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                 BUG_ON(!list_empty(&cache->pending[i]));
 239         BUG_ON(!list_empty(&cache->changed));
 240         BUG_ON(!list_empty(&cache->detached));
 241         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
 242         BUG_ON(cache->nr_nodes);
 243         BUG_ON(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 refernece the block, and then check
 672  * bakcrefs 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 proces */
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                         free_backref_node(cache, lower);
1175                 }
1176                 return ERR_PTR(err);
1177         }
1178         ASSERT(!node || !node->detached);
1179         return node;
1180 }
1181
1182 /*
1183  * helper to add backref node for the newly created snapshot.
1184  * the backref node is created by cloning backref node that
1185  * corresponds to root of source tree
1186  */
1187 static int clone_backref_node(struct btrfs_trans_handle *trans,
1188                               struct reloc_control *rc,
1189                               struct btrfs_root *src,
1190                               struct btrfs_root *dest)
1191 {
1192         struct btrfs_root *reloc_root = src->reloc_root;
1193         struct backref_cache *cache = &rc->backref_cache;
1194         struct backref_node *node = NULL;
1195         struct backref_node *new_node;
1196         struct backref_edge *edge;
1197         struct backref_edge *new_edge;
1198         struct rb_node *rb_node;
1199
1200         if (cache->last_trans > 0)
1201                 update_backref_cache(trans, cache);
1202
1203         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1204         if (rb_node) {
1205                 node = rb_entry(rb_node, struct backref_node, rb_node);
1206                 if (node->detached)
1207                         node = NULL;
1208                 else
1209                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1210         }
1211
1212         if (!node) {
1213                 rb_node = tree_search(&cache->rb_root,
1214                                       reloc_root->commit_root->start);
1215                 if (rb_node) {
1216                         node = rb_entry(rb_node, struct backref_node,
1217                                         rb_node);
1218                         BUG_ON(node->detached);
1219                 }
1220         }
1221
1222         if (!node)
1223                 return 0;
1224
1225         new_node = alloc_backref_node(cache);
1226         if (!new_node)
1227                 return -ENOMEM;
1228
1229         new_node->bytenr = dest->node->start;
1230         new_node->level = node->level;
1231         new_node->lowest = node->lowest;
1232         new_node->checked = 1;
1233         new_node->root = dest;
1234
1235         if (!node->lowest) {
1236                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1237                         new_edge = alloc_backref_edge(cache);
1238                         if (!new_edge)
1239                                 goto fail;
1240
1241                         new_edge->node[UPPER] = new_node;
1242                         new_edge->node[LOWER] = edge->node[LOWER];
1243                         list_add_tail(&new_edge->list[UPPER],
1244                                       &new_node->lower);
1245                 }
1246         } else {
1247                 list_add_tail(&new_node->lower, &cache->leaves);
1248         }
1249
1250         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1251                               &new_node->rb_node);
1252         if (rb_node)
1253                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1254
1255         if (!new_node->lowest) {
1256                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1257                         list_add_tail(&new_edge->list[LOWER],
1258                                       &new_edge->node[LOWER]->upper);
1259                 }
1260         }
1261         return 0;
1262 fail:
1263         while (!list_empty(&new_node->lower)) {
1264                 new_edge = list_entry(new_node->lower.next,
1265                                       struct backref_edge, list[UPPER]);
1266                 list_del(&new_edge->list[UPPER]);
1267                 free_backref_edge(cache, new_edge);
1268         }
1269         free_backref_node(cache, new_node);
1270         return -ENOMEM;
1271 }
1272
1273 /*
1274  * helper to add 'address of tree root -> reloc tree' mapping
1275  */
1276 static int __must_check __add_reloc_root(struct btrfs_root *root)
1277 {
1278         struct rb_node *rb_node;
1279         struct mapping_node *node;
1280         struct reloc_control *rc = root->fs_info->reloc_ctl;
1281
1282         node = kmalloc(sizeof(*node), GFP_NOFS);
1283         if (!node)
1284                 return -ENOMEM;
1285
1286         node->bytenr = root->node->start;
1287         node->data = root;
1288
1289         spin_lock(&rc->reloc_root_tree.lock);
1290         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1291                               node->bytenr, &node->rb_node);
1292         spin_unlock(&rc->reloc_root_tree.lock);
1293         if (rb_node) {
1294                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1295                             "for start=%llu while inserting into relocation "
1296                             "tree", node->bytenr);
1297                 kfree(node);
1298                 return -EEXIST;
1299         }
1300
1301         list_add_tail(&root->root_list, &rc->reloc_roots);
1302         return 0;
1303 }
1304
1305 /*
1306  * helper to delete the 'address of tree root -> reloc tree'
1307  * mapping
1308  */
1309 static void __del_reloc_root(struct btrfs_root *root)
1310 {
1311         struct rb_node *rb_node;
1312         struct mapping_node *node = NULL;
1313         struct reloc_control *rc = root->fs_info->reloc_ctl;
1314
1315         spin_lock(&rc->reloc_root_tree.lock);
1316         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1317                               root->node->start);
1318         if (rb_node) {
1319                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1320                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1321         }
1322         spin_unlock(&rc->reloc_root_tree.lock);
1323
1324         if (!node)
1325                 return;
1326         BUG_ON((struct btrfs_root *)node->data != root);
1327
1328         spin_lock(&root->fs_info->trans_lock);
1329         list_del_init(&root->root_list);
1330         spin_unlock(&root->fs_info->trans_lock);
1331         kfree(node);
1332 }
1333
1334 /*
1335  * helper to update the 'address of tree root -> reloc tree'
1336  * mapping
1337  */
1338 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1339 {
1340         struct rb_node *rb_node;
1341         struct mapping_node *node = NULL;
1342         struct reloc_control *rc = root->fs_info->reloc_ctl;
1343
1344         spin_lock(&rc->reloc_root_tree.lock);
1345         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1346                               root->node->start);
1347         if (rb_node) {
1348                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1349                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1350         }
1351         spin_unlock(&rc->reloc_root_tree.lock);
1352
1353         if (!node)
1354                 return 0;
1355         BUG_ON((struct btrfs_root *)node->data != root);
1356
1357         spin_lock(&rc->reloc_root_tree.lock);
1358         node->bytenr = new_bytenr;
1359         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1360                               node->bytenr, &node->rb_node);
1361         spin_unlock(&rc->reloc_root_tree.lock);
1362         if (rb_node)
1363                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1364         return 0;
1365 }
1366
1367 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1368                                         struct btrfs_root *root, u64 objectid)
1369 {
1370         struct btrfs_root *reloc_root;
1371         struct extent_buffer *eb;
1372         struct btrfs_root_item *root_item;
1373         struct btrfs_key root_key;
1374         u64 last_snap = 0;
1375         int ret;
1376
1377         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1378         BUG_ON(!root_item);
1379
1380         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1381         root_key.type = BTRFS_ROOT_ITEM_KEY;
1382         root_key.offset = objectid;
1383
1384         if (root->root_key.objectid == objectid) {
1385                 /* called by btrfs_init_reloc_root */
1386                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1387                                       BTRFS_TREE_RELOC_OBJECTID);
1388                 BUG_ON(ret);
1389
1390                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1391                 btrfs_set_root_last_snapshot(&root->root_item,
1392                                              trans->transid - 1);
1393         } else {
1394                 /*
1395                  * called by btrfs_reloc_post_snapshot_hook.
1396                  * the source tree is a reloc tree, all tree blocks
1397                  * modified after it was created have RELOC flag
1398                  * set in their headers. so it's OK to not update
1399                  * the 'last_snapshot'.
1400                  */
1401                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1402                                       BTRFS_TREE_RELOC_OBJECTID);
1403                 BUG_ON(ret);
1404         }
1405
1406         memcpy(root_item, &root->root_item, sizeof(*root_item));
1407         btrfs_set_root_bytenr(root_item, eb->start);
1408         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1409         btrfs_set_root_generation(root_item, trans->transid);
1410
1411         if (root->root_key.objectid == objectid) {
1412                 btrfs_set_root_refs(root_item, 0);
1413                 memset(&root_item->drop_progress, 0,
1414                        sizeof(struct btrfs_disk_key));
1415                 root_item->drop_level = 0;
1416                 /*
1417                  * abuse rtransid, it is safe because it is impossible to
1418                  * receive data into a relocation tree.
1419                  */
1420                 btrfs_set_root_rtransid(root_item, last_snap);
1421                 btrfs_set_root_otransid(root_item, trans->transid);
1422         }
1423
1424         btrfs_tree_unlock(eb);
1425         free_extent_buffer(eb);
1426
1427         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1428                                 &root_key, root_item);
1429         BUG_ON(ret);
1430         kfree(root_item);
1431
1432         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1433         BUG_ON(IS_ERR(reloc_root));
1434         reloc_root->last_trans = trans->transid;
1435         return reloc_root;
1436 }
1437
1438 /*
1439  * create reloc tree for a given fs tree. reloc tree is just a
1440  * snapshot of the fs tree with special root objectid.
1441  */
1442 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1443                           struct btrfs_root *root)
1444 {
1445         struct btrfs_root *reloc_root;
1446         struct reloc_control *rc = root->fs_info->reloc_ctl;
1447         struct btrfs_block_rsv *rsv;
1448         int clear_rsv = 0;
1449         int ret;
1450
1451         if (root->reloc_root) {
1452                 reloc_root = root->reloc_root;
1453                 reloc_root->last_trans = trans->transid;
1454                 return 0;
1455         }
1456
1457         if (!rc || !rc->create_reloc_tree ||
1458             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1459                 return 0;
1460
1461         if (!trans->reloc_reserved) {
1462                 rsv = trans->block_rsv;
1463                 trans->block_rsv = rc->block_rsv;
1464                 clear_rsv = 1;
1465         }
1466         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1467         if (clear_rsv)
1468                 trans->block_rsv = rsv;
1469
1470         ret = __add_reloc_root(reloc_root);
1471         BUG_ON(ret < 0);
1472         root->reloc_root = reloc_root;
1473         return 0;
1474 }
1475
1476 /*
1477  * update root item of reloc tree
1478  */
1479 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1480                             struct btrfs_root *root)
1481 {
1482         struct btrfs_root *reloc_root;
1483         struct btrfs_root_item *root_item;
1484         int ret;
1485
1486         if (!root->reloc_root)
1487                 goto out;
1488
1489         reloc_root = root->reloc_root;
1490         root_item = &reloc_root->root_item;
1491
1492         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1493             btrfs_root_refs(root_item) == 0) {
1494                 root->reloc_root = NULL;
1495                 __del_reloc_root(reloc_root);
1496         }
1497
1498         if (reloc_root->commit_root != reloc_root->node) {
1499                 btrfs_set_root_node(root_item, reloc_root->node);
1500                 free_extent_buffer(reloc_root->commit_root);
1501                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1502         }
1503
1504         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1505                                 &reloc_root->root_key, root_item);
1506         BUG_ON(ret);
1507
1508 out:
1509         return 0;
1510 }
1511
1512 /*
1513  * helper to find first cached inode with inode number >= objectid
1514  * in a subvolume
1515  */
1516 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1517 {
1518         struct rb_node *node;
1519         struct rb_node *prev;
1520         struct btrfs_inode *entry;
1521         struct inode *inode;
1522
1523         spin_lock(&root->inode_lock);
1524 again:
1525         node = root->inode_tree.rb_node;
1526         prev = NULL;
1527         while (node) {
1528                 prev = node;
1529                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1530
1531                 if (objectid < btrfs_ino(&entry->vfs_inode))
1532                         node = node->rb_left;
1533                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1534                         node = node->rb_right;
1535                 else
1536                         break;
1537         }
1538         if (!node) {
1539                 while (prev) {
1540                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1541                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1542                                 node = prev;
1543                                 break;
1544                         }
1545                         prev = rb_next(prev);
1546                 }
1547         }
1548         while (node) {
1549                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1550                 inode = igrab(&entry->vfs_inode);
1551                 if (inode) {
1552                         spin_unlock(&root->inode_lock);
1553                         return inode;
1554                 }
1555
1556                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1557                 if (cond_resched_lock(&root->inode_lock))
1558                         goto again;
1559
1560                 node = rb_next(node);
1561         }
1562         spin_unlock(&root->inode_lock);
1563         return NULL;
1564 }
1565
1566 static int in_block_group(u64 bytenr,
1567                           struct btrfs_block_group_cache *block_group)
1568 {
1569         if (bytenr >= block_group->key.objectid &&
1570             bytenr < block_group->key.objectid + block_group->key.offset)
1571                 return 1;
1572         return 0;
1573 }
1574
1575 /*
1576  * get new location of data
1577  */
1578 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1579                             u64 bytenr, u64 num_bytes)
1580 {
1581         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1582         struct btrfs_path *path;
1583         struct btrfs_file_extent_item *fi;
1584         struct extent_buffer *leaf;
1585         int ret;
1586
1587         path = btrfs_alloc_path();
1588         if (!path)
1589                 return -ENOMEM;
1590
1591         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1592         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1593                                        bytenr, 0);
1594         if (ret < 0)
1595                 goto out;
1596         if (ret > 0) {
1597                 ret = -ENOENT;
1598                 goto out;
1599         }
1600
1601         leaf = path->nodes[0];
1602         fi = btrfs_item_ptr(leaf, path->slots[0],
1603                             struct btrfs_file_extent_item);
1604
1605         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1606                btrfs_file_extent_compression(leaf, fi) ||
1607                btrfs_file_extent_encryption(leaf, fi) ||
1608                btrfs_file_extent_other_encoding(leaf, fi));
1609
1610         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1611                 ret = -EINVAL;
1612                 goto out;
1613         }
1614
1615         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1616         ret = 0;
1617 out:
1618         btrfs_free_path(path);
1619         return ret;
1620 }
1621
1622 /*
1623  * update file extent items in the tree leaf to point to
1624  * the new locations.
1625  */
1626 static noinline_for_stack
1627 int replace_file_extents(struct btrfs_trans_handle *trans,
1628                          struct reloc_control *rc,
1629                          struct btrfs_root *root,
1630                          struct extent_buffer *leaf)
1631 {
1632         struct btrfs_key key;
1633         struct btrfs_file_extent_item *fi;
1634         struct inode *inode = NULL;
1635         u64 parent;
1636         u64 bytenr;
1637         u64 new_bytenr = 0;
1638         u64 num_bytes;
1639         u64 end;
1640         u32 nritems;
1641         u32 i;
1642         int ret = 0;
1643         int first = 1;
1644         int dirty = 0;
1645
1646         if (rc->stage != UPDATE_DATA_PTRS)
1647                 return 0;
1648
1649         /* reloc trees always use full backref */
1650         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1651                 parent = leaf->start;
1652         else
1653                 parent = 0;
1654
1655         nritems = btrfs_header_nritems(leaf);
1656         for (i = 0; i < nritems; i++) {
1657                 cond_resched();
1658                 btrfs_item_key_to_cpu(leaf, &key, i);
1659                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1660                         continue;
1661                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1662                 if (btrfs_file_extent_type(leaf, fi) ==
1663                     BTRFS_FILE_EXTENT_INLINE)
1664                         continue;
1665                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1666                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1667                 if (bytenr == 0)
1668                         continue;
1669                 if (!in_block_group(bytenr, rc->block_group))
1670                         continue;
1671
1672                 /*
1673                  * if we are modifying block in fs tree, wait for readpage
1674                  * to complete and drop the extent cache
1675                  */
1676                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1677                         if (first) {
1678                                 inode = find_next_inode(root, key.objectid);
1679                                 first = 0;
1680                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1681                                 btrfs_add_delayed_iput(inode);
1682                                 inode = find_next_inode(root, key.objectid);
1683                         }
1684                         if (inode && btrfs_ino(inode) == key.objectid) {
1685                                 end = key.offset +
1686                                       btrfs_file_extent_num_bytes(leaf, fi);
1687                                 WARN_ON(!IS_ALIGNED(key.offset,
1688                                                     root->sectorsize));
1689                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1690                                 end--;
1691                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1692                                                       key.offset, end);
1693                                 if (!ret)
1694                                         continue;
1695
1696                                 btrfs_drop_extent_cache(inode, key.offset, end,
1697                                                         1);
1698                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1699                                               key.offset, end);
1700                         }
1701                 }
1702
1703                 ret = get_new_location(rc->data_inode, &new_bytenr,
1704                                        bytenr, num_bytes);
1705                 if (ret) {
1706                         /*
1707                          * Don't have to abort since we've not changed anything
1708                          * in the file extent yet.
1709                          */
1710                         break;
1711                 }
1712
1713                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1714                 dirty = 1;
1715
1716                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1717                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1718                                            num_bytes, parent,
1719                                            btrfs_header_owner(leaf),
1720                                            key.objectid, key.offset);
1721                 if (ret) {
1722                         btrfs_abort_transaction(trans, root, ret);
1723                         break;
1724                 }
1725
1726                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1727                                         parent, btrfs_header_owner(leaf),
1728                                         key.objectid, key.offset);
1729                 if (ret) {
1730                         btrfs_abort_transaction(trans, root, ret);
1731                         break;
1732                 }
1733         }
1734         if (dirty)
1735                 btrfs_mark_buffer_dirty(leaf);
1736         if (inode)
1737                 btrfs_add_delayed_iput(inode);
1738         return ret;
1739 }
1740
1741 static noinline_for_stack
1742 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1743                      struct btrfs_path *path, int level)
1744 {
1745         struct btrfs_disk_key key1;
1746         struct btrfs_disk_key key2;
1747         btrfs_node_key(eb, &key1, slot);
1748         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1749         return memcmp(&key1, &key2, sizeof(key1));
1750 }
1751
1752 /*
1753  * try to replace tree blocks in fs tree with the new blocks
1754  * in reloc tree. tree blocks haven't been modified since the
1755  * reloc tree was create can be replaced.
1756  *
1757  * if a block was replaced, level of the block + 1 is returned.
1758  * if no block got replaced, 0 is returned. if there are other
1759  * errors, a negative error number is returned.
1760  */
1761 static noinline_for_stack
1762 int replace_path(struct btrfs_trans_handle *trans,
1763                  struct btrfs_root *dest, struct btrfs_root *src,
1764                  struct btrfs_path *path, struct btrfs_key *next_key,
1765                  int lowest_level, int max_level)
1766 {
1767         struct extent_buffer *eb;
1768         struct extent_buffer *parent;
1769         struct btrfs_key key;
1770         u64 old_bytenr;
1771         u64 new_bytenr;
1772         u64 old_ptr_gen;
1773         u64 new_ptr_gen;
1774         u64 last_snapshot;
1775         u32 blocksize;
1776         int cow = 0;
1777         int level;
1778         int ret;
1779         int slot;
1780
1781         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1782         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1783
1784         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1785 again:
1786         slot = path->slots[lowest_level];
1787         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1788
1789         eb = btrfs_lock_root_node(dest);
1790         btrfs_set_lock_blocking(eb);
1791         level = btrfs_header_level(eb);
1792
1793         if (level < lowest_level) {
1794                 btrfs_tree_unlock(eb);
1795                 free_extent_buffer(eb);
1796                 return 0;
1797         }
1798
1799         if (cow) {
1800                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1801                 BUG_ON(ret);
1802         }
1803         btrfs_set_lock_blocking(eb);
1804
1805         if (next_key) {
1806                 next_key->objectid = (u64)-1;
1807                 next_key->type = (u8)-1;
1808                 next_key->offset = (u64)-1;
1809         }
1810
1811         parent = eb;
1812         while (1) {
1813                 level = btrfs_header_level(parent);
1814                 BUG_ON(level < lowest_level);
1815
1816                 ret = btrfs_bin_search(parent, &key, level, &slot);
1817                 if (ret && slot > 0)
1818                         slot--;
1819
1820                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1821                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1822
1823                 old_bytenr = btrfs_node_blockptr(parent, slot);
1824                 blocksize = dest->nodesize;
1825                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1826
1827                 if (level <= max_level) {
1828                         eb = path->nodes[level];
1829                         new_bytenr = btrfs_node_blockptr(eb,
1830                                                         path->slots[level]);
1831                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1832                                                         path->slots[level]);
1833                 } else {
1834                         new_bytenr = 0;
1835                         new_ptr_gen = 0;
1836                 }
1837
1838                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1839                         ret = level;
1840                         break;
1841                 }
1842
1843                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1844                     memcmp_node_keys(parent, slot, path, level)) {
1845                         if (level <= lowest_level) {
1846                                 ret = 0;
1847                                 break;
1848                         }
1849
1850                         eb = read_tree_block(dest, old_bytenr, old_ptr_gen);
1851                         if (IS_ERR(eb)) {
1852                                 ret = PTR_ERR(eb);
1853                         } else if (!extent_buffer_uptodate(eb)) {
1854                                 ret = -EIO;
1855                                 free_extent_buffer(eb);
1856                                 break;
1857                         }
1858                         btrfs_tree_lock(eb);
1859                         if (cow) {
1860                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1861                                                       slot, &eb);
1862                                 BUG_ON(ret);
1863                         }
1864                         btrfs_set_lock_blocking(eb);
1865
1866                         btrfs_tree_unlock(parent);
1867                         free_extent_buffer(parent);
1868
1869                         parent = eb;
1870                         continue;
1871                 }
1872
1873                 if (!cow) {
1874                         btrfs_tree_unlock(parent);
1875                         free_extent_buffer(parent);
1876                         cow = 1;
1877                         goto again;
1878                 }
1879
1880                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1881                                       path->slots[level]);
1882                 btrfs_release_path(path);
1883
1884                 path->lowest_level = level;
1885                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1886                 path->lowest_level = 0;
1887                 BUG_ON(ret);
1888
1889                 /*
1890                  * swap blocks in fs tree and reloc tree.
1891                  */
1892                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1893                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1894                 btrfs_mark_buffer_dirty(parent);
1895
1896                 btrfs_set_node_blockptr(path->nodes[level],
1897                                         path->slots[level], old_bytenr);
1898                 btrfs_set_node_ptr_generation(path->nodes[level],
1899                                               path->slots[level], old_ptr_gen);
1900                 btrfs_mark_buffer_dirty(path->nodes[level]);
1901
1902                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1903                                         path->nodes[level]->start,
1904                                         src->root_key.objectid, level - 1, 0);
1905                 BUG_ON(ret);
1906                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1907                                         0, dest->root_key.objectid, level - 1,
1908                                         0);
1909                 BUG_ON(ret);
1910
1911                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1912                                         path->nodes[level]->start,
1913                                         src->root_key.objectid, level - 1, 0);
1914                 BUG_ON(ret);
1915
1916                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1917                                         0, dest->root_key.objectid, level - 1,
1918                                         0);
1919                 BUG_ON(ret);
1920
1921                 btrfs_unlock_up_safe(path, 0);
1922
1923                 ret = level;
1924                 break;
1925         }
1926         btrfs_tree_unlock(parent);
1927         free_extent_buffer(parent);
1928         return ret;
1929 }
1930
1931 /*
1932  * helper to find next relocated block in reloc tree
1933  */
1934 static noinline_for_stack
1935 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1936                        int *level)
1937 {
1938         struct extent_buffer *eb;
1939         int i;
1940         u64 last_snapshot;
1941         u32 nritems;
1942
1943         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1944
1945         for (i = 0; i < *level; i++) {
1946                 free_extent_buffer(path->nodes[i]);
1947                 path->nodes[i] = NULL;
1948         }
1949
1950         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1951                 eb = path->nodes[i];
1952                 nritems = btrfs_header_nritems(eb);
1953                 while (path->slots[i] + 1 < nritems) {
1954                         path->slots[i]++;
1955                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1956                             last_snapshot)
1957                                 continue;
1958
1959                         *level = i;
1960                         return 0;
1961                 }
1962                 free_extent_buffer(path->nodes[i]);
1963                 path->nodes[i] = NULL;
1964         }
1965         return 1;
1966 }
1967
1968 /*
1969  * walk down reloc tree to find relocated block of lowest level
1970  */
1971 static noinline_for_stack
1972 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1973                          int *level)
1974 {
1975         struct extent_buffer *eb = NULL;
1976         int i;
1977         u64 bytenr;
1978         u64 ptr_gen = 0;
1979         u64 last_snapshot;
1980         u32 nritems;
1981
1982         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1983
1984         for (i = *level; i > 0; i--) {
1985                 eb = path->nodes[i];
1986                 nritems = btrfs_header_nritems(eb);
1987                 while (path->slots[i] < nritems) {
1988                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1989                         if (ptr_gen > last_snapshot)
1990                                 break;
1991                         path->slots[i]++;
1992                 }
1993                 if (path->slots[i] >= nritems) {
1994                         if (i == *level)
1995                                 break;
1996                         *level = i + 1;
1997                         return 0;
1998                 }
1999                 if (i == 1) {
2000                         *level = i;
2001                         return 0;
2002                 }
2003
2004                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2005                 eb = read_tree_block(root, bytenr, ptr_gen);
2006                 if (IS_ERR(eb)) {
2007                         return PTR_ERR(eb);
2008                 } else if (!extent_buffer_uptodate(eb)) {
2009                         free_extent_buffer(eb);
2010                         return -EIO;
2011                 }
2012                 BUG_ON(btrfs_header_level(eb) != i - 1);
2013                 path->nodes[i - 1] = eb;
2014                 path->slots[i - 1] = 0;
2015         }
2016         return 1;
2017 }
2018
2019 /*
2020  * invalidate extent cache for file extents whose key in range of
2021  * [min_key, max_key)
2022  */
2023 static int invalidate_extent_cache(struct btrfs_root *root,
2024                                    struct btrfs_key *min_key,
2025                                    struct btrfs_key *max_key)
2026 {
2027         struct inode *inode = NULL;
2028         u64 objectid;
2029         u64 start, end;
2030         u64 ino;
2031
2032         objectid = min_key->objectid;
2033         while (1) {
2034                 cond_resched();
2035                 iput(inode);
2036
2037                 if (objectid > max_key->objectid)
2038                         break;
2039
2040                 inode = find_next_inode(root, objectid);
2041                 if (!inode)
2042                         break;
2043                 ino = btrfs_ino(inode);
2044
2045                 if (ino > max_key->objectid) {
2046                         iput(inode);
2047                         break;
2048                 }
2049
2050                 objectid = ino + 1;
2051                 if (!S_ISREG(inode->i_mode))
2052                         continue;
2053
2054                 if (unlikely(min_key->objectid == ino)) {
2055                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2056                                 continue;
2057                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2058                                 start = 0;
2059                         else {
2060                                 start = min_key->offset;
2061                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2062                         }
2063                 } else {
2064                         start = 0;
2065                 }
2066
2067                 if (unlikely(max_key->objectid == ino)) {
2068                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2069                                 continue;
2070                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2071                                 end = (u64)-1;
2072                         } else {
2073                                 if (max_key->offset == 0)
2074                                         continue;
2075                                 end = max_key->offset;
2076                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2077                                 end--;
2078                         }
2079                 } else {
2080                         end = (u64)-1;
2081                 }
2082
2083                 /* the lock_extent waits for readpage to complete */
2084                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2085                 btrfs_drop_extent_cache(inode, start, end, 1);
2086                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2087         }
2088         return 0;
2089 }
2090
2091 static int find_next_key(struct btrfs_path *path, int level,
2092                          struct btrfs_key *key)
2093
2094 {
2095         while (level < BTRFS_MAX_LEVEL) {
2096                 if (!path->nodes[level])
2097                         break;
2098                 if (path->slots[level] + 1 <
2099                     btrfs_header_nritems(path->nodes[level])) {
2100                         btrfs_node_key_to_cpu(path->nodes[level], key,
2101                                               path->slots[level] + 1);
2102                         return 0;
2103                 }
2104                 level++;
2105         }
2106         return 1;
2107 }
2108
2109 /*
2110  * merge the relocated tree blocks in reloc tree with corresponding
2111  * fs tree.
2112  */
2113 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2114                                                struct btrfs_root *root)
2115 {
2116         LIST_HEAD(inode_list);
2117         struct btrfs_key key;
2118         struct btrfs_key next_key;
2119         struct btrfs_trans_handle *trans = NULL;
2120         struct btrfs_root *reloc_root;
2121         struct btrfs_root_item *root_item;
2122         struct btrfs_path *path;
2123         struct extent_buffer *leaf;
2124         int level;
2125         int max_level;
2126         int replaced = 0;
2127         int ret;
2128         int err = 0;
2129         u32 min_reserved;
2130
2131         path = btrfs_alloc_path();
2132         if (!path)
2133                 return -ENOMEM;
2134         path->reada = READA_FORWARD;
2135
2136         reloc_root = root->reloc_root;
2137         root_item = &reloc_root->root_item;
2138
2139         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2140                 level = btrfs_root_level(root_item);
2141                 extent_buffer_get(reloc_root->node);
2142                 path->nodes[level] = reloc_root->node;
2143                 path->slots[level] = 0;
2144         } else {
2145                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2146
2147                 level = root_item->drop_level;
2148                 BUG_ON(level == 0);
2149                 path->lowest_level = level;
2150                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2151                 path->lowest_level = 0;
2152                 if (ret < 0) {
2153                         btrfs_free_path(path);
2154                         return ret;
2155                 }
2156
2157                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2158                                       path->slots[level]);
2159                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2160
2161                 btrfs_unlock_up_safe(path, 0);
2162         }
2163
2164         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2165         memset(&next_key, 0, sizeof(next_key));
2166
2167         while (1) {
2168                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2169                                              BTRFS_RESERVE_FLUSH_ALL);
2170                 if (ret) {
2171                         err = ret;
2172                         goto out;
2173                 }
2174                 trans = btrfs_start_transaction(root, 0);
2175                 if (IS_ERR(trans)) {
2176                         err = PTR_ERR(trans);
2177                         trans = NULL;
2178                         goto out;
2179                 }
2180                 trans->block_rsv = rc->block_rsv;
2181
2182                 replaced = 0;
2183                 max_level = level;
2184
2185                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2186                 if (ret < 0) {
2187                         err = ret;
2188                         goto out;
2189                 }
2190                 if (ret > 0)
2191                         break;
2192
2193                 if (!find_next_key(path, level, &key) &&
2194                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2195                         ret = 0;
2196                 } else {
2197                         ret = replace_path(trans, root, reloc_root, path,
2198                                            &next_key, level, max_level);
2199                 }
2200                 if (ret < 0) {
2201                         err = ret;
2202                         goto out;
2203                 }
2204
2205                 if (ret > 0) {
2206                         level = ret;
2207                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2208                                               path->slots[level]);
2209                         replaced = 1;
2210                 }
2211
2212                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2213                 if (ret > 0)
2214                         break;
2215
2216                 BUG_ON(level == 0);
2217                 /*
2218                  * save the merging progress in the drop_progress.
2219                  * this is OK since root refs == 1 in this case.
2220                  */
2221                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2222                                path->slots[level]);
2223                 root_item->drop_level = level;
2224
2225                 btrfs_end_transaction_throttle(trans, root);
2226                 trans = NULL;
2227
2228                 btrfs_btree_balance_dirty(root);
2229
2230                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2231                         invalidate_extent_cache(root, &key, &next_key);
2232         }
2233
2234         /*
2235          * handle the case only one block in the fs tree need to be
2236          * relocated and the block is tree root.
2237          */
2238         leaf = btrfs_lock_root_node(root);
2239         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2240         btrfs_tree_unlock(leaf);
2241         free_extent_buffer(leaf);
2242         if (ret < 0)
2243                 err = ret;
2244 out:
2245         btrfs_free_path(path);
2246
2247         if (err == 0) {
2248                 memset(&root_item->drop_progress, 0,
2249                        sizeof(root_item->drop_progress));
2250                 root_item->drop_level = 0;
2251                 btrfs_set_root_refs(root_item, 0);
2252                 btrfs_update_reloc_root(trans, root);
2253         }
2254
2255         if (trans)
2256                 btrfs_end_transaction_throttle(trans, root);
2257
2258         btrfs_btree_balance_dirty(root);
2259
2260         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2261                 invalidate_extent_cache(root, &key, &next_key);
2262
2263         return err;
2264 }
2265
2266 static noinline_for_stack
2267 int prepare_to_merge(struct reloc_control *rc, int err)
2268 {
2269         struct btrfs_root *root = rc->extent_root;
2270         struct btrfs_root *reloc_root;
2271         struct btrfs_trans_handle *trans;
2272         LIST_HEAD(reloc_roots);
2273         u64 num_bytes = 0;
2274         int ret;
2275
2276         mutex_lock(&root->fs_info->reloc_mutex);
2277         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2278         rc->merging_rsv_size += rc->nodes_relocated * 2;
2279         mutex_unlock(&root->fs_info->reloc_mutex);
2280
2281 again:
2282         if (!err) {
2283                 num_bytes = rc->merging_rsv_size;
2284                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2285                                           BTRFS_RESERVE_FLUSH_ALL);
2286                 if (ret)
2287                         err = ret;
2288         }
2289
2290         trans = btrfs_join_transaction(rc->extent_root);
2291         if (IS_ERR(trans)) {
2292                 if (!err)
2293                         btrfs_block_rsv_release(rc->extent_root,
2294                                                 rc->block_rsv, num_bytes);
2295                 return PTR_ERR(trans);
2296         }
2297
2298         if (!err) {
2299                 if (num_bytes != rc->merging_rsv_size) {
2300                         btrfs_end_transaction(trans, rc->extent_root);
2301                         btrfs_block_rsv_release(rc->extent_root,
2302                                                 rc->block_rsv, num_bytes);
2303                         goto again;
2304                 }
2305         }
2306
2307         rc->merge_reloc_tree = 1;
2308
2309         while (!list_empty(&rc->reloc_roots)) {
2310                 reloc_root = list_entry(rc->reloc_roots.next,
2311                                         struct btrfs_root, root_list);
2312                 list_del_init(&reloc_root->root_list);
2313
2314                 root = read_fs_root(reloc_root->fs_info,
2315                                     reloc_root->root_key.offset);
2316                 BUG_ON(IS_ERR(root));
2317                 BUG_ON(root->reloc_root != reloc_root);
2318
2319                 /*
2320                  * set reference count to 1, so btrfs_recover_relocation
2321                  * knows it should resumes merging
2322                  */
2323                 if (!err)
2324                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2325                 btrfs_update_reloc_root(trans, root);
2326
2327                 list_add(&reloc_root->root_list, &reloc_roots);
2328         }
2329
2330         list_splice(&reloc_roots, &rc->reloc_roots);
2331
2332         if (!err)
2333                 btrfs_commit_transaction(trans, rc->extent_root);
2334         else
2335                 btrfs_end_transaction(trans, rc->extent_root);
2336         return err;
2337 }
2338
2339 static noinline_for_stack
2340 void free_reloc_roots(struct list_head *list)
2341 {
2342         struct btrfs_root *reloc_root;
2343
2344         while (!list_empty(list)) {
2345                 reloc_root = list_entry(list->next, struct btrfs_root,
2346                                         root_list);
2347                 __del_reloc_root(reloc_root);
2348         }
2349 }
2350
2351 static noinline_for_stack
2352 void merge_reloc_roots(struct reloc_control *rc)
2353 {
2354         struct btrfs_root *root;
2355         struct btrfs_root *reloc_root;
2356         u64 last_snap;
2357         u64 otransid;
2358         u64 objectid;
2359         LIST_HEAD(reloc_roots);
2360         int found = 0;
2361         int ret = 0;
2362 again:
2363         root = rc->extent_root;
2364
2365         /*
2366          * this serializes us with btrfs_record_root_in_transaction,
2367          * we have to make sure nobody is in the middle of
2368          * adding their roots to the list while we are
2369          * doing this splice
2370          */
2371         mutex_lock(&root->fs_info->reloc_mutex);
2372         list_splice_init(&rc->reloc_roots, &reloc_roots);
2373         mutex_unlock(&root->fs_info->reloc_mutex);
2374
2375         while (!list_empty(&reloc_roots)) {
2376                 found = 1;
2377                 reloc_root = list_entry(reloc_roots.next,
2378                                         struct btrfs_root, root_list);
2379
2380                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2381                         root = read_fs_root(reloc_root->fs_info,
2382                                             reloc_root->root_key.offset);
2383                         BUG_ON(IS_ERR(root));
2384                         BUG_ON(root->reloc_root != reloc_root);
2385
2386                         ret = merge_reloc_root(rc, root);
2387                         if (ret) {
2388                                 if (list_empty(&reloc_root->root_list))
2389                                         list_add_tail(&reloc_root->root_list,
2390                                                       &reloc_roots);
2391                                 goto out;
2392                         }
2393                 } else {
2394                         list_del_init(&reloc_root->root_list);
2395                 }
2396
2397                 /*
2398                  * we keep the old last snapshod transid in rtranid when we
2399                  * created the relocation tree.
2400                  */
2401                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2402                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2403                 objectid = reloc_root->root_key.offset;
2404
2405                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2406                 if (ret < 0) {
2407                         if (list_empty(&reloc_root->root_list))
2408                                 list_add_tail(&reloc_root->root_list,
2409                                               &reloc_roots);
2410                         goto out;
2411                 }
2412         }
2413
2414         if (found) {
2415                 found = 0;
2416                 goto again;
2417         }
2418 out:
2419         if (ret) {
2420                 btrfs_std_error(root->fs_info, ret, NULL);
2421                 if (!list_empty(&reloc_roots))
2422                         free_reloc_roots(&reloc_roots);
2423
2424                 /* new reloc root may be added */
2425                 mutex_lock(&root->fs_info->reloc_mutex);
2426                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2427                 mutex_unlock(&root->fs_info->reloc_mutex);
2428                 if (!list_empty(&reloc_roots))
2429                         free_reloc_roots(&reloc_roots);
2430         }
2431
2432         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2433 }
2434
2435 static void free_block_list(struct rb_root *blocks)
2436 {
2437         struct tree_block *block;
2438         struct rb_node *rb_node;
2439         while ((rb_node = rb_first(blocks))) {
2440                 block = rb_entry(rb_node, struct tree_block, rb_node);
2441                 rb_erase(rb_node, blocks);
2442                 kfree(block);
2443         }
2444 }
2445
2446 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2447                                       struct btrfs_root *reloc_root)
2448 {
2449         struct btrfs_root *root;
2450
2451         if (reloc_root->last_trans == trans->transid)
2452                 return 0;
2453
2454         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2455         BUG_ON(IS_ERR(root));
2456         BUG_ON(root->reloc_root != reloc_root);
2457
2458         return btrfs_record_root_in_trans(trans, root);
2459 }
2460
2461 static noinline_for_stack
2462 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2463                                      struct reloc_control *rc,
2464                                      struct backref_node *node,
2465                                      struct backref_edge *edges[])
2466 {
2467         struct backref_node *next;
2468         struct btrfs_root *root;
2469         int index = 0;
2470
2471         next = node;
2472         while (1) {
2473                 cond_resched();
2474                 next = walk_up_backref(next, edges, &index);
2475                 root = next->root;
2476                 BUG_ON(!root);
2477                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2478
2479                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2480                         record_reloc_root_in_trans(trans, root);
2481                         break;
2482                 }
2483
2484                 btrfs_record_root_in_trans(trans, root);
2485                 root = root->reloc_root;
2486
2487                 if (next->new_bytenr != root->node->start) {
2488                         BUG_ON(next->new_bytenr);
2489                         BUG_ON(!list_empty(&next->list));
2490                         next->new_bytenr = root->node->start;
2491                         next->root = root;
2492                         list_add_tail(&next->list,
2493                                       &rc->backref_cache.changed);
2494                         __mark_block_processed(rc, next);
2495                         break;
2496                 }
2497
2498                 WARN_ON(1);
2499                 root = NULL;
2500                 next = walk_down_backref(edges, &index);
2501                 if (!next || next->level <= node->level)
2502                         break;
2503         }
2504         if (!root)
2505                 return NULL;
2506
2507         next = node;
2508         /* setup backref node path for btrfs_reloc_cow_block */
2509         while (1) {
2510                 rc->backref_cache.path[next->level] = next;
2511                 if (--index < 0)
2512                         break;
2513                 next = edges[index]->node[UPPER];
2514         }
2515         return root;
2516 }
2517
2518 /*
2519  * select a tree root for relocation. return NULL if the block
2520  * is reference counted. we should use do_relocation() in this
2521  * case. return a tree root pointer if the block isn't reference
2522  * counted. return -ENOENT if the block is root of reloc tree.
2523  */
2524 static noinline_for_stack
2525 struct btrfs_root *select_one_root(struct backref_node *node)
2526 {
2527         struct backref_node *next;
2528         struct btrfs_root *root;
2529         struct btrfs_root *fs_root = NULL;
2530         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2531         int index = 0;
2532
2533         next = node;
2534         while (1) {
2535                 cond_resched();
2536                 next = walk_up_backref(next, edges, &index);
2537                 root = next->root;
2538                 BUG_ON(!root);
2539
2540                 /* no other choice for non-references counted tree */
2541                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2542                         return root;
2543
2544                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2545                         fs_root = root;
2546
2547                 if (next != node)
2548                         return NULL;
2549
2550                 next = walk_down_backref(edges, &index);
2551                 if (!next || next->level <= node->level)
2552                         break;
2553         }
2554
2555         if (!fs_root)
2556                 return ERR_PTR(-ENOENT);
2557         return fs_root;
2558 }
2559
2560 static noinline_for_stack
2561 u64 calcu_metadata_size(struct reloc_control *rc,
2562                         struct backref_node *node, int reserve)
2563 {
2564         struct backref_node *next = node;
2565         struct backref_edge *edge;
2566         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2567         u64 num_bytes = 0;
2568         int index = 0;
2569
2570         BUG_ON(reserve && node->processed);
2571
2572         while (next) {
2573                 cond_resched();
2574                 while (1) {
2575                         if (next->processed && (reserve || next != node))
2576                                 break;
2577
2578                         num_bytes += rc->extent_root->nodesize;
2579
2580                         if (list_empty(&next->upper))
2581                                 break;
2582
2583                         edge = list_entry(next->upper.next,
2584                                           struct backref_edge, list[LOWER]);
2585                         edges[index++] = edge;
2586                         next = edge->node[UPPER];
2587                 }
2588                 next = walk_down_backref(edges, &index);
2589         }
2590         return num_bytes;
2591 }
2592
2593 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2594                                   struct reloc_control *rc,
2595                                   struct backref_node *node)
2596 {
2597         struct btrfs_root *root = rc->extent_root;
2598         u64 num_bytes;
2599         int ret;
2600         u64 tmp;
2601
2602         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2603
2604         trans->block_rsv = rc->block_rsv;
2605         rc->reserved_bytes += num_bytes;
2606         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2607                                 BTRFS_RESERVE_FLUSH_ALL);
2608         if (ret) {
2609                 if (ret == -EAGAIN) {
2610                         tmp = rc->extent_root->nodesize *
2611                                 RELOCATION_RESERVED_NODES;
2612                         while (tmp <= rc->reserved_bytes)
2613                                 tmp <<= 1;
2614                         /*
2615                          * only one thread can access block_rsv at this point,
2616                          * so we don't need hold lock to protect block_rsv.
2617                          * we expand more reservation size here to allow enough
2618                          * space for relocation and we will return eailer in
2619                          * enospc case.
2620                          */
2621                         rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2622                                               RELOCATION_RESERVED_NODES;
2623                 }
2624                 return ret;
2625         }
2626
2627         return 0;
2628 }
2629
2630 /*
2631  * relocate a block tree, and then update pointers in upper level
2632  * blocks that reference the block to point to the new location.
2633  *
2634  * if called by link_to_upper, the block has already been relocated.
2635  * in that case this function just updates pointers.
2636  */
2637 static int do_relocation(struct btrfs_trans_handle *trans,
2638                          struct reloc_control *rc,
2639                          struct backref_node *node,
2640                          struct btrfs_key *key,
2641                          struct btrfs_path *path, int lowest)
2642 {
2643         struct backref_node *upper;
2644         struct backref_edge *edge;
2645         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2646         struct btrfs_root *root;
2647         struct extent_buffer *eb;
2648         u32 blocksize;
2649         u64 bytenr;
2650         u64 generation;
2651         int slot;
2652         int ret;
2653         int err = 0;
2654
2655         BUG_ON(lowest && node->eb);
2656
2657         path->lowest_level = node->level + 1;
2658         rc->backref_cache.path[node->level] = node;
2659         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2660                 cond_resched();
2661
2662                 upper = edge->node[UPPER];
2663                 root = select_reloc_root(trans, rc, upper, edges);
2664                 BUG_ON(!root);
2665
2666                 if (upper->eb && !upper->locked) {
2667                         if (!lowest) {
2668                                 ret = btrfs_bin_search(upper->eb, key,
2669                                                        upper->level, &slot);
2670                                 BUG_ON(ret);
2671                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2672                                 if (node->eb->start == bytenr)
2673                                         goto next;
2674                         }
2675                         drop_node_buffer(upper);
2676                 }
2677
2678                 if (!upper->eb) {
2679                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2680                         if (ret < 0) {
2681                                 err = ret;
2682                                 break;
2683                         }
2684                         BUG_ON(ret > 0);
2685
2686                         if (!upper->eb) {
2687                                 upper->eb = path->nodes[upper->level];
2688                                 path->nodes[upper->level] = NULL;
2689                         } else {
2690                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2691                         }
2692
2693                         upper->locked = 1;
2694                         path->locks[upper->level] = 0;
2695
2696                         slot = path->slots[upper->level];
2697                         btrfs_release_path(path);
2698                 } else {
2699                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2700                                                &slot);
2701                         BUG_ON(ret);
2702                 }
2703
2704                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2705                 if (lowest) {
2706                         BUG_ON(bytenr != node->bytenr);
2707                 } else {
2708                         if (node->eb->start == bytenr)
2709                                 goto next;
2710                 }
2711
2712                 blocksize = root->nodesize;
2713                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2714                 eb = read_tree_block(root, bytenr, generation);
2715                 if (IS_ERR(eb)) {
2716                         err = PTR_ERR(eb);
2717                         goto next;
2718                 } else if (!extent_buffer_uptodate(eb)) {
2719                         free_extent_buffer(eb);
2720                         err = -EIO;
2721                         goto next;
2722                 }
2723                 btrfs_tree_lock(eb);
2724                 btrfs_set_lock_blocking(eb);
2725
2726                 if (!node->eb) {
2727                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2728                                               slot, &eb);
2729                         btrfs_tree_unlock(eb);
2730                         free_extent_buffer(eb);
2731                         if (ret < 0) {
2732                                 err = ret;
2733                                 goto next;
2734                         }
2735                         BUG_ON(node->eb != eb);
2736                 } else {
2737                         btrfs_set_node_blockptr(upper->eb, slot,
2738                                                 node->eb->start);
2739                         btrfs_set_node_ptr_generation(upper->eb, slot,
2740                                                       trans->transid);
2741                         btrfs_mark_buffer_dirty(upper->eb);
2742
2743                         ret = btrfs_inc_extent_ref(trans, root,
2744                                                 node->eb->start, blocksize,
2745                                                 upper->eb->start,
2746                                                 btrfs_header_owner(upper->eb),
2747                                                 node->level, 0);
2748                         BUG_ON(ret);
2749
2750                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2751                         BUG_ON(ret);
2752                 }
2753 next:
2754                 if (!upper->pending)
2755                         drop_node_buffer(upper);
2756                 else
2757                         unlock_node_buffer(upper);
2758                 if (err)
2759                         break;
2760         }
2761
2762         if (!err && node->pending) {
2763                 drop_node_buffer(node);
2764                 list_move_tail(&node->list, &rc->backref_cache.changed);
2765                 node->pending = 0;
2766         }
2767
2768         path->lowest_level = 0;
2769         BUG_ON(err == -ENOSPC);
2770         return err;
2771 }
2772
2773 static int link_to_upper(struct btrfs_trans_handle *trans,
2774                          struct reloc_control *rc,
2775                          struct backref_node *node,
2776                          struct btrfs_path *path)
2777 {
2778         struct btrfs_key key;
2779
2780         btrfs_node_key_to_cpu(node->eb, &key, 0);
2781         return do_relocation(trans, rc, node, &key, path, 0);
2782 }
2783
2784 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2785                                 struct reloc_control *rc,
2786                                 struct btrfs_path *path, int err)
2787 {
2788         LIST_HEAD(list);
2789         struct backref_cache *cache = &rc->backref_cache;
2790         struct backref_node *node;
2791         int level;
2792         int ret;
2793
2794         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2795                 while (!list_empty(&cache->pending[level])) {
2796                         node = list_entry(cache->pending[level].next,
2797                                           struct backref_node, list);
2798                         list_move_tail(&node->list, &list);
2799                         BUG_ON(!node->pending);
2800
2801                         if (!err) {
2802                                 ret = link_to_upper(trans, rc, node, path);
2803                                 if (ret < 0)
2804                                         err = ret;
2805                         }
2806                 }
2807                 list_splice_init(&list, &cache->pending[level]);
2808         }
2809         return err;
2810 }
2811
2812 static void mark_block_processed(struct reloc_control *rc,
2813                                  u64 bytenr, u32 blocksize)
2814 {
2815         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2816                         EXTENT_DIRTY, GFP_NOFS);
2817 }
2818
2819 static void __mark_block_processed(struct reloc_control *rc,
2820                                    struct backref_node *node)
2821 {
2822         u32 blocksize;
2823         if (node->level == 0 ||
2824             in_block_group(node->bytenr, rc->block_group)) {
2825                 blocksize = rc->extent_root->nodesize;
2826                 mark_block_processed(rc, node->bytenr, blocksize);
2827         }
2828         node->processed = 1;
2829 }
2830
2831 /*
2832  * mark a block and all blocks directly/indirectly reference the block
2833  * as processed.
2834  */
2835 static void update_processed_blocks(struct reloc_control *rc,
2836                                     struct backref_node *node)
2837 {
2838         struct backref_node *next = node;
2839         struct backref_edge *edge;
2840         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2841         int index = 0;
2842
2843         while (next) {
2844                 cond_resched();
2845                 while (1) {
2846                         if (next->processed)
2847                                 break;
2848
2849                         __mark_block_processed(rc, next);
2850
2851                         if (list_empty(&next->upper))
2852                                 break;
2853
2854                         edge = list_entry(next->upper.next,
2855                                           struct backref_edge, list[LOWER]);
2856                         edges[index++] = edge;
2857                         next = edge->node[UPPER];
2858                 }
2859                 next = walk_down_backref(edges, &index);
2860         }
2861 }
2862
2863 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2864 {
2865         u32 blocksize = rc->extent_root->nodesize;
2866
2867         if (test_range_bit(&rc->processed_blocks, bytenr,
2868                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2869                 return 1;
2870         return 0;
2871 }
2872
2873 static int get_tree_block_key(struct reloc_control *rc,
2874                               struct tree_block *block)
2875 {
2876         struct extent_buffer *eb;
2877
2878         BUG_ON(block->key_ready);
2879         eb = read_tree_block(rc->extent_root, block->bytenr,
2880                              block->key.offset);
2881         if (IS_ERR(eb)) {
2882                 return PTR_ERR(eb);
2883         } else if (!extent_buffer_uptodate(eb)) {
2884                 free_extent_buffer(eb);
2885                 return -EIO;
2886         }
2887         WARN_ON(btrfs_header_level(eb) != block->level);
2888         if (block->level == 0)
2889                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2890         else
2891                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2892         free_extent_buffer(eb);
2893         block->key_ready = 1;
2894         return 0;
2895 }
2896
2897 /*
2898  * helper function to relocate a tree block
2899  */
2900 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2901                                 struct reloc_control *rc,
2902                                 struct backref_node *node,
2903                                 struct btrfs_key *key,
2904                                 struct btrfs_path *path)
2905 {
2906         struct btrfs_root *root;
2907         int ret = 0;
2908
2909         if (!node)
2910                 return 0;
2911
2912         BUG_ON(node->processed);
2913         root = select_one_root(node);
2914         if (root == ERR_PTR(-ENOENT)) {
2915                 update_processed_blocks(rc, node);
2916                 goto out;
2917         }
2918
2919         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2920                 ret = reserve_metadata_space(trans, rc, node);
2921                 if (ret)
2922                         goto out;
2923         }
2924
2925         if (root) {
2926                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2927                         BUG_ON(node->new_bytenr);
2928                         BUG_ON(!list_empty(&node->list));
2929                         btrfs_record_root_in_trans(trans, root);
2930                         root = root->reloc_root;
2931                         node->new_bytenr = root->node->start;
2932                         node->root = root;
2933                         list_add_tail(&node->list, &rc->backref_cache.changed);
2934                 } else {
2935                         path->lowest_level = node->level;
2936                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2937                         btrfs_release_path(path);
2938                         if (ret > 0)
2939                                 ret = 0;
2940                 }
2941                 if (!ret)
2942                         update_processed_blocks(rc, node);
2943         } else {
2944                 ret = do_relocation(trans, rc, node, key, path, 1);
2945         }
2946 out:
2947         if (ret || node->level == 0 || node->cowonly)
2948                 remove_backref_node(&rc->backref_cache, node);
2949         return ret;
2950 }
2951
2952 /*
2953  * relocate a list of blocks
2954  */
2955 static noinline_for_stack
2956 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2957                          struct reloc_control *rc, struct rb_root *blocks)
2958 {
2959         struct backref_node *node;
2960         struct btrfs_path *path;
2961         struct tree_block *block;
2962         struct rb_node *rb_node;
2963         int ret;
2964         int err = 0;
2965
2966         path = btrfs_alloc_path();
2967         if (!path) {
2968                 err = -ENOMEM;
2969                 goto out_free_blocks;
2970         }
2971
2972         rb_node = rb_first(blocks);
2973         while (rb_node) {
2974                 block = rb_entry(rb_node, struct tree_block, rb_node);
2975                 if (!block->key_ready)
2976                         readahead_tree_block(rc->extent_root, block->bytenr);
2977                 rb_node = rb_next(rb_node);
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                         err = get_tree_block_key(rc, block);
2985                         if (err)
2986                                 goto out_free_path;
2987                 }
2988                 rb_node = rb_next(rb_node);
2989         }
2990
2991         rb_node = rb_first(blocks);
2992         while (rb_node) {
2993                 block = rb_entry(rb_node, struct tree_block, rb_node);
2994
2995                 node = build_backref_tree(rc, &block->key,
2996                                           block->level, block->bytenr);
2997                 if (IS_ERR(node)) {
2998                         err = PTR_ERR(node);
2999                         goto out;
3000                 }
3001
3002                 ret = relocate_tree_block(trans, rc, node, &block->key,
3003                                           path);
3004                 if (ret < 0) {
3005                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3006                                 err = ret;
3007                         goto out;
3008                 }
3009                 rb_node = rb_next(rb_node);
3010         }
3011 out:
3012         err = finish_pending_nodes(trans, rc, path, err);
3013
3014 out_free_path:
3015         btrfs_free_path(path);
3016 out_free_blocks:
3017         free_block_list(blocks);
3018         return err;
3019 }
3020
3021 static noinline_for_stack
3022 int prealloc_file_extent_cluster(struct inode *inode,
3023                                  struct file_extent_cluster *cluster)
3024 {
3025         u64 alloc_hint = 0;
3026         u64 start;
3027         u64 end;
3028         u64 offset = BTRFS_I(inode)->index_cnt;
3029         u64 num_bytes;
3030         int nr = 0;
3031         int ret = 0;
3032
3033         BUG_ON(cluster->start != cluster->boundary[0]);
3034         inode_lock(inode);
3035
3036         ret = btrfs_check_data_free_space(inode, cluster->start,
3037                                           cluster->end + 1 - cluster->start);
3038         if (ret)
3039                 goto out;
3040
3041         while (nr < cluster->nr) {
3042                 start = cluster->boundary[nr] - offset;
3043                 if (nr + 1 < cluster->nr)
3044                         end = cluster->boundary[nr + 1] - 1 - offset;
3045                 else
3046                         end = cluster->end - offset;
3047
3048                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3049                 num_bytes = end + 1 - start;
3050                 ret = btrfs_prealloc_file_range(inode, 0, start,
3051                                                 num_bytes, num_bytes,
3052                                                 end + 1, &alloc_hint);
3053                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3054                 if (ret)
3055                         break;
3056                 nr++;
3057         }
3058         btrfs_free_reserved_data_space(inode, cluster->start,
3059                                        cluster->end + 1 - cluster->start);
3060 out:
3061         inode_unlock(inode);
3062         return ret;
3063 }
3064
3065 static noinline_for_stack
3066 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3067                          u64 block_start)
3068 {
3069         struct btrfs_root *root = BTRFS_I(inode)->root;
3070         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3071         struct extent_map *em;
3072         int ret = 0;
3073
3074         em = alloc_extent_map();
3075         if (!em)
3076                 return -ENOMEM;
3077
3078         em->start = start;
3079         em->len = end + 1 - start;
3080         em->block_len = em->len;
3081         em->block_start = block_start;
3082         em->bdev = root->fs_info->fs_devices->latest_bdev;
3083         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3084
3085         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3086         while (1) {
3087                 write_lock(&em_tree->lock);
3088                 ret = add_extent_mapping(em_tree, em, 0);
3089                 write_unlock(&em_tree->lock);
3090                 if (ret != -EEXIST) {
3091                         free_extent_map(em);
3092                         break;
3093                 }
3094                 btrfs_drop_extent_cache(inode, start, end, 0);
3095         }
3096         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3097         return ret;
3098 }
3099
3100 static int relocate_file_extent_cluster(struct inode *inode,
3101                                         struct file_extent_cluster *cluster)
3102 {
3103         u64 page_start;
3104         u64 page_end;
3105         u64 offset = BTRFS_I(inode)->index_cnt;
3106         unsigned long index;
3107         unsigned long last_index;
3108         struct page *page;
3109         struct file_ra_state *ra;
3110         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3111         int nr = 0;
3112         int ret = 0;
3113
3114         if (!cluster->nr)
3115                 return 0;
3116
3117         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3118         if (!ra)
3119                 return -ENOMEM;
3120
3121         ret = prealloc_file_extent_cluster(inode, cluster);
3122         if (ret)
3123                 goto out;
3124
3125         file_ra_state_init(ra, inode->i_mapping);
3126
3127         ret = setup_extent_mapping(inode, cluster->start - offset,
3128                                    cluster->end - offset, cluster->start);
3129         if (ret)
3130                 goto out;
3131
3132         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3133         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3134         while (index <= last_index) {
3135                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3136                 if (ret)
3137                         goto out;
3138
3139                 page = find_lock_page(inode->i_mapping, index);
3140                 if (!page) {
3141                         page_cache_sync_readahead(inode->i_mapping,
3142                                                   ra, NULL, index,
3143                                                   last_index + 1 - index);
3144                         page = find_or_create_page(inode->i_mapping, index,
3145                                                    mask);
3146                         if (!page) {
3147                                 btrfs_delalloc_release_metadata(inode,
3148                                                         PAGE_CACHE_SIZE);
3149                                 ret = -ENOMEM;
3150                                 goto out;
3151                         }
3152                 }
3153
3154                 if (PageReadahead(page)) {
3155                         page_cache_async_readahead(inode->i_mapping,
3156                                                    ra, NULL, page, index,
3157                                                    last_index + 1 - index);
3158                 }
3159
3160                 if (!PageUptodate(page)) {
3161                         btrfs_readpage(NULL, page);
3162                         lock_page(page);
3163                         if (!PageUptodate(page)) {
3164                                 unlock_page(page);
3165                                 page_cache_release(page);
3166                                 btrfs_delalloc_release_metadata(inode,
3167                                                         PAGE_CACHE_SIZE);
3168                                 ret = -EIO;
3169                                 goto out;
3170                         }
3171                 }
3172
3173                 page_start = page_offset(page);
3174                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3175
3176                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3177
3178                 set_page_extent_mapped(page);
3179
3180                 if (nr < cluster->nr &&
3181                     page_start + offset == cluster->boundary[nr]) {
3182                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3183                                         page_start, page_end,
3184                                         EXTENT_BOUNDARY, GFP_NOFS);
3185                         nr++;
3186                 }
3187
3188                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3189                 set_page_dirty(page);
3190
3191                 unlock_extent(&BTRFS_I(inode)->io_tree,
3192                               page_start, page_end);
3193                 unlock_page(page);
3194                 page_cache_release(page);
3195
3196                 index++;
3197                 balance_dirty_pages_ratelimited(inode->i_mapping);
3198                 btrfs_throttle(BTRFS_I(inode)->root);
3199         }
3200         WARN_ON(nr != cluster->nr);
3201 out:
3202         kfree(ra);
3203         return ret;
3204 }
3205
3206 static noinline_for_stack
3207 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3208                          struct file_extent_cluster *cluster)
3209 {
3210         int ret;
3211
3212         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3213                 ret = relocate_file_extent_cluster(inode, cluster);
3214                 if (ret)
3215                         return ret;
3216                 cluster->nr = 0;
3217         }
3218
3219         if (!cluster->nr)
3220                 cluster->start = extent_key->objectid;
3221         else
3222                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3223         cluster->end = extent_key->objectid + extent_key->offset - 1;
3224         cluster->boundary[cluster->nr] = extent_key->objectid;
3225         cluster->nr++;
3226
3227         if (cluster->nr >= MAX_EXTENTS) {
3228                 ret = relocate_file_extent_cluster(inode, cluster);
3229                 if (ret)
3230                         return ret;
3231                 cluster->nr = 0;
3232         }
3233         return 0;
3234 }
3235
3236 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3237 static int get_ref_objectid_v0(struct reloc_control *rc,
3238                                struct btrfs_path *path,
3239                                struct btrfs_key *extent_key,
3240                                u64 *ref_objectid, int *path_change)
3241 {
3242         struct btrfs_key key;
3243         struct extent_buffer *leaf;
3244         struct btrfs_extent_ref_v0 *ref0;
3245         int ret;
3246         int slot;
3247
3248         leaf = path->nodes[0];
3249         slot = path->slots[0];
3250         while (1) {
3251                 if (slot >= btrfs_header_nritems(leaf)) {
3252                         ret = btrfs_next_leaf(rc->extent_root, path);
3253                         if (ret < 0)
3254                                 return ret;
3255                         BUG_ON(ret > 0);
3256                         leaf = path->nodes[0];
3257                         slot = path->slots[0];
3258                         if (path_change)
3259                                 *path_change = 1;
3260                 }
3261                 btrfs_item_key_to_cpu(leaf, &key, slot);
3262                 if (key.objectid != extent_key->objectid)
3263                         return -ENOENT;
3264
3265                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3266                         slot++;
3267                         continue;
3268                 }
3269                 ref0 = btrfs_item_ptr(leaf, slot,
3270                                 struct btrfs_extent_ref_v0);
3271                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3272                 break;
3273         }
3274         return 0;
3275 }
3276 #endif
3277
3278 /*
3279  * helper to add a tree block to the list.
3280  * the major work is getting the generation and level of the block
3281  */
3282 static int add_tree_block(struct reloc_control *rc,
3283                           struct btrfs_key *extent_key,
3284                           struct btrfs_path *path,
3285                           struct rb_root *blocks)
3286 {
3287         struct extent_buffer *eb;
3288         struct btrfs_extent_item *ei;
3289         struct btrfs_tree_block_info *bi;
3290         struct tree_block *block;
3291         struct rb_node *rb_node;
3292         u32 item_size;
3293         int level = -1;
3294         u64 generation;
3295
3296         eb =  path->nodes[0];
3297         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3298
3299         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3300             item_size >= sizeof(*ei) + sizeof(*bi)) {
3301                 ei = btrfs_item_ptr(eb, path->slots[0],
3302                                 struct btrfs_extent_item);
3303                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3304                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3305                         level = btrfs_tree_block_level(eb, bi);
3306                 } else {
3307                         level = (int)extent_key->offset;
3308                 }
3309                 generation = btrfs_extent_generation(eb, ei);
3310         } else {
3311 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3312                 u64 ref_owner;
3313                 int ret;
3314
3315                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3316                 ret = get_ref_objectid_v0(rc, path, extent_key,
3317                                           &ref_owner, NULL);
3318                 if (ret < 0)
3319                         return ret;
3320                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3321                 level = (int)ref_owner;
3322                 /* FIXME: get real generation */
3323                 generation = 0;
3324 #else
3325                 BUG();
3326 #endif
3327         }
3328
3329         btrfs_release_path(path);
3330
3331         BUG_ON(level == -1);
3332
3333         block = kmalloc(sizeof(*block), GFP_NOFS);
3334         if (!block)
3335                 return -ENOMEM;
3336
3337         block->bytenr = extent_key->objectid;
3338         block->key.objectid = rc->extent_root->nodesize;
3339         block->key.offset = generation;
3340         block->level = level;
3341         block->key_ready = 0;
3342
3343         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3344         if (rb_node)
3345                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3346
3347         return 0;
3348 }
3349
3350 /*
3351  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3352  */
3353 static int __add_tree_block(struct reloc_control *rc,
3354                             u64 bytenr, u32 blocksize,
3355                             struct rb_root *blocks)
3356 {
3357         struct btrfs_path *path;
3358         struct btrfs_key key;
3359         int ret;
3360         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3361                                         SKINNY_METADATA);
3362
3363         if (tree_block_processed(bytenr, rc))
3364                 return 0;
3365
3366         if (tree_search(blocks, bytenr))
3367                 return 0;
3368
3369         path = btrfs_alloc_path();
3370         if (!path)
3371                 return -ENOMEM;
3372 again:
3373         key.objectid = bytenr;
3374         if (skinny) {
3375                 key.type = BTRFS_METADATA_ITEM_KEY;
3376                 key.offset = (u64)-1;
3377         } else {
3378                 key.type = BTRFS_EXTENT_ITEM_KEY;
3379                 key.offset = blocksize;
3380         }
3381
3382         path->search_commit_root = 1;
3383         path->skip_locking = 1;
3384         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3385         if (ret < 0)
3386                 goto out;
3387
3388         if (ret > 0 && skinny) {
3389                 if (path->slots[0]) {
3390                         path->slots[0]--;
3391                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3392                                               path->slots[0]);
3393                         if (key.objectid == bytenr &&
3394                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3395                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3396                               key.offset == blocksize)))
3397                                 ret = 0;
3398                 }
3399
3400                 if (ret) {
3401                         skinny = false;
3402                         btrfs_release_path(path);
3403                         goto again;
3404                 }
3405         }
3406         BUG_ON(ret);
3407
3408         ret = add_tree_block(rc, &key, path, blocks);
3409 out:
3410         btrfs_free_path(path);
3411         return ret;
3412 }
3413
3414 /*
3415  * helper to check if the block use full backrefs for pointers in it
3416  */
3417 static int block_use_full_backref(struct reloc_control *rc,
3418                                   struct extent_buffer *eb)
3419 {
3420         u64 flags;
3421         int ret;
3422
3423         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3424             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3425                 return 1;
3426
3427         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3428                                        eb->start, btrfs_header_level(eb), 1,
3429                                        NULL, &flags);
3430         BUG_ON(ret);
3431
3432         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3433                 ret = 1;
3434         else
3435                 ret = 0;
3436         return ret;
3437 }
3438
3439 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3440                                     struct btrfs_block_group_cache *block_group,
3441                                     struct inode *inode,
3442                                     u64 ino)
3443 {
3444         struct btrfs_key key;
3445         struct btrfs_root *root = fs_info->tree_root;
3446         struct btrfs_trans_handle *trans;
3447         int ret = 0;
3448
3449         if (inode)
3450                 goto truncate;
3451
3452         key.objectid = ino;
3453         key.type = BTRFS_INODE_ITEM_KEY;
3454         key.offset = 0;
3455
3456         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3457         if (IS_ERR(inode) || is_bad_inode(inode)) {
3458                 if (!IS_ERR(inode))
3459                         iput(inode);
3460                 return -ENOENT;
3461         }
3462
3463 truncate:
3464         ret = btrfs_check_trunc_cache_free_space(root,
3465                                                  &fs_info->global_block_rsv);
3466         if (ret)
3467                 goto out;
3468
3469         trans = btrfs_join_transaction(root);
3470         if (IS_ERR(trans)) {
3471                 ret = PTR_ERR(trans);
3472                 goto out;
3473         }
3474
3475         ret = btrfs_truncate_free_space_cache(root, trans, block_group, inode);
3476
3477         btrfs_end_transaction(trans, root);
3478         btrfs_btree_balance_dirty(root);
3479 out:
3480         iput(inode);
3481         return ret;
3482 }
3483
3484 /*
3485  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3486  * this function scans fs tree to find blocks reference the data extent
3487  */
3488 static int find_data_references(struct reloc_control *rc,
3489                                 struct btrfs_key *extent_key,
3490                                 struct extent_buffer *leaf,
3491                                 struct btrfs_extent_data_ref *ref,
3492                                 struct rb_root *blocks)
3493 {
3494         struct btrfs_path *path;
3495         struct tree_block *block;
3496         struct btrfs_root *root;
3497         struct btrfs_file_extent_item *fi;
3498         struct rb_node *rb_node;
3499         struct btrfs_key key;
3500         u64 ref_root;
3501         u64 ref_objectid;
3502         u64 ref_offset;
3503         u32 ref_count;
3504         u32 nritems;
3505         int err = 0;
3506         int added = 0;
3507         int counted;
3508         int ret;
3509
3510         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3511         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3512         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3513         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3514
3515         /*
3516          * This is an extent belonging to the free space cache, lets just delete
3517          * it and redo the search.
3518          */
3519         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3520                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3521                                                rc->block_group,
3522                                                NULL, ref_objectid);
3523                 if (ret != -ENOENT)
3524                         return ret;
3525                 ret = 0;
3526         }
3527
3528         path = btrfs_alloc_path();
3529         if (!path)
3530                 return -ENOMEM;
3531         path->reada = READA_FORWARD;
3532
3533         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3534         if (IS_ERR(root)) {
3535                 err = PTR_ERR(root);
3536                 goto out;
3537         }
3538
3539         key.objectid = ref_objectid;
3540         key.type = BTRFS_EXTENT_DATA_KEY;
3541         if (ref_offset > ((u64)-1 << 32))
3542                 key.offset = 0;
3543         else
3544                 key.offset = ref_offset;
3545
3546         path->search_commit_root = 1;
3547         path->skip_locking = 1;
3548         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3549         if (ret < 0) {
3550                 err = ret;
3551                 goto out;
3552         }
3553
3554         leaf = path->nodes[0];
3555         nritems = btrfs_header_nritems(leaf);
3556         /*
3557          * the references in tree blocks that use full backrefs
3558          * are not counted in
3559          */
3560         if (block_use_full_backref(rc, leaf))
3561                 counted = 0;
3562         else
3563                 counted = 1;
3564         rb_node = tree_search(blocks, leaf->start);
3565         if (rb_node) {
3566                 if (counted)
3567                         added = 1;
3568                 else
3569                         path->slots[0] = nritems;
3570         }
3571
3572         while (ref_count > 0) {
3573                 while (path->slots[0] >= nritems) {
3574                         ret = btrfs_next_leaf(root, path);
3575                         if (ret < 0) {
3576                                 err = ret;
3577                                 goto out;
3578                         }
3579                         if (WARN_ON(ret > 0))
3580                                 goto out;
3581
3582                         leaf = path->nodes[0];
3583                         nritems = btrfs_header_nritems(leaf);
3584                         added = 0;
3585
3586                         if (block_use_full_backref(rc, leaf))
3587                                 counted = 0;
3588                         else
3589                                 counted = 1;
3590                         rb_node = tree_search(blocks, leaf->start);
3591                         if (rb_node) {
3592                                 if (counted)
3593                                         added = 1;
3594                                 else
3595                                         path->slots[0] = nritems;
3596                         }
3597                 }
3598
3599                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3600                 if (WARN_ON(key.objectid != ref_objectid ||
3601                     key.type != BTRFS_EXTENT_DATA_KEY))
3602                         break;
3603
3604                 fi = btrfs_item_ptr(leaf, path->slots[0],
3605                                     struct btrfs_file_extent_item);
3606
3607                 if (btrfs_file_extent_type(leaf, fi) ==
3608                     BTRFS_FILE_EXTENT_INLINE)
3609                         goto next;
3610
3611                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3612                     extent_key->objectid)
3613                         goto next;
3614
3615                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3616                 if (key.offset != ref_offset)
3617                         goto next;
3618
3619                 if (counted)
3620                         ref_count--;
3621                 if (added)
3622                         goto next;
3623
3624                 if (!tree_block_processed(leaf->start, rc)) {
3625                         block = kmalloc(sizeof(*block), GFP_NOFS);
3626                         if (!block) {
3627                                 err = -ENOMEM;
3628                                 break;
3629                         }
3630                         block->bytenr = leaf->start;
3631                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3632                         block->level = 0;
3633                         block->key_ready = 1;
3634                         rb_node = tree_insert(blocks, block->bytenr,
3635                                               &block->rb_node);
3636                         if (rb_node)
3637                                 backref_tree_panic(rb_node, -EEXIST,
3638                                                    block->bytenr);
3639                 }
3640                 if (counted)
3641                         added = 1;
3642                 else
3643                         path->slots[0] = nritems;
3644 next:
3645                 path->slots[0]++;
3646
3647         }
3648 out:
3649         btrfs_free_path(path);
3650         return err;
3651 }
3652
3653 /*
3654  * helper to find all tree blocks that reference a given data extent
3655  */
3656 static noinline_for_stack
3657 int add_data_references(struct reloc_control *rc,
3658                         struct btrfs_key *extent_key,
3659                         struct btrfs_path *path,
3660                         struct rb_root *blocks)
3661 {
3662         struct btrfs_key key;
3663         struct extent_buffer *eb;
3664         struct btrfs_extent_data_ref *dref;
3665         struct btrfs_extent_inline_ref *iref;
3666         unsigned long ptr;
3667         unsigned long end;
3668         u32 blocksize = rc->extent_root->nodesize;
3669         int ret = 0;
3670         int err = 0;
3671
3672         eb = path->nodes[0];
3673         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3674         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3675 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3676         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3677                 ptr = end;
3678         else
3679 #endif
3680                 ptr += sizeof(struct btrfs_extent_item);
3681
3682         while (ptr < end) {
3683                 iref = (struct btrfs_extent_inline_ref *)ptr;
3684                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3685                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3686                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3687                         ret = __add_tree_block(rc, key.offset, blocksize,
3688                                                blocks);
3689                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3690                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3691                         ret = find_data_references(rc, extent_key,
3692                                                    eb, dref, blocks);
3693                 } else {
3694                         BUG();
3695                 }
3696                 if (ret) {
3697                         err = ret;
3698                         goto out;
3699                 }
3700                 ptr += btrfs_extent_inline_ref_size(key.type);
3701         }
3702         WARN_ON(ptr > end);
3703
3704         while (1) {
3705                 cond_resched();
3706                 eb = path->nodes[0];
3707                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3708                         ret = btrfs_next_leaf(rc->extent_root, path);
3709                         if (ret < 0) {
3710                                 err = ret;
3711                                 break;
3712                         }
3713                         if (ret > 0)
3714                                 break;
3715                         eb = path->nodes[0];
3716                 }
3717
3718                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3719                 if (key.objectid != extent_key->objectid)
3720                         break;
3721
3722 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3723                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3724                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3725 #else
3726                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3727                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3728 #endif
3729                         ret = __add_tree_block(rc, key.offset, blocksize,
3730                                                blocks);
3731                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3732                         dref = btrfs_item_ptr(eb, path->slots[0],
3733                                               struct btrfs_extent_data_ref);
3734                         ret = find_data_references(rc, extent_key,
3735                                                    eb, dref, blocks);
3736                 } else {
3737                         ret = 0;
3738                 }
3739                 if (ret) {
3740                         err = ret;
3741                         break;
3742                 }
3743                 path->slots[0]++;
3744         }
3745 out:
3746         btrfs_release_path(path);
3747         if (err)
3748                 free_block_list(blocks);
3749         return err;
3750 }
3751
3752 /*
3753  * helper to find next unprocessed extent
3754  */
3755 static noinline_for_stack
3756 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3757                      struct btrfs_key *extent_key)
3758 {
3759         struct btrfs_key key;
3760         struct extent_buffer *leaf;
3761         u64 start, end, last;
3762         int ret;
3763
3764         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3765         while (1) {
3766                 cond_resched();
3767                 if (rc->search_start >= last) {
3768                         ret = 1;
3769                         break;
3770                 }
3771
3772                 key.objectid = rc->search_start;
3773                 key.type = BTRFS_EXTENT_ITEM_KEY;
3774                 key.offset = 0;
3775
3776                 path->search_commit_root = 1;
3777                 path->skip_locking = 1;
3778                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3779                                         0, 0);
3780                 if (ret < 0)
3781                         break;
3782 next:
3783                 leaf = path->nodes[0];
3784                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3785                         ret = btrfs_next_leaf(rc->extent_root, path);
3786                         if (ret != 0)
3787                                 break;
3788                         leaf = path->nodes[0];
3789                 }
3790
3791                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3792                 if (key.objectid >= last) {
3793                         ret = 1;
3794                         break;
3795                 }
3796
3797                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3798                     key.type != BTRFS_METADATA_ITEM_KEY) {
3799                         path->slots[0]++;
3800                         goto next;
3801                 }
3802
3803                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3804                     key.objectid + key.offset <= rc->search_start) {
3805                         path->slots[0]++;
3806                         goto next;
3807                 }
3808
3809                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3810                     key.objectid + rc->extent_root->nodesize <=
3811                     rc->search_start) {
3812                         path->slots[0]++;
3813                         goto next;
3814                 }
3815
3816                 ret = find_first_extent_bit(&rc->processed_blocks,
3817                                             key.objectid, &start, &end,
3818                                             EXTENT_DIRTY, NULL);
3819
3820                 if (ret == 0 && start <= key.objectid) {
3821                         btrfs_release_path(path);
3822                         rc->search_start = end + 1;
3823                 } else {
3824                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3825                                 rc->search_start = key.objectid + key.offset;
3826                         else
3827                                 rc->search_start = key.objectid +
3828                                         rc->extent_root->nodesize;
3829                         memcpy(extent_key, &key, sizeof(key));
3830                         return 0;
3831                 }
3832         }
3833         btrfs_release_path(path);
3834         return ret;
3835 }
3836
3837 static void set_reloc_control(struct reloc_control *rc)
3838 {
3839         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3840
3841         mutex_lock(&fs_info->reloc_mutex);
3842         fs_info->reloc_ctl = rc;
3843         mutex_unlock(&fs_info->reloc_mutex);
3844 }
3845
3846 static void unset_reloc_control(struct reloc_control *rc)
3847 {
3848         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3849
3850         mutex_lock(&fs_info->reloc_mutex);
3851         fs_info->reloc_ctl = NULL;
3852         mutex_unlock(&fs_info->reloc_mutex);
3853 }
3854
3855 static int check_extent_flags(u64 flags)
3856 {
3857         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3858             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3859                 return 1;
3860         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3861             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3862                 return 1;
3863         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3864             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3865                 return 1;
3866         return 0;
3867 }
3868
3869 static noinline_for_stack
3870 int prepare_to_relocate(struct reloc_control *rc)
3871 {
3872         struct btrfs_trans_handle *trans;
3873
3874         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3875                                               BTRFS_BLOCK_RSV_TEMP);
3876         if (!rc->block_rsv)
3877                 return -ENOMEM;
3878
3879         memset(&rc->cluster, 0, sizeof(rc->cluster));
3880         rc->search_start = rc->block_group->key.objectid;
3881         rc->extents_found = 0;
3882         rc->nodes_relocated = 0;
3883         rc->merging_rsv_size = 0;
3884         rc->reserved_bytes = 0;
3885         rc->block_rsv->size = rc->extent_root->nodesize *
3886                               RELOCATION_RESERVED_NODES;
3887
3888         rc->create_reloc_tree = 1;
3889         set_reloc_control(rc);
3890
3891         trans = btrfs_join_transaction(rc->extent_root);
3892         if (IS_ERR(trans)) {
3893                 unset_reloc_control(rc);
3894                 /*
3895                  * extent tree is not a ref_cow tree and has no reloc_root to
3896                  * cleanup.  And callers are responsible to free the above
3897                  * block rsv.
3898                  */
3899                 return PTR_ERR(trans);
3900         }
3901         btrfs_commit_transaction(trans, rc->extent_root);
3902         return 0;
3903 }
3904
3905 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3906 {
3907         struct rb_root blocks = RB_ROOT;
3908         struct btrfs_key key;
3909         struct btrfs_trans_handle *trans = NULL;
3910         struct btrfs_path *path;
3911         struct btrfs_extent_item *ei;
3912         u64 flags;
3913         u32 item_size;
3914         int ret;
3915         int err = 0;
3916         int progress = 0;
3917
3918         path = btrfs_alloc_path();
3919         if (!path)
3920                 return -ENOMEM;
3921         path->reada = READA_FORWARD;
3922
3923         ret = prepare_to_relocate(rc);
3924         if (ret) {
3925                 err = ret;
3926                 goto out_free;
3927         }
3928
3929         while (1) {
3930                 rc->reserved_bytes = 0;
3931                 ret = btrfs_block_rsv_refill(rc->extent_root,
3932                                         rc->block_rsv, rc->block_rsv->size,
3933                                         BTRFS_RESERVE_FLUSH_ALL);
3934                 if (ret) {
3935                         err = ret;
3936                         break;
3937                 }
3938                 progress++;
3939                 trans = btrfs_start_transaction(rc->extent_root, 0);
3940                 if (IS_ERR(trans)) {
3941                         err = PTR_ERR(trans);
3942                         trans = NULL;
3943                         break;
3944                 }
3945 restart:
3946                 if (update_backref_cache(trans, &rc->backref_cache)) {
3947                         btrfs_end_transaction(trans, rc->extent_root);
3948                         continue;
3949                 }
3950
3951                 ret = find_next_extent(rc, path, &key);
3952                 if (ret < 0)
3953                         err = ret;
3954                 if (ret != 0)
3955                         break;
3956
3957                 rc->extents_found++;
3958
3959                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3960                                     struct btrfs_extent_item);
3961                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3962                 if (item_size >= sizeof(*ei)) {
3963                         flags = btrfs_extent_flags(path->nodes[0], ei);
3964                         ret = check_extent_flags(flags);
3965                         BUG_ON(ret);
3966
3967                 } else {
3968 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3969                         u64 ref_owner;
3970                         int path_change = 0;
3971
3972                         BUG_ON(item_size !=
3973                                sizeof(struct btrfs_extent_item_v0));
3974                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3975                                                   &path_change);
3976                         if (ret < 0) {
3977                                 err = ret;
3978                                 break;
3979                         }
3980                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3981                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3982                         else
3983                                 flags = BTRFS_EXTENT_FLAG_DATA;
3984
3985                         if (path_change) {
3986                                 btrfs_release_path(path);
3987
3988                                 path->search_commit_root = 1;
3989                                 path->skip_locking = 1;
3990                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3991                                                         &key, path, 0, 0);
3992                                 if (ret < 0) {
3993                                         err = ret;
3994                                         break;
3995                                 }
3996                                 BUG_ON(ret > 0);
3997                         }
3998 #else
3999                         BUG();
4000 #endif
4001                 }
4002
4003                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4004                         ret = add_tree_block(rc, &key, path, &blocks);
4005                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4006                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4007                         ret = add_data_references(rc, &key, path, &blocks);
4008                 } else {
4009                         btrfs_release_path(path);
4010                         ret = 0;
4011                 }
4012                 if (ret < 0) {
4013                         err = ret;
4014                         break;
4015                 }
4016
4017                 if (!RB_EMPTY_ROOT(&blocks)) {
4018                         ret = relocate_tree_blocks(trans, rc, &blocks);
4019                         if (ret < 0) {
4020                                 /*
4021                                  * if we fail to relocate tree blocks, force to update
4022                                  * backref cache when committing transaction.
4023                                  */
4024                                 rc->backref_cache.last_trans = trans->transid - 1;
4025
4026                                 if (ret != -EAGAIN) {
4027                                         err = ret;
4028                                         break;
4029                                 }
4030                                 rc->extents_found--;
4031                                 rc->search_start = key.objectid;
4032                         }
4033                 }
4034
4035                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4036                 btrfs_btree_balance_dirty(rc->extent_root);
4037                 trans = NULL;
4038
4039                 if (rc->stage == MOVE_DATA_EXTENTS &&
4040                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4041                         rc->found_file_extent = 1;
4042                         ret = relocate_data_extent(rc->data_inode,
4043                                                    &key, &rc->cluster);
4044                         if (ret < 0) {
4045                                 err = ret;
4046                                 break;
4047                         }
4048                 }
4049         }
4050         if (trans && progress && err == -ENOSPC) {
4051                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4052                                               rc->block_group->flags);
4053                 if (ret == 1) {
4054                         err = 0;
4055                         progress = 0;
4056                         goto restart;
4057                 }
4058         }
4059
4060         btrfs_release_path(path);
4061         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4062                           GFP_NOFS);
4063
4064         if (trans) {
4065                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4066                 btrfs_btree_balance_dirty(rc->extent_root);
4067         }
4068
4069         if (!err) {
4070                 ret = relocate_file_extent_cluster(rc->data_inode,
4071                                                    &rc->cluster);
4072                 if (ret < 0)
4073                         err = ret;
4074         }
4075
4076         rc->create_reloc_tree = 0;
4077         set_reloc_control(rc);
4078
4079         backref_cache_cleanup(&rc->backref_cache);
4080         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4081
4082         err = prepare_to_merge(rc, err);
4083
4084         merge_reloc_roots(rc);
4085
4086         rc->merge_reloc_tree = 0;
4087         unset_reloc_control(rc);
4088         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4089
4090         /* get rid of pinned extents */
4091         trans = btrfs_join_transaction(rc->extent_root);
4092         if (IS_ERR(trans))
4093                 err = PTR_ERR(trans);
4094         else
4095                 btrfs_commit_transaction(trans, rc->extent_root);
4096 out_free:
4097         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4098         btrfs_free_path(path);
4099         return err;
4100 }
4101
4102 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4103                                  struct btrfs_root *root, u64 objectid)
4104 {
4105         struct btrfs_path *path;
4106         struct btrfs_inode_item *item;
4107         struct extent_buffer *leaf;
4108         int ret;
4109
4110         path = btrfs_alloc_path();
4111         if (!path)
4112                 return -ENOMEM;
4113
4114         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4115         if (ret)
4116                 goto out;
4117
4118         leaf = path->nodes[0];
4119         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4120         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4121         btrfs_set_inode_generation(leaf, item, 1);
4122         btrfs_set_inode_size(leaf, item, 0);
4123         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4124         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4125                                           BTRFS_INODE_PREALLOC);
4126         btrfs_mark_buffer_dirty(leaf);
4127 out:
4128         btrfs_free_path(path);
4129         return ret;
4130 }
4131
4132 /*
4133  * helper to create inode for data relocation.
4134  * the inode is in data relocation tree and its link count is 0
4135  */
4136 static noinline_for_stack
4137 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4138                                  struct btrfs_block_group_cache *group)
4139 {
4140         struct inode *inode = NULL;
4141         struct btrfs_trans_handle *trans;
4142         struct btrfs_root *root;
4143         struct btrfs_key key;
4144         u64 objectid;
4145         int err = 0;
4146
4147         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4148         if (IS_ERR(root))
4149                 return ERR_CAST(root);
4150
4151         trans = btrfs_start_transaction(root, 6);
4152         if (IS_ERR(trans))
4153                 return ERR_CAST(trans);
4154
4155         err = btrfs_find_free_objectid(root, &objectid);
4156         if (err)
4157                 goto out;
4158
4159         err = __insert_orphan_inode(trans, root, objectid);
4160         BUG_ON(err);
4161
4162         key.objectid = objectid;
4163         key.type = BTRFS_INODE_ITEM_KEY;
4164         key.offset = 0;
4165         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4166         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4167         BTRFS_I(inode)->index_cnt = group->key.objectid;
4168
4169         err = btrfs_orphan_add(trans, inode);
4170 out:
4171         btrfs_end_transaction(trans, root);
4172         btrfs_btree_balance_dirty(root);
4173         if (err) {
4174                 if (inode)
4175                         iput(inode);
4176                 inode = ERR_PTR(err);
4177         }
4178         return inode;
4179 }
4180
4181 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4182 {
4183         struct reloc_control *rc;
4184
4185         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4186         if (!rc)
4187                 return NULL;
4188
4189         INIT_LIST_HEAD(&rc->reloc_roots);
4190         backref_cache_init(&rc->backref_cache);
4191         mapping_tree_init(&rc->reloc_root_tree);
4192         extent_io_tree_init(&rc->processed_blocks,
4193                             fs_info->btree_inode->i_mapping);
4194         return rc;
4195 }
4196
4197 /*
4198  * function to relocate all extents in a block group.
4199  */
4200 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4201 {
4202         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4203         struct reloc_control *rc;
4204         struct inode *inode;
4205         struct btrfs_path *path;
4206         int ret;
4207         int rw = 0;
4208         int err = 0;
4209
4210         rc = alloc_reloc_control(fs_info);
4211         if (!rc)
4212                 return -ENOMEM;
4213
4214         rc->extent_root = extent_root;
4215
4216         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4217         BUG_ON(!rc->block_group);
4218
4219         ret = btrfs_inc_block_group_ro(extent_root, rc->block_group);
4220         if (ret) {
4221                 err = ret;
4222                 goto out;
4223         }
4224         rw = 1;
4225
4226         path = btrfs_alloc_path();
4227         if (!path) {
4228                 err = -ENOMEM;
4229                 goto out;
4230         }
4231
4232         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4233                                         path);
4234         btrfs_free_path(path);
4235
4236         if (!IS_ERR(inode))
4237                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4238         else
4239                 ret = PTR_ERR(inode);
4240
4241         if (ret && ret != -ENOENT) {
4242                 err = ret;
4243                 goto out;
4244         }
4245
4246         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4247         if (IS_ERR(rc->data_inode)) {
4248                 err = PTR_ERR(rc->data_inode);
4249                 rc->data_inode = NULL;
4250                 goto out;
4251         }
4252
4253         btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
4254                rc->block_group->key.objectid, rc->block_group->flags);
4255
4256         ret = btrfs_start_delalloc_roots(fs_info, 0, -1);
4257         if (ret < 0) {
4258                 err = ret;
4259                 goto out;
4260         }
4261         btrfs_wait_ordered_roots(fs_info, -1);
4262
4263         while (1) {
4264                 mutex_lock(&fs_info->cleaner_mutex);
4265                 ret = relocate_block_group(rc);
4266                 mutex_unlock(&fs_info->cleaner_mutex);
4267                 if (ret < 0) {
4268                         err = ret;
4269                         goto out;
4270                 }
4271
4272                 if (rc->extents_found == 0)
4273                         break;
4274
4275                 btrfs_info(extent_root->fs_info, "found %llu extents",
4276                         rc->extents_found);
4277
4278                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4279                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4280                                                        (u64)-1);
4281                         if (ret) {
4282                                 err = ret;
4283                                 goto out;
4284                         }
4285                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4286                                                  0, -1);
4287                         rc->stage = UPDATE_DATA_PTRS;
4288                 }
4289         }
4290
4291         WARN_ON(rc->block_group->pinned > 0);
4292         WARN_ON(rc->block_group->reserved > 0);
4293         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4294 out:
4295         if (err && rw)
4296                 btrfs_dec_block_group_ro(extent_root, rc->block_group);
4297         iput(rc->data_inode);
4298         btrfs_put_block_group(rc->block_group);
4299         kfree(rc);
4300         return err;
4301 }
4302
4303 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4304 {
4305         struct btrfs_trans_handle *trans;
4306         int ret, err;
4307
4308         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4309         if (IS_ERR(trans))
4310                 return PTR_ERR(trans);
4311
4312         memset(&root->root_item.drop_progress, 0,
4313                 sizeof(root->root_item.drop_progress));
4314         root->root_item.drop_level = 0;
4315         btrfs_set_root_refs(&root->root_item, 0);
4316         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4317                                 &root->root_key, &root->root_item);
4318
4319         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4320         if (err)
4321                 return err;
4322         return ret;
4323 }
4324
4325 /*
4326  * recover relocation interrupted by system crash.
4327  *
4328  * this function resumes merging reloc trees with corresponding fs trees.
4329  * this is important for keeping the sharing of tree blocks
4330  */
4331 int btrfs_recover_relocation(struct btrfs_root *root)
4332 {
4333         LIST_HEAD(reloc_roots);
4334         struct btrfs_key key;
4335         struct btrfs_root *fs_root;
4336         struct btrfs_root *reloc_root;
4337         struct btrfs_path *path;
4338         struct extent_buffer *leaf;
4339         struct reloc_control *rc = NULL;
4340         struct btrfs_trans_handle *trans;
4341         int ret;
4342         int err = 0;
4343
4344         path = btrfs_alloc_path();
4345         if (!path)
4346                 return -ENOMEM;
4347         path->reada = READA_BACK;
4348
4349         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4350         key.type = BTRFS_ROOT_ITEM_KEY;
4351         key.offset = (u64)-1;
4352
4353         while (1) {
4354                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4355                                         path, 0, 0);
4356                 if (ret < 0) {
4357                         err = ret;
4358                         goto out;
4359                 }
4360                 if (ret > 0) {
4361                         if (path->slots[0] == 0)
4362                                 break;
4363                         path->slots[0]--;
4364                 }
4365                 leaf = path->nodes[0];
4366                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4367                 btrfs_release_path(path);
4368
4369                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4370                     key.type != BTRFS_ROOT_ITEM_KEY)
4371                         break;
4372
4373                 reloc_root = btrfs_read_fs_root(root, &key);
4374                 if (IS_ERR(reloc_root)) {
4375                         err = PTR_ERR(reloc_root);
4376                         goto out;
4377                 }
4378
4379                 list_add(&reloc_root->root_list, &reloc_roots);
4380
4381                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4382                         fs_root = read_fs_root(root->fs_info,
4383                                                reloc_root->root_key.offset);
4384                         if (IS_ERR(fs_root)) {
4385                                 ret = PTR_ERR(fs_root);
4386                                 if (ret != -ENOENT) {
4387                                         err = ret;
4388                                         goto out;
4389                                 }
4390                                 ret = mark_garbage_root(reloc_root);
4391                                 if (ret < 0) {
4392                                         err = ret;
4393                                         goto out;
4394                                 }
4395                         }
4396                 }
4397
4398                 if (key.offset == 0)
4399                         break;
4400
4401                 key.offset--;
4402         }
4403         btrfs_release_path(path);
4404
4405         if (list_empty(&reloc_roots))
4406                 goto out;
4407
4408         rc = alloc_reloc_control(root->fs_info);
4409         if (!rc) {
4410                 err = -ENOMEM;
4411                 goto out;
4412         }
4413
4414         rc->extent_root = root->fs_info->extent_root;
4415
4416         set_reloc_control(rc);
4417
4418         trans = btrfs_join_transaction(rc->extent_root);
4419         if (IS_ERR(trans)) {
4420                 unset_reloc_control(rc);
4421                 err = PTR_ERR(trans);
4422                 goto out_free;
4423         }
4424
4425         rc->merge_reloc_tree = 1;
4426
4427         while (!list_empty(&reloc_roots)) {
4428                 reloc_root = list_entry(reloc_roots.next,
4429                                         struct btrfs_root, root_list);
4430                 list_del(&reloc_root->root_list);
4431
4432                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4433                         list_add_tail(&reloc_root->root_list,
4434                                       &rc->reloc_roots);
4435                         continue;
4436                 }
4437
4438                 fs_root = read_fs_root(root->fs_info,
4439                                        reloc_root->root_key.offset);
4440                 if (IS_ERR(fs_root)) {
4441                         err = PTR_ERR(fs_root);
4442                         goto out_free;
4443                 }
4444
4445                 err = __add_reloc_root(reloc_root);
4446                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4447                 fs_root->reloc_root = reloc_root;
4448         }
4449
4450         err = btrfs_commit_transaction(trans, rc->extent_root);
4451         if (err)
4452                 goto out_free;
4453
4454         merge_reloc_roots(rc);
4455
4456         unset_reloc_control(rc);
4457
4458         trans = btrfs_join_transaction(rc->extent_root);
4459         if (IS_ERR(trans))
4460                 err = PTR_ERR(trans);
4461         else
4462                 err = btrfs_commit_transaction(trans, rc->extent_root);
4463 out_free:
4464         kfree(rc);
4465 out:
4466         if (!list_empty(&reloc_roots))
4467                 free_reloc_roots(&reloc_roots);
4468
4469         btrfs_free_path(path);
4470
4471         if (err == 0) {
4472                 /* cleanup orphan inode in data relocation tree */
4473                 fs_root = read_fs_root(root->fs_info,
4474                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4475                 if (IS_ERR(fs_root))
4476                         err = PTR_ERR(fs_root);
4477                 else
4478                         err = btrfs_orphan_cleanup(fs_root);
4479         }
4480         return err;
4481 }
4482
4483 /*
4484  * helper to add ordered checksum for data relocation.
4485  *
4486  * cloning checksum properly handles the nodatasum extents.
4487  * it also saves CPU time to re-calculate the checksum.
4488  */
4489 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4490 {
4491         struct btrfs_ordered_sum *sums;
4492         struct btrfs_ordered_extent *ordered;
4493         struct btrfs_root *root = BTRFS_I(inode)->root;
4494         int ret;
4495         u64 disk_bytenr;
4496         u64 new_bytenr;
4497         LIST_HEAD(list);
4498
4499         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4500         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4501
4502         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4503         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4504                                        disk_bytenr + len - 1, &list, 0);
4505         if (ret)
4506                 goto out;
4507
4508         while (!list_empty(&list)) {
4509                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4510                 list_del_init(&sums->list);
4511
4512                 /*
4513                  * We need to offset the new_bytenr based on where the csum is.
4514                  * We need to do this because we will read in entire prealloc
4515                  * extents but we may have written to say the middle of the
4516                  * prealloc extent, so we need to make sure the csum goes with
4517                  * the right disk offset.
4518                  *
4519                  * We can do this because the data reloc inode refers strictly
4520                  * to the on disk bytes, so we don't have to worry about
4521                  * disk_len vs real len like with real inodes since it's all
4522                  * disk length.
4523                  */
4524                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4525                 sums->bytenr = new_bytenr;
4526
4527                 btrfs_add_ordered_sum(inode, ordered, sums);
4528         }
4529 out:
4530         btrfs_put_ordered_extent(ordered);
4531         return ret;
4532 }
4533
4534 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4535                           struct btrfs_root *root, struct extent_buffer *buf,
4536                           struct extent_buffer *cow)
4537 {
4538         struct reloc_control *rc;
4539         struct backref_node *node;
4540         int first_cow = 0;
4541         int level;
4542         int ret = 0;
4543
4544         rc = root->fs_info->reloc_ctl;
4545         if (!rc)
4546                 return 0;
4547
4548         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4549                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4550
4551         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4552                 if (buf == root->node)
4553                         __update_reloc_root(root, cow->start);
4554         }
4555
4556         level = btrfs_header_level(buf);
4557         if (btrfs_header_generation(buf) <=
4558             btrfs_root_last_snapshot(&root->root_item))
4559                 first_cow = 1;
4560
4561         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4562             rc->create_reloc_tree) {
4563                 WARN_ON(!first_cow && level == 0);
4564
4565                 node = rc->backref_cache.path[level];
4566                 BUG_ON(node->bytenr != buf->start &&
4567                        node->new_bytenr != buf->start);
4568
4569                 drop_node_buffer(node);
4570                 extent_buffer_get(cow);
4571                 node->eb = cow;
4572                 node->new_bytenr = cow->start;
4573
4574                 if (!node->pending) {
4575                         list_move_tail(&node->list,
4576                                        &rc->backref_cache.pending[level]);
4577                         node->pending = 1;
4578                 }
4579
4580                 if (first_cow)
4581                         __mark_block_processed(rc, node);
4582
4583                 if (first_cow && level > 0)
4584                         rc->nodes_relocated += buf->len;
4585         }
4586
4587         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4588                 ret = replace_file_extents(trans, rc, root, cow);
4589         return ret;
4590 }
4591
4592 /*
4593  * called before creating snapshot. it calculates metadata reservation
4594  * requried for relocating tree blocks in the snapshot
4595  */
4596 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4597                               u64 *bytes_to_reserve)
4598 {
4599         struct btrfs_root *root;
4600         struct reloc_control *rc;
4601
4602         root = pending->root;
4603         if (!root->reloc_root)
4604                 return;
4605
4606         rc = root->fs_info->reloc_ctl;
4607         if (!rc->merge_reloc_tree)
4608                 return;
4609
4610         root = root->reloc_root;
4611         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4612         /*
4613          * relocation is in the stage of merging trees. the space
4614          * used by merging a reloc tree is twice the size of
4615          * relocated tree nodes in the worst case. half for cowing
4616          * the reloc tree, half for cowing the fs tree. the space
4617          * used by cowing the reloc tree will be freed after the
4618          * tree is dropped. if we create snapshot, cowing the fs
4619          * tree may use more space than it frees. so we need
4620          * reserve extra space.
4621          */
4622         *bytes_to_reserve += rc->nodes_relocated;
4623 }
4624
4625 /*
4626  * called after snapshot is created. migrate block reservation
4627  * and create reloc root for the newly created snapshot
4628  */
4629 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4630                                struct btrfs_pending_snapshot *pending)
4631 {
4632         struct btrfs_root *root = pending->root;
4633         struct btrfs_root *reloc_root;
4634         struct btrfs_root *new_root;
4635         struct reloc_control *rc;
4636         int ret;
4637
4638         if (!root->reloc_root)
4639                 return 0;
4640
4641         rc = root->fs_info->reloc_ctl;
4642         rc->merging_rsv_size += rc->nodes_relocated;
4643
4644         if (rc->merge_reloc_tree) {
4645                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4646                                               rc->block_rsv,
4647                                               rc->nodes_relocated);
4648                 if (ret)
4649                         return ret;
4650         }
4651
4652         new_root = pending->snap;
4653         reloc_root = create_reloc_root(trans, root->reloc_root,
4654                                        new_root->root_key.objectid);
4655         if (IS_ERR(reloc_root))
4656                 return PTR_ERR(reloc_root);
4657
4658         ret = __add_reloc_root(reloc_root);
4659         BUG_ON(ret < 0);
4660         new_root->reloc_root = reloc_root;
4661
4662         if (rc->create_reloc_tree)
4663                 ret = clone_backref_node(trans, rc, root, reloc_root);
4664         return ret;
4665 }