fs/btrfs/relocation.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2009 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include "ctree.h"
  13 #include "disk-io.h"
  14 #include "transaction.h"
  15 #include "volumes.h"
  16 #include "locking.h"
  17 #include "btrfs_inode.h"
  18 #include "async-thread.h"
  19 #include "free-space-cache.h"
  20 #include "inode-map.h"
  21 #include "qgroup.h"
  22 #include "print-tree.h"
  23
  24 /*
  25  * backref_node, mapping_node and tree_block start with this
  26  */
  27 struct tree_entry {
  28         struct rb_node rb_node;
  29         u64 bytenr;
  30 };
  31
  32 /*
  33  * present a tree block in the backref cache
  34  */
  35 struct backref_node {
  36         struct rb_node rb_node;
  37         u64 bytenr;
  38
  39         u64 new_bytenr;
  40         /* objectid of tree block owner, can be not uptodate */
  41         u64 owner;
  42         /* link to pending, changed or detached list */
  43         struct list_head list;
  44         /* list of upper level blocks reference this block */
  45         struct list_head upper;
  46         /* list of child blocks in the cache */
  47         struct list_head lower;
  48         /* NULL if this node is not tree root */
  49         struct btrfs_root *root;
  50         /* extent buffer got by COW the block */
  51         struct extent_buffer *eb;
  52         /* level of tree block */
  53         unsigned int level:8;
  54         /* is the block in non-reference counted tree */
  55         unsigned int cowonly:1;
  56         /* 1 if no child node in the cache */
  57         unsigned int lowest:1;
  58         /* is the extent buffer locked */
  59         unsigned int locked:1;
  60         /* has the block been processed */
  61         unsigned int processed:1;
  62         /* have backrefs of this block been checked */
  63         unsigned int checked:1;
  64         /*
  65          * 1 if corresponding block has been cowed but some upper
  66          * level block pointers may not point to the new location
  67          */
  68         unsigned int pending:1;
  69         /*
  70          * 1 if the backref node isn't connected to any other
  71          * backref node.
  72          */
  73         unsigned int detached:1;
  74 };
  75
  76 /*
  77  * present a block pointer in the backref cache
  78  */
  79 struct backref_edge {
  80         struct list_head list[2];
  81         struct backref_node *node[2];
  82 };
  83
  84 #define LOWER   0
  85 #define UPPER   1
  86 #define RELOCATION_RESERVED_NODES       256
  87
  88 struct backref_cache {
  89         /* red black tree of all backref nodes in the cache */
  90         struct rb_root rb_root;
  91         /* for passing backref nodes to btrfs_reloc_cow_block */
  92         struct backref_node *path[BTRFS_MAX_LEVEL];
  93         /*
  94          * list of blocks that have been cowed but some block
  95          * pointers in upper level blocks may not reflect the
  96          * new location
  97          */
  98         struct list_head pending[BTRFS_MAX_LEVEL];
  99         /* list of backref nodes with no child node */
 100         struct list_head leaves;
 101         /* list of blocks that have been cowed in current transaction */
 102         struct list_head changed;
 103         /* list of detached backref node. */
 104         struct list_head detached;
 105
 106         u64 last_trans;
 107
 108         int nr_nodes;
 109         int nr_edges;
 110 };
 111
 112 /*
 113  * map address of tree root to tree
 114  */
 115 struct mapping_node {
 116         struct rb_node rb_node;
 117         u64 bytenr;
 118         void *data;
 119 };
 120
 121 struct mapping_tree {
 122         struct rb_root rb_root;
 123         spinlock_t lock;
 124 };
 125
 126 /*
 127  * present a tree block to process
 128  */
 129 struct tree_block {
 130         struct rb_node rb_node;
 131         u64 bytenr;
 132         struct btrfs_key key;
 133         unsigned int level:8;
 134         unsigned int key_ready:1;
 135 };
 136
 137 #define MAX_EXTENTS 128
 138
 139 struct file_extent_cluster {
 140         u64 start;
 141         u64 end;
 142         u64 boundary[MAX_EXTENTS];
 143         unsigned int nr;
 144 };
 145
 146 struct reloc_control {
 147         /* block group to relocate */
 148         struct btrfs_block_group_cache *block_group;
 149         /* extent tree */
 150         struct btrfs_root *extent_root;
 151         /* inode for moving data */
 152         struct inode *data_inode;
 153
 154         struct btrfs_block_rsv *block_rsv;
 155
 156         struct backref_cache backref_cache;
 157
 158         struct file_extent_cluster cluster;
 159         /* tree blocks have been processed */
 160         struct extent_io_tree processed_blocks;
 161         /* map start of tree root to corresponding reloc tree */
 162         struct mapping_tree reloc_root_tree;
 163         /* list of reloc trees */
 164         struct list_head reloc_roots;
 165         /* list of subvolume trees that get relocated */
 166         struct list_head dirty_subvol_roots;
 167         /* size of metadata reservation for merging reloc trees */
 168         u64 merging_rsv_size;
 169         /* size of relocated tree nodes */
 170         u64 nodes_relocated;
 171         /* reserved size for block group relocation*/
 172         u64 reserved_bytes;
 173
 174         u64 search_start;
 175         u64 extents_found;
 176
 177         unsigned int stage:8;
 178         unsigned int create_reloc_tree:1;
 179         unsigned int merge_reloc_tree:1;
 180         unsigned int found_file_extent:1;
 181 };
 182
 183 /* stages of data relocation */
 184 #define MOVE_DATA_EXTENTS       0
 185 #define UPDATE_DATA_PTRS        1
 186
 187 static void remove_backref_node(struct backref_cache *cache,
 188                                 struct backref_node *node);
 189 static void __mark_block_processed(struct reloc_control *rc,
 190                                    struct backref_node *node);
 191
 192 static void mapping_tree_init(struct mapping_tree *tree)
 193 {
 194         tree->rb_root = RB_ROOT;
 195         spin_lock_init(&tree->lock);
 196 }
 197
 198 static void backref_cache_init(struct backref_cache *cache)
 199 {
 200         int i;
 201         cache->rb_root = RB_ROOT;
 202         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 203                 INIT_LIST_HEAD(&cache->pending[i]);
 204         INIT_LIST_HEAD(&cache->changed);
 205         INIT_LIST_HEAD(&cache->detached);
 206         INIT_LIST_HEAD(&cache->leaves);
 207 }
 208
 209 static void backref_cache_cleanup(struct backref_cache *cache)
 210 {
 211         struct backref_node *node;
 212         int i;
 213
 214         while (!list_empty(&cache->detached)) {
 215                 node = list_entry(cache->detached.next,
 216                                   struct backref_node, list);
 217                 remove_backref_node(cache, node);
 218         }
 219
 220         while (!list_empty(&cache->leaves)) {
 221                 node = list_entry(cache->leaves.next,
 222                                   struct backref_node, lower);
 223                 remove_backref_node(cache, node);
 224         }
 225
 226         cache->last_trans = 0;
 227
 228         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 229                 ASSERT(list_empty(&cache->pending[i]));
 230         ASSERT(list_empty(&cache->changed));
 231         ASSERT(list_empty(&cache->detached));
 232         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
 233         ASSERT(!cache->nr_nodes);
 234         ASSERT(!cache->nr_edges);
 235 }
 236
 237 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 238 {
 239         struct backref_node *node;
 240
 241         node = kzalloc(sizeof(*node), GFP_NOFS);
 242         if (node) {
 243                 INIT_LIST_HEAD(&node->list);
 244                 INIT_LIST_HEAD(&node->upper);
 245                 INIT_LIST_HEAD(&node->lower);
 246                 RB_CLEAR_NODE(&node->rb_node);
 247                 cache->nr_nodes++;
 248         }
 249         return node;
 250 }
 251
 252 static void free_backref_node(struct backref_cache *cache,
 253                               struct backref_node *node)
 254 {
 255         if (node) {
 256                 cache->nr_nodes--;
 257                 kfree(node);
 258         }
 259 }
 260
 261 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 262 {
 263         struct backref_edge *edge;
 264
 265         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 266         if (edge)
 267                 cache->nr_edges++;
 268         return edge;
 269 }
 270
 271 static void free_backref_edge(struct backref_cache *cache,
 272                               struct backref_edge *edge)
 273 {
 274         if (edge) {
 275                 cache->nr_edges--;
 276                 kfree(edge);
 277         }
 278 }
 279
 280 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 281                                    struct rb_node *node)
 282 {
 283         struct rb_node **p = &root->rb_node;
 284         struct rb_node *parent = NULL;
 285         struct tree_entry *entry;
 286
 287         while (*p) {
 288                 parent = *p;
 289                 entry = rb_entry(parent, struct tree_entry, rb_node);
 290
 291                 if (bytenr < entry->bytenr)
 292                         p = &(*p)->rb_left;
 293                 else if (bytenr > entry->bytenr)
 294                         p = &(*p)->rb_right;
 295                 else
 296                         return parent;
 297         }
 298
 299         rb_link_node(node, parent, p);
 300         rb_insert_color(node, root);
 301         return NULL;
 302 }
 303
 304 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 305 {
 306         struct rb_node *n = root->rb_node;
 307         struct tree_entry *entry;
 308
 309         while (n) {
 310                 entry = rb_entry(n, struct tree_entry, rb_node);
 311
 312                 if (bytenr < entry->bytenr)
 313                         n = n->rb_left;
 314                 else if (bytenr > entry->bytenr)
 315                         n = n->rb_right;
 316                 else
 317                         return n;
 318         }
 319         return NULL;
 320 }
 321
 322 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 323 {
 324
 325         struct btrfs_fs_info *fs_info = NULL;
 326         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 327                                               rb_node);
 328         if (bnode->root)
 329                 fs_info = bnode->root->fs_info;
 330         btrfs_panic(fs_info, errno,
 331                     "Inconsistency in backref cache found at offset %llu",
 332                     bytenr);
 333 }
 334
 335 /*
 336  * walk up backref nodes until reach node presents tree root
 337  */
 338 static struct backref_node *walk_up_backref(struct backref_node *node,
 339                                             struct backref_edge *edges[],
 340                                             int *index)
 341 {
 342         struct backref_edge *edge;
 343         int idx = *index;
 344
 345         while (!list_empty(&node->upper)) {
 346                 edge = list_entry(node->upper.next,
 347                                   struct backref_edge, list[LOWER]);
 348                 edges[idx++] = edge;
 349                 node = edge->node[UPPER];
 350         }
 351         BUG_ON(node->detached);
 352         *index = idx;
 353         return node;
 354 }
 355
 356 /*
 357  * walk down backref nodes to find start of next reference path
 358  */
 359 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 360                                               int *index)
 361 {
 362         struct backref_edge *edge;
 363         struct backref_node *lower;
 364         int idx = *index;
 365
 366         while (idx > 0) {
 367                 edge = edges[idx - 1];
 368                 lower = edge->node[LOWER];
 369                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 370                         idx--;
 371                         continue;
 372                 }
 373                 edge = list_entry(edge->list[LOWER].next,
 374                                   struct backref_edge, list[LOWER]);
 375                 edges[idx - 1] = edge;
 376                 *index = idx;
 377                 return edge->node[UPPER];
 378         }
 379         *index = 0;
 380         return NULL;
 381 }
 382
 383 static void unlock_node_buffer(struct backref_node *node)
 384 {
 385         if (node->locked) {
 386                 btrfs_tree_unlock(node->eb);
 387                 node->locked = 0;
 388         }
 389 }
 390
 391 static void drop_node_buffer(struct backref_node *node)
 392 {
 393         if (node->eb) {
 394                 unlock_node_buffer(node);
 395                 free_extent_buffer(node->eb);
 396                 node->eb = NULL;
 397         }
 398 }
 399
 400 static void drop_backref_node(struct backref_cache *tree,
 401                               struct backref_node *node)
 402 {
 403         BUG_ON(!list_empty(&node->upper));
 404
 405         drop_node_buffer(node);
 406         list_del(&node->list);
 407         list_del(&node->lower);
 408         if (!RB_EMPTY_NODE(&node->rb_node))
 409                 rb_erase(&node->rb_node, &tree->rb_root);
 410         free_backref_node(tree, node);
 411 }
 412
 413 /*
 414  * remove a backref node from the backref cache
 415  */
 416 static void remove_backref_node(struct backref_cache *cache,
 417                                 struct backref_node *node)
 418 {
 419         struct backref_node *upper;
 420         struct backref_edge *edge;
 421
 422         if (!node)
 423                 return;
 424
 425         BUG_ON(!node->lowest && !node->detached);
 426         while (!list_empty(&node->upper)) {
 427                 edge = list_entry(node->upper.next, struct backref_edge,
 428                                   list[LOWER]);
 429                 upper = edge->node[UPPER];
 430                 list_del(&edge->list[LOWER]);
 431                 list_del(&edge->list[UPPER]);
 432                 free_backref_edge(cache, edge);
 433
 434                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 435                         BUG_ON(!list_empty(&node->upper));
 436                         drop_backref_node(cache, node);
 437                         node = upper;
 438                         node->lowest = 1;
 439                         continue;
 440                 }
 441                 /*
 442                  * add the node to leaf node list if no other
 443                  * child block cached.
 444                  */
 445                 if (list_empty(&upper->lower)) {
 446                         list_add_tail(&upper->lower, &cache->leaves);
 447                         upper->lowest = 1;
 448                 }
 449         }
 450
 451         drop_backref_node(cache, node);
 452 }
 453
 454 static void update_backref_node(struct backref_cache *cache,
 455                                 struct backref_node *node, u64 bytenr)
 456 {
 457         struct rb_node *rb_node;
 458         rb_erase(&node->rb_node, &cache->rb_root);
 459         node->bytenr = bytenr;
 460         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 461         if (rb_node)
 462                 backref_tree_panic(rb_node, -EEXIST, bytenr);
 463 }
 464
 465 /*
 466  * update backref cache after a transaction commit
 467  */
 468 static int update_backref_cache(struct btrfs_trans_handle *trans,
 469                                 struct backref_cache *cache)
 470 {
 471         struct backref_node *node;
 472         int level = 0;
 473
 474         if (cache->last_trans == 0) {
 475                 cache->last_trans = trans->transid;
 476                 return 0;
 477         }
 478
 479         if (cache->last_trans == trans->transid)
 480                 return 0;
 481
 482         /*
 483          * detached nodes are used to avoid unnecessary backref
 484          * lookup. transaction commit changes the extent tree.
 485          * so the detached nodes are no longer useful.
 486          */
 487         while (!list_empty(&cache->detached)) {
 488                 node = list_entry(cache->detached.next,
 489                                   struct backref_node, list);
 490                 remove_backref_node(cache, node);
 491         }
 492
 493         while (!list_empty(&cache->changed)) {
 494                 node = list_entry(cache->changed.next,
 495                                   struct backref_node, list);
 496                 list_del_init(&node->list);
 497                 BUG_ON(node->pending);
 498                 update_backref_node(cache, node, node->new_bytenr);
 499         }
 500
 501         /*
 502          * some nodes can be left in the pending list if there were
 503          * errors during processing the pending nodes.
 504          */
 505         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 506                 list_for_each_entry(node, &cache->pending[level], list) {
 507                         BUG_ON(!node->pending);
 508                         if (node->bytenr == node->new_bytenr)
 509                                 continue;
 510                         update_backref_node(cache, node, node->new_bytenr);
 511                 }
 512         }
 513
 514         cache->last_trans = 0;
 515         return 1;
 516 }
 517
 518
 519 static int should_ignore_root(struct btrfs_root *root)
 520 {
 521         struct btrfs_root *reloc_root;
 522
 523         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 524                 return 0;
 525
 526         reloc_root = root->reloc_root;
 527         if (!reloc_root)
 528                 return 0;
 529
 530         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 531             root->fs_info->running_transaction->transid - 1)
 532                 return 0;
 533         /*
 534          * if there is reloc tree and it was created in previous
 535          * transaction backref lookup can find the reloc tree,
 536          * so backref node for the fs tree root is useless for
 537          * relocation.
 538          */
 539         return 1;
 540 }
 541 /*
 542  * find reloc tree by address of tree root
 543  */
 544 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 545                                           u64 bytenr)
 546 {
 547         struct rb_node *rb_node;
 548         struct mapping_node *node;
 549         struct btrfs_root *root = NULL;
 550
 551         spin_lock(&rc->reloc_root_tree.lock);
 552         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 553         if (rb_node) {
 554                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 555                 root = (struct btrfs_root *)node->data;
 556         }
 557         spin_unlock(&rc->reloc_root_tree.lock);
 558         return root;
 559 }
 560
 561 static int is_cowonly_root(u64 root_objectid)
 562 {
 563         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 564             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 565             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 566             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 567             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 568             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 569             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 570             root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
 571             root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
 572                 return 1;
 573         return 0;
 574 }
 575
 576 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 577                                         u64 root_objectid)
 578 {
 579         struct btrfs_key key;
 580
 581         key.objectid = root_objectid;
 582         key.type = BTRFS_ROOT_ITEM_KEY;
 583         if (is_cowonly_root(root_objectid))
 584                 key.offset = 0;
 585         else
 586                 key.offset = (u64)-1;
 587
 588         return btrfs_get_fs_root(fs_info, &key, false);
 589 }
 590
 591 static noinline_for_stack
 592 int find_inline_backref(struct extent_buffer *leaf, int slot,
 593                         unsigned long *ptr, unsigned long *end)
 594 {
 595         struct btrfs_key key;
 596         struct btrfs_extent_item *ei;
 597         struct btrfs_tree_block_info *bi;
 598         u32 item_size;
 599
 600         btrfs_item_key_to_cpu(leaf, &key, slot);
 601
 602         item_size = btrfs_item_size_nr(leaf, slot);
 603         if (item_size < sizeof(*ei)) {
 604                 btrfs_print_v0_err(leaf->fs_info);
 605                 btrfs_handle_fs_error(leaf->fs_info, -EINVAL, NULL);
 606                 return 1;
 607         }
 608         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 609         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 610                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 611
 612         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 613             item_size <= sizeof(*ei) + sizeof(*bi)) {
 614                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 615                 return 1;
 616         }
 617         if (key.type == BTRFS_METADATA_ITEM_KEY &&
 618             item_size <= sizeof(*ei)) {
 619                 WARN_ON(item_size < sizeof(*ei));
 620                 return 1;
 621         }
 622
 623         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 624                 bi = (struct btrfs_tree_block_info *)(ei + 1);
 625                 *ptr = (unsigned long)(bi + 1);
 626         } else {
 627                 *ptr = (unsigned long)(ei + 1);
 628         }
 629         *end = (unsigned long)ei + item_size;
 630         return 0;
 631 }
 632
 633 /*
 634  * build backref tree for a given tree block. root of the backref tree
 635  * corresponds the tree block, leaves of the backref tree correspond
 636  * roots of b-trees that reference the tree block.
 637  *
 638  * the basic idea of this function is check backrefs of a given block
 639  * to find upper level blocks that reference the block, and then check
 640  * backrefs of these upper level blocks recursively. the recursion stop
 641  * when tree root is reached or backrefs for the block is cached.
 642  *
 643  * NOTE: if we find backrefs for a block are cached, we know backrefs
 644  * for all upper level blocks that directly/indirectly reference the
 645  * block are also cached.
 646  */
 647 static noinline_for_stack
 648 struct backref_node *build_backref_tree(struct reloc_control *rc,
 649                                         struct btrfs_key *node_key,
 650                                         int level, u64 bytenr)
 651 {
 652         struct backref_cache *cache = &rc->backref_cache;
 653         struct btrfs_path *path1; /* For searching extent root */
 654         struct btrfs_path *path2; /* For searching parent of TREE_BLOCK_REF */
 655         struct extent_buffer *eb;
 656         struct btrfs_root *root;
 657         struct backref_node *cur;
 658         struct backref_node *upper;
 659         struct backref_node *lower;
 660         struct backref_node *node = NULL;
 661         struct backref_node *exist = NULL;
 662         struct backref_edge *edge;
 663         struct rb_node *rb_node;
 664         struct btrfs_key key;
 665         unsigned long end;
 666         unsigned long ptr;
 667         LIST_HEAD(list); /* Pending edge list, upper node needs to be checked */
 668         LIST_HEAD(useless);
 669         int cowonly;
 670         int ret;
 671         int err = 0;
 672         bool need_check = true;
 673
 674         path1 = btrfs_alloc_path();
 675         path2 = btrfs_alloc_path();
 676         if (!path1 || !path2) {
 677                 err = -ENOMEM;
 678                 goto out;
 679         }
 680         path1->reada = READA_FORWARD;
 681         path2->reada = READA_FORWARD;
 682
 683         node = alloc_backref_node(cache);
 684         if (!node) {
 685                 err = -ENOMEM;
 686                 goto out;
 687         }
 688
 689         node->bytenr = bytenr;
 690         node->level = level;
 691         node->lowest = 1;
 692         cur = node;
 693 again:
 694         end = 0;
 695         ptr = 0;
 696         key.objectid = cur->bytenr;
 697         key.type = BTRFS_METADATA_ITEM_KEY;
 698         key.offset = (u64)-1;
 699
 700         path1->search_commit_root = 1;
 701         path1->skip_locking = 1;
 702         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 703                                 0, 0);
 704         if (ret < 0) {
 705                 err = ret;
 706                 goto out;
 707         }
 708         ASSERT(ret);
 709         ASSERT(path1->slots[0]);
 710
 711         path1->slots[0]--;
 712
 713         WARN_ON(cur->checked);
 714         if (!list_empty(&cur->upper)) {
 715                 /*
 716                  * the backref was added previously when processing
 717                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 718                  */
 719                 ASSERT(list_is_singular(&cur->upper));
 720                 edge = list_entry(cur->upper.next, struct backref_edge,
 721                                   list[LOWER]);
 722                 ASSERT(list_empty(&edge->list[UPPER]));
 723                 exist = edge->node[UPPER];
 724                 /*
 725                  * add the upper level block to pending list if we need
 726                  * check its backrefs
 727                  */
 728                 if (!exist->checked)
 729                         list_add_tail(&edge->list[UPPER], &list);
 730         } else {
 731                 exist = NULL;
 732         }
 733
 734         while (1) {
 735                 cond_resched();
 736                 eb = path1->nodes[0];
 737
 738                 if (ptr >= end) {
 739                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 740                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 741                                 if (ret < 0) {
 742                                         err = ret;
 743                                         goto out;
 744                                 }
 745                                 if (ret > 0)
 746                                         break;
 747                                 eb = path1->nodes[0];
 748                         }
 749
 750                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 751                         if (key.objectid != cur->bytenr) {
 752                                 WARN_ON(exist);
 753                                 break;
 754                         }
 755
 756                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 757                             key.type == BTRFS_METADATA_ITEM_KEY) {
 758                                 ret = find_inline_backref(eb, path1->slots[0],
 759                                                           &ptr, &end);
 760                                 if (ret)
 761                                         goto next;
 762                         }
 763                 }
 764
 765                 if (ptr < end) {
 766                         /* update key for inline back ref */
 767                         struct btrfs_extent_inline_ref *iref;
 768                         int type;
 769                         iref = (struct btrfs_extent_inline_ref *)ptr;
 770                         type = btrfs_get_extent_inline_ref_type(eb, iref,
 771                                                         BTRFS_REF_TYPE_BLOCK);
 772                         if (type == BTRFS_REF_TYPE_INVALID) {
 773                                 err = -EUCLEAN;
 774                                 goto out;
 775                         }
 776                         key.type = type;
 777                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 778
 779                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 780                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 781                 }
 782
 783                 /*
 784                  * Parent node found and matches current inline ref, no need to
 785                  * rebuild this node for this inline ref.
 786                  */
 787                 if (exist &&
 788                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 789                       exist->owner == key.offset) ||
 790                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 791                       exist->bytenr == key.offset))) {
 792                         exist = NULL;
 793                         goto next;
 794                 }
 795
 796                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
 797                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 798                         if (key.objectid == key.offset) {
 799                                 /*
 800                                  * Only root blocks of reloc trees use backref
 801                                  * pointing to itself.
 802                                  */
 803                                 root = find_reloc_root(rc, cur->bytenr);
 804                                 ASSERT(root);
 805                                 cur->root = root;
 806                                 break;
 807                         }
 808
 809                         edge = alloc_backref_edge(cache);
 810                         if (!edge) {
 811                                 err = -ENOMEM;
 812                                 goto out;
 813                         }
 814                         rb_node = tree_search(&cache->rb_root, key.offset);
 815                         if (!rb_node) {
 816                                 upper = alloc_backref_node(cache);
 817                                 if (!upper) {
 818                                         free_backref_edge(cache, edge);
 819                                         err = -ENOMEM;
 820                                         goto out;
 821                                 }
 822                                 upper->bytenr = key.offset;
 823                                 upper->level = cur->level + 1;
 824                                 /*
 825                                  *  backrefs for the upper level block isn't
 826                                  *  cached, add the block to pending list
 827                                  */
 828                                 list_add_tail(&edge->list[UPPER], &list);
 829                         } else {
 830                                 upper = rb_entry(rb_node, struct backref_node,
 831                                                  rb_node);
 832                                 ASSERT(upper->checked);
 833                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 834                         }
 835                         list_add_tail(&edge->list[LOWER], &cur->upper);
 836                         edge->node[LOWER] = cur;
 837                         edge->node[UPPER] = upper;
 838
 839                         goto next;
 840                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 841                         err = -EINVAL;
 842                         btrfs_print_v0_err(rc->extent_root->fs_info);
 843                         btrfs_handle_fs_error(rc->extent_root->fs_info, err,
 844                                               NULL);
 845                         goto out;
 846                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 847                         goto next;
 848                 }
 849
 850                 /*
 851                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
 852                  * means the root objectid. We need to search the tree to get
 853                  * its parent bytenr.
 854                  */
 855                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 856                 if (IS_ERR(root)) {
 857                         err = PTR_ERR(root);
 858                         goto out;
 859                 }
 860
 861                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 862                         cur->cowonly = 1;
 863
 864                 if (btrfs_root_level(&root->root_item) == cur->level) {
 865                         /* tree root */
 866                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
 867                                cur->bytenr);
 868                         if (should_ignore_root(root))
 869                                 list_add(&cur->list, &useless);
 870                         else
 871                                 cur->root = root;
 872                         break;
 873                 }
 874
 875                 level = cur->level + 1;
 876
 877                 /* Search the tree to find parent blocks referring the block. */
 878                 path2->search_commit_root = 1;
 879                 path2->skip_locking = 1;
 880                 path2->lowest_level = level;
 881                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 882                 path2->lowest_level = 0;
 883                 if (ret < 0) {
 884                         err = ret;
 885                         goto out;
 886                 }
 887                 if (ret > 0 && path2->slots[level] > 0)
 888                         path2->slots[level]--;
 889
 890                 eb = path2->nodes[level];
 891                 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
 892                     cur->bytenr) {
 893                         btrfs_err(root->fs_info,
 894         "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
 895                                   cur->bytenr, level - 1,
 896                                   root->root_key.objectid,
 897                                   node_key->objectid, node_key->type,
 898                                   node_key->offset);
 899                         err = -ENOENT;
 900                         goto out;
 901                 }
 902                 lower = cur;
 903                 need_check = true;
 904
 905                 /* Add all nodes and edges in the path */
 906                 for (; level < BTRFS_MAX_LEVEL; level++) {
 907                         if (!path2->nodes[level]) {
 908                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
 909                                        lower->bytenr);
 910                                 if (should_ignore_root(root))
 911                                         list_add(&lower->list, &useless);
 912                                 else
 913                                         lower->root = root;
 914                                 break;
 915                         }
 916
 917                         edge = alloc_backref_edge(cache);
 918                         if (!edge) {
 919                                 err = -ENOMEM;
 920                                 goto out;
 921                         }
 922
 923                         eb = path2->nodes[level];
 924                         rb_node = tree_search(&cache->rb_root, eb->start);
 925                         if (!rb_node) {
 926                                 upper = alloc_backref_node(cache);
 927                                 if (!upper) {
 928                                         free_backref_edge(cache, edge);
 929                                         err = -ENOMEM;
 930                                         goto out;
 931                                 }
 932                                 upper->bytenr = eb->start;
 933                                 upper->owner = btrfs_header_owner(eb);
 934                                 upper->level = lower->level + 1;
 935                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
 936                                               &root->state))
 937                                         upper->cowonly = 1;
 938
 939                                 /*
 940                                  * if we know the block isn't shared
 941                                  * we can void checking its backrefs.
 942                                  */
 943                                 if (btrfs_block_can_be_shared(root, eb))
 944                                         upper->checked = 0;
 945                                 else
 946                                         upper->checked = 1;
 947
 948                                 /*
 949                                  * add the block to pending list if we
 950                                  * need check its backrefs, we only do this once
 951                                  * while walking up a tree as we will catch
 952                                  * anything else later on.
 953                                  */
 954                                 if (!upper->checked && need_check) {
 955                                         need_check = false;
 956                                         list_add_tail(&edge->list[UPPER],
 957                                                       &list);
 958                                 } else {
 959                                         if (upper->checked)
 960                                                 need_check = true;
 961                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 962                                 }
 963                         } else {
 964                                 upper = rb_entry(rb_node, struct backref_node,
 965                                                  rb_node);
 966                                 ASSERT(upper->checked);
 967                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 968                                 if (!upper->owner)
 969                                         upper->owner = btrfs_header_owner(eb);
 970                         }
 971                         list_add_tail(&edge->list[LOWER], &lower->upper);
 972                         edge->node[LOWER] = lower;
 973                         edge->node[UPPER] = upper;
 974
 975                         if (rb_node)
 976                                 break;
 977                         lower = upper;
 978                         upper = NULL;
 979                 }
 980                 btrfs_release_path(path2);
 981 next:
 982                 if (ptr < end) {
 983                         ptr += btrfs_extent_inline_ref_size(key.type);
 984                         if (ptr >= end) {
 985                                 WARN_ON(ptr > end);
 986                                 ptr = 0;
 987                                 end = 0;
 988                         }
 989                 }
 990                 if (ptr >= end)
 991                         path1->slots[0]++;
 992         }
 993         btrfs_release_path(path1);
 994
 995         cur->checked = 1;
 996         WARN_ON(exist);
 997
 998         /* the pending list isn't empty, take the first block to process */
 999         if (!list_empty(&list)) {
1000                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1001                 list_del_init(&edge->list[UPPER]);
1002                 cur = edge->node[UPPER];
1003                 goto again;
1004         }
1005
1006         /*
1007          * everything goes well, connect backref nodes and insert backref nodes
1008          * into the cache.
1009          */
1010         ASSERT(node->checked);
1011         cowonly = node->cowonly;
1012         if (!cowonly) {
1013                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1014                                       &node->rb_node);
1015                 if (rb_node)
1016                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1017                 list_add_tail(&node->lower, &cache->leaves);
1018         }
1019
1020         list_for_each_entry(edge, &node->upper, list[LOWER])
1021                 list_add_tail(&edge->list[UPPER], &list);
1022
1023         while (!list_empty(&list)) {
1024                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1025                 list_del_init(&edge->list[UPPER]);
1026                 upper = edge->node[UPPER];
1027                 if (upper->detached) {
1028                         list_del(&edge->list[LOWER]);
1029                         lower = edge->node[LOWER];
1030                         free_backref_edge(cache, edge);
1031                         if (list_empty(&lower->upper))
1032                                 list_add(&lower->list, &useless);
1033                         continue;
1034                 }
1035
1036                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1037                         if (upper->lowest) {
1038                                 list_del_init(&upper->lower);
1039                                 upper->lowest = 0;
1040                         }
1041
1042                         list_add_tail(&edge->list[UPPER], &upper->lower);
1043                         continue;
1044                 }
1045
1046                 if (!upper->checked) {
1047                         /*
1048                          * Still want to blow up for developers since this is a
1049                          * logic bug.
1050                          */
1051                         ASSERT(0);
1052                         err = -EINVAL;
1053                         goto out;
1054                 }
1055                 if (cowonly != upper->cowonly) {
1056                         ASSERT(0);
1057                         err = -EINVAL;
1058                         goto out;
1059                 }
1060
1061                 if (!cowonly) {
1062                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1063                                               &upper->rb_node);
1064                         if (rb_node)
1065                                 backref_tree_panic(rb_node, -EEXIST,
1066                                                    upper->bytenr);
1067                 }
1068
1069                 list_add_tail(&edge->list[UPPER], &upper->lower);
1070
1071                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1072                         list_add_tail(&edge->list[UPPER], &list);
1073         }
1074         /*
1075          * process useless backref nodes. backref nodes for tree leaves
1076          * are deleted from the cache. backref nodes for upper level
1077          * tree blocks are left in the cache to avoid unnecessary backref
1078          * lookup.
1079          */
1080         while (!list_empty(&useless)) {
1081                 upper = list_entry(useless.next, struct backref_node, list);
1082                 list_del_init(&upper->list);
1083                 ASSERT(list_empty(&upper->upper));
1084                 if (upper == node)
1085                         node = NULL;
1086                 if (upper->lowest) {
1087                         list_del_init(&upper->lower);
1088                         upper->lowest = 0;
1089                 }
1090                 while (!list_empty(&upper->lower)) {
1091                         edge = list_entry(upper->lower.next,
1092                                           struct backref_edge, list[UPPER]);
1093                         list_del(&edge->list[UPPER]);
1094                         list_del(&edge->list[LOWER]);
1095                         lower = edge->node[LOWER];
1096                         free_backref_edge(cache, edge);
1097
1098                         if (list_empty(&lower->upper))
1099                                 list_add(&lower->list, &useless);
1100                 }
1101                 __mark_block_processed(rc, upper);
1102                 if (upper->level > 0) {
1103                         list_add(&upper->list, &cache->detached);
1104                         upper->detached = 1;
1105                 } else {
1106                         rb_erase(&upper->rb_node, &cache->rb_root);
1107                         free_backref_node(cache, upper);
1108                 }
1109         }
1110 out:
1111         btrfs_free_path(path1);
1112         btrfs_free_path(path2);
1113         if (err) {
1114                 while (!list_empty(&useless)) {
1115                         lower = list_entry(useless.next,
1116                                            struct backref_node, list);
1117                         list_del_init(&lower->list);
1118                 }
1119                 while (!list_empty(&list)) {
1120                         edge = list_first_entry(&list, struct backref_edge,
1121                                                 list[UPPER]);
1122                         list_del(&edge->list[UPPER]);
1123                         list_del(&edge->list[LOWER]);
1124                         lower = edge->node[LOWER];
1125                         upper = edge->node[UPPER];
1126                         free_backref_edge(cache, edge);
1127
1128                         /*
1129                          * Lower is no longer linked to any upper backref nodes
1130                          * and isn't in the cache, we can free it ourselves.
1131                          */
1132                         if (list_empty(&lower->upper) &&
1133                             RB_EMPTY_NODE(&lower->rb_node))
1134                                 list_add(&lower->list, &useless);
1135
1136                         if (!RB_EMPTY_NODE(&upper->rb_node))
1137                                 continue;
1138
1139                         /* Add this guy's upper edges to the list to process */
1140                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1141                                 list_add_tail(&edge->list[UPPER], &list);
1142                         if (list_empty(&upper->upper))
1143                                 list_add(&upper->list, &useless);
1144                 }
1145
1146                 while (!list_empty(&useless)) {
1147                         lower = list_entry(useless.next,
1148                                            struct backref_node, list);
1149                         list_del_init(&lower->list);
1150                         if (lower == node)
1151                                 node = NULL;
1152                         free_backref_node(cache, lower);
1153                 }
1154
1155                 free_backref_node(cache, node);
1156                 return ERR_PTR(err);
1157         }
1158         ASSERT(!node || !node->detached);
1159         return node;
1160 }
1161
1162 /*
1163  * helper to add backref node for the newly created snapshot.
1164  * the backref node is created by cloning backref node that
1165  * corresponds to root of source tree
1166  */
1167 static int clone_backref_node(struct btrfs_trans_handle *trans,
1168                               struct reloc_control *rc,
1169                               struct btrfs_root *src,
1170                               struct btrfs_root *dest)
1171 {
1172         struct btrfs_root *reloc_root = src->reloc_root;
1173         struct backref_cache *cache = &rc->backref_cache;
1174         struct backref_node *node = NULL;
1175         struct backref_node *new_node;
1176         struct backref_edge *edge;
1177         struct backref_edge *new_edge;
1178         struct rb_node *rb_node;
1179
1180         if (cache->last_trans > 0)
1181                 update_backref_cache(trans, cache);
1182
1183         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1184         if (rb_node) {
1185                 node = rb_entry(rb_node, struct backref_node, rb_node);
1186                 if (node->detached)
1187                         node = NULL;
1188                 else
1189                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1190         }
1191
1192         if (!node) {
1193                 rb_node = tree_search(&cache->rb_root,
1194                                       reloc_root->commit_root->start);
1195                 if (rb_node) {
1196                         node = rb_entry(rb_node, struct backref_node,
1197                                         rb_node);
1198                         BUG_ON(node->detached);
1199                 }
1200         }
1201
1202         if (!node)
1203                 return 0;
1204
1205         new_node = alloc_backref_node(cache);
1206         if (!new_node)
1207                 return -ENOMEM;
1208
1209         new_node->bytenr = dest->node->start;
1210         new_node->level = node->level;
1211         new_node->lowest = node->lowest;
1212         new_node->checked = 1;
1213         new_node->root = dest;
1214
1215         if (!node->lowest) {
1216                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1217                         new_edge = alloc_backref_edge(cache);
1218                         if (!new_edge)
1219                                 goto fail;
1220
1221                         new_edge->node[UPPER] = new_node;
1222                         new_edge->node[LOWER] = edge->node[LOWER];
1223                         list_add_tail(&new_edge->list[UPPER],
1224                                       &new_node->lower);
1225                 }
1226         } else {
1227                 list_add_tail(&new_node->lower, &cache->leaves);
1228         }
1229
1230         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1231                               &new_node->rb_node);
1232         if (rb_node)
1233                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1234
1235         if (!new_node->lowest) {
1236                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1237                         list_add_tail(&new_edge->list[LOWER],
1238                                       &new_edge->node[LOWER]->upper);
1239                 }
1240         }
1241         return 0;
1242 fail:
1243         while (!list_empty(&new_node->lower)) {
1244                 new_edge = list_entry(new_node->lower.next,
1245                                       struct backref_edge, list[UPPER]);
1246                 list_del(&new_edge->list[UPPER]);
1247                 free_backref_edge(cache, new_edge);
1248         }
1249         free_backref_node(cache, new_node);
1250         return -ENOMEM;
1251 }
1252
1253 /*
1254  * helper to add 'address of tree root -> reloc tree' mapping
1255  */
1256 static int __must_check __add_reloc_root(struct btrfs_root *root)
1257 {
1258         struct btrfs_fs_info *fs_info = root->fs_info;
1259         struct rb_node *rb_node;
1260         struct mapping_node *node;
1261         struct reloc_control *rc = fs_info->reloc_ctl;
1262
1263         node = kmalloc(sizeof(*node), GFP_NOFS);
1264         if (!node)
1265                 return -ENOMEM;
1266
1267         node->bytenr = root->node->start;
1268         node->data = root;
1269
1270         spin_lock(&rc->reloc_root_tree.lock);
1271         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1272                               node->bytenr, &node->rb_node);
1273         spin_unlock(&rc->reloc_root_tree.lock);
1274         if (rb_node) {
1275                 btrfs_panic(fs_info, -EEXIST,
1276                             "Duplicate root found for start=%llu while inserting into relocation tree",
1277                             node->bytenr);
1278         }
1279
1280         list_add_tail(&root->root_list, &rc->reloc_roots);
1281         return 0;
1282 }
1283
1284 /*
1285  * helper to delete the 'address of tree root -> reloc tree'
1286  * mapping
1287  */
1288 static void __del_reloc_root(struct btrfs_root *root)
1289 {
1290         struct btrfs_fs_info *fs_info = root->fs_info;
1291         struct rb_node *rb_node;
1292         struct mapping_node *node = NULL;
1293         struct reloc_control *rc = fs_info->reloc_ctl;
1294
1295         if (rc && root->node) {
1296                 spin_lock(&rc->reloc_root_tree.lock);
1297                 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1298                                       root->node->start);
1299                 if (rb_node) {
1300                         node = rb_entry(rb_node, struct mapping_node, rb_node);
1301                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1302                 }
1303                 spin_unlock(&rc->reloc_root_tree.lock);
1304                 if (!node)
1305                         return;
1306                 BUG_ON((struct btrfs_root *)node->data != root);
1307         }
1308
1309         spin_lock(&fs_info->trans_lock);
1310         list_del_init(&root->root_list);
1311         spin_unlock(&fs_info->trans_lock);
1312         kfree(node);
1313 }
1314
1315 /*
1316  * helper to update the 'address of tree root -> reloc tree'
1317  * mapping
1318  */
1319 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1320 {
1321         struct btrfs_fs_info *fs_info = root->fs_info;
1322         struct rb_node *rb_node;
1323         struct mapping_node *node = NULL;
1324         struct reloc_control *rc = fs_info->reloc_ctl;
1325
1326         spin_lock(&rc->reloc_root_tree.lock);
1327         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1328                               root->node->start);
1329         if (rb_node) {
1330                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1331                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1332         }
1333         spin_unlock(&rc->reloc_root_tree.lock);
1334
1335         if (!node)
1336                 return 0;
1337         BUG_ON((struct btrfs_root *)node->data != root);
1338
1339         spin_lock(&rc->reloc_root_tree.lock);
1340         node->bytenr = new_bytenr;
1341         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1342                               node->bytenr, &node->rb_node);
1343         spin_unlock(&rc->reloc_root_tree.lock);
1344         if (rb_node)
1345                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1346         return 0;
1347 }
1348
1349 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1350                                         struct btrfs_root *root, u64 objectid)
1351 {
1352         struct btrfs_fs_info *fs_info = root->fs_info;
1353         struct btrfs_root *reloc_root;
1354         struct extent_buffer *eb;
1355         struct btrfs_root_item *root_item;
1356         struct btrfs_key root_key;
1357         int ret;
1358
1359         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1360         BUG_ON(!root_item);
1361
1362         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1363         root_key.type = BTRFS_ROOT_ITEM_KEY;
1364         root_key.offset = objectid;
1365
1366         if (root->root_key.objectid == objectid) {
1367                 u64 commit_root_gen;
1368
1369                 /* called by btrfs_init_reloc_root */
1370                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1371                                       BTRFS_TREE_RELOC_OBJECTID);
1372                 BUG_ON(ret);
1373                 /*
1374                  * Set the last_snapshot field to the generation of the commit
1375                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1376                  * correctly (returns true) when the relocation root is created
1377                  * either inside the critical section of a transaction commit
1378                  * (through transaction.c:qgroup_account_snapshot()) and when
1379                  * it's created before the transaction commit is started.
1380                  */
1381                 commit_root_gen = btrfs_header_generation(root->commit_root);
1382                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1383         } else {
1384                 /*
1385                  * called by btrfs_reloc_post_snapshot_hook.
1386                  * the source tree is a reloc tree, all tree blocks
1387                  * modified after it was created have RELOC flag
1388                  * set in their headers. so it's OK to not update
1389                  * the 'last_snapshot'.
1390                  */
1391                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1392                                       BTRFS_TREE_RELOC_OBJECTID);
1393                 BUG_ON(ret);
1394         }
1395
1396         memcpy(root_item, &root->root_item, sizeof(*root_item));
1397         btrfs_set_root_bytenr(root_item, eb->start);
1398         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1399         btrfs_set_root_generation(root_item, trans->transid);
1400
1401         if (root->root_key.objectid == objectid) {
1402                 btrfs_set_root_refs(root_item, 0);
1403                 memset(&root_item->drop_progress, 0,
1404                        sizeof(struct btrfs_disk_key));
1405                 root_item->drop_level = 0;
1406         }
1407
1408         btrfs_tree_unlock(eb);
1409         free_extent_buffer(eb);
1410
1411         ret = btrfs_insert_root(trans, fs_info->tree_root,
1412                                 &root_key, root_item);
1413         BUG_ON(ret);
1414         kfree(root_item);
1415
1416         reloc_root = btrfs_read_fs_root(fs_info->tree_root, &root_key);
1417         BUG_ON(IS_ERR(reloc_root));
1418         reloc_root->last_trans = trans->transid;
1419         return reloc_root;
1420 }
1421
1422 /*
1423  * create reloc tree for a given fs tree. reloc tree is just a
1424  * snapshot of the fs tree with special root objectid.
1425  */
1426 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1427                           struct btrfs_root *root)
1428 {
1429         struct btrfs_fs_info *fs_info = root->fs_info;
1430         struct btrfs_root *reloc_root;
1431         struct reloc_control *rc = fs_info->reloc_ctl;
1432         struct btrfs_block_rsv *rsv;
1433         int clear_rsv = 0;
1434         int ret;
1435
1436         if (root->reloc_root) {
1437                 reloc_root = root->reloc_root;
1438                 reloc_root->last_trans = trans->transid;
1439                 return 0;
1440         }
1441
1442         if (!rc || !rc->create_reloc_tree ||
1443             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1444                 return 0;
1445
1446         if (!trans->reloc_reserved) {
1447                 rsv = trans->block_rsv;
1448                 trans->block_rsv = rc->block_rsv;
1449                 clear_rsv = 1;
1450         }
1451         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1452         if (clear_rsv)
1453                 trans->block_rsv = rsv;
1454
1455         ret = __add_reloc_root(reloc_root);
1456         BUG_ON(ret < 0);
1457         root->reloc_root = reloc_root;
1458         return 0;
1459 }
1460
1461 /*
1462  * update root item of reloc tree
1463  */
1464 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1465                             struct btrfs_root *root)
1466 {
1467         struct btrfs_fs_info *fs_info = root->fs_info;
1468         struct btrfs_root *reloc_root;
1469         struct btrfs_root_item *root_item;
1470         int ret;
1471
1472         if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state) ||
1473             !root->reloc_root)
1474                 goto out;
1475
1476         reloc_root = root->reloc_root;
1477         root_item = &reloc_root->root_item;
1478
1479         /* root->reloc_root will stay until current relocation finished */
1480         if (fs_info->reloc_ctl->merge_reloc_tree &&
1481             btrfs_root_refs(root_item) == 0) {
1482                 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1483                 __del_reloc_root(reloc_root);
1484         }
1485
1486         if (reloc_root->commit_root != reloc_root->node) {
1487                 btrfs_set_root_node(root_item, reloc_root->node);
1488                 free_extent_buffer(reloc_root->commit_root);
1489                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1490         }
1491
1492         ret = btrfs_update_root(trans, fs_info->tree_root,
1493                                 &reloc_root->root_key, root_item);
1494         BUG_ON(ret);
1495
1496 out:
1497         return 0;
1498 }
1499
1500 /*
1501  * helper to find first cached inode with inode number >= objectid
1502  * in a subvolume
1503  */
1504 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1505 {
1506         struct rb_node *node;
1507         struct rb_node *prev;
1508         struct btrfs_inode *entry;
1509         struct inode *inode;
1510
1511         spin_lock(&root->inode_lock);
1512 again:
1513         node = root->inode_tree.rb_node;
1514         prev = NULL;
1515         while (node) {
1516                 prev = node;
1517                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1518
1519                 if (objectid < btrfs_ino(entry))
1520                         node = node->rb_left;
1521                 else if (objectid > btrfs_ino(entry))
1522                         node = node->rb_right;
1523                 else
1524                         break;
1525         }
1526         if (!node) {
1527                 while (prev) {
1528                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1529                         if (objectid <= btrfs_ino(entry)) {
1530                                 node = prev;
1531                                 break;
1532                         }
1533                         prev = rb_next(prev);
1534                 }
1535         }
1536         while (node) {
1537                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1538                 inode = igrab(&entry->vfs_inode);
1539                 if (inode) {
1540                         spin_unlock(&root->inode_lock);
1541                         return inode;
1542                 }
1543
1544                 objectid = btrfs_ino(entry) + 1;
1545                 if (cond_resched_lock(&root->inode_lock))
1546                         goto again;
1547
1548                 node = rb_next(node);
1549         }
1550         spin_unlock(&root->inode_lock);
1551         return NULL;
1552 }
1553
1554 static int in_block_group(u64 bytenr,
1555                           struct btrfs_block_group_cache *block_group)
1556 {
1557         if (bytenr >= block_group->key.objectid &&
1558             bytenr < block_group->key.objectid + block_group->key.offset)
1559                 return 1;
1560         return 0;
1561 }
1562
1563 /*
1564  * get new location of data
1565  */
1566 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1567                             u64 bytenr, u64 num_bytes)
1568 {
1569         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1570         struct btrfs_path *path;
1571         struct btrfs_file_extent_item *fi;
1572         struct extent_buffer *leaf;
1573         int ret;
1574
1575         path = btrfs_alloc_path();
1576         if (!path)
1577                 return -ENOMEM;
1578
1579         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1580         ret = btrfs_lookup_file_extent(NULL, root, path,
1581                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1582         if (ret < 0)
1583                 goto out;
1584         if (ret > 0) {
1585                 ret = -ENOENT;
1586                 goto out;
1587         }
1588
1589         leaf = path->nodes[0];
1590         fi = btrfs_item_ptr(leaf, path->slots[0],
1591                             struct btrfs_file_extent_item);
1592
1593         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1594                btrfs_file_extent_compression(leaf, fi) ||
1595                btrfs_file_extent_encryption(leaf, fi) ||
1596                btrfs_file_extent_other_encoding(leaf, fi));
1597
1598         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1599                 ret = -EINVAL;
1600                 goto out;
1601         }
1602
1603         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1604         ret = 0;
1605 out:
1606         btrfs_free_path(path);
1607         return ret;
1608 }
1609
1610 /*
1611  * update file extent items in the tree leaf to point to
1612  * the new locations.
1613  */
1614 static noinline_for_stack
1615 int replace_file_extents(struct btrfs_trans_handle *trans,
1616                          struct reloc_control *rc,
1617                          struct btrfs_root *root,
1618                          struct extent_buffer *leaf)
1619 {
1620         struct btrfs_fs_info *fs_info = root->fs_info;
1621         struct btrfs_key key;
1622         struct btrfs_file_extent_item *fi;
1623         struct inode *inode = NULL;
1624         u64 parent;
1625         u64 bytenr;
1626         u64 new_bytenr = 0;
1627         u64 num_bytes;
1628         u64 end;
1629         u32 nritems;
1630         u32 i;
1631         int ret = 0;
1632         int first = 1;
1633         int dirty = 0;
1634
1635         if (rc->stage != UPDATE_DATA_PTRS)
1636                 return 0;
1637
1638         /* reloc trees always use full backref */
1639         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1640                 parent = leaf->start;
1641         else
1642                 parent = 0;
1643
1644         nritems = btrfs_header_nritems(leaf);
1645         for (i = 0; i < nritems; i++) {
1646                 cond_resched();
1647                 btrfs_item_key_to_cpu(leaf, &key, i);
1648                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1649                         continue;
1650                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1651                 if (btrfs_file_extent_type(leaf, fi) ==
1652                     BTRFS_FILE_EXTENT_INLINE)
1653                         continue;
1654                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1655                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1656                 if (bytenr == 0)
1657                         continue;
1658                 if (!in_block_group(bytenr, rc->block_group))
1659                         continue;
1660
1661                 /*
1662                  * if we are modifying block in fs tree, wait for readpage
1663                  * to complete and drop the extent cache
1664                  */
1665                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1666                         if (first) {
1667                                 inode = find_next_inode(root, key.objectid);
1668                                 first = 0;
1669                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1670                                 btrfs_add_delayed_iput(inode);
1671                                 inode = find_next_inode(root, key.objectid);
1672                         }
1673                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1674                                 end = key.offset +
1675                                       btrfs_file_extent_num_bytes(leaf, fi);
1676                                 WARN_ON(!IS_ALIGNED(key.offset,
1677                                                     fs_info->sectorsize));
1678                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1679                                 end--;
1680                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1681                                                       key.offset, end);
1682                                 if (!ret)
1683                                         continue;
1684
1685                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1686                                                 key.offset,     end, 1);
1687                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1688                                               key.offset, end);
1689                         }
1690                 }
1691
1692                 ret = get_new_location(rc->data_inode, &new_bytenr,
1693                                        bytenr, num_bytes);
1694                 if (ret) {
1695                         /*
1696                          * Don't have to abort since we've not changed anything
1697                          * in the file extent yet.
1698                          */
1699                         break;
1700                 }
1701
1702                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1703                 dirty = 1;
1704
1705                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1706                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1707                                            num_bytes, parent,
1708                                            btrfs_header_owner(leaf),
1709                                            key.objectid, key.offset);
1710                 if (ret) {
1711                         btrfs_abort_transaction(trans, ret);
1712                         break;
1713                 }
1714
1715                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1716                                         parent, btrfs_header_owner(leaf),
1717                                         key.objectid, key.offset);
1718                 if (ret) {
1719                         btrfs_abort_transaction(trans, ret);
1720                         break;
1721                 }
1722         }
1723         if (dirty)
1724                 btrfs_mark_buffer_dirty(leaf);
1725         if (inode)
1726                 btrfs_add_delayed_iput(inode);
1727         return ret;
1728 }
1729
1730 static noinline_for_stack
1731 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1732                      struct btrfs_path *path, int level)
1733 {
1734         struct btrfs_disk_key key1;
1735         struct btrfs_disk_key key2;
1736         btrfs_node_key(eb, &key1, slot);
1737         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1738         return memcmp(&key1, &key2, sizeof(key1));
1739 }
1740
1741 /*
1742  * try to replace tree blocks in fs tree with the new blocks
1743  * in reloc tree. tree blocks haven't been modified since the
1744  * reloc tree was create can be replaced.
1745  *
1746  * if a block was replaced, level of the block + 1 is returned.
1747  * if no block got replaced, 0 is returned. if there are other
1748  * errors, a negative error number is returned.
1749  */
1750 static noinline_for_stack
1751 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1752                  struct btrfs_root *dest, struct btrfs_root *src,
1753                  struct btrfs_path *path, struct btrfs_key *next_key,
1754                  int lowest_level, int max_level)
1755 {
1756         struct btrfs_fs_info *fs_info = dest->fs_info;
1757         struct extent_buffer *eb;
1758         struct extent_buffer *parent;
1759         struct btrfs_key key;
1760         u64 old_bytenr;
1761         u64 new_bytenr;
1762         u64 old_ptr_gen;
1763         u64 new_ptr_gen;
1764         u64 last_snapshot;
1765         u32 blocksize;
1766         int cow = 0;
1767         int level;
1768         int ret;
1769         int slot;
1770
1771         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1772         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1773
1774         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1775 again:
1776         slot = path->slots[lowest_level];
1777         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1778
1779         eb = btrfs_lock_root_node(dest);
1780         btrfs_set_lock_blocking(eb);
1781         level = btrfs_header_level(eb);
1782
1783         if (level < lowest_level) {
1784                 btrfs_tree_unlock(eb);
1785                 free_extent_buffer(eb);
1786                 return 0;
1787         }
1788
1789         if (cow) {
1790                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1791                 BUG_ON(ret);
1792         }
1793         btrfs_set_lock_blocking(eb);
1794
1795         if (next_key) {
1796                 next_key->objectid = (u64)-1;
1797                 next_key->type = (u8)-1;
1798                 next_key->offset = (u64)-1;
1799         }
1800
1801         parent = eb;
1802         while (1) {
1803                 struct btrfs_key first_key;
1804
1805                 level = btrfs_header_level(parent);
1806                 BUG_ON(level < lowest_level);
1807
1808                 ret = btrfs_bin_search(parent, &key, level, &slot);
1809                 if (ret && slot > 0)
1810                         slot--;
1811
1812                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1813                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1814
1815                 old_bytenr = btrfs_node_blockptr(parent, slot);
1816                 blocksize = fs_info->nodesize;
1817                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1818                 btrfs_node_key_to_cpu(parent, &first_key, slot);
1819
1820                 if (level <= max_level) {
1821                         eb = path->nodes[level];
1822                         new_bytenr = btrfs_node_blockptr(eb,
1823                                                         path->slots[level]);
1824                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1825                                                         path->slots[level]);
1826                 } else {
1827                         new_bytenr = 0;
1828                         new_ptr_gen = 0;
1829                 }
1830
1831                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1832                         ret = level;
1833                         break;
1834                 }
1835
1836                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1837                     memcmp_node_keys(parent, slot, path, level)) {
1838                         if (level <= lowest_level) {
1839                                 ret = 0;
1840                                 break;
1841                         }
1842
1843                         eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen,
1844                                              level - 1, &first_key);
1845                         if (IS_ERR(eb)) {
1846                                 ret = PTR_ERR(eb);
1847                                 break;
1848                         } else if (!extent_buffer_uptodate(eb)) {
1849                                 ret = -EIO;
1850                                 free_extent_buffer(eb);
1851                                 break;
1852                         }
1853                         btrfs_tree_lock(eb);
1854                         if (cow) {
1855                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1856                                                       slot, &eb);
1857                                 BUG_ON(ret);
1858                         }
1859                         btrfs_set_lock_blocking(eb);
1860
1861                         btrfs_tree_unlock(parent);
1862                         free_extent_buffer(parent);
1863
1864                         parent = eb;
1865                         continue;
1866                 }
1867
1868                 if (!cow) {
1869                         btrfs_tree_unlock(parent);
1870                         free_extent_buffer(parent);
1871                         cow = 1;
1872                         goto again;
1873                 }
1874
1875                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1876                                       path->slots[level]);
1877                 btrfs_release_path(path);
1878
1879                 path->lowest_level = level;
1880                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1881                 path->lowest_level = 0;
1882                 BUG_ON(ret);
1883
1884                 /*
1885                  * Info qgroup to trace both subtrees.
1886                  *
1887                  * We must trace both trees.
1888                  * 1) Tree reloc subtree
1889                  *    If not traced, we will leak data numbers
1890                  * 2) Fs subtree
1891                  *    If not traced, we will double count old data
1892                  *    and tree block numbers, if current trans doesn't free
1893                  *    data reloc tree inode.
1894                  */
1895                 ret = btrfs_qgroup_trace_subtree_swap(trans, rc->block_group,
1896                                 parent, slot, path->nodes[level],
1897                                 path->slots[level], last_snapshot);
1898                 if (ret < 0)
1899                         break;
1900
1901                 /*
1902                  * swap blocks in fs tree and reloc tree.
1903                  */
1904                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1905                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1906                 btrfs_mark_buffer_dirty(parent);
1907
1908                 btrfs_set_node_blockptr(path->nodes[level],
1909                                         path->slots[level], old_bytenr);
1910                 btrfs_set_node_ptr_generation(path->nodes[level],
1911                                               path->slots[level], old_ptr_gen);
1912                 btrfs_mark_buffer_dirty(path->nodes[level]);
1913
1914                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr,
1915                                         blocksize, path->nodes[level]->start,
1916                                         src->root_key.objectid, level - 1, 0);
1917                 BUG_ON(ret);
1918                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr,
1919                                         blocksize, 0, dest->root_key.objectid,
1920                                         level - 1, 0);
1921                 BUG_ON(ret);
1922
1923                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1924                                         path->nodes[level]->start,
1925                                         src->root_key.objectid, level - 1, 0);
1926                 BUG_ON(ret);
1927
1928                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1929                                         0, dest->root_key.objectid, level - 1,
1930                                         0);
1931                 BUG_ON(ret);
1932
1933                 btrfs_unlock_up_safe(path, 0);
1934
1935                 ret = level;
1936                 break;
1937         }
1938         btrfs_tree_unlock(parent);
1939         free_extent_buffer(parent);
1940         return ret;
1941 }
1942
1943 /*
1944  * helper to find next relocated block in reloc tree
1945  */
1946 static noinline_for_stack
1947 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1948                        int *level)
1949 {
1950         struct extent_buffer *eb;
1951         int i;
1952         u64 last_snapshot;
1953         u32 nritems;
1954
1955         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1956
1957         for (i = 0; i < *level; i++) {
1958                 free_extent_buffer(path->nodes[i]);
1959                 path->nodes[i] = NULL;
1960         }
1961
1962         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1963                 eb = path->nodes[i];
1964                 nritems = btrfs_header_nritems(eb);
1965                 while (path->slots[i] + 1 < nritems) {
1966                         path->slots[i]++;
1967                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1968                             last_snapshot)
1969                                 continue;
1970
1971                         *level = i;
1972                         return 0;
1973                 }
1974                 free_extent_buffer(path->nodes[i]);
1975                 path->nodes[i] = NULL;
1976         }
1977         return 1;
1978 }
1979
1980 /*
1981  * walk down reloc tree to find relocated block of lowest level
1982  */
1983 static noinline_for_stack
1984 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1985                          int *level)
1986 {
1987         struct btrfs_fs_info *fs_info = root->fs_info;
1988         struct extent_buffer *eb = NULL;
1989         int i;
1990         u64 bytenr;
1991         u64 ptr_gen = 0;
1992         u64 last_snapshot;
1993         u32 nritems;
1994
1995         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1996
1997         for (i = *level; i > 0; i--) {
1998                 struct btrfs_key first_key;
1999
2000                 eb = path->nodes[i];
2001                 nritems = btrfs_header_nritems(eb);
2002                 while (path->slots[i] < nritems) {
2003                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2004                         if (ptr_gen > last_snapshot)
2005                                 break;
2006                         path->slots[i]++;
2007                 }
2008                 if (path->slots[i] >= nritems) {
2009                         if (i == *level)
2010                                 break;
2011                         *level = i + 1;
2012                         return 0;
2013                 }
2014                 if (i == 1) {
2015                         *level = i;
2016                         return 0;
2017                 }
2018
2019                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2020                 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
2021                 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
2022                                      &first_key);
2023                 if (IS_ERR(eb)) {
2024                         return PTR_ERR(eb);
2025                 } else if (!extent_buffer_uptodate(eb)) {
2026                         free_extent_buffer(eb);
2027                         return -EIO;
2028                 }
2029                 BUG_ON(btrfs_header_level(eb) != i - 1);
2030                 path->nodes[i - 1] = eb;
2031                 path->slots[i - 1] = 0;
2032         }
2033         return 1;
2034 }
2035
2036 /*
2037  * invalidate extent cache for file extents whose key in range of
2038  * [min_key, max_key)
2039  */
2040 static int invalidate_extent_cache(struct btrfs_root *root,
2041                                    struct btrfs_key *min_key,
2042                                    struct btrfs_key *max_key)
2043 {
2044         struct btrfs_fs_info *fs_info = root->fs_info;
2045         struct inode *inode = NULL;
2046         u64 objectid;
2047         u64 start, end;
2048         u64 ino;
2049
2050         objectid = min_key->objectid;
2051         while (1) {
2052                 cond_resched();
2053                 iput(inode);
2054
2055                 if (objectid > max_key->objectid)
2056                         break;
2057
2058                 inode = find_next_inode(root, objectid);
2059                 if (!inode)
2060                         break;
2061                 ino = btrfs_ino(BTRFS_I(inode));
2062
2063                 if (ino > max_key->objectid) {
2064                         iput(inode);
2065                         break;
2066                 }
2067
2068                 objectid = ino + 1;
2069                 if (!S_ISREG(inode->i_mode))
2070                         continue;
2071
2072                 if (unlikely(min_key->objectid == ino)) {
2073                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2074                                 continue;
2075                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2076                                 start = 0;
2077                         else {
2078                                 start = min_key->offset;
2079                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
2080                         }
2081                 } else {
2082                         start = 0;
2083                 }
2084
2085                 if (unlikely(max_key->objectid == ino)) {
2086                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2087                                 continue;
2088                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2089                                 end = (u64)-1;
2090                         } else {
2091                                 if (max_key->offset == 0)
2092                                         continue;
2093                                 end = max_key->offset;
2094                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
2095                                 end--;
2096                         }
2097                 } else {
2098                         end = (u64)-1;
2099                 }
2100
2101                 /* the lock_extent waits for readpage to complete */
2102                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2103                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
2104                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2105         }
2106         return 0;
2107 }
2108
2109 static int find_next_key(struct btrfs_path *path, int level,
2110                          struct btrfs_key *key)
2111
2112 {
2113         while (level < BTRFS_MAX_LEVEL) {
2114                 if (!path->nodes[level])
2115                         break;
2116                 if (path->slots[level] + 1 <
2117                     btrfs_header_nritems(path->nodes[level])) {
2118                         btrfs_node_key_to_cpu(path->nodes[level], key,
2119                                               path->slots[level] + 1);
2120                         return 0;
2121                 }
2122                 level++;
2123         }
2124         return 1;
2125 }
2126
2127 /*
2128  * Insert current subvolume into reloc_control::dirty_subvol_roots
2129  */
2130 static void insert_dirty_subvol(struct btrfs_trans_handle *trans,
2131                                 struct reloc_control *rc,
2132                                 struct btrfs_root *root)
2133 {
2134         struct btrfs_root *reloc_root = root->reloc_root;
2135         struct btrfs_root_item *reloc_root_item;
2136
2137         /* @root must be a subvolume tree root with a valid reloc tree */
2138         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
2139         ASSERT(reloc_root);
2140
2141         reloc_root_item = &reloc_root->root_item;
2142         memset(&reloc_root_item->drop_progress, 0,
2143                 sizeof(reloc_root_item->drop_progress));
2144         reloc_root_item->drop_level = 0;
2145         btrfs_set_root_refs(reloc_root_item, 0);
2146         btrfs_update_reloc_root(trans, root);
2147
2148         if (list_empty(&root->reloc_dirty_list)) {
2149                 btrfs_grab_fs_root(root);
2150                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
2151         }
2152 }
2153
2154 static int clean_dirty_subvols(struct reloc_control *rc)
2155 {
2156         struct btrfs_root *root;
2157         struct btrfs_root *next;
2158         int ret = 0;
2159
2160         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
2161                                  reloc_dirty_list) {
2162                 struct btrfs_root *reloc_root = root->reloc_root;
2163
2164                 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
2165                 list_del_init(&root->reloc_dirty_list);
2166                 root->reloc_root = NULL;
2167                 if (reloc_root) {
2168                         int ret2;
2169
2170                         ret2 = btrfs_drop_snapshot(reloc_root, NULL, 0, 1);
2171                         if (ret2 < 0 && !ret)
2172                                 ret = ret2;
2173                 }
2174                 btrfs_put_fs_root(root);
2175         }
2176         return ret;
2177 }
2178
2179 /*
2180  * merge the relocated tree blocks in reloc tree with corresponding
2181  * fs tree.
2182  */
2183 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2184                                                struct btrfs_root *root)
2185 {
2186         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2187         struct btrfs_key key;
2188         struct btrfs_key next_key;
2189         struct btrfs_trans_handle *trans = NULL;
2190         struct btrfs_root *reloc_root;
2191         struct btrfs_root_item *root_item;
2192         struct btrfs_path *path;
2193         struct extent_buffer *leaf;
2194         int level;
2195         int max_level;
2196         int replaced = 0;
2197         int ret;
2198         int err = 0;
2199         u32 min_reserved;
2200
2201         path = btrfs_alloc_path();
2202         if (!path)
2203                 return -ENOMEM;
2204         path->reada = READA_FORWARD;
2205
2206         reloc_root = root->reloc_root;
2207         root_item = &reloc_root->root_item;
2208
2209         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2210                 level = btrfs_root_level(root_item);
2211                 extent_buffer_get(reloc_root->node);
2212                 path->nodes[level] = reloc_root->node;
2213                 path->slots[level] = 0;
2214         } else {
2215                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2216
2217                 level = root_item->drop_level;
2218                 BUG_ON(level == 0);
2219                 path->lowest_level = level;
2220                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2221                 path->lowest_level = 0;
2222                 if (ret < 0) {
2223                         btrfs_free_path(path);
2224                         return ret;
2225                 }
2226
2227                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2228                                       path->slots[level]);
2229                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2230
2231                 btrfs_unlock_up_safe(path, 0);
2232         }
2233
2234         min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2235         memset(&next_key, 0, sizeof(next_key));
2236
2237         while (1) {
2238                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2239                                              BTRFS_RESERVE_FLUSH_ALL);
2240                 if (ret) {
2241                         err = ret;
2242                         goto out;
2243                 }
2244                 trans = btrfs_start_transaction(root, 0);
2245                 if (IS_ERR(trans)) {
2246                         err = PTR_ERR(trans);
2247                         trans = NULL;
2248                         goto out;
2249                 }
2250                 trans->block_rsv = rc->block_rsv;
2251
2252                 replaced = 0;
2253                 max_level = level;
2254
2255                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2256                 if (ret < 0) {
2257                         err = ret;
2258                         goto out;
2259                 }
2260                 if (ret > 0)
2261                         break;
2262
2263                 if (!find_next_key(path, level, &key) &&
2264                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2265                         ret = 0;
2266                 } else {
2267                         ret = replace_path(trans, rc, root, reloc_root, path,
2268                                            &next_key, level, max_level);
2269                 }
2270                 if (ret < 0) {
2271                         err = ret;
2272                         goto out;
2273                 }
2274
2275                 if (ret > 0) {
2276                         level = ret;
2277                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2278                                               path->slots[level]);
2279                         replaced = 1;
2280                 }
2281
2282                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2283                 if (ret > 0)
2284                         break;
2285
2286                 BUG_ON(level == 0);
2287                 /*
2288                  * save the merging progress in the drop_progress.
2289                  * this is OK since root refs == 1 in this case.
2290                  */
2291                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2292                                path->slots[level]);
2293                 root_item->drop_level = level;
2294
2295                 btrfs_end_transaction_throttle(trans);
2296                 trans = NULL;
2297
2298                 btrfs_btree_balance_dirty(fs_info);
2299
2300                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2301                         invalidate_extent_cache(root, &key, &next_key);
2302         }
2303
2304         /*
2305          * handle the case only one block in the fs tree need to be
2306          * relocated and the block is tree root.
2307          */
2308         leaf = btrfs_lock_root_node(root);
2309         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2310         btrfs_tree_unlock(leaf);
2311         free_extent_buffer(leaf);
2312         if (ret < 0)
2313                 err = ret;
2314 out:
2315         btrfs_free_path(path);
2316
2317         if (err == 0)
2318                 insert_dirty_subvol(trans, rc, root);
2319
2320         if (trans)
2321                 btrfs_end_transaction_throttle(trans);
2322
2323         btrfs_btree_balance_dirty(fs_info);
2324
2325         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2326                 invalidate_extent_cache(root, &key, &next_key);
2327
2328         return err;
2329 }
2330
2331 static noinline_for_stack
2332 int prepare_to_merge(struct reloc_control *rc, int err)
2333 {
2334         struct btrfs_root *root = rc->extent_root;
2335         struct btrfs_fs_info *fs_info = root->fs_info;
2336         struct btrfs_root *reloc_root;
2337         struct btrfs_trans_handle *trans;
2338         LIST_HEAD(reloc_roots);
2339         u64 num_bytes = 0;
2340         int ret;
2341
2342         mutex_lock(&fs_info->reloc_mutex);
2343         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2344         rc->merging_rsv_size += rc->nodes_relocated * 2;
2345         mutex_unlock(&fs_info->reloc_mutex);
2346
2347 again:
2348         if (!err) {
2349                 num_bytes = rc->merging_rsv_size;
2350                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2351                                           BTRFS_RESERVE_FLUSH_ALL);
2352                 if (ret)
2353                         err = ret;
2354         }
2355
2356         trans = btrfs_join_transaction(rc->extent_root);
2357         if (IS_ERR(trans)) {
2358                 if (!err)
2359                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2360                                                 num_bytes);
2361                 return PTR_ERR(trans);
2362         }
2363
2364         if (!err) {
2365                 if (num_bytes != rc->merging_rsv_size) {
2366                         btrfs_end_transaction(trans);
2367                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2368                                                 num_bytes);
2369                         goto again;
2370                 }
2371         }
2372
2373         rc->merge_reloc_tree = 1;
2374
2375         while (!list_empty(&rc->reloc_roots)) {
2376                 reloc_root = list_entry(rc->reloc_roots.next,
2377                                         struct btrfs_root, root_list);
2378                 list_del_init(&reloc_root->root_list);
2379
2380                 root = read_fs_root(fs_info, reloc_root->root_key.offset);
2381                 BUG_ON(IS_ERR(root));
2382                 BUG_ON(root->reloc_root != reloc_root);
2383
2384                 /*
2385                  * set reference count to 1, so btrfs_recover_relocation
2386                  * knows it should resumes merging
2387                  */
2388                 if (!err)
2389                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2390                 btrfs_update_reloc_root(trans, root);
2391
2392                 list_add(&reloc_root->root_list, &reloc_roots);
2393         }
2394
2395         list_splice(&reloc_roots, &rc->reloc_roots);
2396
2397         if (!err)
2398                 btrfs_commit_transaction(trans);
2399         else
2400                 btrfs_end_transaction(trans);
2401         return err;
2402 }
2403
2404 static noinline_for_stack
2405 void free_reloc_roots(struct list_head *list)
2406 {
2407         struct btrfs_root *reloc_root;
2408
2409         while (!list_empty(list)) {
2410                 reloc_root = list_entry(list->next, struct btrfs_root,
2411                                         root_list);
2412                 __del_reloc_root(reloc_root);
2413                 free_extent_buffer(reloc_root->node);
2414                 free_extent_buffer(reloc_root->commit_root);
2415                 reloc_root->node = NULL;
2416                 reloc_root->commit_root = NULL;
2417         }
2418 }
2419
2420 static noinline_for_stack
2421 void merge_reloc_roots(struct reloc_control *rc)
2422 {
2423         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2424         struct btrfs_root *root;
2425         struct btrfs_root *reloc_root;
2426         LIST_HEAD(reloc_roots);
2427         int found = 0;
2428         int ret = 0;
2429 again:
2430         root = rc->extent_root;
2431
2432         /*
2433          * this serializes us with btrfs_record_root_in_transaction,
2434          * we have to make sure nobody is in the middle of
2435          * adding their roots to the list while we are
2436          * doing this splice
2437          */
2438         mutex_lock(&fs_info->reloc_mutex);
2439         list_splice_init(&rc->reloc_roots, &reloc_roots);
2440         mutex_unlock(&fs_info->reloc_mutex);
2441
2442         while (!list_empty(&reloc_roots)) {
2443                 found = 1;
2444                 reloc_root = list_entry(reloc_roots.next,
2445                                         struct btrfs_root, root_list);
2446
2447                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2448                         root = read_fs_root(fs_info,
2449                                             reloc_root->root_key.offset);
2450                         BUG_ON(IS_ERR(root));
2451                         BUG_ON(root->reloc_root != reloc_root);
2452
2453                         ret = merge_reloc_root(rc, root);
2454                         if (ret) {
2455                                 if (list_empty(&reloc_root->root_list))
2456                                         list_add_tail(&reloc_root->root_list,
2457                                                       &reloc_roots);
2458                                 goto out;
2459                         }
2460                 } else {
2461                         list_del_init(&reloc_root->root_list);
2462                 }
2463         }
2464
2465         if (found) {
2466                 found = 0;
2467                 goto again;
2468         }
2469 out:
2470         if (ret) {
2471                 btrfs_handle_fs_error(fs_info, ret, NULL);
2472                 if (!list_empty(&reloc_roots))
2473                         free_reloc_roots(&reloc_roots);
2474
2475                 /* new reloc root may be added */
2476                 mutex_lock(&fs_info->reloc_mutex);
2477                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2478                 mutex_unlock(&fs_info->reloc_mutex);
2479                 if (!list_empty(&reloc_roots))
2480                         free_reloc_roots(&reloc_roots);
2481         }
2482
2483         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2484 }
2485
2486 static void free_block_list(struct rb_root *blocks)
2487 {
2488         struct tree_block *block;
2489         struct rb_node *rb_node;
2490         while ((rb_node = rb_first(blocks))) {
2491                 block = rb_entry(rb_node, struct tree_block, rb_node);
2492                 rb_erase(rb_node, blocks);
2493                 kfree(block);
2494         }
2495 }
2496
2497 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2498                                       struct btrfs_root *reloc_root)
2499 {
2500         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2501         struct btrfs_root *root;
2502
2503         if (reloc_root->last_trans == trans->transid)
2504                 return 0;
2505
2506         root = read_fs_root(fs_info, reloc_root->root_key.offset);
2507         BUG_ON(IS_ERR(root));
2508         BUG_ON(root->reloc_root != reloc_root);
2509
2510         return btrfs_record_root_in_trans(trans, root);
2511 }
2512
2513 static noinline_for_stack
2514 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2515                                      struct reloc_control *rc,
2516                                      struct backref_node *node,
2517                                      struct backref_edge *edges[])
2518 {
2519         struct backref_node *next;
2520         struct btrfs_root *root;
2521         int index = 0;
2522
2523         next = node;
2524         while (1) {
2525                 cond_resched();
2526                 next = walk_up_backref(next, edges, &index);
2527                 root = next->root;
2528                 BUG_ON(!root);
2529                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2530
2531                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2532                         record_reloc_root_in_trans(trans, root);
2533                         break;
2534                 }
2535
2536                 btrfs_record_root_in_trans(trans, root);
2537                 root = root->reloc_root;
2538
2539                 if (next->new_bytenr != root->node->start) {
2540                         BUG_ON(next->new_bytenr);
2541                         BUG_ON(!list_empty(&next->list));
2542                         next->new_bytenr = root->node->start;
2543                         next->root = root;
2544                         list_add_tail(&next->list,
2545                                       &rc->backref_cache.changed);
2546                         __mark_block_processed(rc, next);
2547                         break;
2548                 }
2549
2550                 WARN_ON(1);
2551                 root = NULL;
2552                 next = walk_down_backref(edges, &index);
2553                 if (!next || next->level <= node->level)
2554                         break;
2555         }
2556         if (!root)
2557                 return NULL;
2558
2559         next = node;
2560         /* setup backref node path for btrfs_reloc_cow_block */
2561         while (1) {
2562                 rc->backref_cache.path[next->level] = next;
2563                 if (--index < 0)
2564                         break;
2565                 next = edges[index]->node[UPPER];
2566         }
2567         return root;
2568 }
2569
2570 /*
2571  * select a tree root for relocation. return NULL if the block
2572  * is reference counted. we should use do_relocation() in this
2573  * case. return a tree root pointer if the block isn't reference
2574  * counted. return -ENOENT if the block is root of reloc tree.
2575  */
2576 static noinline_for_stack
2577 struct btrfs_root *select_one_root(struct backref_node *node)
2578 {
2579         struct backref_node *next;
2580         struct btrfs_root *root;
2581         struct btrfs_root *fs_root = NULL;
2582         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2583         int index = 0;
2584
2585         next = node;
2586         while (1) {
2587                 cond_resched();
2588                 next = walk_up_backref(next, edges, &index);
2589                 root = next->root;
2590                 BUG_ON(!root);
2591
2592                 /* no other choice for non-references counted tree */
2593                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2594                         return root;
2595
2596                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2597                         fs_root = root;
2598
2599                 if (next != node)
2600                         return NULL;
2601
2602                 next = walk_down_backref(edges, &index);
2603                 if (!next || next->level <= node->level)
2604                         break;
2605         }
2606
2607         if (!fs_root)
2608                 return ERR_PTR(-ENOENT);
2609         return fs_root;
2610 }
2611
2612 static noinline_for_stack
2613 u64 calcu_metadata_size(struct reloc_control *rc,
2614                         struct backref_node *node, int reserve)
2615 {
2616         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2617         struct backref_node *next = node;
2618         struct backref_edge *edge;
2619         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2620         u64 num_bytes = 0;
2621         int index = 0;
2622
2623         BUG_ON(reserve && node->processed);
2624
2625         while (next) {
2626                 cond_resched();
2627                 while (1) {
2628                         if (next->processed && (reserve || next != node))
2629                                 break;
2630
2631                         num_bytes += fs_info->nodesize;
2632
2633                         if (list_empty(&next->upper))
2634                                 break;
2635
2636                         edge = list_entry(next->upper.next,
2637                                           struct backref_edge, list[LOWER]);
2638                         edges[index++] = edge;
2639                         next = edge->node[UPPER];
2640                 }
2641                 next = walk_down_backref(edges, &index);
2642         }
2643         return num_bytes;
2644 }
2645
2646 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2647                                   struct reloc_control *rc,
2648                                   struct backref_node *node)
2649 {
2650         struct btrfs_root *root = rc->extent_root;
2651         struct btrfs_fs_info *fs_info = root->fs_info;
2652         u64 num_bytes;
2653         int ret;
2654         u64 tmp;
2655
2656         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2657
2658         trans->block_rsv = rc->block_rsv;
2659         rc->reserved_bytes += num_bytes;
2660
2661         /*
2662          * We are under a transaction here so we can only do limited flushing.
2663          * If we get an enospc just kick back -EAGAIN so we know to drop the
2664          * transaction and try to refill when we can flush all the things.
2665          */
2666         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2667                                 BTRFS_RESERVE_FLUSH_LIMIT);
2668         if (ret) {
2669                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2670                 while (tmp <= rc->reserved_bytes)
2671                         tmp <<= 1;
2672                 /*
2673                  * only one thread can access block_rsv at this point,
2674                  * so we don't need hold lock to protect block_rsv.
2675                  * we expand more reservation size here to allow enough
2676                  * space for relocation and we will return earlier in
2677                  * enospc case.
2678                  */
2679                 rc->block_rsv->size = tmp + fs_info->nodesize *
2680                                       RELOCATION_RESERVED_NODES;
2681                 return -EAGAIN;
2682         }
2683
2684         return 0;
2685 }
2686
2687 /*
2688  * relocate a block tree, and then update pointers in upper level
2689  * blocks that reference the block to point to the new location.
2690  *
2691  * if called by link_to_upper, the block has already been relocated.
2692  * in that case this function just updates pointers.
2693  */
2694 static int do_relocation(struct btrfs_trans_handle *trans,
2695                          struct reloc_control *rc,
2696                          struct backref_node *node,
2697                          struct btrfs_key *key,
2698                          struct btrfs_path *path, int lowest)
2699 {
2700         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2701         struct backref_node *upper;
2702         struct backref_edge *edge;
2703         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2704         struct btrfs_root *root;
2705         struct extent_buffer *eb;
2706         u32 blocksize;
2707         u64 bytenr;
2708         u64 generation;
2709         int slot;
2710         int ret;
2711         int err = 0;
2712
2713         BUG_ON(lowest && node->eb);
2714
2715         path->lowest_level = node->level + 1;
2716         rc->backref_cache.path[node->level] = node;
2717         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2718                 struct btrfs_key first_key;
2719
2720                 cond_resched();
2721
2722                 upper = edge->node[UPPER];
2723                 root = select_reloc_root(trans, rc, upper, edges);
2724                 BUG_ON(!root);
2725
2726                 if (upper->eb && !upper->locked) {
2727                         if (!lowest) {
2728                                 ret = btrfs_bin_search(upper->eb, key,
2729                                                        upper->level, &slot);
2730                                 BUG_ON(ret);
2731                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2732                                 if (node->eb->start == bytenr)
2733                                         goto next;
2734                         }
2735                         drop_node_buffer(upper);
2736                 }
2737
2738                 if (!upper->eb) {
2739                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2740                         if (ret) {
2741                                 if (ret < 0)
2742                                         err = ret;
2743                                 else
2744                                         err = -ENOENT;
2745
2746                                 btrfs_release_path(path);
2747                                 break;
2748                         }
2749
2750                         if (!upper->eb) {
2751                                 upper->eb = path->nodes[upper->level];
2752                                 path->nodes[upper->level] = NULL;
2753                         } else {
2754                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2755                         }
2756
2757                         upper->locked = 1;
2758                         path->locks[upper->level] = 0;
2759
2760                         slot = path->slots[upper->level];
2761                         btrfs_release_path(path);
2762                 } else {
2763                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2764                                                &slot);
2765                         BUG_ON(ret);
2766                 }
2767
2768                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2769                 if (lowest) {
2770                         if (bytenr != node->bytenr) {
2771                                 btrfs_err(root->fs_info,
2772                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2773                                           bytenr, node->bytenr, slot,
2774                                           upper->eb->start);
2775                                 err = -EIO;
2776                                 goto next;
2777                         }
2778                 } else {
2779                         if (node->eb->start == bytenr)
2780                                 goto next;
2781                 }
2782
2783                 blocksize = root->fs_info->nodesize;
2784                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2785                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2786                 eb = read_tree_block(fs_info, bytenr, generation,
2787                                      upper->level - 1, &first_key);
2788                 if (IS_ERR(eb)) {
2789                         err = PTR_ERR(eb);
2790                         goto next;
2791                 } else if (!extent_buffer_uptodate(eb)) {
2792                         free_extent_buffer(eb);
2793                         err = -EIO;
2794                         goto next;
2795                 }
2796                 btrfs_tree_lock(eb);
2797                 btrfs_set_lock_blocking(eb);
2798
2799                 if (!node->eb) {
2800                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2801                                               slot, &eb);
2802                         btrfs_tree_unlock(eb);
2803                         free_extent_buffer(eb);
2804                         if (ret < 0) {
2805                                 err = ret;
2806                                 goto next;
2807                         }
2808                         BUG_ON(node->eb != eb);
2809                 } else {
2810                         btrfs_set_node_blockptr(upper->eb, slot,
2811                                                 node->eb->start);
2812                         btrfs_set_node_ptr_generation(upper->eb, slot,
2813                                                       trans->transid);
2814                         btrfs_mark_buffer_dirty(upper->eb);
2815
2816                         ret = btrfs_inc_extent_ref(trans, root,
2817                                                 node->eb->start, blocksize,
2818                                                 upper->eb->start,
2819                                                 btrfs_header_owner(upper->eb),
2820                                                 node->level, 0);
2821                         BUG_ON(ret);
2822
2823                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2824                         BUG_ON(ret);
2825                 }
2826 next:
2827                 if (!upper->pending)
2828                         drop_node_buffer(upper);
2829                 else
2830                         unlock_node_buffer(upper);
2831                 if (err)
2832                         break;
2833         }
2834
2835         if (!err && node->pending) {
2836                 drop_node_buffer(node);
2837                 list_move_tail(&node->list, &rc->backref_cache.changed);
2838                 node->pending = 0;
2839         }
2840
2841         path->lowest_level = 0;
2842         BUG_ON(err == -ENOSPC);
2843         return err;
2844 }
2845
2846 static int link_to_upper(struct btrfs_trans_handle *trans,
2847                          struct reloc_control *rc,
2848                          struct backref_node *node,
2849                          struct btrfs_path *path)
2850 {
2851         struct btrfs_key key;
2852
2853         btrfs_node_key_to_cpu(node->eb, &key, 0);
2854         return do_relocation(trans, rc, node, &key, path, 0);
2855 }
2856
2857 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2858                                 struct reloc_control *rc,
2859                                 struct btrfs_path *path, int err)
2860 {
2861         LIST_HEAD(list);
2862         struct backref_cache *cache = &rc->backref_cache;
2863         struct backref_node *node;
2864         int level;
2865         int ret;
2866
2867         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2868                 while (!list_empty(&cache->pending[level])) {
2869                         node = list_entry(cache->pending[level].next,
2870                                           struct backref_node, list);
2871                         list_move_tail(&node->list, &list);
2872                         BUG_ON(!node->pending);
2873
2874                         if (!err) {
2875                                 ret = link_to_upper(trans, rc, node, path);
2876                                 if (ret < 0)
2877                                         err = ret;
2878                         }
2879                 }
2880                 list_splice_init(&list, &cache->pending[level]);
2881         }
2882         return err;
2883 }
2884
2885 static void mark_block_processed(struct reloc_control *rc,
2886                                  u64 bytenr, u32 blocksize)
2887 {
2888         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2889                         EXTENT_DIRTY);
2890 }
2891
2892 static void __mark_block_processed(struct reloc_control *rc,
2893                                    struct backref_node *node)
2894 {
2895         u32 blocksize;
2896         if (node->level == 0 ||
2897             in_block_group(node->bytenr, rc->block_group)) {
2898                 blocksize = rc->extent_root->fs_info->nodesize;
2899                 mark_block_processed(rc, node->bytenr, blocksize);
2900         }
2901         node->processed = 1;
2902 }
2903
2904 /*
2905  * mark a block and all blocks directly/indirectly reference the block
2906  * as processed.
2907  */
2908 static void update_processed_blocks(struct reloc_control *rc,
2909                                     struct backref_node *node)
2910 {
2911         struct backref_node *next = node;
2912         struct backref_edge *edge;
2913         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2914         int index = 0;
2915
2916         while (next) {
2917                 cond_resched();
2918                 while (1) {
2919                         if (next->processed)
2920                                 break;
2921
2922                         __mark_block_processed(rc, next);
2923
2924                         if (list_empty(&next->upper))
2925                                 break;
2926
2927                         edge = list_entry(next->upper.next,
2928                                           struct backref_edge, list[LOWER]);
2929                         edges[index++] = edge;
2930                         next = edge->node[UPPER];
2931                 }
2932                 next = walk_down_backref(edges, &index);
2933         }
2934 }
2935
2936 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2937 {
2938         u32 blocksize = rc->extent_root->fs_info->nodesize;
2939
2940         if (test_range_bit(&rc->processed_blocks, bytenr,
2941                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2942                 return 1;
2943         return 0;
2944 }
2945
2946 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2947                               struct tree_block *block)
2948 {
2949         struct extent_buffer *eb;
2950
2951         BUG_ON(block->key_ready);
2952         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2953                              block->level, NULL);
2954         if (IS_ERR(eb)) {
2955                 return PTR_ERR(eb);
2956         } else if (!extent_buffer_uptodate(eb)) {
2957                 free_extent_buffer(eb);
2958                 return -EIO;
2959         }
2960         if (block->level == 0)
2961                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2962         else
2963                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2964         free_extent_buffer(eb);
2965         block->key_ready = 1;
2966         return 0;
2967 }
2968
2969 /*
2970  * helper function to relocate a tree block
2971  */
2972 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2973                                 struct reloc_control *rc,
2974                                 struct backref_node *node,
2975                                 struct btrfs_key *key,
2976                                 struct btrfs_path *path)
2977 {
2978         struct btrfs_root *root;
2979         int ret = 0;
2980
2981         if (!node)
2982                 return 0;
2983
2984         BUG_ON(node->processed);
2985         root = select_one_root(node);
2986         if (root == ERR_PTR(-ENOENT)) {
2987                 update_processed_blocks(rc, node);
2988                 goto out;
2989         }
2990
2991         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2992                 ret = reserve_metadata_space(trans, rc, node);
2993                 if (ret)
2994                         goto out;
2995         }
2996
2997         if (root) {
2998                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2999                         BUG_ON(node->new_bytenr);
3000                         BUG_ON(!list_empty(&node->list));
3001                         btrfs_record_root_in_trans(trans, root);
3002                         root = root->reloc_root;
3003                         node->new_bytenr = root->node->start;
3004                         node->root = root;
3005                         list_add_tail(&node->list, &rc->backref_cache.changed);
3006                 } else {
3007                         path->lowest_level = node->level;
3008                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
3009                         btrfs_release_path(path);
3010                         if (ret > 0)
3011                                 ret = 0;
3012                 }
3013                 if (!ret)
3014                         update_processed_blocks(rc, node);
3015         } else {
3016                 ret = do_relocation(trans, rc, node, key, path, 1);
3017         }
3018 out:
3019         if (ret || node->level == 0 || node->cowonly)
3020                 remove_backref_node(&rc->backref_cache, node);
3021         return ret;
3022 }
3023
3024 /*
3025  * relocate a list of blocks
3026  */
3027 static noinline_for_stack
3028 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
3029                          struct reloc_control *rc, struct rb_root *blocks)
3030 {
3031         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3032         struct backref_node *node;
3033         struct btrfs_path *path;
3034         struct tree_block *block;
3035         struct tree_block *next;
3036         int ret;
3037         int err = 0;
3038
3039         path = btrfs_alloc_path();
3040         if (!path) {
3041                 err = -ENOMEM;
3042                 goto out_free_blocks;
3043         }
3044
3045         /* Kick in readahead for tree blocks with missing keys */
3046         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3047                 if (!block->key_ready)
3048                         readahead_tree_block(fs_info, block->bytenr);
3049         }
3050
3051         /* Get first keys */
3052         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3053                 if (!block->key_ready) {
3054                         err = get_tree_block_key(fs_info, block);
3055                         if (err)
3056                                 goto out_free_path;
3057                 }
3058         }
3059
3060         /* Do tree relocation */
3061         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3062                 node = build_backref_tree(rc, &block->key,
3063                                           block->level, block->bytenr);
3064                 if (IS_ERR(node)) {
3065                         err = PTR_ERR(node);
3066                         goto out;
3067                 }
3068
3069                 ret = relocate_tree_block(trans, rc, node, &block->key,
3070                                           path);
3071                 if (ret < 0) {
3072                         if (ret != -EAGAIN || &block->rb_node == rb_first(blocks))
3073                                 err = ret;
3074                         goto out;
3075                 }
3076         }
3077 out:
3078         err = finish_pending_nodes(trans, rc, path, err);
3079
3080 out_free_path:
3081         btrfs_free_path(path);
3082 out_free_blocks:
3083         free_block_list(blocks);
3084         return err;
3085 }
3086
3087 static noinline_for_stack
3088 int prealloc_file_extent_cluster(struct inode *inode,
3089                                  struct file_extent_cluster *cluster)
3090 {
3091         u64 alloc_hint = 0;
3092         u64 start;
3093         u64 end;
3094         u64 offset = BTRFS_I(inode)->index_cnt;
3095         u64 num_bytes;
3096         int nr = 0;
3097         int ret = 0;
3098         u64 prealloc_start = cluster->start - offset;
3099         u64 prealloc_end = cluster->end - offset;
3100         u64 cur_offset;
3101         struct extent_changeset *data_reserved = NULL;
3102
3103         BUG_ON(cluster->start != cluster->boundary[0]);
3104         inode_lock(inode);
3105
3106         ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
3107                                           prealloc_end + 1 - prealloc_start);
3108         if (ret)
3109                 goto out;
3110
3111         cur_offset = prealloc_start;
3112         while (nr < cluster->nr) {
3113                 start = cluster->boundary[nr] - offset;
3114                 if (nr + 1 < cluster->nr)
3115                         end = cluster->boundary[nr + 1] - 1 - offset;
3116                 else
3117                         end = cluster->end - offset;
3118
3119                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3120                 num_bytes = end + 1 - start;
3121                 if (cur_offset < start)
3122                         btrfs_free_reserved_data_space(inode, data_reserved,
3123                                         cur_offset, start - cur_offset);
3124                 ret = btrfs_prealloc_file_range(inode, 0, start,
3125                                                 num_bytes, num_bytes,
3126                                                 end + 1, &alloc_hint);
3127                 cur_offset = end + 1;
3128                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3129                 if (ret)
3130                         break;
3131                 nr++;
3132         }
3133         if (cur_offset < prealloc_end)
3134                 btrfs_free_reserved_data_space(inode, data_reserved,
3135                                 cur_offset, prealloc_end + 1 - cur_offset);
3136 out:
3137         inode_unlock(inode);
3138         extent_changeset_free(data_reserved);
3139         return ret;
3140 }
3141
3142 static noinline_for_stack
3143 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3144                          u64 block_start)
3145 {
3146         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3147         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3148         struct extent_map *em;
3149         int ret = 0;
3150
3151         em = alloc_extent_map();
3152         if (!em)
3153                 return -ENOMEM;
3154
3155         em->start = start;
3156         em->len = end + 1 - start;
3157         em->block_len = em->len;
3158         em->block_start = block_start;
3159         em->bdev = fs_info->fs_devices->latest_bdev;
3160         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3161
3162         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3163         while (1) {
3164                 write_lock(&em_tree->lock);
3165                 ret = add_extent_mapping(em_tree, em, 0);
3166                 write_unlock(&em_tree->lock);
3167                 if (ret != -EEXIST) {
3168                         free_extent_map(em);
3169                         break;
3170                 }
3171                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3172         }
3173         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3174         return ret;
3175 }
3176
3177 static int relocate_file_extent_cluster(struct inode *inode,
3178                                         struct file_extent_cluster *cluster)
3179 {
3180         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3181         u64 page_start;
3182         u64 page_end;
3183         u64 offset = BTRFS_I(inode)->index_cnt;
3184         unsigned long index;
3185         unsigned long last_index;
3186         struct page *page;
3187         struct file_ra_state *ra;
3188         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3189         int nr = 0;
3190         int ret = 0;
3191
3192         if (!cluster->nr)
3193                 return 0;
3194
3195         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3196         if (!ra)
3197                 return -ENOMEM;
3198
3199         ret = prealloc_file_extent_cluster(inode, cluster);
3200         if (ret)
3201                 goto out;
3202
3203         file_ra_state_init(ra, inode->i_mapping);
3204
3205         ret = setup_extent_mapping(inode, cluster->start - offset,
3206                                    cluster->end - offset, cluster->start);
3207         if (ret)
3208                 goto out;
3209
3210         index = (cluster->start - offset) >> PAGE_SHIFT;
3211         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3212         while (index <= last_index) {
3213                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3214                                 PAGE_SIZE);
3215                 if (ret)
3216                         goto out;
3217
3218                 page = find_lock_page(inode->i_mapping, index);
3219                 if (!page) {
3220                         page_cache_sync_readahead(inode->i_mapping,
3221                                                   ra, NULL, index,
3222                                                   last_index + 1 - index);
3223                         page = find_or_create_page(inode->i_mapping, index,
3224                                                    mask);
3225                         if (!page) {
3226                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3227                                                         PAGE_SIZE, true);
3228                                 ret = -ENOMEM;
3229                                 goto out;
3230                         }
3231                 }
3232
3233                 if (PageReadahead(page)) {
3234                         page_cache_async_readahead(inode->i_mapping,
3235                                                    ra, NULL, page, index,
3236                                                    last_index + 1 - index);
3237                 }
3238
3239                 if (!PageUptodate(page)) {
3240                         btrfs_readpage(NULL, page);
3241                         lock_page(page);
3242                         if (!PageUptodate(page)) {
3243                                 unlock_page(page);
3244                                 put_page(page);
3245                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3246                                                         PAGE_SIZE, true);
3247                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
3248                                                                PAGE_SIZE, true);
3249                                 ret = -EIO;
3250                                 goto out;
3251                         }
3252                 }
3253
3254                 page_start = page_offset(page);
3255                 page_end = page_start + PAGE_SIZE - 1;
3256
3257                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3258
3259                 set_page_extent_mapped(page);
3260
3261                 if (nr < cluster->nr &&
3262                     page_start + offset == cluster->boundary[nr]) {
3263                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3264                                         page_start, page_end,
3265                                         EXTENT_BOUNDARY);
3266                         nr++;
3267                 }
3268
3269                 ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
3270                                                 NULL, 0);
3271                 if (ret) {
3272                         unlock_page(page);
3273                         put_page(page);
3274                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3275                                                          PAGE_SIZE, true);
3276                         btrfs_delalloc_release_extents(BTRFS_I(inode),
3277                                                        PAGE_SIZE, true);
3278
3279                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
3280                                           page_start, page_end,
3281                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
3282                         goto out;
3283
3284                 }
3285                 set_page_dirty(page);
3286
3287                 unlock_extent(&BTRFS_I(inode)->io_tree,
3288                               page_start, page_end);
3289                 unlock_page(page);
3290                 put_page(page);
3291
3292                 index++;
3293                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE,
3294                                                false);
3295                 balance_dirty_pages_ratelimited(inode->i_mapping);
3296                 btrfs_throttle(fs_info);
3297         }
3298         WARN_ON(nr != cluster->nr);
3299 out:
3300         kfree(ra);
3301         return ret;
3302 }
3303
3304 static noinline_for_stack
3305 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3306                          struct file_extent_cluster *cluster)
3307 {
3308         int ret;
3309
3310         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3311                 ret = relocate_file_extent_cluster(inode, cluster);
3312                 if (ret)
3313                         return ret;
3314                 cluster->nr = 0;
3315         }
3316
3317         if (!cluster->nr)
3318                 cluster->start = extent_key->objectid;
3319         else
3320                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3321         cluster->end = extent_key->objectid + extent_key->offset - 1;
3322         cluster->boundary[cluster->nr] = extent_key->objectid;
3323         cluster->nr++;
3324
3325         if (cluster->nr >= MAX_EXTENTS) {
3326                 ret = relocate_file_extent_cluster(inode, cluster);
3327                 if (ret)
3328                         return ret;
3329                 cluster->nr = 0;
3330         }
3331         return 0;
3332 }
3333
3334 /*
3335  * helper to add a tree block to the list.
3336  * the major work is getting the generation and level of the block
3337  */
3338 static int add_tree_block(struct reloc_control *rc,
3339                           struct btrfs_key *extent_key,
3340                           struct btrfs_path *path,
3341                           struct rb_root *blocks)
3342 {
3343         struct extent_buffer *eb;
3344         struct btrfs_extent_item *ei;
3345         struct btrfs_tree_block_info *bi;
3346         struct tree_block *block;
3347         struct rb_node *rb_node;
3348         u32 item_size;
3349         int level = -1;
3350         u64 generation;
3351
3352         eb =  path->nodes[0];
3353         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3354
3355         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3356             item_size >= sizeof(*ei) + sizeof(*bi)) {
3357                 ei = btrfs_item_ptr(eb, path->slots[0],
3358                                 struct btrfs_extent_item);
3359                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3360                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3361                         level = btrfs_tree_block_level(eb, bi);
3362                 } else {
3363                         level = (int)extent_key->offset;
3364                 }
3365                 generation = btrfs_extent_generation(eb, ei);
3366         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3367                 btrfs_print_v0_err(eb->fs_info);
3368                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3369                 return -EINVAL;
3370         } else {
3371                 BUG();
3372         }
3373
3374         btrfs_release_path(path);
3375
3376         BUG_ON(level == -1);
3377
3378         block = kmalloc(sizeof(*block), GFP_NOFS);
3379         if (!block)
3380                 return -ENOMEM;
3381
3382         block->bytenr = extent_key->objectid;
3383         block->key.objectid = rc->extent_root->fs_info->nodesize;
3384         block->key.offset = generation;
3385         block->level = level;
3386         block->key_ready = 0;
3387
3388         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3389         if (rb_node)
3390                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3391
3392         return 0;
3393 }
3394
3395 /*
3396  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3397  */
3398 static int __add_tree_block(struct reloc_control *rc,
3399                             u64 bytenr, u32 blocksize,
3400                             struct rb_root *blocks)
3401 {
3402         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3403         struct btrfs_path *path;
3404         struct btrfs_key key;
3405         int ret;
3406         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3407
3408         if (tree_block_processed(bytenr, rc))
3409                 return 0;
3410
3411         if (tree_search(blocks, bytenr))
3412                 return 0;
3413
3414         path = btrfs_alloc_path();
3415         if (!path)
3416                 return -ENOMEM;
3417 again:
3418         key.objectid = bytenr;
3419         if (skinny) {
3420                 key.type = BTRFS_METADATA_ITEM_KEY;
3421                 key.offset = (u64)-1;
3422         } else {
3423                 key.type = BTRFS_EXTENT_ITEM_KEY;
3424                 key.offset = blocksize;
3425         }
3426
3427         path->search_commit_root = 1;
3428         path->skip_locking = 1;
3429         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3430         if (ret < 0)
3431                 goto out;
3432
3433         if (ret > 0 && skinny) {
3434                 if (path->slots[0]) {
3435                         path->slots[0]--;
3436                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3437                                               path->slots[0]);
3438                         if (key.objectid == bytenr &&
3439                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3440                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3441                               key.offset == blocksize)))
3442                                 ret = 0;
3443                 }
3444
3445                 if (ret) {
3446                         skinny = false;
3447                         btrfs_release_path(path);
3448                         goto again;
3449                 }
3450         }
3451         if (ret) {
3452                 ASSERT(ret == 1);
3453                 btrfs_print_leaf(path->nodes[0]);
3454                 btrfs_err(fs_info,
3455              "tree block extent item (%llu) is not found in extent tree",
3456                      bytenr);
3457                 WARN_ON(1);
3458                 ret = -EINVAL;
3459                 goto out;
3460         }
3461
3462         ret = add_tree_block(rc, &key, path, blocks);
3463 out:
3464         btrfs_free_path(path);
3465         return ret;
3466 }
3467
3468 /*
3469  * helper to check if the block use full backrefs for pointers in it
3470  */
3471 static int block_use_full_backref(struct reloc_control *rc,
3472                                   struct extent_buffer *eb)
3473 {
3474         u64 flags;
3475         int ret;
3476
3477         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3478             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3479                 return 1;
3480
3481         ret = btrfs_lookup_extent_info(NULL, rc->extent_root->fs_info,
3482                                        eb->start, btrfs_header_level(eb), 1,
3483                                        NULL, &flags);
3484         BUG_ON(ret);
3485
3486         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3487                 ret = 1;
3488         else
3489                 ret = 0;
3490         return ret;
3491 }
3492
3493 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3494                                     struct btrfs_block_group_cache *block_group,
3495                                     struct inode *inode,
3496                                     u64 ino)
3497 {
3498         struct btrfs_key key;
3499         struct btrfs_root *root = fs_info->tree_root;
3500         struct btrfs_trans_handle *trans;
3501         int ret = 0;
3502
3503         if (inode)
3504                 goto truncate;
3505
3506         key.objectid = ino;
3507         key.type = BTRFS_INODE_ITEM_KEY;
3508         key.offset = 0;
3509
3510         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3511         if (IS_ERR(inode))
3512                 return -ENOENT;
3513
3514 truncate:
3515         ret = btrfs_check_trunc_cache_free_space(fs_info,
3516                                                  &fs_info->global_block_rsv);
3517         if (ret)
3518                 goto out;
3519
3520         trans = btrfs_join_transaction(root);
3521         if (IS_ERR(trans)) {
3522                 ret = PTR_ERR(trans);
3523                 goto out;
3524         }
3525
3526         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3527
3528         btrfs_end_transaction(trans);
3529         btrfs_btree_balance_dirty(fs_info);
3530 out:
3531         iput(inode);
3532         return ret;
3533 }
3534
3535 /*
3536  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3537  * this function scans fs tree to find blocks reference the data extent
3538  */
3539 static int find_data_references(struct reloc_control *rc,
3540                                 struct btrfs_key *extent_key,
3541                                 struct extent_buffer *leaf,
3542                                 struct btrfs_extent_data_ref *ref,
3543                                 struct rb_root *blocks)
3544 {
3545         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3546         struct btrfs_path *path;
3547         struct tree_block *block;
3548         struct btrfs_root *root;
3549         struct btrfs_file_extent_item *fi;
3550         struct rb_node *rb_node;
3551         struct btrfs_key key;
3552         u64 ref_root;
3553         u64 ref_objectid;
3554         u64 ref_offset;
3555         u32 ref_count;
3556         u32 nritems;
3557         int err = 0;
3558         int added = 0;
3559         int counted;
3560         int ret;
3561
3562         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3563         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3564         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3565         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3566
3567         /*
3568          * This is an extent belonging to the free space cache, lets just delete
3569          * it and redo the search.
3570          */
3571         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3572                 ret = delete_block_group_cache(fs_info, rc->block_group,
3573                                                NULL, ref_objectid);
3574                 if (ret != -ENOENT)
3575                         return ret;
3576                 ret = 0;
3577         }
3578
3579         path = btrfs_alloc_path();
3580         if (!path)
3581                 return -ENOMEM;
3582         path->reada = READA_FORWARD;
3583
3584         root = read_fs_root(fs_info, ref_root);
3585         if (IS_ERR(root)) {
3586                 err = PTR_ERR(root);
3587                 goto out;
3588         }
3589
3590         key.objectid = ref_objectid;
3591         key.type = BTRFS_EXTENT_DATA_KEY;
3592         if (ref_offset > ((u64)-1 << 32))
3593                 key.offset = 0;
3594         else
3595                 key.offset = ref_offset;
3596
3597         path->search_commit_root = 1;
3598         path->skip_locking = 1;
3599         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3600         if (ret < 0) {
3601                 err = ret;
3602                 goto out;
3603         }
3604
3605         leaf = path->nodes[0];
3606         nritems = btrfs_header_nritems(leaf);
3607         /*
3608          * the references in tree blocks that use full backrefs
3609          * are not counted in
3610          */
3611         if (block_use_full_backref(rc, leaf))
3612                 counted = 0;
3613         else
3614                 counted = 1;
3615         rb_node = tree_search(blocks, leaf->start);
3616         if (rb_node) {
3617                 if (counted)
3618                         added = 1;
3619                 else
3620                         path->slots[0] = nritems;
3621         }
3622
3623         while (ref_count > 0) {
3624                 while (path->slots[0] >= nritems) {
3625                         ret = btrfs_next_leaf(root, path);
3626                         if (ret < 0) {
3627                                 err = ret;
3628                                 goto out;
3629                         }
3630                         if (WARN_ON(ret > 0))
3631                                 goto out;
3632
3633                         leaf = path->nodes[0];
3634                         nritems = btrfs_header_nritems(leaf);
3635                         added = 0;
3636
3637                         if (block_use_full_backref(rc, leaf))
3638                                 counted = 0;
3639                         else
3640                                 counted = 1;
3641                         rb_node = tree_search(blocks, leaf->start);
3642                         if (rb_node) {
3643                                 if (counted)
3644                                         added = 1;
3645                                 else
3646                                         path->slots[0] = nritems;
3647                         }
3648                 }
3649
3650                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3651                 if (WARN_ON(key.objectid != ref_objectid ||
3652                     key.type != BTRFS_EXTENT_DATA_KEY))
3653                         break;
3654
3655                 fi = btrfs_item_ptr(leaf, path->slots[0],
3656                                     struct btrfs_file_extent_item);
3657
3658                 if (btrfs_file_extent_type(leaf, fi) ==
3659                     BTRFS_FILE_EXTENT_INLINE)
3660                         goto next;
3661
3662                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3663                     extent_key->objectid)
3664                         goto next;
3665
3666                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3667                 if (key.offset != ref_offset)
3668                         goto next;
3669
3670                 if (counted)
3671                         ref_count--;
3672                 if (added)
3673                         goto next;
3674
3675                 if (!tree_block_processed(leaf->start, rc)) {
3676                         block = kmalloc(sizeof(*block), GFP_NOFS);
3677                         if (!block) {
3678                                 err = -ENOMEM;
3679                                 break;
3680                         }
3681                         block->bytenr = leaf->start;
3682                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3683                         block->level = 0;
3684                         block->key_ready = 1;
3685                         rb_node = tree_insert(blocks, block->bytenr,
3686                                               &block->rb_node);
3687                         if (rb_node)
3688                                 backref_tree_panic(rb_node, -EEXIST,
3689                                                    block->bytenr);
3690                 }
3691                 if (counted)
3692                         added = 1;
3693                 else
3694                         path->slots[0] = nritems;
3695 next:
3696                 path->slots[0]++;
3697
3698         }
3699 out:
3700         btrfs_free_path(path);
3701         return err;
3702 }
3703
3704 /*
3705  * helper to find all tree blocks that reference a given data extent
3706  */
3707 static noinline_for_stack
3708 int add_data_references(struct reloc_control *rc,
3709                         struct btrfs_key *extent_key,
3710                         struct btrfs_path *path,
3711                         struct rb_root *blocks)
3712 {
3713         struct btrfs_key key;
3714         struct extent_buffer *eb;
3715         struct btrfs_extent_data_ref *dref;
3716         struct btrfs_extent_inline_ref *iref;
3717         unsigned long ptr;
3718         unsigned long end;
3719         u32 blocksize = rc->extent_root->fs_info->nodesize;
3720         int ret = 0;
3721         int err = 0;
3722
3723         eb = path->nodes[0];
3724         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3725         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3726         ptr += sizeof(struct btrfs_extent_item);
3727
3728         while (ptr < end) {
3729                 iref = (struct btrfs_extent_inline_ref *)ptr;
3730                 key.type = btrfs_get_extent_inline_ref_type(eb, iref,
3731                                                         BTRFS_REF_TYPE_DATA);
3732                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3733                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3734                         ret = __add_tree_block(rc, key.offset, blocksize,
3735                                                blocks);
3736                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3737                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3738                         ret = find_data_references(rc, extent_key,
3739                                                    eb, dref, blocks);
3740                 } else {
3741                         ret = -EUCLEAN;
3742                         btrfs_err(rc->extent_root->fs_info,
3743                      "extent %llu slot %d has an invalid inline ref type",
3744                              eb->start, path->slots[0]);
3745                 }
3746                 if (ret) {
3747                         err = ret;
3748                         goto out;
3749                 }
3750                 ptr += btrfs_extent_inline_ref_size(key.type);
3751         }
3752         WARN_ON(ptr > end);
3753
3754         while (1) {
3755                 cond_resched();
3756                 eb = path->nodes[0];
3757                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3758                         ret = btrfs_next_leaf(rc->extent_root, path);
3759                         if (ret < 0) {
3760                                 err = ret;
3761                                 break;
3762                         }
3763                         if (ret > 0)
3764                                 break;
3765                         eb = path->nodes[0];
3766                 }
3767
3768                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3769                 if (key.objectid != extent_key->objectid)
3770                         break;
3771
3772                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3773                         ret = __add_tree_block(rc, key.offset, blocksize,
3774                                                blocks);
3775                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3776                         dref = btrfs_item_ptr(eb, path->slots[0],
3777                                               struct btrfs_extent_data_ref);
3778                         ret = find_data_references(rc, extent_key,
3779                                                    eb, dref, blocks);
3780                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
3781                         btrfs_print_v0_err(eb->fs_info);
3782                         btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3783                         ret = -EINVAL;
3784                 } else {
3785                         ret = 0;
3786                 }
3787                 if (ret) {
3788                         err = ret;
3789                         break;
3790                 }
3791                 path->slots[0]++;
3792         }
3793 out:
3794         btrfs_release_path(path);
3795         if (err)
3796                 free_block_list(blocks);
3797         return err;
3798 }
3799
3800 /*
3801  * helper to find next unprocessed extent
3802  */
3803 static noinline_for_stack
3804 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3805                      struct btrfs_key *extent_key)
3806 {
3807         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3808         struct btrfs_key key;
3809         struct extent_buffer *leaf;
3810         u64 start, end, last;
3811         int ret;
3812
3813         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3814         while (1) {
3815                 cond_resched();
3816                 if (rc->search_start >= last) {
3817                         ret = 1;
3818                         break;
3819                 }
3820
3821                 key.objectid = rc->search_start;
3822                 key.type = BTRFS_EXTENT_ITEM_KEY;
3823                 key.offset = 0;
3824
3825                 path->search_commit_root = 1;
3826                 path->skip_locking = 1;
3827                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3828                                         0, 0);
3829                 if (ret < 0)
3830                         break;
3831 next:
3832                 leaf = path->nodes[0];
3833                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3834                         ret = btrfs_next_leaf(rc->extent_root, path);
3835                         if (ret != 0)
3836                                 break;
3837                         leaf = path->nodes[0];
3838                 }
3839
3840                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3841                 if (key.objectid >= last) {
3842                         ret = 1;
3843                         break;
3844                 }
3845
3846                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3847                     key.type != BTRFS_METADATA_ITEM_KEY) {
3848                         path->slots[0]++;
3849                         goto next;
3850                 }
3851
3852                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3853                     key.objectid + key.offset <= rc->search_start) {
3854                         path->slots[0]++;
3855                         goto next;
3856                 }
3857
3858                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3859                     key.objectid + fs_info->nodesize <=
3860                     rc->search_start) {
3861                         path->slots[0]++;
3862                         goto next;
3863                 }
3864
3865                 ret = find_first_extent_bit(&rc->processed_blocks,
3866                                             key.objectid, &start, &end,
3867                                             EXTENT_DIRTY, NULL);
3868
3869                 if (ret == 0 && start <= key.objectid) {
3870                         btrfs_release_path(path);
3871                         rc->search_start = end + 1;
3872                 } else {
3873                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3874                                 rc->search_start = key.objectid + key.offset;
3875                         else
3876                                 rc->search_start = key.objectid +
3877                                         fs_info->nodesize;
3878                         memcpy(extent_key, &key, sizeof(key));
3879                         return 0;
3880                 }
3881         }
3882         btrfs_release_path(path);
3883         return ret;
3884 }
3885
3886 static void set_reloc_control(struct reloc_control *rc)
3887 {
3888         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3889
3890         mutex_lock(&fs_info->reloc_mutex);
3891         fs_info->reloc_ctl = rc;
3892         mutex_unlock(&fs_info->reloc_mutex);
3893 }
3894
3895 static void unset_reloc_control(struct reloc_control *rc)
3896 {
3897         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3898
3899         mutex_lock(&fs_info->reloc_mutex);
3900         fs_info->reloc_ctl = NULL;
3901         mutex_unlock(&fs_info->reloc_mutex);
3902 }
3903
3904 static int check_extent_flags(u64 flags)
3905 {
3906         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3907             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3908                 return 1;
3909         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3910             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3911                 return 1;
3912         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3913             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3914                 return 1;
3915         return 0;
3916 }
3917
3918 static noinline_for_stack
3919 int prepare_to_relocate(struct reloc_control *rc)
3920 {
3921         struct btrfs_trans_handle *trans;
3922         int ret;
3923
3924         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3925                                               BTRFS_BLOCK_RSV_TEMP);
3926         if (!rc->block_rsv)
3927                 return -ENOMEM;
3928
3929         memset(&rc->cluster, 0, sizeof(rc->cluster));
3930         rc->search_start = rc->block_group->key.objectid;
3931         rc->extents_found = 0;
3932         rc->nodes_relocated = 0;
3933         rc->merging_rsv_size = 0;
3934         rc->reserved_bytes = 0;
3935         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3936                               RELOCATION_RESERVED_NODES;
3937         ret = btrfs_block_rsv_refill(rc->extent_root,
3938                                      rc->block_rsv, rc->block_rsv->size,
3939                                      BTRFS_RESERVE_FLUSH_ALL);
3940         if (ret)
3941                 return ret;
3942
3943         rc->create_reloc_tree = 1;
3944         set_reloc_control(rc);
3945
3946         trans = btrfs_join_transaction(rc->extent_root);
3947         if (IS_ERR(trans)) {
3948                 unset_reloc_control(rc);
3949                 /*
3950                  * extent tree is not a ref_cow tree and has no reloc_root to
3951                  * cleanup.  And callers are responsible to free the above
3952                  * block rsv.
3953                  */
3954                 return PTR_ERR(trans);
3955         }
3956         btrfs_commit_transaction(trans);
3957         return 0;
3958 }
3959
3960 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3961 {
3962         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3963         struct rb_root blocks = RB_ROOT;
3964         struct btrfs_key key;
3965         struct btrfs_trans_handle *trans = NULL;
3966         struct btrfs_path *path;
3967         struct btrfs_extent_item *ei;
3968         u64 flags;
3969         u32 item_size;
3970         int ret;
3971         int err = 0;
3972         int progress = 0;
3973
3974         path = btrfs_alloc_path();
3975         if (!path)
3976                 return -ENOMEM;
3977         path->reada = READA_FORWARD;
3978
3979         ret = prepare_to_relocate(rc);
3980         if (ret) {
3981                 err = ret;
3982                 goto out_free;
3983         }
3984
3985         while (1) {
3986                 rc->reserved_bytes = 0;
3987                 ret = btrfs_block_rsv_refill(rc->extent_root,
3988                                         rc->block_rsv, rc->block_rsv->size,
3989                                         BTRFS_RESERVE_FLUSH_ALL);
3990                 if (ret) {
3991                         err = ret;
3992                         break;
3993                 }
3994                 progress++;
3995                 trans = btrfs_start_transaction(rc->extent_root, 0);
3996                 if (IS_ERR(trans)) {
3997                         err = PTR_ERR(trans);
3998                         trans = NULL;
3999                         break;
4000                 }
4001 restart:
4002                 if (update_backref_cache(trans, &rc->backref_cache)) {
4003                         btrfs_end_transaction(trans);
4004                         trans = NULL;
4005                         continue;
4006                 }
4007
4008                 ret = find_next_extent(rc, path, &key);
4009                 if (ret < 0)
4010                         err = ret;
4011                 if (ret != 0)
4012                         break;
4013
4014                 rc->extents_found++;
4015
4016                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4017                                     struct btrfs_extent_item);
4018                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4019                 if (item_size >= sizeof(*ei)) {
4020                         flags = btrfs_extent_flags(path->nodes[0], ei);
4021                         ret = check_extent_flags(flags);
4022                         BUG_ON(ret);
4023                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
4024                         err = -EINVAL;
4025                         btrfs_print_v0_err(trans->fs_info);
4026                         btrfs_abort_transaction(trans, err);
4027                         break;
4028                 } else {
4029                         BUG();
4030                 }
4031
4032                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4033                         ret = add_tree_block(rc, &key, path, &blocks);
4034                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4035                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4036                         ret = add_data_references(rc, &key, path, &blocks);
4037                 } else {
4038                         btrfs_release_path(path);
4039                         ret = 0;
4040                 }
4041                 if (ret < 0) {
4042                         err = ret;
4043                         break;
4044                 }
4045
4046                 if (!RB_EMPTY_ROOT(&blocks)) {
4047                         ret = relocate_tree_blocks(trans, rc, &blocks);
4048                         if (ret < 0) {
4049                                 /*
4050                                  * if we fail to relocate tree blocks, force to update
4051                                  * backref cache when committing transaction.
4052                                  */
4053                                 rc->backref_cache.last_trans = trans->transid - 1;
4054
4055                                 if (ret != -EAGAIN) {
4056                                         err = ret;
4057                                         break;
4058                                 }
4059                                 rc->extents_found--;
4060                                 rc->search_start = key.objectid;
4061                         }
4062                 }
4063
4064                 btrfs_end_transaction_throttle(trans);
4065                 btrfs_btree_balance_dirty(fs_info);
4066                 trans = NULL;
4067
4068                 if (rc->stage == MOVE_DATA_EXTENTS &&
4069                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4070                         rc->found_file_extent = 1;
4071                         ret = relocate_data_extent(rc->data_inode,
4072                                                    &key, &rc->cluster);
4073                         if (ret < 0) {
4074                                 err = ret;
4075                                 break;
4076                         }
4077                 }
4078         }
4079         if (trans && progress && err == -ENOSPC) {
4080                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
4081                 if (ret == 1) {
4082                         err = 0;
4083                         progress = 0;
4084                         goto restart;
4085                 }
4086         }
4087
4088         btrfs_release_path(path);
4089         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4090
4091         if (trans) {
4092                 btrfs_end_transaction_throttle(trans);
4093                 btrfs_btree_balance_dirty(fs_info);
4094         }
4095
4096         if (!err) {
4097                 ret = relocate_file_extent_cluster(rc->data_inode,
4098                                                    &rc->cluster);
4099                 if (ret < 0)
4100                         err = ret;
4101         }
4102
4103         rc->create_reloc_tree = 0;
4104         set_reloc_control(rc);
4105
4106         backref_cache_cleanup(&rc->backref_cache);
4107         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4108
4109         err = prepare_to_merge(rc, err);
4110
4111         merge_reloc_roots(rc);
4112
4113         rc->merge_reloc_tree = 0;
4114         unset_reloc_control(rc);
4115         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4116
4117         /* get rid of pinned extents */
4118         trans = btrfs_join_transaction(rc->extent_root);
4119         if (IS_ERR(trans)) {
4120                 err = PTR_ERR(trans);
4121                 goto out_free;
4122         }
4123         btrfs_commit_transaction(trans);
4124         ret = clean_dirty_subvols(rc);
4125         if (ret < 0 && !err)
4126                 err = ret;
4127 out_free:
4128         btrfs_free_block_rsv(fs_info, rc->block_rsv);
4129         btrfs_free_path(path);
4130         return err;
4131 }
4132
4133 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4134                                  struct btrfs_root *root, u64 objectid)
4135 {
4136         struct btrfs_path *path;
4137         struct btrfs_inode_item *item;
4138         struct extent_buffer *leaf;
4139         int ret;
4140
4141         path = btrfs_alloc_path();
4142         if (!path)
4143                 return -ENOMEM;
4144
4145         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4146         if (ret)
4147                 goto out;
4148
4149         leaf = path->nodes[0];
4150         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4151         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
4152         btrfs_set_inode_generation(leaf, item, 1);
4153         btrfs_set_inode_size(leaf, item, 0);
4154         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4155         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4156                                           BTRFS_INODE_PREALLOC);
4157         btrfs_mark_buffer_dirty(leaf);
4158 out:
4159         btrfs_free_path(path);
4160         return ret;
4161 }
4162
4163 /*
4164  * helper to create inode for data relocation.
4165  * the inode is in data relocation tree and its link count is 0
4166  */
4167 static noinline_for_stack
4168 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4169                                  struct btrfs_block_group_cache *group)
4170 {
4171         struct inode *inode = NULL;
4172         struct btrfs_trans_handle *trans;
4173         struct btrfs_root *root;
4174         struct btrfs_key key;
4175         u64 objectid;
4176         int err = 0;
4177
4178         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4179         if (IS_ERR(root))
4180                 return ERR_CAST(root);
4181
4182         trans = btrfs_start_transaction(root, 6);
4183         if (IS_ERR(trans))
4184                 return ERR_CAST(trans);
4185
4186         err = btrfs_find_free_objectid(root, &objectid);
4187         if (err)
4188                 goto out;
4189
4190         err = __insert_orphan_inode(trans, root, objectid);
4191         BUG_ON(err);
4192
4193         key.objectid = objectid;
4194         key.type = BTRFS_INODE_ITEM_KEY;
4195         key.offset = 0;
4196         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4197         BUG_ON(IS_ERR(inode));
4198         BTRFS_I(inode)->index_cnt = group->key.objectid;
4199
4200         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4201 out:
4202         btrfs_end_transaction(trans);
4203         btrfs_btree_balance_dirty(fs_info);
4204         if (err) {
4205                 if (inode)
4206                         iput(inode);
4207                 inode = ERR_PTR(err);
4208         }
4209         return inode;
4210 }
4211
4212 static struct reloc_control *alloc_reloc_control(void)
4213 {
4214         struct reloc_control *rc;
4215
4216         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4217         if (!rc)
4218                 return NULL;
4219
4220         INIT_LIST_HEAD(&rc->reloc_roots);
4221         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
4222         backref_cache_init(&rc->backref_cache);
4223         mapping_tree_init(&rc->reloc_root_tree);
4224         extent_io_tree_init(&rc->processed_blocks, NULL);
4225         return rc;
4226 }
4227
4228 /*
4229  * Print the block group being relocated
4230  */
4231 static void describe_relocation(struct btrfs_fs_info *fs_info,
4232                                 struct btrfs_block_group_cache *block_group)
4233 {
4234         char buf[128] = {'\0'};
4235
4236         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
4237
4238         btrfs_info(fs_info,
4239                    "relocating block group %llu flags %s",
4240                    block_group->key.objectid, buf);
4241 }
4242
4243 /*
4244  * function to relocate all extents in a block group.
4245  */
4246 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4247 {
4248         struct btrfs_block_group_cache *bg;
4249         struct btrfs_root *extent_root = fs_info->extent_root;
4250         struct reloc_control *rc;
4251         struct inode *inode;
4252         struct btrfs_path *path;
4253         int ret;
4254         int rw = 0;
4255         int err = 0;
4256
4257         bg = btrfs_lookup_block_group(fs_info, group_start);
4258         if (!bg)
4259                 return -ENOENT;
4260
4261         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4262                 btrfs_put_block_group(bg);
4263                 return -ETXTBSY;
4264         }
4265
4266         rc = alloc_reloc_control();
4267         if (!rc) {
4268                 btrfs_put_block_group(bg);
4269                 return -ENOMEM;
4270         }
4271
4272         rc->extent_root = extent_root;
4273         rc->block_group = bg;
4274
4275         ret = btrfs_inc_block_group_ro(rc->block_group);
4276         if (ret) {
4277                 err = ret;
4278                 goto out;
4279         }
4280         rw = 1;
4281
4282         path = btrfs_alloc_path();
4283         if (!path) {
4284                 err = -ENOMEM;
4285                 goto out;
4286         }
4287
4288         inode = lookup_free_space_inode(fs_info, rc->block_group, path);
4289         btrfs_free_path(path);
4290
4291         if (!IS_ERR(inode))
4292                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4293         else
4294                 ret = PTR_ERR(inode);
4295
4296         if (ret && ret != -ENOENT) {
4297                 err = ret;
4298                 goto out;
4299         }
4300
4301         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4302         if (IS_ERR(rc->data_inode)) {
4303                 err = PTR_ERR(rc->data_inode);
4304                 rc->data_inode = NULL;
4305                 goto out;
4306         }
4307
4308         describe_relocation(fs_info, rc->block_group);
4309
4310         btrfs_wait_block_group_reservations(rc->block_group);
4311         btrfs_wait_nocow_writers(rc->block_group);
4312         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4313                                  rc->block_group->key.objectid,
4314                                  rc->block_group->key.offset);
4315
4316         while (1) {
4317                 mutex_lock(&fs_info->cleaner_mutex);
4318                 ret = relocate_block_group(rc);
4319                 mutex_unlock(&fs_info->cleaner_mutex);
4320                 if (ret < 0) {
4321                         err = ret;
4322                         goto out;
4323                 }
4324
4325                 if (rc->extents_found == 0)
4326                         break;
4327
4328                 btrfs_info(fs_info, "found %llu extents", rc->extents_found);
4329
4330                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4331                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4332                                                        (u64)-1);
4333                         if (ret) {
4334                                 err = ret;
4335                                 goto out;
4336                         }
4337                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4338                                                  0, -1);
4339                         rc->stage = UPDATE_DATA_PTRS;
4340                 }
4341         }
4342
4343         WARN_ON(rc->block_group->pinned > 0);
4344         WARN_ON(rc->block_group->reserved > 0);
4345         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4346 out:
4347         if (err && rw)
4348                 btrfs_dec_block_group_ro(rc->block_group);
4349         iput(rc->data_inode);
4350         btrfs_put_block_group(rc->block_group);
4351         kfree(rc);
4352         return err;
4353 }
4354
4355 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4356 {
4357         struct btrfs_fs_info *fs_info = root->fs_info;
4358         struct btrfs_trans_handle *trans;
4359         int ret, err;
4360
4361         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4362         if (IS_ERR(trans))
4363                 return PTR_ERR(trans);
4364
4365         memset(&root->root_item.drop_progress, 0,
4366                 sizeof(root->root_item.drop_progress));
4367         root->root_item.drop_level = 0;
4368         btrfs_set_root_refs(&root->root_item, 0);
4369         ret = btrfs_update_root(trans, fs_info->tree_root,
4370                                 &root->root_key, &root->root_item);
4371
4372         err = btrfs_end_transaction(trans);
4373         if (err)
4374                 return err;
4375         return ret;
4376 }
4377
4378 /*
4379  * recover relocation interrupted by system crash.
4380  *
4381  * this function resumes merging reloc trees with corresponding fs trees.
4382  * this is important for keeping the sharing of tree blocks
4383  */
4384 int btrfs_recover_relocation(struct btrfs_root *root)
4385 {
4386         struct btrfs_fs_info *fs_info = root->fs_info;
4387         LIST_HEAD(reloc_roots);
4388         struct btrfs_key key;
4389         struct btrfs_root *fs_root;
4390         struct btrfs_root *reloc_root;
4391         struct btrfs_path *path;
4392         struct extent_buffer *leaf;
4393         struct reloc_control *rc = NULL;
4394         struct btrfs_trans_handle *trans;
4395         int ret;
4396         int err = 0;
4397
4398         path = btrfs_alloc_path();
4399         if (!path)
4400                 return -ENOMEM;
4401         path->reada = READA_BACK;
4402
4403         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4404         key.type = BTRFS_ROOT_ITEM_KEY;
4405         key.offset = (u64)-1;
4406
4407         while (1) {
4408                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4409                                         path, 0, 0);
4410                 if (ret < 0) {
4411                         err = ret;
4412                         goto out;
4413                 }
4414                 if (ret > 0) {
4415                         if (path->slots[0] == 0)
4416                                 break;
4417                         path->slots[0]--;
4418                 }
4419                 leaf = path->nodes[0];
4420                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4421                 btrfs_release_path(path);
4422
4423                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4424                     key.type != BTRFS_ROOT_ITEM_KEY)
4425                         break;
4426
4427                 reloc_root = btrfs_read_fs_root(root, &key);
4428                 if (IS_ERR(reloc_root)) {
4429                         err = PTR_ERR(reloc_root);
4430                         goto out;
4431                 }
4432
4433                 list_add(&reloc_root->root_list, &reloc_roots);
4434
4435                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4436                         fs_root = read_fs_root(fs_info,
4437                                                reloc_root->root_key.offset);
4438                         if (IS_ERR(fs_root)) {
4439                                 ret = PTR_ERR(fs_root);
4440                                 if (ret != -ENOENT) {
4441                                         err = ret;
4442                                         goto out;
4443                                 }
4444                                 ret = mark_garbage_root(reloc_root);
4445                                 if (ret < 0) {
4446                                         err = ret;
4447                                         goto out;
4448                                 }
4449                         }
4450                 }
4451
4452                 if (key.offset == 0)
4453                         break;
4454
4455                 key.offset--;
4456         }
4457         btrfs_release_path(path);
4458
4459         if (list_empty(&reloc_roots))
4460                 goto out;
4461
4462         rc = alloc_reloc_control();
4463         if (!rc) {
4464                 err = -ENOMEM;
4465                 goto out;
4466         }
4467
4468         rc->extent_root = fs_info->extent_root;
4469
4470         set_reloc_control(rc);
4471
4472         trans = btrfs_join_transaction(rc->extent_root);
4473         if (IS_ERR(trans)) {
4474                 unset_reloc_control(rc);
4475                 err = PTR_ERR(trans);
4476                 goto out_free;
4477         }
4478
4479         rc->merge_reloc_tree = 1;
4480
4481         while (!list_empty(&reloc_roots)) {
4482                 reloc_root = list_entry(reloc_roots.next,
4483                                         struct btrfs_root, root_list);
4484                 list_del(&reloc_root->root_list);
4485
4486                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4487                         list_add_tail(&reloc_root->root_list,
4488                                       &rc->reloc_roots);
4489                         continue;
4490                 }
4491
4492                 fs_root = read_fs_root(fs_info, reloc_root->root_key.offset);
4493                 if (IS_ERR(fs_root)) {
4494                         err = PTR_ERR(fs_root);
4495                         goto out_free;
4496                 }
4497
4498                 err = __add_reloc_root(reloc_root);
4499                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4500                 fs_root->reloc_root = reloc_root;
4501         }
4502
4503         err = btrfs_commit_transaction(trans);
4504         if (err)
4505                 goto out_free;
4506
4507         merge_reloc_roots(rc);
4508
4509         unset_reloc_control(rc);
4510
4511         trans = btrfs_join_transaction(rc->extent_root);
4512         if (IS_ERR(trans)) {
4513                 err = PTR_ERR(trans);
4514                 goto out_free;
4515         }
4516         err = btrfs_commit_transaction(trans);
4517
4518         ret = clean_dirty_subvols(rc);
4519         if (ret < 0 && !err)
4520                 err = ret;
4521 out_free:
4522         kfree(rc);
4523 out:
4524         if (!list_empty(&reloc_roots))
4525                 free_reloc_roots(&reloc_roots);
4526
4527         btrfs_free_path(path);
4528
4529         if (err == 0) {
4530                 /* cleanup orphan inode in data relocation tree */
4531                 fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4532                 if (IS_ERR(fs_root))
4533                         err = PTR_ERR(fs_root);
4534                 else
4535                         err = btrfs_orphan_cleanup(fs_root);
4536         }
4537         return err;
4538 }
4539
4540 /*
4541  * helper to add ordered checksum for data relocation.
4542  *
4543  * cloning checksum properly handles the nodatasum extents.
4544  * it also saves CPU time to re-calculate the checksum.
4545  */
4546 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4547 {
4548         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4549         struct btrfs_ordered_sum *sums;
4550         struct btrfs_ordered_extent *ordered;
4551         int ret;
4552         u64 disk_bytenr;
4553         u64 new_bytenr;
4554         LIST_HEAD(list);
4555
4556         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4557         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4558
4559         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4560         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4561                                        disk_bytenr + len - 1, &list, 0);
4562         if (ret)
4563                 goto out;
4564
4565         while (!list_empty(&list)) {
4566                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4567                 list_del_init(&sums->list);
4568
4569                 /*
4570                  * We need to offset the new_bytenr based on where the csum is.
4571                  * We need to do this because we will read in entire prealloc
4572                  * extents but we may have written to say the middle of the
4573                  * prealloc extent, so we need to make sure the csum goes with
4574                  * the right disk offset.
4575                  *
4576                  * We can do this because the data reloc inode refers strictly
4577                  * to the on disk bytes, so we don't have to worry about
4578                  * disk_len vs real len like with real inodes since it's all
4579                  * disk length.
4580                  */
4581                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4582                 sums->bytenr = new_bytenr;
4583
4584                 btrfs_add_ordered_sum(inode, ordered, sums);
4585         }
4586 out:
4587         btrfs_put_ordered_extent(ordered);
4588         return ret;
4589 }
4590
4591 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4592                           struct btrfs_root *root, struct extent_buffer *buf,
4593                           struct extent_buffer *cow)
4594 {
4595         struct btrfs_fs_info *fs_info = root->fs_info;
4596         struct reloc_control *rc;
4597         struct backref_node *node;
4598         int first_cow = 0;
4599         int level;
4600         int ret = 0;
4601
4602         rc = fs_info->reloc_ctl;
4603         if (!rc)
4604                 return 0;
4605
4606         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4607                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4608
4609         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4610                 if (buf == root->node)
4611                         __update_reloc_root(root, cow->start);
4612         }
4613
4614         level = btrfs_header_level(buf);
4615         if (btrfs_header_generation(buf) <=
4616             btrfs_root_last_snapshot(&root->root_item))
4617                 first_cow = 1;
4618
4619         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4620             rc->create_reloc_tree) {
4621                 WARN_ON(!first_cow && level == 0);
4622
4623                 node = rc->backref_cache.path[level];
4624                 BUG_ON(node->bytenr != buf->start &&
4625                        node->new_bytenr != buf->start);
4626
4627                 drop_node_buffer(node);
4628                 extent_buffer_get(cow);
4629                 node->eb = cow;
4630                 node->new_bytenr = cow->start;
4631
4632                 if (!node->pending) {
4633                         list_move_tail(&node->list,
4634                                        &rc->backref_cache.pending[level]);
4635                         node->pending = 1;
4636                 }
4637
4638                 if (first_cow)
4639                         __mark_block_processed(rc, node);
4640
4641                 if (first_cow && level > 0)
4642                         rc->nodes_relocated += buf->len;
4643         }
4644
4645         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4646                 ret = replace_file_extents(trans, rc, root, cow);
4647         return ret;
4648 }
4649
4650 /*
4651  * called before creating snapshot. it calculates metadata reservation
4652  * required for relocating tree blocks in the snapshot
4653  */
4654 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4655                               u64 *bytes_to_reserve)
4656 {
4657         struct btrfs_root *root;
4658         struct reloc_control *rc;
4659
4660         root = pending->root;
4661         if (!root->reloc_root)
4662                 return;
4663
4664         rc = root->fs_info->reloc_ctl;
4665         if (!rc->merge_reloc_tree)
4666                 return;
4667
4668         root = root->reloc_root;
4669         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4670         /*
4671          * relocation is in the stage of merging trees. the space
4672          * used by merging a reloc tree is twice the size of
4673          * relocated tree nodes in the worst case. half for cowing
4674          * the reloc tree, half for cowing the fs tree. the space
4675          * used by cowing the reloc tree will be freed after the
4676          * tree is dropped. if we create snapshot, cowing the fs
4677          * tree may use more space than it frees. so we need
4678          * reserve extra space.
4679          */
4680         *bytes_to_reserve += rc->nodes_relocated;
4681 }
4682
4683 /*
4684  * called after snapshot is created. migrate block reservation
4685  * and create reloc root for the newly created snapshot
4686  */
4687 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4688                                struct btrfs_pending_snapshot *pending)
4689 {
4690         struct btrfs_root *root = pending->root;
4691         struct btrfs_root *reloc_root;
4692         struct btrfs_root *new_root;
4693         struct reloc_control *rc;
4694         int ret;
4695
4696         if (!root->reloc_root)
4697                 return 0;
4698
4699         rc = root->fs_info->reloc_ctl;
4700         rc->merging_rsv_size += rc->nodes_relocated;
4701
4702         if (rc->merge_reloc_tree) {
4703                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4704                                               rc->block_rsv,
4705                                               rc->nodes_relocated, true);
4706                 if (ret)
4707                         return ret;
4708         }
4709
4710         new_root = pending->snap;
4711         reloc_root = create_reloc_root(trans, root->reloc_root,
4712                                        new_root->root_key.objectid);
4713         if (IS_ERR(reloc_root))
4714                 return PTR_ERR(reloc_root);
4715
4716         ret = __add_reloc_root(reloc_root);
4717         BUG_ON(ret < 0);
4718         new_root->reloc_root = reloc_root;
4719
4720         if (rc->create_reloc_tree)
4721                 ret = clone_backref_node(trans, rc, root, reloc_root);
4722         return ret;
4723 }