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_write(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_write(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_write(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                  *
1893                  * We don't scan the subtree right now, but only record
1894                  * the swapped tree blocks.
1895                  * The real subtree rescan is delayed until we have new
1896                  * CoW on the subtree root node before transaction commit.
1897                  */
1898                 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1899                                 rc->block_group, parent, slot,
1900                                 path->nodes[level], path->slots[level],
1901                                 last_snapshot);
1902                 if (ret < 0)
1903                         break;
1904                 /*
1905                  * swap blocks in fs tree and reloc tree.
1906                  */
1907                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1908                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1909                 btrfs_mark_buffer_dirty(parent);
1910
1911                 btrfs_set_node_blockptr(path->nodes[level],
1912                                         path->slots[level], old_bytenr);
1913                 btrfs_set_node_ptr_generation(path->nodes[level],
1914                                               path->slots[level], old_ptr_gen);
1915                 btrfs_mark_buffer_dirty(path->nodes[level]);
1916
1917                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr,
1918                                         blocksize, path->nodes[level]->start,
1919                                         src->root_key.objectid, level - 1, 0);
1920                 BUG_ON(ret);
1921                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr,
1922                                         blocksize, 0, dest->root_key.objectid,
1923                                         level - 1, 0);
1924                 BUG_ON(ret);
1925
1926                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1927                                         path->nodes[level]->start,
1928                                         src->root_key.objectid, level - 1, 0);
1929                 BUG_ON(ret);
1930
1931                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1932                                         0, dest->root_key.objectid, level - 1,
1933                                         0);
1934                 BUG_ON(ret);
1935
1936                 btrfs_unlock_up_safe(path, 0);
1937
1938                 ret = level;
1939                 break;
1940         }
1941         btrfs_tree_unlock(parent);
1942         free_extent_buffer(parent);
1943         return ret;
1944 }
1945
1946 /*
1947  * helper to find next relocated block in reloc tree
1948  */
1949 static noinline_for_stack
1950 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1951                        int *level)
1952 {
1953         struct extent_buffer *eb;
1954         int i;
1955         u64 last_snapshot;
1956         u32 nritems;
1957
1958         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1959
1960         for (i = 0; i < *level; i++) {
1961                 free_extent_buffer(path->nodes[i]);
1962                 path->nodes[i] = NULL;
1963         }
1964
1965         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1966                 eb = path->nodes[i];
1967                 nritems = btrfs_header_nritems(eb);
1968                 while (path->slots[i] + 1 < nritems) {
1969                         path->slots[i]++;
1970                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1971                             last_snapshot)
1972                                 continue;
1973
1974                         *level = i;
1975                         return 0;
1976                 }
1977                 free_extent_buffer(path->nodes[i]);
1978                 path->nodes[i] = NULL;
1979         }
1980         return 1;
1981 }
1982
1983 /*
1984  * walk down reloc tree to find relocated block of lowest level
1985  */
1986 static noinline_for_stack
1987 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1988                          int *level)
1989 {
1990         struct btrfs_fs_info *fs_info = root->fs_info;
1991         struct extent_buffer *eb = NULL;
1992         int i;
1993         u64 bytenr;
1994         u64 ptr_gen = 0;
1995         u64 last_snapshot;
1996         u32 nritems;
1997
1998         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1999
2000         for (i = *level; i > 0; i--) {
2001                 struct btrfs_key first_key;
2002
2003                 eb = path->nodes[i];
2004                 nritems = btrfs_header_nritems(eb);
2005                 while (path->slots[i] < nritems) {
2006                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2007                         if (ptr_gen > last_snapshot)
2008                                 break;
2009                         path->slots[i]++;
2010                 }
2011                 if (path->slots[i] >= nritems) {
2012                         if (i == *level)
2013                                 break;
2014                         *level = i + 1;
2015                         return 0;
2016                 }
2017                 if (i == 1) {
2018                         *level = i;
2019                         return 0;
2020                 }
2021
2022                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2023                 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
2024                 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
2025                                      &first_key);
2026                 if (IS_ERR(eb)) {
2027                         return PTR_ERR(eb);
2028                 } else if (!extent_buffer_uptodate(eb)) {
2029                         free_extent_buffer(eb);
2030                         return -EIO;
2031                 }
2032                 BUG_ON(btrfs_header_level(eb) != i - 1);
2033                 path->nodes[i - 1] = eb;
2034                 path->slots[i - 1] = 0;
2035         }
2036         return 1;
2037 }
2038
2039 /*
2040  * invalidate extent cache for file extents whose key in range of
2041  * [min_key, max_key)
2042  */
2043 static int invalidate_extent_cache(struct btrfs_root *root,
2044                                    struct btrfs_key *min_key,
2045                                    struct btrfs_key *max_key)
2046 {
2047         struct btrfs_fs_info *fs_info = root->fs_info;
2048         struct inode *inode = NULL;
2049         u64 objectid;
2050         u64 start, end;
2051         u64 ino;
2052
2053         objectid = min_key->objectid;
2054         while (1) {
2055                 cond_resched();
2056                 iput(inode);
2057
2058                 if (objectid > max_key->objectid)
2059                         break;
2060
2061                 inode = find_next_inode(root, objectid);
2062                 if (!inode)
2063                         break;
2064                 ino = btrfs_ino(BTRFS_I(inode));
2065
2066                 if (ino > max_key->objectid) {
2067                         iput(inode);
2068                         break;
2069                 }
2070
2071                 objectid = ino + 1;
2072                 if (!S_ISREG(inode->i_mode))
2073                         continue;
2074
2075                 if (unlikely(min_key->objectid == ino)) {
2076                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2077                                 continue;
2078                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2079                                 start = 0;
2080                         else {
2081                                 start = min_key->offset;
2082                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
2083                         }
2084                 } else {
2085                         start = 0;
2086                 }
2087
2088                 if (unlikely(max_key->objectid == ino)) {
2089                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2090                                 continue;
2091                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2092                                 end = (u64)-1;
2093                         } else {
2094                                 if (max_key->offset == 0)
2095                                         continue;
2096                                 end = max_key->offset;
2097                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
2098                                 end--;
2099                         }
2100                 } else {
2101                         end = (u64)-1;
2102                 }
2103
2104                 /* the lock_extent waits for readpage to complete */
2105                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2106                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
2107                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2108         }
2109         return 0;
2110 }
2111
2112 static int find_next_key(struct btrfs_path *path, int level,
2113                          struct btrfs_key *key)
2114
2115 {
2116         while (level < BTRFS_MAX_LEVEL) {
2117                 if (!path->nodes[level])
2118                         break;
2119                 if (path->slots[level] + 1 <
2120                     btrfs_header_nritems(path->nodes[level])) {
2121                         btrfs_node_key_to_cpu(path->nodes[level], key,
2122                                               path->slots[level] + 1);
2123                         return 0;
2124                 }
2125                 level++;
2126         }
2127         return 1;
2128 }
2129
2130 /*
2131  * Insert current subvolume into reloc_control::dirty_subvol_roots
2132  */
2133 static void insert_dirty_subvol(struct btrfs_trans_handle *trans,
2134                                 struct reloc_control *rc,
2135                                 struct btrfs_root *root)
2136 {
2137         struct btrfs_root *reloc_root = root->reloc_root;
2138         struct btrfs_root_item *reloc_root_item;
2139
2140         /* @root must be a subvolume tree root with a valid reloc tree */
2141         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
2142         ASSERT(reloc_root);
2143
2144         reloc_root_item = &reloc_root->root_item;
2145         memset(&reloc_root_item->drop_progress, 0,
2146                 sizeof(reloc_root_item->drop_progress));
2147         reloc_root_item->drop_level = 0;
2148         btrfs_set_root_refs(reloc_root_item, 0);
2149         btrfs_update_reloc_root(trans, root);
2150
2151         if (list_empty(&root->reloc_dirty_list)) {
2152                 btrfs_grab_fs_root(root);
2153                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
2154         }
2155 }
2156
2157 static int clean_dirty_subvols(struct reloc_control *rc)
2158 {
2159         struct btrfs_root *root;
2160         struct btrfs_root *next;
2161         int ret = 0;
2162
2163         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
2164                                  reloc_dirty_list) {
2165                 struct btrfs_root *reloc_root = root->reloc_root;
2166
2167                 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
2168                 list_del_init(&root->reloc_dirty_list);
2169                 root->reloc_root = NULL;
2170                 if (reloc_root) {
2171                         int ret2;
2172
2173                         ret2 = btrfs_drop_snapshot(reloc_root, NULL, 0, 1);
2174                         if (ret2 < 0 && !ret)
2175                                 ret = ret2;
2176                 }
2177                 btrfs_put_fs_root(root);
2178         }
2179         return ret;
2180 }
2181
2182 /*
2183  * merge the relocated tree blocks in reloc tree with corresponding
2184  * fs tree.
2185  */
2186 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2187                                                struct btrfs_root *root)
2188 {
2189         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2190         struct btrfs_key key;
2191         struct btrfs_key next_key;
2192         struct btrfs_trans_handle *trans = NULL;
2193         struct btrfs_root *reloc_root;
2194         struct btrfs_root_item *root_item;
2195         struct btrfs_path *path;
2196         struct extent_buffer *leaf;
2197         int level;
2198         int max_level;
2199         int replaced = 0;
2200         int ret;
2201         int err = 0;
2202         u32 min_reserved;
2203
2204         path = btrfs_alloc_path();
2205         if (!path)
2206                 return -ENOMEM;
2207         path->reada = READA_FORWARD;
2208
2209         reloc_root = root->reloc_root;
2210         root_item = &reloc_root->root_item;
2211
2212         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2213                 level = btrfs_root_level(root_item);
2214                 extent_buffer_get(reloc_root->node);
2215                 path->nodes[level] = reloc_root->node;
2216                 path->slots[level] = 0;
2217         } else {
2218                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2219
2220                 level = root_item->drop_level;
2221                 BUG_ON(level == 0);
2222                 path->lowest_level = level;
2223                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2224                 path->lowest_level = 0;
2225                 if (ret < 0) {
2226                         btrfs_free_path(path);
2227                         return ret;
2228                 }
2229
2230                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2231                                       path->slots[level]);
2232                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2233
2234                 btrfs_unlock_up_safe(path, 0);
2235         }
2236
2237         min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2238         memset(&next_key, 0, sizeof(next_key));
2239
2240         while (1) {
2241                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2242                                              BTRFS_RESERVE_FLUSH_ALL);
2243                 if (ret) {
2244                         err = ret;
2245                         goto out;
2246                 }
2247                 trans = btrfs_start_transaction(root, 0);
2248                 if (IS_ERR(trans)) {
2249                         err = PTR_ERR(trans);
2250                         trans = NULL;
2251                         goto out;
2252                 }
2253                 trans->block_rsv = rc->block_rsv;
2254
2255                 replaced = 0;
2256                 max_level = level;
2257
2258                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2259                 if (ret < 0) {
2260                         err = ret;
2261                         goto out;
2262                 }
2263                 if (ret > 0)
2264                         break;
2265
2266                 if (!find_next_key(path, level, &key) &&
2267                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2268                         ret = 0;
2269                 } else {
2270                         ret = replace_path(trans, rc, root, reloc_root, path,
2271                                            &next_key, level, max_level);
2272                 }
2273                 if (ret < 0) {
2274                         err = ret;
2275                         goto out;
2276                 }
2277
2278                 if (ret > 0) {
2279                         level = ret;
2280                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2281                                               path->slots[level]);
2282                         replaced = 1;
2283                 }
2284
2285                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2286                 if (ret > 0)
2287                         break;
2288
2289                 BUG_ON(level == 0);
2290                 /*
2291                  * save the merging progress in the drop_progress.
2292                  * this is OK since root refs == 1 in this case.
2293                  */
2294                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2295                                path->slots[level]);
2296                 root_item->drop_level = level;
2297
2298                 btrfs_end_transaction_throttle(trans);
2299                 trans = NULL;
2300
2301                 btrfs_btree_balance_dirty(fs_info);
2302
2303                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2304                         invalidate_extent_cache(root, &key, &next_key);
2305         }
2306
2307         /*
2308          * handle the case only one block in the fs tree need to be
2309          * relocated and the block is tree root.
2310          */
2311         leaf = btrfs_lock_root_node(root);
2312         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2313         btrfs_tree_unlock(leaf);
2314         free_extent_buffer(leaf);
2315         if (ret < 0)
2316                 err = ret;
2317 out:
2318         btrfs_free_path(path);
2319
2320         if (err == 0)
2321                 insert_dirty_subvol(trans, rc, root);
2322
2323         if (trans)
2324                 btrfs_end_transaction_throttle(trans);
2325
2326         btrfs_btree_balance_dirty(fs_info);
2327
2328         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2329                 invalidate_extent_cache(root, &key, &next_key);
2330
2331         return err;
2332 }
2333
2334 static noinline_for_stack
2335 int prepare_to_merge(struct reloc_control *rc, int err)
2336 {
2337         struct btrfs_root *root = rc->extent_root;
2338         struct btrfs_fs_info *fs_info = root->fs_info;
2339         struct btrfs_root *reloc_root;
2340         struct btrfs_trans_handle *trans;
2341         LIST_HEAD(reloc_roots);
2342         u64 num_bytes = 0;
2343         int ret;
2344
2345         mutex_lock(&fs_info->reloc_mutex);
2346         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2347         rc->merging_rsv_size += rc->nodes_relocated * 2;
2348         mutex_unlock(&fs_info->reloc_mutex);
2349
2350 again:
2351         if (!err) {
2352                 num_bytes = rc->merging_rsv_size;
2353                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2354                                           BTRFS_RESERVE_FLUSH_ALL);
2355                 if (ret)
2356                         err = ret;
2357         }
2358
2359         trans = btrfs_join_transaction(rc->extent_root);
2360         if (IS_ERR(trans)) {
2361                 if (!err)
2362                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2363                                                 num_bytes);
2364                 return PTR_ERR(trans);
2365         }
2366
2367         if (!err) {
2368                 if (num_bytes != rc->merging_rsv_size) {
2369                         btrfs_end_transaction(trans);
2370                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
2371                                                 num_bytes);
2372                         goto again;
2373                 }
2374         }
2375
2376         rc->merge_reloc_tree = 1;
2377
2378         while (!list_empty(&rc->reloc_roots)) {
2379                 reloc_root = list_entry(rc->reloc_roots.next,
2380                                         struct btrfs_root, root_list);
2381                 list_del_init(&reloc_root->root_list);
2382
2383                 root = read_fs_root(fs_info, reloc_root->root_key.offset);
2384                 BUG_ON(IS_ERR(root));
2385                 BUG_ON(root->reloc_root != reloc_root);
2386
2387                 /*
2388                  * set reference count to 1, so btrfs_recover_relocation
2389                  * knows it should resumes merging
2390                  */
2391                 if (!err)
2392                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2393                 btrfs_update_reloc_root(trans, root);
2394
2395                 list_add(&reloc_root->root_list, &reloc_roots);
2396         }
2397
2398         list_splice(&reloc_roots, &rc->reloc_roots);
2399
2400         if (!err)
2401                 btrfs_commit_transaction(trans);
2402         else
2403                 btrfs_end_transaction(trans);
2404         return err;
2405 }
2406
2407 static noinline_for_stack
2408 void free_reloc_roots(struct list_head *list)
2409 {
2410         struct btrfs_root *reloc_root;
2411
2412         while (!list_empty(list)) {
2413                 reloc_root = list_entry(list->next, struct btrfs_root,
2414                                         root_list);
2415                 __del_reloc_root(reloc_root);
2416                 free_extent_buffer(reloc_root->node);
2417                 free_extent_buffer(reloc_root->commit_root);
2418                 reloc_root->node = NULL;
2419                 reloc_root->commit_root = NULL;
2420         }
2421 }
2422
2423 static noinline_for_stack
2424 void merge_reloc_roots(struct reloc_control *rc)
2425 {
2426         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2427         struct btrfs_root *root;
2428         struct btrfs_root *reloc_root;
2429         LIST_HEAD(reloc_roots);
2430         int found = 0;
2431         int ret = 0;
2432 again:
2433         root = rc->extent_root;
2434
2435         /*
2436          * this serializes us with btrfs_record_root_in_transaction,
2437          * we have to make sure nobody is in the middle of
2438          * adding their roots to the list while we are
2439          * doing this splice
2440          */
2441         mutex_lock(&fs_info->reloc_mutex);
2442         list_splice_init(&rc->reloc_roots, &reloc_roots);
2443         mutex_unlock(&fs_info->reloc_mutex);
2444
2445         while (!list_empty(&reloc_roots)) {
2446                 found = 1;
2447                 reloc_root = list_entry(reloc_roots.next,
2448                                         struct btrfs_root, root_list);
2449
2450                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2451                         root = read_fs_root(fs_info,
2452                                             reloc_root->root_key.offset);
2453                         BUG_ON(IS_ERR(root));
2454                         BUG_ON(root->reloc_root != reloc_root);
2455
2456                         ret = merge_reloc_root(rc, root);
2457                         if (ret) {
2458                                 if (list_empty(&reloc_root->root_list))
2459                                         list_add_tail(&reloc_root->root_list,
2460                                                       &reloc_roots);
2461                                 goto out;
2462                         }
2463                 } else {
2464                         list_del_init(&reloc_root->root_list);
2465                 }
2466         }
2467
2468         if (found) {
2469                 found = 0;
2470                 goto again;
2471         }
2472 out:
2473         if (ret) {
2474                 btrfs_handle_fs_error(fs_info, ret, NULL);
2475                 if (!list_empty(&reloc_roots))
2476                         free_reloc_roots(&reloc_roots);
2477
2478                 /* new reloc root may be added */
2479                 mutex_lock(&fs_info->reloc_mutex);
2480                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2481                 mutex_unlock(&fs_info->reloc_mutex);
2482                 if (!list_empty(&reloc_roots))
2483                         free_reloc_roots(&reloc_roots);
2484         }
2485
2486         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2487 }
2488
2489 static void free_block_list(struct rb_root *blocks)
2490 {
2491         struct tree_block *block;
2492         struct rb_node *rb_node;
2493         while ((rb_node = rb_first(blocks))) {
2494                 block = rb_entry(rb_node, struct tree_block, rb_node);
2495                 rb_erase(rb_node, blocks);
2496                 kfree(block);
2497         }
2498 }
2499
2500 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2501                                       struct btrfs_root *reloc_root)
2502 {
2503         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2504         struct btrfs_root *root;
2505
2506         if (reloc_root->last_trans == trans->transid)
2507                 return 0;
2508
2509         root = read_fs_root(fs_info, reloc_root->root_key.offset);
2510         BUG_ON(IS_ERR(root));
2511         BUG_ON(root->reloc_root != reloc_root);
2512
2513         return btrfs_record_root_in_trans(trans, root);
2514 }
2515
2516 static noinline_for_stack
2517 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2518                                      struct reloc_control *rc,
2519                                      struct backref_node *node,
2520                                      struct backref_edge *edges[])
2521 {
2522         struct backref_node *next;
2523         struct btrfs_root *root;
2524         int index = 0;
2525
2526         next = node;
2527         while (1) {
2528                 cond_resched();
2529                 next = walk_up_backref(next, edges, &index);
2530                 root = next->root;
2531                 BUG_ON(!root);
2532                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2533
2534                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2535                         record_reloc_root_in_trans(trans, root);
2536                         break;
2537                 }
2538
2539                 btrfs_record_root_in_trans(trans, root);
2540                 root = root->reloc_root;
2541
2542                 if (next->new_bytenr != root->node->start) {
2543                         BUG_ON(next->new_bytenr);
2544                         BUG_ON(!list_empty(&next->list));
2545                         next->new_bytenr = root->node->start;
2546                         next->root = root;
2547                         list_add_tail(&next->list,
2548                                       &rc->backref_cache.changed);
2549                         __mark_block_processed(rc, next);
2550                         break;
2551                 }
2552
2553                 WARN_ON(1);
2554                 root = NULL;
2555                 next = walk_down_backref(edges, &index);
2556                 if (!next || next->level <= node->level)
2557                         break;
2558         }
2559         if (!root)
2560                 return NULL;
2561
2562         next = node;
2563         /* setup backref node path for btrfs_reloc_cow_block */
2564         while (1) {
2565                 rc->backref_cache.path[next->level] = next;
2566                 if (--index < 0)
2567                         break;
2568                 next = edges[index]->node[UPPER];
2569         }
2570         return root;
2571 }
2572
2573 /*
2574  * select a tree root for relocation. return NULL if the block
2575  * is reference counted. we should use do_relocation() in this
2576  * case. return a tree root pointer if the block isn't reference
2577  * counted. return -ENOENT if the block is root of reloc tree.
2578  */
2579 static noinline_for_stack
2580 struct btrfs_root *select_one_root(struct backref_node *node)
2581 {
2582         struct backref_node *next;
2583         struct btrfs_root *root;
2584         struct btrfs_root *fs_root = NULL;
2585         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2586         int index = 0;
2587
2588         next = node;
2589         while (1) {
2590                 cond_resched();
2591                 next = walk_up_backref(next, edges, &index);
2592                 root = next->root;
2593                 BUG_ON(!root);
2594
2595                 /* no other choice for non-references counted tree */
2596                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2597                         return root;
2598
2599                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2600                         fs_root = root;
2601
2602                 if (next != node)
2603                         return NULL;
2604
2605                 next = walk_down_backref(edges, &index);
2606                 if (!next || next->level <= node->level)
2607                         break;
2608         }
2609
2610         if (!fs_root)
2611                 return ERR_PTR(-ENOENT);
2612         return fs_root;
2613 }
2614
2615 static noinline_for_stack
2616 u64 calcu_metadata_size(struct reloc_control *rc,
2617                         struct backref_node *node, int reserve)
2618 {
2619         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2620         struct backref_node *next = node;
2621         struct backref_edge *edge;
2622         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2623         u64 num_bytes = 0;
2624         int index = 0;
2625
2626         BUG_ON(reserve && node->processed);
2627
2628         while (next) {
2629                 cond_resched();
2630                 while (1) {
2631                         if (next->processed && (reserve || next != node))
2632                                 break;
2633
2634                         num_bytes += fs_info->nodesize;
2635
2636                         if (list_empty(&next->upper))
2637                                 break;
2638
2639                         edge = list_entry(next->upper.next,
2640                                           struct backref_edge, list[LOWER]);
2641                         edges[index++] = edge;
2642                         next = edge->node[UPPER];
2643                 }
2644                 next = walk_down_backref(edges, &index);
2645         }
2646         return num_bytes;
2647 }
2648
2649 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2650                                   struct reloc_control *rc,
2651                                   struct backref_node *node)
2652 {
2653         struct btrfs_root *root = rc->extent_root;
2654         struct btrfs_fs_info *fs_info = root->fs_info;
2655         u64 num_bytes;
2656         int ret;
2657         u64 tmp;
2658
2659         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2660
2661         trans->block_rsv = rc->block_rsv;
2662         rc->reserved_bytes += num_bytes;
2663
2664         /*
2665          * We are under a transaction here so we can only do limited flushing.
2666          * If we get an enospc just kick back -EAGAIN so we know to drop the
2667          * transaction and try to refill when we can flush all the things.
2668          */
2669         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2670                                 BTRFS_RESERVE_FLUSH_LIMIT);
2671         if (ret) {
2672                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2673                 while (tmp <= rc->reserved_bytes)
2674                         tmp <<= 1;
2675                 /*
2676                  * only one thread can access block_rsv at this point,
2677                  * so we don't need hold lock to protect block_rsv.
2678                  * we expand more reservation size here to allow enough
2679                  * space for relocation and we will return earlier in
2680                  * enospc case.
2681                  */
2682                 rc->block_rsv->size = tmp + fs_info->nodesize *
2683                                       RELOCATION_RESERVED_NODES;
2684                 return -EAGAIN;
2685         }
2686
2687         return 0;
2688 }
2689
2690 /*
2691  * relocate a block tree, and then update pointers in upper level
2692  * blocks that reference the block to point to the new location.
2693  *
2694  * if called by link_to_upper, the block has already been relocated.
2695  * in that case this function just updates pointers.
2696  */
2697 static int do_relocation(struct btrfs_trans_handle *trans,
2698                          struct reloc_control *rc,
2699                          struct backref_node *node,
2700                          struct btrfs_key *key,
2701                          struct btrfs_path *path, int lowest)
2702 {
2703         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2704         struct backref_node *upper;
2705         struct backref_edge *edge;
2706         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2707         struct btrfs_root *root;
2708         struct extent_buffer *eb;
2709         u32 blocksize;
2710         u64 bytenr;
2711         u64 generation;
2712         int slot;
2713         int ret;
2714         int err = 0;
2715
2716         BUG_ON(lowest && node->eb);
2717
2718         path->lowest_level = node->level + 1;
2719         rc->backref_cache.path[node->level] = node;
2720         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2721                 struct btrfs_key first_key;
2722
2723                 cond_resched();
2724
2725                 upper = edge->node[UPPER];
2726                 root = select_reloc_root(trans, rc, upper, edges);
2727                 BUG_ON(!root);
2728
2729                 if (upper->eb && !upper->locked) {
2730                         if (!lowest) {
2731                                 ret = btrfs_bin_search(upper->eb, key,
2732                                                        upper->level, &slot);
2733                                 BUG_ON(ret);
2734                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2735                                 if (node->eb->start == bytenr)
2736                                         goto next;
2737                         }
2738                         drop_node_buffer(upper);
2739                 }
2740
2741                 if (!upper->eb) {
2742                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2743                         if (ret) {
2744                                 if (ret < 0)
2745                                         err = ret;
2746                                 else
2747                                         err = -ENOENT;
2748
2749                                 btrfs_release_path(path);
2750                                 break;
2751                         }
2752
2753                         if (!upper->eb) {
2754                                 upper->eb = path->nodes[upper->level];
2755                                 path->nodes[upper->level] = NULL;
2756                         } else {
2757                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2758                         }
2759
2760                         upper->locked = 1;
2761                         path->locks[upper->level] = 0;
2762
2763                         slot = path->slots[upper->level];
2764                         btrfs_release_path(path);
2765                 } else {
2766                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2767                                                &slot);
2768                         BUG_ON(ret);
2769                 }
2770
2771                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2772                 if (lowest) {
2773                         if (bytenr != node->bytenr) {
2774                                 btrfs_err(root->fs_info,
2775                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2776                                           bytenr, node->bytenr, slot,
2777                                           upper->eb->start);
2778                                 err = -EIO;
2779                                 goto next;
2780                         }
2781                 } else {
2782                         if (node->eb->start == bytenr)
2783                                 goto next;
2784                 }
2785
2786                 blocksize = root->fs_info->nodesize;
2787                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2788                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2789                 eb = read_tree_block(fs_info, bytenr, generation,
2790                                      upper->level - 1, &first_key);
2791                 if (IS_ERR(eb)) {
2792                         err = PTR_ERR(eb);
2793                         goto next;
2794                 } else if (!extent_buffer_uptodate(eb)) {
2795                         free_extent_buffer(eb);
2796                         err = -EIO;
2797                         goto next;
2798                 }
2799                 btrfs_tree_lock(eb);
2800                 btrfs_set_lock_blocking_write(eb);
2801
2802                 if (!node->eb) {
2803                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2804                                               slot, &eb);
2805                         btrfs_tree_unlock(eb);
2806                         free_extent_buffer(eb);
2807                         if (ret < 0) {
2808                                 err = ret;
2809                                 goto next;
2810                         }
2811                         BUG_ON(node->eb != eb);
2812                 } else {
2813                         btrfs_set_node_blockptr(upper->eb, slot,
2814                                                 node->eb->start);
2815                         btrfs_set_node_ptr_generation(upper->eb, slot,
2816                                                       trans->transid);
2817                         btrfs_mark_buffer_dirty(upper->eb);
2818
2819                         ret = btrfs_inc_extent_ref(trans, root,
2820                                                 node->eb->start, blocksize,
2821                                                 upper->eb->start,
2822                                                 btrfs_header_owner(upper->eb),
2823                                                 node->level, 0);
2824                         BUG_ON(ret);
2825
2826                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2827                         BUG_ON(ret);
2828                 }
2829 next:
2830                 if (!upper->pending)
2831                         drop_node_buffer(upper);
2832                 else
2833                         unlock_node_buffer(upper);
2834                 if (err)
2835                         break;
2836         }
2837
2838         if (!err && node->pending) {
2839                 drop_node_buffer(node);
2840                 list_move_tail(&node->list, &rc->backref_cache.changed);
2841                 node->pending = 0;
2842         }
2843
2844         path->lowest_level = 0;
2845         BUG_ON(err == -ENOSPC);
2846         return err;
2847 }
2848
2849 static int link_to_upper(struct btrfs_trans_handle *trans,
2850                          struct reloc_control *rc,
2851                          struct backref_node *node,
2852                          struct btrfs_path *path)
2853 {
2854         struct btrfs_key key;
2855
2856         btrfs_node_key_to_cpu(node->eb, &key, 0);
2857         return do_relocation(trans, rc, node, &key, path, 0);
2858 }
2859
2860 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2861                                 struct reloc_control *rc,
2862                                 struct btrfs_path *path, int err)
2863 {
2864         LIST_HEAD(list);
2865         struct backref_cache *cache = &rc->backref_cache;
2866         struct backref_node *node;
2867         int level;
2868         int ret;
2869
2870         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2871                 while (!list_empty(&cache->pending[level])) {
2872                         node = list_entry(cache->pending[level].next,
2873                                           struct backref_node, list);
2874                         list_move_tail(&node->list, &list);
2875                         BUG_ON(!node->pending);
2876
2877                         if (!err) {
2878                                 ret = link_to_upper(trans, rc, node, path);
2879                                 if (ret < 0)
2880                                         err = ret;
2881                         }
2882                 }
2883                 list_splice_init(&list, &cache->pending[level]);
2884         }
2885         return err;
2886 }
2887
2888 static void mark_block_processed(struct reloc_control *rc,
2889                                  u64 bytenr, u32 blocksize)
2890 {
2891         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2892                         EXTENT_DIRTY);
2893 }
2894
2895 static void __mark_block_processed(struct reloc_control *rc,
2896                                    struct backref_node *node)
2897 {
2898         u32 blocksize;
2899         if (node->level == 0 ||
2900             in_block_group(node->bytenr, rc->block_group)) {
2901                 blocksize = rc->extent_root->fs_info->nodesize;
2902                 mark_block_processed(rc, node->bytenr, blocksize);
2903         }
2904         node->processed = 1;
2905 }
2906
2907 /*
2908  * mark a block and all blocks directly/indirectly reference the block
2909  * as processed.
2910  */
2911 static void update_processed_blocks(struct reloc_control *rc,
2912                                     struct backref_node *node)
2913 {
2914         struct backref_node *next = node;
2915         struct backref_edge *edge;
2916         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2917         int index = 0;
2918
2919         while (next) {
2920                 cond_resched();
2921                 while (1) {
2922                         if (next->processed)
2923                                 break;
2924
2925                         __mark_block_processed(rc, next);
2926
2927                         if (list_empty(&next->upper))
2928                                 break;
2929
2930                         edge = list_entry(next->upper.next,
2931                                           struct backref_edge, list[LOWER]);
2932                         edges[index++] = edge;
2933                         next = edge->node[UPPER];
2934                 }
2935                 next = walk_down_backref(edges, &index);
2936         }
2937 }
2938
2939 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2940 {
2941         u32 blocksize = rc->extent_root->fs_info->nodesize;
2942
2943         if (test_range_bit(&rc->processed_blocks, bytenr,
2944                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2945                 return 1;
2946         return 0;
2947 }
2948
2949 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2950                               struct tree_block *block)
2951 {
2952         struct extent_buffer *eb;
2953
2954         BUG_ON(block->key_ready);
2955         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2956                              block->level, NULL);
2957         if (IS_ERR(eb)) {
2958                 return PTR_ERR(eb);
2959         } else if (!extent_buffer_uptodate(eb)) {
2960                 free_extent_buffer(eb);
2961                 return -EIO;
2962         }
2963         if (block->level == 0)
2964                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2965         else
2966                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2967         free_extent_buffer(eb);
2968         block->key_ready = 1;
2969         return 0;
2970 }
2971
2972 /*
2973  * helper function to relocate a tree block
2974  */
2975 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2976                                 struct reloc_control *rc,
2977                                 struct backref_node *node,
2978                                 struct btrfs_key *key,
2979                                 struct btrfs_path *path)
2980 {
2981         struct btrfs_root *root;
2982         int ret = 0;
2983
2984         if (!node)
2985                 return 0;
2986
2987         BUG_ON(node->processed);
2988         root = select_one_root(node);
2989         if (root == ERR_PTR(-ENOENT)) {
2990                 update_processed_blocks(rc, node);
2991                 goto out;
2992         }
2993
2994         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2995                 ret = reserve_metadata_space(trans, rc, node);
2996                 if (ret)
2997                         goto out;
2998         }
2999
3000         if (root) {
3001                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3002                         BUG_ON(node->new_bytenr);
3003                         BUG_ON(!list_empty(&node->list));
3004                         btrfs_record_root_in_trans(trans, root);
3005                         root = root->reloc_root;
3006                         node->new_bytenr = root->node->start;
3007                         node->root = root;
3008                         list_add_tail(&node->list, &rc->backref_cache.changed);
3009                 } else {
3010                         path->lowest_level = node->level;
3011                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
3012                         btrfs_release_path(path);
3013                         if (ret > 0)
3014                                 ret = 0;
3015                 }
3016                 if (!ret)
3017                         update_processed_blocks(rc, node);
3018         } else {
3019                 ret = do_relocation(trans, rc, node, key, path, 1);
3020         }
3021 out:
3022         if (ret || node->level == 0 || node->cowonly)
3023                 remove_backref_node(&rc->backref_cache, node);
3024         return ret;
3025 }
3026
3027 /*
3028  * relocate a list of blocks
3029  */
3030 static noinline_for_stack
3031 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
3032                          struct reloc_control *rc, struct rb_root *blocks)
3033 {
3034         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3035         struct backref_node *node;
3036         struct btrfs_path *path;
3037         struct tree_block *block;
3038         struct tree_block *next;
3039         int ret;
3040         int err = 0;
3041
3042         path = btrfs_alloc_path();
3043         if (!path) {
3044                 err = -ENOMEM;
3045                 goto out_free_blocks;
3046         }
3047
3048         /* Kick in readahead for tree blocks with missing keys */
3049         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3050                 if (!block->key_ready)
3051                         readahead_tree_block(fs_info, block->bytenr);
3052         }
3053
3054         /* Get first keys */
3055         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3056                 if (!block->key_ready) {
3057                         err = get_tree_block_key(fs_info, block);
3058                         if (err)
3059                                 goto out_free_path;
3060                 }
3061         }
3062
3063         /* Do tree relocation */
3064         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
3065                 node = build_backref_tree(rc, &block->key,
3066                                           block->level, block->bytenr);
3067                 if (IS_ERR(node)) {
3068                         err = PTR_ERR(node);
3069                         goto out;
3070                 }
3071
3072                 ret = relocate_tree_block(trans, rc, node, &block->key,
3073                                           path);
3074                 if (ret < 0) {
3075                         if (ret != -EAGAIN || &block->rb_node == rb_first(blocks))
3076                                 err = ret;
3077                         goto out;
3078                 }
3079         }
3080 out:
3081         err = finish_pending_nodes(trans, rc, path, err);
3082
3083 out_free_path:
3084         btrfs_free_path(path);
3085 out_free_blocks:
3086         free_block_list(blocks);
3087         return err;
3088 }
3089
3090 static noinline_for_stack
3091 int prealloc_file_extent_cluster(struct inode *inode,
3092                                  struct file_extent_cluster *cluster)
3093 {
3094         u64 alloc_hint = 0;
3095         u64 start;
3096         u64 end;
3097         u64 offset = BTRFS_I(inode)->index_cnt;
3098         u64 num_bytes;
3099         int nr = 0;
3100         int ret = 0;
3101         u64 prealloc_start = cluster->start - offset;
3102         u64 prealloc_end = cluster->end - offset;
3103         u64 cur_offset;
3104         struct extent_changeset *data_reserved = NULL;
3105
3106         BUG_ON(cluster->start != cluster->boundary[0]);
3107         inode_lock(inode);
3108
3109         ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
3110                                           prealloc_end + 1 - prealloc_start);
3111         if (ret)
3112                 goto out;
3113
3114         cur_offset = prealloc_start;
3115         while (nr < cluster->nr) {
3116                 start = cluster->boundary[nr] - offset;
3117                 if (nr + 1 < cluster->nr)
3118                         end = cluster->boundary[nr + 1] - 1 - offset;
3119                 else
3120                         end = cluster->end - offset;
3121
3122                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3123                 num_bytes = end + 1 - start;
3124                 if (cur_offset < start)
3125                         btrfs_free_reserved_data_space(inode, data_reserved,
3126                                         cur_offset, start - cur_offset);
3127                 ret = btrfs_prealloc_file_range(inode, 0, start,
3128                                                 num_bytes, num_bytes,
3129                                                 end + 1, &alloc_hint);
3130                 cur_offset = end + 1;
3131                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3132                 if (ret)
3133                         break;
3134                 nr++;
3135         }
3136         if (cur_offset < prealloc_end)
3137                 btrfs_free_reserved_data_space(inode, data_reserved,
3138                                 cur_offset, prealloc_end + 1 - cur_offset);
3139 out:
3140         inode_unlock(inode);
3141         extent_changeset_free(data_reserved);
3142         return ret;
3143 }
3144
3145 static noinline_for_stack
3146 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3147                          u64 block_start)
3148 {
3149         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3150         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3151         struct extent_map *em;
3152         int ret = 0;
3153
3154         em = alloc_extent_map();
3155         if (!em)
3156                 return -ENOMEM;
3157
3158         em->start = start;
3159         em->len = end + 1 - start;
3160         em->block_len = em->len;
3161         em->block_start = block_start;
3162         em->bdev = fs_info->fs_devices->latest_bdev;
3163         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3164
3165         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3166         while (1) {
3167                 write_lock(&em_tree->lock);
3168                 ret = add_extent_mapping(em_tree, em, 0);
3169                 write_unlock(&em_tree->lock);
3170                 if (ret != -EEXIST) {
3171                         free_extent_map(em);
3172                         break;
3173                 }
3174                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3175         }
3176         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3177         return ret;
3178 }
3179
3180 static int relocate_file_extent_cluster(struct inode *inode,
3181                                         struct file_extent_cluster *cluster)
3182 {
3183         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3184         u64 page_start;
3185         u64 page_end;
3186         u64 offset = BTRFS_I(inode)->index_cnt;
3187         unsigned long index;
3188         unsigned long last_index;
3189         struct page *page;
3190         struct file_ra_state *ra;
3191         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3192         int nr = 0;
3193         int ret = 0;
3194
3195         if (!cluster->nr)
3196                 return 0;
3197
3198         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3199         if (!ra)
3200                 return -ENOMEM;
3201
3202         ret = prealloc_file_extent_cluster(inode, cluster);
3203         if (ret)
3204                 goto out;
3205
3206         file_ra_state_init(ra, inode->i_mapping);
3207
3208         ret = setup_extent_mapping(inode, cluster->start - offset,
3209                                    cluster->end - offset, cluster->start);
3210         if (ret)
3211                 goto out;
3212
3213         index = (cluster->start - offset) >> PAGE_SHIFT;
3214         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3215         while (index <= last_index) {
3216                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3217                                 PAGE_SIZE);
3218                 if (ret)
3219                         goto out;
3220
3221                 page = find_lock_page(inode->i_mapping, index);
3222                 if (!page) {
3223                         page_cache_sync_readahead(inode->i_mapping,
3224                                                   ra, NULL, index,
3225                                                   last_index + 1 - index);
3226                         page = find_or_create_page(inode->i_mapping, index,
3227                                                    mask);
3228                         if (!page) {
3229                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3230                                                         PAGE_SIZE, true);
3231                                 ret = -ENOMEM;
3232                                 goto out;
3233                         }
3234                 }
3235
3236                 if (PageReadahead(page)) {
3237                         page_cache_async_readahead(inode->i_mapping,
3238                                                    ra, NULL, page, index,
3239                                                    last_index + 1 - index);
3240                 }
3241
3242                 if (!PageUptodate(page)) {
3243                         btrfs_readpage(NULL, page);
3244                         lock_page(page);
3245                         if (!PageUptodate(page)) {
3246                                 unlock_page(page);
3247                                 put_page(page);
3248                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3249                                                         PAGE_SIZE, true);
3250                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
3251                                                                PAGE_SIZE, true);
3252                                 ret = -EIO;
3253                                 goto out;
3254                         }
3255                 }
3256
3257                 page_start = page_offset(page);
3258                 page_end = page_start + PAGE_SIZE - 1;
3259
3260                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3261
3262                 set_page_extent_mapped(page);
3263
3264                 if (nr < cluster->nr &&
3265                     page_start + offset == cluster->boundary[nr]) {
3266                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3267                                         page_start, page_end,
3268                                         EXTENT_BOUNDARY);
3269                         nr++;
3270                 }
3271
3272                 ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
3273                                                 NULL, 0);
3274                 if (ret) {
3275                         unlock_page(page);
3276                         put_page(page);
3277                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3278                                                          PAGE_SIZE, true);
3279                         btrfs_delalloc_release_extents(BTRFS_I(inode),
3280                                                        PAGE_SIZE, true);
3281
3282                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
3283                                           page_start, page_end,
3284                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
3285                         goto out;
3286
3287                 }
3288                 set_page_dirty(page);
3289
3290                 unlock_extent(&BTRFS_I(inode)->io_tree,
3291                               page_start, page_end);
3292                 unlock_page(page);
3293                 put_page(page);
3294
3295                 index++;
3296                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE,
3297                                                false);
3298                 balance_dirty_pages_ratelimited(inode->i_mapping);
3299                 btrfs_throttle(fs_info);
3300         }
3301         WARN_ON(nr != cluster->nr);
3302 out:
3303         kfree(ra);
3304         return ret;
3305 }
3306
3307 static noinline_for_stack
3308 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3309                          struct file_extent_cluster *cluster)
3310 {
3311         int ret;
3312
3313         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3314                 ret = relocate_file_extent_cluster(inode, cluster);
3315                 if (ret)
3316                         return ret;
3317                 cluster->nr = 0;
3318         }
3319
3320         if (!cluster->nr)
3321                 cluster->start = extent_key->objectid;
3322         else
3323                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3324         cluster->end = extent_key->objectid + extent_key->offset - 1;
3325         cluster->boundary[cluster->nr] = extent_key->objectid;
3326         cluster->nr++;
3327
3328         if (cluster->nr >= MAX_EXTENTS) {
3329                 ret = relocate_file_extent_cluster(inode, cluster);
3330                 if (ret)
3331                         return ret;
3332                 cluster->nr = 0;
3333         }
3334         return 0;
3335 }
3336
3337 /*
3338  * helper to add a tree block to the list.
3339  * the major work is getting the generation and level of the block
3340  */
3341 static int add_tree_block(struct reloc_control *rc,
3342                           struct btrfs_key *extent_key,
3343                           struct btrfs_path *path,
3344                           struct rb_root *blocks)
3345 {
3346         struct extent_buffer *eb;
3347         struct btrfs_extent_item *ei;
3348         struct btrfs_tree_block_info *bi;
3349         struct tree_block *block;
3350         struct rb_node *rb_node;
3351         u32 item_size;
3352         int level = -1;
3353         u64 generation;
3354
3355         eb =  path->nodes[0];
3356         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3357
3358         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3359             item_size >= sizeof(*ei) + sizeof(*bi)) {
3360                 ei = btrfs_item_ptr(eb, path->slots[0],
3361                                 struct btrfs_extent_item);
3362                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3363                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3364                         level = btrfs_tree_block_level(eb, bi);
3365                 } else {
3366                         level = (int)extent_key->offset;
3367                 }
3368                 generation = btrfs_extent_generation(eb, ei);
3369         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3370                 btrfs_print_v0_err(eb->fs_info);
3371                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3372                 return -EINVAL;
3373         } else {
3374                 BUG();
3375         }
3376
3377         btrfs_release_path(path);
3378
3379         BUG_ON(level == -1);
3380
3381         block = kmalloc(sizeof(*block), GFP_NOFS);
3382         if (!block)
3383                 return -ENOMEM;
3384
3385         block->bytenr = extent_key->objectid;
3386         block->key.objectid = rc->extent_root->fs_info->nodesize;
3387         block->key.offset = generation;
3388         block->level = level;
3389         block->key_ready = 0;
3390
3391         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3392         if (rb_node)
3393                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3394
3395         return 0;
3396 }
3397
3398 /*
3399  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3400  */
3401 static int __add_tree_block(struct reloc_control *rc,
3402                             u64 bytenr, u32 blocksize,
3403                             struct rb_root *blocks)
3404 {
3405         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3406         struct btrfs_path *path;
3407         struct btrfs_key key;
3408         int ret;
3409         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3410
3411         if (tree_block_processed(bytenr, rc))
3412                 return 0;
3413
3414         if (tree_search(blocks, bytenr))
3415                 return 0;
3416
3417         path = btrfs_alloc_path();
3418         if (!path)
3419                 return -ENOMEM;
3420 again:
3421         key.objectid = bytenr;
3422         if (skinny) {
3423                 key.type = BTRFS_METADATA_ITEM_KEY;
3424                 key.offset = (u64)-1;
3425         } else {
3426                 key.type = BTRFS_EXTENT_ITEM_KEY;
3427                 key.offset = blocksize;
3428         }
3429
3430         path->search_commit_root = 1;
3431         path->skip_locking = 1;
3432         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3433         if (ret < 0)
3434                 goto out;
3435
3436         if (ret > 0 && skinny) {
3437                 if (path->slots[0]) {
3438                         path->slots[0]--;
3439                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3440                                               path->slots[0]);
3441                         if (key.objectid == bytenr &&
3442                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3443                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3444                               key.offset == blocksize)))
3445                                 ret = 0;
3446                 }
3447
3448                 if (ret) {
3449                         skinny = false;
3450                         btrfs_release_path(path);
3451                         goto again;
3452                 }
3453         }
3454         if (ret) {
3455                 ASSERT(ret == 1);
3456                 btrfs_print_leaf(path->nodes[0]);
3457                 btrfs_err(fs_info,
3458              "tree block extent item (%llu) is not found in extent tree",
3459                      bytenr);
3460                 WARN_ON(1);
3461                 ret = -EINVAL;
3462                 goto out;
3463         }
3464
3465         ret = add_tree_block(rc, &key, path, blocks);
3466 out:
3467         btrfs_free_path(path);
3468         return ret;
3469 }
3470
3471 /*
3472  * helper to check if the block use full backrefs for pointers in it
3473  */
3474 static int block_use_full_backref(struct reloc_control *rc,
3475                                   struct extent_buffer *eb)
3476 {
3477         u64 flags;
3478         int ret;
3479
3480         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3481             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3482                 return 1;
3483
3484         ret = btrfs_lookup_extent_info(NULL, rc->extent_root->fs_info,
3485                                        eb->start, btrfs_header_level(eb), 1,
3486                                        NULL, &flags);
3487         BUG_ON(ret);
3488
3489         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3490                 ret = 1;
3491         else
3492                 ret = 0;
3493         return ret;
3494 }
3495
3496 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3497                                     struct btrfs_block_group_cache *block_group,
3498                                     struct inode *inode,
3499                                     u64 ino)
3500 {
3501         struct btrfs_key key;
3502         struct btrfs_root *root = fs_info->tree_root;
3503         struct btrfs_trans_handle *trans;
3504         int ret = 0;
3505
3506         if (inode)
3507                 goto truncate;
3508
3509         key.objectid = ino;
3510         key.type = BTRFS_INODE_ITEM_KEY;
3511         key.offset = 0;
3512
3513         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3514         if (IS_ERR(inode))
3515                 return -ENOENT;
3516
3517 truncate:
3518         ret = btrfs_check_trunc_cache_free_space(fs_info,
3519                                                  &fs_info->global_block_rsv);
3520         if (ret)
3521                 goto out;
3522
3523         trans = btrfs_join_transaction(root);
3524         if (IS_ERR(trans)) {
3525                 ret = PTR_ERR(trans);
3526                 goto out;
3527         }
3528
3529         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3530
3531         btrfs_end_transaction(trans);
3532         btrfs_btree_balance_dirty(fs_info);
3533 out:
3534         iput(inode);
3535         return ret;
3536 }
3537
3538 /*
3539  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3540  * this function scans fs tree to find blocks reference the data extent
3541  */
3542 static int find_data_references(struct reloc_control *rc,
3543                                 struct btrfs_key *extent_key,
3544                                 struct extent_buffer *leaf,
3545                                 struct btrfs_extent_data_ref *ref,
3546                                 struct rb_root *blocks)
3547 {
3548         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3549         struct btrfs_path *path;
3550         struct tree_block *block;
3551         struct btrfs_root *root;
3552         struct btrfs_file_extent_item *fi;
3553         struct rb_node *rb_node;
3554         struct btrfs_key key;
3555         u64 ref_root;
3556         u64 ref_objectid;
3557         u64 ref_offset;
3558         u32 ref_count;
3559         u32 nritems;
3560         int err = 0;
3561         int added = 0;
3562         int counted;
3563         int ret;
3564
3565         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3566         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3567         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3568         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3569
3570         /*
3571          * This is an extent belonging to the free space cache, lets just delete
3572          * it and redo the search.
3573          */
3574         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3575                 ret = delete_block_group_cache(fs_info, rc->block_group,
3576                                                NULL, ref_objectid);
3577                 if (ret != -ENOENT)
3578                         return ret;
3579                 ret = 0;
3580         }
3581
3582         path = btrfs_alloc_path();
3583         if (!path)
3584                 return -ENOMEM;
3585         path->reada = READA_FORWARD;
3586
3587         root = read_fs_root(fs_info, ref_root);
3588         if (IS_ERR(root)) {
3589                 err = PTR_ERR(root);
3590                 goto out;
3591         }
3592
3593         key.objectid = ref_objectid;
3594         key.type = BTRFS_EXTENT_DATA_KEY;
3595         if (ref_offset > ((u64)-1 << 32))
3596                 key.offset = 0;
3597         else
3598                 key.offset = ref_offset;
3599
3600         path->search_commit_root = 1;
3601         path->skip_locking = 1;
3602         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3603         if (ret < 0) {
3604                 err = ret;
3605                 goto out;
3606         }
3607
3608         leaf = path->nodes[0];
3609         nritems = btrfs_header_nritems(leaf);
3610         /*
3611          * the references in tree blocks that use full backrefs
3612          * are not counted in
3613          */
3614         if (block_use_full_backref(rc, leaf))
3615                 counted = 0;
3616         else
3617                 counted = 1;
3618         rb_node = tree_search(blocks, leaf->start);
3619         if (rb_node) {
3620                 if (counted)
3621                         added = 1;
3622                 else
3623                         path->slots[0] = nritems;
3624         }
3625
3626         while (ref_count > 0) {
3627                 while (path->slots[0] >= nritems) {
3628                         ret = btrfs_next_leaf(root, path);
3629                         if (ret < 0) {
3630                                 err = ret;
3631                                 goto out;
3632                         }
3633                         if (WARN_ON(ret > 0))
3634                                 goto out;
3635
3636                         leaf = path->nodes[0];
3637                         nritems = btrfs_header_nritems(leaf);
3638                         added = 0;
3639
3640                         if (block_use_full_backref(rc, leaf))
3641                                 counted = 0;
3642                         else
3643                                 counted = 1;
3644                         rb_node = tree_search(blocks, leaf->start);
3645                         if (rb_node) {
3646                                 if (counted)
3647                                         added = 1;
3648                                 else
3649                                         path->slots[0] = nritems;
3650                         }
3651                 }
3652
3653                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3654                 if (WARN_ON(key.objectid != ref_objectid ||
3655                     key.type != BTRFS_EXTENT_DATA_KEY))
3656                         break;
3657
3658                 fi = btrfs_item_ptr(leaf, path->slots[0],
3659                                     struct btrfs_file_extent_item);
3660
3661                 if (btrfs_file_extent_type(leaf, fi) ==
3662                     BTRFS_FILE_EXTENT_INLINE)
3663                         goto next;
3664
3665                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3666                     extent_key->objectid)
3667                         goto next;
3668
3669                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3670                 if (key.offset != ref_offset)
3671                         goto next;
3672
3673                 if (counted)
3674                         ref_count--;
3675                 if (added)
3676                         goto next;
3677
3678                 if (!tree_block_processed(leaf->start, rc)) {
3679                         block = kmalloc(sizeof(*block), GFP_NOFS);
3680                         if (!block) {
3681                                 err = -ENOMEM;
3682                                 break;
3683                         }
3684                         block->bytenr = leaf->start;
3685                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3686                         block->level = 0;
3687                         block->key_ready = 1;
3688                         rb_node = tree_insert(blocks, block->bytenr,
3689                                               &block->rb_node);
3690                         if (rb_node)
3691                                 backref_tree_panic(rb_node, -EEXIST,
3692                                                    block->bytenr);
3693                 }
3694                 if (counted)
3695                         added = 1;
3696                 else
3697                         path->slots[0] = nritems;
3698 next:
3699                 path->slots[0]++;
3700
3701         }
3702 out:
3703         btrfs_free_path(path);
3704         return err;
3705 }
3706
3707 /*
3708  * helper to find all tree blocks that reference a given data extent
3709  */
3710 static noinline_for_stack
3711 int add_data_references(struct reloc_control *rc,
3712                         struct btrfs_key *extent_key,
3713                         struct btrfs_path *path,
3714                         struct rb_root *blocks)
3715 {
3716         struct btrfs_key key;
3717         struct extent_buffer *eb;
3718         struct btrfs_extent_data_ref *dref;
3719         struct btrfs_extent_inline_ref *iref;
3720         unsigned long ptr;
3721         unsigned long end;
3722         u32 blocksize = rc->extent_root->fs_info->nodesize;
3723         int ret = 0;
3724         int err = 0;
3725
3726         eb = path->nodes[0];
3727         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3728         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3729         ptr += sizeof(struct btrfs_extent_item);
3730
3731         while (ptr < end) {
3732                 iref = (struct btrfs_extent_inline_ref *)ptr;
3733                 key.type = btrfs_get_extent_inline_ref_type(eb, iref,
3734                                                         BTRFS_REF_TYPE_DATA);
3735                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3736                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3737                         ret = __add_tree_block(rc, key.offset, blocksize,
3738                                                blocks);
3739                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3740                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3741                         ret = find_data_references(rc, extent_key,
3742                                                    eb, dref, blocks);
3743                 } else {
3744                         ret = -EUCLEAN;
3745                         btrfs_err(rc->extent_root->fs_info,
3746                      "extent %llu slot %d has an invalid inline ref type",
3747                              eb->start, path->slots[0]);
3748                 }
3749                 if (ret) {
3750                         err = ret;
3751                         goto out;
3752                 }
3753                 ptr += btrfs_extent_inline_ref_size(key.type);
3754         }
3755         WARN_ON(ptr > end);
3756
3757         while (1) {
3758                 cond_resched();
3759                 eb = path->nodes[0];
3760                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3761                         ret = btrfs_next_leaf(rc->extent_root, path);
3762                         if (ret < 0) {
3763                                 err = ret;
3764                                 break;
3765                         }
3766                         if (ret > 0)
3767                                 break;
3768                         eb = path->nodes[0];
3769                 }
3770
3771                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3772                 if (key.objectid != extent_key->objectid)
3773                         break;
3774
3775                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3776                         ret = __add_tree_block(rc, key.offset, blocksize,
3777                                                blocks);
3778                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3779                         dref = btrfs_item_ptr(eb, path->slots[0],
3780                                               struct btrfs_extent_data_ref);
3781                         ret = find_data_references(rc, extent_key,
3782                                                    eb, dref, blocks);
3783                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
3784                         btrfs_print_v0_err(eb->fs_info);
3785                         btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3786                         ret = -EINVAL;
3787                 } else {
3788                         ret = 0;
3789                 }
3790                 if (ret) {
3791                         err = ret;
3792                         break;
3793                 }
3794                 path->slots[0]++;
3795         }
3796 out:
3797         btrfs_release_path(path);
3798         if (err)
3799                 free_block_list(blocks);
3800         return err;
3801 }
3802
3803 /*
3804  * helper to find next unprocessed extent
3805  */
3806 static noinline_for_stack
3807 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3808                      struct btrfs_key *extent_key)
3809 {
3810         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3811         struct btrfs_key key;
3812         struct extent_buffer *leaf;
3813         u64 start, end, last;
3814         int ret;
3815
3816         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3817         while (1) {
3818                 cond_resched();
3819                 if (rc->search_start >= last) {
3820                         ret = 1;
3821                         break;
3822                 }
3823
3824                 key.objectid = rc->search_start;
3825                 key.type = BTRFS_EXTENT_ITEM_KEY;
3826                 key.offset = 0;
3827
3828                 path->search_commit_root = 1;
3829                 path->skip_locking = 1;
3830                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3831                                         0, 0);
3832                 if (ret < 0)
3833                         break;
3834 next:
3835                 leaf = path->nodes[0];
3836                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3837                         ret = btrfs_next_leaf(rc->extent_root, path);
3838                         if (ret != 0)
3839                                 break;
3840                         leaf = path->nodes[0];
3841                 }
3842
3843                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3844                 if (key.objectid >= last) {
3845                         ret = 1;
3846                         break;
3847                 }
3848
3849                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3850                     key.type != BTRFS_METADATA_ITEM_KEY) {
3851                         path->slots[0]++;
3852                         goto next;
3853                 }
3854
3855                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3856                     key.objectid + key.offset <= rc->search_start) {
3857                         path->slots[0]++;
3858                         goto next;
3859                 }
3860
3861                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3862                     key.objectid + fs_info->nodesize <=
3863                     rc->search_start) {
3864                         path->slots[0]++;
3865                         goto next;
3866                 }
3867
3868                 ret = find_first_extent_bit(&rc->processed_blocks,
3869                                             key.objectid, &start, &end,
3870                                             EXTENT_DIRTY, NULL);
3871
3872                 if (ret == 0 && start <= key.objectid) {
3873                         btrfs_release_path(path);
3874                         rc->search_start = end + 1;
3875                 } else {
3876                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3877                                 rc->search_start = key.objectid + key.offset;
3878                         else
3879                                 rc->search_start = key.objectid +
3880                                         fs_info->nodesize;
3881                         memcpy(extent_key, &key, sizeof(key));
3882                         return 0;
3883                 }
3884         }
3885         btrfs_release_path(path);
3886         return ret;
3887 }
3888
3889 static void set_reloc_control(struct reloc_control *rc)
3890 {
3891         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3892
3893         mutex_lock(&fs_info->reloc_mutex);
3894         fs_info->reloc_ctl = rc;
3895         mutex_unlock(&fs_info->reloc_mutex);
3896 }
3897
3898 static void unset_reloc_control(struct reloc_control *rc)
3899 {
3900         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3901
3902         mutex_lock(&fs_info->reloc_mutex);
3903         fs_info->reloc_ctl = NULL;
3904         mutex_unlock(&fs_info->reloc_mutex);
3905 }
3906
3907 static int check_extent_flags(u64 flags)
3908 {
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_EXTENT_FLAG_TREE_BLOCK))
3914                 return 1;
3915         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3916             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3917                 return 1;
3918         return 0;
3919 }
3920
3921 static noinline_for_stack
3922 int prepare_to_relocate(struct reloc_control *rc)
3923 {
3924         struct btrfs_trans_handle *trans;
3925         int ret;
3926
3927         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3928                                               BTRFS_BLOCK_RSV_TEMP);
3929         if (!rc->block_rsv)
3930                 return -ENOMEM;
3931
3932         memset(&rc->cluster, 0, sizeof(rc->cluster));
3933         rc->search_start = rc->block_group->key.objectid;
3934         rc->extents_found = 0;
3935         rc->nodes_relocated = 0;
3936         rc->merging_rsv_size = 0;
3937         rc->reserved_bytes = 0;
3938         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3939                               RELOCATION_RESERVED_NODES;
3940         ret = btrfs_block_rsv_refill(rc->extent_root,
3941                                      rc->block_rsv, rc->block_rsv->size,
3942                                      BTRFS_RESERVE_FLUSH_ALL);
3943         if (ret)
3944                 return ret;
3945
3946         rc->create_reloc_tree = 1;
3947         set_reloc_control(rc);
3948
3949         trans = btrfs_join_transaction(rc->extent_root);
3950         if (IS_ERR(trans)) {
3951                 unset_reloc_control(rc);
3952                 /*
3953                  * extent tree is not a ref_cow tree and has no reloc_root to
3954                  * cleanup.  And callers are responsible to free the above
3955                  * block rsv.
3956                  */
3957                 return PTR_ERR(trans);
3958         }
3959         btrfs_commit_transaction(trans);
3960         return 0;
3961 }
3962
3963 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3964 {
3965         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3966         struct rb_root blocks = RB_ROOT;
3967         struct btrfs_key key;
3968         struct btrfs_trans_handle *trans = NULL;
3969         struct btrfs_path *path;
3970         struct btrfs_extent_item *ei;
3971         u64 flags;
3972         u32 item_size;
3973         int ret;
3974         int err = 0;
3975         int progress = 0;
3976
3977         path = btrfs_alloc_path();
3978         if (!path)
3979                 return -ENOMEM;
3980         path->reada = READA_FORWARD;
3981
3982         ret = prepare_to_relocate(rc);
3983         if (ret) {
3984                 err = ret;
3985                 goto out_free;
3986         }
3987
3988         while (1) {
3989                 rc->reserved_bytes = 0;
3990                 ret = btrfs_block_rsv_refill(rc->extent_root,
3991                                         rc->block_rsv, rc->block_rsv->size,
3992                                         BTRFS_RESERVE_FLUSH_ALL);
3993                 if (ret) {
3994                         err = ret;
3995                         break;
3996                 }
3997                 progress++;
3998                 trans = btrfs_start_transaction(rc->extent_root, 0);
3999                 if (IS_ERR(trans)) {
4000                         err = PTR_ERR(trans);
4001                         trans = NULL;
4002                         break;
4003                 }
4004 restart:
4005                 if (update_backref_cache(trans, &rc->backref_cache)) {
4006                         btrfs_end_transaction(trans);
4007                         trans = NULL;
4008                         continue;
4009                 }
4010
4011                 ret = find_next_extent(rc, path, &key);
4012                 if (ret < 0)
4013                         err = ret;
4014                 if (ret != 0)
4015                         break;
4016
4017                 rc->extents_found++;
4018
4019                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4020                                     struct btrfs_extent_item);
4021                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4022                 if (item_size >= sizeof(*ei)) {
4023                         flags = btrfs_extent_flags(path->nodes[0], ei);
4024                         ret = check_extent_flags(flags);
4025                         BUG_ON(ret);
4026                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
4027                         err = -EINVAL;
4028                         btrfs_print_v0_err(trans->fs_info);
4029                         btrfs_abort_transaction(trans, err);
4030                         break;
4031                 } else {
4032                         BUG();
4033                 }
4034
4035                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4036                         ret = add_tree_block(rc, &key, path, &blocks);
4037                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4038                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4039                         ret = add_data_references(rc, &key, path, &blocks);
4040                 } else {
4041                         btrfs_release_path(path);
4042                         ret = 0;
4043                 }
4044                 if (ret < 0) {
4045                         err = ret;
4046                         break;
4047                 }
4048
4049                 if (!RB_EMPTY_ROOT(&blocks)) {
4050                         ret = relocate_tree_blocks(trans, rc, &blocks);
4051                         if (ret < 0) {
4052                                 /*
4053                                  * if we fail to relocate tree blocks, force to update
4054                                  * backref cache when committing transaction.
4055                                  */
4056                                 rc->backref_cache.last_trans = trans->transid - 1;
4057
4058                                 if (ret != -EAGAIN) {
4059                                         err = ret;
4060                                         break;
4061                                 }
4062                                 rc->extents_found--;
4063                                 rc->search_start = key.objectid;
4064                         }
4065                 }
4066
4067                 btrfs_end_transaction_throttle(trans);
4068                 btrfs_btree_balance_dirty(fs_info);
4069                 trans = NULL;
4070
4071                 if (rc->stage == MOVE_DATA_EXTENTS &&
4072                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4073                         rc->found_file_extent = 1;
4074                         ret = relocate_data_extent(rc->data_inode,
4075                                                    &key, &rc->cluster);
4076                         if (ret < 0) {
4077                                 err = ret;
4078                                 break;
4079                         }
4080                 }
4081         }
4082         if (trans && progress && err == -ENOSPC) {
4083                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
4084                 if (ret == 1) {
4085                         err = 0;
4086                         progress = 0;
4087                         goto restart;
4088                 }
4089         }
4090
4091         btrfs_release_path(path);
4092         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4093
4094         if (trans) {
4095                 btrfs_end_transaction_throttle(trans);
4096                 btrfs_btree_balance_dirty(fs_info);
4097         }
4098
4099         if (!err) {
4100                 ret = relocate_file_extent_cluster(rc->data_inode,
4101                                                    &rc->cluster);
4102                 if (ret < 0)
4103                         err = ret;
4104         }
4105
4106         rc->create_reloc_tree = 0;
4107         set_reloc_control(rc);
4108
4109         backref_cache_cleanup(&rc->backref_cache);
4110         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4111
4112         err = prepare_to_merge(rc, err);
4113
4114         merge_reloc_roots(rc);
4115
4116         rc->merge_reloc_tree = 0;
4117         unset_reloc_control(rc);
4118         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4119
4120         /* get rid of pinned extents */
4121         trans = btrfs_join_transaction(rc->extent_root);
4122         if (IS_ERR(trans)) {
4123                 err = PTR_ERR(trans);
4124                 goto out_free;
4125         }
4126         btrfs_commit_transaction(trans);
4127         ret = clean_dirty_subvols(rc);
4128         if (ret < 0 && !err)
4129                 err = ret;
4130 out_free:
4131         btrfs_free_block_rsv(fs_info, rc->block_rsv);
4132         btrfs_free_path(path);
4133         return err;
4134 }
4135
4136 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4137                                  struct btrfs_root *root, u64 objectid)
4138 {
4139         struct btrfs_path *path;
4140         struct btrfs_inode_item *item;
4141         struct extent_buffer *leaf;
4142         int ret;
4143
4144         path = btrfs_alloc_path();
4145         if (!path)
4146                 return -ENOMEM;
4147
4148         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4149         if (ret)
4150                 goto out;
4151
4152         leaf = path->nodes[0];
4153         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4154         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
4155         btrfs_set_inode_generation(leaf, item, 1);
4156         btrfs_set_inode_size(leaf, item, 0);
4157         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4158         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4159                                           BTRFS_INODE_PREALLOC);
4160         btrfs_mark_buffer_dirty(leaf);
4161 out:
4162         btrfs_free_path(path);
4163         return ret;
4164 }
4165
4166 /*
4167  * helper to create inode for data relocation.
4168  * the inode is in data relocation tree and its link count is 0
4169  */
4170 static noinline_for_stack
4171 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4172                                  struct btrfs_block_group_cache *group)
4173 {
4174         struct inode *inode = NULL;
4175         struct btrfs_trans_handle *trans;
4176         struct btrfs_root *root;
4177         struct btrfs_key key;
4178         u64 objectid;
4179         int err = 0;
4180
4181         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4182         if (IS_ERR(root))
4183                 return ERR_CAST(root);
4184
4185         trans = btrfs_start_transaction(root, 6);
4186         if (IS_ERR(trans))
4187                 return ERR_CAST(trans);
4188
4189         err = btrfs_find_free_objectid(root, &objectid);
4190         if (err)
4191                 goto out;
4192
4193         err = __insert_orphan_inode(trans, root, objectid);
4194         BUG_ON(err);
4195
4196         key.objectid = objectid;
4197         key.type = BTRFS_INODE_ITEM_KEY;
4198         key.offset = 0;
4199         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4200         BUG_ON(IS_ERR(inode));
4201         BTRFS_I(inode)->index_cnt = group->key.objectid;
4202
4203         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4204 out:
4205         btrfs_end_transaction(trans);
4206         btrfs_btree_balance_dirty(fs_info);
4207         if (err) {
4208                 if (inode)
4209                         iput(inode);
4210                 inode = ERR_PTR(err);
4211         }
4212         return inode;
4213 }
4214
4215 static struct reloc_control *alloc_reloc_control(void)
4216 {
4217         struct reloc_control *rc;
4218
4219         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4220         if (!rc)
4221                 return NULL;
4222
4223         INIT_LIST_HEAD(&rc->reloc_roots);
4224         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
4225         backref_cache_init(&rc->backref_cache);
4226         mapping_tree_init(&rc->reloc_root_tree);
4227         extent_io_tree_init(&rc->processed_blocks, NULL);
4228         return rc;
4229 }
4230
4231 /*
4232  * Print the block group being relocated
4233  */
4234 static void describe_relocation(struct btrfs_fs_info *fs_info,
4235                                 struct btrfs_block_group_cache *block_group)
4236 {
4237         char buf[128] = {'\0'};
4238
4239         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
4240
4241         btrfs_info(fs_info,
4242                    "relocating block group %llu flags %s",
4243                    block_group->key.objectid, buf);
4244 }
4245
4246 /*
4247  * function to relocate all extents in a block group.
4248  */
4249 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4250 {
4251         struct btrfs_block_group_cache *bg;
4252         struct btrfs_root *extent_root = fs_info->extent_root;
4253         struct reloc_control *rc;
4254         struct inode *inode;
4255         struct btrfs_path *path;
4256         int ret;
4257         int rw = 0;
4258         int err = 0;
4259
4260         bg = btrfs_lookup_block_group(fs_info, group_start);
4261         if (!bg)
4262                 return -ENOENT;
4263
4264         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4265                 btrfs_put_block_group(bg);
4266                 return -ETXTBSY;
4267         }
4268
4269         rc = alloc_reloc_control();
4270         if (!rc) {
4271                 btrfs_put_block_group(bg);
4272                 return -ENOMEM;
4273         }
4274
4275         rc->extent_root = extent_root;
4276         rc->block_group = bg;
4277
4278         ret = btrfs_inc_block_group_ro(rc->block_group);
4279         if (ret) {
4280                 err = ret;
4281                 goto out;
4282         }
4283         rw = 1;
4284
4285         path = btrfs_alloc_path();
4286         if (!path) {
4287                 err = -ENOMEM;
4288                 goto out;
4289         }
4290
4291         inode = lookup_free_space_inode(fs_info, rc->block_group, path);
4292         btrfs_free_path(path);
4293
4294         if (!IS_ERR(inode))
4295                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4296         else
4297                 ret = PTR_ERR(inode);
4298
4299         if (ret && ret != -ENOENT) {
4300                 err = ret;
4301                 goto out;
4302         }
4303
4304         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4305         if (IS_ERR(rc->data_inode)) {
4306                 err = PTR_ERR(rc->data_inode);
4307                 rc->data_inode = NULL;
4308                 goto out;
4309         }
4310
4311         describe_relocation(fs_info, rc->block_group);
4312
4313         btrfs_wait_block_group_reservations(rc->block_group);
4314         btrfs_wait_nocow_writers(rc->block_group);
4315         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4316                                  rc->block_group->key.objectid,
4317                                  rc->block_group->key.offset);
4318
4319         while (1) {
4320                 mutex_lock(&fs_info->cleaner_mutex);
4321                 ret = relocate_block_group(rc);
4322                 mutex_unlock(&fs_info->cleaner_mutex);
4323                 if (ret < 0) {
4324                         err = ret;
4325                         goto out;
4326                 }
4327
4328                 if (rc->extents_found == 0)
4329                         break;
4330
4331                 btrfs_info(fs_info, "found %llu extents", rc->extents_found);
4332
4333                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4334                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4335                                                        (u64)-1);
4336                         if (ret) {
4337                                 err = ret;
4338                                 goto out;
4339                         }
4340                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4341                                                  0, -1);
4342                         rc->stage = UPDATE_DATA_PTRS;
4343                 }
4344         }
4345
4346         WARN_ON(rc->block_group->pinned > 0);
4347         WARN_ON(rc->block_group->reserved > 0);
4348         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4349 out:
4350         if (err && rw)
4351                 btrfs_dec_block_group_ro(rc->block_group);
4352         iput(rc->data_inode);
4353         btrfs_put_block_group(rc->block_group);
4354         kfree(rc);
4355         return err;
4356 }
4357
4358 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4359 {
4360         struct btrfs_fs_info *fs_info = root->fs_info;
4361         struct btrfs_trans_handle *trans;
4362         int ret, err;
4363
4364         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4365         if (IS_ERR(trans))
4366                 return PTR_ERR(trans);
4367
4368         memset(&root->root_item.drop_progress, 0,
4369                 sizeof(root->root_item.drop_progress));
4370         root->root_item.drop_level = 0;
4371         btrfs_set_root_refs(&root->root_item, 0);
4372         ret = btrfs_update_root(trans, fs_info->tree_root,
4373                                 &root->root_key, &root->root_item);
4374
4375         err = btrfs_end_transaction(trans);
4376         if (err)
4377                 return err;
4378         return ret;
4379 }
4380
4381 /*
4382  * recover relocation interrupted by system crash.
4383  *
4384  * this function resumes merging reloc trees with corresponding fs trees.
4385  * this is important for keeping the sharing of tree blocks
4386  */
4387 int btrfs_recover_relocation(struct btrfs_root *root)
4388 {
4389         struct btrfs_fs_info *fs_info = root->fs_info;
4390         LIST_HEAD(reloc_roots);
4391         struct btrfs_key key;
4392         struct btrfs_root *fs_root;
4393         struct btrfs_root *reloc_root;
4394         struct btrfs_path *path;
4395         struct extent_buffer *leaf;
4396         struct reloc_control *rc = NULL;
4397         struct btrfs_trans_handle *trans;
4398         int ret;
4399         int err = 0;
4400
4401         path = btrfs_alloc_path();
4402         if (!path)
4403                 return -ENOMEM;
4404         path->reada = READA_BACK;
4405
4406         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4407         key.type = BTRFS_ROOT_ITEM_KEY;
4408         key.offset = (u64)-1;
4409
4410         while (1) {
4411                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4412                                         path, 0, 0);
4413                 if (ret < 0) {
4414                         err = ret;
4415                         goto out;
4416                 }
4417                 if (ret > 0) {
4418                         if (path->slots[0] == 0)
4419                                 break;
4420                         path->slots[0]--;
4421                 }
4422                 leaf = path->nodes[0];
4423                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4424                 btrfs_release_path(path);
4425
4426                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4427                     key.type != BTRFS_ROOT_ITEM_KEY)
4428                         break;
4429
4430                 reloc_root = btrfs_read_fs_root(root, &key);
4431                 if (IS_ERR(reloc_root)) {
4432                         err = PTR_ERR(reloc_root);
4433                         goto out;
4434                 }
4435
4436                 list_add(&reloc_root->root_list, &reloc_roots);
4437
4438                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4439                         fs_root = read_fs_root(fs_info,
4440                                                reloc_root->root_key.offset);
4441                         if (IS_ERR(fs_root)) {
4442                                 ret = PTR_ERR(fs_root);
4443                                 if (ret != -ENOENT) {
4444                                         err = ret;
4445                                         goto out;
4446                                 }
4447                                 ret = mark_garbage_root(reloc_root);
4448                                 if (ret < 0) {
4449                                         err = ret;
4450                                         goto out;
4451                                 }
4452                         }
4453                 }
4454
4455                 if (key.offset == 0)
4456                         break;
4457
4458                 key.offset--;
4459         }
4460         btrfs_release_path(path);
4461
4462         if (list_empty(&reloc_roots))
4463                 goto out;
4464
4465         rc = alloc_reloc_control();
4466         if (!rc) {
4467                 err = -ENOMEM;
4468                 goto out;
4469         }
4470
4471         rc->extent_root = fs_info->extent_root;
4472
4473         set_reloc_control(rc);
4474
4475         trans = btrfs_join_transaction(rc->extent_root);
4476         if (IS_ERR(trans)) {
4477                 unset_reloc_control(rc);
4478                 err = PTR_ERR(trans);
4479                 goto out_free;
4480         }
4481
4482         rc->merge_reloc_tree = 1;
4483
4484         while (!list_empty(&reloc_roots)) {
4485                 reloc_root = list_entry(reloc_roots.next,
4486                                         struct btrfs_root, root_list);
4487                 list_del(&reloc_root->root_list);
4488
4489                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4490                         list_add_tail(&reloc_root->root_list,
4491                                       &rc->reloc_roots);
4492                         continue;
4493                 }
4494
4495                 fs_root = read_fs_root(fs_info, reloc_root->root_key.offset);
4496                 if (IS_ERR(fs_root)) {
4497                         err = PTR_ERR(fs_root);
4498                         goto out_free;
4499                 }
4500
4501                 err = __add_reloc_root(reloc_root);
4502                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4503                 fs_root->reloc_root = reloc_root;
4504         }
4505
4506         err = btrfs_commit_transaction(trans);
4507         if (err)
4508                 goto out_free;
4509
4510         merge_reloc_roots(rc);
4511
4512         unset_reloc_control(rc);
4513
4514         trans = btrfs_join_transaction(rc->extent_root);
4515         if (IS_ERR(trans)) {
4516                 err = PTR_ERR(trans);
4517                 goto out_free;
4518         }
4519         err = btrfs_commit_transaction(trans);
4520
4521         ret = clean_dirty_subvols(rc);
4522         if (ret < 0 && !err)
4523                 err = ret;
4524 out_free:
4525         kfree(rc);
4526 out:
4527         if (!list_empty(&reloc_roots))
4528                 free_reloc_roots(&reloc_roots);
4529
4530         btrfs_free_path(path);
4531
4532         if (err == 0) {
4533                 /* cleanup orphan inode in data relocation tree */
4534                 fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4535                 if (IS_ERR(fs_root))
4536                         err = PTR_ERR(fs_root);
4537                 else
4538                         err = btrfs_orphan_cleanup(fs_root);
4539         }
4540         return err;
4541 }
4542
4543 /*
4544  * helper to add ordered checksum for data relocation.
4545  *
4546  * cloning checksum properly handles the nodatasum extents.
4547  * it also saves CPU time to re-calculate the checksum.
4548  */
4549 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4550 {
4551         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4552         struct btrfs_ordered_sum *sums;
4553         struct btrfs_ordered_extent *ordered;
4554         int ret;
4555         u64 disk_bytenr;
4556         u64 new_bytenr;
4557         LIST_HEAD(list);
4558
4559         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4560         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4561
4562         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4563         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4564                                        disk_bytenr + len - 1, &list, 0);
4565         if (ret)
4566                 goto out;
4567
4568         while (!list_empty(&list)) {
4569                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4570                 list_del_init(&sums->list);
4571
4572                 /*
4573                  * We need to offset the new_bytenr based on where the csum is.
4574                  * We need to do this because we will read in entire prealloc
4575                  * extents but we may have written to say the middle of the
4576                  * prealloc extent, so we need to make sure the csum goes with
4577                  * the right disk offset.
4578                  *
4579                  * We can do this because the data reloc inode refers strictly
4580                  * to the on disk bytes, so we don't have to worry about
4581                  * disk_len vs real len like with real inodes since it's all
4582                  * disk length.
4583                  */
4584                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4585                 sums->bytenr = new_bytenr;
4586
4587                 btrfs_add_ordered_sum(inode, ordered, sums);
4588         }
4589 out:
4590         btrfs_put_ordered_extent(ordered);
4591         return ret;
4592 }
4593
4594 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4595                           struct btrfs_root *root, struct extent_buffer *buf,
4596                           struct extent_buffer *cow)
4597 {
4598         struct btrfs_fs_info *fs_info = root->fs_info;
4599         struct reloc_control *rc;
4600         struct backref_node *node;
4601         int first_cow = 0;
4602         int level;
4603         int ret = 0;
4604
4605         rc = fs_info->reloc_ctl;
4606         if (!rc)
4607                 return 0;
4608
4609         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4610                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4611
4612         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4613                 if (buf == root->node)
4614                         __update_reloc_root(root, cow->start);
4615         }
4616
4617         level = btrfs_header_level(buf);
4618         if (btrfs_header_generation(buf) <=
4619             btrfs_root_last_snapshot(&root->root_item))
4620                 first_cow = 1;
4621
4622         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4623             rc->create_reloc_tree) {
4624                 WARN_ON(!first_cow && level == 0);
4625
4626                 node = rc->backref_cache.path[level];
4627                 BUG_ON(node->bytenr != buf->start &&
4628                        node->new_bytenr != buf->start);
4629
4630                 drop_node_buffer(node);
4631                 extent_buffer_get(cow);
4632                 node->eb = cow;
4633                 node->new_bytenr = cow->start;
4634
4635                 if (!node->pending) {
4636                         list_move_tail(&node->list,
4637                                        &rc->backref_cache.pending[level]);
4638                         node->pending = 1;
4639                 }
4640
4641                 if (first_cow)
4642                         __mark_block_processed(rc, node);
4643
4644                 if (first_cow && level > 0)
4645                         rc->nodes_relocated += buf->len;
4646         }
4647
4648         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4649                 ret = replace_file_extents(trans, rc, root, cow);
4650         return ret;
4651 }
4652
4653 /*
4654  * called before creating snapshot. it calculates metadata reservation
4655  * required for relocating tree blocks in the snapshot
4656  */
4657 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4658                               u64 *bytes_to_reserve)
4659 {
4660         struct btrfs_root *root;
4661         struct reloc_control *rc;
4662
4663         root = pending->root;
4664         if (!root->reloc_root)
4665                 return;
4666
4667         rc = root->fs_info->reloc_ctl;
4668         if (!rc->merge_reloc_tree)
4669                 return;
4670
4671         root = root->reloc_root;
4672         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4673         /*
4674          * relocation is in the stage of merging trees. the space
4675          * used by merging a reloc tree is twice the size of
4676          * relocated tree nodes in the worst case. half for cowing
4677          * the reloc tree, half for cowing the fs tree. the space
4678          * used by cowing the reloc tree will be freed after the
4679          * tree is dropped. if we create snapshot, cowing the fs
4680          * tree may use more space than it frees. so we need
4681          * reserve extra space.
4682          */
4683         *bytes_to_reserve += rc->nodes_relocated;
4684 }
4685
4686 /*
4687  * called after snapshot is created. migrate block reservation
4688  * and create reloc root for the newly created snapshot
4689  */
4690 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4691                                struct btrfs_pending_snapshot *pending)
4692 {
4693         struct btrfs_root *root = pending->root;
4694         struct btrfs_root *reloc_root;
4695         struct btrfs_root *new_root;
4696         struct reloc_control *rc;
4697         int ret;
4698
4699         if (!root->reloc_root)
4700                 return 0;
4701
4702         rc = root->fs_info->reloc_ctl;
4703         rc->merging_rsv_size += rc->nodes_relocated;
4704
4705         if (rc->merge_reloc_tree) {
4706                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4707                                               rc->block_rsv,
4708                                               rc->nodes_relocated, true);
4709                 if (ret)
4710                         return ret;
4711         }
4712
4713         new_root = pending->snap;
4714         reloc_root = create_reloc_root(trans, root->reloc_root,
4715                                        new_root->root_key.objectid);
4716         if (IS_ERR(reloc_root))
4717                 return PTR_ERR(reloc_root);
4718
4719         ret = __add_reloc_root(reloc_root);
4720         BUG_ON(ret < 0);
4721         new_root->reloc_root = reloc_root;
4722
4723         if (rc->create_reloc_tree)
4724                 ret = clone_backref_node(trans, rc, root, reloc_root);
4725         return ret;
4726 }