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 #define RELOCATION_RESERVED_NODES       256
  98
  99 struct backref_cache {
 100         /* red black tree of all backref nodes in the cache */
 101         struct rb_root rb_root;
 102         /* for passing backref nodes to btrfs_reloc_cow_block */
 103         struct backref_node *path[BTRFS_MAX_LEVEL];
 104         /*
 105          * list of blocks that have been cowed but some block
 106          * pointers in upper level blocks may not reflect the
 107          * new location
 108          */
 109         struct list_head pending[BTRFS_MAX_LEVEL];
 110         /* list of backref nodes with no child node */
 111         struct list_head leaves;
 112         /* list of blocks that have been cowed in current transaction */
 113         struct list_head changed;
 114         /* list of detached backref node. */
 115         struct list_head detached;
 116
 117         u64 last_trans;
 118
 119         int nr_nodes;
 120         int nr_edges;
 121 };
 122
 123 /*
 124  * map address of tree root to tree
 125  */
 126 struct mapping_node {
 127         struct rb_node rb_node;
 128         u64 bytenr;
 129         void *data;
 130 };
 131
 132 struct mapping_tree {
 133         struct rb_root rb_root;
 134         spinlock_t lock;
 135 };
 136
 137 /*
 138  * present a tree block to process
 139  */
 140 struct tree_block {
 141         struct rb_node rb_node;
 142         u64 bytenr;
 143         struct btrfs_key key;
 144         unsigned int level:8;
 145         unsigned int key_ready:1;
 146 };
 147
 148 #define MAX_EXTENTS 128
 149
 150 struct file_extent_cluster {
 151         u64 start;
 152         u64 end;
 153         u64 boundary[MAX_EXTENTS];
 154         unsigned int nr;
 155 };
 156
 157 struct reloc_control {
 158         /* block group to relocate */
 159         struct btrfs_block_group_cache *block_group;
 160         /* extent tree */
 161         struct btrfs_root *extent_root;
 162         /* inode for moving data */
 163         struct inode *data_inode;
 164
 165         struct btrfs_block_rsv *block_rsv;
 166
 167         struct backref_cache backref_cache;
 168
 169         struct file_extent_cluster cluster;
 170         /* tree blocks have been processed */
 171         struct extent_io_tree processed_blocks;
 172         /* map start of tree root to corresponding reloc tree */
 173         struct mapping_tree reloc_root_tree;
 174         /* list of reloc trees */
 175         struct list_head reloc_roots;
 176         /* size of metadata reservation for merging reloc trees */
 177         u64 merging_rsv_size;
 178         /* size of relocated tree nodes */
 179         u64 nodes_relocated;
 180         /* reserved size for block group relocation*/
 181         u64 reserved_bytes;
 182
 183         u64 search_start;
 184         u64 extents_found;
 185
 186         unsigned int stage:8;
 187         unsigned int create_reloc_tree:1;
 188         unsigned int merge_reloc_tree:1;
 189         unsigned int found_file_extent:1;
 190 };
 191
 192 /* stages of data relocation */
 193 #define MOVE_DATA_EXTENTS       0
 194 #define UPDATE_DATA_PTRS        1
 195
 196 static void remove_backref_node(struct backref_cache *cache,
 197                                 struct backref_node *node);
 198 static void __mark_block_processed(struct reloc_control *rc,
 199                                    struct backref_node *node);
 200
 201 static void mapping_tree_init(struct mapping_tree *tree)
 202 {
 203         tree->rb_root = RB_ROOT;
 204         spin_lock_init(&tree->lock);
 205 }
 206
 207 static void backref_cache_init(struct backref_cache *cache)
 208 {
 209         int i;
 210         cache->rb_root = RB_ROOT;
 211         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 212                 INIT_LIST_HEAD(&cache->pending[i]);
 213         INIT_LIST_HEAD(&cache->changed);
 214         INIT_LIST_HEAD(&cache->detached);
 215         INIT_LIST_HEAD(&cache->leaves);
 216 }
 217
 218 static void backref_cache_cleanup(struct backref_cache *cache)
 219 {
 220         struct backref_node *node;
 221         int i;
 222
 223         while (!list_empty(&cache->detached)) {
 224                 node = list_entry(cache->detached.next,
 225                                   struct backref_node, list);
 226                 remove_backref_node(cache, node);
 227         }
 228
 229         while (!list_empty(&cache->leaves)) {
 230                 node = list_entry(cache->leaves.next,
 231                                   struct backref_node, lower);
 232                 remove_backref_node(cache, node);
 233         }
 234
 235         cache->last_trans = 0;
 236
 237         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 238                 ASSERT(list_empty(&cache->pending[i]));
 239         ASSERT(list_empty(&cache->changed));
 240         ASSERT(list_empty(&cache->detached));
 241         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
 242         ASSERT(!cache->nr_nodes);
 243         ASSERT(!cache->nr_edges);
 244 }
 245
 246 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 247 {
 248         struct backref_node *node;
 249
 250         node = kzalloc(sizeof(*node), GFP_NOFS);
 251         if (node) {
 252                 INIT_LIST_HEAD(&node->list);
 253                 INIT_LIST_HEAD(&node->upper);
 254                 INIT_LIST_HEAD(&node->lower);
 255                 RB_CLEAR_NODE(&node->rb_node);
 256                 cache->nr_nodes++;
 257         }
 258         return node;
 259 }
 260
 261 static void free_backref_node(struct backref_cache *cache,
 262                               struct backref_node *node)
 263 {
 264         if (node) {
 265                 cache->nr_nodes--;
 266                 kfree(node);
 267         }
 268 }
 269
 270 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 271 {
 272         struct backref_edge *edge;
 273
 274         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 275         if (edge)
 276                 cache->nr_edges++;
 277         return edge;
 278 }
 279
 280 static void free_backref_edge(struct backref_cache *cache,
 281                               struct backref_edge *edge)
 282 {
 283         if (edge) {
 284                 cache->nr_edges--;
 285                 kfree(edge);
 286         }
 287 }
 288
 289 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 290                                    struct rb_node *node)
 291 {
 292         struct rb_node **p = &root->rb_node;
 293         struct rb_node *parent = NULL;
 294         struct tree_entry *entry;
 295
 296         while (*p) {
 297                 parent = *p;
 298                 entry = rb_entry(parent, struct tree_entry, rb_node);
 299
 300                 if (bytenr < entry->bytenr)
 301                         p = &(*p)->rb_left;
 302                 else if (bytenr > entry->bytenr)
 303                         p = &(*p)->rb_right;
 304                 else
 305                         return parent;
 306         }
 307
 308         rb_link_node(node, parent, p);
 309         rb_insert_color(node, root);
 310         return NULL;
 311 }
 312
 313 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 314 {
 315         struct rb_node *n = root->rb_node;
 316         struct tree_entry *entry;
 317
 318         while (n) {
 319                 entry = rb_entry(n, struct tree_entry, rb_node);
 320
 321                 if (bytenr < entry->bytenr)
 322                         n = n->rb_left;
 323                 else if (bytenr > entry->bytenr)
 324                         n = n->rb_right;
 325                 else
 326                         return n;
 327         }
 328         return NULL;
 329 }
 330
 331 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 332 {
 333
 334         struct btrfs_fs_info *fs_info = NULL;
 335         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 336                                               rb_node);
 337         if (bnode->root)
 338                 fs_info = bnode->root->fs_info;
 339         btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
 340                     "found at offset %llu", bytenr);
 341 }
 342
 343 /*
 344  * walk up backref nodes until reach node presents tree root
 345  */
 346 static struct backref_node *walk_up_backref(struct backref_node *node,
 347                                             struct backref_edge *edges[],
 348                                             int *index)
 349 {
 350         struct backref_edge *edge;
 351         int idx = *index;
 352
 353         while (!list_empty(&node->upper)) {
 354                 edge = list_entry(node->upper.next,
 355                                   struct backref_edge, list[LOWER]);
 356                 edges[idx++] = edge;
 357                 node = edge->node[UPPER];
 358         }
 359         BUG_ON(node->detached);
 360         *index = idx;
 361         return node;
 362 }
 363
 364 /*
 365  * walk down backref nodes to find start of next reference path
 366  */
 367 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 368                                               int *index)
 369 {
 370         struct backref_edge *edge;
 371         struct backref_node *lower;
 372         int idx = *index;
 373
 374         while (idx > 0) {
 375                 edge = edges[idx - 1];
 376                 lower = edge->node[LOWER];
 377                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 378                         idx--;
 379                         continue;
 380                 }
 381                 edge = list_entry(edge->list[LOWER].next,
 382                                   struct backref_edge, list[LOWER]);
 383                 edges[idx - 1] = edge;
 384                 *index = idx;
 385                 return edge->node[UPPER];
 386         }
 387         *index = 0;
 388         return NULL;
 389 }
 390
 391 static void unlock_node_buffer(struct backref_node *node)
 392 {
 393         if (node->locked) {
 394                 btrfs_tree_unlock(node->eb);
 395                 node->locked = 0;
 396         }
 397 }
 398
 399 static void drop_node_buffer(struct backref_node *node)
 400 {
 401         if (node->eb) {
 402                 unlock_node_buffer(node);
 403                 free_extent_buffer(node->eb);
 404                 node->eb = NULL;
 405         }
 406 }
 407
 408 static void drop_backref_node(struct backref_cache *tree,
 409                               struct backref_node *node)
 410 {
 411         BUG_ON(!list_empty(&node->upper));
 412
 413         drop_node_buffer(node);
 414         list_del(&node->list);
 415         list_del(&node->lower);
 416         if (!RB_EMPTY_NODE(&node->rb_node))
 417                 rb_erase(&node->rb_node, &tree->rb_root);
 418         free_backref_node(tree, node);
 419 }
 420
 421 /*
 422  * remove a backref node from the backref cache
 423  */
 424 static void remove_backref_node(struct backref_cache *cache,
 425                                 struct backref_node *node)
 426 {
 427         struct backref_node *upper;
 428         struct backref_edge *edge;
 429
 430         if (!node)
 431                 return;
 432
 433         BUG_ON(!node->lowest && !node->detached);
 434         while (!list_empty(&node->upper)) {
 435                 edge = list_entry(node->upper.next, struct backref_edge,
 436                                   list[LOWER]);
 437                 upper = edge->node[UPPER];
 438                 list_del(&edge->list[LOWER]);
 439                 list_del(&edge->list[UPPER]);
 440                 free_backref_edge(cache, edge);
 441
 442                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 443                         BUG_ON(!list_empty(&node->upper));
 444                         drop_backref_node(cache, node);
 445                         node = upper;
 446                         node->lowest = 1;
 447                         continue;
 448                 }
 449                 /*
 450                  * add the node to leaf node list if no other
 451                  * child block cached.
 452                  */
 453                 if (list_empty(&upper->lower)) {
 454                         list_add_tail(&upper->lower, &cache->leaves);
 455                         upper->lowest = 1;
 456                 }
 457         }
 458
 459         drop_backref_node(cache, node);
 460 }
 461
 462 static void update_backref_node(struct backref_cache *cache,
 463                                 struct backref_node *node, u64 bytenr)
 464 {
 465         struct rb_node *rb_node;
 466         rb_erase(&node->rb_node, &cache->rb_root);
 467         node->bytenr = bytenr;
 468         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 469         if (rb_node)
 470                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 471 }
 472
 473 /*
 474  * update backref cache after a transaction commit
 475  */
 476 static int update_backref_cache(struct btrfs_trans_handle *trans,
 477                                 struct backref_cache *cache)
 478 {
 479         struct backref_node *node;
 480         int level = 0;
 481
 482         if (cache->last_trans == 0) {
 483                 cache->last_trans = trans->transid;
 484                 return 0;
 485         }
 486
 487         if (cache->last_trans == trans->transid)
 488                 return 0;
 489
 490         /*
 491          * detached nodes are used to avoid unnecessary backref
 492          * lookup. transaction commit changes the extent tree.
 493          * so the detached nodes are no longer useful.
 494          */
 495         while (!list_empty(&cache->detached)) {
 496                 node = list_entry(cache->detached.next,
 497                                   struct backref_node, list);
 498                 remove_backref_node(cache, node);
 499         }
 500
 501         while (!list_empty(&cache->changed)) {
 502                 node = list_entry(cache->changed.next,
 503                                   struct backref_node, list);
 504                 list_del_init(&node->list);
 505                 BUG_ON(node->pending);
 506                 update_backref_node(cache, node, node->new_bytenr);
 507         }
 508
 509         /*
 510          * some nodes can be left in the pending list if there were
 511          * errors during processing the pending nodes.
 512          */
 513         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 514                 list_for_each_entry(node, &cache->pending[level], list) {
 515                         BUG_ON(!node->pending);
 516                         if (node->bytenr == node->new_bytenr)
 517                                 continue;
 518                         update_backref_node(cache, node, node->new_bytenr);
 519                 }
 520         }
 521
 522         cache->last_trans = 0;
 523         return 1;
 524 }
 525
 526
 527 static int should_ignore_root(struct btrfs_root *root)
 528 {
 529         struct btrfs_root *reloc_root;
 530
 531         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 532                 return 0;
 533
 534         reloc_root = root->reloc_root;
 535         if (!reloc_root)
 536                 return 0;
 537
 538         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 539             root->fs_info->running_transaction->transid - 1)
 540                 return 0;
 541         /*
 542          * if there is reloc tree and it was created in previous
 543          * transaction backref lookup can find the reloc tree,
 544          * so backref node for the fs tree root is useless for
 545          * relocation.
 546          */
 547         return 1;
 548 }
 549 /*
 550  * find reloc tree by address of tree root
 551  */
 552 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 553                                           u64 bytenr)
 554 {
 555         struct rb_node *rb_node;
 556         struct mapping_node *node;
 557         struct btrfs_root *root = NULL;
 558
 559         spin_lock(&rc->reloc_root_tree.lock);
 560         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 561         if (rb_node) {
 562                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 563                 root = (struct btrfs_root *)node->data;
 564         }
 565         spin_unlock(&rc->reloc_root_tree.lock);
 566         return root;
 567 }
 568
 569 static int is_cowonly_root(u64 root_objectid)
 570 {
 571         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 572             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 573             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 574             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 575             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 576             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 577             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 578             root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
 579             root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
 580                 return 1;
 581         return 0;
 582 }
 583
 584 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 585                                         u64 root_objectid)
 586 {
 587         struct btrfs_key key;
 588
 589         key.objectid = root_objectid;
 590         key.type = BTRFS_ROOT_ITEM_KEY;
 591         if (is_cowonly_root(root_objectid))
 592                 key.offset = 0;
 593         else
 594                 key.offset = (u64)-1;
 595
 596         return btrfs_get_fs_root(fs_info, &key, false);
 597 }
 598
 599 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 600 static noinline_for_stack
 601 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 602                                   struct extent_buffer *leaf,
 603                                   struct btrfs_extent_ref_v0 *ref0)
 604 {
 605         struct btrfs_root *root;
 606         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 607         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 608
 609         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 610
 611         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 612         BUG_ON(IS_ERR(root));
 613
 614         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 615             generation != btrfs_root_generation(&root->root_item))
 616                 return NULL;
 617
 618         return root;
 619 }
 620 #endif
 621
 622 static noinline_for_stack
 623 int find_inline_backref(struct extent_buffer *leaf, int slot,
 624                         unsigned long *ptr, unsigned long *end)
 625 {
 626         struct btrfs_key key;
 627         struct btrfs_extent_item *ei;
 628         struct btrfs_tree_block_info *bi;
 629         u32 item_size;
 630
 631         btrfs_item_key_to_cpu(leaf, &key, slot);
 632
 633         item_size = btrfs_item_size_nr(leaf, slot);
 634 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 635         if (item_size < sizeof(*ei)) {
 636                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 637                 return 1;
 638         }
 639 #endif
 640         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 641         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 642                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 643
 644         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 645             item_size <= sizeof(*ei) + sizeof(*bi)) {
 646                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 647                 return 1;
 648         }
 649         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 650             item_size <= sizeof(*ei)) {
 651                 WARN_ON(item_size < sizeof(*ei));
 652                 return 1;
 653         }
 654
 655         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 656                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 657                 *ptr = (unsigned long)(bi + 1);
 658         } else {
 659                 *ptr = (unsigned long)(ei + 1);
 660         }
 661         *end = (unsigned long)ei + item_size;
 662         return 0;
 663 }
 664
 665 /*
 666  * build backref tree for a given tree block. root of the backref tree
 667  * corresponds the tree block, leaves of the backref tree correspond
 668  * roots of b-trees that reference the tree block.
 669  *
 670  * the basic idea of this function is check backrefs of a given block
 671  * to find upper level blocks that reference the block, and then check
 672  * backrefs of these upper level blocks recursively. the recursion stop
 673  * when tree root is reached or backrefs for the block is cached.
 674  *
 675  * NOTE: if we find backrefs for a block are cached, we know backrefs
 676  * for all upper level blocks that directly/indirectly reference the
 677  * block are also cached.
 678  */
 679 static noinline_for_stack
 680 struct backref_node *build_backref_tree(struct reloc_control *rc,
 681                                         struct btrfs_key *node_key,
 682                                         int level, u64 bytenr)
 683 {
 684         struct backref_cache *cache = &rc->backref_cache;
 685         struct btrfs_path *path1;
 686         struct btrfs_path *path2;
 687         struct extent_buffer *eb;
 688         struct btrfs_root *root;
 689         struct backref_node *cur;
 690         struct backref_node *upper;
 691         struct backref_node *lower;
 692         struct backref_node *node = NULL;
 693         struct backref_node *exist = NULL;
 694         struct backref_edge *edge;
 695         struct rb_node *rb_node;
 696         struct btrfs_key key;
 697         unsigned long end;
 698         unsigned long ptr;
 699         LIST_HEAD(list);
 700         LIST_HEAD(useless);
 701         int cowonly;
 702         int ret;
 703         int err = 0;
 704         bool need_check = true;
 705
 706         path1 = btrfs_alloc_path();
 707         path2 = btrfs_alloc_path();
 708         if (!path1 || !path2) {
 709                 err = -ENOMEM;
 710                 goto out;
 711         }
 712         path1->reada = READA_FORWARD;
 713         path2->reada = READA_FORWARD;
 714
 715         node = alloc_backref_node(cache);
 716         if (!node) {
 717                 err = -ENOMEM;
 718                 goto out;
 719         }
 720
 721         node->bytenr = bytenr;
 722         node->level = level;
 723         node->lowest = 1;
 724         cur = node;
 725 again:
 726         end = 0;
 727         ptr = 0;
 728         key.objectid = cur->bytenr;
 729         key.type = BTRFS_METADATA_ITEM_KEY;
 730         key.offset = (u64)-1;
 731
 732         path1->search_commit_root = 1;
 733         path1->skip_locking = 1;
 734         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 735                                 0, 0);
 736         if (ret < 0) {
 737                 err = ret;
 738                 goto out;
 739         }
 740         ASSERT(ret);
 741         ASSERT(path1->slots[0]);
 742
 743         path1->slots[0]--;
 744
 745         WARN_ON(cur->checked);
 746         if (!list_empty(&cur->upper)) {
 747                 /*
 748                  * the backref was added previously when processing
 749                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 750                  */
 751                 ASSERT(list_is_singular(&cur->upper));
 752                 edge = list_entry(cur->upper.next, struct backref_edge,
 753                                   list[LOWER]);
 754                 ASSERT(list_empty(&edge->list[UPPER]));
 755                 exist = edge->node[UPPER];
 756                 /*
 757                  * add the upper level block to pending list if we need
 758                  * check its backrefs
 759                  */
 760                 if (!exist->checked)
 761                         list_add_tail(&edge->list[UPPER], &list);
 762         } else {
 763                 exist = NULL;
 764         }
 765
 766         while (1) {
 767                 cond_resched();
 768                 eb = path1->nodes[0];
 769
 770                 if (ptr >= end) {
 771                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 772                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 773                                 if (ret < 0) {
 774                                         err = ret;
 775                                         goto out;
 776                                 }
 777                                 if (ret > 0)
 778                                         break;
 779                                 eb = path1->nodes[0];
 780                         }
 781
 782                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 783                         if (key.objectid != cur->bytenr) {
 784                                 WARN_ON(exist);
 785                                 break;
 786                         }
 787
 788                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 789                             key.type == BTRFS_METADATA_ITEM_KEY) {
 790                                 ret = find_inline_backref(eb, path1->slots[0],
 791                                                           &ptr, &end);
 792                                 if (ret)
 793                                         goto next;
 794                         }
 795                 }
 796
 797                 if (ptr < end) {
 798                         /* update key for inline back ref */
 799                         struct btrfs_extent_inline_ref *iref;
 800                         iref = (struct btrfs_extent_inline_ref *)ptr;
 801                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 802                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 803                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 804                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 805                 }
 806
 807                 if (exist &&
 808                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 809                       exist->owner == key.offset) ||
 810                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 811                       exist->bytenr == key.offset))) {
 812                         exist = NULL;
 813                         goto next;
 814                 }
 815
 816 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 817                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 818                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 819                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 820                                 struct btrfs_extent_ref_v0 *ref0;
 821                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 822                                                 struct btrfs_extent_ref_v0);
 823                                 if (key.objectid == key.offset) {
 824                                         root = find_tree_root(rc, eb, ref0);
 825                                         if (root && !should_ignore_root(root))
 826                                                 cur->root = root;
 827                                         else
 828                                                 list_add(&cur->list, &useless);
 829                                         break;
 830                                 }
 831                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 832                                                                       ref0)))
 833                                         cur->cowonly = 1;
 834                         }
 835 #else
 836                 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
 837                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 838 #endif
 839                         if (key.objectid == key.offset) {
 840                                 /*
 841                                  * only root blocks of reloc trees use
 842                                  * backref of this type.
 843                                  */
 844                                 root = find_reloc_root(rc, cur->bytenr);
 845                                 ASSERT(root);
 846                                 cur->root = root;
 847                                 break;
 848                         }
 849
 850                         edge = alloc_backref_edge(cache);
 851                         if (!edge) {
 852                                 err = -ENOMEM;
 853                                 goto out;
 854                         }
 855                         rb_node = tree_search(&cache->rb_root, key.offset);
 856                         if (!rb_node) {
 857                                 upper = alloc_backref_node(cache);
 858                                 if (!upper) {
 859                                         free_backref_edge(cache, edge);
 860                                         err = -ENOMEM;
 861                                         goto out;
 862                                 }
 863                                 upper->bytenr = key.offset;
 864                                 upper->level = cur->level + 1;
 865                                 /*
 866                                  *  backrefs for the upper level block isn't
 867                                  *  cached, add the block to pending list
 868                                  */
 869                                 list_add_tail(&edge->list[UPPER], &list);
 870                         } else {
 871                                 upper = rb_entry(rb_node, struct backref_node,
 872                                                  rb_node);
 873                                 ASSERT(upper->checked);
 874                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 875                         }
 876                         list_add_tail(&edge->list[LOWER], &cur->upper);
 877                         edge->node[LOWER] = cur;
 878                         edge->node[UPPER] = upper;
 879
 880                         goto next;
 881                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 882                         goto next;
 883                 }
 884
 885                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 886                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 887                 if (IS_ERR(root)) {
 888                         err = PTR_ERR(root);
 889                         goto out;
 890                 }
 891
 892                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 893                         cur->cowonly = 1;
 894
 895                 if (btrfs_root_level(&root->root_item) == cur->level) {
 896                         /* tree root */
 897                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
 898                                cur->bytenr);
 899                         if (should_ignore_root(root))
 900                                 list_add(&cur->list, &useless);
 901                         else
 902                                 cur->root = root;
 903                         break;
 904                 }
 905
 906                 level = cur->level + 1;
 907
 908                 /*
 909                  * searching the tree to find upper level blocks
 910                  * reference the block.
 911                  */
 912                 path2->search_commit_root = 1;
 913                 path2->skip_locking = 1;
 914                 path2->lowest_level = level;
 915                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 916                 path2->lowest_level = 0;
 917                 if (ret < 0) {
 918                         err = ret;
 919                         goto out;
 920                 }
 921                 if (ret > 0 && path2->slots[level] > 0)
 922                         path2->slots[level]--;
 923
 924                 eb = path2->nodes[level];
 925                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 926                         cur->bytenr);
 927
 928                 lower = cur;
 929                 need_check = true;
 930                 for (; level < BTRFS_MAX_LEVEL; level++) {
 931                         if (!path2->nodes[level]) {
 932                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
 933                                        lower->bytenr);
 934                                 if (should_ignore_root(root))
 935                                         list_add(&lower->list, &useless);
 936                                 else
 937                                         lower->root = root;
 938                                 break;
 939                         }
 940
 941                         edge = alloc_backref_edge(cache);
 942                         if (!edge) {
 943                                 err = -ENOMEM;
 944                                 goto out;
 945                         }
 946
 947                         eb = path2->nodes[level];
 948                         rb_node = tree_search(&cache->rb_root, eb->start);
 949                         if (!rb_node) {
 950                                 upper = alloc_backref_node(cache);
 951                                 if (!upper) {
 952                                         free_backref_edge(cache, edge);
 953                                         err = -ENOMEM;
 954                                         goto out;
 955                                 }
 956                                 upper->bytenr = eb->start;
 957                                 upper->owner = btrfs_header_owner(eb);
 958                                 upper->level = lower->level + 1;
 959                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
 960                                               &root->state))
 961                                         upper->cowonly = 1;
 962
 963                                 /*
 964                                  * if we know the block isn't shared
 965                                  * we can void checking its backrefs.
 966                                  */
 967                                 if (btrfs_block_can_be_shared(root, eb))
 968                                         upper->checked = 0;
 969                                 else
 970                                         upper->checked = 1;
 971
 972                                 /*
 973                                  * add the block to pending list if we
 974                                  * need check its backrefs, we only do this once
 975                                  * while walking up a tree as we will catch
 976                                  * anything else later on.
 977                                  */
 978                                 if (!upper->checked && need_check) {
 979                                         need_check = false;
 980                                         list_add_tail(&edge->list[UPPER],
 981                                                       &list);
 982                                 } else {
 983                                         if (upper->checked)
 984                                                 need_check = true;
 985                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 986                                 }
 987                         } else {
 988                                 upper = rb_entry(rb_node, struct backref_node,
 989                                                  rb_node);
 990                                 ASSERT(upper->checked);
 991                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 992                                 if (!upper->owner)
 993                                         upper->owner = btrfs_header_owner(eb);
 994                         }
 995                         list_add_tail(&edge->list[LOWER], &lower->upper);
 996                         edge->node[LOWER] = lower;
 997                         edge->node[UPPER] = upper;
 998
 999                         if (rb_node)
1000                                 break;
1001                         lower = upper;
1002                         upper = NULL;
1003                 }
1004                 btrfs_release_path(path2);
1005 next:
1006                 if (ptr < end) {
1007                         ptr += btrfs_extent_inline_ref_size(key.type);
1008                         if (ptr >= end) {
1009                                 WARN_ON(ptr > end);
1010                                 ptr = 0;
1011                                 end = 0;
1012                         }
1013                 }
1014                 if (ptr >= end)
1015                         path1->slots[0]++;
1016         }
1017         btrfs_release_path(path1);
1018
1019         cur->checked = 1;
1020         WARN_ON(exist);
1021
1022         /* the pending list isn't empty, take the first block to process */
1023         if (!list_empty(&list)) {
1024                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1025                 list_del_init(&edge->list[UPPER]);
1026                 cur = edge->node[UPPER];
1027                 goto again;
1028         }
1029
1030         /*
1031          * everything goes well, connect backref nodes and insert backref nodes
1032          * into the cache.
1033          */
1034         ASSERT(node->checked);
1035         cowonly = node->cowonly;
1036         if (!cowonly) {
1037                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1038                                       &node->rb_node);
1039                 if (rb_node)
1040                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1041                 list_add_tail(&node->lower, &cache->leaves);
1042         }
1043
1044         list_for_each_entry(edge, &node->upper, list[LOWER])
1045                 list_add_tail(&edge->list[UPPER], &list);
1046
1047         while (!list_empty(&list)) {
1048                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1049                 list_del_init(&edge->list[UPPER]);
1050                 upper = edge->node[UPPER];
1051                 if (upper->detached) {
1052                         list_del(&edge->list[LOWER]);
1053                         lower = edge->node[LOWER];
1054                         free_backref_edge(cache, edge);
1055                         if (list_empty(&lower->upper))
1056                                 list_add(&lower->list, &useless);
1057                         continue;
1058                 }
1059
1060                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1061                         if (upper->lowest) {
1062                                 list_del_init(&upper->lower);
1063                                 upper->lowest = 0;
1064                         }
1065
1066                         list_add_tail(&edge->list[UPPER], &upper->lower);
1067                         continue;
1068                 }
1069
1070                 if (!upper->checked) {
1071                         /*
1072                          * Still want to blow up for developers since this is a
1073                          * logic bug.
1074                          */
1075                         ASSERT(0);
1076                         err = -EINVAL;
1077                         goto out;
1078                 }
1079                 if (cowonly != upper->cowonly) {
1080                         ASSERT(0);
1081                         err = -EINVAL;
1082                         goto out;
1083                 }
1084
1085                 if (!cowonly) {
1086                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1087                                               &upper->rb_node);
1088                         if (rb_node)
1089                                 backref_tree_panic(rb_node, -EEXIST,
1090                                                    upper->bytenr);
1091                 }
1092
1093                 list_add_tail(&edge->list[UPPER], &upper->lower);
1094
1095                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1096                         list_add_tail(&edge->list[UPPER], &list);
1097         }
1098         /*
1099          * process useless backref nodes. backref nodes for tree leaves
1100          * are deleted from the cache. backref nodes for upper level
1101          * tree blocks are left in the cache to avoid unnecessary backref
1102          * lookup.
1103          */
1104         while (!list_empty(&useless)) {
1105                 upper = list_entry(useless.next, struct backref_node, list);
1106                 list_del_init(&upper->list);
1107                 ASSERT(list_empty(&upper->upper));
1108                 if (upper == node)
1109                         node = NULL;
1110                 if (upper->lowest) {
1111                         list_del_init(&upper->lower);
1112                         upper->lowest = 0;
1113                 }
1114                 while (!list_empty(&upper->lower)) {
1115                         edge = list_entry(upper->lower.next,
1116                                           struct backref_edge, list[UPPER]);
1117                         list_del(&edge->list[UPPER]);
1118                         list_del(&edge->list[LOWER]);
1119                         lower = edge->node[LOWER];
1120                         free_backref_edge(cache, edge);
1121
1122                         if (list_empty(&lower->upper))
1123                                 list_add(&lower->list, &useless);
1124                 }
1125                 __mark_block_processed(rc, upper);
1126                 if (upper->level > 0) {
1127                         list_add(&upper->list, &cache->detached);
1128                         upper->detached = 1;
1129                 } else {
1130                         rb_erase(&upper->rb_node, &cache->rb_root);
1131                         free_backref_node(cache, upper);
1132                 }
1133         }
1134 out:
1135         btrfs_free_path(path1);
1136         btrfs_free_path(path2);
1137         if (err) {
1138                 while (!list_empty(&useless)) {
1139                         lower = list_entry(useless.next,
1140                                            struct backref_node, list);
1141                         list_del_init(&lower->list);
1142                 }
1143                 while (!list_empty(&list)) {
1144                         edge = list_first_entry(&list, struct backref_edge,
1145                                                 list[UPPER]);
1146                         list_del(&edge->list[UPPER]);
1147                         list_del(&edge->list[LOWER]);
1148                         lower = edge->node[LOWER];
1149                         upper = edge->node[UPPER];
1150                         free_backref_edge(cache, edge);
1151
1152                         /*
1153                          * Lower is no longer linked to any upper backref nodes
1154                          * and isn't in the cache, we can free it ourselves.
1155                          */
1156                         if (list_empty(&lower->upper) &&
1157                             RB_EMPTY_NODE(&lower->rb_node))
1158                                 list_add(&lower->list, &useless);
1159
1160                         if (!RB_EMPTY_NODE(&upper->rb_node))
1161                                 continue;
1162
1163                         /* Add this guy's upper edges to the list to process */
1164                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1165                                 list_add_tail(&edge->list[UPPER], &list);
1166                         if (list_empty(&upper->upper))
1167                                 list_add(&upper->list, &useless);
1168                 }
1169
1170                 while (!list_empty(&useless)) {
1171                         lower = list_entry(useless.next,
1172                                            struct backref_node, list);
1173                         list_del_init(&lower->list);
1174                         if (lower == node)
1175                                 node = NULL;
1176                         free_backref_node(cache, lower);
1177                 }
1178
1179                 free_backref_node(cache, node);
1180                 return ERR_PTR(err);
1181         }
1182         ASSERT(!node || !node->detached);
1183         return node;
1184 }
1185
1186 /*
1187  * helper to add backref node for the newly created snapshot.
1188  * the backref node is created by cloning backref node that
1189  * corresponds to root of source tree
1190  */
1191 static int clone_backref_node(struct btrfs_trans_handle *trans,
1192                               struct reloc_control *rc,
1193                               struct btrfs_root *src,
1194                               struct btrfs_root *dest)
1195 {
1196         struct btrfs_root *reloc_root = src->reloc_root;
1197         struct backref_cache *cache = &rc->backref_cache;
1198         struct backref_node *node = NULL;
1199         struct backref_node *new_node;
1200         struct backref_edge *edge;
1201         struct backref_edge *new_edge;
1202         struct rb_node *rb_node;
1203
1204         if (cache->last_trans > 0)
1205                 update_backref_cache(trans, cache);
1206
1207         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1208         if (rb_node) {
1209                 node = rb_entry(rb_node, struct backref_node, rb_node);
1210                 if (node->detached)
1211                         node = NULL;
1212                 else
1213                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1214         }
1215
1216         if (!node) {
1217                 rb_node = tree_search(&cache->rb_root,
1218                                       reloc_root->commit_root->start);
1219                 if (rb_node) {
1220                         node = rb_entry(rb_node, struct backref_node,
1221                                         rb_node);
1222                         BUG_ON(node->detached);
1223                 }
1224         }
1225
1226         if (!node)
1227                 return 0;
1228
1229         new_node = alloc_backref_node(cache);
1230         if (!new_node)
1231                 return -ENOMEM;
1232
1233         new_node->bytenr = dest->node->start;
1234         new_node->level = node->level;
1235         new_node->lowest = node->lowest;
1236         new_node->checked = 1;
1237         new_node->root = dest;
1238
1239         if (!node->lowest) {
1240                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1241                         new_edge = alloc_backref_edge(cache);
1242                         if (!new_edge)
1243                                 goto fail;
1244
1245                         new_edge->node[UPPER] = new_node;
1246                         new_edge->node[LOWER] = edge->node[LOWER];
1247                         list_add_tail(&new_edge->list[UPPER],
1248                                       &new_node->lower);
1249                 }
1250         } else {
1251                 list_add_tail(&new_node->lower, &cache->leaves);
1252         }
1253
1254         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1255                               &new_node->rb_node);
1256         if (rb_node)
1257                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1258
1259         if (!new_node->lowest) {
1260                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1261                         list_add_tail(&new_edge->list[LOWER],
1262                                       &new_edge->node[LOWER]->upper);
1263                 }
1264         }
1265         return 0;
1266 fail:
1267         while (!list_empty(&new_node->lower)) {
1268                 new_edge = list_entry(new_node->lower.next,
1269                                       struct backref_edge, list[UPPER]);
1270                 list_del(&new_edge->list[UPPER]);
1271                 free_backref_edge(cache, new_edge);
1272         }
1273         free_backref_node(cache, new_node);
1274         return -ENOMEM;
1275 }
1276
1277 /*
1278  * helper to add 'address of tree root -> reloc tree' mapping
1279  */
1280 static int __must_check __add_reloc_root(struct btrfs_root *root)
1281 {
1282         struct rb_node *rb_node;
1283         struct mapping_node *node;
1284         struct reloc_control *rc = root->fs_info->reloc_ctl;
1285
1286         node = kmalloc(sizeof(*node), GFP_NOFS);
1287         if (!node)
1288                 return -ENOMEM;
1289
1290         node->bytenr = root->node->start;
1291         node->data = root;
1292
1293         spin_lock(&rc->reloc_root_tree.lock);
1294         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1295                               node->bytenr, &node->rb_node);
1296         spin_unlock(&rc->reloc_root_tree.lock);
1297         if (rb_node) {
1298                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1299                             "for start=%llu while inserting into relocation "
1300                             "tree", node->bytenr);
1301                 kfree(node);
1302                 return -EEXIST;
1303         }
1304
1305         list_add_tail(&root->root_list, &rc->reloc_roots);
1306         return 0;
1307 }
1308
1309 /*
1310  * helper to delete the 'address of tree root -> reloc tree'
1311  * mapping
1312  */
1313 static void __del_reloc_root(struct btrfs_root *root)
1314 {
1315         struct rb_node *rb_node;
1316         struct mapping_node *node = NULL;
1317         struct reloc_control *rc = root->fs_info->reloc_ctl;
1318
1319         spin_lock(&rc->reloc_root_tree.lock);
1320         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1321                               root->node->start);
1322         if (rb_node) {
1323                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1324                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1325         }
1326         spin_unlock(&rc->reloc_root_tree.lock);
1327
1328         if (!node)
1329                 return;
1330         BUG_ON((struct btrfs_root *)node->data != root);
1331
1332         spin_lock(&root->fs_info->trans_lock);
1333         list_del_init(&root->root_list);
1334         spin_unlock(&root->fs_info->trans_lock);
1335         kfree(node);
1336 }
1337
1338 /*
1339  * helper to update the 'address of tree root -> reloc tree'
1340  * mapping
1341  */
1342 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1343 {
1344         struct rb_node *rb_node;
1345         struct mapping_node *node = NULL;
1346         struct reloc_control *rc = root->fs_info->reloc_ctl;
1347
1348         spin_lock(&rc->reloc_root_tree.lock);
1349         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1350                               root->node->start);
1351         if (rb_node) {
1352                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1353                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1354         }
1355         spin_unlock(&rc->reloc_root_tree.lock);
1356
1357         if (!node)
1358                 return 0;
1359         BUG_ON((struct btrfs_root *)node->data != root);
1360
1361         spin_lock(&rc->reloc_root_tree.lock);
1362         node->bytenr = new_bytenr;
1363         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1364                               node->bytenr, &node->rb_node);
1365         spin_unlock(&rc->reloc_root_tree.lock);
1366         if (rb_node)
1367                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1368         return 0;
1369 }
1370
1371 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1372                                         struct btrfs_root *root, u64 objectid)
1373 {
1374         struct btrfs_root *reloc_root;
1375         struct extent_buffer *eb;
1376         struct btrfs_root_item *root_item;
1377         struct btrfs_key root_key;
1378         u64 last_snap = 0;
1379         int ret;
1380
1381         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1382         BUG_ON(!root_item);
1383
1384         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1385         root_key.type = BTRFS_ROOT_ITEM_KEY;
1386         root_key.offset = objectid;
1387
1388         if (root->root_key.objectid == objectid) {
1389                 /* called by btrfs_init_reloc_root */
1390                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1391                                       BTRFS_TREE_RELOC_OBJECTID);
1392                 BUG_ON(ret);
1393
1394                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1395                 btrfs_set_root_last_snapshot(&root->root_item,
1396                                              trans->transid - 1);
1397         } else {
1398                 /*
1399                  * called by btrfs_reloc_post_snapshot_hook.
1400                  * the source tree is a reloc tree, all tree blocks
1401                  * modified after it was created have RELOC flag
1402                  * set in their headers. so it's OK to not update
1403                  * the 'last_snapshot'.
1404                  */
1405                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1406                                       BTRFS_TREE_RELOC_OBJECTID);
1407                 BUG_ON(ret);
1408         }
1409
1410         memcpy(root_item, &root->root_item, sizeof(*root_item));
1411         btrfs_set_root_bytenr(root_item, eb->start);
1412         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1413         btrfs_set_root_generation(root_item, trans->transid);
1414
1415         if (root->root_key.objectid == objectid) {
1416                 btrfs_set_root_refs(root_item, 0);
1417                 memset(&root_item->drop_progress, 0,
1418                        sizeof(struct btrfs_disk_key));
1419                 root_item->drop_level = 0;
1420                 /*
1421                  * abuse rtransid, it is safe because it is impossible to
1422                  * receive data into a relocation tree.
1423                  */
1424                 btrfs_set_root_rtransid(root_item, last_snap);
1425                 btrfs_set_root_otransid(root_item, trans->transid);
1426         }
1427
1428         btrfs_tree_unlock(eb);
1429         free_extent_buffer(eb);
1430
1431         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1432                                 &root_key, root_item);
1433         BUG_ON(ret);
1434         kfree(root_item);
1435
1436         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1437         BUG_ON(IS_ERR(reloc_root));
1438         reloc_root->last_trans = trans->transid;
1439         return reloc_root;
1440 }
1441
1442 /*
1443  * create reloc tree for a given fs tree. reloc tree is just a
1444  * snapshot of the fs tree with special root objectid.
1445  */
1446 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1447                           struct btrfs_root *root)
1448 {
1449         struct btrfs_root *reloc_root;
1450         struct reloc_control *rc = root->fs_info->reloc_ctl;
1451         struct btrfs_block_rsv *rsv;
1452         int clear_rsv = 0;
1453         int ret;
1454
1455         if (root->reloc_root) {
1456                 reloc_root = root->reloc_root;
1457                 reloc_root->last_trans = trans->transid;
1458                 return 0;
1459         }
1460
1461         if (!rc || !rc->create_reloc_tree ||
1462             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1463                 return 0;
1464
1465         if (!trans->reloc_reserved) {
1466                 rsv = trans->block_rsv;
1467                 trans->block_rsv = rc->block_rsv;
1468                 clear_rsv = 1;
1469         }
1470         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1471         if (clear_rsv)
1472                 trans->block_rsv = rsv;
1473
1474         ret = __add_reloc_root(reloc_root);
1475         BUG_ON(ret < 0);
1476         root->reloc_root = reloc_root;
1477         return 0;
1478 }
1479
1480 /*
1481  * update root item of reloc tree
1482  */
1483 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1484                             struct btrfs_root *root)
1485 {
1486         struct btrfs_root *reloc_root;
1487         struct btrfs_root_item *root_item;
1488         int ret;
1489
1490         if (!root->reloc_root)
1491                 goto out;
1492
1493         reloc_root = root->reloc_root;
1494         root_item = &reloc_root->root_item;
1495
1496         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1497             btrfs_root_refs(root_item) == 0) {
1498                 root->reloc_root = NULL;
1499                 __del_reloc_root(reloc_root);
1500         }
1501
1502         if (reloc_root->commit_root != reloc_root->node) {
1503                 btrfs_set_root_node(root_item, reloc_root->node);
1504                 free_extent_buffer(reloc_root->commit_root);
1505                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1506         }
1507
1508         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1509                                 &reloc_root->root_key, root_item);
1510         BUG_ON(ret);
1511
1512 out:
1513         return 0;
1514 }
1515
1516 /*
1517  * helper to find first cached inode with inode number >= objectid
1518  * in a subvolume
1519  */
1520 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1521 {
1522         struct rb_node *node;
1523         struct rb_node *prev;
1524         struct btrfs_inode *entry;
1525         struct inode *inode;
1526
1527         spin_lock(&root->inode_lock);
1528 again:
1529         node = root->inode_tree.rb_node;
1530         prev = NULL;
1531         while (node) {
1532                 prev = node;
1533                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1534
1535                 if (objectid < btrfs_ino(&entry->vfs_inode))
1536                         node = node->rb_left;
1537                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1538                         node = node->rb_right;
1539                 else
1540                         break;
1541         }
1542         if (!node) {
1543                 while (prev) {
1544                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1545                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1546                                 node = prev;
1547                                 break;
1548                         }
1549                         prev = rb_next(prev);
1550                 }
1551         }
1552         while (node) {
1553                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1554                 inode = igrab(&entry->vfs_inode);
1555                 if (inode) {
1556                         spin_unlock(&root->inode_lock);
1557                         return inode;
1558                 }
1559
1560                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1561                 if (cond_resched_lock(&root->inode_lock))
1562                         goto again;
1563
1564                 node = rb_next(node);
1565         }
1566         spin_unlock(&root->inode_lock);
1567         return NULL;
1568 }
1569
1570 static int in_block_group(u64 bytenr,
1571                           struct btrfs_block_group_cache *block_group)
1572 {
1573         if (bytenr >= block_group->key.objectid &&
1574             bytenr < block_group->key.objectid + block_group->key.offset)
1575                 return 1;
1576         return 0;
1577 }
1578
1579 /*
1580  * get new location of data
1581  */
1582 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1583                             u64 bytenr, u64 num_bytes)
1584 {
1585         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1586         struct btrfs_path *path;
1587         struct btrfs_file_extent_item *fi;
1588         struct extent_buffer *leaf;
1589         int ret;
1590
1591         path = btrfs_alloc_path();
1592         if (!path)
1593                 return -ENOMEM;
1594
1595         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1596         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1597                                        bytenr, 0);
1598         if (ret < 0)
1599                 goto out;
1600         if (ret > 0) {
1601                 ret = -ENOENT;
1602                 goto out;
1603         }
1604
1605         leaf = path->nodes[0];
1606         fi = btrfs_item_ptr(leaf, path->slots[0],
1607                             struct btrfs_file_extent_item);
1608
1609         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1610                btrfs_file_extent_compression(leaf, fi) ||
1611                btrfs_file_extent_encryption(leaf, fi) ||
1612                btrfs_file_extent_other_encoding(leaf, fi));
1613
1614         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1615                 ret = -EINVAL;
1616                 goto out;
1617         }
1618
1619         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1620         ret = 0;
1621 out:
1622         btrfs_free_path(path);
1623         return ret;
1624 }
1625
1626 /*
1627  * update file extent items in the tree leaf to point to
1628  * the new locations.
1629  */
1630 static noinline_for_stack
1631 int replace_file_extents(struct btrfs_trans_handle *trans,
1632                          struct reloc_control *rc,
1633                          struct btrfs_root *root,
1634                          struct extent_buffer *leaf)
1635 {
1636         struct btrfs_key key;
1637         struct btrfs_file_extent_item *fi;
1638         struct inode *inode = NULL;
1639         u64 parent;
1640         u64 bytenr;
1641         u64 new_bytenr = 0;
1642         u64 num_bytes;
1643         u64 end;
1644         u32 nritems;
1645         u32 i;
1646         int ret = 0;
1647         int first = 1;
1648         int dirty = 0;
1649
1650         if (rc->stage != UPDATE_DATA_PTRS)
1651                 return 0;
1652
1653         /* reloc trees always use full backref */
1654         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1655                 parent = leaf->start;
1656         else
1657                 parent = 0;
1658
1659         nritems = btrfs_header_nritems(leaf);
1660         for (i = 0; i < nritems; i++) {
1661                 cond_resched();
1662                 btrfs_item_key_to_cpu(leaf, &key, i);
1663                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1664                         continue;
1665                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1666                 if (btrfs_file_extent_type(leaf, fi) ==
1667                     BTRFS_FILE_EXTENT_INLINE)
1668                         continue;
1669                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1670                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1671                 if (bytenr == 0)
1672                         continue;
1673                 if (!in_block_group(bytenr, rc->block_group))
1674                         continue;
1675
1676                 /*
1677                  * if we are modifying block in fs tree, wait for readpage
1678                  * to complete and drop the extent cache
1679                  */
1680                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1681                         if (first) {
1682                                 inode = find_next_inode(root, key.objectid);
1683                                 first = 0;
1684                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1685                                 btrfs_add_delayed_iput(inode);
1686                                 inode = find_next_inode(root, key.objectid);
1687                         }
1688                         if (inode && btrfs_ino(inode) == key.objectid) {
1689                                 end = key.offset +
1690                                       btrfs_file_extent_num_bytes(leaf, fi);
1691                                 WARN_ON(!IS_ALIGNED(key.offset,
1692                                                     root->sectorsize));
1693                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1694                                 end--;
1695                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1696                                                       key.offset, end);
1697                                 if (!ret)
1698                                         continue;
1699
1700                                 btrfs_drop_extent_cache(inode, key.offset, end,
1701                                                         1);
1702                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1703                                               key.offset, end);
1704                         }
1705                 }
1706
1707                 ret = get_new_location(rc->data_inode, &new_bytenr,
1708                                        bytenr, num_bytes);
1709                 if (ret) {
1710                         /*
1711                          * Don't have to abort since we've not changed anything
1712                          * in the file extent yet.
1713                          */
1714                         break;
1715                 }
1716
1717                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1718                 dirty = 1;
1719
1720                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1721                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1722                                            num_bytes, parent,
1723                                            btrfs_header_owner(leaf),
1724                                            key.objectid, key.offset);
1725                 if (ret) {
1726                         btrfs_abort_transaction(trans, ret);
1727                         break;
1728                 }
1729
1730                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1731                                         parent, btrfs_header_owner(leaf),
1732                                         key.objectid, key.offset);
1733                 if (ret) {
1734                         btrfs_abort_transaction(trans, ret);
1735                         break;
1736                 }
1737         }
1738         if (dirty)
1739                 btrfs_mark_buffer_dirty(leaf);
1740         if (inode)
1741                 btrfs_add_delayed_iput(inode);
1742         return ret;
1743 }
1744
1745 static noinline_for_stack
1746 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1747                      struct btrfs_path *path, int level)
1748 {
1749         struct btrfs_disk_key key1;
1750         struct btrfs_disk_key key2;
1751         btrfs_node_key(eb, &key1, slot);
1752         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1753         return memcmp(&key1, &key2, sizeof(key1));
1754 }
1755
1756 /*
1757  * try to replace tree blocks in fs tree with the new blocks
1758  * in reloc tree. tree blocks haven't been modified since the
1759  * reloc tree was create can be replaced.
1760  *
1761  * if a block was replaced, level of the block + 1 is returned.
1762  * if no block got replaced, 0 is returned. if there are other
1763  * errors, a negative error number is returned.
1764  */
1765 static noinline_for_stack
1766 int replace_path(struct btrfs_trans_handle *trans,
1767                  struct btrfs_root *dest, struct btrfs_root *src,
1768                  struct btrfs_path *path, struct btrfs_key *next_key,
1769                  int lowest_level, int max_level)
1770 {
1771         struct extent_buffer *eb;
1772         struct extent_buffer *parent;
1773         struct btrfs_key key;
1774         u64 old_bytenr;
1775         u64 new_bytenr;
1776         u64 old_ptr_gen;
1777         u64 new_ptr_gen;
1778         u64 last_snapshot;
1779         u32 blocksize;
1780         int cow = 0;
1781         int level;
1782         int ret;
1783         int slot;
1784
1785         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1786         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1787
1788         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1789 again:
1790         slot = path->slots[lowest_level];
1791         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1792
1793         eb = btrfs_lock_root_node(dest);
1794         btrfs_set_lock_blocking(eb);
1795         level = btrfs_header_level(eb);
1796
1797         if (level < lowest_level) {
1798                 btrfs_tree_unlock(eb);
1799                 free_extent_buffer(eb);
1800                 return 0;
1801         }
1802
1803         if (cow) {
1804                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1805                 BUG_ON(ret);
1806         }
1807         btrfs_set_lock_blocking(eb);
1808
1809         if (next_key) {
1810                 next_key->objectid = (u64)-1;
1811                 next_key->type = (u8)-1;
1812                 next_key->offset = (u64)-1;
1813         }
1814
1815         parent = eb;
1816         while (1) {
1817                 level = btrfs_header_level(parent);
1818                 BUG_ON(level < lowest_level);
1819
1820                 ret = btrfs_bin_search(parent, &key, level, &slot);
1821                 if (ret && slot > 0)
1822                         slot--;
1823
1824                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1825                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1826
1827                 old_bytenr = btrfs_node_blockptr(parent, slot);
1828                 blocksize = dest->nodesize;
1829                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1830
1831                 if (level <= max_level) {
1832                         eb = path->nodes[level];
1833                         new_bytenr = btrfs_node_blockptr(eb,
1834                                                         path->slots[level]);
1835                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1836                                                         path->slots[level]);
1837                 } else {
1838                         new_bytenr = 0;
1839                         new_ptr_gen = 0;
1840                 }
1841
1842                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1843                         ret = level;
1844                         break;
1845                 }
1846
1847                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1848                     memcmp_node_keys(parent, slot, path, level)) {
1849                         if (level <= lowest_level) {
1850                                 ret = 0;
1851                                 break;
1852                         }
1853
1854                         eb = read_tree_block(dest, old_bytenr, old_ptr_gen);
1855                         if (IS_ERR(eb)) {
1856                                 ret = PTR_ERR(eb);
1857                                 break;
1858                         } else if (!extent_buffer_uptodate(eb)) {
1859                                 ret = -EIO;
1860                                 free_extent_buffer(eb);
1861                                 break;
1862                         }
1863                         btrfs_tree_lock(eb);
1864                         if (cow) {
1865                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1866                                                       slot, &eb);
1867                                 BUG_ON(ret);
1868                         }
1869                         btrfs_set_lock_blocking(eb);
1870
1871                         btrfs_tree_unlock(parent);
1872                         free_extent_buffer(parent);
1873
1874                         parent = eb;
1875                         continue;
1876                 }
1877
1878                 if (!cow) {
1879                         btrfs_tree_unlock(parent);
1880                         free_extent_buffer(parent);
1881                         cow = 1;
1882                         goto again;
1883                 }
1884
1885                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1886                                       path->slots[level]);
1887                 btrfs_release_path(path);
1888
1889                 path->lowest_level = level;
1890                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1891                 path->lowest_level = 0;
1892                 BUG_ON(ret);
1893
1894                 /*
1895                  * swap blocks in fs tree and reloc tree.
1896                  */
1897                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1898                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1899                 btrfs_mark_buffer_dirty(parent);
1900
1901                 btrfs_set_node_blockptr(path->nodes[level],
1902                                         path->slots[level], old_bytenr);
1903                 btrfs_set_node_ptr_generation(path->nodes[level],
1904                                               path->slots[level], old_ptr_gen);
1905                 btrfs_mark_buffer_dirty(path->nodes[level]);
1906
1907                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1908                                         path->nodes[level]->start,
1909                                         src->root_key.objectid, level - 1, 0);
1910                 BUG_ON(ret);
1911                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1912                                         0, dest->root_key.objectid, level - 1,
1913                                         0);
1914                 BUG_ON(ret);
1915
1916                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1917                                         path->nodes[level]->start,
1918                                         src->root_key.objectid, level - 1, 0);
1919                 BUG_ON(ret);
1920
1921                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1922                                         0, dest->root_key.objectid, level - 1,
1923                                         0);
1924                 BUG_ON(ret);
1925
1926                 btrfs_unlock_up_safe(path, 0);
1927
1928                 ret = level;
1929                 break;
1930         }
1931         btrfs_tree_unlock(parent);
1932         free_extent_buffer(parent);
1933         return ret;
1934 }
1935
1936 /*
1937  * helper to find next relocated block in reloc tree
1938  */
1939 static noinline_for_stack
1940 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1941                        int *level)
1942 {
1943         struct extent_buffer *eb;
1944         int i;
1945         u64 last_snapshot;
1946         u32 nritems;
1947
1948         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1949
1950         for (i = 0; i < *level; i++) {
1951                 free_extent_buffer(path->nodes[i]);
1952                 path->nodes[i] = NULL;
1953         }
1954
1955         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1956                 eb = path->nodes[i];
1957                 nritems = btrfs_header_nritems(eb);
1958                 while (path->slots[i] + 1 < nritems) {
1959                         path->slots[i]++;
1960                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1961                             last_snapshot)
1962                                 continue;
1963
1964                         *level = i;
1965                         return 0;
1966                 }
1967                 free_extent_buffer(path->nodes[i]);
1968                 path->nodes[i] = NULL;
1969         }
1970         return 1;
1971 }
1972
1973 /*
1974  * walk down reloc tree to find relocated block of lowest level
1975  */
1976 static noinline_for_stack
1977 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1978                          int *level)
1979 {
1980         struct extent_buffer *eb = NULL;
1981         int i;
1982         u64 bytenr;
1983         u64 ptr_gen = 0;
1984         u64 last_snapshot;
1985         u32 nritems;
1986
1987         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1988
1989         for (i = *level; i > 0; i--) {
1990                 eb = path->nodes[i];
1991                 nritems = btrfs_header_nritems(eb);
1992                 while (path->slots[i] < nritems) {
1993                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1994                         if (ptr_gen > last_snapshot)
1995                                 break;
1996                         path->slots[i]++;
1997                 }
1998                 if (path->slots[i] >= nritems) {
1999                         if (i == *level)
2000                                 break;
2001                         *level = i + 1;
2002                         return 0;
2003                 }
2004                 if (i == 1) {
2005                         *level = i;
2006                         return 0;
2007                 }
2008
2009                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2010                 eb = read_tree_block(root, bytenr, ptr_gen);
2011                 if (IS_ERR(eb)) {
2012                         return PTR_ERR(eb);
2013                 } else if (!extent_buffer_uptodate(eb)) {
2014                         free_extent_buffer(eb);
2015                         return -EIO;
2016                 }
2017                 BUG_ON(btrfs_header_level(eb) != i - 1);
2018                 path->nodes[i - 1] = eb;
2019                 path->slots[i - 1] = 0;
2020         }
2021         return 1;
2022 }
2023
2024 /*
2025  * invalidate extent cache for file extents whose key in range of
2026  * [min_key, max_key)
2027  */
2028 static int invalidate_extent_cache(struct btrfs_root *root,
2029                                    struct btrfs_key *min_key,
2030                                    struct btrfs_key *max_key)
2031 {
2032         struct inode *inode = NULL;
2033         u64 objectid;
2034         u64 start, end;
2035         u64 ino;
2036
2037         objectid = min_key->objectid;
2038         while (1) {
2039                 cond_resched();
2040                 iput(inode);
2041
2042                 if (objectid > max_key->objectid)
2043                         break;
2044
2045                 inode = find_next_inode(root, objectid);
2046                 if (!inode)
2047                         break;
2048                 ino = btrfs_ino(inode);
2049
2050                 if (ino > max_key->objectid) {
2051                         iput(inode);
2052                         break;
2053                 }
2054
2055                 objectid = ino + 1;
2056                 if (!S_ISREG(inode->i_mode))
2057                         continue;
2058
2059                 if (unlikely(min_key->objectid == ino)) {
2060                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2061                                 continue;
2062                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2063                                 start = 0;
2064                         else {
2065                                 start = min_key->offset;
2066                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2067                         }
2068                 } else {
2069                         start = 0;
2070                 }
2071
2072                 if (unlikely(max_key->objectid == ino)) {
2073                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2074                                 continue;
2075                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2076                                 end = (u64)-1;
2077                         } else {
2078                                 if (max_key->offset == 0)
2079                                         continue;
2080                                 end = max_key->offset;
2081                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2082                                 end--;
2083                         }
2084                 } else {
2085                         end = (u64)-1;
2086                 }
2087
2088                 /* the lock_extent waits for readpage to complete */
2089                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2090                 btrfs_drop_extent_cache(inode, start, end, 1);
2091                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2092         }
2093         return 0;
2094 }
2095
2096 static int find_next_key(struct btrfs_path *path, int level,
2097                          struct btrfs_key *key)
2098
2099 {
2100         while (level < BTRFS_MAX_LEVEL) {
2101                 if (!path->nodes[level])
2102                         break;
2103                 if (path->slots[level] + 1 <
2104                     btrfs_header_nritems(path->nodes[level])) {
2105                         btrfs_node_key_to_cpu(path->nodes[level], key,
2106                                               path->slots[level] + 1);
2107                         return 0;
2108                 }
2109                 level++;
2110         }
2111         return 1;
2112 }
2113
2114 /*
2115  * merge the relocated tree blocks in reloc tree with corresponding
2116  * fs tree.
2117  */
2118 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2119                                                struct btrfs_root *root)
2120 {
2121         LIST_HEAD(inode_list);
2122         struct btrfs_key key;
2123         struct btrfs_key next_key;
2124         struct btrfs_trans_handle *trans = NULL;
2125         struct btrfs_root *reloc_root;
2126         struct btrfs_root_item *root_item;
2127         struct btrfs_path *path;
2128         struct extent_buffer *leaf;
2129         int level;
2130         int max_level;
2131         int replaced = 0;
2132         int ret;
2133         int err = 0;
2134         u32 min_reserved;
2135
2136         path = btrfs_alloc_path();
2137         if (!path)
2138                 return -ENOMEM;
2139         path->reada = READA_FORWARD;
2140
2141         reloc_root = root->reloc_root;
2142         root_item = &reloc_root->root_item;
2143
2144         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2145                 level = btrfs_root_level(root_item);
2146                 extent_buffer_get(reloc_root->node);
2147                 path->nodes[level] = reloc_root->node;
2148                 path->slots[level] = 0;
2149         } else {
2150                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2151
2152                 level = root_item->drop_level;
2153                 BUG_ON(level == 0);
2154                 path->lowest_level = level;
2155                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2156                 path->lowest_level = 0;
2157                 if (ret < 0) {
2158                         btrfs_free_path(path);
2159                         return ret;
2160                 }
2161
2162                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2163                                       path->slots[level]);
2164                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2165
2166                 btrfs_unlock_up_safe(path, 0);
2167         }
2168
2169         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2170         memset(&next_key, 0, sizeof(next_key));
2171
2172         while (1) {
2173                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2174                                              BTRFS_RESERVE_FLUSH_ALL);
2175                 if (ret) {
2176                         err = ret;
2177                         goto out;
2178                 }
2179                 trans = btrfs_start_transaction(root, 0);
2180                 if (IS_ERR(trans)) {
2181                         err = PTR_ERR(trans);
2182                         trans = NULL;
2183                         goto out;
2184                 }
2185                 trans->block_rsv = rc->block_rsv;
2186
2187                 replaced = 0;
2188                 max_level = level;
2189
2190                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2191                 if (ret < 0) {
2192                         err = ret;
2193                         goto out;
2194                 }
2195                 if (ret > 0)
2196                         break;
2197
2198                 if (!find_next_key(path, level, &key) &&
2199                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2200                         ret = 0;
2201                 } else {
2202                         ret = replace_path(trans, root, reloc_root, path,
2203                                            &next_key, level, max_level);
2204                 }
2205                 if (ret < 0) {
2206                         err = ret;
2207                         goto out;
2208                 }
2209
2210                 if (ret > 0) {
2211                         level = ret;
2212                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2213                                               path->slots[level]);
2214                         replaced = 1;
2215                 }
2216
2217                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2218                 if (ret > 0)
2219                         break;
2220
2221                 BUG_ON(level == 0);
2222                 /*
2223                  * save the merging progress in the drop_progress.
2224                  * this is OK since root refs == 1 in this case.
2225                  */
2226                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2227                                path->slots[level]);
2228                 root_item->drop_level = level;
2229
2230                 btrfs_end_transaction_throttle(trans, root);
2231                 trans = NULL;
2232
2233                 btrfs_btree_balance_dirty(root);
2234
2235                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2236                         invalidate_extent_cache(root, &key, &next_key);
2237         }
2238
2239         /*
2240          * handle the case only one block in the fs tree need to be
2241          * relocated and the block is tree root.
2242          */
2243         leaf = btrfs_lock_root_node(root);
2244         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2245         btrfs_tree_unlock(leaf);
2246         free_extent_buffer(leaf);
2247         if (ret < 0)
2248                 err = ret;
2249 out:
2250         btrfs_free_path(path);
2251
2252         if (err == 0) {
2253                 memset(&root_item->drop_progress, 0,
2254                        sizeof(root_item->drop_progress));
2255                 root_item->drop_level = 0;
2256                 btrfs_set_root_refs(root_item, 0);
2257                 btrfs_update_reloc_root(trans, root);
2258         }
2259
2260         if (trans)
2261                 btrfs_end_transaction_throttle(trans, root);
2262
2263         btrfs_btree_balance_dirty(root);
2264
2265         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2266                 invalidate_extent_cache(root, &key, &next_key);
2267
2268         return err;
2269 }
2270
2271 static noinline_for_stack
2272 int prepare_to_merge(struct reloc_control *rc, int err)
2273 {
2274         struct btrfs_root *root = rc->extent_root;
2275         struct btrfs_root *reloc_root;
2276         struct btrfs_trans_handle *trans;
2277         LIST_HEAD(reloc_roots);
2278         u64 num_bytes = 0;
2279         int ret;
2280
2281         mutex_lock(&root->fs_info->reloc_mutex);
2282         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2283         rc->merging_rsv_size += rc->nodes_relocated * 2;
2284         mutex_unlock(&root->fs_info->reloc_mutex);
2285
2286 again:
2287         if (!err) {
2288                 num_bytes = rc->merging_rsv_size;
2289                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2290                                           BTRFS_RESERVE_FLUSH_ALL);
2291                 if (ret)
2292                         err = ret;
2293         }
2294
2295         trans = btrfs_join_transaction(rc->extent_root);
2296         if (IS_ERR(trans)) {
2297                 if (!err)
2298                         btrfs_block_rsv_release(rc->extent_root,
2299                                                 rc->block_rsv, num_bytes);
2300                 return PTR_ERR(trans);
2301         }
2302
2303         if (!err) {
2304                 if (num_bytes != rc->merging_rsv_size) {
2305                         btrfs_end_transaction(trans, rc->extent_root);
2306                         btrfs_block_rsv_release(rc->extent_root,
2307                                                 rc->block_rsv, num_bytes);
2308                         goto again;
2309                 }
2310         }
2311
2312         rc->merge_reloc_tree = 1;
2313
2314         while (!list_empty(&rc->reloc_roots)) {
2315                 reloc_root = list_entry(rc->reloc_roots.next,
2316                                         struct btrfs_root, root_list);
2317                 list_del_init(&reloc_root->root_list);
2318
2319                 root = read_fs_root(reloc_root->fs_info,
2320                                     reloc_root->root_key.offset);
2321                 BUG_ON(IS_ERR(root));
2322                 BUG_ON(root->reloc_root != reloc_root);
2323
2324                 /*
2325                  * set reference count to 1, so btrfs_recover_relocation
2326                  * knows it should resumes merging
2327                  */
2328                 if (!err)
2329                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2330                 btrfs_update_reloc_root(trans, root);
2331
2332                 list_add(&reloc_root->root_list, &reloc_roots);
2333         }
2334
2335         list_splice(&reloc_roots, &rc->reloc_roots);
2336
2337         if (!err)
2338                 btrfs_commit_transaction(trans, rc->extent_root);
2339         else
2340                 btrfs_end_transaction(trans, rc->extent_root);
2341         return err;
2342 }
2343
2344 static noinline_for_stack
2345 void free_reloc_roots(struct list_head *list)
2346 {
2347         struct btrfs_root *reloc_root;
2348
2349         while (!list_empty(list)) {
2350                 reloc_root = list_entry(list->next, struct btrfs_root,
2351                                         root_list);
2352                 __del_reloc_root(reloc_root);
2353         }
2354 }
2355
2356 static noinline_for_stack
2357 void merge_reloc_roots(struct reloc_control *rc)
2358 {
2359         struct btrfs_root *root;
2360         struct btrfs_root *reloc_root;
2361         u64 last_snap;
2362         u64 otransid;
2363         u64 objectid;
2364         LIST_HEAD(reloc_roots);
2365         int found = 0;
2366         int ret = 0;
2367 again:
2368         root = rc->extent_root;
2369
2370         /*
2371          * this serializes us with btrfs_record_root_in_transaction,
2372          * we have to make sure nobody is in the middle of
2373          * adding their roots to the list while we are
2374          * doing this splice
2375          */
2376         mutex_lock(&root->fs_info->reloc_mutex);
2377         list_splice_init(&rc->reloc_roots, &reloc_roots);
2378         mutex_unlock(&root->fs_info->reloc_mutex);
2379
2380         while (!list_empty(&reloc_roots)) {
2381                 found = 1;
2382                 reloc_root = list_entry(reloc_roots.next,
2383                                         struct btrfs_root, root_list);
2384
2385                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2386                         root = read_fs_root(reloc_root->fs_info,
2387                                             reloc_root->root_key.offset);
2388                         BUG_ON(IS_ERR(root));
2389                         BUG_ON(root->reloc_root != reloc_root);
2390
2391                         ret = merge_reloc_root(rc, root);
2392                         if (ret) {
2393                                 if (list_empty(&reloc_root->root_list))
2394                                         list_add_tail(&reloc_root->root_list,
2395                                                       &reloc_roots);
2396                                 goto out;
2397                         }
2398                 } else {
2399                         list_del_init(&reloc_root->root_list);
2400                 }
2401
2402                 /*
2403                  * we keep the old last snapshot transid in rtranid when we
2404                  * created the relocation tree.
2405                  */
2406                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2407                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2408                 objectid = reloc_root->root_key.offset;
2409
2410                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2411                 if (ret < 0) {
2412                         if (list_empty(&reloc_root->root_list))
2413                                 list_add_tail(&reloc_root->root_list,
2414                                               &reloc_roots);
2415                         goto out;
2416                 }
2417         }
2418
2419         if (found) {
2420                 found = 0;
2421                 goto again;
2422         }
2423 out:
2424         if (ret) {
2425                 btrfs_handle_fs_error(root->fs_info, ret, NULL);
2426                 if (!list_empty(&reloc_roots))
2427                         free_reloc_roots(&reloc_roots);
2428
2429                 /* new reloc root may be added */
2430                 mutex_lock(&root->fs_info->reloc_mutex);
2431                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2432                 mutex_unlock(&root->fs_info->reloc_mutex);
2433                 if (!list_empty(&reloc_roots))
2434                         free_reloc_roots(&reloc_roots);
2435         }
2436
2437         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2438 }
2439
2440 static void free_block_list(struct rb_root *blocks)
2441 {
2442         struct tree_block *block;
2443         struct rb_node *rb_node;
2444         while ((rb_node = rb_first(blocks))) {
2445                 block = rb_entry(rb_node, struct tree_block, rb_node);
2446                 rb_erase(rb_node, blocks);
2447                 kfree(block);
2448         }
2449 }
2450
2451 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2452                                       struct btrfs_root *reloc_root)
2453 {
2454         struct btrfs_root *root;
2455
2456         if (reloc_root->last_trans == trans->transid)
2457                 return 0;
2458
2459         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2460         BUG_ON(IS_ERR(root));
2461         BUG_ON(root->reloc_root != reloc_root);
2462
2463         return btrfs_record_root_in_trans(trans, root);
2464 }
2465
2466 static noinline_for_stack
2467 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2468                                      struct reloc_control *rc,
2469                                      struct backref_node *node,
2470                                      struct backref_edge *edges[])
2471 {
2472         struct backref_node *next;
2473         struct btrfs_root *root;
2474         int index = 0;
2475
2476         next = node;
2477         while (1) {
2478                 cond_resched();
2479                 next = walk_up_backref(next, edges, &index);
2480                 root = next->root;
2481                 BUG_ON(!root);
2482                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2483
2484                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2485                         record_reloc_root_in_trans(trans, root);
2486                         break;
2487                 }
2488
2489                 btrfs_record_root_in_trans(trans, root);
2490                 root = root->reloc_root;
2491
2492                 if (next->new_bytenr != root->node->start) {
2493                         BUG_ON(next->new_bytenr);
2494                         BUG_ON(!list_empty(&next->list));
2495                         next->new_bytenr = root->node->start;
2496                         next->root = root;
2497                         list_add_tail(&next->list,
2498                                       &rc->backref_cache.changed);
2499                         __mark_block_processed(rc, next);
2500                         break;
2501                 }
2502
2503                 WARN_ON(1);
2504                 root = NULL;
2505                 next = walk_down_backref(edges, &index);
2506                 if (!next || next->level <= node->level)
2507                         break;
2508         }
2509         if (!root)
2510                 return NULL;
2511
2512         next = node;
2513         /* setup backref node path for btrfs_reloc_cow_block */
2514         while (1) {
2515                 rc->backref_cache.path[next->level] = next;
2516                 if (--index < 0)
2517                         break;
2518                 next = edges[index]->node[UPPER];
2519         }
2520         return root;
2521 }
2522
2523 /*
2524  * select a tree root for relocation. return NULL if the block
2525  * is reference counted. we should use do_relocation() in this
2526  * case. return a tree root pointer if the block isn't reference
2527  * counted. return -ENOENT if the block is root of reloc tree.
2528  */
2529 static noinline_for_stack
2530 struct btrfs_root *select_one_root(struct backref_node *node)
2531 {
2532         struct backref_node *next;
2533         struct btrfs_root *root;
2534         struct btrfs_root *fs_root = NULL;
2535         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2536         int index = 0;
2537
2538         next = node;
2539         while (1) {
2540                 cond_resched();
2541                 next = walk_up_backref(next, edges, &index);
2542                 root = next->root;
2543                 BUG_ON(!root);
2544
2545                 /* no other choice for non-references counted tree */
2546                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2547                         return root;
2548
2549                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2550                         fs_root = root;
2551
2552                 if (next != node)
2553                         return NULL;
2554
2555                 next = walk_down_backref(edges, &index);
2556                 if (!next || next->level <= node->level)
2557                         break;
2558         }
2559
2560         if (!fs_root)
2561                 return ERR_PTR(-ENOENT);
2562         return fs_root;
2563 }
2564
2565 static noinline_for_stack
2566 u64 calcu_metadata_size(struct reloc_control *rc,
2567                         struct backref_node *node, int reserve)
2568 {
2569         struct backref_node *next = node;
2570         struct backref_edge *edge;
2571         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2572         u64 num_bytes = 0;
2573         int index = 0;
2574
2575         BUG_ON(reserve && node->processed);
2576
2577         while (next) {
2578                 cond_resched();
2579                 while (1) {
2580                         if (next->processed && (reserve || next != node))
2581                                 break;
2582
2583                         num_bytes += rc->extent_root->nodesize;
2584
2585                         if (list_empty(&next->upper))
2586                                 break;
2587
2588                         edge = list_entry(next->upper.next,
2589                                           struct backref_edge, list[LOWER]);
2590                         edges[index++] = edge;
2591                         next = edge->node[UPPER];
2592                 }
2593                 next = walk_down_backref(edges, &index);
2594         }
2595         return num_bytes;
2596 }
2597
2598 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2599                                   struct reloc_control *rc,
2600                                   struct backref_node *node)
2601 {
2602         struct btrfs_root *root = rc->extent_root;
2603         u64 num_bytes;
2604         int ret;
2605         u64 tmp;
2606
2607         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2608
2609         trans->block_rsv = rc->block_rsv;
2610         rc->reserved_bytes += num_bytes;
2611
2612         /*
2613          * We are under a transaction here so we can only do limited flushing.
2614          * If we get an enospc just kick back -EAGAIN so we know to drop the
2615          * transaction and try to refill when we can flush all the things.
2616          */
2617         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2618                                 BTRFS_RESERVE_FLUSH_LIMIT);
2619         if (ret) {
2620                 tmp = rc->extent_root->nodesize * RELOCATION_RESERVED_NODES;
2621                 while (tmp <= rc->reserved_bytes)
2622                         tmp <<= 1;
2623                 /*
2624                  * only one thread can access block_rsv at this point,
2625                  * so we don't need hold lock to protect block_rsv.
2626                  * we expand more reservation size here to allow enough
2627                  * space for relocation and we will return eailer in
2628                  * enospc case.
2629                  */
2630                 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2631                         RELOCATION_RESERVED_NODES;
2632                 return -EAGAIN;
2633         }
2634
2635         return 0;
2636 }
2637
2638 /*
2639  * relocate a block tree, and then update pointers in upper level
2640  * blocks that reference the block to point to the new location.
2641  *
2642  * if called by link_to_upper, the block has already been relocated.
2643  * in that case this function just updates pointers.
2644  */
2645 static int do_relocation(struct btrfs_trans_handle *trans,
2646                          struct reloc_control *rc,
2647                          struct backref_node *node,
2648                          struct btrfs_key *key,
2649                          struct btrfs_path *path, int lowest)
2650 {
2651         struct backref_node *upper;
2652         struct backref_edge *edge;
2653         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2654         struct btrfs_root *root;
2655         struct extent_buffer *eb;
2656         u32 blocksize;
2657         u64 bytenr;
2658         u64 generation;
2659         int slot;
2660         int ret;
2661         int err = 0;
2662
2663         BUG_ON(lowest && node->eb);
2664
2665         path->lowest_level = node->level + 1;
2666         rc->backref_cache.path[node->level] = node;
2667         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2668                 cond_resched();
2669
2670                 upper = edge->node[UPPER];
2671                 root = select_reloc_root(trans, rc, upper, edges);
2672                 BUG_ON(!root);
2673
2674                 if (upper->eb && !upper->locked) {
2675                         if (!lowest) {
2676                                 ret = btrfs_bin_search(upper->eb, key,
2677                                                        upper->level, &slot);
2678                                 BUG_ON(ret);
2679                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2680                                 if (node->eb->start == bytenr)
2681                                         goto next;
2682                         }
2683                         drop_node_buffer(upper);
2684                 }
2685
2686                 if (!upper->eb) {
2687                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2688                         if (ret < 0) {
2689                                 err = ret;
2690                                 break;
2691                         }
2692                         BUG_ON(ret > 0);
2693
2694                         if (!upper->eb) {
2695                                 upper->eb = path->nodes[upper->level];
2696                                 path->nodes[upper->level] = NULL;
2697                         } else {
2698                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2699                         }
2700
2701                         upper->locked = 1;
2702                         path->locks[upper->level] = 0;
2703
2704                         slot = path->slots[upper->level];
2705                         btrfs_release_path(path);
2706                 } else {
2707                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2708                                                &slot);
2709                         BUG_ON(ret);
2710                 }
2711
2712                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2713                 if (lowest) {
2714                         BUG_ON(bytenr != node->bytenr);
2715                 } else {
2716                         if (node->eb->start == bytenr)
2717                                 goto next;
2718                 }
2719
2720                 blocksize = root->nodesize;
2721                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2722                 eb = read_tree_block(root, bytenr, generation);
2723                 if (IS_ERR(eb)) {
2724                         err = PTR_ERR(eb);
2725                         goto next;
2726                 } else if (!extent_buffer_uptodate(eb)) {
2727                         free_extent_buffer(eb);
2728                         err = -EIO;
2729                         goto next;
2730                 }
2731                 btrfs_tree_lock(eb);
2732                 btrfs_set_lock_blocking(eb);
2733
2734                 if (!node->eb) {
2735                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2736                                               slot, &eb);
2737                         btrfs_tree_unlock(eb);
2738                         free_extent_buffer(eb);
2739                         if (ret < 0) {
2740                                 err = ret;
2741                                 goto next;
2742                         }
2743                         BUG_ON(node->eb != eb);
2744                 } else {
2745                         btrfs_set_node_blockptr(upper->eb, slot,
2746                                                 node->eb->start);
2747                         btrfs_set_node_ptr_generation(upper->eb, slot,
2748                                                       trans->transid);
2749                         btrfs_mark_buffer_dirty(upper->eb);
2750
2751                         ret = btrfs_inc_extent_ref(trans, root,
2752                                                 node->eb->start, blocksize,
2753                                                 upper->eb->start,
2754                                                 btrfs_header_owner(upper->eb),
2755                                                 node->level, 0);
2756                         BUG_ON(ret);
2757
2758                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2759                         BUG_ON(ret);
2760                 }
2761 next:
2762                 if (!upper->pending)
2763                         drop_node_buffer(upper);
2764                 else
2765                         unlock_node_buffer(upper);
2766                 if (err)
2767                         break;
2768         }
2769
2770         if (!err && node->pending) {
2771                 drop_node_buffer(node);
2772                 list_move_tail(&node->list, &rc->backref_cache.changed);
2773                 node->pending = 0;
2774         }
2775
2776         path->lowest_level = 0;
2777         BUG_ON(err == -ENOSPC);
2778         return err;
2779 }
2780
2781 static int link_to_upper(struct btrfs_trans_handle *trans,
2782                          struct reloc_control *rc,
2783                          struct backref_node *node,
2784                          struct btrfs_path *path)
2785 {
2786         struct btrfs_key key;
2787
2788         btrfs_node_key_to_cpu(node->eb, &key, 0);
2789         return do_relocation(trans, rc, node, &key, path, 0);
2790 }
2791
2792 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2793                                 struct reloc_control *rc,
2794                                 struct btrfs_path *path, int err)
2795 {
2796         LIST_HEAD(list);
2797         struct backref_cache *cache = &rc->backref_cache;
2798         struct backref_node *node;
2799         int level;
2800         int ret;
2801
2802         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2803                 while (!list_empty(&cache->pending[level])) {
2804                         node = list_entry(cache->pending[level].next,
2805                                           struct backref_node, list);
2806                         list_move_tail(&node->list, &list);
2807                         BUG_ON(!node->pending);
2808
2809                         if (!err) {
2810                                 ret = link_to_upper(trans, rc, node, path);
2811                                 if (ret < 0)
2812                                         err = ret;
2813                         }
2814                 }
2815                 list_splice_init(&list, &cache->pending[level]);
2816         }
2817         return err;
2818 }
2819
2820 static void mark_block_processed(struct reloc_control *rc,
2821                                  u64 bytenr, u32 blocksize)
2822 {
2823         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2824                         EXTENT_DIRTY);
2825 }
2826
2827 static void __mark_block_processed(struct reloc_control *rc,
2828                                    struct backref_node *node)
2829 {
2830         u32 blocksize;
2831         if (node->level == 0 ||
2832             in_block_group(node->bytenr, rc->block_group)) {
2833                 blocksize = rc->extent_root->nodesize;
2834                 mark_block_processed(rc, node->bytenr, blocksize);
2835         }
2836         node->processed = 1;
2837 }
2838
2839 /*
2840  * mark a block and all blocks directly/indirectly reference the block
2841  * as processed.
2842  */
2843 static void update_processed_blocks(struct reloc_control *rc,
2844                                     struct backref_node *node)
2845 {
2846         struct backref_node *next = node;
2847         struct backref_edge *edge;
2848         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2849         int index = 0;
2850
2851         while (next) {
2852                 cond_resched();
2853                 while (1) {
2854                         if (next->processed)
2855                                 break;
2856
2857                         __mark_block_processed(rc, next);
2858
2859                         if (list_empty(&next->upper))
2860                                 break;
2861
2862                         edge = list_entry(next->upper.next,
2863                                           struct backref_edge, list[LOWER]);
2864                         edges[index++] = edge;
2865                         next = edge->node[UPPER];
2866                 }
2867                 next = walk_down_backref(edges, &index);
2868         }
2869 }
2870
2871 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2872 {
2873         u32 blocksize = rc->extent_root->nodesize;
2874
2875         if (test_range_bit(&rc->processed_blocks, bytenr,
2876                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2877                 return 1;
2878         return 0;
2879 }
2880
2881 static int get_tree_block_key(struct reloc_control *rc,
2882                               struct tree_block *block)
2883 {
2884         struct extent_buffer *eb;
2885
2886         BUG_ON(block->key_ready);
2887         eb = read_tree_block(rc->extent_root, block->bytenr,
2888                              block->key.offset);
2889         if (IS_ERR(eb)) {
2890                 return PTR_ERR(eb);
2891         } else if (!extent_buffer_uptodate(eb)) {
2892                 free_extent_buffer(eb);
2893                 return -EIO;
2894         }
2895         WARN_ON(btrfs_header_level(eb) != block->level);
2896         if (block->level == 0)
2897                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2898         else
2899                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2900         free_extent_buffer(eb);
2901         block->key_ready = 1;
2902         return 0;
2903 }
2904
2905 /*
2906  * helper function to relocate a tree block
2907  */
2908 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2909                                 struct reloc_control *rc,
2910                                 struct backref_node *node,
2911                                 struct btrfs_key *key,
2912                                 struct btrfs_path *path)
2913 {
2914         struct btrfs_root *root;
2915         int ret = 0;
2916
2917         if (!node)
2918                 return 0;
2919
2920         BUG_ON(node->processed);
2921         root = select_one_root(node);
2922         if (root == ERR_PTR(-ENOENT)) {
2923                 update_processed_blocks(rc, node);
2924                 goto out;
2925         }
2926
2927         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2928                 ret = reserve_metadata_space(trans, rc, node);
2929                 if (ret)
2930                         goto out;
2931         }
2932
2933         if (root) {
2934                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2935                         BUG_ON(node->new_bytenr);
2936                         BUG_ON(!list_empty(&node->list));
2937                         btrfs_record_root_in_trans(trans, root);
2938                         root = root->reloc_root;
2939                         node->new_bytenr = root->node->start;
2940                         node->root = root;
2941                         list_add_tail(&node->list, &rc->backref_cache.changed);
2942                 } else {
2943                         path->lowest_level = node->level;
2944                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2945                         btrfs_release_path(path);
2946                         if (ret > 0)
2947                                 ret = 0;
2948                 }
2949                 if (!ret)
2950                         update_processed_blocks(rc, node);
2951         } else {
2952                 ret = do_relocation(trans, rc, node, key, path, 1);
2953         }
2954 out:
2955         if (ret || node->level == 0 || node->cowonly)
2956                 remove_backref_node(&rc->backref_cache, node);
2957         return ret;
2958 }
2959
2960 /*
2961  * relocate a list of blocks
2962  */
2963 static noinline_for_stack
2964 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2965                          struct reloc_control *rc, struct rb_root *blocks)
2966 {
2967         struct backref_node *node;
2968         struct btrfs_path *path;
2969         struct tree_block *block;
2970         struct rb_node *rb_node;
2971         int ret;
2972         int err = 0;
2973
2974         path = btrfs_alloc_path();
2975         if (!path) {
2976                 err = -ENOMEM;
2977                 goto out_free_blocks;
2978         }
2979
2980         rb_node = rb_first(blocks);
2981         while (rb_node) {
2982                 block = rb_entry(rb_node, struct tree_block, rb_node);
2983                 if (!block->key_ready)
2984                         readahead_tree_block(rc->extent_root, block->bytenr);
2985                 rb_node = rb_next(rb_node);
2986         }
2987
2988         rb_node = rb_first(blocks);
2989         while (rb_node) {
2990                 block = rb_entry(rb_node, struct tree_block, rb_node);
2991                 if (!block->key_ready) {
2992                         err = get_tree_block_key(rc, block);
2993                         if (err)
2994                                 goto out_free_path;
2995                 }
2996                 rb_node = rb_next(rb_node);
2997         }
2998
2999         rb_node = rb_first(blocks);
3000         while (rb_node) {
3001                 block = rb_entry(rb_node, struct tree_block, rb_node);
3002
3003                 node = build_backref_tree(rc, &block->key,
3004                                           block->level, block->bytenr);
3005                 if (IS_ERR(node)) {
3006                         err = PTR_ERR(node);
3007                         goto out;
3008                 }
3009
3010                 ret = relocate_tree_block(trans, rc, node, &block->key,
3011                                           path);
3012                 if (ret < 0) {
3013                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3014                                 err = ret;
3015                         goto out;
3016                 }
3017                 rb_node = rb_next(rb_node);
3018         }
3019 out:
3020         err = finish_pending_nodes(trans, rc, path, err);
3021
3022 out_free_path:
3023         btrfs_free_path(path);
3024 out_free_blocks:
3025         free_block_list(blocks);
3026         return err;
3027 }
3028
3029 static noinline_for_stack
3030 int prealloc_file_extent_cluster(struct inode *inode,
3031                                  struct file_extent_cluster *cluster)
3032 {
3033         u64 alloc_hint = 0;
3034         u64 start;
3035         u64 end;
3036         u64 offset = BTRFS_I(inode)->index_cnt;
3037         u64 num_bytes;
3038         int nr = 0;
3039         int ret = 0;
3040
3041         BUG_ON(cluster->start != cluster->boundary[0]);
3042         inode_lock(inode);
3043
3044         ret = btrfs_check_data_free_space(inode, cluster->start,
3045                                           cluster->end + 1 - cluster->start);
3046         if (ret)
3047                 goto out;
3048
3049         while (nr < cluster->nr) {
3050                 start = cluster->boundary[nr] - offset;
3051                 if (nr + 1 < cluster->nr)
3052                         end = cluster->boundary[nr + 1] - 1 - offset;
3053                 else
3054                         end = cluster->end - offset;
3055
3056                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3057                 num_bytes = end + 1 - start;
3058                 ret = btrfs_prealloc_file_range(inode, 0, start,
3059                                                 num_bytes, num_bytes,
3060                                                 end + 1, &alloc_hint);
3061                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3062                 if (ret)
3063                         break;
3064                 nr++;
3065         }
3066         btrfs_free_reserved_data_space(inode, cluster->start,
3067                                        cluster->end + 1 - cluster->start);
3068 out:
3069         inode_unlock(inode);
3070         return ret;
3071 }
3072
3073 static noinline_for_stack
3074 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3075                          u64 block_start)
3076 {
3077         struct btrfs_root *root = BTRFS_I(inode)->root;
3078         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3079         struct extent_map *em;
3080         int ret = 0;
3081
3082         em = alloc_extent_map();
3083         if (!em)
3084                 return -ENOMEM;
3085
3086         em->start = start;
3087         em->len = end + 1 - start;
3088         em->block_len = em->len;
3089         em->block_start = block_start;
3090         em->bdev = root->fs_info->fs_devices->latest_bdev;
3091         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3092
3093         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3094         while (1) {
3095                 write_lock(&em_tree->lock);
3096                 ret = add_extent_mapping(em_tree, em, 0);
3097                 write_unlock(&em_tree->lock);
3098                 if (ret != -EEXIST) {
3099                         free_extent_map(em);
3100                         break;
3101                 }
3102                 btrfs_drop_extent_cache(inode, start, end, 0);
3103         }
3104         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3105         return ret;
3106 }
3107
3108 static int relocate_file_extent_cluster(struct inode *inode,
3109                                         struct file_extent_cluster *cluster)
3110 {
3111         u64 page_start;
3112         u64 page_end;
3113         u64 offset = BTRFS_I(inode)->index_cnt;
3114         unsigned long index;
3115         unsigned long last_index;
3116         struct page *page;
3117         struct file_ra_state *ra;
3118         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3119         int nr = 0;
3120         int ret = 0;
3121
3122         if (!cluster->nr)
3123                 return 0;
3124
3125         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3126         if (!ra)
3127                 return -ENOMEM;
3128
3129         ret = prealloc_file_extent_cluster(inode, cluster);
3130         if (ret)
3131                 goto out;
3132
3133         file_ra_state_init(ra, inode->i_mapping);
3134
3135         ret = setup_extent_mapping(inode, cluster->start - offset,
3136                                    cluster->end - offset, cluster->start);
3137         if (ret)
3138                 goto out;
3139
3140         index = (cluster->start - offset) >> PAGE_SHIFT;
3141         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3142         while (index <= last_index) {
3143                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_SIZE);
3144                 if (ret)
3145                         goto out;
3146
3147                 page = find_lock_page(inode->i_mapping, index);
3148                 if (!page) {
3149                         page_cache_sync_readahead(inode->i_mapping,
3150                                                   ra, NULL, index,
3151                                                   last_index + 1 - index);
3152                         page = find_or_create_page(inode->i_mapping, index,
3153                                                    mask);
3154                         if (!page) {
3155                                 btrfs_delalloc_release_metadata(inode,
3156                                                         PAGE_SIZE);
3157                                 ret = -ENOMEM;
3158                                 goto out;
3159                         }
3160                 }
3161
3162                 if (PageReadahead(page)) {
3163                         page_cache_async_readahead(inode->i_mapping,
3164                                                    ra, NULL, page, index,
3165                                                    last_index + 1 - index);
3166                 }
3167
3168                 if (!PageUptodate(page)) {
3169                         btrfs_readpage(NULL, page);
3170                         lock_page(page);
3171                         if (!PageUptodate(page)) {
3172                                 unlock_page(page);
3173                                 put_page(page);
3174                                 btrfs_delalloc_release_metadata(inode,
3175                                                         PAGE_SIZE);
3176                                 ret = -EIO;
3177                                 goto out;
3178                         }
3179                 }
3180
3181                 page_start = page_offset(page);
3182                 page_end = page_start + PAGE_SIZE - 1;
3183
3184                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3185
3186                 set_page_extent_mapped(page);
3187
3188                 if (nr < cluster->nr &&
3189                     page_start + offset == cluster->boundary[nr]) {
3190                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3191                                         page_start, page_end,
3192                                         EXTENT_BOUNDARY);
3193                         nr++;
3194                 }
3195
3196                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3197                 set_page_dirty(page);
3198
3199                 unlock_extent(&BTRFS_I(inode)->io_tree,
3200                               page_start, page_end);
3201                 unlock_page(page);
3202                 put_page(page);
3203
3204                 index++;
3205                 balance_dirty_pages_ratelimited(inode->i_mapping);
3206                 btrfs_throttle(BTRFS_I(inode)->root);
3207         }
3208         WARN_ON(nr != cluster->nr);
3209 out:
3210         kfree(ra);
3211         return ret;
3212 }
3213
3214 static noinline_for_stack
3215 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3216                          struct file_extent_cluster *cluster)
3217 {
3218         int ret;
3219
3220         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3221                 ret = relocate_file_extent_cluster(inode, cluster);
3222                 if (ret)
3223                         return ret;
3224                 cluster->nr = 0;
3225         }
3226
3227         if (!cluster->nr)
3228                 cluster->start = extent_key->objectid;
3229         else
3230                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3231         cluster->end = extent_key->objectid + extent_key->offset - 1;
3232         cluster->boundary[cluster->nr] = extent_key->objectid;
3233         cluster->nr++;
3234
3235         if (cluster->nr >= MAX_EXTENTS) {
3236                 ret = relocate_file_extent_cluster(inode, cluster);
3237                 if (ret)
3238                         return ret;
3239                 cluster->nr = 0;
3240         }
3241         return 0;
3242 }
3243
3244 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3245 static int get_ref_objectid_v0(struct reloc_control *rc,
3246                                struct btrfs_path *path,
3247                                struct btrfs_key *extent_key,
3248                                u64 *ref_objectid, int *path_change)
3249 {
3250         struct btrfs_key key;
3251         struct extent_buffer *leaf;
3252         struct btrfs_extent_ref_v0 *ref0;
3253         int ret;
3254         int slot;
3255
3256         leaf = path->nodes[0];
3257         slot = path->slots[0];
3258         while (1) {
3259                 if (slot >= btrfs_header_nritems(leaf)) {
3260                         ret = btrfs_next_leaf(rc->extent_root, path);
3261                         if (ret < 0)
3262                                 return ret;
3263                         BUG_ON(ret > 0);
3264                         leaf = path->nodes[0];
3265                         slot = path->slots[0];
3266                         if (path_change)
3267                                 *path_change = 1;
3268                 }
3269                 btrfs_item_key_to_cpu(leaf, &key, slot);
3270                 if (key.objectid != extent_key->objectid)
3271                         return -ENOENT;
3272
3273                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3274                         slot++;
3275                         continue;
3276                 }
3277                 ref0 = btrfs_item_ptr(leaf, slot,
3278                                 struct btrfs_extent_ref_v0);
3279                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3280                 break;
3281         }
3282         return 0;
3283 }
3284 #endif
3285
3286 /*
3287  * helper to add a tree block to the list.
3288  * the major work is getting the generation and level of the block
3289  */
3290 static int add_tree_block(struct reloc_control *rc,
3291                           struct btrfs_key *extent_key,
3292                           struct btrfs_path *path,
3293                           struct rb_root *blocks)
3294 {
3295         struct extent_buffer *eb;
3296         struct btrfs_extent_item *ei;
3297         struct btrfs_tree_block_info *bi;
3298         struct tree_block *block;
3299         struct rb_node *rb_node;
3300         u32 item_size;
3301         int level = -1;
3302         u64 generation;
3303
3304         eb =  path->nodes[0];
3305         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3306
3307         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3308             item_size >= sizeof(*ei) + sizeof(*bi)) {
3309                 ei = btrfs_item_ptr(eb, path->slots[0],
3310                                 struct btrfs_extent_item);
3311                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3312                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3313                         level = btrfs_tree_block_level(eb, bi);
3314                 } else {
3315                         level = (int)extent_key->offset;
3316                 }
3317                 generation = btrfs_extent_generation(eb, ei);
3318         } else {
3319 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3320                 u64 ref_owner;
3321                 int ret;
3322
3323                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3324                 ret = get_ref_objectid_v0(rc, path, extent_key,
3325                                           &ref_owner, NULL);
3326                 if (ret < 0)
3327                         return ret;
3328                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3329                 level = (int)ref_owner;
3330                 /* FIXME: get real generation */
3331                 generation = 0;
3332 #else
3333                 BUG();
3334 #endif
3335         }
3336
3337         btrfs_release_path(path);
3338
3339         BUG_ON(level == -1);
3340
3341         block = kmalloc(sizeof(*block), GFP_NOFS);
3342         if (!block)
3343                 return -ENOMEM;
3344
3345         block->bytenr = extent_key->objectid;
3346         block->key.objectid = rc->extent_root->nodesize;
3347         block->key.offset = generation;
3348         block->level = level;
3349         block->key_ready = 0;
3350
3351         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3352         if (rb_node)
3353                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3354
3355         return 0;
3356 }
3357
3358 /*
3359  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3360  */
3361 static int __add_tree_block(struct reloc_control *rc,
3362                             u64 bytenr, u32 blocksize,
3363                             struct rb_root *blocks)
3364 {
3365         struct btrfs_path *path;
3366         struct btrfs_key key;
3367         int ret;
3368         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3369                                         SKINNY_METADATA);
3370
3371         if (tree_block_processed(bytenr, rc))
3372                 return 0;
3373
3374         if (tree_search(blocks, bytenr))
3375                 return 0;
3376
3377         path = btrfs_alloc_path();
3378         if (!path)
3379                 return -ENOMEM;
3380 again:
3381         key.objectid = bytenr;
3382         if (skinny) {
3383                 key.type = BTRFS_METADATA_ITEM_KEY;
3384                 key.offset = (u64)-1;
3385         } else {
3386                 key.type = BTRFS_EXTENT_ITEM_KEY;
3387                 key.offset = blocksize;
3388         }
3389
3390         path->search_commit_root = 1;
3391         path->skip_locking = 1;
3392         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3393         if (ret < 0)
3394                 goto out;
3395
3396         if (ret > 0 && skinny) {
3397                 if (path->slots[0]) {
3398                         path->slots[0]--;
3399                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3400                                               path->slots[0]);
3401                         if (key.objectid == bytenr &&
3402                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3403                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3404                               key.offset == blocksize)))
3405                                 ret = 0;
3406                 }
3407
3408                 if (ret) {
3409                         skinny = false;
3410                         btrfs_release_path(path);
3411                         goto again;
3412                 }
3413         }
3414         BUG_ON(ret);
3415
3416         ret = add_tree_block(rc, &key, path, blocks);
3417 out:
3418         btrfs_free_path(path);
3419         return ret;
3420 }
3421
3422 /*
3423  * helper to check if the block use full backrefs for pointers in it
3424  */
3425 static int block_use_full_backref(struct reloc_control *rc,
3426                                   struct extent_buffer *eb)
3427 {
3428         u64 flags;
3429         int ret;
3430
3431         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3432             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3433                 return 1;
3434
3435         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3436                                        eb->start, btrfs_header_level(eb), 1,
3437                                        NULL, &flags);
3438         BUG_ON(ret);
3439
3440         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3441                 ret = 1;
3442         else
3443                 ret = 0;
3444         return ret;
3445 }
3446
3447 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3448                                     struct btrfs_block_group_cache *block_group,
3449                                     struct inode *inode,
3450                                     u64 ino)
3451 {
3452         struct btrfs_key key;
3453         struct btrfs_root *root = fs_info->tree_root;
3454         struct btrfs_trans_handle *trans;
3455         int ret = 0;
3456
3457         if (inode)
3458                 goto truncate;
3459
3460         key.objectid = ino;
3461         key.type = BTRFS_INODE_ITEM_KEY;
3462         key.offset = 0;
3463
3464         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3465         if (IS_ERR(inode) || is_bad_inode(inode)) {
3466                 if (!IS_ERR(inode))
3467                         iput(inode);
3468                 return -ENOENT;
3469         }
3470
3471 truncate:
3472         ret = btrfs_check_trunc_cache_free_space(root,
3473                                                  &fs_info->global_block_rsv);
3474         if (ret)
3475                 goto out;
3476
3477         trans = btrfs_join_transaction(root);
3478         if (IS_ERR(trans)) {
3479                 ret = PTR_ERR(trans);
3480                 goto out;
3481         }
3482
3483         ret = btrfs_truncate_free_space_cache(root, trans, block_group, inode);
3484
3485         btrfs_end_transaction(trans, root);
3486         btrfs_btree_balance_dirty(root);
3487 out:
3488         iput(inode);
3489         return ret;
3490 }
3491
3492 /*
3493  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3494  * this function scans fs tree to find blocks reference the data extent
3495  */
3496 static int find_data_references(struct reloc_control *rc,
3497                                 struct btrfs_key *extent_key,
3498                                 struct extent_buffer *leaf,
3499                                 struct btrfs_extent_data_ref *ref,
3500                                 struct rb_root *blocks)
3501 {
3502         struct btrfs_path *path;
3503         struct tree_block *block;
3504         struct btrfs_root *root;
3505         struct btrfs_file_extent_item *fi;
3506         struct rb_node *rb_node;
3507         struct btrfs_key key;
3508         u64 ref_root;
3509         u64 ref_objectid;
3510         u64 ref_offset;
3511         u32 ref_count;
3512         u32 nritems;
3513         int err = 0;
3514         int added = 0;
3515         int counted;
3516         int ret;
3517
3518         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3519         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3520         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3521         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3522
3523         /*
3524          * This is an extent belonging to the free space cache, lets just delete
3525          * it and redo the search.
3526          */
3527         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3528                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3529                                                rc->block_group,
3530                                                NULL, ref_objectid);
3531                 if (ret != -ENOENT)
3532                         return ret;
3533                 ret = 0;
3534         }
3535
3536         path = btrfs_alloc_path();
3537         if (!path)
3538                 return -ENOMEM;
3539         path->reada = READA_FORWARD;
3540
3541         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3542         if (IS_ERR(root)) {
3543                 err = PTR_ERR(root);
3544                 goto out;
3545         }
3546
3547         key.objectid = ref_objectid;
3548         key.type = BTRFS_EXTENT_DATA_KEY;
3549         if (ref_offset > ((u64)-1 << 32))
3550                 key.offset = 0;
3551         else
3552                 key.offset = ref_offset;
3553
3554         path->search_commit_root = 1;
3555         path->skip_locking = 1;
3556         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3557         if (ret < 0) {
3558                 err = ret;
3559                 goto out;
3560         }
3561
3562         leaf = path->nodes[0];
3563         nritems = btrfs_header_nritems(leaf);
3564         /*
3565          * the references in tree blocks that use full backrefs
3566          * are not counted in
3567          */
3568         if (block_use_full_backref(rc, leaf))
3569                 counted = 0;
3570         else
3571                 counted = 1;
3572         rb_node = tree_search(blocks, leaf->start);
3573         if (rb_node) {
3574                 if (counted)
3575                         added = 1;
3576                 else
3577                         path->slots[0] = nritems;
3578         }
3579
3580         while (ref_count > 0) {
3581                 while (path->slots[0] >= nritems) {
3582                         ret = btrfs_next_leaf(root, path);
3583                         if (ret < 0) {
3584                                 err = ret;
3585                                 goto out;
3586                         }
3587                         if (WARN_ON(ret > 0))
3588                                 goto out;
3589
3590                         leaf = path->nodes[0];
3591                         nritems = btrfs_header_nritems(leaf);
3592                         added = 0;
3593
3594                         if (block_use_full_backref(rc, leaf))
3595                                 counted = 0;
3596                         else
3597                                 counted = 1;
3598                         rb_node = tree_search(blocks, leaf->start);
3599                         if (rb_node) {
3600                                 if (counted)
3601                                         added = 1;
3602                                 else
3603                                         path->slots[0] = nritems;
3604                         }
3605                 }
3606
3607                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3608                 if (WARN_ON(key.objectid != ref_objectid ||
3609                     key.type != BTRFS_EXTENT_DATA_KEY))
3610                         break;
3611
3612                 fi = btrfs_item_ptr(leaf, path->slots[0],
3613                                     struct btrfs_file_extent_item);
3614
3615                 if (btrfs_file_extent_type(leaf, fi) ==
3616                     BTRFS_FILE_EXTENT_INLINE)
3617                         goto next;
3618
3619                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3620                     extent_key->objectid)
3621                         goto next;
3622
3623                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3624                 if (key.offset != ref_offset)
3625                         goto next;
3626
3627                 if (counted)
3628                         ref_count--;
3629                 if (added)
3630                         goto next;
3631
3632                 if (!tree_block_processed(leaf->start, rc)) {
3633                         block = kmalloc(sizeof(*block), GFP_NOFS);
3634                         if (!block) {
3635                                 err = -ENOMEM;
3636                                 break;
3637                         }
3638                         block->bytenr = leaf->start;
3639                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3640                         block->level = 0;
3641                         block->key_ready = 1;
3642                         rb_node = tree_insert(blocks, block->bytenr,
3643                                               &block->rb_node);
3644                         if (rb_node)
3645                                 backref_tree_panic(rb_node, -EEXIST,
3646                                                    block->bytenr);
3647                 }
3648                 if (counted)
3649                         added = 1;
3650                 else
3651                         path->slots[0] = nritems;
3652 next:
3653                 path->slots[0]++;
3654
3655         }
3656 out:
3657         btrfs_free_path(path);
3658         return err;
3659 }
3660
3661 /*
3662  * helper to find all tree blocks that reference a given data extent
3663  */
3664 static noinline_for_stack
3665 int add_data_references(struct reloc_control *rc,
3666                         struct btrfs_key *extent_key,
3667                         struct btrfs_path *path,
3668                         struct rb_root *blocks)
3669 {
3670         struct btrfs_key key;
3671         struct extent_buffer *eb;
3672         struct btrfs_extent_data_ref *dref;
3673         struct btrfs_extent_inline_ref *iref;
3674         unsigned long ptr;
3675         unsigned long end;
3676         u32 blocksize = rc->extent_root->nodesize;
3677         int ret = 0;
3678         int err = 0;
3679
3680         eb = path->nodes[0];
3681         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3682         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3683 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3684         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3685                 ptr = end;
3686         else
3687 #endif
3688                 ptr += sizeof(struct btrfs_extent_item);
3689
3690         while (ptr < end) {
3691                 iref = (struct btrfs_extent_inline_ref *)ptr;
3692                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3693                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3694                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3695                         ret = __add_tree_block(rc, key.offset, blocksize,
3696                                                blocks);
3697                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3698                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3699                         ret = find_data_references(rc, extent_key,
3700                                                    eb, dref, blocks);
3701                 } else {
3702                         BUG();
3703                 }
3704                 if (ret) {
3705                         err = ret;
3706                         goto out;
3707                 }
3708                 ptr += btrfs_extent_inline_ref_size(key.type);
3709         }
3710         WARN_ON(ptr > end);
3711
3712         while (1) {
3713                 cond_resched();
3714                 eb = path->nodes[0];
3715                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3716                         ret = btrfs_next_leaf(rc->extent_root, path);
3717                         if (ret < 0) {
3718                                 err = ret;
3719                                 break;
3720                         }
3721                         if (ret > 0)
3722                                 break;
3723                         eb = path->nodes[0];
3724                 }
3725
3726                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3727                 if (key.objectid != extent_key->objectid)
3728                         break;
3729
3730 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3731                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3732                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3733 #else
3734                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3735                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3736 #endif
3737                         ret = __add_tree_block(rc, key.offset, blocksize,
3738                                                blocks);
3739                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3740                         dref = btrfs_item_ptr(eb, path->slots[0],
3741                                               struct btrfs_extent_data_ref);
3742                         ret = find_data_references(rc, extent_key,
3743                                                    eb, dref, blocks);
3744                 } else {
3745                         ret = 0;
3746                 }
3747                 if (ret) {
3748                         err = ret;
3749                         break;
3750                 }
3751                 path->slots[0]++;
3752         }
3753 out:
3754         btrfs_release_path(path);
3755         if (err)
3756                 free_block_list(blocks);
3757         return err;
3758 }
3759
3760 /*
3761  * helper to find next unprocessed extent
3762  */
3763 static noinline_for_stack
3764 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3765                      struct btrfs_key *extent_key)
3766 {
3767         struct btrfs_key key;
3768         struct extent_buffer *leaf;
3769         u64 start, end, last;
3770         int ret;
3771
3772         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3773         while (1) {
3774                 cond_resched();
3775                 if (rc->search_start >= last) {
3776                         ret = 1;
3777                         break;
3778                 }
3779
3780                 key.objectid = rc->search_start;
3781                 key.type = BTRFS_EXTENT_ITEM_KEY;
3782                 key.offset = 0;
3783
3784                 path->search_commit_root = 1;
3785                 path->skip_locking = 1;
3786                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3787                                         0, 0);
3788                 if (ret < 0)
3789                         break;
3790 next:
3791                 leaf = path->nodes[0];
3792                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3793                         ret = btrfs_next_leaf(rc->extent_root, path);
3794                         if (ret != 0)
3795                                 break;
3796                         leaf = path->nodes[0];
3797                 }
3798
3799                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3800                 if (key.objectid >= last) {
3801                         ret = 1;
3802                         break;
3803                 }
3804
3805                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3806                     key.type != BTRFS_METADATA_ITEM_KEY) {
3807                         path->slots[0]++;
3808                         goto next;
3809                 }
3810
3811                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3812                     key.objectid + key.offset <= rc->search_start) {
3813                         path->slots[0]++;
3814                         goto next;
3815                 }
3816
3817                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3818                     key.objectid + rc->extent_root->nodesize <=
3819                     rc->search_start) {
3820                         path->slots[0]++;
3821                         goto next;
3822                 }
3823
3824                 ret = find_first_extent_bit(&rc->processed_blocks,
3825                                             key.objectid, &start, &end,
3826                                             EXTENT_DIRTY, NULL);
3827
3828                 if (ret == 0 && start <= key.objectid) {
3829                         btrfs_release_path(path);
3830                         rc->search_start = end + 1;
3831                 } else {
3832                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3833                                 rc->search_start = key.objectid + key.offset;
3834                         else
3835                                 rc->search_start = key.objectid +
3836                                         rc->extent_root->nodesize;
3837                         memcpy(extent_key, &key, sizeof(key));
3838                         return 0;
3839                 }
3840         }
3841         btrfs_release_path(path);
3842         return ret;
3843 }
3844
3845 static void set_reloc_control(struct reloc_control *rc)
3846 {
3847         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3848
3849         mutex_lock(&fs_info->reloc_mutex);
3850         fs_info->reloc_ctl = rc;
3851         mutex_unlock(&fs_info->reloc_mutex);
3852 }
3853
3854 static void unset_reloc_control(struct reloc_control *rc)
3855 {
3856         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3857
3858         mutex_lock(&fs_info->reloc_mutex);
3859         fs_info->reloc_ctl = NULL;
3860         mutex_unlock(&fs_info->reloc_mutex);
3861 }
3862
3863 static int check_extent_flags(u64 flags)
3864 {
3865         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3866             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3867                 return 1;
3868         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3869             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3870                 return 1;
3871         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3872             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3873                 return 1;
3874         return 0;
3875 }
3876
3877 static noinline_for_stack
3878 int prepare_to_relocate(struct reloc_control *rc)
3879 {
3880         struct btrfs_trans_handle *trans;
3881         int ret;
3882
3883         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3884                                               BTRFS_BLOCK_RSV_TEMP);
3885         if (!rc->block_rsv)
3886                 return -ENOMEM;
3887
3888         memset(&rc->cluster, 0, sizeof(rc->cluster));
3889         rc->search_start = rc->block_group->key.objectid;
3890         rc->extents_found = 0;
3891         rc->nodes_relocated = 0;
3892         rc->merging_rsv_size = 0;
3893         rc->reserved_bytes = 0;
3894         rc->block_rsv->size = rc->extent_root->nodesize *
3895                               RELOCATION_RESERVED_NODES;
3896         ret = btrfs_block_rsv_refill(rc->extent_root,
3897                                      rc->block_rsv, rc->block_rsv->size,
3898                                      BTRFS_RESERVE_FLUSH_ALL);
3899         if (ret)
3900                 return ret;
3901
3902         rc->create_reloc_tree = 1;
3903         set_reloc_control(rc);
3904
3905         trans = btrfs_join_transaction(rc->extent_root);
3906         if (IS_ERR(trans)) {
3907                 unset_reloc_control(rc);
3908                 /*
3909                  * extent tree is not a ref_cow tree and has no reloc_root to
3910                  * cleanup.  And callers are responsible to free the above
3911                  * block rsv.
3912                  */
3913                 return PTR_ERR(trans);
3914         }
3915         btrfs_commit_transaction(trans, rc->extent_root);
3916         return 0;
3917 }
3918
3919 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3920 {
3921         struct rb_root blocks = RB_ROOT;
3922         struct btrfs_key key;
3923         struct btrfs_trans_handle *trans = NULL;
3924         struct btrfs_path *path;
3925         struct btrfs_extent_item *ei;
3926         u64 flags;
3927         u32 item_size;
3928         int ret;
3929         int err = 0;
3930         int progress = 0;
3931
3932         path = btrfs_alloc_path();
3933         if (!path)
3934                 return -ENOMEM;
3935         path->reada = READA_FORWARD;
3936
3937         ret = prepare_to_relocate(rc);
3938         if (ret) {
3939                 err = ret;
3940                 goto out_free;
3941         }
3942
3943         while (1) {
3944                 rc->reserved_bytes = 0;
3945                 ret = btrfs_block_rsv_refill(rc->extent_root,
3946                                         rc->block_rsv, rc->block_rsv->size,
3947                                         BTRFS_RESERVE_FLUSH_ALL);
3948                 if (ret) {
3949                         err = ret;
3950                         break;
3951                 }
3952                 progress++;
3953                 trans = btrfs_start_transaction(rc->extent_root, 0);
3954                 if (IS_ERR(trans)) {
3955                         err = PTR_ERR(trans);
3956                         trans = NULL;
3957                         break;
3958                 }
3959 restart:
3960                 if (update_backref_cache(trans, &rc->backref_cache)) {
3961                         btrfs_end_transaction(trans, rc->extent_root);
3962                         continue;
3963                 }
3964
3965                 ret = find_next_extent(rc, path, &key);
3966                 if (ret < 0)
3967                         err = ret;
3968                 if (ret != 0)
3969                         break;
3970
3971                 rc->extents_found++;
3972
3973                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3974                                     struct btrfs_extent_item);
3975                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3976                 if (item_size >= sizeof(*ei)) {
3977                         flags = btrfs_extent_flags(path->nodes[0], ei);
3978                         ret = check_extent_flags(flags);
3979                         BUG_ON(ret);
3980
3981                 } else {
3982 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3983                         u64 ref_owner;
3984                         int path_change = 0;
3985
3986                         BUG_ON(item_size !=
3987                                sizeof(struct btrfs_extent_item_v0));
3988                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3989                                                   &path_change);
3990                         if (ret < 0) {
3991                                 err = ret;
3992                                 break;
3993                         }
3994                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3995                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3996                         else
3997                                 flags = BTRFS_EXTENT_FLAG_DATA;
3998
3999                         if (path_change) {
4000                                 btrfs_release_path(path);
4001
4002                                 path->search_commit_root = 1;
4003                                 path->skip_locking = 1;
4004                                 ret = btrfs_search_slot(NULL, rc->extent_root,
4005                                                         &key, path, 0, 0);
4006                                 if (ret < 0) {
4007                                         err = ret;
4008                                         break;
4009                                 }
4010                                 BUG_ON(ret > 0);
4011                         }
4012 #else
4013                         BUG();
4014 #endif
4015                 }
4016
4017                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4018                         ret = add_tree_block(rc, &key, path, &blocks);
4019                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4020                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4021                         ret = add_data_references(rc, &key, path, &blocks);
4022                 } else {
4023                         btrfs_release_path(path);
4024                         ret = 0;
4025                 }
4026                 if (ret < 0) {
4027                         err = ret;
4028                         break;
4029                 }
4030
4031                 if (!RB_EMPTY_ROOT(&blocks)) {
4032                         ret = relocate_tree_blocks(trans, rc, &blocks);
4033                         if (ret < 0) {
4034                                 /*
4035                                  * if we fail to relocate tree blocks, force to update
4036                                  * backref cache when committing transaction.
4037                                  */
4038                                 rc->backref_cache.last_trans = trans->transid - 1;
4039
4040                                 if (ret != -EAGAIN) {
4041                                         err = ret;
4042                                         break;
4043                                 }
4044                                 rc->extents_found--;
4045                                 rc->search_start = key.objectid;
4046                         }
4047                 }
4048
4049                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4050                 btrfs_btree_balance_dirty(rc->extent_root);
4051                 trans = NULL;
4052
4053                 if (rc->stage == MOVE_DATA_EXTENTS &&
4054                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4055                         rc->found_file_extent = 1;
4056                         ret = relocate_data_extent(rc->data_inode,
4057                                                    &key, &rc->cluster);
4058                         if (ret < 0) {
4059                                 err = ret;
4060                                 break;
4061                         }
4062                 }
4063         }
4064         if (trans && progress && err == -ENOSPC) {
4065                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4066                                               rc->block_group->flags);
4067                 if (ret == 1) {
4068                         err = 0;
4069                         progress = 0;
4070                         goto restart;
4071                 }
4072         }
4073
4074         btrfs_release_path(path);
4075         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4076
4077         if (trans) {
4078                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4079                 btrfs_btree_balance_dirty(rc->extent_root);
4080         }
4081
4082         if (!err) {
4083                 ret = relocate_file_extent_cluster(rc->data_inode,
4084                                                    &rc->cluster);
4085                 if (ret < 0)
4086                         err = ret;
4087         }
4088
4089         rc->create_reloc_tree = 0;
4090         set_reloc_control(rc);
4091
4092         backref_cache_cleanup(&rc->backref_cache);
4093         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4094
4095         err = prepare_to_merge(rc, err);
4096
4097         merge_reloc_roots(rc);
4098
4099         rc->merge_reloc_tree = 0;
4100         unset_reloc_control(rc);
4101         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4102
4103         /* get rid of pinned extents */
4104         trans = btrfs_join_transaction(rc->extent_root);
4105         if (IS_ERR(trans))
4106                 err = PTR_ERR(trans);
4107         else
4108                 btrfs_commit_transaction(trans, rc->extent_root);
4109 out_free:
4110         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4111         btrfs_free_path(path);
4112         return err;
4113 }
4114
4115 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4116                                  struct btrfs_root *root, u64 objectid)
4117 {
4118         struct btrfs_path *path;
4119         struct btrfs_inode_item *item;
4120         struct extent_buffer *leaf;
4121         int ret;
4122
4123         path = btrfs_alloc_path();
4124         if (!path)
4125                 return -ENOMEM;
4126
4127         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4128         if (ret)
4129                 goto out;
4130
4131         leaf = path->nodes[0];
4132         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4133         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4134         btrfs_set_inode_generation(leaf, item, 1);
4135         btrfs_set_inode_size(leaf, item, 0);
4136         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4137         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4138                                           BTRFS_INODE_PREALLOC);
4139         btrfs_mark_buffer_dirty(leaf);
4140 out:
4141         btrfs_free_path(path);
4142         return ret;
4143 }
4144
4145 /*
4146  * helper to create inode for data relocation.
4147  * the inode is in data relocation tree and its link count is 0
4148  */
4149 static noinline_for_stack
4150 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4151                                  struct btrfs_block_group_cache *group)
4152 {
4153         struct inode *inode = NULL;
4154         struct btrfs_trans_handle *trans;
4155         struct btrfs_root *root;
4156         struct btrfs_key key;
4157         u64 objectid;
4158         int err = 0;
4159
4160         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4161         if (IS_ERR(root))
4162                 return ERR_CAST(root);
4163
4164         trans = btrfs_start_transaction(root, 6);
4165         if (IS_ERR(trans))
4166                 return ERR_CAST(trans);
4167
4168         err = btrfs_find_free_objectid(root, &objectid);
4169         if (err)
4170                 goto out;
4171
4172         err = __insert_orphan_inode(trans, root, objectid);
4173         BUG_ON(err);
4174
4175         key.objectid = objectid;
4176         key.type = BTRFS_INODE_ITEM_KEY;
4177         key.offset = 0;
4178         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4179         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4180         BTRFS_I(inode)->index_cnt = group->key.objectid;
4181
4182         err = btrfs_orphan_add(trans, inode);
4183 out:
4184         btrfs_end_transaction(trans, root);
4185         btrfs_btree_balance_dirty(root);
4186         if (err) {
4187                 if (inode)
4188                         iput(inode);
4189                 inode = ERR_PTR(err);
4190         }
4191         return inode;
4192 }
4193
4194 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4195 {
4196         struct reloc_control *rc;
4197
4198         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4199         if (!rc)
4200                 return NULL;
4201
4202         INIT_LIST_HEAD(&rc->reloc_roots);
4203         backref_cache_init(&rc->backref_cache);
4204         mapping_tree_init(&rc->reloc_root_tree);
4205         extent_io_tree_init(&rc->processed_blocks,
4206                             fs_info->btree_inode->i_mapping);
4207         return rc;
4208 }
4209
4210 /*
4211  * function to relocate all extents in a block group.
4212  */
4213 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4214 {
4215         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4216         struct reloc_control *rc;
4217         struct inode *inode;
4218         struct btrfs_path *path;
4219         int ret;
4220         int rw = 0;
4221         int err = 0;
4222
4223         rc = alloc_reloc_control(fs_info);
4224         if (!rc)
4225                 return -ENOMEM;
4226
4227         rc->extent_root = extent_root;
4228
4229         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4230         BUG_ON(!rc->block_group);
4231
4232         ret = btrfs_inc_block_group_ro(extent_root, rc->block_group);
4233         if (ret) {
4234                 err = ret;
4235                 goto out;
4236         }
4237         rw = 1;
4238
4239         path = btrfs_alloc_path();
4240         if (!path) {
4241                 err = -ENOMEM;
4242                 goto out;
4243         }
4244
4245         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4246                                         path);
4247         btrfs_free_path(path);
4248
4249         if (!IS_ERR(inode))
4250                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4251         else
4252                 ret = PTR_ERR(inode);
4253
4254         if (ret && ret != -ENOENT) {
4255                 err = ret;
4256                 goto out;
4257         }
4258
4259         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4260         if (IS_ERR(rc->data_inode)) {
4261                 err = PTR_ERR(rc->data_inode);
4262                 rc->data_inode = NULL;
4263                 goto out;
4264         }
4265
4266         btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
4267                rc->block_group->key.objectid, rc->block_group->flags);
4268
4269         btrfs_wait_block_group_reservations(rc->block_group);
4270         btrfs_wait_nocow_writers(rc->block_group);
4271         btrfs_wait_ordered_roots(fs_info, -1,
4272                                  rc->block_group->key.objectid,
4273                                  rc->block_group->key.offset);
4274
4275         while (1) {
4276                 mutex_lock(&fs_info->cleaner_mutex);
4277                 ret = relocate_block_group(rc);
4278                 mutex_unlock(&fs_info->cleaner_mutex);
4279                 if (ret < 0) {
4280                         err = ret;
4281                         goto out;
4282                 }
4283
4284                 if (rc->extents_found == 0)
4285                         break;
4286
4287                 btrfs_info(extent_root->fs_info, "found %llu extents",
4288                         rc->extents_found);
4289
4290                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4291                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4292                                                        (u64)-1);
4293                         if (ret) {
4294                                 err = ret;
4295                                 goto out;
4296                         }
4297                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4298                                                  0, -1);
4299                         rc->stage = UPDATE_DATA_PTRS;
4300                 }
4301         }
4302
4303         WARN_ON(rc->block_group->pinned > 0);
4304         WARN_ON(rc->block_group->reserved > 0);
4305         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4306 out:
4307         if (err && rw)
4308                 btrfs_dec_block_group_ro(extent_root, rc->block_group);
4309         iput(rc->data_inode);
4310         btrfs_put_block_group(rc->block_group);
4311         kfree(rc);
4312         return err;
4313 }
4314
4315 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4316 {
4317         struct btrfs_trans_handle *trans;
4318         int ret, err;
4319
4320         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4321         if (IS_ERR(trans))
4322                 return PTR_ERR(trans);
4323
4324         memset(&root->root_item.drop_progress, 0,
4325                 sizeof(root->root_item.drop_progress));
4326         root->root_item.drop_level = 0;
4327         btrfs_set_root_refs(&root->root_item, 0);
4328         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4329                                 &root->root_key, &root->root_item);
4330
4331         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4332         if (err)
4333                 return err;
4334         return ret;
4335 }
4336
4337 /*
4338  * recover relocation interrupted by system crash.
4339  *
4340  * this function resumes merging reloc trees with corresponding fs trees.
4341  * this is important for keeping the sharing of tree blocks
4342  */
4343 int btrfs_recover_relocation(struct btrfs_root *root)
4344 {
4345         LIST_HEAD(reloc_roots);
4346         struct btrfs_key key;
4347         struct btrfs_root *fs_root;
4348         struct btrfs_root *reloc_root;
4349         struct btrfs_path *path;
4350         struct extent_buffer *leaf;
4351         struct reloc_control *rc = NULL;
4352         struct btrfs_trans_handle *trans;
4353         int ret;
4354         int err = 0;
4355
4356         path = btrfs_alloc_path();
4357         if (!path)
4358                 return -ENOMEM;
4359         path->reada = READA_BACK;
4360
4361         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4362         key.type = BTRFS_ROOT_ITEM_KEY;
4363         key.offset = (u64)-1;
4364
4365         while (1) {
4366                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4367                                         path, 0, 0);
4368                 if (ret < 0) {
4369                         err = ret;
4370                         goto out;
4371                 }
4372                 if (ret > 0) {
4373                         if (path->slots[0] == 0)
4374                                 break;
4375                         path->slots[0]--;
4376                 }
4377                 leaf = path->nodes[0];
4378                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4379                 btrfs_release_path(path);
4380
4381                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4382                     key.type != BTRFS_ROOT_ITEM_KEY)
4383                         break;
4384
4385                 reloc_root = btrfs_read_fs_root(root, &key);
4386                 if (IS_ERR(reloc_root)) {
4387                         err = PTR_ERR(reloc_root);
4388                         goto out;
4389                 }
4390
4391                 list_add(&reloc_root->root_list, &reloc_roots);
4392
4393                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4394                         fs_root = read_fs_root(root->fs_info,
4395                                                reloc_root->root_key.offset);
4396                         if (IS_ERR(fs_root)) {
4397                                 ret = PTR_ERR(fs_root);
4398                                 if (ret != -ENOENT) {
4399                                         err = ret;
4400                                         goto out;
4401                                 }
4402                                 ret = mark_garbage_root(reloc_root);
4403                                 if (ret < 0) {
4404                                         err = ret;
4405                                         goto out;
4406                                 }
4407                         }
4408                 }
4409
4410                 if (key.offset == 0)
4411                         break;
4412
4413                 key.offset--;
4414         }
4415         btrfs_release_path(path);
4416
4417         if (list_empty(&reloc_roots))
4418                 goto out;
4419
4420         rc = alloc_reloc_control(root->fs_info);
4421         if (!rc) {
4422                 err = -ENOMEM;
4423                 goto out;
4424         }
4425
4426         rc->extent_root = root->fs_info->extent_root;
4427
4428         set_reloc_control(rc);
4429
4430         trans = btrfs_join_transaction(rc->extent_root);
4431         if (IS_ERR(trans)) {
4432                 unset_reloc_control(rc);
4433                 err = PTR_ERR(trans);
4434                 goto out_free;
4435         }
4436
4437         rc->merge_reloc_tree = 1;
4438
4439         while (!list_empty(&reloc_roots)) {
4440                 reloc_root = list_entry(reloc_roots.next,
4441                                         struct btrfs_root, root_list);
4442                 list_del(&reloc_root->root_list);
4443
4444                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4445                         list_add_tail(&reloc_root->root_list,
4446                                       &rc->reloc_roots);
4447                         continue;
4448                 }
4449
4450                 fs_root = read_fs_root(root->fs_info,
4451                                        reloc_root->root_key.offset);
4452                 if (IS_ERR(fs_root)) {
4453                         err = PTR_ERR(fs_root);
4454                         goto out_free;
4455                 }
4456
4457                 err = __add_reloc_root(reloc_root);
4458                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4459                 fs_root->reloc_root = reloc_root;
4460         }
4461
4462         err = btrfs_commit_transaction(trans, rc->extent_root);
4463         if (err)
4464                 goto out_free;
4465
4466         merge_reloc_roots(rc);
4467
4468         unset_reloc_control(rc);
4469
4470         trans = btrfs_join_transaction(rc->extent_root);
4471         if (IS_ERR(trans))
4472                 err = PTR_ERR(trans);
4473         else
4474                 err = btrfs_commit_transaction(trans, rc->extent_root);
4475 out_free:
4476         kfree(rc);
4477 out:
4478         if (!list_empty(&reloc_roots))
4479                 free_reloc_roots(&reloc_roots);
4480
4481         btrfs_free_path(path);
4482
4483         if (err == 0) {
4484                 /* cleanup orphan inode in data relocation tree */
4485                 fs_root = read_fs_root(root->fs_info,
4486                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4487                 if (IS_ERR(fs_root))
4488                         err = PTR_ERR(fs_root);
4489                 else
4490                         err = btrfs_orphan_cleanup(fs_root);
4491         }
4492         return err;
4493 }
4494
4495 /*
4496  * helper to add ordered checksum for data relocation.
4497  *
4498  * cloning checksum properly handles the nodatasum extents.
4499  * it also saves CPU time to re-calculate the checksum.
4500  */
4501 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4502 {
4503         struct btrfs_ordered_sum *sums;
4504         struct btrfs_ordered_extent *ordered;
4505         struct btrfs_root *root = BTRFS_I(inode)->root;
4506         int ret;
4507         u64 disk_bytenr;
4508         u64 new_bytenr;
4509         LIST_HEAD(list);
4510
4511         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4512         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4513
4514         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4515         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4516                                        disk_bytenr + len - 1, &list, 0);
4517         if (ret)
4518                 goto out;
4519
4520         while (!list_empty(&list)) {
4521                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4522                 list_del_init(&sums->list);
4523
4524                 /*
4525                  * We need to offset the new_bytenr based on where the csum is.
4526                  * We need to do this because we will read in entire prealloc
4527                  * extents but we may have written to say the middle of the
4528                  * prealloc extent, so we need to make sure the csum goes with
4529                  * the right disk offset.
4530                  *
4531                  * We can do this because the data reloc inode refers strictly
4532                  * to the on disk bytes, so we don't have to worry about
4533                  * disk_len vs real len like with real inodes since it's all
4534                  * disk length.
4535                  */
4536                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4537                 sums->bytenr = new_bytenr;
4538
4539                 btrfs_add_ordered_sum(inode, ordered, sums);
4540         }
4541 out:
4542         btrfs_put_ordered_extent(ordered);
4543         return ret;
4544 }
4545
4546 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4547                           struct btrfs_root *root, struct extent_buffer *buf,
4548                           struct extent_buffer *cow)
4549 {
4550         struct reloc_control *rc;
4551         struct backref_node *node;
4552         int first_cow = 0;
4553         int level;
4554         int ret = 0;
4555
4556         rc = root->fs_info->reloc_ctl;
4557         if (!rc)
4558                 return 0;
4559
4560         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4561                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4562
4563         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4564                 if (buf == root->node)
4565                         __update_reloc_root(root, cow->start);
4566         }
4567
4568         level = btrfs_header_level(buf);
4569         if (btrfs_header_generation(buf) <=
4570             btrfs_root_last_snapshot(&root->root_item))
4571                 first_cow = 1;
4572
4573         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4574             rc->create_reloc_tree) {
4575                 WARN_ON(!first_cow && level == 0);
4576
4577                 node = rc->backref_cache.path[level];
4578                 BUG_ON(node->bytenr != buf->start &&
4579                        node->new_bytenr != buf->start);
4580
4581                 drop_node_buffer(node);
4582                 extent_buffer_get(cow);
4583                 node->eb = cow;
4584                 node->new_bytenr = cow->start;
4585
4586                 if (!node->pending) {
4587                         list_move_tail(&node->list,
4588                                        &rc->backref_cache.pending[level]);
4589                         node->pending = 1;
4590                 }
4591
4592                 if (first_cow)
4593                         __mark_block_processed(rc, node);
4594
4595                 if (first_cow && level > 0)
4596                         rc->nodes_relocated += buf->len;
4597         }
4598
4599         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4600                 ret = replace_file_extents(trans, rc, root, cow);
4601         return ret;
4602 }
4603
4604 /*
4605  * called before creating snapshot. it calculates metadata reservation
4606  * required for relocating tree blocks in the snapshot
4607  */
4608 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4609                               u64 *bytes_to_reserve)
4610 {
4611         struct btrfs_root *root;
4612         struct reloc_control *rc;
4613
4614         root = pending->root;
4615         if (!root->reloc_root)
4616                 return;
4617
4618         rc = root->fs_info->reloc_ctl;
4619         if (!rc->merge_reloc_tree)
4620                 return;
4621
4622         root = root->reloc_root;
4623         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4624         /*
4625          * relocation is in the stage of merging trees. the space
4626          * used by merging a reloc tree is twice the size of
4627          * relocated tree nodes in the worst case. half for cowing
4628          * the reloc tree, half for cowing the fs tree. the space
4629          * used by cowing the reloc tree will be freed after the
4630          * tree is dropped. if we create snapshot, cowing the fs
4631          * tree may use more space than it frees. so we need
4632          * reserve extra space.
4633          */
4634         *bytes_to_reserve += rc->nodes_relocated;
4635 }
4636
4637 /*
4638  * called after snapshot is created. migrate block reservation
4639  * and create reloc root for the newly created snapshot
4640  */
4641 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4642                                struct btrfs_pending_snapshot *pending)
4643 {
4644         struct btrfs_root *root = pending->root;
4645         struct btrfs_root *reloc_root;
4646         struct btrfs_root *new_root;
4647         struct reloc_control *rc;
4648         int ret;
4649
4650         if (!root->reloc_root)
4651                 return 0;
4652
4653         rc = root->fs_info->reloc_ctl;
4654         rc->merging_rsv_size += rc->nodes_relocated;
4655
4656         if (rc->merge_reloc_tree) {
4657                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4658                                               rc->block_rsv,
4659                                               rc->nodes_relocated, 1);
4660                 if (ret)
4661                         return ret;
4662         }
4663
4664         new_root = pending->snap;
4665         reloc_root = create_reloc_root(trans, root->reloc_root,
4666                                        new_root->root_key.objectid);
4667         if (IS_ERR(reloc_root))
4668                 return PTR_ERR(reloc_root);
4669
4670         ret = __add_reloc_root(reloc_root);
4671         BUG_ON(ret < 0);
4672         new_root->reloc_root = reloc_root;
4673
4674         if (rc->create_reloc_tree)
4675                 ret = clone_backref_node(trans, rc, root, reloc_root);
4676         return ret;
4677 }