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