 err = -ENOMEM;
 705                 goto out;
 706         }
 707         path1->reada = 1;
 708         path2->reada = 2;
 709
 710         node = alloc_backref_node(cache);
 711         if (!node) {
 712                 err = -ENOMEM;
 713                 goto out;
 714         }
 715
 716         node->bytenr = bytenr;
 717         node->level = level;
 718         node->lowest = 1;
 719         cur = node;
 720 again:
 721         end = 0;
 722         ptr = 0;
 723         key.objectid = cur->bytenr;
 724         key.type = BTRFS_METADATA_ITEM_KEY;
 725         key.offset = (u64)-1;
 726
 727         path1->search_commit_root = 1;
 728         path1->skip_locking = 1;
 729         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 730                                 0, 0);
 731         if (ret < 0) {
 732                 err = ret;
 733                 goto out;
 734         }
 735         BUG_ON(!ret || !path1->slots[0]);
 736
 737         path1->slots[0]--;
 738
 739         WARN_ON(cur->checked);
 740         if (!list_empty(&cur->upper)) {
 741                 /*
 742                  * the backref was added previously when processing
 743                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 744                  */
 745                 BUG_ON(!list_is_singular(&cur->upper));
 746                 edge = list_entry(cur->upper.next, struct backref_edge,
 747                                   list[LOWER]);
 748                 BUG_ON(!list_empty(&edge->list[UPPER]));
 749                 exist = edge->node[UPPER];
 750                 /*
 751                  * add the upper level block to pending list if we need
 752                  * check its backrefs
 753                  */
 754                 if (!exist->checked)
 755                         list_add_tail(&edge->list[UPPER], &list);
 756         } else {
 757                 exist = NULL;
 758         }
 759
 760         while (1) {
 761                 cond_resched();
 762                 eb = path1->nodes[0];
 763
 764                 if (ptr >= end) {
 765                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 766                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 767                                 if (ret < 0) {
 768                                         err = ret;
 769                                         goto out;
 770                                 }
 771                                 if (ret > 0)
 772                                         break;
 773                                 eb = path1->nodes[0];
 774                         }
 775
 776                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 777                         if (key.objectid != cur->bytenr) {
 778                                 WARN_ON(exist);
 779                                 break;
 780                         }
 781
 782                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 783                             key.type == BTRFS_METADATA_ITEM_KEY) {
 784                                 ret = find_inline_backref(eb, path1->slots[0],
 785                                                           &ptr, &end);
 786                                 if (ret)
 787                                         goto next;
 788                         }
 789                 }
 790
 791                 if (ptr < end) {
 792                         /* update key for inline back ref */
 793                         struct btrfs_extent_inline_ref *iref;
 794                         iref = (struct btrfs_extent_inline_ref *)ptr;
 795                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 796                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 797                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 798                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 799                 }
 800
 801                 if (exist &&
 802                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 803                       exist->owner == key.offset) ||
 804                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 805                       exist->bytenr == key.offset))) {
 806                         exist = NULL;
 807                         goto next;
 808                 }
 809
 810 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 811                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 812                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 813                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 814                                 struct btrfs_extent_ref_v0 *ref0;
 815                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 816                                                 struct btrfs_extent_ref_v0);
 817                                 if (key.objectid == key.offset) {
 818                                         root = find_tree_root(rc, eb, ref0);
 819                                         if (root && !should_ignore_root(root))
 820                                                 cur->root = root;
 821                                         else
 822                                                 list_add(&cur->list, &useless);
 823                                         break;
 824                                 }
 825                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 826                                                                       ref0)))
 827                                         cur->cowonly = 1;
 828                         }
 829 #else
 830                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 831                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 832 #endif
 833                         if (key.objectid == key.offset) {
 834                                 /*
 835                                  * only root blocks of reloc trees use
 836                                  * backref of this type.
 837                                  */
 838                                 root = find_reloc_root(rc, cur->bytenr);
 839                                 BUG_ON(!root);
 840                                 cur->root = root;
 841                                 break;
 842                         }
 843
 844                         edge = alloc_backref_edge(cache);
 845                         if (!edge) {
 846                                 err = -ENOMEM;
 847                                 goto out;
 848                         }
 849                         rb_node = tree_search(&cache->rb_root, key.offset);
 850                         if (!rb_node) {
 851                                 upper = alloc_backref_node(cache);
 852                                 if (!upper) {
 853                                         free_backref_edge(cache, edge);
 854                                         err = -ENOMEM;
 855                                         goto out;
 856                                 }
 857                                 upper->bytenr = key.offset;
 858                                 upper->level = cur->level + 1;
 859                                 /*
 860                                  *  backrefs for the upper level block isn't
 861                                  *  cached, add the block to pending list
 862                                  */
 863                                 list_add_tail(&edge->list[UPPER], &list);
 864                         } else {
 865                                 upper = rb_entry(rb_node, struct backref_node,
 866                                                  rb_node);
 867                                 BUG_ON(!upper->checked);
 868                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 869                         }
 870                         list_add_tail(&edge->list[LOWER], &cur->upper);
 871                         edge->node[LOWER] = cur;
 872                         edge->node[UPPER] = upper;
 873
 874                         goto next;
 875                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 876                         goto next;
 877                 }
 878
 879                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 880                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 881                 if (IS_ERR(root)) {
 882                         err = PTR_ERR(root);
 883                         goto out;
 884                 }
 885
 886                 if (!root->ref_cows)
 887                         cur->cowonly = 1;
 888
 889                 if (btrfs_root_level(&root->root_item) == cur->level) {
 890                         /* tree root */
 891                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 892                                cur->bytenr);
 893                         if (should_ignore_root(root))
 894                                 list_add(&cur->list, &useless);
 895                         else
 896                                 cur->root = root;
 897                         break;
 898                 }
 899
 900                 level = cur->level + 1;
 901
 902                 /*
 903                  * searching the tree to find upper level blocks
 904                  * reference the block.
 905                  */
 906                 path2->search_commit_root = 1;
 907                 path2->skip_locking = 1;
 908                 path2->lowest_level = level;
 909                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 910                 path2->lowest_level = 0;
 911                 if (ret < 0) {
 912                         err = ret;
 913                         goto out;
 914                 }
 915                 if (ret > 0 && path2->slots[level] > 0)
 916                         path2->slots[level]--;
 917
 918                 eb = path2->nodes[level];
 919                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 920                         cur->bytenr);
 921
 922                 lower = cur;
 923                 need_check = true;
 924                 for (; level < BTRFS_MAX_LEVEL; level++) {
 925                         if (!path2->nodes[level]) {
 926                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 927                                        lower->bytenr);
 928                                 if (should_ignore_root(root))
 929                                         list_add(&lower->list, &useless);
 930                                 else
 931                                         lower->root = root;
 932                                 break;
 933                         }
 934
 935                         edge = alloc_backref_edge(cache);
 936                         if (!edge) {
 937                                 err = -ENOMEM;
 938                                 goto out;
 939                         }
 940
 941                         eb = path2->nodes[level];
 942                         rb_node = tree_search(&cache->rb_root, eb->start);
 943                         if (!rb_node) {
 944                                 upper = alloc_backref_node(cache);
 945                                 if (!upper) {
 946                                         free_backref_edge(cache, edge);
 947                                         err = -ENOMEM;
 948                                         goto out;
 949                                 }
 950                                 upper->bytenr = eb->start;
 951                                 upper->owner = btrfs_header_owner(eb);
 952                                 upper->level = lower->level + 1;
 953                                 if (!root->ref_cows)
 954                                         upper->cowonly = 1;
 955
 956                                 /*
 957                                  * if we know the block isn't shared
 958                                  * we can void checking its backrefs.
 959                                  */
 960                                 if (btrfs_block_can_be_shared(root, eb))
 961                                         upper->checked = 0;
 962                                 else
 963                                         upper->checked = 1;
 964
 965                                 /*
 966                                  * add the block to pending list if we
 967                                  * need check its backrefs, we only do this once
 968                                  * while walking up a tree as we will catch
 969                                  * anything else later on.
 970                                  */
 971                                 if (!upper->checked && need_check) {
 972                                         need_check = false;
 973                                         list_add_tail(&edge->list[UPPER],
 974                                                       &list);
 975                                 } else
 976                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 977                         } else {
 978                                 upper = rb_entry(rb_node, struct backref_node,
 979                                                  rb_node);
 980                                 BUG_ON(!upper->checked);
 981                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 982                                 if (!upper->owner)
 983                                         upper->owner = btrfs_header_owner(eb);
 984                         }
 985                         list_add_tail(&edge->list[LOWER], &lower->upper);
 986                         edge->node[LOWER] = lower;
 987                         edge->node[UPPER] = upper;
 988
 989                         if (rb_node)
 990                                 break;
 991                         lower = upper;
 992                         upper = NULL;
 993                 }
 994                 btrfs_release_path(path2);
 995 next:
 996                 if (ptr < end) {
 997                         ptr += btrfs_extent_inline_ref_size(key.type);
 998                         if (ptr >= end) {
 999                                 WARN_ON(ptr > end);
1000                                 ptr = 0;
1001                                 end = 0;
1002                         }
1003                 }
1004                 if (ptr >= end)
1005                         path1->slots[0]++;
1006         }
1007         btrfs_release_path(path1);
1008
1009         cur->checked = 1;
1010         WARN_ON(exist);
1011
1012         /* the pending list isn't empty, take the first block to process */
1013         if (!list_empty(&list)) {
1014                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1015                 list_del_init(&edge->list[UPPER]);
1016                 cur = edge->node[UPPER];
1017                 goto again;
1018         }
1019
1020         /*
1021          * everything goes well, connect backref nodes and insert backref nodes
1022          * into the cache.
1023          */
1024         BUG_ON(!node->checked);
1025         cowonly = node->cowonly;
1026         if (!cowonly) {
1027                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1028                                       &node->rb_node);
1029                 if (rb_node)
1030                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1031                 list_add_tail(&node->lower, &cache->leaves);
1032         }
1033
1034         list_for_each_entry(edge, &node->upper, list[LOWER])
1035                 list_add_tail(&edge->list[UPPER], &list);
1036
1037         while (!list_empty(&list)) {
1038                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1039                 list_del_init(&edge->list[UPPER]);
1040                 upper = edge->node[UPPER];
1041                 if (upper->detached) {
1042                         list_del(&edge->list[LOWER]);
1043                         lower = edge->node[LOWER];
1044                         free_backref_edge(cache, edge);
1045                         if (list_empty(&lower->upper))
1046                                 list_add(&lower->list, &useless);
1047                         continue;
1048                 }
1049
1050                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1051                         if (upper->lowest) {
1052                                 list_del_init(&upper->lower);
1053                                 upper->lowest = 0;
1054                         }
1055
1056                         list_add_tail(&edge->list[UPPER], &upper->lower);
1057                         continue;
1058                 }
1059
1060                 BUG_ON(!upper->checked);
1061                 BUG_ON(cowonly != upper->cowonly);
1062                 if (!cowonly) {
1063                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1064                                               &upper->rb_node);
1065                         if (rb_node)
1066                                 backref_tree_panic(rb_node, -EEXIST,
1067                                                    upper->bytenr);
1068                 }
1069
1070                 list_add_tail(&edge->list[UPPER], &upper->lower);
1071
1072                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1073                         list_add_tail(&edge->list[UPPER], &list);
1074         }
1075         /*
1076          * process useless backref nodes. backref nodes for tree leaves
1077          * are deleted from the cache. backref nodes for upper level
1078          * tree blocks are left in the cache to avoid unnecessary backref
1079          * lookup.
1080          */
1081         while (!list_empty(&useless)) {
1082                 upper = list_entry(useless.next, struct backref_node, list);
1083                 list_del_init(&upper->list);
1084                 BUG_ON(!list_empty(&upper->upper));
1085                 if (upper == node)
1086                         node = NULL;
1087                 if (upper->lowest) {
1088                         list_del_init(&upper->lower);
1089                         upper->lowest = 0;
1090                 }
1091                 while (!list_empty(&upper->lower)) {
1092                         edge = list_entry(upper->lower.next,
1093                                           struct backref_edge, list[UPPER]);
1094                         list_del(&edge->list[UPPER]);
1095                         list_del(&edge->list[LOWER]);
1096                         lower = edge->node[LOWER];
1097                         free_backref_edge(cache, edge);
1098
1099                         if (list_empty(&lower->upper))
1100                                 list_add(&lower->list, &useless);
1101                 }
1102                 __mark_block_processed(rc, upper);
1103                 if (upper->level > 0) {
1104                         list_add(&upper->list, &cache->detached);
1105                         upper->detached = 1;
1106                 } else {
1107                         rb_erase(&upper->rb_node, &cache->rb_root);
1108                         free_backref_node(cache, upper);
1109                 }
1110         }
1111 out:
1112         btrfs_free_path(path1);
1113         btrfs_free_path(path2);
1114         if (err) {
1115                 while (!list_empty(&useless)) {
1116                         lower = list_entry(useless.next,
1117                                            struct backref_node, upper);
1118                         list_del_init(&lower->upper);
1119                 }
1120                 upper = node;
1121                 INIT_LIST_HEAD(&list);
1122                 while (upper) {
1123                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1124                                 list_splice_tail(&upper->upper, &list);
1125                                 free_backref_node(cache, upper);
1126                         }
1127
1128                         if (list_empty(&list))
1129                                 break;
1130
1131                         edge = list_entry(list.next, struct backref_edge,
1132                                           list[LOWER]);
1133                         list_del(&edge->list[LOWER]);
1134                         upper = edge->node[UPPER];
1135                         free_backref_edge(cache, edge);
1136                 }
1137                 return ERR_PTR(err);
1138         }
1139         BUG_ON(node && node->detached);
1140         return node;
1141 }
1142
1143 /*
1144  * helper to add backref node for the newly created snapshot.
1145  * the backref node is created by cloning backref node that
1146  * corresponds to root of source tree
1147  */
1148 static int clone_backref_node(struct btrfs_trans_handle *trans,
1149                               struct reloc_control *rc,
1150                               struct btrfs_root *src,
1151                               struct btrfs_root *dest)
1152 {
1153         struct btrfs_root *reloc_root = src->reloc_root;
1154         struct backref_cache *cache = &rc->backref_cache;
1155         struct backref_node *node = NULL;
1156         struct backref_node *new_node;
1157         struct backref_edge *edge;
1158         struct backref_edge *new_edge;
1159         struct rb_node *rb_node;
1160
1161         if (cache->last_trans > 0)
1162                 update_backref_cache(trans, cache);
1163
1164         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1165         if (rb_node) {
1166                 node = rb_entry(rb_node, struct backref_node, rb_node);
1167                 if (node->detached)
1168                         node = NULL;
1169                 else
1170                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1171         }
1172
1173         if (!node) {
1174                 rb_node = tree_search(&cache->rb_root,
1175                                       reloc_root->commit_root->start);
1176                 if (rb_node) {
1177                         node = rb_entry(rb_node, struct backref_node,
1178                                         rb_node);
1179                         BUG_ON(node->detached);
1180                 }
1181         }
1182
1183         if (!node)
1184                 return 0;
1185
1186         new_node = alloc_backref_node(cache);
1187         if (!new_node)
1188                 return -ENOMEM;
1189
1190         new_node->bytenr = dest->node->start;
1191         new_node->level = node->level;
1192         new_node->lowest = node->lowest;
1193         new_node->checked = 1;
1194         new_node->root = dest;
1195
1196         if (!node->lowest) {
1197                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1198                         new_edge = alloc_backref_edge(cache);
1199                         if (!new_edge)
1200                                 goto fail;
1201
1202                         new_edge->node[UPPER] = new_node;
1203                         new_edge->node[LOWER] = edge->node[LOWER];
1204                         list_add_tail(&new_edge->list[UPPER],
1205                                       &new_node->lower);
1206                 }
1207         } else {
1208                 list_add_tail(&new_node->lower, &cache->leaves);
1209         }
1210
1211         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1212                               &new_node->rb_node);
1213         if (rb_node)
1214                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1215
1216         if (!new_node->lowest) {
1217                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1218                         list_add_tail(&new_edge->list[LOWER],
1219                                       &new_edge->node[LOWER]->upper);
1220                 }
1221         }
1222         return 0;
1223 fail:
1224         while (!list_empty(&new_node->lower)) {
1225                 new_edge = list_entry(new_node->lower.next,
1226                                       struct backref_edge, list[UPPER]);
1227                 list_del(&new_edge->list[UPPER]);
1228                 free_backref_edge(cache, new_edge);
1229         }
1230         free_backref_node(cache, new_node);
1231         return -ENOMEM;
1232 }
1233
1234 /*
1235  * helper to add 'address of tree root -> reloc tree' mapping
1236  */
1237 static int __must_check __add_reloc_root(struct btrfs_root *root)
1238 {
1239         struct rb_node *rb_node;
1240         struct mapping_node *node;
1241         struct reloc_control *rc = root->fs_info->reloc_ctl;
1242
1243         node = kmalloc(sizeof(*node), GFP_NOFS);
1244         if (!node)
1245                 return -ENOMEM;
1246
1247         node->bytenr = root->node->start;
1248         node->data = root;
1249
1250         spin_lock(&rc->reloc_root_tree.lock);
1251         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1252                               node->bytenr, &node->rb_node);
1253         spin_unlock(&rc->reloc_root_tree.lock);
1254         if (rb_node) {
1255                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1256                             "for start=%llu while inserting into relocation "
1257                             "tree\n", node->bytenr);
1258                 kfree(node);
1259                 return -EEXIST;
1260         }
1261
1262         list_add_tail(&root->root_list, &rc->reloc_roots);
1263         return 0;
1264 }
1265
1266 /*
1267  * helper to update/delete the 'address of tree root -> reloc tree'
1268  * mapping
1269  */
1270 static int __update_reloc_root(struct btrfs_root *root, int del)
1271 {
1272         struct rb_node *rb_node;
1273         struct mapping_node *node = NULL;
1274         struct reloc_control *rc = root->fs_info->reloc_ctl;
1275
1276         spin_lock(&rc->reloc_root_tree.lock);
1277         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1278                               root->commit_root->start);
1279         if (rb_node) {
1280                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1281                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1282         }
1283         spin_unlock(&rc->reloc_root_tree.lock);
1284
1285         if (!node)
1286                 return 0;
1287         BUG_ON((struct btrfs_root *)node->data != root);
1288
1289         if (!del) {
1290                 spin_lock(&rc->reloc_root_tree.lock);
1291                 node->bytenr = root->node->start;
1292                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1293                                       node->bytenr, &node->rb_node);
1294                 spin_unlock(&rc->reloc_root_tree.lock);
1295                 if (rb_node)
1296                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1297         } else {
1298                 spin_lock(&root->fs_info->trans_lock);
1299                 list_del_init(&root->root_list);
1300                 spin_unlock(&root->fs_info->trans_lock);
1301                 kfree(node);
1302         }
1303         return 0;
1304 }
1305
1306 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1307                                         struct btrfs_root *root, u64 objectid)
1308 {
1309         struct btrfs_root *reloc_root;
1310         struct extent_buffer *eb;
1311         struct btrfs_root_item *root_item;
1312         struct btrfs_key root_key;
1313         u64 last_snap = 0;
1314         int ret;
1315
1316         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1317         BUG_ON(!root_item);
1318
1319         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1320         root_key.type = BTRFS_ROOT_ITEM_KEY;
1321         root_key.offset = objectid;
1322
1323         if (root->root_key.objectid == objectid) {
1324                 /* called by btrfs_init_reloc_root */
1325                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1326                                       BTRFS_TREE_RELOC_OBJECTID);
1327                 BUG_ON(ret);
1328
1329                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1330                 btrfs_set_root_last_snapshot(&root->root_item,
1331                                              trans->transid - 1);
1332         } else {
1333                 /*
1334                  * called by btrfs_reloc_post_snapshot_hook.
1335                  * the source tree is a reloc tree, all tree blocks
1336                  * modified after it was created have RELOC flag
1337                  * set in their headers. so it's OK to not update
1338                  * the 'last_snapshot'.
1339                  */
1340                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1341                                       BTRFS_TREE_RELOC_OBJECTID);
1342                 BUG_ON(ret);
1343         }
1344
1345         memcpy(root_item, &root->root_item, sizeof(*root_item));
1346         btrfs_set_root_bytenr(root_item, eb->start);
1347         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1348         btrfs_set_root_generation(root_item, trans->transid);
1349
1350         if (root->root_key.objectid == objectid) {
1351                 btrfs_set_root_refs(root_item, 0);
1352                 memset(&root_item->drop_progress, 0,
1353                        sizeof(struct btrfs_disk_key));
1354                 root_item->drop_level = 0;
1355                 /*
1356                  * abuse rtransid, it is safe because it is impossible to
1357                  * receive data into a relocation tree.
1358                  */
1359                 btrfs_set_root_rtransid(root_item, last_snap);
1360                 btrfs_set_root_otransid(root_item, trans->transid);
1361         }
1362
1363         btrfs_tree_unlock(eb);
1364         free_extent_buffer(eb);
1365
1366         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1367                                 &root_key, root_item);
1368         BUG_ON(ret);
1369         kfree(root_item);
1370
1371         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1372         BUG_ON(IS_ERR(reloc_root));
1373         reloc_root->last_trans = trans->transid;
1374         return reloc_root;
1375 }
1376
1377 /*
1378  * create reloc tree for a given fs tree. reloc tree is just a
1379  * snapshot of the fs tree with special root objectid.
1380  */
1381 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1382                           struct btrfs_root *root)
1383 {
1384         struct btrfs_root *reloc_root;
1385         struct reloc_control *rc = root->fs_info->reloc_ctl;
1386         int clear_rsv = 0;
1387         int ret;
1388
1389         if (root->reloc_root) {
1390                 reloc_root = root->reloc_root;
1391                 reloc_root->last_trans = trans->transid;
1392                 return 0;
1393         }
1394
1395         if (!rc || !rc->create_reloc_tree ||
1396             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1397                 return 0;
1398
1399         if (!trans->block_rsv) {
1400                 trans->block_rsv = rc->block_rsv;
1401                 clear_rsv = 1;
1402         }
1403         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1404         if (clear_rsv)
1405                 trans->block_rsv = NULL;
1406
1407         ret = __add_reloc_root(reloc_root);
1408         BUG_ON(ret < 0);
1409         root->reloc_root = reloc_root;
1410         return 0;
1411 }
1412
1413 /*
1414  * update root item of reloc tree
1415  */
1416 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1417                             struct btrfs_root *root)
1418 {
1419         struct btrfs_root *reloc_root;
1420         struct btrfs_root_item *root_item;
1421         int del = 0;
1422         int ret;
1423
1424         if (!root->reloc_root)
1425                 goto out;
1426
1427         reloc_root = root->reloc_root;
1428         root_item = &reloc_root->root_item;
1429
1430         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1431             btrfs_root_refs(root_item) == 0) {
1432                 root->reloc_root = NULL;
1433                 del = 1;
1434         }
1435
1436         __update_reloc_root(reloc_root, del);
1437
1438         if (reloc_root->commit_root != reloc_root->node) {
1439                 btrfs_set_root_node(root_item, reloc_root->node);
1440                 free_extent_buffer(reloc_root->commit_root);
1441                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1442         }
1443
1444         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1445                                 &reloc_root->root_key, root_item);
1446         BUG_ON(ret);
1447
1448 out:
1449         return 0;
1450 }
1451
1452 /*
1453  * helper to find first cached inode with inode number >= objectid
1454  * in a subvolume
1455  */
1456 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1457 {
1458         struct rb_node *node;
1459         struct rb_node *prev;
1460         struct btrfs_inode *entry;
1461         struct inode *inode;
1462
1463         spin_lock(&root->inode_lock);
1464 again:
1465         node = root->inode_tree.rb_node;
1466         prev = NULL;
1467         while (node) {
1468                 prev = node;
1469                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1470
1471                 if (objectid < btrfs_ino(&entry->vfs_inode))
1472                         node = node->rb_left;
1473                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1474                         node = node->rb_right;
1475                 else
1476                         break;
1477         }
1478         if (!node) {
1479                 while (prev) {
1480                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1481                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1482                                 node = prev;
1483                                 break;
1484                         }
1485                         prev = rb_next(prev);
1486                 }
1487         }
1488         while (node) {
1489                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1490                 inode = igrab(&entry->vfs_inode);
1491                 if (inode) {
1492                         spin_unlock(&root->inode_lock);
1493                         return inode;
1494                 }
1495
1496                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1497                 if (cond_resched_lock(&root->inode_lock))
1498                         goto again;
1499
1500                 node = rb_next(node);
1501         }
1502         spin_unlock(&root->inode_lock);
1503         return NULL;
1504 }
1505
1506 static int in_block_group(u64 bytenr,
1507                           struct btrfs_block_group_cache *block_group)
1508 {
1509         if (bytenr >= block_group->key.objectid &&
1510             bytenr < block_group->key.objectid + block_group->key.offset)
1511                 return 1;
1512         return 0;
1513 }
1514
1515 /*
1516  * get new location of data
1517  */
1518 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1519                             u64 bytenr, u64 num_bytes)
1520 {
1521         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1522         struct btrfs_path *path;
1523         struct btrfs_file_extent_item *fi;
1524         struct extent_buffer *leaf;
1525         int ret;
1526
1527         path = btrfs_alloc_path();
1528         if (!path)
1529                 return -ENOMEM;
1530
1531         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1532         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1533                                        bytenr, 0);
1534         if (ret < 0)
1535                 goto out;
1536         if (ret > 0) {
1537                 ret = -ENOENT;
1538                 goto out;
1539         }
1540
1541         leaf = path->nodes[0];
1542         fi = btrfs_item_ptr(leaf, path->slots[0],
1543                             struct btrfs_file_extent_item);
1544
1545         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1546                btrfs_file_extent_compression(leaf, fi) ||
1547                btrfs_file_extent_encryption(leaf, fi) ||
1548                btrfs_file_extent_other_encoding(leaf, fi));
1549
1550         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1551                 ret = -EINVAL;
1552                 goto out;
1553         }
1554
1555         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1556         ret = 0;
1557 out:
1558         btrfs_free_path(path);
1559         return ret;
1560 }
1561
1562 /*
1563  * update file extent items in the tree leaf to point to
1564  * the new locations.
1565  */
1566 static noinline_for_stack
1567 int replace_file_extents(struct btrfs_trans_handle *trans,
1568                          struct reloc_control *rc,
1569                          struct btrfs_root *root,
1570                          struct extent_buffer *leaf)
1571 {
1572         struct btrfs_key key;
1573         struct btrfs_file_extent_item *fi;
1574         struct inode *inode = NULL;
1575         u64 parent;
1576         u64 bytenr;
1577         u64 new_bytenr = 0;
1578         u64 num_bytes;
1579         u64 end;
1580         u32 nritems;
1581         u32 i;
1582         int ret = 0;
1583         int first = 1;
1584         int dirty = 0;
1585
1586         if (rc->stage != UPDATE_DATA_PTRS)
1587                 return 0;
1588
1589         /* reloc trees always use full backref */
1590         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1591                 parent = leaf->start;
1592         else
1593                 parent = 0;
1594
1595         nritems = btrfs_header_nritems(leaf);
1596         for (i = 0; i < nritems; i++) {
1597                 cond_resched();
1598                 btrfs_item_key_to_cpu(leaf, &key, i);
1599                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1600                         continue;
1601                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1602                 if (btrfs_file_extent_type(leaf, fi) ==
1603                     BTRFS_FILE_EXTENT_INLINE)
1604                         continue;
1605                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1606                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1607                 if (bytenr == 0)
1608                         continue;
1609                 if (!in_block_group(bytenr, rc->block_group))
1610                         continue;
1611
1612                 /*
1613                  * if we are modifying block in fs tree, wait for readpage
1614                  * to complete and drop the extent cache
1615                  */
1616                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1617                         if (first) {
1618                                 inode = find_next_inode(root, key.objectid);
1619                                 first = 0;
1620                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1621                                 btrfs_add_delayed_iput(inode);
1622                                 inode = find_next_inode(root, key.objectid);
1623                         }
1624                         if (inode && btrfs_ino(inode) == key.objectid) {
1625                                 end = key.offset +
1626                                       btrfs_file_extent_num_bytes(leaf, fi);
1627                                 WARN_ON(!IS_ALIGNED(key.offset,
1628                                                     root->sectorsize));
1629                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1630                                 end--;
1631                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1632                                                       key.offset, end);
1633                                 if (!ret)
1634                                         continue;
1635
1636                                 btrfs_drop_extent_cache(inode, key.offset, end,
1637                                                         1);
1638                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1639                                               key.offset, end);
1640                         }
1641                 }
1642
1643                 ret = get_new_location(rc->data_inode, &new_bytenr,
1644                                        bytenr, num_bytes);
1645                 if (ret) {
1646                         /*
1647                          * Don't have to abort since we've not changed anything
1648                          * in the file extent yet.
1649                          */
1650                         break;
1651                 }
1652
1653                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1654                 dirty = 1;
1655
1656                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1657                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1658                                            num_bytes, parent,
1659                                            btrfs_header_owner(leaf),
1660                                            key.objectid, key.offset, 1);
1661                 if (ret) {
1662                         btrfs_abort_transaction(trans, root, ret);
1663                         break;
1664                 }
1665
1666                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1667                                         parent, btrfs_header_owner(leaf),
1668                                         key.objectid, key.offset, 1);
1669                 if (ret) {
1670                         btrfs_abort_transaction(trans, root, ret);
1671                         break;
1672                 }
1673         }
1674         if (dirty)
1675                 btrfs_mark_buffer_dirty(leaf);
1676         if (inode)
1677                 btrfs_add_delayed_iput(inode);
1678         return ret;
1679 }
1680
1681 static noinline_for_stack
1682 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1683                      struct btrfs_path *path, int level)
1684 {
1685         struct btrfs_disk_key key1;
1686         struct btrfs_disk_key key2;
1687         btrfs_node_key(eb, &key1, slot);
1688         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1689         return memcmp(&key1, &key2, sizeof(key1));
1690 }
1691
1692 /*
1693  * try to replace tree blocks in fs tree with the new blocks
1694  * in reloc tree. tree blocks haven't been modified since the
1695  * reloc tree was create can be replaced.
1696  *
1697  * if a block was replaced, level of the block + 1 is returned.
1698  * if no block got replaced, 0 is returned. if there are other
1699  * errors, a negative error number is returned.
1700  */
1701 static noinline_for_stack
1702 int replace_path(struct btrfs_trans_handle *trans,
1703                  struct btrfs_root *dest, struct btrfs_root *src,
1704                  struct btrfs_path *path, struct btrfs_key *next_key,
1705                  int lowest_level, int max_level)
1706 {
1707         struct extent_buffer *eb;
1708         struct extent_buffer *parent;
1709         struct btrfs_key key;
1710         u64 old_bytenr;
1711         u64 new_bytenr;
1712         u64 old_ptr_gen;
1713         u64 new_ptr_gen;
1714         u64 last_snapshot;
1715         u32 blocksize;
1716         int cow = 0;
1717         int level;
1718         int ret;
1719         int slot;
1720
1721         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1722         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1723
1724         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1725 again:
1726         slot = path->slots[lowest_level];
1727         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1728
1729         eb = btrfs_lock_root_node(dest);
1730         btrfs_set_lock_blocking(eb);
1731         level = btrfs_header_level(eb);
1732
1733         if (level < lowest_level) {
1734                 btrfs_tree_unlock(eb);
1735                 free_extent_buffer(eb);
1736                 return 0;
1737         }
1738
1739         if (cow) {
1740                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1741                 BUG_ON(ret);
1742         }
1743         btrfs_set_lock_blocking(eb);
1744
1745         if (next_key) {
1746                 next_key->objectid = (u64)-1;
1747                 next_key->type = (u8)-1;
1748                 next_key->offset = (u64)-1;
1749         }
1750
1751         parent = eb;
1752         while (1) {
1753                 level = btrfs_header_level(parent);
1754                 BUG_ON(level < lowest_level);
1755
1756                 ret = btrfs_bin_search(parent, &key, level, &slot);
1757                 if (ret && slot > 0)
1758                         slot--;
1759
1760                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1761                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1762
1763                 old_bytenr = btrfs_node_blockptr(parent, slot);
1764                 blocksize = btrfs_level_size(dest, level - 1);
1765                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1766
1767                 if (level <= max_level) {
1768                         eb = path->nodes[level];
1769                         new_bytenr = btrfs_node_blockptr(eb,
1770                                                         path->slots[level]);
1771                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1772                                                         path->slots[level]);
1773                 } else {
1774                         new_bytenr = 0;
1775                         new_ptr_gen = 0;
1776                 }
1777
1778                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1779                         WARN_ON(1);
1780                         ret = level;
1781                         break;
1782                 }
1783
1784                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1785                     memcmp_node_keys(parent, slot, path, level)) {
1786                         if (level <= lowest_level) {
1787                                 ret = 0;
1788                                 break;
1789                         }
1790
1791                         eb = read_tree_block(dest, old_bytenr, blocksize,
1792                                              old_ptr_gen);
1793                         if (!eb || !extent_buffer_uptodate(eb)) {
1794                                 ret = (!eb) ? -ENOMEM : -EIO;
1795                                 free_extent_buffer(eb);
1796                                 break;
1797                         }
1798                         btrfs_tree_lock(eb);
1799                         if (cow) {
1800                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1801                                                       slot, &eb);
1802                                 BUG_ON(ret);
1803                         }
1804                         btrfs_set_lock_blocking(eb);
1805
1806                         btrfs_tree_unlock(parent);
1807                         free_extent_buffer(parent);
1808
1809                         parent = eb;
1810                         continue;
1811                 }
1812
1813                 if (!cow) {
1814                         btrfs_tree_unlock(parent);
1815                         free_extent_buffer(parent);
1816                         cow = 1;
1817                         goto again;
1818                 }
1819
1820                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1821                                       path->slots[level]);
1822                 btrfs_release_path(path);
1823
1824                 path->lowest_level = level;
1825                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1826                 path->lowest_level = 0;
1827                 BUG_ON(ret);
1828
1829                 /*
1830                  * swap blocks in fs tree and reloc tree.
1831                  */
1832                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1833                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1834                 btrfs_mark_buffer_dirty(parent);
1835
1836                 btrfs_set_node_blockptr(path->nodes[level],
1837                                         path->slots[level], old_bytenr);
1838                 btrfs_set_node_ptr_generation(path->nodes[level],
1839                                               path->slots[level], old_ptr_gen);
1840                 btrfs_mark_buffer_dirty(path->nodes[level]);
1841
1842                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1843                                         path->nodes[level]->start,
1844                                         src->root_key.objectid, level - 1, 0,
1845                                         1);
1846                 BUG_ON(ret);
1847                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1848                                         0, dest->root_key.objectid, level - 1,
1849                                         0, 1);
1850                 BUG_ON(ret);
1851
1852                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1853                                         path->nodes[level]->start,
1854                                         src->root_key.objectid, level - 1, 0,
1855                                         1);
1856                 BUG_ON(ret);
1857
1858                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1859                                         0, dest->root_key.objectid, level - 1,
1860                                         0, 1);
1861                 BUG_ON(ret);
1862
1863                 btrfs_unlock_up_safe(path, 0);
1864
1865                 ret = level;
1866                 break;
1867         }
1868         btrfs_tree_unlock(parent);
1869         free_extent_buffer(parent);
1870         return ret;
1871 }
1872
1873 /*
1874  * helper to find next relocated block in reloc tree
1875  */
1876 static noinline_for_stack
1877 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1878                        int *level)
1879 {
1880         struct extent_buffer *eb;
1881         int i;
1882         u64 last_snapshot;
1883         u32 nritems;
1884
1885         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1886
1887         for (i = 0; i < *level; i++) {
1888                 free_extent_buffer(path->nodes[i]);
1889                 path->nodes[i] = NULL;
1890         }
1891
1892         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1893                 eb = path->nodes[i];
1894                 nritems = btrfs_header_nritems(eb);
1895                 while (path->slots[i] + 1 < nritems) {
1896                         path->slots[i]++;
1897                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1898                             last_snapshot)
1899                                 continue;
1900
1901                         *level = i;
1902                         return 0;
1903                 }
1904                 free_extent_buffer(path->nodes[i]);
1905                 path->nodes[i] = NULL;
1906         }
1907         return 1;
1908 }
1909
1910 /*
1911  * walk down reloc tree to find relocated block of lowest level
1912  */
1913 static noinline_for_stack
1914 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1915                          int *level)
1916 {
1917         struct extent_buffer *eb = NULL;
1918         int i;
1919         u64 bytenr;
1920         u64 ptr_gen = 0;
1921         u64 last_snapshot;
1922         u32 blocksize;
1923         u32 nritems;
1924
1925         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1926
1927         for (i = *level; i > 0; i--) {
1928                 eb = path->nodes[i];
1929                 nritems = btrfs_header_nritems(eb);
1930                 while (path->slots[i] < nritems) {
1931                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1932                         if (ptr_gen > last_snapshot)
1933                                 break;
1934                         path->slots[i]++;
1935                 }
1936                 if (path->slots[i] >= nritems) {
1937                         if (i == *level)
1938                                 break;
1939                         *level = i + 1;
1940                         return 0;
1941                 }
1942                 if (i == 1) {
1943                         *level = i;
1944                         return 0;
1945                 }
1946
1947                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1948                 blocksize = btrfs_level_size(root, i - 1);
1949                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1950                 if (!eb || !extent_buffer_uptodate(eb)) {
1951                         free_extent_buffer(eb);
1952                         return -EIO;
1953                 }
1954                 BUG_ON(btrfs_header_level(eb) != i - 1);
1955                 path->nodes[i - 1] = eb;
1956                 path->slots[i - 1] = 0;
1957         }
1958         return 1;
1959 }
1960
1961 /*
1962  * invalidate extent cache for file extents whose key in range of
1963  * [min_key, max_key)
1964  */
1965 static int invalidate_extent_cache(struct btrfs_root *root,
1966                                    struct btrfs_key *min_key,
1967                                    struct btrfs_key *max_key)
1968 {
1969         struct inode *inode = NULL;
1970         u64 objectid;
1971         u64 start, end;
1972         u64 ino;
1973
1974         objectid = min_key->objectid;
1975         while (1) {
1976                 cond_resched();
1977                 iput(inode);
1978
1979                 if (objectid > max_key->objectid)
1980                         break;
1981
1982                 inode = find_next_inode(root, objectid);
1983                 if (!inode)
1984                         break;
1985                 ino = btrfs_ino(inode);
1986
1987                 if (ino > max_key->objectid) {
1988                         iput(inode);
1989                         break;
1990                 }
1991
1992                 objectid = ino + 1;
1993                 if (!S_ISREG(inode->i_mode))
1994                         continue;
1995
1996                 if (unlikely(min_key->objectid == ino)) {
1997                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1998                                 continue;
1999                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2000                                 start = 0;
2001                         else {
2002                                 start = min_key->offset;
2003                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2004                         }
2005                 } else {
2006                         start = 0;
2007                 }
2008
2009                 if (unlikely(max_key->objectid == ino)) {
2010                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2011                                 continue;
2012                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2013                                 end = (u64)-1;
2014                         } else {
2015                                 if (max_key->offset == 0)
2016                                         continue;
2017                                 end = max_key->offset;
2018                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2019                                 end--;
2020                         }
2021                 } else {
2022                         end = (u64)-1;
2023                 }
2024
2025                 /* the lock_extent waits for readpage to complete */
2026                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2027                 btrfs_drop_extent_cache(inode, start, end, 1);
2028                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2029         }
2030         return 0;
2031 }
2032
2033 static int find_next_key(struct btrfs_path *path, int level,
2034                          struct btrfs_key *key)
2035
2036 {
2037         while (level < BTRFS_MAX_LEVEL) {
2038                 if (!path->nodes[level])
2039                         break;
2040                 if (path->slots[level] + 1 <
2041                     btrfs_header_nritems(path->nodes[level])) {
2042                         btrfs_node_key_to_cpu(path->nodes[level], key,
2043                                               path->slots[level] + 1);
2044                         return 0;
2045                 }
2046                 level++;
2047         }
2048         return 1;
2049 }
2050
2051 /*
2052  * merge the relocated tree blocks in reloc tree with corresponding
2053  * fs tree.
2054  */
2055 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2056                                                struct btrfs_root *root)
2057 {
2058         LIST_HEAD(inode_list);
2059         struct btrfs_key key;
2060         struct btrfs_key next_key;
2061         struct btrfs_trans_handle *trans;
2062         struct btrfs_root *reloc_root;
2063         struct btrfs_root_item *root_item;
2064         struct btrfs_path *path;
2065         struct extent_buffer *leaf;
2066         int level;
2067         int max_level;
2068         int replaced = 0;
2069         int ret;
2070         int err = 0;
2071         u32 min_reserved;
2072
2073         path = btrfs_alloc_path();
2074         if (!path)
2075                 return -ENOMEM;
2076         path->reada = 1;
2077
2078         reloc_root = root->reloc_root;
2079         root_item = &reloc_root->root_item;
2080
2081         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2082                 level = btrfs_root_level(root_item);
2083                 extent_buffer_get(reloc_root->node);
2084                 path->nodes[level] = reloc_root->node;
2085                 path->slots[level] = 0;
2086         } else {
2087                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2088
2089                 level = root_item->drop_level;
2090                 BUG_ON(level == 0);
2091                 path->lowest_level = level;
2092                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2093                 path->lowest_level = 0;
2094                 if (ret < 0) {
2095                         btrfs_free_path(path);
2096                         return ret;
2097                 }
2098
2099                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2100                                       path->slots[level]);
2101                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2102
2103                 btrfs_unlock_up_safe(path, 0);
2104         }
2105
2106         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2107         memset(&next_key, 0, sizeof(next_key));
2108
2109         while (1) {
2110                 trans = btrfs_start_transaction(root, 0);
2111                 BUG_ON(IS_ERR(trans));
2112                 trans->block_rsv = rc->block_rsv;
2113
2114                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2115                                              BTRFS_RESERVE_FLUSH_ALL);
2116                 if (ret) {
2117                         BUG_ON(ret != -EAGAIN);
2118                         ret = btrfs_commit_transaction(trans, root);
2119                         BUG_ON(ret);
2120                         continue;
2121                 }
2122
2123                 replaced = 0;
2124                 max_level = level;
2125
2126                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2127                 if (ret < 0) {
2128                         err = ret;
2129                         goto out;
2130                 }
2131                 if (ret > 0)
2132                         break;
2133
2134                 if (!find_next_key(path, level, &key) &&
2135                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2136                         ret = 0;
2137                 } else {
2138                         ret = replace_path(trans, root, reloc_root, path,
2139                                            &next_key, level, max_level);
2140                 }
2141                 if (ret < 0) {
2142                         err = ret;
2143                         goto out;
2144                 }
2145
2146                 if (ret > 0) {
2147                         level = ret;
2148                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2149                                               path->slots[level]);
2150                         replaced = 1;
2151                 }
2152
2153                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2154                 if (ret > 0)
2155                         break;
2156
2157                 BUG_ON(level == 0);
2158                 /*
2159                  * save the merging progress in the drop_progress.
2160                  * this is OK since root refs == 1 in this case.
2161                  */
2162                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2163                                path->slots[level]);
2164                 root_item->drop_level = level;
2165
2166                 btrfs_end_transaction_throttle(trans, root);
2167
2168                 btrfs_btree_balance_dirty(root);
2169
2170                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2171                         invalidate_extent_cache(root, &key, &next_key);
2172         }
2173
2174         /*
2175          * handle the case only one block in the fs tree need to be
2176          * relocated and the block is tree root.
2177          */
2178         leaf = btrfs_lock_root_node(root);
2179         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2180         btrfs_tree_unlock(leaf);
2181         free_extent_buffer(leaf);
2182         if (ret < 0)
2183                 err = ret;
2184 out:
2185         btrfs_free_path(path);
2186
2187         if (err == 0) {
2188                 memset(&root_item->drop_progress, 0,
2189                        sizeof(root_item->drop_progress));
2190                 root_item->drop_level = 0;
2191                 btrfs_set_root_refs(root_item, 0);
2192                 btrfs_update_reloc_root(trans, root);
2193         }
2194
2195         btrfs_end_transaction_throttle(trans, root);
2196
2197         btrfs_btree_balance_dirty(root);
2198
2199         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2200                 invalidate_extent_cache(root, &key, &next_key);
2201
2202         return err;
2203 }
2204
2205 static noinline_for_stack
2206 int prepare_to_merge(struct reloc_control *rc, int err)
2207 {
2208         struct btrfs_root *root = rc->extent_root;
2209         struct btrfs_root *reloc_root;
2210         struct btrfs_trans_handle *trans;
2211         LIST_HEAD(reloc_roots);
2212         u64 num_bytes = 0;
2213         int ret;
2214
2215         mutex_lock(&root->fs_info->reloc_mutex);
2216         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2217         rc->merging_rsv_size += rc->nodes_relocated * 2;
2218         mutex_unlock(&root->fs_info->reloc_mutex);
2219
2220 again:
2221         if (!err) {
2222                 num_bytes = rc->merging_rsv_size;
2223                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2224                                           BTRFS_RESERVE_FLUSH_ALL);
2225                 if (ret)
2226                         err = ret;
2227         }
2228
2229         trans = btrfs_join_transaction(rc->extent_root);
2230         if (IS_ERR(trans)) {
2231                 if (!err)
2232                         btrfs_block_rsv_release(rc->extent_root,
2233                                                 rc->block_rsv, num_bytes);
2234                 return PTR_ERR(trans);
2235         }
2236
2237         if (!err) {
2238                 if (num_bytes != rc->merging_rsv_size) {
2239                         btrfs_end_transaction(trans, rc->extent_root);
2240                         btrfs_block_rsv_release(rc->extent_root,
2241                                                 rc->block_rsv, num_bytes);
2242                         goto again;
2243                 }
2244         }
2245
2246         rc->merge_reloc_tree = 1;
2247
2248         while (!list_empty(&rc->reloc_roots)) {
2249                 reloc_root = list_entry(rc->reloc_roots.next,
2250                                         struct btrfs_root, root_list);
2251                 list_del_init(&reloc_root->root_list);
2252
2253                 root = read_fs_root(reloc_root->fs_info,
2254                                     reloc_root->root_key.offset);
2255                 BUG_ON(IS_ERR(root));
2256                 BUG_ON(root->reloc_root != reloc_root);
2257
2258                 /*
2259                  * set reference count to 1, so btrfs_recover_relocation
2260                  * knows it should resumes merging
2261                  */
2262                 if (!err)
2263                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2264                 btrfs_update_reloc_root(trans, root);
2265
2266                 list_add(&reloc_root->root_list, &reloc_roots);
2267         }
2268
2269         list_splice(&reloc_roots, &rc->reloc_roots);
2270
2271         if (!err)
2272                 btrfs_commit_transaction(trans, rc->extent_root);
2273         else
2274                 btrfs_end_transaction(trans, rc->extent_root);
2275         return err;
2276 }
2277
2278 static noinline_for_stack
2279 void free_reloc_roots(struct list_head *list)
2280 {
2281         struct btrfs_root *reloc_root;
2282
2283         while (!list_empty(list)) {
2284                 reloc_root = list_entry(list->next, struct btrfs_root,
2285                                         root_list);
2286                 __update_reloc_root(reloc_root, 1);
2287                 free_extent_buffer(reloc_root->node);
2288                 free_extent_buffer(reloc_root->commit_root);
2289                 kfree(reloc_root);
2290         }
2291 }
2292
2293 static noinline_for_stack
2294 int merge_reloc_roots(struct reloc_control *rc)
2295 {
2296         struct btrfs_trans_handle *trans;
2297         struct btrfs_root *root;
2298         struct btrfs_root *reloc_root;
2299         u64 last_snap;
2300         u64 otransid;
2301         u64 objectid;
2302         LIST_HEAD(reloc_roots);
2303         int found = 0;
2304         int ret = 0;
2305 again:
2306         root = rc->extent_root;
2307
2308         /*
2309          * this serializes us with btrfs_record_root_in_transaction,
2310          * we have to make sure nobody is in the middle of
2311          * adding their roots to the list while we are
2312          * doing this splice
2313          */
2314         mutex_lock(&root->fs_info->reloc_mutex);
2315         list_splice_init(&rc->reloc_roots, &reloc_roots);
2316         mutex_unlock(&root->fs_info->reloc_mutex);
2317
2318         while (!list_empty(&reloc_roots)) {
2319                 found = 1;
2320                 reloc_root = list_entry(reloc_roots.next,
2321                                         struct btrfs_root, root_list);
2322
2323                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2324                         root = read_fs_root(reloc_root->fs_info,
2325                                             reloc_root->root_key.offset);
2326                         BUG_ON(IS_ERR(root));
2327                         BUG_ON(root->reloc_root != reloc_root);
2328
2329                         ret = merge_reloc_root(rc, root);
2330                         if (ret) {
2331                                 __update_reloc_root(reloc_root, 1);
2332                                 free_extent_buffer(reloc_root->node);
2333                                 free_extent_buffer(reloc_root->commit_root);
2334                                 kfree(reloc_root);
2335                                 goto out;
2336                         }
2337                 } else {
2338                         list_del_init(&reloc_root->root_list);
2339                 }
2340
2341                 /*
2342                  * we keep the old last snapshod transid in rtranid when we
2343                  * created the relocation tree.
2344                  */
2345                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2346                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2347                 objectid = reloc_root->root_key.offset;
2348
2349                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2350                 if (ret < 0) {
2351                         if (list_empty(&reloc_root->root_list))
2352                                 list_add_tail(&reloc_root->root_list,
2353                                               &reloc_roots);
2354                         goto out;
2355                 } else if (!ret) {
2356                         /*
2357                          * recover the last snapshot tranid to avoid
2358                          * the space balance break NOCOW.
2359                          */
2360                         root = read_fs_root(rc->extent_root->fs_info,
2361                                             objectid);
2362                         if (IS_ERR(root))
2363                                 continue;
2364
2365                         trans = btrfs_join_transaction(root);
2366                         BUG_ON(IS_ERR(trans));
2367
2368                         /* Check if the fs/file tree was snapshoted or not. */
2369                         if (btrfs_root_last_snapshot(&root->root_item) ==
2370                             otransid - 1)
2371                                 btrfs_set_root_last_snapshot(&root->root_item,
2372                                                              last_snap);
2373
2374                         btrfs_end_transaction(trans, root);
2375                 }
2376         }
2377
2378         if (found) {
2379                 found = 0;
2380                 goto again;
2381         }
2382 out:
2383         if (ret) {
2384                 btrfs_std_error(root->fs_info, ret);
2385                 if (!list_empty(&reloc_roots))
2386                         free_reloc_roots(&reloc_roots);
2387         }
2388
2389         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2390         return ret;
2391 }
2392
2393 static void free_block_list(struct rb_root *blocks)
2394 {
2395         struct tree_block *block;
2396         struct rb_node *rb_node;
2397         while ((rb_node = rb_first(blocks))) {
2398                 block = rb_entry(rb_node, struct tree_block, rb_node);
2399                 rb_erase(rb_node, blocks);
2400                 kfree(block);
2401         }
2402 }
2403
2404 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2405                                       struct btrfs_root *reloc_root)
2406 {
2407         struct btrfs_root *root;
2408
2409         if (reloc_root->last_trans == trans->transid)
2410                 return 0;
2411
2412         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2413         BUG_ON(IS_ERR(root));
2414         BUG_ON(root->reloc_root != reloc_root);
2415
2416         return btrfs_record_root_in_trans(trans, root);
2417 }
2418
2419 static noinline_for_stack
2420 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2421                                      struct reloc_control *rc,
2422                                      struct backref_node *node,
2423                                      struct backref_edge *edges[], int *nr)
2424 {
2425         struct backref_node *next;
2426         struct btrfs_root *root;
2427         int index = 0;
2428
2429         next = node;
2430         while (1) {
2431                 cond_resched();
2432                 next = walk_up_backref(next, edges, &index);
2433                 root = next->root;
2434                 BUG_ON(!root);
2435                 BUG_ON(!root->ref_cows);
2436
2437                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2438                         record_reloc_root_in_trans(trans, root);
2439                         break;
2440                 }
2441
2442                 btrfs_record_root_in_trans(trans, root);
2443                 root = root->reloc_root;
2444
2445                 if (next->new_bytenr != root->node->start) {
2446                         BUG_ON(next->new_bytenr);
2447                         BUG_ON(!list_empty(&next->list));
2448                         next->new_bytenr = root->node->start;
2449                         next->root = root;
2450                         list_add_tail(&next->list,
2451                                       &rc->backref_cache.changed);
2452                         __mark_block_processed(rc, next);
2453                         break;
2454                 }
2455
2456                 WARN_ON(1);
2457                 root = NULL;
2458                 next = walk_down_backref(edges, &index);
2459                 if (!next || next->level <= node->level)
2460                         break;
2461         }
2462         if (!root)
2463                 return NULL;
2464
2465         *nr = index;
2466         next = node;
2467         /* setup backref node path for btrfs_reloc_cow_block */
2468         while (1) {
2469                 rc->backref_cache.path[next->level] = next;
2470                 if (--index < 0)
2471                         break;
2472                 next = edges[index]->node[UPPER];
2473         }
2474         return root;
2475 }
2476
2477 /*
2478  * select a tree root for relocation. return NULL if the block
2479  * is reference counted. we should use do_relocation() in this
2480  * case. return a tree root pointer if the block isn't reference
2481  * counted. return -ENOENT if the block is root of reloc tree.
2482  */
2483 static noinline_for_stack
2484 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2485                                    struct backref_node *node)
2486 {
2487         struct backref_node *next;
2488         struct btrfs_root *root;
2489         struct btrfs_root *fs_root = NULL;
2490         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2491         int index = 0;
2492
2493         next = node;
2494         while (1) {
2495                 cond_resched();
2496                 next = walk_up_backref(next, edges, &index);
2497                 root = next->root;
2498                 BUG_ON(!root);
2499
2500                 /* no other choice for non-references counted tree */
2501                 if (!root->ref_cows)
2502                         return root;
2503
2504                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2505                         fs_root = root;
2506
2507                 if (next != node)
2508                         return NULL;
2509
2510                 next = walk_down_backref(edges, &index);
2511                 if (!next || next->level <= node->level)
2512                         break;
2513         }
2514
2515         if (!fs_root)
2516                 return ERR_PTR(-ENOENT);
2517         return fs_root;
2518 }
2519
2520 static noinline_for_stack
2521 u64 calcu_metadata_size(struct reloc_control *rc,
2522                         struct backref_node *node, int reserve)
2523 {
2524         struct backref_node *next = node;
2525         struct backref_edge *edge;
2526         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2527         u64 num_bytes = 0;
2528         int index = 0;
2529
2530         BUG_ON(reserve && node->processed);
2531
2532         while (next) {
2533                 cond_resched();
2534                 while (1) {
2535                         if (next->processed && (reserve || next != node))
2536                                 break;
2537
2538                         num_bytes += btrfs_level_size(rc->extent_root,
2539                                                       next->level);
2540
2541                         if (list_empty(&next->upper))
2542                                 break;
2543
2544                         edge = list_entry(next->upper.next,
2545                                           struct backref_edge, list[LOWER]);
2546                         edges[index++] = edge;
2547                         next = edge->node[UPPER];
2548                 }
2549                 next = walk_down_backref(edges, &index);
2550         }
2551         return num_bytes;
2552 }
2553
2554 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2555                                   struct reloc_control *rc,
2556                                   struct backref_node *node)
2557 {
2558         struct btrfs_root *root = rc->extent_root;
2559         u64 num_bytes;
2560         int ret;
2561
2562         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2563
2564         trans->block_rsv = rc->block_rsv;
2565         ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2566                                   BTRFS_RESERVE_FLUSH_ALL);
2567         if (ret) {
2568                 if (ret == -EAGAIN)
2569                         rc->commit_transaction = 1;
2570                 return ret;
2571         }
2572
2573         return 0;
2574 }
2575
2576 static void release_metadata_space(struct reloc_control *rc,
2577                                    struct backref_node *node)
2578 {
2579         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2580         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2581 }
2582
2583 /*
2584  * relocate a block tree, and then update pointers in upper level
2585  * blocks that reference the block to point to the new location.
2586  *
2587  * if called by link_to_upper, the block has already been relocated.
2588  * in that case this function just updates pointers.
2589  */
2590 static int do_relocation(struct btrfs_trans_handle *trans,
2591                          struct reloc_control *rc,
2592                          struct backref_node *node,
2593                          struct btrfs_key *key,
2594                          struct btrfs_path *path, int lowest)
2595 {
2596         struct backref_node *upper;
2597         struct backref_edge *edge;
2598         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2599         struct btrfs_root *root;
2600         struct extent_buffer *eb;
2601         u32 blocksize;
2602         u64 bytenr;
2603         u64 generation;
2604         int nr;
2605         int slot;
2606         int ret;
2607         int err = 0;
2608
2609         BUG_ON(lowest && node->eb);
2610
2611         path->lowest_level = node->level + 1;
2612         rc->backref_cache.path[node->level] = node;
2613         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2614                 cond_resched();
2615
2616                 upper = edge->node[UPPER];
2617                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2618                 BUG_ON(!root);
2619
2620                 if (upper->eb && !upper->locked) {
2621                         if (!lowest) {
2622                                 ret = btrfs_bin_search(upper->eb, key,
2623                                                        upper->level, &slot);
2624                                 BUG_ON(ret);
2625                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2626                                 if (node->eb->start == bytenr)
2627                                         goto next;
2628                         }
2629                         drop_node_buffer(upper);
2630                 }
2631
2632                 if (!upper->eb) {
2633                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2634                         if (ret < 0) {
2635                                 err = ret;
2636                                 break;
2637                         }
2638                         BUG_ON(ret > 0);
2639
2640                         if (!upper->eb) {
2641                                 upper->eb = path->nodes[upper->level];
2642                                 path->nodes[upper->level] = NULL;
2643                         } else {
2644                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2645                         }
2646
2647                         upper->locked = 1;
2648                         path->locks[upper->level] = 0;
2649
2650                         slot = path->slots[upper->level];
2651                         btrfs_release_path(path);
2652                 } else {
2653                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2654                                                &slot);
2655                         BUG_ON(ret);
2656                 }
2657
2658                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2659                 if (lowest) {
2660                         BUG_ON(bytenr != node->bytenr);
2661                 } else {
2662                         if (node->eb->start == bytenr)
2663                                 goto next;
2664                 }
2665
2666                 blocksize = btrfs_level_size(root, node->level);
2667                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2668                 eb = read_tree_block(root, bytenr, blocksize, generation);
2669                 if (!eb || !extent_buffer_uptodate(eb)) {
2670                         free_extent_buffer(eb);
2671                         err = -EIO;
2672                         goto next;
2673                 }
2674                 btrfs_tree_lock(eb);
2675                 btrfs_set_lock_blocking(eb);
2676
2677                 if (!node->eb) {
2678                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2679                                               slot, &eb);
2680                         btrfs_tree_unlock(eb);
2681                         free_extent_buffer(eb);
2682                         if (ret < 0) {
2683                                 err = ret;
2684                                 goto next;
2685                         }
2686                         BUG_ON(node->eb != eb);
2687                 } else {
2688                         btrfs_set_node_blockptr(upper->eb, slot,
2689                                                 node->eb->start);
2690                         btrfs_set_node_ptr_generation(upper->eb, slot,
2691                                                       trans->transid);
2692                         btrfs_mark_buffer_dirty(upper->eb);
2693
2694                         ret = btrfs_inc_extent_ref(trans, root,
2695                                                 node->eb->start, blocksize,
2696                                                 upper->eb->start,
2697                                                 btrfs_header_owner(upper->eb),
2698                                                 node->level, 0, 1);
2699                         BUG_ON(ret);
2700
2701                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2702                         BUG_ON(ret);
2703                 }
2704 next:
2705                 if (!upper->pending)
2706                         drop_node_buffer(upper);
2707                 else
2708                         unlock_node_buffer(upper);
2709                 if (err)
2710                         break;
2711         }
2712
2713         if (!err && node->pending) {
2714                 drop_node_buffer(node);
2715                 list_move_tail(&node->list, &rc->backref_cache.changed);
2716                 node->pending = 0;
2717         }
2718
2719         path->lowest_level = 0;
2720         BUG_ON(err == -ENOSPC);
2721         return err;
2722 }
2723
2724 static int link_to_upper(struct btrfs_trans_handle *trans,
2725                          struct reloc_control *rc,
2726                          struct backref_node *node,
2727                          struct btrfs_path *path)
2728 {
2729         struct btrfs_key key;
2730
2731         btrfs_node_key_to_cpu(node->eb, &key, 0);
2732         return do_relocation(trans, rc, node, &key, path, 0);
2733 }
2734
2735 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2736                                 struct reloc_control *rc,
2737                                 struct btrfs_path *path, int err)
2738 {
2739         LIST_HEAD(list);
2740         struct backref_cache *cache = &rc->backref_cache;
2741         struct backref_node *node;
2742         int level;
2743         int ret;
2744
2745         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2746                 while (!list_empty(&cache->pending[level])) {
2747                         node = list_entry(cache->pending[level].next,
2748                                           struct backref_node, list);
2749                         list_move_tail(&node->list, &list);
2750                         BUG_ON(!node->pending);
2751
2752                         if (!err) {
2753                                 ret = link_to_upper(trans, rc, node, path);
2754                                 if (ret < 0)
2755                                         err = ret;
2756                         }
2757                 }
2758                 list_splice_init(&list, &cache->pending[level]);
2759         }
2760         return err;
2761 }
2762
2763 static void mark_block_processed(struct reloc_control *rc,
2764                                  u64 bytenr, u32 blocksize)
2765 {
2766         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2767                         EXTENT_DIRTY, GFP_NOFS);
2768 }
2769
2770 static void __mark_block_processed(struct reloc_control *rc,
2771                                    struct backref_node *node)
2772 {
2773         u32 blocksize;
2774         if (node->level == 0 ||
2775             in_block_group(node->bytenr, rc->block_group)) {
2776                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2777                 mark_block_processed(rc, node->bytenr, blocksize);
2778         }
2779         node->processed = 1;
2780 }
2781
2782 /*
2783  * mark a block and all blocks directly/indirectly reference the block
2784  * as processed.
2785  */
2786 static void update_processed_blocks(struct reloc_control *rc,
2787                                     struct backref_node *node)
2788 {
2789         struct backref_node *next = node;
2790         struct backref_edge *edge;
2791         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2792         int index = 0;
2793
2794         while (next) {
2795                 cond_resched();
2796                 while (1) {
2797                         if (next->processed)
2798                                 break;
2799
2800                         __mark_block_processed(rc, next);
2801
2802                         if (list_empty(&next->upper))
2803                                 break;
2804
2805                         edge = list_entry(next->upper.next,
2806                                           struct backref_edge, list[LOWER]);
2807                         edges[index++] = edge;
2808                         next = edge->node[UPPER];
2809                 }
2810                 next = walk_down_backref(edges, &index);
2811         }
2812 }
2813
2814 static int tree_block_processed(u64 bytenr, u32 blocksize,
2815                                 struct reloc_control *rc)
2816 {
2817         if (test_range_bit(&rc->processed_blocks, bytenr,
2818                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2819                 return 1;
2820         return 0;
2821 }
2822
2823 static int get_tree_block_key(struct reloc_control *rc,
2824                               struct tree_block *block)
2825 {
2826         struct extent_buffer *eb;
2827
2828         BUG_ON(block->key_ready);
2829         eb = read_tree_block(rc->extent_root, block->bytenr,
2830                              block->key.objectid, block->key.offset);
2831         if (!eb || !extent_buffer_uptodate(eb)) {
2832                 free_extent_buffer(eb);
2833                 return -EIO;
2834         }
2835         WARN_ON(btrfs_header_level(eb) != block->level);
2836         if (block->level == 0)
2837                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2838         else
2839                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2840         free_extent_buffer(eb);
2841         block->key_ready = 1;
2842         return 0;
2843 }
2844
2845 static int reada_tree_block(struct reloc_control *rc,
2846                             struct tree_block *block)
2847 {
2848         BUG_ON(block->key_ready);
2849         if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2850                 readahead_tree_block(rc->extent_root, block->bytenr,
2851                                      block->key.objectid,
2852                                      rc->extent_root->leafsize);
2853         else
2854                 readahead_tree_block(rc->extent_root, block->bytenr,
2855                                      block->key.objectid, block->key.offset);
2856         return 0;
2857 }
2858
2859 /*
2860  * helper function to relocate a tree block
2861  */
2862 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2863                                 struct reloc_control *rc,
2864                                 struct backref_node *node,
2865                                 struct btrfs_key *key,
2866                                 struct btrfs_path *path)
2867 {
2868         struct btrfs_root *root;
2869         int release = 0;
2870         int ret = 0;
2871
2872         if (!node)
2873                 return 0;
2874
2875         BUG_ON(node->processed);
2876         root = select_one_root(trans, node);
2877         if (root == ERR_PTR(-ENOENT)) {
2878                 update_processed_blocks(rc, node);
2879                 goto out;
2880         }
2881
2882         if (!root || root->ref_cows) {
2883                 ret = reserve_metadata_space(trans, rc, node);
2884                 if (ret)
2885                         goto out;
2886                 release = 1;
2887         }
2888
2889         if (root) {
2890                 if (root->ref_cows) {
2891                         BUG_ON(node->new_bytenr);
2892                         BUG_ON(!list_empty(&node->list));
2893                         btrfs_record_root_in_trans(trans, root);
2894                         root = root->reloc_root;
2895                         node->new_bytenr = root->node->start;
2896                         node->root = root;
2897                         list_add_tail(&node->list, &rc->backref_cache.changed);
2898                 } else {
2899                         path->lowest_level = node->level;
2900                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2901                         btrfs_release_path(path);
2902                         if (ret > 0)
2903                                 ret = 0;
2904                 }
2905                 if (!ret)
2906                         update_processed_blocks(rc, node);
2907         } else {
2908                 ret = do_relocation(trans, rc, node, key, path, 1);
2909         }
2910 out:
2911         if (ret || node->level == 0 || node->cowonly) {
2912                 if (release)
2913                         release_metadata_space(rc, node);
2914                 remove_backref_node(&rc->backref_cache, node);
2915         }
2916         return ret;
2917 }
2918
2919 /*
2920  * relocate a list of blocks
2921  */
2922 static noinline_for_stack
2923 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2924                          struct reloc_control *rc, struct rb_root *blocks)
2925 {
2926         struct backref_node *node;
2927         struct btrfs_path *path;
2928         struct tree_block *block;
2929         struct rb_node *rb_node;
2930         int ret;
2931         int err = 0;
2932
2933         path = btrfs_alloc_path();
2934         if (!path) {
2935                 err = -ENOMEM;
2936                 goto out_free_blocks;
2937         }
2938
2939         rb_node = rb_first(blocks);
2940         while (rb_node) {
2941                 block = rb_entry(rb_node, struct tree_block, rb_node);
2942                 if (!block->key_ready)
2943                         reada_tree_block(rc, block);
2944                 rb_node = rb_next(rb_node);
2945         }
2946
2947         rb_node = rb_first(blocks);
2948         while (rb_node) {
2949                 block = rb_entry(rb_node, struct tree_block, rb_node);
2950                 if (!block->key_ready) {
2951                         err = get_tree_block_key(rc, block);
2952                         if (err)
2953                                 goto out_free_path;
2954                 }
2955                 rb_node = rb_next(rb_node);
2956         }
2957
2958         rb_node = rb_first(blocks);
2959         while (rb_node) {
2960                 block = rb_entry(rb_node, struct tree_block, rb_node);
2961
2962                 node = build_backref_tree(rc, &block->key,
2963                                           block->level, block->bytenr);
2964                 if (IS_ERR(node)) {
2965                         err = PTR_ERR(node);
2966                         goto out;
2967                 }
2968
2969                 ret = relocate_tree_block(trans, rc, node, &block->key,
2970                                           path);
2971                 if (ret < 0) {
2972                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2973                                 err = ret;
2974                         goto out;
2975                 }
2976                 rb_node = rb_next(rb_node);
2977         }
2978 out:
2979         err = finish_pending_nodes(trans, rc, path, err);
2980
2981 out_free_path:
2982         btrfs_free_path(path);
2983 out_free_blocks:
2984         free_block_list(blocks);
2985         return err;
2986 }
2987
2988 static noinline_for_stack
2989 int prealloc_file_extent_cluster(struct inode *inode,
2990                                  struct file_extent_cluster *cluster)
2991 {
2992         u64 alloc_hint = 0;
2993         u64 start;
2994         u64 end;
2995         u64 offset = BTRFS_I(inode)->index_cnt;
2996         u64 num_bytes;
2997         int nr = 0;
2998         int ret = 0;
2999
3000         BUG_ON(cluster->start != cluster->boundary[0]);
3001         mutex_lock(&inode->i_mutex);
3002
3003         ret = btrfs_check_data_free_space(inode, cluster->end +
3004                                           1 - cluster->start);
3005         if (ret)
3006                 goto out;
3007
3008         while (nr < cluster->nr) {
3009                 start = cluster->boundary[nr] - offset;
3010                 if (nr + 1 < cluster->nr)
3011                         end = cluster->boundary[nr + 1] - 1 - offset;
3012                 else
3013                         end = cluster->end - offset;
3014
3015                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3016                 num_bytes = end + 1 - start;
3017                 ret = btrfs_prealloc_file_range(inode, 0, start,
3018                                                 num_bytes, num_bytes,
3019                                                 end + 1, &alloc_hint);
3020                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3021                 if (ret)
3022                         break;
3023                 nr++;
3024         }
3025         btrfs_free_reserved_data_space(inode, cluster->end +
3026                                        1 - cluster->start);
3027 out:
3028         mutex_unlock(&inode->i_mutex);
3029         return ret;
3030 }
3031
3032 static noinline_for_stack
3033 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3034                          u64 block_start)
3035 {
3036         struct btrfs_root *root = BTRFS_I(inode)->root;
3037         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3038         struct extent_map *em;
3039         int ret = 0;
3040
3041         em = alloc_extent_map();
3042         if (!em)
3043                 return -ENOMEM;
3044
3045         em->start = start;
3046         em->len = end + 1 - start;
3047         em->block_len = em->len;
3048         em->block_start = block_start;
3049         em->bdev = root->fs_info->fs_devices->latest_bdev;
3050         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3051
3052         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3053         while (1) {
3054                 write_lock(&em_tree->lock);
3055                 ret = add_extent_mapping(em_tree, em, 0);
3056                 write_unlock(&em_tree->lock);
3057                 if (ret != -EEXIST) {
3058                         free_extent_map(em);
3059                         break;
3060                 }
3061                 btrfs_drop_extent_cache(inode, start, end, 0);
3062         }
3063         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3064         return ret;
3065 }
3066
3067 static int relocate_file_extent_cluster(struct inode *inode,
3068                                         struct file_extent_cluster *cluster)
3069 {
3070         u64 page_start;
3071         u64 page_end;
3072         u64 offset = BTRFS_I(inode)->index_cnt;
3073         unsigned long index;
3074         unsigned long last_index;
3075         struct page *page;
3076         struct file_ra_state *ra;
3077         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3078         int nr = 0;
3079         int ret = 0;
3080
3081         if (!cluster->nr)
3082                 return 0;
3083
3084         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3085         if (!ra)
3086                 return -ENOMEM;
3087
3088         ret = prealloc_file_extent_cluster(inode, cluster);
3089         if (ret)
3090                 goto out;
3091
3092         file_ra_state_init(ra, inode->i_mapping);
3093
3094         ret = setup_extent_mapping(inode, cluster->start - offset,
3095                                    cluster->end - offset, cluster->start);
3096         if (ret)
3097                 goto out;
3098
3099         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3100         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3101         while (index <= last_index) {
3102                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3103                 if (ret)
3104                         goto out;
3105
3106                 page = find_lock_page(inode->i_mapping, index);
3107                 if (!page) {
3108                         page_cache_sync_readahead(inode->i_mapping,
3109                                                   ra, NULL, index,
3110                                                   last_index + 1 - index);
3111                         page = find_or_create_page(inode->i_mapping, index,
3112                                                    mask);
3113                         if (!page) {
3114                                 btrfs_delalloc_release_metadata(inode,
3115                                                         PAGE_CACHE_SIZE);
3116                                 ret = -ENOMEM;
3117                                 goto out;
3118                         }
3119                 }
3120
3121                 if (PageReadahead(page)) {
3122                         page_cache_async_readahead(inode->i_mapping,
3123                                                    ra, NULL, page, index,
3124                                                    last_index + 1 - index);
3125                 }
3126
3127                 if (!PageUptodate(page)) {
3128                         btrfs_readpage(NULL, page);
3129                         lock_page(page);
3130                         if (!PageUptodate(page)) {
3131                                 unlock_page(page);
3132                                 page_cache_release(page);
3133                                 btrfs_delalloc_release_metadata(inode,
3134                                                         PAGE_CACHE_SIZE);
3135                                 ret = -EIO;
3136                                 goto out;
3137                         }
3138                 }
3139
3140                 page_start = page_offset(page);
3141                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3142
3143                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3144
3145                 set_page_extent_mapped(page);
3146
3147                 if (nr < cluster->nr &&
3148                     page_start + offset == cluster->boundary[nr]) {
3149                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3150                                         page_start, page_end,
3151                                         EXTENT_BOUNDARY, GFP_NOFS);
3152                         nr++;
3153                 }
3154
3155                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3156                 set_page_dirty(page);
3157
3158                 unlock_extent(&BTRFS_I(inode)->io_tree,
3159                               page_start, page_end);
3160                 unlock_page(page);
3161                 page_cache_release(page);
3162
3163                 index++;
3164                 balance_dirty_pages_ratelimited(inode->i_mapping);
3165                 btrfs_throttle(BTRFS_I(inode)->root);
3166         }
3167         WARN_ON(nr != cluster->nr);
3168 out:
3169         kfree(ra);
3170         return ret;
3171 }
3172
3173 static noinline_for_stack
3174 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3175                          struct file_extent_cluster *cluster)
3176 {
3177         int ret;
3178
3179         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3180                 ret = relocate_file_extent_cluster(inode, cluster);
3181                 if (ret)
3182                         return ret;
3183                 cluster->nr = 0;
3184         }
3185
3186         if (!cluster->nr)
3187                 cluster->start = extent_key->objectid;
3188         else
3189                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3190         cluster->end = extent_key->objectid + extent_key->offset - 1;
3191         cluster->boundary[cluster->nr] = extent_key->objectid;
3192         cluster->nr++;
3193
3194         if (cluster->nr >= MAX_EXTENTS) {
3195                 ret = relocate_file_extent_cluster(inode, cluster);
3196                 if (ret)
3197                         return ret;
3198                 cluster->nr = 0;
3199         }
3200         return 0;
3201 }
3202
3203 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3204 static int get_ref_objectid_v0(struct reloc_control *rc,
3205                                struct btrfs_path *path,
3206                                struct btrfs_key *extent_key,
3207                                u64 *ref_objectid, int *path_change)
3208 {
3209         struct btrfs_key key;
3210         struct extent_buffer *leaf;
3211         struct btrfs_extent_ref_v0 *ref0;
3212         int ret;
3213         int slot;
3214
3215         leaf = path->nodes[0];
3216         slot = path->slots[0];
3217         while (1) {
3218                 if (slot >= btrfs_header_nritems(leaf)) {
3219                         ret = btrfs_next_leaf(rc->extent_root, path);
3220                         if (ret < 0)
3221                                 return ret;
3222                         BUG_ON(ret > 0);
3223                         leaf = path->nodes[0];
3224                         slot = path->slots[0];
3225                         if (path_change)
3226                                 *path_change = 1;
3227                 }
3228                 btrfs_item_key_to_cpu(leaf, &key, slot);
3229                 if (key.objectid != extent_key->objectid)
3230                         return -ENOENT;
3231
3232                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3233                         slot++;
3234                         continue;
3235                 }
3236                 ref0 = btrfs_item_ptr(leaf, slot,
3237                                 struct btrfs_extent_ref_v0);
3238                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3239                 break;
3240         }
3241         return 0;
3242 }
3243 #endif
3244
3245 /*
3246  * helper to add a tree block to the list.
3247  * the major work is getting the generation and level of the block
3248  */
3249 static int add_tree_block(struct reloc_control *rc,
3250                           struct btrfs_key *extent_key,
3251                           struct btrfs_path *path,
3252                           struct rb_root *blocks)
3253 {
3254         struct extent_buffer *eb;
3255         struct btrfs_extent_item *ei;
3256         struct btrfs_tree_block_info *bi;
3257         struct tree_block *block;
3258         struct rb_node *rb_node;
3259         u32 item_size;
3260         int level = -1;
3261         int generation;
3262
3263         eb =  path->nodes[0];
3264         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3265
3266         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3267             item_size >= sizeof(*ei) + sizeof(*bi)) {
3268                 ei = btrfs_item_ptr(eb, path->slots[0],
3269                                 struct btrfs_extent_item);
3270                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3271                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3272                         level = btrfs_tree_block_level(eb, bi);
3273                 } else {
3274                         level = (int)extent_key->offset;
3275                 }
3276                 generation = btrfs_extent_generation(eb, ei);
3277         } else {
3278 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3279                 u64 ref_owner;
3280                 int ret;
3281
3282                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3283                 ret = get_ref_objectid_v0(rc, path, extent_key,
3284                                           &ref_owner, NULL);
3285                 if (ret < 0)
3286                         return ret;
3287                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3288                 level = (int)ref_owner;
3289                 /* FIXME: get real generation */
3290                 generation = 0;
3291 #else
3292                 BUG();
3293 #endif
3294         }
3295
3296         btrfs_release_path(path);
3297
3298         BUG_ON(level == -1);
3299
3300         block = kmalloc(sizeof(*block), GFP_NOFS);
3301         if (!block)
3302                 return -ENOMEM;
3303
3304         block->bytenr = extent_key->objectid;
3305         block->key.objectid = rc->extent_root->leafsize;
3306         block->key.offset = generation;
3307         block->level = level;
3308         block->key_ready = 0;
3309
3310         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3311         if (rb_node)
3312                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3313
3314         return 0;
3315 }
3316
3317 /*
3318  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3319  */
3320 static int __add_tree_block(struct reloc_control *rc,
3321                             u64 bytenr, u32 blocksize,
3322                             struct rb_root *blocks)
3323 {
3324         struct btrfs_path *path;
3325         struct btrfs_key key;
3326         int ret;
3327         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3328                                         SKINNY_METADATA);
3329
3330         if (tree_block_processed(bytenr, blocksize, rc))
3331                 return 0;
3332
3333         if (tree_search(blocks, bytenr))
3334                 return 0;
3335
3336         path = btrfs_alloc_path();
3337         if (!path)
3338                 return -ENOMEM;
3339 again:
3340         key.objectid = bytenr;
3341         if (skinny) {
3342                 key.type = BTRFS_METADATA_ITEM_KEY;
3343                 key.offset = (u64)-1;
3344         } else {
3345                 key.type = BTRFS_EXTENT_ITEM_KEY;
3346                 key.offset = blocksize;
3347         }
3348
3349         path->search_commit_root = 1;
3350         path->skip_locking = 1;
3351         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3352         if (ret < 0)
3353                 goto out;
3354
3355         if (ret > 0 && skinny) {
3356                 if (path->slots[0]) {
3357                         path->slots[0]--;
3358                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3359                                               path->slots[0]);
3360                         if (key.objectid == bytenr &&
3361                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3362                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3363                               key.offset == blocksize)))
3364                                 ret = 0;
3365                 }
3366
3367                 if (ret) {
3368                         skinny = false;
3369                         btrfs_release_path(path);
3370                         goto again;
3371                 }
3372         }
3373         BUG_ON(ret);
3374
3375         ret = add_tree_block(rc, &key, path, blocks);
3376 out:
3377         btrfs_free_path(path);
3378         return ret;
3379 }
3380
3381 /*
3382  * helper to check if the block use full backrefs for pointers in it
3383  */
3384 static int block_use_full_backref(struct reloc_control *rc,
3385                                   struct extent_buffer *eb)
3386 {
3387         u64 flags;
3388         int ret;
3389
3390         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3391             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3392                 return 1;
3393
3394         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3395                                        eb->start, btrfs_header_level(eb), 1,
3396                                        NULL, &flags);
3397         BUG_ON(ret);
3398
3399         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3400                 ret = 1;
3401         else
3402                 ret = 0;
3403         return ret;
3404 }
3405
3406 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3407                                     struct inode *inode, u64 ino)
3408 {
3409         struct btrfs_key key;
3410         struct btrfs_path *path;
3411         struct btrfs_root *root = fs_info->tree_root;
3412         struct btrfs_trans_handle *trans;
3413         int ret = 0;
3414
3415         if (inode)
3416                 goto truncate;
3417
3418         key.objectid = ino;
3419         key.type = BTRFS_INODE_ITEM_KEY;
3420         key.offset = 0;
3421
3422         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3423         if (IS_ERR(inode) || is_bad_inode(inode)) {
3424                 if (!IS_ERR(inode))
3425                         iput(inode);
3426                 return -ENOENT;
3427         }
3428
3429 truncate:
3430         ret = btrfs_check_trunc_cache_free_space(root,
3431                                                  &fs_info->global_block_rsv);
3432         if (ret)
3433                 goto out;
3434
3435         path = btrfs_alloc_path();
3436         if (!path) {
3437                 ret = -ENOMEM;
3438                 goto out;
3439         }
3440
3441         trans = btrfs_join_transaction(root);
3442         if (IS_ERR(trans)) {
3443                 btrfs_free_path(path);
3444                 ret = PTR_ERR(trans);
3445                 goto out;
3446         }
3447
3448         ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3449
3450         btrfs_free_path(path);
3451         btrfs_end_transaction(trans, root);
3452         btrfs_btree_balance_dirty(root);
3453 out:
3454         iput(inode);
3455         return ret;
3456 }
3457
3458 /*
3459  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3460  * this function scans fs tree to find blocks reference the data extent
3461  */
3462 static int find_data_references(struct reloc_control *rc,
3463                                 struct btrfs_key *extent_key,
3464                                 struct extent_buffer *leaf,
3465                                 struct btrfs_extent_data_ref *ref,
3466                                 struct rb_root *blocks)
3467 {
3468         struct btrfs_path *path;
3469         struct tree_block *block;
3470         struct btrfs_root *root;
3471         struct btrfs_file_extent_item *fi;
3472         struct rb_node *rb_node;
3473         struct btrfs_key key;
3474         u64 ref_root;
3475         u64 ref_objectid;
3476         u64 ref_offset;
3477         u32 ref_count;
3478         u32 nritems;
3479         int err = 0;
3480         int added = 0;
3481         int counted;
3482         int ret;
3483
3484         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3485         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3486         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3487         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3488
3489         /*
3490          * This is an extent belonging to the free space cache, lets just delete
3491          * it and redo the search.
3492          */
3493         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3494                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3495                                                NULL, ref_objectid);
3496                 if (ret != -ENOENT)
3497                         return ret;
3498                 ret = 0;
3499         }
3500
3501         path = btrfs_alloc_path();
3502         if (!path)
3503                 return -ENOMEM;
3504         path->reada = 1;
3505
3506         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3507         if (IS_ERR(root)) {
3508                 err = PTR_ERR(root);
3509                 goto out;
3510         }
3511
3512         key.objectid = ref_objectid;
3513         key.type = BTRFS_EXTENT_DATA_KEY;
3514         if (ref_offset > ((u64)-1 << 32))
3515                 key.offset = 0;
3516         else
3517                 key.offset = ref_offset;
3518
3519         path->search_commit_root = 1;
3520         path->skip_locking = 1;
3521         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3522         if (ret < 0) {
3523                 err = ret;
3524                 goto out;
3525         }
3526
3527         leaf = path->nodes[0];
3528         nritems = btrfs_header_nritems(leaf);
3529         /*
3530          * the references in tree blocks that use full backrefs
3531          * are not counted in
3532          */
3533         if (block_use_full_backref(rc, leaf))
3534                 counted = 0;
3535         else
3536                 counted = 1;
3537         rb_node = tree_search(blocks, leaf->start);
3538         if (rb_node) {
3539                 if (counted)
3540                         added = 1;
3541                 else
3542                         path->slots[0] = nritems;
3543         }
3544
3545         while (ref_count > 0) {
3546                 while (path->slots[0] >= nritems) {
3547                         ret = btrfs_next_leaf(root, path);
3548                         if (ret < 0) {
3549                                 err = ret;
3550                                 goto out;
3551                         }
3552                         if (ret > 0) {
3553                                 WARN_ON(1);
3554                                 goto out;
3555                         }
3556
3557                         leaf = path->nodes[0];
3558                         nritems = btrfs_header_nritems(leaf);
3559                         added = 0;
3560
3561                         if (block_use_full_backref(rc, leaf))
3562                                 counted = 0;
3563                         else
3564                                 counted = 1;
3565                         rb_node = tree_search(blocks, leaf->start);
3566                         if (rb_node) {
3567                                 if (counted)
3568                                         added = 1;
3569                                 else
3570                                         path->slots[0] = nritems;
3571                         }
3572                 }
3573
3574                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3575                 if (key.objectid != ref_objectid ||
3576                     key.type != BTRFS_EXTENT_DATA_KEY) {
3577                         WARN_ON(1);
3578                         break;
3579                 }
3580
3581                 fi = btrfs_item_ptr(leaf, path->slots[0],
3582                                     struct btrfs_file_extent_item);
3583
3584                 if (btrfs_file_extent_type(leaf, fi) ==
3585                     BTRFS_FILE_EXTENT_INLINE)
3586                         goto next;
3587
3588                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3589                     extent_key->objectid)
3590                         goto next;
3591
3592                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3593                 if (key.offset != ref_offset)
3594                         goto next;
3595
3596                 if (counted)
3597                         ref_count--;
3598                 if (added)
3599                         goto next;
3600
3601                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3602                         block = kmalloc(sizeof(*block), GFP_NOFS);
3603                         if (!block) {
3604                                 err = -ENOMEM;
3605                                 break;
3606                         }
3607                         block->bytenr = leaf->start;
3608                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3609                         block->level = 0;
3610                         block->key_ready = 1;
3611                         rb_node = tree_insert(blocks, block->bytenr,
3612                                               &block->rb_node);
3613                         if (rb_node)
3614                                 backref_tree_panic(rb_node, -EEXIST,
3615                                                    block->bytenr);
3616                 }
3617                 if (counted)
3618                         added = 1;
3619                 else
3620                         path->slots[0] = nritems;
3621 next:
3622                 path->slots[0]++;
3623
3624         }
3625 out:
3626         btrfs_free_path(path);
3627         return err;
3628 }
3629
3630 /*
3631  * helper to find all tree blocks that reference a given data extent
3632  */
3633 static noinline_for_stack
3634 int add_data_references(struct reloc_control *rc,
3635                         struct btrfs_key *extent_key,
3636                         struct btrfs_path *path,
3637                         struct rb_root *blocks)
3638 {
3639         struct btrfs_key key;
3640         struct extent_buffer *eb;
3641         struct btrfs_extent_data_ref *dref;
3642         struct btrfs_extent_inline_ref *iref;
3643         unsigned long ptr;
3644         unsigned long end;
3645         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3646         int ret = 0;
3647         int err = 0;
3648
3649         eb = path->nodes[0];
3650         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3651         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3652 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3653         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3654                 ptr = end;
3655         else
3656 #endif
3657                 ptr += sizeof(struct btrfs_extent_item);
3658
3659         while (ptr < end) {
3660                 iref = (struct btrfs_extent_inline_ref *)ptr;
3661                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3662                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3663                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3664                         ret = __add_tree_block(rc, key.offset, blocksize,
3665                                                blocks);
3666                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3667                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3668                         ret = find_data_references(rc, extent_key,
3669                                                    eb, dref, blocks);
3670                 } else {
3671                         BUG();
3672                 }
3673                 if (ret) {
3674                         err = ret;
3675                         goto out;
3676                 }
3677                 ptr += btrfs_extent_inline_ref_size(key.type);
3678         }
3679         WARN_ON(ptr > end);
3680
3681         while (1) {
3682                 cond_resched();
3683                 eb = path->nodes[0];
3684                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3685                         ret = btrfs_next_leaf(rc->extent_root, path);
3686                         if (ret < 0) {
3687                                 err = ret;
3688                                 break;
3689                         }
3690                         if (ret > 0)
3691                                 break;
3692                         eb = path->nodes[0];
3693                 }
3694
3695                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3696                 if (key.objectid != extent_key->objectid)
3697                         break;
3698
3699 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3700                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3701                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3702 #else
3703                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3704                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3705 #endif
3706                         ret = __add_tree_block(rc, key.offset, blocksize,
3707                                                blocks);
3708                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3709                         dref = btrfs_item_ptr(eb, path->slots[0],
3710                                               struct btrfs_extent_data_ref);
3711                         ret = find_data_references(rc, extent_key,
3712                                                    eb, dref, blocks);
3713                 } else {
3714                         ret = 0;
3715                 }
3716                 if (ret) {
3717                         err = ret;
3718                         break;
3719                 }
3720                 path->slots[0]++;
3721         }
3722 out:
3723         btrfs_release_path(path);
3724         if (err)
3725                 free_block_list(blocks);
3726         return err;
3727 }
3728
3729 /*
3730  * helper to find next unprocessed extent
3731  */
3732 static noinline_for_stack
3733 int find_next_extent(struct btrfs_trans_handle *trans,
3734                      struct reloc_control *rc, struct btrfs_path *path,
3735                      struct btrfs_key *extent_key)
3736 {
3737         struct btrfs_key key;
3738         struct extent_buffer *leaf;
3739         u64 start, end, last;
3740         int ret;
3741
3742         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3743         while (1) {
3744                 cond_resched();
3745                 if (rc->search_start >= last) {
3746                         ret = 1;
3747                         break;
3748                 }
3749
3750                 key.objectid = rc->search_start;
3751                 key.type = BTRFS_EXTENT_ITEM_KEY;
3752                 key.offset = 0;
3753
3754                 path->search_commit_root = 1;
3755                 path->skip_locking = 1;
3756                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3757                                         0, 0);
3758                 if (ret < 0)
3759                         break;
3760 next:
3761                 leaf = path->nodes[0];
3762                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3763                         ret = btrfs_next_leaf(rc->extent_root, path);
3764                         if (ret != 0)
3765                                 break;
3766                         leaf = path->nodes[0];
3767                 }
3768
3769                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3770                 if (key.objectid >= last) {
3771                         ret = 1;
3772                         break;
3773                 }
3774
3775                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3776                     key.type != BTRFS_METADATA_ITEM_KEY) {
3777                         path->slots[0]++;
3778                         goto next;
3779                 }
3780
3781                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3782                     key.objectid + key.offset <= rc->search_start) {
3783                         path->slots[0]++;
3784                         goto next;
3785                 }
3786
3787                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3788                     key.objectid + rc->extent_root->leafsize <=
3789                     rc->search_start) {
3790                         path->slots[0]++;
3791                         goto next;
3792                 }
3793
3794                 ret = find_first_extent_bit(&rc->processed_blocks,
3795                                             key.objectid, &start, &end,
3796                                             EXTENT_DIRTY, NULL);
3797
3798                 if (ret == 0 && start <= key.objectid) {
3799                         btrfs_release_path(path);
3800                         rc->search_start = end + 1;
3801                 } else {
3802                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3803                                 rc->search_start = key.objectid + key.offset;
3804                         else
3805                                 rc->search_start = key.objectid +
3806                                         rc->extent_root->leafsize;
3807                         memcpy(extent_key, &key, sizeof(key));
3808                         return 0;
3809                 }
3810         }
3811         btrfs_release_path(path);
3812         return ret;
3813 }
3814
3815 static void set_reloc_control(struct reloc_control *rc)
3816 {
3817         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3818
3819         mutex_lock(&fs_info->reloc_mutex);
3820         fs_info->reloc_ctl = rc;
3821         mutex_unlock(&fs_info->reloc_mutex);
3822 }
3823
3824 static void unset_reloc_control(struct reloc_control *rc)
3825 {
3826         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3827
3828         mutex_lock(&fs_info->reloc_mutex);
3829         fs_info->reloc_ctl = NULL;
3830         mutex_unlock(&fs_info->reloc_mutex);
3831 }
3832
3833 static int check_extent_flags(u64 flags)
3834 {
3835         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3836             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3837                 return 1;
3838         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3839             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3840                 return 1;
3841         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3842             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3843                 return 1;
3844         return 0;
3845 }
3846
3847 static noinline_for_stack
3848 int prepare_to_relocate(struct reloc_control *rc)
3849 {
3850         struct btrfs_trans_handle *trans;
3851         int ret;
3852
3853         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3854                                               BTRFS_BLOCK_RSV_TEMP);
3855         if (!rc->block_rsv)
3856                 return -ENOMEM;
3857
3858         /*
3859          * reserve some space for creating reloc trees.
3860          * btrfs_init_reloc_root will use them when there
3861          * is no reservation in transaction handle.
3862          */
3863         ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3864                                   rc->extent_root->nodesize * 256,
3865                                   BTRFS_RESERVE_FLUSH_ALL);
3866         if (ret)
3867                 return ret;
3868
3869         memset(&rc->cluster, 0, sizeof(rc->cluster));
3870         rc->search_start = rc->block_group->key.objectid;
3871         rc->extents_found = 0;
3872         rc->nodes_relocated = 0;
3873         rc->merging_rsv_size = 0;
3874
3875         rc->create_reloc_tree = 1;
3876         set_reloc_control(rc);
3877
3878         trans = btrfs_join_transaction(rc->extent_root);
3879         if (IS_ERR(trans)) {
3880                 unset_reloc_control(rc);
3881                 /*
3882                  * extent tree is not a ref_cow tree and has no reloc_root to
3883                  * cleanup.  And callers are responsible to free the above
3884                  * block rsv.
3885                  */
3886                 return PTR_ERR(trans);
3887         }
3888         btrfs_commit_transaction(trans, rc->extent_root);
3889         return 0;
3890 }
3891
3892 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3893 {
3894         struct rb_root blocks = RB_ROOT;
3895         struct btrfs_key key;
3896         struct btrfs_trans_handle *trans = NULL;
3897         struct btrfs_path *path;
3898         struct btrfs_extent_item *ei;
3899         u64 flags;
3900         u32 item_size;
3901         int ret;
3902         int err = 0;
3903         int progress = 0;
3904
3905         path = btrfs_alloc_path();
3906         if (!path)
3907                 return -ENOMEM;
3908         path->reada = 1;
3909
3910         ret = prepare_to_relocate(rc);
3911         if (ret) {
3912                 err = ret;
3913                 goto out_free;
3914         }
3915
3916         while (1) {
3917                 progress++;
3918                 trans = btrfs_start_transaction(rc->extent_root, 0);
3919                 if (IS_ERR(trans)) {
3920                         err = PTR_ERR(trans);
3921                         trans = NULL;
3922                         break;
3923                 }
3924 restart:
3925                 if (update_backref_cache(trans, &rc->backref_cache)) {
3926                         btrfs_end_transaction(trans, rc->extent_root);
3927                         continue;
3928                 }
3929
3930                 ret = find_next_extent(trans, rc, path, &key);
3931                 if (ret < 0)
3932                         err = ret;
3933                 if (ret != 0)
3934                         break;
3935
3936                 rc->extents_found++;
3937
3938                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3939                                     struct btrfs_extent_item);
3940                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3941                 if (item_size >= sizeof(*ei)) {
3942                         flags = btrfs_extent_flags(path->nodes[0], ei);
3943                         ret = check_extent_flags(flags);
3944                         BUG_ON(ret);
3945
3946                 } else {
3947 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3948                         u64 ref_owner;
3949                         int path_change = 0;
3950
3951                         BUG_ON(item_size !=
3952                                sizeof(struct btrfs_extent_item_v0));
3953                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3954                                                   &path_change);
3955                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3956                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3957                         else
3958                                 flags = BTRFS_EXTENT_FLAG_DATA;
3959
3960                         if (path_change) {
3961                                 btrfs_release_path(path);
3962
3963                                 path->search_commit_root = 1;
3964                                 path->skip_locking = 1;
3965                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3966                                                         &key, path, 0, 0);
3967                                 if (ret < 0) {
3968                                         err = ret;
3969                                         break;
3970                                 }
3971                                 BUG_ON(ret > 0);
3972                         }
3973 #else
3974                         BUG();
3975 #endif
3976                 }
3977
3978                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3979                         ret = add_tree_block(rc, &key, path, &blocks);
3980                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3981                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3982                         ret = add_data_references(rc, &key, path, &blocks);
3983                 } else {
3984                         btrfs_release_path(path);
3985                         ret = 0;
3986                 }
3987                 if (ret < 0) {
3988                         err = ret;
3989                         break;
3990                 }
3991
3992                 if (!RB_EMPTY_ROOT(&blocks)) {
3993                         ret = relocate_tree_blocks(trans, rc, &blocks);
3994                         if (ret < 0) {
3995                                 if (ret != -EAGAIN) {
3996                                         err = ret;
3997                                         break;
3998                                 }
3999                                 rc->extents_found--;
4000                                 rc->search_start = key.objectid;
4001                         }
4002                 }
4003
4004                 ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5);
4005                 if (ret < 0) {
4006                         if (ret != -ENOSPC) {
4007                                 err = ret;
4008                                 WARN_ON(1);
4009                                 break;
4010                         }
4011                         rc->commit_transaction = 1;
4012                 }
4013
4014                 if (rc->commit_transaction) {
4015                         rc->commit_transaction = 0;
4016                         ret = btrfs_commit_transaction(trans, rc->extent_root);
4017                         BUG_ON(ret);
4018                 } else {
4019                         btrfs_end_transaction_throttle(trans, rc->extent_root);
4020                         btrfs_btree_balance_dirty(rc->extent_root);
4021                 }
4022                 trans = NULL;
4023
4024                 if (rc->stage == MOVE_DATA_EXTENTS &&
4025                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4026                         rc->found_file_extent = 1;
4027                         ret = relocate_data_extent(rc->data_inode,
4028                                                    &key, &rc->cluster);
4029                         if (ret < 0) {
4030                                 err = ret;
4031                                 break;
4032                         }
4033                 }
4034         }
4035         if (trans && progress && err == -ENOSPC) {
4036                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4037                                               rc->block_group->flags);
4038                 if (ret == 0) {
4039                         err = 0;
4040                         progress = 0;
4041                         goto restart;
4042                 }
4043         }
4044
4045         btrfs_release_path(path);
4046         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4047                           GFP_NOFS);
4048
4049         if (trans) {
4050                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4051                 btrfs_btree_balance_dirty(rc->extent_root);
4052         }
4053
4054         if (!err) {
4055                 ret = relocate_file_extent_cluster(rc->data_inode,
4056                                                    &rc->cluster);
4057                 if (ret < 0)
4058                         err = ret;
4059         }
4060
4061         rc->create_reloc_tree = 0;
4062         set_reloc_control(rc);
4063
4064         backref_cache_cleanup(&rc->backref_cache);
4065         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4066
4067         err = prepare_to_merge(rc, err);
4068
4069         merge_reloc_roots(rc);
4070
4071         rc->merge_reloc_tree = 0;
4072         unset_reloc_control(rc);
4073         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4074
4075         /* get rid of pinned extents */
4076         trans = btrfs_join_transaction(rc->extent_root);
4077         if (IS_ERR(trans))
4078                 err = PTR_ERR(trans);
4079         else
4080                 btrfs_commit_transaction(trans, rc->extent_root);
4081 out_free:
4082         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4083         btrfs_free_path(path);
4084         return err;
4085 }
4086
4087 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4088                                  struct btrfs_root *root, u64 objectid)
4089 {
4090         struct btrfs_path *path;
4091         struct btrfs_inode_item *item;
4092         struct extent_buffer *leaf;
4093         int ret;
4094
4095         path = btrfs_alloc_path();
4096         if (!path)
4097                 return -ENOMEM;
4098
4099         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4100         if (ret)
4101                 goto out;
4102
4103         leaf = path->nodes[0];
4104         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4105         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4106         btrfs_set_inode_generation(leaf, item, 1);
4107         btrfs_set_inode_size(leaf, item, 0);
4108         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4109         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4110                                           BTRFS_INODE_PREALLOC);
4111         btrfs_mark_buffer_dirty(leaf);
4112         btrfs_release_path(path);
4113 out:
4114         btrfs_free_path(path);
4115         return ret;
4116 }
4117
4118 /*
4119  * helper to create inode for data relocation.
4120  * the inode is in data relocation tree and its link count is 0
4121  */
4122 static noinline_for_stack
4123 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4124                                  struct btrfs_block_group_cache *group)
4125 {
4126         struct inode *inode = NULL;
4127         struct btrfs_trans_handle *trans;
4128         struct btrfs_root *root;
4129         struct btrfs_key key;
4130         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4131         int err = 0;
4132
4133         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4134         if (IS_ERR(root))
4135                 return ERR_CAST(root);
4136
4137         trans = btrfs_start_transaction(root, 6);
4138         if (IS_ERR(trans))
4139                 return ERR_CAST(trans);
4140
4141         err = btrfs_find_free_objectid(root, &objectid);
4142         if (err)
4143                 goto out;
4144
4145         err = __insert_orphan_inode(trans, root, objectid);
4146         BUG_ON(err);
4147
4148         key.objectid = objectid;
4149         key.type = BTRFS_INODE_ITEM_KEY;
4150         key.offset = 0;
4151         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4152         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4153         BTRFS_I(inode)->index_cnt = group->key.objectid;
4154
4155         err = btrfs_orphan_add(trans, inode);
4156 out:
4157         btrfs_end_transaction(trans, root);
4158         btrfs_btree_balance_dirty(root);
4159         if (err) {
4160                 if (inode)
4161                         iput(inode);
4162                 inode = ERR_PTR(err);
4163         }
4164         return inode;
4165 }
4166
4167 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4168 {
4169         struct reloc_control *rc;
4170
4171         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4172         if (!rc)
4173                 return NULL;
4174
4175         INIT_LIST_HEAD(&rc->reloc_roots);
4176         backref_cache_init(&rc->backref_cache);
4177         mapping_tree_init(&rc->reloc_root_tree);
4178         extent_io_tree_init(&rc->processed_blocks,
4179                             fs_info->btree_inode->i_mapping);
4180         return rc;
4181 }
4182
4183 /*
4184  * function to relocate all extents in a block group.
4185  */
4186 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4187 {
4188         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4189         struct reloc_control *rc;
4190         struct inode *inode;
4191         struct btrfs_path *path;
4192         int ret;
4193         int rw = 0;
4194         int err = 0;
4195
4196         rc = alloc_reloc_control(fs_info);
4197         if (!rc)
4198                 return -ENOMEM;
4199
4200         rc->extent_root = extent_root;
4201
4202         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4203         BUG_ON(!rc->block_group);
4204
4205         if (!rc->block_group->ro) {
4206                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4207                 if (ret) {
4208                         err = ret;
4209                         goto out;
4210                 }
4211                 rw = 1;
4212         }
4213
4214         path = btrfs_alloc_path();
4215         if (!path) {
4216                 err = -ENOMEM;
4217                 goto out;
4218         }
4219
4220         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4221                                         path);
4222         btrfs_free_path(path);
4223
4224         if (!IS_ERR(inode))
4225                 ret = delete_block_group_cache(fs_info, inode, 0);
4226         else
4227                 ret = PTR_ERR(inode);
4228
4229         if (ret && ret != -ENOENT) {
4230                 err = ret;
4231                 goto out;
4232         }
4233
4234         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4235         if (IS_ERR(rc->data_inode)) {
4236                 err = PTR_ERR(rc->data_inode);
4237                 rc->data_inode = NULL;
4238                 goto out;
4239         }
4240
4241         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4242                rc->block_group->key.objectid, rc->block_group->flags);
4243
4244         ret = btrfs_start_all_delalloc_inodes(fs_info, 0);
4245         if (ret < 0) {
4246                 err = ret;
4247                 goto out;
4248         }
4249         btrfs_wait_all_ordered_extents(fs_info);
4250
4251         while (1) {
4252                 mutex_lock(&fs_info->cleaner_mutex);
4253                 ret = relocate_block_group(rc);
4254                 mutex_unlock(&fs_info->cleaner_mutex);
4255                 if (ret < 0) {
4256                         err = ret;
4257                         goto out;
4258                 }
4259
4260                 if (rc->extents_found == 0)
4261                         break;
4262
4263                 printk(KERN_INFO "btrfs: found %llu extents\n",
4264                         rc->extents_found);
4265
4266                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4267                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4268                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4269                                                  0, -1);
4270                         rc->stage = UPDATE_DATA_PTRS;
4271                 }
4272         }
4273
4274         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4275                                      rc->block_group->key.objectid,
4276                                      rc->block_group->key.objectid +
4277                                      rc->block_group->key.offset - 1);
4278
4279         WARN_ON(rc->block_group->pinned > 0);
4280         WARN_ON(rc->block_group->reserved > 0);
4281         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4282 out:
4283         if (err && rw)
4284                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4285         iput(rc->data_inode);
4286         btrfs_put_block_group(rc->block_group);
4287         kfree(rc);
4288         return err;
4289 }
4290
4291 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4292 {
4293         struct btrfs_trans_handle *trans;
4294         int ret, err;
4295
4296         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4297         if (IS_ERR(trans))
4298                 return PTR_ERR(trans);
4299
4300         memset(&root->root_item.drop_progress, 0,
4301                 sizeof(root->root_item.drop_progress));
4302         root->root_item.drop_level = 0;
4303         btrfs_set_root_refs(&root->root_item, 0);
4304         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4305                                 &root->root_key, &root->root_item);
4306
4307         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4308         if (err)
4309                 return err;
4310         return ret;
4311 }
4312
4313 /*
4314  * recover relocation interrupted by system crash.
4315  *
4316  * this function resumes merging reloc trees with corresponding fs trees.
4317  * this is important for keeping the sharing of tree blocks
4318  */
4319 int btrfs_recover_relocation(struct btrfs_root *root)
4320 {
4321         LIST_HEAD(reloc_roots);
4322         struct btrfs_key key;
4323         struct btrfs_root *fs_root;
4324         struct btrfs_root *reloc_root;
4325         struct btrfs_path *path;
4326         struct extent_buffer *leaf;
4327         struct reloc_control *rc = NULL;
4328         struct btrfs_trans_handle *trans;
4329         int ret;
4330         int err = 0;
4331
4332         path = btrfs_alloc_path();
4333         if (!path)
4334                 return -ENOMEM;
4335         path->reada = -1;
4336
4337         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4338         key.type = BTRFS_ROOT_ITEM_KEY;
4339         key.offset = (u64)-1;
4340
4341         while (1) {
4342                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4343                                         path, 0, 0);
4344                 if (ret < 0) {
4345                         err = ret;
4346                         goto out;
4347                 }
4348                 if (ret > 0) {
4349                         if (path->slots[0] == 0)
4350                                 break;
4351                         path->slots[0]--;
4352                 }
4353                 leaf = path->nodes[0];
4354                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4355                 btrfs_release_path(path);
4356
4357                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4358                     key.type != BTRFS_ROOT_ITEM_KEY)
4359                         break;
4360
4361                 reloc_root = btrfs_read_fs_root(root, &key);
4362                 if (IS_ERR(reloc_root)) {
4363                         err = PTR_ERR(reloc_root);
4364                         goto out;
4365                 }
4366
4367                 list_add(&reloc_root->root_list, &reloc_roots);
4368
4369                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4370                         fs_root = read_fs_root(root->fs_info,
4371                                                reloc_root->root_key.offset);
4372                         if (IS_ERR(fs_root)) {
4373                                 ret = PTR_ERR(fs_root);
4374                                 if (ret != -ENOENT) {
4375                                         err = ret;
4376                                         goto out;
4377                                 }
4378                                 ret = mark_garbage_root(reloc_root);
4379                                 if (ret < 0) {
4380                                         err = ret;
4381                                         goto out;
4382                                 }
4383                         }
4384                 }
4385
4386                 if (key.offset == 0)
4387                         break;
4388
4389                 key.offset--;
4390         }
4391         btrfs_release_path(path);
4392
4393         if (list_empty(&reloc_roots))
4394                 goto out;
4395
4396         rc = alloc_reloc_control(root->fs_info);
4397         if (!rc) {
4398                 err = -ENOMEM;
4399                 goto out;
4400         }
4401
4402         rc->extent_root = root->fs_info->extent_root;
4403
4404         set_reloc_control(rc);
4405
4406         trans = btrfs_join_transaction(rc->extent_root);
4407         if (IS_ERR(trans)) {
4408                 unset_reloc_control(rc);
4409                 err = PTR_ERR(trans);
4410                 goto out_free;
4411         }
4412
4413         rc->merge_reloc_tree = 1;
4414
4415         while (!list_empty(&reloc_roots)) {
4416                 reloc_root = list_entry(reloc_roots.next,
4417                                         struct btrfs_root, root_list);
4418                 list_del(&reloc_root->root_list);
4419
4420                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4421                         list_add_tail(&reloc_root->root_list,
4422                                       &rc->reloc_roots);
4423                         continue;
4424                 }
4425
4426                 fs_root = read_fs_root(root->fs_info,
4427                                        reloc_root->root_key.offset);
4428                 if (IS_ERR(fs_root)) {
4429                         err = PTR_ERR(fs_root);
4430                         goto out_free;
4431                 }
4432
4433                 err = __add_reloc_root(reloc_root);
4434                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4435                 fs_root->reloc_root = reloc_root;
4436         }
4437
4438         err = btrfs_commit_transaction(trans, rc->extent_root);
4439         if (err)
4440                 goto out_free;
4441
4442         merge_reloc_roots(rc);
4443
4444         unset_reloc_control(rc);
4445
4446         trans = btrfs_join_transaction(rc->extent_root);
4447         if (IS_ERR(trans))
4448                 err = PTR_ERR(trans);
4449         else
4450                 err = btrfs_commit_transaction(trans, rc->extent_root);
4451 out_free:
4452         kfree(rc);
4453 out:
4454         if (!list_empty(&reloc_roots))
4455                 free_reloc_roots(&reloc_roots);
4456
4457         btrfs_free_path(path);
4458
4459         if (err == 0) {
4460                 /* cleanup orphan inode in data relocation tree */
4461                 fs_root = read_fs_root(root->fs_info,
4462                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4463                 if (IS_ERR(fs_root))
4464                         err = PTR_ERR(fs_root);
4465                 else
4466                         err = btrfs_orphan_cleanup(fs_root);
4467         }
4468         return err;
4469 }
4470
4471 /*
4472  * helper to add ordered checksum for data relocation.
4473  *
4474  * cloning checksum properly handles the nodatasum extents.
4475  * it also saves CPU time to re-calculate the checksum.
4476  */
4477 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4478 {
4479         struct btrfs_ordered_sum *sums;
4480         struct btrfs_ordered_extent *ordered;
4481         struct btrfs_root *root = BTRFS_I(inode)->root;
4482         int ret;
4483         u64 disk_bytenr;
4484         LIST_HEAD(list);
4485
4486         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4487         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4488
4489         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4490         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4491                                        disk_bytenr + len - 1, &list, 0);
4492         if (ret)
4493                 goto out;
4494
4495         disk_bytenr = ordered->start;
4496         while (!list_empty(&list)) {
4497                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4498                 list_del_init(&sums->list);
4499
4500                 sums->bytenr = disk_bytenr;
4501                 disk_bytenr += sums->len;
4502
4503                 btrfs_add_ordered_sum(inode, ordered, sums);
4504         }
4505 out:
4506         btrfs_put_ordered_extent(ordered);
4507         return ret;
4508 }
4509
4510 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4511                           struct btrfs_root *root, struct extent_buffer *buf,
4512                           struct extent_buffer *cow)
4513 {
4514         struct reloc_control *rc;
4515         struct backref_node *node;
4516         int first_cow = 0;
4517         int level;
4518         int ret = 0;
4519
4520         rc = root->fs_info->reloc_ctl;
4521         if (!rc)
4522                 return 0;
4523
4524         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4525                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4526
4527         level = btrfs_header_level(buf);
4528         if (btrfs_header_generation(buf) <=
4529             btrfs_root_last_snapshot(&root->root_item))
4530                 first_cow = 1;
4531
4532         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4533             rc->create_reloc_tree) {
4534                 WARN_ON(!first_cow && level == 0);
4535
4536                 node = rc->backref_cache.path[level];
4537                 BUG_ON(node->bytenr != buf->start &&
4538                        node->new_bytenr != buf->start);
4539
4540                 drop_node_buffer(node);
4541                 extent_buffer_get(cow);
4542                 node->eb = cow;
4543                 node->new_bytenr = cow->start;
4544
4545                 if (!node->pending) {
4546                         list_move_tail(&node->list,
4547                                        &rc->backref_cache.pending[level]);
4548                         node->pending = 1;
4549                 }
4550
4551                 if (first_cow)
4552                         __mark_block_processed(rc, node);
4553
4554                 if (first_cow && level > 0)
4555                         rc->nodes_relocated += buf->len;
4556         }
4557
4558         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4559                 ret = replace_file_extents(trans, rc, root, cow);
4560         return ret;
4561 }
4562
4563 /*
4564  * called before creating snapshot. it calculates metadata reservation
4565  * requried for relocating tree blocks in the snapshot
4566  */
4567 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4568                               struct btrfs_pending_snapshot *pending,
4569                               u64 *bytes_to_reserve)
4570 {
4571         struct btrfs_root *root;
4572         struct reloc_control *rc;
4573
4574         root = pending->root;
4575         if (!root->reloc_root)
4576                 return;
4577
4578         rc = root->fs_info->reloc_ctl;
4579         if (!rc->merge_reloc_tree)
4580                 return;
4581
4582         root = root->reloc_root;
4583         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4584         /*
4585          * relocation is in the stage of merging trees. the space
4586          * used by merging a reloc tree is twice the size of
4587          * relocated tree nodes in the worst case. half for cowing
4588          * the reloc tree, half for cowing the fs tree. the space
4589          * used by cowing the reloc tree will be freed after the
4590          * tree is dropped. if we create snapshot, cowing the fs
4591          * tree may use more space than it frees. so we need
4592          * reserve extra space.
4593          */
4594         *bytes_to_reserve += rc->nodes_relocated;
4595 }
4596
4597 /*
4598  * called after snapshot is created. migrate block reservation
4599  * and create reloc root for the newly created snapshot
4600  */
4601 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4602                                struct btrfs_pending_snapshot *pending)
4603 {
4604         struct btrfs_root *root = pending->root;
4605         struct btrfs_root *reloc_root;
4606         struct btrfs_root *new_root;
4607         struct reloc_control *rc;
4608         int ret;
4609
4610         if (!root->reloc_root)
4611                 return 0;
4612
4613         rc = root->fs_info->reloc_ctl;
4614         rc->merging_rsv_size += rc->nodes_relocated;
4615
4616         if (rc->merge_reloc_tree) {
4617                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4618                                               rc->block_rsv,
4619                                               rc->nodes_relocated);
4620                 if (ret)
4621                         return ret;
4622         }
4623
4624         new_root = pending->snap;
4625         reloc_root = create_reloc_root(trans, root->reloc_root,
4626                                        new_root->root_key.objectid);
4627         if (IS_ERR(reloc_root))
4628                 return PTR_ERR(reloc_root);
4629
4630         ret = __add_reloc_root(reloc_root);
4631         BUG_ON(ret < 0);
4632         new_root->reloc_root = reloc_root;
4633
4634         if (rc->create_reloc_tree)
4635                 ret = clone_backref_node(trans, rc, root, reloc_root);
4636         return ret;
4637 }