linux/fs/btrfs/relocation.c
<<
>>
Prefs
   1/*
   2 * Copyright (C) 2009 Oracle.  All rights reserved.
   3 *
   4 * This program is free software; you can redistribute it and/or
   5 * modify it under the terms of the GNU General Public
   6 * License v2 as published by the Free Software Foundation.
   7 *
   8 * This program is distributed in the hope that it will be useful,
   9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11 * General Public License for more details.
  12 *
  13 * You should have received a copy of the GNU General Public
  14 * License along with this program; if not, write to the
  15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16 * Boston, MA 021110-1307, USA.
  17 */
  18
  19#include <linux/sched.h>
  20#include <linux/pagemap.h>
  21#include <linux/writeback.h>
  22#include <linux/blkdev.h>
  23#include <linux/rbtree.h>
  24#include <linux/slab.h>
  25#include "ctree.h"
  26#include "disk-io.h"
  27#include "transaction.h"
  28#include "volumes.h"
  29#include "locking.h"
  30#include "btrfs_inode.h"
  31#include "async-thread.h"
  32#include "free-space-cache.h"
  33#include "inode-map.h"
  34#include "qgroup.h"
  35#include "print-tree.h"
  36
  37/*
  38 * backref_node, mapping_node and tree_block start with this
  39 */
  40struct tree_entry {
  41        struct rb_node rb_node;
  42        u64 bytenr;
  43};
  44
  45/*
  46 * present a tree block in the backref cache
  47 */
  48struct backref_node {
  49        struct rb_node rb_node;
  50        u64 bytenr;
  51
  52        u64 new_bytenr;
  53        /* objectid of tree block owner, can be not uptodate */
  54        u64 owner;
  55        /* link to pending, changed or detached list */
  56        struct list_head list;
  57        /* list of upper level blocks reference this block */
  58        struct list_head upper;
  59        /* list of child blocks in the cache */
  60        struct list_head lower;
  61        /* NULL if this node is not tree root */
  62        struct btrfs_root *root;
  63        /* extent buffer got by COW the block */
  64        struct extent_buffer *eb;
  65        /* level of tree block */
  66        unsigned int level:8;
  67        /* is the block in non-reference counted tree */
  68        unsigned int cowonly:1;
  69        /* 1 if no child node in the cache */
  70        unsigned int lowest:1;
  71        /* is the extent buffer locked */
  72        unsigned int locked:1;
  73        /* has the block been processed */
  74        unsigned int processed:1;
  75        /* have backrefs of this block been checked */
  76        unsigned int checked:1;
  77        /*
  78         * 1 if corresponding block has been cowed but some upper
  79         * level block pointers may not point to the new location
  80         */
  81        unsigned int pending:1;
  82        /*
  83         * 1 if the backref node isn't connected to any other
  84         * backref node.
  85         */
  86        unsigned int detached:1;
  87};
  88
  89/*
  90 * present a block pointer in the backref cache
  91 */
  92struct backref_edge {
  93        struct list_head list[2];
  94        struct backref_node *node[2];
  95};
  96
  97#define LOWER   0
  98#define UPPER   1
  99#define RELOCATION_RESERVED_NODES       256
 100
 101struct backref_cache {
 102        /* red black tree of all backref nodes in the cache */
 103        struct rb_root rb_root;
 104        /* for passing backref nodes to btrfs_reloc_cow_block */
 105        struct backref_node *path[BTRFS_MAX_LEVEL];
 106        /*
 107         * list of blocks that have been cowed but some block
 108         * pointers in upper level blocks may not reflect the
 109         * new location
 110         */
 111        struct list_head pending[BTRFS_MAX_LEVEL];
 112        /* list of backref nodes with no child node */
 113        struct list_head leaves;
 114        /* list of blocks that have been cowed in current transaction */
 115        struct list_head changed;
 116        /* list of detached backref node. */
 117        struct list_head detached;
 118
 119        u64 last_trans;
 120
 121        int nr_nodes;
 122        int nr_edges;
 123};
 124
 125/*
 126 * map address of tree root to tree
 127 */
 128struct mapping_node {
 129        struct rb_node rb_node;
 130        u64 bytenr;
 131        void *data;
 132};
 133
 134struct mapping_tree {
 135        struct rb_root rb_root;
 136        spinlock_t lock;
 137};
 138
 139/*
 140 * present a tree block to process
 141 */
 142struct tree_block {
 143        struct rb_node rb_node;
 144        u64 bytenr;
 145        struct btrfs_key key;
 146        unsigned int level:8;
 147        unsigned int key_ready:1;
 148};
 149
 150#define MAX_EXTENTS 128
 151
 152struct file_extent_cluster {
 153        u64 start;
 154        u64 end;
 155        u64 boundary[MAX_EXTENTS];
 156        unsigned int nr;
 157};
 158
 159struct reloc_control {
 160        /* block group to relocate */
 161        struct btrfs_block_group_cache *block_group;
 162        /* extent tree */
 163        struct btrfs_root *extent_root;
 164        /* inode for moving data */
 165        struct inode *data_inode;
 166
 167        struct btrfs_block_rsv *block_rsv;
 168
 169        struct backref_cache backref_cache;
 170
 171        struct file_extent_cluster cluster;
 172        /* tree blocks have been processed */
 173        struct extent_io_tree processed_blocks;
 174        /* map start of tree root to corresponding reloc tree */
 175        struct mapping_tree reloc_root_tree;
 176        /* list of reloc trees */
 177        struct list_head reloc_roots;
 178        /* size of metadata reservation for merging reloc trees */
 179        u64 merging_rsv_size;
 180        /* size of relocated tree nodes */
 181        u64 nodes_relocated;
 182        /* reserved size for block group relocation*/
 183        u64 reserved_bytes;
 184
 185        u64 search_start;
 186        u64 extents_found;
 187
 188        unsigned int stage:8;
 189        unsigned int create_reloc_tree:1;
 190        unsigned int merge_reloc_tree:1;
 191        unsigned int found_file_extent:1;
 192};
 193
 194/* stages of data relocation */
 195#define MOVE_DATA_EXTENTS       0
 196#define UPDATE_DATA_PTRS        1
 197
 198static void remove_backref_node(struct backref_cache *cache,
 199                                struct backref_node *node);
 200static void __mark_block_processed(struct reloc_control *rc,
 201                                   struct backref_node *node);
 202
 203static void mapping_tree_init(struct mapping_tree *tree)
 204{
 205        tree->rb_root = RB_ROOT;
 206        spin_lock_init(&tree->lock);
 207}
 208
 209static void backref_cache_init(struct backref_cache *cache)
 210{
 211        int i;
 212        cache->rb_root = RB_ROOT;
 213        for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 214                INIT_LIST_HEAD(&cache->pending[i]);
 215        INIT_LIST_HEAD(&cache->changed);
 216        INIT_LIST_HEAD(&cache->detached);
 217        INIT_LIST_HEAD(&cache->leaves);
 218}
 219
 220static void backref_cache_cleanup(struct backref_cache *cache)
 221{
 222        struct backref_node *node;
 223        int i;
 224
 225        while (!list_empty(&cache->detached)) {
 226                node = list_entry(cache->detached.next,
 227                                  struct backref_node, list);
 228                remove_backref_node(cache, node);
 229        }
 230
 231        while (!list_empty(&cache->leaves)) {
 232                node = list_entry(cache->leaves.next,
 233                                  struct backref_node, lower);
 234                remove_backref_node(cache, node);
 235        }
 236
 237        cache->last_trans = 0;
 238
 239        for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 240                ASSERT(list_empty(&cache->pending[i]));
 241        ASSERT(list_empty(&cache->changed));
 242        ASSERT(list_empty(&cache->detached));
 243        ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
 244        ASSERT(!cache->nr_nodes);
 245        ASSERT(!cache->nr_edges);
 246}
 247
 248static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 249{
 250        struct backref_node *node;
 251
 252        node = kzalloc(sizeof(*node), GFP_NOFS);
 253        if (node) {
 254                INIT_LIST_HEAD(&node->list);
 255                INIT_LIST_HEAD(&node->upper);
 256                INIT_LIST_HEAD(&node->lower);
 257                RB_CLEAR_NODE(&node->rb_node);
 258                cache->nr_nodes++;
 259        }
 260        return node;
 261}
 262
 263static void free_backref_node(struct backref_cache *cache,
 264                              struct backref_node *node)
 265{
 266        if (node) {
 267                cache->nr_nodes--;
 268                kfree(node);
 269        }
 270}
 271
 272static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 273{
 274        struct backref_edge *edge;
 275
 276        edge = kzalloc(sizeof(*edge), GFP_NOFS);
 277        if (edge)
 278                cache->nr_edges++;
 279        return edge;
 280}
 281
 282static void free_backref_edge(struct backref_cache *cache,
 283                              struct backref_edge *edge)
 284{
 285        if (edge) {
 286                cache->nr_edges--;
 287                kfree(edge);
 288        }
 289}
 290
 291static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 292                                   struct rb_node *node)
 293{
 294        struct rb_node **p = &root->rb_node;
 295        struct rb_node *parent = NULL;
 296        struct tree_entry *entry;
 297
 298        while (*p) {
 299                parent = *p;
 300                entry = rb_entry(parent, struct tree_entry, rb_node);
 301
 302                if (bytenr < entry->bytenr)
 303                        p = &(*p)->rb_left;
 304                else if (bytenr > entry->bytenr)
 305                        p = &(*p)->rb_right;
 306                else
 307                        return parent;
 308        }
 309
 310        rb_link_node(node, parent, p);
 311        rb_insert_color(node, root);
 312        return NULL;
 313}
 314
 315static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 316{
 317        struct rb_node *n = root->rb_node;
 318        struct tree_entry *entry;
 319
 320        while (n) {
 321                entry = rb_entry(n, struct tree_entry, rb_node);
 322
 323                if (bytenr < entry->bytenr)
 324                        n = n->rb_left;
 325                else if (bytenr > entry->bytenr)
 326                        n = n->rb_right;
 327                else
 328                        return n;
 329        }
 330        return NULL;
 331}
 332
 333static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
 334{
 335
 336        struct btrfs_fs_info *fs_info = NULL;
 337        struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
 338                                              rb_node);
 339        if (bnode->root)
 340                fs_info = bnode->root->fs_info;
 341        btrfs_panic(fs_info, errno,
 342                    "Inconsistency in backref cache found at offset %llu",
 343                    bytenr);
 344}
 345
 346/*
 347 * walk up backref nodes until reach node presents tree root
 348 */
 349static struct backref_node *walk_up_backref(struct backref_node *node,
 350                                            struct backref_edge *edges[],
 351                                            int *index)
 352{
 353        struct backref_edge *edge;
 354        int idx = *index;
 355
 356        while (!list_empty(&node->upper)) {
 357                edge = list_entry(node->upper.next,
 358                                  struct backref_edge, list[LOWER]);
 359                edges[idx++] = edge;
 360                node = edge->node[UPPER];
 361        }
 362        BUG_ON(node->detached);
 363        *index = idx;
 364        return node;
 365}
 366
 367/*
 368 * walk down backref nodes to find start of next reference path
 369 */
 370static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 371                                              int *index)
 372{
 373        struct backref_edge *edge;
 374        struct backref_node *lower;
 375        int idx = *index;
 376
 377        while (idx > 0) {
 378                edge = edges[idx - 1];
 379                lower = edge->node[LOWER];
 380                if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 381                        idx--;
 382                        continue;
 383                }
 384                edge = list_entry(edge->list[LOWER].next,
 385                                  struct backref_edge, list[LOWER]);
 386                edges[idx - 1] = edge;
 387                *index = idx;
 388                return edge->node[UPPER];
 389        }
 390        *index = 0;
 391        return NULL;
 392}
 393
 394static void unlock_node_buffer(struct backref_node *node)
 395{
 396        if (node->locked) {
 397                btrfs_tree_unlock(node->eb);
 398                node->locked = 0;
 399        }
 400}
 401
 402static void drop_node_buffer(struct backref_node *node)
 403{
 404        if (node->eb) {
 405                unlock_node_buffer(node);
 406                free_extent_buffer(node->eb);
 407                node->eb = NULL;
 408        }
 409}
 410
 411static void drop_backref_node(struct backref_cache *tree,
 412                              struct backref_node *node)
 413{
 414        BUG_ON(!list_empty(&node->upper));
 415
 416        drop_node_buffer(node);
 417        list_del(&node->list);
 418        list_del(&node->lower);
 419        if (!RB_EMPTY_NODE(&node->rb_node))
 420                rb_erase(&node->rb_node, &tree->rb_root);
 421        free_backref_node(tree, node);
 422}
 423
 424/*
 425 * remove a backref node from the backref cache
 426 */
 427static void remove_backref_node(struct backref_cache *cache,
 428                                struct backref_node *node)
 429{
 430        struct backref_node *upper;
 431        struct backref_edge *edge;
 432
 433        if (!node)
 434                return;
 435
 436        BUG_ON(!node->lowest && !node->detached);
 437        while (!list_empty(&node->upper)) {
 438                edge = list_entry(node->upper.next, struct backref_edge,
 439                                  list[LOWER]);
 440                upper = edge->node[UPPER];
 441                list_del(&edge->list[LOWER]);
 442                list_del(&edge->list[UPPER]);
 443                free_backref_edge(cache, edge);
 444
 445                if (RB_EMPTY_NODE(&upper->rb_node)) {
 446                        BUG_ON(!list_empty(&node->upper));
 447                        drop_backref_node(cache, node);
 448                        node = upper;
 449                        node->lowest = 1;
 450                        continue;
 451                }
 452                /*
 453                 * add the node to leaf node list if no other
 454                 * child block cached.
 455                 */
 456                if (list_empty(&upper->lower)) {
 457                        list_add_tail(&upper->lower, &cache->leaves);
 458                        upper->lowest = 1;
 459                }
 460        }
 461
 462        drop_backref_node(cache, node);
 463}
 464
 465static void update_backref_node(struct backref_cache *cache,
 466                                struct backref_node *node, u64 bytenr)
 467{
 468        struct rb_node *rb_node;
 469        rb_erase(&node->rb_node, &cache->rb_root);
 470        node->bytenr = bytenr;
 471        rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 472        if (rb_node)
 473                backref_tree_panic(rb_node, -EEXIST, bytenr);
 474}
 475
 476/*
 477 * update backref cache after a transaction commit
 478 */
 479static int update_backref_cache(struct btrfs_trans_handle *trans,
 480                                struct backref_cache *cache)
 481{
 482        struct backref_node *node;
 483        int level = 0;
 484
 485        if (cache->last_trans == 0) {
 486                cache->last_trans = trans->transid;
 487                return 0;
 488        }
 489
 490        if (cache->last_trans == trans->transid)
 491                return 0;
 492
 493        /*
 494         * detached nodes are used to avoid unnecessary backref
 495         * lookup. transaction commit changes the extent tree.
 496         * so the detached nodes are no longer useful.
 497         */
 498        while (!list_empty(&cache->detached)) {
 499                node = list_entry(cache->detached.next,
 500                                  struct backref_node, list);
 501                remove_backref_node(cache, node);
 502        }
 503
 504        while (!list_empty(&cache->changed)) {
 505                node = list_entry(cache->changed.next,
 506                                  struct backref_node, list);
 507                list_del_init(&node->list);
 508                BUG_ON(node->pending);
 509                update_backref_node(cache, node, node->new_bytenr);
 510        }
 511
 512        /*
 513         * some nodes can be left in the pending list if there were
 514         * errors during processing the pending nodes.
 515         */
 516        for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 517                list_for_each_entry(node, &cache->pending[level], list) {
 518                        BUG_ON(!node->pending);
 519                        if (node->bytenr == node->new_bytenr)
 520                                continue;
 521                        update_backref_node(cache, node, node->new_bytenr);
 522                }
 523        }
 524
 525        cache->last_trans = 0;
 526        return 1;
 527}
 528
 529
 530static int should_ignore_root(struct btrfs_root *root)
 531{
 532        struct btrfs_root *reloc_root;
 533
 534        if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 535                return 0;
 536
 537        reloc_root = root->reloc_root;
 538        if (!reloc_root)
 539                return 0;
 540
 541        if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 542            root->fs_info->running_transaction->transid - 1)
 543                return 0;
 544        /*
 545         * if there is reloc tree and it was created in previous
 546         * transaction backref lookup can find the reloc tree,
 547         * so backref node for the fs tree root is useless for
 548         * relocation.
 549         */
 550        return 1;
 551}
 552/*
 553 * find reloc tree by address of tree root
 554 */
 555static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 556                                          u64 bytenr)
 557{
 558        struct rb_node *rb_node;
 559        struct mapping_node *node;
 560        struct btrfs_root *root = NULL;
 561
 562        spin_lock(&rc->reloc_root_tree.lock);
 563        rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 564        if (rb_node) {
 565                node = rb_entry(rb_node, struct mapping_node, rb_node);
 566                root = (struct btrfs_root *)node->data;
 567        }
 568        spin_unlock(&rc->reloc_root_tree.lock);
 569        return root;
 570}
 571
 572static int is_cowonly_root(u64 root_objectid)
 573{
 574        if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 575            root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 576            root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 577            root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 578            root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 579            root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
 580            root_objectid == BTRFS_UUID_TREE_OBJECTID ||
 581            root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
 582            root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
 583                return 1;
 584        return 0;
 585}
 586
 587static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 588                                        u64 root_objectid)
 589{
 590        struct btrfs_key key;
 591
 592        key.objectid = root_objectid;
 593        key.type = BTRFS_ROOT_ITEM_KEY;
 594        if (is_cowonly_root(root_objectid))
 595                key.offset = 0;
 596        else
 597                key.offset = (u64)-1;
 598
 599        return btrfs_get_fs_root(fs_info, &key, false);
 600}
 601
 602#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 603static noinline_for_stack
 604struct btrfs_root *find_tree_root(struct reloc_control *rc,
 605                                  struct extent_buffer *leaf,
 606                                  struct btrfs_extent_ref_v0 *ref0)
 607{
 608        struct btrfs_root *root;
 609        u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 610        u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 611
 612        BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 613
 614        root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 615        BUG_ON(IS_ERR(root));
 616
 617        if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 618            generation != btrfs_root_generation(&root->root_item))
 619                return NULL;
 620
 621        return root;
 622}
 623#endif
 624
 625static noinline_for_stack
 626int find_inline_backref(struct extent_buffer *leaf, int slot,
 627                        unsigned long *ptr, unsigned long *end)
 628{
 629        struct btrfs_key key;
 630        struct btrfs_extent_item *ei;
 631        struct btrfs_tree_block_info *bi;
 632        u32 item_size;
 633
 634        btrfs_item_key_to_cpu(leaf, &key, slot);
 635
 636        item_size = btrfs_item_size_nr(leaf, slot);
 637#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 638        if (item_size < sizeof(*ei)) {
 639                WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 640                return 1;
 641        }
 642#endif
 643        ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 644        WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 645                  BTRFS_EXTENT_FLAG_TREE_BLOCK));
 646
 647        if (key.type == BTRFS_EXTENT_ITEM_KEY &&
 648            item_size <= sizeof(*ei) + sizeof(*bi)) {
 649                WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 650                return 1;
 651        }
 652        if (key.type == BTRFS_METADATA_ITEM_KEY &&
 653            item_size <= sizeof(*ei)) {
 654                WARN_ON(item_size < sizeof(*ei));
 655                return 1;
 656        }
 657
 658        if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 659                bi = (struct btrfs_tree_block_info *)(ei + 1);
 660                *ptr = (unsigned long)(bi + 1);
 661        } else {
 662                *ptr = (unsigned long)(ei + 1);
 663        }
 664        *end = (unsigned long)ei + item_size;
 665        return 0;
 666}
 667
 668/*
 669 * build backref tree for a given tree block. root of the backref tree
 670 * corresponds the tree block, leaves of the backref tree correspond
 671 * roots of b-trees that reference the tree block.
 672 *
 673 * the basic idea of this function is check backrefs of a given block
 674 * to find upper level blocks that reference the block, and then check
 675 * backrefs of these upper level blocks recursively. the recursion stop
 676 * when tree root is reached or backrefs for the block is cached.
 677 *
 678 * NOTE: if we find backrefs for a block are cached, we know backrefs
 679 * for all upper level blocks that directly/indirectly reference the
 680 * block are also cached.
 681 */
 682static noinline_for_stack
 683struct backref_node *build_backref_tree(struct reloc_control *rc,
 684                                        struct btrfs_key *node_key,
 685                                        int level, u64 bytenr)
 686{
 687        struct backref_cache *cache = &rc->backref_cache;
 688        struct btrfs_path *path1;
 689        struct btrfs_path *path2;
 690        struct extent_buffer *eb;
 691        struct btrfs_root *root;
 692        struct backref_node *cur;
 693        struct backref_node *upper;
 694        struct backref_node *lower;
 695        struct backref_node *node = NULL;
 696        struct backref_node *exist = NULL;
 697        struct backref_edge *edge;
 698        struct rb_node *rb_node;
 699        struct btrfs_key key;
 700        unsigned long end;
 701        unsigned long ptr;
 702        LIST_HEAD(list);
 703        LIST_HEAD(useless);
 704        int cowonly;
 705        int ret;
 706        int err = 0;
 707        bool need_check = true;
 708
 709        path1 = btrfs_alloc_path();
 710        path2 = btrfs_alloc_path();
 711        if (!path1 || !path2) {
 712                err = -ENOMEM;
 713                goto out;
 714        }
 715        path1->reada = READA_FORWARD;
 716        path2->reada = READA_FORWARD;
 717
 718        node = alloc_backref_node(cache);
 719        if (!node) {
 720                err = -ENOMEM;
 721                goto out;
 722        }
 723
 724        node->bytenr = bytenr;
 725        node->level = level;
 726        node->lowest = 1;
 727        cur = node;
 728again:
 729        end = 0;
 730        ptr = 0;
 731        key.objectid = cur->bytenr;
 732        key.type = BTRFS_METADATA_ITEM_KEY;
 733        key.offset = (u64)-1;
 734
 735        path1->search_commit_root = 1;
 736        path1->skip_locking = 1;
 737        ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 738                                0, 0);
 739        if (ret < 0) {
 740                err = ret;
 741                goto out;
 742        }
 743        ASSERT(ret);
 744        ASSERT(path1->slots[0]);
 745
 746        path1->slots[0]--;
 747
 748        WARN_ON(cur->checked);
 749        if (!list_empty(&cur->upper)) {
 750                /*
 751                 * the backref was added previously when processing
 752                 * backref of type BTRFS_TREE_BLOCK_REF_KEY
 753                 */
 754                ASSERT(list_is_singular(&cur->upper));
 755                edge = list_entry(cur->upper.next, struct backref_edge,
 756                                  list[LOWER]);
 757                ASSERT(list_empty(&edge->list[UPPER]));
 758                exist = edge->node[UPPER];
 759                /*
 760                 * add the upper level block to pending list if we need
 761                 * check its backrefs
 762                 */
 763                if (!exist->checked)
 764                        list_add_tail(&edge->list[UPPER], &list);
 765        } else {
 766                exist = NULL;
 767        }
 768
 769        while (1) {
 770                cond_resched();
 771                eb = path1->nodes[0];
 772
 773                if (ptr >= end) {
 774                        if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 775                                ret = btrfs_next_leaf(rc->extent_root, path1);
 776                                if (ret < 0) {
 777                                        err = ret;
 778                                        goto out;
 779                                }
 780                                if (ret > 0)
 781                                        break;
 782                                eb = path1->nodes[0];
 783                        }
 784
 785                        btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 786                        if (key.objectid != cur->bytenr) {
 787                                WARN_ON(exist);
 788                                break;
 789                        }
 790
 791                        if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 792                            key.type == BTRFS_METADATA_ITEM_KEY) {
 793                                ret = find_inline_backref(eb, path1->slots[0],
 794                                                          &ptr, &end);
 795                                if (ret)
 796                                        goto next;
 797                        }
 798                }
 799
 800                if (ptr < end) {
 801                        /* update key for inline back ref */
 802                        struct btrfs_extent_inline_ref *iref;
 803                        int type;
 804                        iref = (struct btrfs_extent_inline_ref *)ptr;
 805                        type = btrfs_get_extent_inline_ref_type(eb, iref,
 806                                                        BTRFS_REF_TYPE_BLOCK);
 807                        if (type == BTRFS_REF_TYPE_INVALID) {
 808                                err = -EINVAL;
 809                                goto out;
 810                        }
 811                        key.type = type;
 812                        key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 813
 814                        WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 815                                key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 816                }
 817
 818                if (exist &&
 819                    ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 820                      exist->owner == key.offset) ||
 821                     (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 822                      exist->bytenr == key.offset))) {
 823                        exist = NULL;
 824                        goto next;
 825                }
 826
 827#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 828                if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 829                    key.type == BTRFS_EXTENT_REF_V0_KEY) {
 830                        if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 831                                struct btrfs_extent_ref_v0 *ref0;
 832                                ref0 = btrfs_item_ptr(eb, path1->slots[0],
 833                                                struct btrfs_extent_ref_v0);
 834                                if (key.objectid == key.offset) {
 835                                        root = find_tree_root(rc, eb, ref0);
 836                                        if (root && !should_ignore_root(root))
 837                                                cur->root = root;
 838                                        else
 839                                                list_add(&cur->list, &useless);
 840                                        break;
 841                                }
 842                                if (is_cowonly_root(btrfs_ref_root_v0(eb,
 843                                                                      ref0)))
 844                                        cur->cowonly = 1;
 845                        }
 846#else
 847                ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
 848                if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 849#endif
 850                        if (key.objectid == key.offset) {
 851                                /*
 852                                 * only root blocks of reloc trees use
 853                                 * backref of this type.
 854                                 */
 855                                root = find_reloc_root(rc, cur->bytenr);
 856                                ASSERT(root);
 857                                cur->root = root;
 858                                break;
 859                        }
 860
 861                        edge = alloc_backref_edge(cache);
 862                        if (!edge) {
 863                                err = -ENOMEM;
 864                                goto out;
 865                        }
 866                        rb_node = tree_search(&cache->rb_root, key.offset);
 867                        if (!rb_node) {
 868                                upper = alloc_backref_node(cache);
 869                                if (!upper) {
 870                                        free_backref_edge(cache, edge);
 871                                        err = -ENOMEM;
 872                                        goto out;
 873                                }
 874                                upper->bytenr = key.offset;
 875                                upper->level = cur->level + 1;
 876                                /*
 877                                 *  backrefs for the upper level block isn't
 878                                 *  cached, add the block to pending list
 879                                 */
 880                                list_add_tail(&edge->list[UPPER], &list);
 881                        } else {
 882                                upper = rb_entry(rb_node, struct backref_node,
 883                                                 rb_node);
 884                                ASSERT(upper->checked);
 885                                INIT_LIST_HEAD(&edge->list[UPPER]);
 886                        }
 887                        list_add_tail(&edge->list[LOWER], &cur->upper);
 888                        edge->node[LOWER] = cur;
 889                        edge->node[UPPER] = upper;
 890
 891                        goto next;
 892                } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 893                        goto next;
 894                }
 895
 896                /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 897                root = read_fs_root(rc->extent_root->fs_info, key.offset);
 898                if (IS_ERR(root)) {
 899                        err = PTR_ERR(root);
 900                        goto out;
 901                }
 902
 903                if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 904                        cur->cowonly = 1;
 905
 906                if (btrfs_root_level(&root->root_item) == cur->level) {
 907                        /* tree root */
 908                        ASSERT(btrfs_root_bytenr(&root->root_item) ==
 909                               cur->bytenr);
 910                        if (should_ignore_root(root))
 911                                list_add(&cur->list, &useless);
 912                        else
 913                                cur->root = root;
 914                        break;
 915                }
 916
 917                level = cur->level + 1;
 918
 919                /*
 920                 * searching the tree to find upper level blocks
 921                 * reference the block.
 922                 */
 923                path2->search_commit_root = 1;
 924                path2->skip_locking = 1;
 925                path2->lowest_level = level;
 926                ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 927                path2->lowest_level = 0;
 928                if (ret < 0) {
 929                        err = ret;
 930                        goto out;
 931                }
 932                if (ret > 0 && path2->slots[level] > 0)
 933                        path2->slots[level]--;
 934
 935                eb = path2->nodes[level];
 936                if (btrfs_node_blockptr(eb, path2->slots[level]) !=
 937                    cur->bytenr) {
 938                        btrfs_err(root->fs_info,
 939        "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
 940                                  cur->bytenr, level - 1, root->objectid,
 941                                  node_key->objectid, node_key->type,
 942                                  node_key->offset);
 943                        err = -ENOENT;
 944                        goto out;
 945                }
 946                lower = cur;
 947                need_check = true;
 948                for (; level < BTRFS_MAX_LEVEL; level++) {
 949                        if (!path2->nodes[level]) {
 950                                ASSERT(btrfs_root_bytenr(&root->root_item) ==
 951                                       lower->bytenr);
 952                                if (should_ignore_root(root))
 953                                        list_add(&lower->list, &useless);
 954                                else
 955                                        lower->root = root;
 956                                break;
 957                        }
 958
 959                        edge = alloc_backref_edge(cache);
 960                        if (!edge) {
 961                                err = -ENOMEM;
 962                                goto out;
 963                        }
 964
 965                        eb = path2->nodes[level];
 966                        rb_node = tree_search(&cache->rb_root, eb->start);
 967                        if (!rb_node) {
 968                                upper = alloc_backref_node(cache);
 969                                if (!upper) {
 970                                        free_backref_edge(cache, edge);
 971                                        err = -ENOMEM;
 972                                        goto out;
 973                                }
 974                                upper->bytenr = eb->start;
 975                                upper->owner = btrfs_header_owner(eb);
 976                                upper->level = lower->level + 1;
 977                                if (!test_bit(BTRFS_ROOT_REF_COWS,
 978                                              &root->state))
 979                                        upper->cowonly = 1;
 980
 981                                /*
 982                                 * if we know the block isn't shared
 983                                 * we can void checking its backrefs.
 984                                 */
 985                                if (btrfs_block_can_be_shared(root, eb))
 986                                        upper->checked = 0;
 987                                else
 988                                        upper->checked = 1;
 989
 990                                /*
 991                                 * add the block to pending list if we
 992                                 * need check its backrefs, we only do this once
 993                                 * while walking up a tree as we will catch
 994                                 * anything else later on.
 995                                 */
 996                                if (!upper->checked && need_check) {
 997                                        need_check = false;
 998                                        list_add_tail(&edge->list[UPPER],
 999                                                      &list);
1000                                } else {
1001                                        if (upper->checked)
1002                                                need_check = true;
1003                                        INIT_LIST_HEAD(&edge->list[UPPER]);
1004                                }
1005                        } else {
1006                                upper = rb_entry(rb_node, struct backref_node,
1007                                                 rb_node);
1008                                ASSERT(upper->checked);
1009                                INIT_LIST_HEAD(&edge->list[UPPER]);
1010                                if (!upper->owner)
1011                                        upper->owner = btrfs_header_owner(eb);
1012                        }
1013                        list_add_tail(&edge->list[LOWER], &lower->upper);
1014                        edge->node[LOWER] = lower;
1015                        edge->node[UPPER] = upper;
1016
1017                        if (rb_node)
1018                                break;
1019                        lower = upper;
1020                        upper = NULL;
1021                }
1022                btrfs_release_path(path2);
1023next:
1024                if (ptr < end) {
1025                        ptr += btrfs_extent_inline_ref_size(key.type);
1026                        if (ptr >= end) {
1027                                WARN_ON(ptr > end);
1028                                ptr = 0;
1029                                end = 0;
1030                        }
1031                }
1032                if (ptr >= end)
1033                        path1->slots[0]++;
1034        }
1035        btrfs_release_path(path1);
1036
1037        cur->checked = 1;
1038        WARN_ON(exist);
1039
1040        /* the pending list isn't empty, take the first block to process */
1041        if (!list_empty(&list)) {
1042                edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1043                list_del_init(&edge->list[UPPER]);
1044                cur = edge->node[UPPER];
1045                goto again;
1046        }
1047
1048        /*
1049         * everything goes well, connect backref nodes and insert backref nodes
1050         * into the cache.
1051         */
1052        ASSERT(node->checked);
1053        cowonly = node->cowonly;
1054        if (!cowonly) {
1055                rb_node = tree_insert(&cache->rb_root, node->bytenr,
1056                                      &node->rb_node);
1057                if (rb_node)
1058                        backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1059                list_add_tail(&node->lower, &cache->leaves);
1060        }
1061
1062        list_for_each_entry(edge, &node->upper, list[LOWER])
1063                list_add_tail(&edge->list[UPPER], &list);
1064
1065        while (!list_empty(&list)) {
1066                edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1067                list_del_init(&edge->list[UPPER]);
1068                upper = edge->node[UPPER];
1069                if (upper->detached) {
1070                        list_del(&edge->list[LOWER]);
1071                        lower = edge->node[LOWER];
1072                        free_backref_edge(cache, edge);
1073                        if (list_empty(&lower->upper))
1074                                list_add(&lower->list, &useless);
1075                        continue;
1076                }
1077
1078                if (!RB_EMPTY_NODE(&upper->rb_node)) {
1079                        if (upper->lowest) {
1080                                list_del_init(&upper->lower);
1081                                upper->lowest = 0;
1082                        }
1083
1084                        list_add_tail(&edge->list[UPPER], &upper->lower);
1085                        continue;
1086                }
1087
1088                if (!upper->checked) {
1089                        /*
1090                         * Still want to blow up for developers since this is a
1091                         * logic bug.
1092                         */
1093                        ASSERT(0);
1094                        err = -EINVAL;
1095                        goto out;
1096                }
1097                if (cowonly != upper->cowonly) {
1098                        ASSERT(0);
1099                        err = -EINVAL;
1100                        goto out;
1101                }
1102
1103                if (!cowonly) {
1104                        rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1105                                              &upper->rb_node);
1106                        if (rb_node)
1107                                backref_tree_panic(rb_node, -EEXIST,
1108                                                   upper->bytenr);
1109                }
1110
1111                list_add_tail(&edge->list[UPPER], &upper->lower);
1112
1113                list_for_each_entry(edge, &upper->upper, list[LOWER])
1114                        list_add_tail(&edge->list[UPPER], &list);
1115        }
1116        /*
1117         * process useless backref nodes. backref nodes for tree leaves
1118         * are deleted from the cache. backref nodes for upper level
1119         * tree blocks are left in the cache to avoid unnecessary backref
1120         * lookup.
1121         */
1122        while (!list_empty(&useless)) {
1123                upper = list_entry(useless.next, struct backref_node, list);
1124                list_del_init(&upper->list);
1125                ASSERT(list_empty(&upper->upper));
1126                if (upper == node)
1127                        node = NULL;
1128                if (upper->lowest) {
1129                        list_del_init(&upper->lower);
1130                        upper->lowest = 0;
1131                }
1132                while (!list_empty(&upper->lower)) {
1133                        edge = list_entry(upper->lower.next,
1134                                          struct backref_edge, list[UPPER]);
1135                        list_del(&edge->list[UPPER]);
1136                        list_del(&edge->list[LOWER]);
1137                        lower = edge->node[LOWER];
1138                        free_backref_edge(cache, edge);
1139
1140                        if (list_empty(&lower->upper))
1141                                list_add(&lower->list, &useless);
1142                }
1143                __mark_block_processed(rc, upper);
1144                if (upper->level > 0) {
1145                        list_add(&upper->list, &cache->detached);
1146                        upper->detached = 1;
1147                } else {
1148                        rb_erase(&upper->rb_node, &cache->rb_root);
1149                        free_backref_node(cache, upper);
1150                }
1151        }
1152out:
1153        btrfs_free_path(path1);
1154        btrfs_free_path(path2);
1155        if (err) {
1156                while (!list_empty(&useless)) {
1157                        lower = list_entry(useless.next,
1158                                           struct backref_node, list);
1159                        list_del_init(&lower->list);
1160                }
1161                while (!list_empty(&list)) {
1162                        edge = list_first_entry(&list, struct backref_edge,
1163                                                list[UPPER]);
1164                        list_del(&edge->list[UPPER]);
1165                        list_del(&edge->list[LOWER]);
1166                        lower = edge->node[LOWER];
1167                        upper = edge->node[UPPER];
1168                        free_backref_edge(cache, edge);
1169
1170                        /*
1171                         * Lower is no longer linked to any upper backref nodes
1172                         * and isn't in the cache, we can free it ourselves.
1173                         */
1174                        if (list_empty(&lower->upper) &&
1175                            RB_EMPTY_NODE(&lower->rb_node))
1176                                list_add(&lower->list, &useless);
1177
1178                        if (!RB_EMPTY_NODE(&upper->rb_node))
1179                                continue;
1180
1181                        /* Add this guy's upper edges to the list to process */
1182                        list_for_each_entry(edge, &upper->upper, list[LOWER])
1183                                list_add_tail(&edge->list[UPPER], &list);
1184                        if (list_empty(&upper->upper))
1185                                list_add(&upper->list, &useless);
1186                }
1187
1188                while (!list_empty(&useless)) {
1189                        lower = list_entry(useless.next,
1190                                           struct backref_node, list);
1191                        list_del_init(&lower->list);
1192                        if (lower == node)
1193                                node = NULL;
1194                        free_backref_node(cache, lower);
1195                }
1196
1197                free_backref_node(cache, node);
1198                return ERR_PTR(err);
1199        }
1200        ASSERT(!node || !node->detached);
1201        return node;
1202}
1203
1204/*
1205 * helper to add backref node for the newly created snapshot.
1206 * the backref node is created by cloning backref node that
1207 * corresponds to root of source tree
1208 */
1209static int clone_backref_node(struct btrfs_trans_handle *trans,
1210                              struct reloc_control *rc,
1211                              struct btrfs_root *src,
1212                              struct btrfs_root *dest)
1213{
1214        struct btrfs_root *reloc_root = src->reloc_root;
1215        struct backref_cache *cache = &rc->backref_cache;
1216        struct backref_node *node = NULL;
1217        struct backref_node *new_node;
1218        struct backref_edge *edge;
1219        struct backref_edge *new_edge;
1220        struct rb_node *rb_node;
1221
1222        if (cache->last_trans > 0)
1223                update_backref_cache(trans, cache);
1224
1225        rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1226        if (rb_node) {
1227                node = rb_entry(rb_node, struct backref_node, rb_node);
1228                if (node->detached)
1229                        node = NULL;
1230                else
1231                        BUG_ON(node->new_bytenr != reloc_root->node->start);
1232        }
1233
1234        if (!node) {
1235                rb_node = tree_search(&cache->rb_root,
1236                                      reloc_root->commit_root->start);
1237                if (rb_node) {
1238                        node = rb_entry(rb_node, struct backref_node,
1239                                        rb_node);
1240                        BUG_ON(node->detached);
1241                }
1242        }
1243
1244        if (!node)
1245                return 0;
1246
1247        new_node = alloc_backref_node(cache);
1248        if (!new_node)
1249                return -ENOMEM;
1250
1251        new_node->bytenr = dest->node->start;
1252        new_node->level = node->level;
1253        new_node->lowest = node->lowest;
1254        new_node->checked = 1;
1255        new_node->root = dest;
1256
1257        if (!node->lowest) {
1258                list_for_each_entry(edge, &node->lower, list[UPPER]) {
1259                        new_edge = alloc_backref_edge(cache);
1260                        if (!new_edge)
1261                                goto fail;
1262
1263                        new_edge->node[UPPER] = new_node;
1264                        new_edge->node[LOWER] = edge->node[LOWER];
1265                        list_add_tail(&new_edge->list[UPPER],
1266                                      &new_node->lower);
1267                }
1268        } else {
1269                list_add_tail(&new_node->lower, &cache->leaves);
1270        }
1271
1272        rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1273                              &new_node->rb_node);
1274        if (rb_node)
1275                backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1276
1277        if (!new_node->lowest) {
1278                list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1279                        list_add_tail(&new_edge->list[LOWER],
1280                                      &new_edge->node[LOWER]->upper);
1281                }
1282        }
1283        return 0;
1284fail:
1285        while (!list_empty(&new_node->lower)) {
1286                new_edge = list_entry(new_node->lower.next,
1287                                      struct backref_edge, list[UPPER]);
1288                list_del(&new_edge->list[UPPER]);
1289                free_backref_edge(cache, new_edge);
1290        }
1291        free_backref_node(cache, new_node);
1292        return -ENOMEM;
1293}
1294
1295/*
1296 * helper to add 'address of tree root -> reloc tree' mapping
1297 */
1298static int __must_check __add_reloc_root(struct btrfs_root *root)
1299{
1300        struct btrfs_fs_info *fs_info = root->fs_info;
1301        struct rb_node *rb_node;
1302        struct mapping_node *node;
1303        struct reloc_control *rc = fs_info->reloc_ctl;
1304
1305        node = kmalloc(sizeof(*node), GFP_NOFS);
1306        if (!node)
1307                return -ENOMEM;
1308
1309        node->bytenr = root->node->start;
1310        node->data = root;
1311
1312        spin_lock(&rc->reloc_root_tree.lock);
1313        rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1314                              node->bytenr, &node->rb_node);
1315        spin_unlock(&rc->reloc_root_tree.lock);
1316        if (rb_node) {
1317                btrfs_panic(fs_info, -EEXIST,
1318                            "Duplicate root found for start=%llu while inserting into relocation tree",
1319                            node->bytenr);
1320        }
1321
1322        list_add_tail(&root->root_list, &rc->reloc_roots);
1323        return 0;
1324}
1325
1326/*
1327 * helper to delete the 'address of tree root -> reloc tree'
1328 * mapping
1329 */
1330static void __del_reloc_root(struct btrfs_root *root)
1331{
1332        struct btrfs_fs_info *fs_info = root->fs_info;
1333        struct rb_node *rb_node;
1334        struct mapping_node *node = NULL;
1335        struct reloc_control *rc = fs_info->reloc_ctl;
1336
1337        spin_lock(&rc->reloc_root_tree.lock);
1338        rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1339                              root->node->start);
1340        if (rb_node) {
1341                node = rb_entry(rb_node, struct mapping_node, rb_node);
1342                rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1343        }
1344        spin_unlock(&rc->reloc_root_tree.lock);
1345
1346        if (!node)
1347                return;
1348        BUG_ON((struct btrfs_root *)node->data != root);
1349
1350        spin_lock(&fs_info->trans_lock);
1351        list_del_init(&root->root_list);
1352        spin_unlock(&fs_info->trans_lock);
1353        kfree(node);
1354}
1355
1356/*
1357 * helper to update the 'address of tree root -> reloc tree'
1358 * mapping
1359 */
1360static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1361{
1362        struct btrfs_fs_info *fs_info = root->fs_info;
1363        struct rb_node *rb_node;
1364        struct mapping_node *node = NULL;
1365        struct reloc_control *rc = fs_info->reloc_ctl;
1366
1367        spin_lock(&rc->reloc_root_tree.lock);
1368        rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1369                              root->node->start);
1370        if (rb_node) {
1371                node = rb_entry(rb_node, struct mapping_node, rb_node);
1372                rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1373        }
1374        spin_unlock(&rc->reloc_root_tree.lock);
1375
1376        if (!node)
1377                return 0;
1378        BUG_ON((struct btrfs_root *)node->data != root);
1379
1380        spin_lock(&rc->reloc_root_tree.lock);
1381        node->bytenr = new_bytenr;
1382        rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1383                              node->bytenr, &node->rb_node);
1384        spin_unlock(&rc->reloc_root_tree.lock);
1385        if (rb_node)
1386                backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1387        return 0;
1388}
1389
1390static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1391                                        struct btrfs_root *root, u64 objectid)
1392{
1393        struct btrfs_fs_info *fs_info = root->fs_info;
1394        struct btrfs_root *reloc_root;
1395        struct extent_buffer *eb;
1396        struct btrfs_root_item *root_item;
1397        struct btrfs_key root_key;
1398        int ret;
1399
1400        root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1401        BUG_ON(!root_item);
1402
1403        root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1404        root_key.type = BTRFS_ROOT_ITEM_KEY;
1405        root_key.offset = objectid;
1406
1407        if (root->root_key.objectid == objectid) {
1408                u64 commit_root_gen;
1409
1410                /* called by btrfs_init_reloc_root */
1411                ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1412                                      BTRFS_TREE_RELOC_OBJECTID);
1413                BUG_ON(ret);
1414                /*
1415                 * Set the last_snapshot field to the generation of the commit
1416                 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1417                 * correctly (returns true) when the relocation root is created
1418                 * either inside the critical section of a transaction commit
1419                 * (through transaction.c:qgroup_account_snapshot()) and when
1420                 * it's created before the transaction commit is started.
1421                 */
1422                commit_root_gen = btrfs_header_generation(root->commit_root);
1423                btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1424        } else {
1425                /*
1426                 * called by btrfs_reloc_post_snapshot_hook.
1427                 * the source tree is a reloc tree, all tree blocks
1428                 * modified after it was created have RELOC flag
1429                 * set in their headers. so it's OK to not update
1430                 * the 'last_snapshot'.
1431                 */
1432                ret = btrfs_copy_root(trans, root, root->node, &eb,
1433                                      BTRFS_TREE_RELOC_OBJECTID);
1434                BUG_ON(ret);
1435        }
1436
1437        memcpy(root_item, &root->root_item, sizeof(*root_item));
1438        btrfs_set_root_bytenr(root_item, eb->start);
1439        btrfs_set_root_level(root_item, btrfs_header_level(eb));
1440        btrfs_set_root_generation(root_item, trans->transid);
1441
1442        if (root->root_key.objectid == objectid) {
1443                btrfs_set_root_refs(root_item, 0);
1444                memset(&root_item->drop_progress, 0,
1445                       sizeof(struct btrfs_disk_key));
1446                root_item->drop_level = 0;
1447        }
1448
1449        btrfs_tree_unlock(eb);
1450        free_extent_buffer(eb);
1451
1452        ret = btrfs_insert_root(trans, fs_info->tree_root,
1453                                &root_key, root_item);
1454        BUG_ON(ret);
1455        kfree(root_item);
1456
1457        reloc_root = btrfs_read_fs_root(fs_info->tree_root, &root_key);
1458        BUG_ON(IS_ERR(reloc_root));
1459        reloc_root->last_trans = trans->transid;
1460        return reloc_root;
1461}
1462
1463/*
1464 * create reloc tree for a given fs tree. reloc tree is just a
1465 * snapshot of the fs tree with special root objectid.
1466 */
1467int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1468                          struct btrfs_root *root)
1469{
1470        struct btrfs_fs_info *fs_info = root->fs_info;
1471        struct btrfs_root *reloc_root;
1472        struct reloc_control *rc = fs_info->reloc_ctl;
1473        struct btrfs_block_rsv *rsv;
1474        int clear_rsv = 0;
1475        int ret;
1476
1477        if (root->reloc_root) {
1478                reloc_root = root->reloc_root;
1479                reloc_root->last_trans = trans->transid;
1480                return 0;
1481        }
1482
1483        if (!rc || !rc->create_reloc_tree ||
1484            root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1485                return 0;
1486
1487        if (!trans->reloc_reserved) {
1488                rsv = trans->block_rsv;
1489                trans->block_rsv = rc->block_rsv;
1490                clear_rsv = 1;
1491        }
1492        reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1493        if (clear_rsv)
1494                trans->block_rsv = rsv;
1495
1496        ret = __add_reloc_root(reloc_root);
1497        BUG_ON(ret < 0);
1498        root->reloc_root = reloc_root;
1499        return 0;
1500}
1501
1502/*
1503 * update root item of reloc tree
1504 */
1505int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1506                            struct btrfs_root *root)
1507{
1508        struct btrfs_fs_info *fs_info = root->fs_info;
1509        struct btrfs_root *reloc_root;
1510        struct btrfs_root_item *root_item;
1511        int ret;
1512
1513        if (!root->reloc_root)
1514                goto out;
1515
1516        reloc_root = root->reloc_root;
1517        root_item = &reloc_root->root_item;
1518
1519        if (fs_info->reloc_ctl->merge_reloc_tree &&
1520            btrfs_root_refs(root_item) == 0) {
1521                root->reloc_root = NULL;
1522                __del_reloc_root(reloc_root);
1523        }
1524
1525        if (reloc_root->commit_root != reloc_root->node) {
1526                btrfs_set_root_node(root_item, reloc_root->node);
1527                free_extent_buffer(reloc_root->commit_root);
1528                reloc_root->commit_root = btrfs_root_node(reloc_root);
1529        }
1530
1531        ret = btrfs_update_root(trans, fs_info->tree_root,
1532                                &reloc_root->root_key, root_item);
1533        BUG_ON(ret);
1534
1535out:
1536        return 0;
1537}
1538
1539/*
1540 * helper to find first cached inode with inode number >= objectid
1541 * in a subvolume
1542 */
1543static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1544{
1545        struct rb_node *node;
1546        struct rb_node *prev;
1547        struct btrfs_inode *entry;
1548        struct inode *inode;
1549
1550        spin_lock(&root->inode_lock);
1551again:
1552        node = root->inode_tree.rb_node;
1553        prev = NULL;
1554        while (node) {
1555                prev = node;
1556                entry = rb_entry(node, struct btrfs_inode, rb_node);
1557
1558                if (objectid < btrfs_ino(entry))
1559                        node = node->rb_left;
1560                else if (objectid > btrfs_ino(entry))
1561                        node = node->rb_right;
1562                else
1563                        break;
1564        }
1565        if (!node) {
1566                while (prev) {
1567                        entry = rb_entry(prev, struct btrfs_inode, rb_node);
1568                        if (objectid <= btrfs_ino(entry)) {
1569                                node = prev;
1570                                break;
1571                        }
1572                        prev = rb_next(prev);
1573                }
1574        }
1575        while (node) {
1576                entry = rb_entry(node, struct btrfs_inode, rb_node);
1577                inode = igrab(&entry->vfs_inode);
1578                if (inode) {
1579                        spin_unlock(&root->inode_lock);
1580                        return inode;
1581                }
1582
1583                objectid = btrfs_ino(entry) + 1;
1584                if (cond_resched_lock(&root->inode_lock))
1585                        goto again;
1586
1587                node = rb_next(node);
1588        }
1589        spin_unlock(&root->inode_lock);
1590        return NULL;
1591}
1592
1593static int in_block_group(u64 bytenr,
1594                          struct btrfs_block_group_cache *block_group)
1595{
1596        if (bytenr >= block_group->key.objectid &&
1597            bytenr < block_group->key.objectid + block_group->key.offset)
1598                return 1;
1599        return 0;
1600}
1601
1602/*
1603 * get new location of data
1604 */
1605static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1606                            u64 bytenr, u64 num_bytes)
1607{
1608        struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1609        struct btrfs_path *path;
1610        struct btrfs_file_extent_item *fi;
1611        struct extent_buffer *leaf;
1612        int ret;
1613
1614        path = btrfs_alloc_path();
1615        if (!path)
1616                return -ENOMEM;
1617
1618        bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1619        ret = btrfs_lookup_file_extent(NULL, root, path,
1620                        btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1621        if (ret < 0)
1622                goto out;
1623        if (ret > 0) {
1624                ret = -ENOENT;
1625                goto out;
1626        }
1627
1628        leaf = path->nodes[0];
1629        fi = btrfs_item_ptr(leaf, path->slots[0],
1630                            struct btrfs_file_extent_item);
1631
1632        BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1633               btrfs_file_extent_compression(leaf, fi) ||
1634               btrfs_file_extent_encryption(leaf, fi) ||
1635               btrfs_file_extent_other_encoding(leaf, fi));
1636
1637        if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1638                ret = -EINVAL;
1639                goto out;
1640        }
1641
1642        *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1643        ret = 0;
1644out:
1645        btrfs_free_path(path);
1646        return ret;
1647}
1648
1649/*
1650 * update file extent items in the tree leaf to point to
1651 * the new locations.
1652 */
1653static noinline_for_stack
1654int replace_file_extents(struct btrfs_trans_handle *trans,
1655                         struct reloc_control *rc,
1656                         struct btrfs_root *root,
1657                         struct extent_buffer *leaf)
1658{
1659        struct btrfs_fs_info *fs_info = root->fs_info;
1660        struct btrfs_key key;
1661        struct btrfs_file_extent_item *fi;
1662        struct inode *inode = NULL;
1663        u64 parent;
1664        u64 bytenr;
1665        u64 new_bytenr = 0;
1666        u64 num_bytes;
1667        u64 end;
1668        u32 nritems;
1669        u32 i;
1670        int ret = 0;
1671        int first = 1;
1672        int dirty = 0;
1673
1674        if (rc->stage != UPDATE_DATA_PTRS)
1675                return 0;
1676
1677        /* reloc trees always use full backref */
1678        if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1679                parent = leaf->start;
1680        else
1681                parent = 0;
1682
1683        nritems = btrfs_header_nritems(leaf);
1684        for (i = 0; i < nritems; i++) {
1685                cond_resched();
1686                btrfs_item_key_to_cpu(leaf, &key, i);
1687                if (key.type != BTRFS_EXTENT_DATA_KEY)
1688                        continue;
1689                fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1690                if (btrfs_file_extent_type(leaf, fi) ==
1691                    BTRFS_FILE_EXTENT_INLINE)
1692                        continue;
1693                bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1694                num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1695                if (bytenr == 0)
1696                        continue;
1697                if (!in_block_group(bytenr, rc->block_group))
1698                        continue;
1699
1700                /*
1701                 * if we are modifying block in fs tree, wait for readpage
1702                 * to complete and drop the extent cache
1703                 */
1704                if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1705                        if (first) {
1706                                inode = find_next_inode(root, key.objectid);
1707                                first = 0;
1708                        } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1709                                btrfs_add_delayed_iput(inode);
1710                                inode = find_next_inode(root, key.objectid);
1711                        }
1712                        if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1713                                end = key.offset +
1714                                      btrfs_file_extent_num_bytes(leaf, fi);
1715                                WARN_ON(!IS_ALIGNED(key.offset,
1716                                                    fs_info->sectorsize));
1717                                WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1718                                end--;
1719                                ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1720                                                      key.offset, end);
1721                                if (!ret)
1722                                        continue;
1723
1724                                btrfs_drop_extent_cache(BTRFS_I(inode),
1725                                                key.offset,     end, 1);
1726                                unlock_extent(&BTRFS_I(inode)->io_tree,
1727                                              key.offset, end);
1728                        }
1729                }
1730
1731                ret = get_new_location(rc->data_inode, &new_bytenr,
1732                                       bytenr, num_bytes);
1733                if (ret) {
1734                        /*
1735                         * Don't have to abort since we've not changed anything
1736                         * in the file extent yet.
1737                         */
1738                        break;
1739                }
1740
1741                btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1742                dirty = 1;
1743
1744                key.offset -= btrfs_file_extent_offset(leaf, fi);
1745                ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1746                                           num_bytes, parent,
1747                                           btrfs_header_owner(leaf),
1748                                           key.objectid, key.offset);
1749                if (ret) {
1750                        btrfs_abort_transaction(trans, ret);
1751                        break;
1752                }
1753
1754                ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1755                                        parent, btrfs_header_owner(leaf),
1756                                        key.objectid, key.offset);
1757                if (ret) {
1758                        btrfs_abort_transaction(trans, ret);
1759                        break;
1760                }
1761        }
1762        if (dirty)
1763                btrfs_mark_buffer_dirty(leaf);
1764        if (inode)
1765                btrfs_add_delayed_iput(inode);
1766        return ret;
1767}
1768
1769static noinline_for_stack
1770int memcmp_node_keys(struct extent_buffer *eb, int slot,
1771                     struct btrfs_path *path, int level)
1772{
1773        struct btrfs_disk_key key1;
1774        struct btrfs_disk_key key2;
1775        btrfs_node_key(eb, &key1, slot);
1776        btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1777        return memcmp(&key1, &key2, sizeof(key1));
1778}
1779
1780/*
1781 * try to replace tree blocks in fs tree with the new blocks
1782 * in reloc tree. tree blocks haven't been modified since the
1783 * reloc tree was create can be replaced.
1784 *
1785 * if a block was replaced, level of the block + 1 is returned.
1786 * if no block got replaced, 0 is returned. if there are other
1787 * errors, a negative error number is returned.
1788 */
1789static noinline_for_stack
1790int replace_path(struct btrfs_trans_handle *trans,
1791                 struct btrfs_root *dest, struct btrfs_root *src,
1792                 struct btrfs_path *path, struct btrfs_key *next_key,
1793                 int lowest_level, int max_level)
1794{
1795        struct btrfs_fs_info *fs_info = dest->fs_info;
1796        struct extent_buffer *eb;
1797        struct extent_buffer *parent;
1798        struct btrfs_key key;
1799        u64 old_bytenr;
1800        u64 new_bytenr;
1801        u64 old_ptr_gen;
1802        u64 new_ptr_gen;
1803        u64 last_snapshot;
1804        u32 blocksize;
1805        int cow = 0;
1806        int level;
1807        int ret;
1808        int slot;
1809
1810        BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1811        BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1812
1813        last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1814again:
1815        slot = path->slots[lowest_level];
1816        btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1817
1818        eb = btrfs_lock_root_node(dest);
1819        btrfs_set_lock_blocking(eb);
1820        level = btrfs_header_level(eb);
1821
1822        if (level < lowest_level) {
1823                btrfs_tree_unlock(eb);
1824                free_extent_buffer(eb);
1825                return 0;
1826        }
1827
1828        if (cow) {
1829                ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1830                BUG_ON(ret);
1831        }
1832        btrfs_set_lock_blocking(eb);
1833
1834        if (next_key) {
1835                next_key->objectid = (u64)-1;
1836                next_key->type = (u8)-1;
1837                next_key->offset = (u64)-1;
1838        }
1839
1840        parent = eb;
1841        while (1) {
1842                level = btrfs_header_level(parent);
1843                BUG_ON(level < lowest_level);
1844
1845                ret = btrfs_bin_search(parent, &key, level, &slot);
1846                if (ret && slot > 0)
1847                        slot--;
1848
1849                if (next_key && slot + 1 < btrfs_header_nritems(parent))
1850                        btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1851
1852                old_bytenr = btrfs_node_blockptr(parent, slot);
1853                blocksize = fs_info->nodesize;
1854                old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1855
1856                if (level <= max_level) {
1857                        eb = path->nodes[level];
1858                        new_bytenr = btrfs_node_blockptr(eb,
1859                                                        path->slots[level]);
1860                        new_ptr_gen = btrfs_node_ptr_generation(eb,
1861                                                        path->slots[level]);
1862                } else {
1863                        new_bytenr = 0;
1864                        new_ptr_gen = 0;
1865                }
1866
1867                if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1868                        ret = level;
1869                        break;
1870                }
1871
1872                if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1873                    memcmp_node_keys(parent, slot, path, level)) {
1874                        if (level <= lowest_level) {
1875                                ret = 0;
1876                                break;
1877                        }
1878
1879                        eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen);
1880                        if (IS_ERR(eb)) {
1881                                ret = PTR_ERR(eb);
1882                                break;
1883                        } else if (!extent_buffer_uptodate(eb)) {
1884                                ret = -EIO;
1885                                free_extent_buffer(eb);
1886                                break;
1887                        }
1888                        btrfs_tree_lock(eb);
1889                        if (cow) {
1890                                ret = btrfs_cow_block(trans, dest, eb, parent,
1891                                                      slot, &eb);
1892                                BUG_ON(ret);
1893                        }
1894                        btrfs_set_lock_blocking(eb);
1895
1896                        btrfs_tree_unlock(parent);
1897                        free_extent_buffer(parent);
1898
1899                        parent = eb;
1900                        continue;
1901                }
1902
1903                if (!cow) {
1904                        btrfs_tree_unlock(parent);
1905                        free_extent_buffer(parent);
1906                        cow = 1;
1907                        goto again;
1908                }
1909
1910                btrfs_node_key_to_cpu(path->nodes[level], &key,
1911                                      path->slots[level]);
1912                btrfs_release_path(path);
1913
1914                path->lowest_level = level;
1915                ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1916                path->lowest_level = 0;
1917                BUG_ON(ret);
1918
1919                /*
1920                 * Info qgroup to trace both subtrees.
1921                 *
1922                 * We must trace both trees.
1923                 * 1) Tree reloc subtree
1924                 *    If not traced, we will leak data numbers
1925                 * 2) Fs subtree
1926                 *    If not traced, we will double count old data
1927                 *    and tree block numbers, if current trans doesn't free
1928                 *    data reloc tree inode.
1929                 */
1930                ret = btrfs_qgroup_trace_subtree(trans, src, parent,
1931                                btrfs_header_generation(parent),
1932                                btrfs_header_level(parent));
1933                if (ret < 0)
1934                        break;
1935                ret = btrfs_qgroup_trace_subtree(trans, dest,
1936                                path->nodes[level],
1937                                btrfs_header_generation(path->nodes[level]),
1938                                btrfs_header_level(path->nodes[level]));
1939                if (ret < 0)
1940                        break;
1941
1942                /*
1943                 * swap blocks in fs tree and reloc tree.
1944                 */
1945                btrfs_set_node_blockptr(parent, slot, new_bytenr);
1946                btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1947                btrfs_mark_buffer_dirty(parent);
1948
1949                btrfs_set_node_blockptr(path->nodes[level],
1950                                        path->slots[level], old_bytenr);
1951                btrfs_set_node_ptr_generation(path->nodes[level],
1952                                              path->slots[level], old_ptr_gen);
1953                btrfs_mark_buffer_dirty(path->nodes[level]);
1954
1955                ret = btrfs_inc_extent_ref(trans, src, old_bytenr,
1956                                        blocksize, path->nodes[level]->start,
1957                                        src->root_key.objectid, level - 1, 0);
1958                BUG_ON(ret);
1959                ret = btrfs_inc_extent_ref(trans, dest, new_bytenr,
1960                                        blocksize, 0, dest->root_key.objectid,
1961                                        level - 1, 0);
1962                BUG_ON(ret);
1963
1964                ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1965                                        path->nodes[level]->start,
1966                                        src->root_key.objectid, level - 1, 0);
1967                BUG_ON(ret);
1968
1969                ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1970                                        0, dest->root_key.objectid, level - 1,
1971                                        0);
1972                BUG_ON(ret);
1973
1974                btrfs_unlock_up_safe(path, 0);
1975
1976                ret = level;
1977                break;
1978        }
1979        btrfs_tree_unlock(parent);
1980        free_extent_buffer(parent);
1981        return ret;
1982}
1983
1984/*
1985 * helper to find next relocated block in reloc tree
1986 */
1987static noinline_for_stack
1988int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1989                       int *level)
1990{
1991        struct extent_buffer *eb;
1992        int i;
1993        u64 last_snapshot;
1994        u32 nritems;
1995
1996        last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1997
1998        for (i = 0; i < *level; i++) {
1999                free_extent_buffer(path->nodes[i]);
2000                path->nodes[i] = NULL;
2001        }
2002
2003        for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
2004                eb = path->nodes[i];
2005                nritems = btrfs_header_nritems(eb);
2006                while (path->slots[i] + 1 < nritems) {
2007                        path->slots[i]++;
2008                        if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
2009                            last_snapshot)
2010                                continue;
2011
2012                        *level = i;
2013                        return 0;
2014                }
2015                free_extent_buffer(path->nodes[i]);
2016                path->nodes[i] = NULL;
2017        }
2018        return 1;
2019}
2020
2021/*
2022 * walk down reloc tree to find relocated block of lowest level
2023 */
2024static noinline_for_stack
2025int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
2026                         int *level)
2027{
2028        struct btrfs_fs_info *fs_info = root->fs_info;
2029        struct extent_buffer *eb = NULL;
2030        int i;
2031        u64 bytenr;
2032        u64 ptr_gen = 0;
2033        u64 last_snapshot;
2034        u32 nritems;
2035
2036        last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2037
2038        for (i = *level; i > 0; i--) {
2039                eb = path->nodes[i];
2040                nritems = btrfs_header_nritems(eb);
2041                while (path->slots[i] < nritems) {
2042                        ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2043                        if (ptr_gen > last_snapshot)
2044                                break;
2045                        path->slots[i]++;
2046                }
2047                if (path->slots[i] >= nritems) {
2048                        if (i == *level)
2049                                break;
2050                        *level = i + 1;
2051                        return 0;
2052                }
2053                if (i == 1) {
2054                        *level = i;
2055                        return 0;
2056                }
2057
2058                bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2059                eb = read_tree_block(fs_info, bytenr, ptr_gen);
2060                if (IS_ERR(eb)) {
2061                        return PTR_ERR(eb);
2062                } else if (!extent_buffer_uptodate(eb)) {
2063                        free_extent_buffer(eb);
2064                        return -EIO;
2065                }
2066                BUG_ON(btrfs_header_level(eb) != i - 1);
2067                path->nodes[i - 1] = eb;
2068                path->slots[i - 1] = 0;
2069        }
2070        return 1;
2071}
2072
2073/*
2074 * invalidate extent cache for file extents whose key in range of
2075 * [min_key, max_key)
2076 */
2077static int invalidate_extent_cache(struct btrfs_root *root,
2078                                   struct btrfs_key *min_key,
2079                                   struct btrfs_key *max_key)
2080{
2081        struct btrfs_fs_info *fs_info = root->fs_info;
2082        struct inode *inode = NULL;
2083        u64 objectid;
2084        u64 start, end;
2085        u64 ino;
2086
2087        objectid = min_key->objectid;
2088        while (1) {
2089                cond_resched();
2090                iput(inode);
2091
2092                if (objectid > max_key->objectid)
2093                        break;
2094
2095                inode = find_next_inode(root, objectid);
2096                if (!inode)
2097                        break;
2098                ino = btrfs_ino(BTRFS_I(inode));
2099
2100                if (ino > max_key->objectid) {
2101                        iput(inode);
2102                        break;
2103                }
2104
2105                objectid = ino + 1;
2106                if (!S_ISREG(inode->i_mode))
2107                        continue;
2108
2109                if (unlikely(min_key->objectid == ino)) {
2110                        if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2111                                continue;
2112                        if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2113                                start = 0;
2114                        else {
2115                                start = min_key->offset;
2116                                WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
2117                        }
2118                } else {
2119                        start = 0;
2120                }
2121
2122                if (unlikely(max_key->objectid == ino)) {
2123                        if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2124                                continue;
2125                        if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2126                                end = (u64)-1;
2127                        } else {
2128                                if (max_key->offset == 0)
2129                                        continue;
2130                                end = max_key->offset;
2131                                WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
2132                                end--;
2133                        }
2134                } else {
2135                        end = (u64)-1;
2136                }
2137
2138                /* the lock_extent waits for readpage to complete */
2139                lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2140                btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
2141                unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2142        }
2143        return 0;
2144}
2145
2146static int find_next_key(struct btrfs_path *path, int level,
2147                         struct btrfs_key *key)
2148
2149{
2150        while (level < BTRFS_MAX_LEVEL) {
2151                if (!path->nodes[level])
2152                        break;
2153                if (path->slots[level] + 1 <
2154                    btrfs_header_nritems(path->nodes[level])) {
2155                        btrfs_node_key_to_cpu(path->nodes[level], key,
2156                                              path->slots[level] + 1);
2157                        return 0;
2158                }
2159                level++;
2160        }
2161        return 1;
2162}
2163
2164/*
2165 * merge the relocated tree blocks in reloc tree with corresponding
2166 * fs tree.
2167 */
2168static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2169                                               struct btrfs_root *root)
2170{
2171        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2172        LIST_HEAD(inode_list);
2173        struct btrfs_key key;
2174        struct btrfs_key next_key;
2175        struct btrfs_trans_handle *trans = NULL;
2176        struct btrfs_root *reloc_root;
2177        struct btrfs_root_item *root_item;
2178        struct btrfs_path *path;
2179        struct extent_buffer *leaf;
2180        int level;
2181        int max_level;
2182        int replaced = 0;
2183        int ret;
2184        int err = 0;
2185        u32 min_reserved;
2186
2187        path = btrfs_alloc_path();
2188        if (!path)
2189                return -ENOMEM;
2190        path->reada = READA_FORWARD;
2191
2192        reloc_root = root->reloc_root;
2193        root_item = &reloc_root->root_item;
2194
2195        if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2196                level = btrfs_root_level(root_item);
2197                extent_buffer_get(reloc_root->node);
2198                path->nodes[level] = reloc_root->node;
2199                path->slots[level] = 0;
2200        } else {
2201                btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2202
2203                level = root_item->drop_level;
2204                BUG_ON(level == 0);
2205                path->lowest_level = level;
2206                ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2207                path->lowest_level = 0;
2208                if (ret < 0) {
2209                        btrfs_free_path(path);
2210                        return ret;
2211                }
2212
2213                btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2214                                      path->slots[level]);
2215                WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2216
2217                btrfs_unlock_up_safe(path, 0);
2218        }
2219
2220        min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2221        memset(&next_key, 0, sizeof(next_key));
2222
2223        while (1) {
2224                ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2225                                             BTRFS_RESERVE_FLUSH_ALL);
2226                if (ret) {
2227                        err = ret;
2228                        goto out;
2229                }
2230                trans = btrfs_start_transaction(root, 0);
2231                if (IS_ERR(trans)) {
2232                        err = PTR_ERR(trans);
2233                        trans = NULL;
2234                        goto out;
2235                }
2236                trans->block_rsv = rc->block_rsv;
2237
2238                replaced = 0;
2239                max_level = level;
2240
2241                ret = walk_down_reloc_tree(reloc_root, path, &level);
2242                if (ret < 0) {
2243                        err = ret;
2244                        goto out;
2245                }
2246                if (ret > 0)
2247                        break;
2248
2249                if (!find_next_key(path, level, &key) &&
2250                    btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2251                        ret = 0;
2252                } else {
2253                        ret = replace_path(trans, root, reloc_root, path,
2254                                           &next_key, level, max_level);
2255                }
2256                if (ret < 0) {
2257                        err = ret;
2258                        goto out;
2259                }
2260
2261                if (ret > 0) {
2262                        level = ret;
2263                        btrfs_node_key_to_cpu(path->nodes[level], &key,
2264                                              path->slots[level]);
2265                        replaced = 1;
2266                }
2267
2268                ret = walk_up_reloc_tree(reloc_root, path, &level);
2269                if (ret > 0)
2270                        break;
2271
2272                BUG_ON(level == 0);
2273                /*
2274                 * save the merging progress in the drop_progress.
2275                 * this is OK since root refs == 1 in this case.
2276                 */
2277                btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2278                               path->slots[level]);
2279                root_item->drop_level = level;
2280
2281                btrfs_end_transaction_throttle(trans);
2282                trans = NULL;
2283
2284                btrfs_btree_balance_dirty(fs_info);
2285
2286                if (replaced && rc->stage == UPDATE_DATA_PTRS)
2287                        invalidate_extent_cache(root, &key, &next_key);
2288        }
2289
2290        /*
2291         * handle the case only one block in the fs tree need to be
2292         * relocated and the block is tree root.
2293         */
2294        leaf = btrfs_lock_root_node(root);
2295        ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2296        btrfs_tree_unlock(leaf);
2297        free_extent_buffer(leaf);
2298        if (ret < 0)
2299                err = ret;
2300out:
2301        btrfs_free_path(path);
2302
2303        if (err == 0) {
2304                memset(&root_item->drop_progress, 0,
2305                       sizeof(root_item->drop_progress));
2306                root_item->drop_level = 0;
2307                btrfs_set_root_refs(root_item, 0);
2308                btrfs_update_reloc_root(trans, root);
2309        }
2310
2311        if (trans)
2312                btrfs_end_transaction_throttle(trans);
2313
2314        btrfs_btree_balance_dirty(fs_info);
2315
2316        if (replaced && rc->stage == UPDATE_DATA_PTRS)
2317                invalidate_extent_cache(root, &key, &next_key);
2318
2319        return err;
2320}
2321
2322static noinline_for_stack
2323int prepare_to_merge(struct reloc_control *rc, int err)
2324{
2325        struct btrfs_root *root = rc->extent_root;
2326        struct btrfs_fs_info *fs_info = root->fs_info;
2327        struct btrfs_root *reloc_root;
2328        struct btrfs_trans_handle *trans;
2329        LIST_HEAD(reloc_roots);
2330        u64 num_bytes = 0;
2331        int ret;
2332
2333        mutex_lock(&fs_info->reloc_mutex);
2334        rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2335        rc->merging_rsv_size += rc->nodes_relocated * 2;
2336        mutex_unlock(&fs_info->reloc_mutex);
2337
2338again:
2339        if (!err) {
2340                num_bytes = rc->merging_rsv_size;
2341                ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2342                                          BTRFS_RESERVE_FLUSH_ALL);
2343                if (ret)
2344                        err = ret;
2345        }
2346
2347        trans = btrfs_join_transaction(rc->extent_root);
2348        if (IS_ERR(trans)) {
2349                if (!err)
2350                        btrfs_block_rsv_release(fs_info, rc->block_rsv,
2351                                                num_bytes);
2352                return PTR_ERR(trans);
2353        }
2354
2355        if (!err) {
2356                if (num_bytes != rc->merging_rsv_size) {
2357                        btrfs_end_transaction(trans);
2358                        btrfs_block_rsv_release(fs_info, rc->block_rsv,
2359                                                num_bytes);
2360                        goto again;
2361                }
2362        }
2363
2364        rc->merge_reloc_tree = 1;
2365
2366        while (!list_empty(&rc->reloc_roots)) {
2367                reloc_root = list_entry(rc->reloc_roots.next,
2368                                        struct btrfs_root, root_list);
2369                list_del_init(&reloc_root->root_list);
2370
2371                root = read_fs_root(fs_info, reloc_root->root_key.offset);
2372                BUG_ON(IS_ERR(root));
2373                BUG_ON(root->reloc_root != reloc_root);
2374
2375                /*
2376                 * set reference count to 1, so btrfs_recover_relocation
2377                 * knows it should resumes merging
2378                 */
2379                if (!err)
2380                        btrfs_set_root_refs(&reloc_root->root_item, 1);
2381                btrfs_update_reloc_root(trans, root);
2382
2383                list_add(&reloc_root->root_list, &reloc_roots);
2384        }
2385
2386        list_splice(&reloc_roots, &rc->reloc_roots);
2387
2388        if (!err)
2389                btrfs_commit_transaction(trans);
2390        else
2391                btrfs_end_transaction(trans);
2392        return err;
2393}
2394
2395static noinline_for_stack
2396void free_reloc_roots(struct list_head *list)
2397{
2398        struct btrfs_root *reloc_root;
2399
2400        while (!list_empty(list)) {
2401                reloc_root = list_entry(list->next, struct btrfs_root,
2402                                        root_list);
2403                __del_reloc_root(reloc_root);
2404                free_extent_buffer(reloc_root->node);
2405                free_extent_buffer(reloc_root->commit_root);
2406                reloc_root->node = NULL;
2407                reloc_root->commit_root = NULL;
2408        }
2409}
2410
2411static noinline_for_stack
2412void merge_reloc_roots(struct reloc_control *rc)
2413{
2414        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2415        struct btrfs_root *root;
2416        struct btrfs_root *reloc_root;
2417        LIST_HEAD(reloc_roots);
2418        int found = 0;
2419        int ret = 0;
2420again:
2421        root = rc->extent_root;
2422
2423        /*
2424         * this serializes us with btrfs_record_root_in_transaction,
2425         * we have to make sure nobody is in the middle of
2426         * adding their roots to the list while we are
2427         * doing this splice
2428         */
2429        mutex_lock(&fs_info->reloc_mutex);
2430        list_splice_init(&rc->reloc_roots, &reloc_roots);
2431        mutex_unlock(&fs_info->reloc_mutex);
2432
2433        while (!list_empty(&reloc_roots)) {
2434                found = 1;
2435                reloc_root = list_entry(reloc_roots.next,
2436                                        struct btrfs_root, root_list);
2437
2438                if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2439                        root = read_fs_root(fs_info,
2440                                            reloc_root->root_key.offset);
2441                        BUG_ON(IS_ERR(root));
2442                        BUG_ON(root->reloc_root != reloc_root);
2443
2444                        ret = merge_reloc_root(rc, root);
2445                        if (ret) {
2446                                if (list_empty(&reloc_root->root_list))
2447                                        list_add_tail(&reloc_root->root_list,
2448                                                      &reloc_roots);
2449                                goto out;
2450                        }
2451                } else {
2452                        list_del_init(&reloc_root->root_list);
2453                }
2454
2455                ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2456                if (ret < 0) {
2457                        if (list_empty(&reloc_root->root_list))
2458                                list_add_tail(&reloc_root->root_list,
2459                                              &reloc_roots);
2460                        goto out;
2461                }
2462        }
2463
2464        if (found) {
2465                found = 0;
2466                goto again;
2467        }
2468out:
2469        if (ret) {
2470                btrfs_handle_fs_error(fs_info, ret, NULL);
2471                if (!list_empty(&reloc_roots))
2472                        free_reloc_roots(&reloc_roots);
2473
2474                /* new reloc root may be added */
2475                mutex_lock(&fs_info->reloc_mutex);
2476                list_splice_init(&rc->reloc_roots, &reloc_roots);
2477                mutex_unlock(&fs_info->reloc_mutex);
2478                if (!list_empty(&reloc_roots))
2479                        free_reloc_roots(&reloc_roots);
2480        }
2481
2482        BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2483}
2484
2485static void free_block_list(struct rb_root *blocks)
2486{
2487        struct tree_block *block;
2488        struct rb_node *rb_node;
2489        while ((rb_node = rb_first(blocks))) {
2490                block = rb_entry(rb_node, struct tree_block, rb_node);
2491                rb_erase(rb_node, blocks);
2492                kfree(block);
2493        }
2494}
2495
2496static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2497                                      struct btrfs_root *reloc_root)
2498{
2499        struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2500        struct btrfs_root *root;
2501
2502        if (reloc_root->last_trans == trans->transid)
2503                return 0;
2504
2505        root = read_fs_root(fs_info, reloc_root->root_key.offset);
2506        BUG_ON(IS_ERR(root));
2507        BUG_ON(root->reloc_root != reloc_root);
2508
2509        return btrfs_record_root_in_trans(trans, root);
2510}
2511
2512static noinline_for_stack
2513struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2514                                     struct reloc_control *rc,
2515                                     struct backref_node *node,
2516                                     struct backref_edge *edges[])
2517{
2518        struct backref_node *next;
2519        struct btrfs_root *root;
2520        int index = 0;
2521
2522        next = node;
2523        while (1) {
2524                cond_resched();
2525                next = walk_up_backref(next, edges, &index);
2526                root = next->root;
2527                BUG_ON(!root);
2528                BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2529
2530                if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2531                        record_reloc_root_in_trans(trans, root);
2532                        break;
2533                }
2534
2535                btrfs_record_root_in_trans(trans, root);
2536                root = root->reloc_root;
2537
2538                if (next->new_bytenr != root->node->start) {
2539                        BUG_ON(next->new_bytenr);
2540                        BUG_ON(!list_empty(&next->list));
2541                        next->new_bytenr = root->node->start;
2542                        next->root = root;
2543                        list_add_tail(&next->list,
2544                                      &rc->backref_cache.changed);
2545                        __mark_block_processed(rc, next);
2546                        break;
2547                }
2548
2549                WARN_ON(1);
2550                root = NULL;
2551                next = walk_down_backref(edges, &index);
2552                if (!next || next->level <= node->level)
2553                        break;
2554        }
2555        if (!root)
2556                return NULL;
2557
2558        next = node;
2559        /* setup backref node path for btrfs_reloc_cow_block */
2560        while (1) {
2561                rc->backref_cache.path[next->level] = next;
2562                if (--index < 0)
2563                        break;
2564                next = edges[index]->node[UPPER];
2565        }
2566        return root;
2567}
2568
2569/*
2570 * select a tree root for relocation. return NULL if the block
2571 * is reference counted. we should use do_relocation() in this
2572 * case. return a tree root pointer if the block isn't reference
2573 * counted. return -ENOENT if the block is root of reloc tree.
2574 */
2575static noinline_for_stack
2576struct btrfs_root *select_one_root(struct backref_node *node)
2577{
2578        struct backref_node *next;
2579        struct btrfs_root *root;
2580        struct btrfs_root *fs_root = NULL;
2581        struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2582        int index = 0;
2583
2584        next = node;
2585        while (1) {
2586                cond_resched();
2587                next = walk_up_backref(next, edges, &index);
2588                root = next->root;
2589                BUG_ON(!root);
2590
2591                /* no other choice for non-references counted tree */
2592                if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2593                        return root;
2594
2595                if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2596                        fs_root = root;
2597
2598                if (next != node)
2599                        return NULL;
2600
2601                next = walk_down_backref(edges, &index);
2602                if (!next || next->level <= node->level)
2603                        break;
2604        }
2605
2606        if (!fs_root)
2607                return ERR_PTR(-ENOENT);
2608        return fs_root;
2609}
2610
2611static noinline_for_stack
2612u64 calcu_metadata_size(struct reloc_control *rc,
2613                        struct backref_node *node, int reserve)
2614{
2615        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2616        struct backref_node *next = node;
2617        struct backref_edge *edge;
2618        struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2619        u64 num_bytes = 0;
2620        int index = 0;
2621
2622        BUG_ON(reserve && node->processed);
2623
2624        while (next) {
2625                cond_resched();
2626                while (1) {
2627                        if (next->processed && (reserve || next != node))
2628                                break;
2629
2630                        num_bytes += fs_info->nodesize;
2631
2632                        if (list_empty(&next->upper))
2633                                break;
2634
2635                        edge = list_entry(next->upper.next,
2636                                          struct backref_edge, list[LOWER]);
2637                        edges[index++] = edge;
2638                        next = edge->node[UPPER];
2639                }
2640                next = walk_down_backref(edges, &index);
2641        }
2642        return num_bytes;
2643}
2644
2645static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2646                                  struct reloc_control *rc,
2647                                  struct backref_node *node)
2648{
2649        struct btrfs_root *root = rc->extent_root;
2650        struct btrfs_fs_info *fs_info = root->fs_info;
2651        u64 num_bytes;
2652        int ret;
2653        u64 tmp;
2654
2655        num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2656
2657        trans->block_rsv = rc->block_rsv;
2658        rc->reserved_bytes += num_bytes;
2659
2660        /*
2661         * We are under a transaction here so we can only do limited flushing.
2662         * If we get an enospc just kick back -EAGAIN so we know to drop the
2663         * transaction and try to refill when we can flush all the things.
2664         */
2665        ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2666                                BTRFS_RESERVE_FLUSH_LIMIT);
2667        if (ret) {
2668                tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2669                while (tmp <= rc->reserved_bytes)
2670                        tmp <<= 1;
2671                /*
2672                 * only one thread can access block_rsv at this point,
2673                 * so we don't need hold lock to protect block_rsv.
2674                 * we expand more reservation size here to allow enough
2675                 * space for relocation and we will return eailer in
2676                 * enospc case.
2677                 */
2678                rc->block_rsv->size = tmp + fs_info->nodesize *
2679                                      RELOCATION_RESERVED_NODES;
2680                return -EAGAIN;
2681        }
2682
2683        return 0;
2684}
2685
2686/*
2687 * relocate a block tree, and then update pointers in upper level
2688 * blocks that reference the block to point to the new location.
2689 *
2690 * if called by link_to_upper, the block has already been relocated.
2691 * in that case this function just updates pointers.
2692 */
2693static int do_relocation(struct btrfs_trans_handle *trans,
2694                         struct reloc_control *rc,
2695                         struct backref_node *node,
2696                         struct btrfs_key *key,
2697                         struct btrfs_path *path, int lowest)
2698{
2699        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2700        struct backref_node *upper;
2701        struct backref_edge *edge;
2702        struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2703        struct btrfs_root *root;
2704        struct extent_buffer *eb;
2705        u32 blocksize;
2706        u64 bytenr;
2707        u64 generation;
2708        int slot;
2709        int ret;
2710        int err = 0;
2711
2712        BUG_ON(lowest && node->eb);
2713
2714        path->lowest_level = node->level + 1;
2715        rc->backref_cache.path[node->level] = node;
2716        list_for_each_entry(edge, &node->upper, list[LOWER]) {
2717                cond_resched();
2718
2719                upper = edge->node[UPPER];
2720                root = select_reloc_root(trans, rc, upper, edges);
2721                BUG_ON(!root);
2722
2723                if (upper->eb && !upper->locked) {
2724                        if (!lowest) {
2725                                ret = btrfs_bin_search(upper->eb, key,
2726                                                       upper->level, &slot);
2727                                BUG_ON(ret);
2728                                bytenr = btrfs_node_blockptr(upper->eb, slot);
2729                                if (node->eb->start == bytenr)
2730                                        goto next;
2731                        }
2732                        drop_node_buffer(upper);
2733                }
2734
2735                if (!upper->eb) {
2736                        ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2737                        if (ret) {
2738                                if (ret < 0)
2739                                        err = ret;
2740                                else
2741                                        err = -ENOENT;
2742
2743                                btrfs_release_path(path);
2744                                break;
2745                        }
2746
2747                        if (!upper->eb) {
2748                                upper->eb = path->nodes[upper->level];
2749                                path->nodes[upper->level] = NULL;
2750                        } else {
2751                                BUG_ON(upper->eb != path->nodes[upper->level]);
2752                        }
2753
2754                        upper->locked = 1;
2755                        path->locks[upper->level] = 0;
2756
2757                        slot = path->slots[upper->level];
2758                        btrfs_release_path(path);
2759                } else {
2760                        ret = btrfs_bin_search(upper->eb, key, upper->level,
2761                                               &slot);
2762                        BUG_ON(ret);
2763                }
2764
2765                bytenr = btrfs_node_blockptr(upper->eb, slot);
2766                if (lowest) {
2767                        if (bytenr != node->bytenr) {
2768                                btrfs_err(root->fs_info,
2769                "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2770                                          bytenr, node->bytenr, slot,
2771                                          upper->eb->start);
2772                                err = -EIO;
2773                                goto next;
2774                        }
2775                } else {
2776                        if (node->eb->start == bytenr)
2777                                goto next;
2778                }
2779
2780                blocksize = root->fs_info->nodesize;
2781                generation = btrfs_node_ptr_generation(upper->eb, slot);
2782                eb = read_tree_block(fs_info, bytenr, generation);
2783                if (IS_ERR(eb)) {
2784                        err = PTR_ERR(eb);
2785                        goto next;
2786                } else if (!extent_buffer_uptodate(eb)) {
2787                        free_extent_buffer(eb);
2788                        err = -EIO;
2789                        goto next;
2790                }
2791                btrfs_tree_lock(eb);
2792                btrfs_set_lock_blocking(eb);
2793
2794                if (!node->eb) {
2795                        ret = btrfs_cow_block(trans, root, eb, upper->eb,
2796                                              slot, &eb);
2797                        btrfs_tree_unlock(eb);
2798                        free_extent_buffer(eb);
2799                        if (ret < 0) {
2800                                err = ret;
2801                                goto next;
2802                        }
2803                        BUG_ON(node->eb != eb);
2804                } else {
2805                        btrfs_set_node_blockptr(upper->eb, slot,
2806                                                node->eb->start);
2807                        btrfs_set_node_ptr_generation(upper->eb, slot,
2808                                                      trans->transid);
2809                        btrfs_mark_buffer_dirty(upper->eb);
2810
2811                        ret = btrfs_inc_extent_ref(trans, root,
2812                                                node->eb->start, blocksize,
2813                                                upper->eb->start,
2814                                                btrfs_header_owner(upper->eb),
2815                                                node->level, 0);
2816                        BUG_ON(ret);
2817
2818                        ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2819                        BUG_ON(ret);
2820                }
2821next:
2822                if (!upper->pending)
2823                        drop_node_buffer(upper);
2824                else
2825                        unlock_node_buffer(upper);
2826                if (err)
2827                        break;
2828        }
2829
2830        if (!err && node->pending) {
2831                drop_node_buffer(node);
2832                list_move_tail(&node->list, &rc->backref_cache.changed);
2833                node->pending = 0;
2834        }
2835
2836        path->lowest_level = 0;
2837        BUG_ON(err == -ENOSPC);
2838        return err;
2839}
2840
2841static int link_to_upper(struct btrfs_trans_handle *trans,
2842                         struct reloc_control *rc,
2843                         struct backref_node *node,
2844                         struct btrfs_path *path)
2845{
2846        struct btrfs_key key;
2847
2848        btrfs_node_key_to_cpu(node->eb, &key, 0);
2849        return do_relocation(trans, rc, node, &key, path, 0);
2850}
2851
2852static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2853                                struct reloc_control *rc,
2854                                struct btrfs_path *path, int err)
2855{
2856        LIST_HEAD(list);
2857        struct backref_cache *cache = &rc->backref_cache;
2858        struct backref_node *node;
2859        int level;
2860        int ret;
2861
2862        for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2863                while (!list_empty(&cache->pending[level])) {
2864                        node = list_entry(cache->pending[level].next,
2865                                          struct backref_node, list);
2866                        list_move_tail(&node->list, &list);
2867                        BUG_ON(!node->pending);
2868
2869                        if (!err) {
2870                                ret = link_to_upper(trans, rc, node, path);
2871                                if (ret < 0)
2872                                        err = ret;
2873                        }
2874                }
2875                list_splice_init(&list, &cache->pending[level]);
2876        }
2877        return err;
2878}
2879
2880static void mark_block_processed(struct reloc_control *rc,
2881                                 u64 bytenr, u32 blocksize)
2882{
2883        set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2884                        EXTENT_DIRTY);
2885}
2886
2887static void __mark_block_processed(struct reloc_control *rc,
2888                                   struct backref_node *node)
2889{
2890        u32 blocksize;
2891        if (node->level == 0 ||
2892            in_block_group(node->bytenr, rc->block_group)) {
2893                blocksize = rc->extent_root->fs_info->nodesize;
2894                mark_block_processed(rc, node->bytenr, blocksize);
2895        }
2896        node->processed = 1;
2897}
2898
2899/*
2900 * mark a block and all blocks directly/indirectly reference the block
2901 * as processed.
2902 */
2903static void update_processed_blocks(struct reloc_control *rc,
2904                                    struct backref_node *node)
2905{
2906        struct backref_node *next = node;
2907        struct backref_edge *edge;
2908        struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2909        int index = 0;
2910
2911        while (next) {
2912                cond_resched();
2913                while (1) {
2914                        if (next->processed)
2915                                break;
2916
2917                        __mark_block_processed(rc, next);
2918
2919                        if (list_empty(&next->upper))
2920                                break;
2921
2922                        edge = list_entry(next->upper.next,
2923                                          struct backref_edge, list[LOWER]);
2924                        edges[index++] = edge;
2925                        next = edge->node[UPPER];
2926                }
2927                next = walk_down_backref(edges, &index);
2928        }
2929}
2930
2931static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2932{
2933        u32 blocksize = rc->extent_root->fs_info->nodesize;
2934
2935        if (test_range_bit(&rc->processed_blocks, bytenr,
2936                           bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2937                return 1;
2938        return 0;
2939}
2940
2941static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2942                              struct tree_block *block)
2943{
2944        struct extent_buffer *eb;
2945
2946        BUG_ON(block->key_ready);
2947        eb = read_tree_block(fs_info, block->bytenr, block->key.offset);
2948        if (IS_ERR(eb)) {
2949                return PTR_ERR(eb);
2950        } else if (!extent_buffer_uptodate(eb)) {
2951                free_extent_buffer(eb);
2952                return -EIO;
2953        }
2954        WARN_ON(btrfs_header_level(eb) != block->level);
2955        if (block->level == 0)
2956                btrfs_item_key_to_cpu(eb, &block->key, 0);
2957        else
2958                btrfs_node_key_to_cpu(eb, &block->key, 0);
2959        free_extent_buffer(eb);
2960        block->key_ready = 1;
2961        return 0;
2962}
2963
2964/*
2965 * helper function to relocate a tree block
2966 */
2967static int relocate_tree_block(struct btrfs_trans_handle *trans,
2968                                struct reloc_control *rc,
2969                                struct backref_node *node,
2970                                struct btrfs_key *key,
2971                                struct btrfs_path *path)
2972{
2973        struct btrfs_root *root;
2974        int ret = 0;
2975
2976        if (!node)
2977                return 0;
2978
2979        BUG_ON(node->processed);
2980        root = select_one_root(node);
2981        if (root == ERR_PTR(-ENOENT)) {
2982                update_processed_blocks(rc, node);
2983                goto out;
2984        }
2985
2986        if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2987                ret = reserve_metadata_space(trans, rc, node);
2988                if (ret)
2989                        goto out;
2990        }
2991
2992        if (root) {
2993                if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2994                        BUG_ON(node->new_bytenr);
2995                        BUG_ON(!list_empty(&node->list));
2996                        btrfs_record_root_in_trans(trans, root);
2997                        root = root->reloc_root;
2998                        node->new_bytenr = root->node->start;
2999                        node->root = root;
3000                        list_add_tail(&node->list, &rc->backref_cache.changed);
3001                } else {
3002                        path->lowest_level = node->level;
3003                        ret = btrfs_search_slot(trans, root, key, path, 0, 1);
3004                        btrfs_release_path(path);
3005                        if (ret > 0)
3006                                ret = 0;
3007                }
3008                if (!ret)
3009                        update_processed_blocks(rc, node);
3010        } else {
3011                ret = do_relocation(trans, rc, node, key, path, 1);
3012        }
3013out:
3014        if (ret || node->level == 0 || node->cowonly)
3015                remove_backref_node(&rc->backref_cache, node);
3016        return ret;
3017}
3018
3019/*
3020 * relocate a list of blocks
3021 */
3022static noinline_for_stack
3023int relocate_tree_blocks(struct btrfs_trans_handle *trans,
3024                         struct reloc_control *rc, struct rb_root *blocks)
3025{
3026        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3027        struct backref_node *node;
3028        struct btrfs_path *path;
3029        struct tree_block *block;
3030        struct rb_node *rb_node;
3031        int ret;
3032        int err = 0;
3033
3034        path = btrfs_alloc_path();
3035        if (!path) {
3036                err = -ENOMEM;
3037                goto out_free_blocks;
3038        }
3039
3040        rb_node = rb_first(blocks);
3041        while (rb_node) {
3042                block = rb_entry(rb_node, struct tree_block, rb_node);
3043                if (!block->key_ready)
3044                        readahead_tree_block(fs_info, block->bytenr);
3045                rb_node = rb_next(rb_node);
3046        }
3047
3048        rb_node = rb_first(blocks);
3049        while (rb_node) {
3050                block = rb_entry(rb_node, struct tree_block, rb_node);
3051                if (!block->key_ready) {
3052                        err = get_tree_block_key(fs_info, block);
3053                        if (err)
3054                                goto out_free_path;
3055                }
3056                rb_node = rb_next(rb_node);
3057        }
3058
3059        rb_node = rb_first(blocks);
3060        while (rb_node) {
3061                block = rb_entry(rb_node, struct tree_block, rb_node);
3062
3063                node = build_backref_tree(rc, &block->key,
3064                                          block->level, block->bytenr);
3065                if (IS_ERR(node)) {
3066                        err = PTR_ERR(node);
3067                        goto out;
3068                }
3069
3070                ret = relocate_tree_block(trans, rc, node, &block->key,
3071                                          path);
3072                if (ret < 0) {
3073                        if (ret != -EAGAIN || rb_node == rb_first(blocks))
3074                                err = ret;
3075                        goto out;
3076                }
3077                rb_node = rb_next(rb_node);
3078        }
3079out:
3080        err = finish_pending_nodes(trans, rc, path, err);
3081
3082out_free_path:
3083        btrfs_free_path(path);
3084out_free_blocks:
3085        free_block_list(blocks);
3086        return err;
3087}
3088
3089static noinline_for_stack
3090int prealloc_file_extent_cluster(struct inode *inode,
3091                                 struct file_extent_cluster *cluster)
3092{
3093        u64 alloc_hint = 0;
3094        u64 start;
3095        u64 end;
3096        u64 offset = BTRFS_I(inode)->index_cnt;
3097        u64 num_bytes;
3098        int nr = 0;
3099        int ret = 0;
3100        u64 prealloc_start = cluster->start - offset;
3101        u64 prealloc_end = cluster->end - offset;
3102        u64 cur_offset;
3103        struct extent_changeset *data_reserved = NULL;
3104
3105        BUG_ON(cluster->start != cluster->boundary[0]);
3106        inode_lock(inode);
3107
3108        ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start,
3109                                          prealloc_end + 1 - prealloc_start);
3110        if (ret)
3111                goto out;
3112
3113        cur_offset = prealloc_start;
3114        while (nr < cluster->nr) {
3115                start = cluster->boundary[nr] - offset;
3116                if (nr + 1 < cluster->nr)
3117                        end = cluster->boundary[nr + 1] - 1 - offset;
3118                else
3119                        end = cluster->end - offset;
3120
3121                lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3122                num_bytes = end + 1 - start;
3123                if (cur_offset < start)
3124                        btrfs_free_reserved_data_space(inode, data_reserved,
3125                                        cur_offset, start - cur_offset);
3126                ret = btrfs_prealloc_file_range(inode, 0, start,
3127                                                num_bytes, num_bytes,
3128                                                end + 1, &alloc_hint);
3129                cur_offset = end + 1;
3130                unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3131                if (ret)
3132                        break;
3133                nr++;
3134        }
3135        if (cur_offset < prealloc_end)
3136                btrfs_free_reserved_data_space(inode, data_reserved,
3137                                cur_offset, prealloc_end + 1 - cur_offset);
3138out:
3139        inode_unlock(inode);
3140        extent_changeset_free(data_reserved);
3141        return ret;
3142}
3143
3144static noinline_for_stack
3145int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3146                         u64 block_start)
3147{
3148        struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3149        struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3150        struct extent_map *em;
3151        int ret = 0;
3152
3153        em = alloc_extent_map();
3154        if (!em)
3155                return -ENOMEM;
3156
3157        em->start = start;
3158        em->len = end + 1 - start;
3159        em->block_len = em->len;
3160        em->block_start = block_start;
3161        em->bdev = fs_info->fs_devices->latest_bdev;
3162        set_bit(EXTENT_FLAG_PINNED, &em->flags);
3163
3164        lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3165        while (1) {
3166                write_lock(&em_tree->lock);
3167                ret = add_extent_mapping(em_tree, em, 0);
3168                write_unlock(&em_tree->lock);
3169                if (ret != -EEXIST) {
3170                        free_extent_map(em);
3171                        break;
3172                }
3173                btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3174        }
3175        unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3176        return ret;
3177}
3178
3179static int relocate_file_extent_cluster(struct inode *inode,
3180                                        struct file_extent_cluster *cluster)
3181{
3182        struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3183        u64 page_start;
3184        u64 page_end;
3185        u64 offset = BTRFS_I(inode)->index_cnt;
3186        unsigned long index;
3187        unsigned long last_index;
3188        struct page *page;
3189        struct file_ra_state *ra;
3190        gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3191        int nr = 0;
3192        int ret = 0;
3193
3194        if (!cluster->nr)
3195                return 0;
3196
3197        ra = kzalloc(sizeof(*ra), GFP_NOFS);
3198        if (!ra)
3199                return -ENOMEM;
3200
3201        ret = prealloc_file_extent_cluster(inode, cluster);
3202        if (ret)
3203                goto out;
3204
3205        file_ra_state_init(ra, inode->i_mapping);
3206
3207        ret = setup_extent_mapping(inode, cluster->start - offset,
3208                                   cluster->end - offset, cluster->start);
3209        if (ret)
3210                goto out;
3211
3212        index = (cluster->start - offset) >> PAGE_SHIFT;
3213        last_index = (cluster->end - offset) >> PAGE_SHIFT;
3214        while (index <= last_index) {
3215                ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3216                                PAGE_SIZE);
3217                if (ret)
3218                        goto out;
3219
3220                page = find_lock_page(inode->i_mapping, index);
3221                if (!page) {
3222                        page_cache_sync_readahead(inode->i_mapping,
3223                                                  ra, NULL, index,
3224                                                  last_index + 1 - index);
3225                        page = find_or_create_page(inode->i_mapping, index,
3226                                                   mask);
3227                        if (!page) {
3228                                btrfs_delalloc_release_metadata(BTRFS_I(inode),
3229                                                        PAGE_SIZE);
3230                                ret = -ENOMEM;
3231                                goto out;
3232                        }
3233                }
3234
3235                if (PageReadahead(page)) {
3236                        page_cache_async_readahead(inode->i_mapping,
3237                                                   ra, NULL, page, index,
3238                                                   last_index + 1 - index);
3239                }
3240
3241                if (!PageUptodate(page)) {
3242                        btrfs_readpage(NULL, page);
3243                        lock_page(page);
3244                        if (!PageUptodate(page)) {
3245                                unlock_page(page);
3246                                put_page(page);
3247                                btrfs_delalloc_release_metadata(BTRFS_I(inode),
3248                                                        PAGE_SIZE);
3249                                btrfs_delalloc_release_extents(BTRFS_I(inode),
3250                                                               PAGE_SIZE);
3251                                ret = -EIO;
3252                                goto out;
3253                        }
3254                }
3255
3256                page_start = page_offset(page);
3257                page_end = page_start + PAGE_SIZE - 1;
3258
3259                lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3260
3261                set_page_extent_mapped(page);
3262
3263                if (nr < cluster->nr &&
3264                    page_start + offset == cluster->boundary[nr]) {
3265                        set_extent_bits(&BTRFS_I(inode)->io_tree,
3266                                        page_start, page_end,
3267                                        EXTENT_BOUNDARY);
3268                        nr++;
3269                }
3270
3271                btrfs_set_extent_delalloc(inode, page_start, page_end, 0, NULL,
3272                                          0);
3273                set_page_dirty(page);
3274
3275                unlock_extent(&BTRFS_I(inode)->io_tree,
3276                              page_start, page_end);
3277                unlock_page(page);
3278                put_page(page);
3279
3280                index++;
3281                btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
3282                balance_dirty_pages_ratelimited(inode->i_mapping);
3283                btrfs_throttle(fs_info);
3284        }
3285        WARN_ON(nr != cluster->nr);
3286out:
3287        kfree(ra);
3288        return ret;
3289}
3290
3291static noinline_for_stack
3292int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3293                         struct file_extent_cluster *cluster)
3294{
3295        int ret;
3296
3297        if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3298                ret = relocate_file_extent_cluster(inode, cluster);
3299                if (ret)
3300                        return ret;
3301                cluster->nr = 0;
3302        }
3303
3304        if (!cluster->nr)
3305                cluster->start = extent_key->objectid;
3306        else
3307                BUG_ON(cluster->nr >= MAX_EXTENTS);
3308        cluster->end = extent_key->objectid + extent_key->offset - 1;
3309        cluster->boundary[cluster->nr] = extent_key->objectid;
3310        cluster->nr++;
3311
3312        if (cluster->nr >= MAX_EXTENTS) {
3313                ret = relocate_file_extent_cluster(inode, cluster);
3314                if (ret)
3315                        return ret;
3316                cluster->nr = 0;
3317        }
3318        return 0;
3319}
3320
3321#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3322static int get_ref_objectid_v0(struct reloc_control *rc,
3323                               struct btrfs_path *path,
3324                               struct btrfs_key *extent_key,
3325                               u64 *ref_objectid, int *path_change)
3326{
3327        struct btrfs_key key;
3328        struct extent_buffer *leaf;
3329        struct btrfs_extent_ref_v0 *ref0;
3330        int ret;
3331        int slot;
3332
3333        leaf = path->nodes[0];
3334        slot = path->slots[0];
3335        while (1) {
3336                if (slot >= btrfs_header_nritems(leaf)) {
3337                        ret = btrfs_next_leaf(rc->extent_root, path);
3338                        if (ret < 0)
3339                                return ret;
3340                        BUG_ON(ret > 0);
3341                        leaf = path->nodes[0];
3342                        slot = path->slots[0];
3343                        if (path_change)
3344                                *path_change = 1;
3345                }
3346                btrfs_item_key_to_cpu(leaf, &key, slot);
3347                if (key.objectid != extent_key->objectid)
3348                        return -ENOENT;
3349
3350                if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3351                        slot++;
3352                        continue;
3353                }
3354                ref0 = btrfs_item_ptr(leaf, slot,
3355                                struct btrfs_extent_ref_v0);
3356                *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3357                break;
3358        }
3359        return 0;
3360}
3361#endif
3362
3363/*
3364 * helper to add a tree block to the list.
3365 * the major work is getting the generation and level of the block
3366 */
3367static int add_tree_block(struct reloc_control *rc,
3368                          struct btrfs_key *extent_key,
3369                          struct btrfs_path *path,
3370                          struct rb_root *blocks)
3371{
3372        struct extent_buffer *eb;
3373        struct btrfs_extent_item *ei;
3374        struct btrfs_tree_block_info *bi;
3375        struct tree_block *block;
3376        struct rb_node *rb_node;
3377        u32 item_size;
3378        int level = -1;
3379        u64 generation;
3380
3381        eb =  path->nodes[0];
3382        item_size = btrfs_item_size_nr(eb, path->slots[0]);
3383
3384        if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3385            item_size >= sizeof(*ei) + sizeof(*bi)) {
3386                ei = btrfs_item_ptr(eb, path->slots[0],
3387                                struct btrfs_extent_item);
3388                if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3389                        bi = (struct btrfs_tree_block_info *)(ei + 1);
3390                        level = btrfs_tree_block_level(eb, bi);
3391                } else {
3392                        level = (int)extent_key->offset;
3393                }
3394                generation = btrfs_extent_generation(eb, ei);
3395        } else {
3396#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3397                u64 ref_owner;
3398                int ret;
3399
3400                BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3401                ret = get_ref_objectid_v0(rc, path, extent_key,
3402                                          &ref_owner, NULL);
3403                if (ret < 0)
3404                        return ret;
3405                BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3406                level = (int)ref_owner;
3407                /* FIXME: get real generation */
3408                generation = 0;
3409#else
3410                BUG();
3411#endif
3412        }
3413
3414        btrfs_release_path(path);
3415
3416        BUG_ON(level == -1);
3417
3418        block = kmalloc(sizeof(*block), GFP_NOFS);
3419        if (!block)
3420                return -ENOMEM;
3421
3422        block->bytenr = extent_key->objectid;
3423        block->key.objectid = rc->extent_root->fs_info->nodesize;
3424        block->key.offset = generation;
3425        block->level = level;
3426        block->key_ready = 0;
3427
3428        rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3429        if (rb_node)
3430                backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3431
3432        return 0;
3433}
3434
3435/*
3436 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3437 */
3438static int __add_tree_block(struct reloc_control *rc,
3439                            u64 bytenr, u32 blocksize,
3440                            struct rb_root *blocks)
3441{
3442        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3443        struct btrfs_path *path;
3444        struct btrfs_key key;
3445        int ret;
3446        bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3447
3448        if (tree_block_processed(bytenr, rc))
3449                return 0;
3450
3451        if (tree_search(blocks, bytenr))
3452                return 0;
3453
3454        path = btrfs_alloc_path();
3455        if (!path)
3456                return -ENOMEM;
3457again:
3458        key.objectid = bytenr;
3459        if (skinny) {
3460                key.type = BTRFS_METADATA_ITEM_KEY;
3461                key.offset = (u64)-1;
3462        } else {
3463                key.type = BTRFS_EXTENT_ITEM_KEY;
3464                key.offset = blocksize;
3465        }
3466
3467        path->search_commit_root = 1;
3468        path->skip_locking = 1;
3469        ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3470        if (ret < 0)
3471                goto out;
3472
3473        if (ret > 0 && skinny) {
3474                if (path->slots[0]) {
3475                        path->slots[0]--;
3476                        btrfs_item_key_to_cpu(path->nodes[0], &key,
3477                                              path->slots[0]);
3478                        if (key.objectid == bytenr &&
3479                            (key.type == BTRFS_METADATA_ITEM_KEY ||
3480                             (key.type == BTRFS_EXTENT_ITEM_KEY &&
3481                              key.offset == blocksize)))
3482                                ret = 0;
3483                }
3484
3485                if (ret) {
3486                        skinny = false;
3487                        btrfs_release_path(path);
3488                        goto again;
3489                }
3490        }
3491        if (ret) {
3492                ASSERT(ret == 1);
3493                btrfs_print_leaf(path->nodes[0]);
3494                btrfs_err(fs_info,
3495             "tree block extent item (%llu) is not found in extent tree",
3496                     bytenr);
3497                WARN_ON(1);
3498                ret = -EINVAL;
3499                goto out;
3500        }
3501
3502        ret = add_tree_block(rc, &key, path, blocks);
3503out:
3504        btrfs_free_path(path);
3505        return ret;
3506}
3507
3508/*
3509 * helper to check if the block use full backrefs for pointers in it
3510 */
3511static int block_use_full_backref(struct reloc_control *rc,
3512                                  struct extent_buffer *eb)
3513{
3514        u64 flags;
3515        int ret;
3516
3517        if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3518            btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3519                return 1;
3520
3521        ret = btrfs_lookup_extent_info(NULL, rc->extent_root->fs_info,
3522                                       eb->start, btrfs_header_level(eb), 1,
3523                                       NULL, &flags);
3524        BUG_ON(ret);
3525
3526        if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3527                ret = 1;
3528        else
3529                ret = 0;
3530        return ret;
3531}
3532
3533static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3534                                    struct btrfs_block_group_cache *block_group,
3535                                    struct inode *inode,
3536                                    u64 ino)
3537{
3538        struct btrfs_key key;
3539        struct btrfs_root *root = fs_info->tree_root;
3540        struct btrfs_trans_handle *trans;
3541        int ret = 0;
3542
3543        if (inode)
3544                goto truncate;
3545
3546        key.objectid = ino;
3547        key.type = BTRFS_INODE_ITEM_KEY;
3548        key.offset = 0;
3549
3550        inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3551        if (IS_ERR(inode) || is_bad_inode(inode)) {
3552                if (!IS_ERR(inode))
3553                        iput(inode);
3554                return -ENOENT;
3555        }
3556
3557truncate:
3558        ret = btrfs_check_trunc_cache_free_space(fs_info,
3559                                                 &fs_info->global_block_rsv);
3560        if (ret)
3561                goto out;
3562
3563        trans = btrfs_join_transaction(root);
3564        if (IS_ERR(trans)) {
3565                ret = PTR_ERR(trans);
3566                goto out;
3567        }
3568
3569        ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3570
3571        btrfs_end_transaction(trans);
3572        btrfs_btree_balance_dirty(fs_info);
3573out:
3574        iput(inode);
3575        return ret;
3576}
3577
3578/*
3579 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3580 * this function scans fs tree to find blocks reference the data extent
3581 */
3582static int find_data_references(struct reloc_control *rc,
3583                                struct btrfs_key *extent_key,
3584                                struct extent_buffer *leaf,
3585                                struct btrfs_extent_data_ref *ref,
3586                                struct rb_root *blocks)
3587{
3588        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3589        struct btrfs_path *path;
3590        struct tree_block *block;
3591        struct btrfs_root *root;
3592        struct btrfs_file_extent_item *fi;
3593        struct rb_node *rb_node;
3594        struct btrfs_key key;
3595        u64 ref_root;
3596        u64 ref_objectid;
3597        u64 ref_offset;
3598        u32 ref_count;
3599        u32 nritems;
3600        int err = 0;
3601        int added = 0;
3602        int counted;
3603        int ret;
3604
3605        ref_root = btrfs_extent_data_ref_root(leaf, ref);
3606        ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3607        ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3608        ref_count = btrfs_extent_data_ref_count(leaf, ref);
3609
3610        /*
3611         * This is an extent belonging to the free space cache, lets just delete
3612         * it and redo the search.
3613         */
3614        if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3615                ret = delete_block_group_cache(fs_info, rc->block_group,
3616                                               NULL, ref_objectid);
3617                if (ret != -ENOENT)
3618                        return ret;
3619                ret = 0;
3620        }
3621
3622        path = btrfs_alloc_path();
3623        if (!path)
3624                return -ENOMEM;
3625        path->reada = READA_FORWARD;
3626
3627        root = read_fs_root(fs_info, ref_root);
3628        if (IS_ERR(root)) {
3629                err = PTR_ERR(root);
3630                goto out;
3631        }
3632
3633        key.objectid = ref_objectid;
3634        key.type = BTRFS_EXTENT_DATA_KEY;
3635        if (ref_offset > ((u64)-1 << 32))
3636                key.offset = 0;
3637        else
3638                key.offset = ref_offset;
3639
3640        path->search_commit_root = 1;
3641        path->skip_locking = 1;
3642        ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3643        if (ret < 0) {
3644                err = ret;
3645                goto out;
3646        }
3647
3648        leaf = path->nodes[0];
3649        nritems = btrfs_header_nritems(leaf);
3650        /*
3651         * the references in tree blocks that use full backrefs
3652         * are not counted in
3653         */
3654        if (block_use_full_backref(rc, leaf))
3655                counted = 0;
3656        else
3657                counted = 1;
3658        rb_node = tree_search(blocks, leaf->start);
3659        if (rb_node) {
3660                if (counted)
3661                        added = 1;
3662                else
3663                        path->slots[0] = nritems;
3664        }
3665
3666        while (ref_count > 0) {
3667                while (path->slots[0] >= nritems) {
3668                        ret = btrfs_next_leaf(root, path);
3669                        if (ret < 0) {
3670                                err = ret;
3671                                goto out;
3672                        }
3673                        if (WARN_ON(ret > 0))
3674                                goto out;
3675
3676                        leaf = path->nodes[0];
3677                        nritems = btrfs_header_nritems(leaf);
3678                        added = 0;
3679
3680                        if (block_use_full_backref(rc, leaf))
3681                                counted = 0;
3682                        else
3683                                counted = 1;
3684                        rb_node = tree_search(blocks, leaf->start);
3685                        if (rb_node) {
3686                                if (counted)
3687                                        added = 1;
3688                                else
3689                                        path->slots[0] = nritems;
3690                        }
3691                }
3692
3693                btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3694                if (WARN_ON(key.objectid != ref_objectid ||
3695                    key.type != BTRFS_EXTENT_DATA_KEY))
3696                        break;
3697
3698                fi = btrfs_item_ptr(leaf, path->slots[0],
3699                                    struct btrfs_file_extent_item);
3700
3701                if (btrfs_file_extent_type(leaf, fi) ==
3702                    BTRFS_FILE_EXTENT_INLINE)
3703                        goto next;
3704
3705                if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3706                    extent_key->objectid)
3707                        goto next;
3708
3709                key.offset -= btrfs_file_extent_offset(leaf, fi);
3710                if (key.offset != ref_offset)
3711                        goto next;
3712
3713                if (counted)
3714                        ref_count--;
3715                if (added)
3716                        goto next;
3717
3718                if (!tree_block_processed(leaf->start, rc)) {
3719                        block = kmalloc(sizeof(*block), GFP_NOFS);
3720                        if (!block) {
3721                                err = -ENOMEM;
3722                                break;
3723                        }
3724                        block->bytenr = leaf->start;
3725                        btrfs_item_key_to_cpu(leaf, &block->key, 0);
3726                        block->level = 0;
3727                        block->key_ready = 1;
3728                        rb_node = tree_insert(blocks, block->bytenr,
3729                                              &block->rb_node);
3730                        if (rb_node)
3731                                backref_tree_panic(rb_node, -EEXIST,
3732                                                   block->bytenr);
3733                }
3734                if (counted)
3735                        added = 1;
3736                else
3737                        path->slots[0] = nritems;
3738next:
3739                path->slots[0]++;
3740
3741        }
3742out:
3743        btrfs_free_path(path);
3744        return err;
3745}
3746
3747/*
3748 * helper to find all tree blocks that reference a given data extent
3749 */
3750static noinline_for_stack
3751int add_data_references(struct reloc_control *rc,
3752                        struct btrfs_key *extent_key,
3753                        struct btrfs_path *path,
3754                        struct rb_root *blocks)
3755{
3756        struct btrfs_key key;
3757        struct extent_buffer *eb;
3758        struct btrfs_extent_data_ref *dref;
3759        struct btrfs_extent_inline_ref *iref;
3760        unsigned long ptr;
3761        unsigned long end;
3762        u32 blocksize = rc->extent_root->fs_info->nodesize;
3763        int ret = 0;
3764        int err = 0;
3765
3766        eb = path->nodes[0];
3767        ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3768        end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3769#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3770        if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3771                ptr = end;
3772        else
3773#endif
3774                ptr += sizeof(struct btrfs_extent_item);
3775
3776        while (ptr < end) {
3777                iref = (struct btrfs_extent_inline_ref *)ptr;
3778                key.type = btrfs_get_extent_inline_ref_type(eb, iref,
3779                                                        BTRFS_REF_TYPE_DATA);
3780                if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3781                        key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3782                        ret = __add_tree_block(rc, key.offset, blocksize,
3783                                               blocks);
3784                } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3785                        dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3786                        ret = find_data_references(rc, extent_key,
3787                                                   eb, dref, blocks);
3788                } else {
3789                        ret = -EINVAL;
3790                        btrfs_err(rc->extent_root->fs_info,
3791                     "extent %llu slot %d has an invalid inline ref type",
3792                             eb->start, path->slots[0]);
3793                }
3794                if (ret) {
3795                        err = ret;
3796                        goto out;
3797                }
3798                ptr += btrfs_extent_inline_ref_size(key.type);
3799        }
3800        WARN_ON(ptr > end);
3801
3802        while (1) {
3803                cond_resched();
3804                eb = path->nodes[0];
3805                if (path->slots[0] >= btrfs_header_nritems(eb)) {
3806                        ret = btrfs_next_leaf(rc->extent_root, path);
3807                        if (ret < 0) {
3808                                err = ret;
3809                                break;
3810                        }
3811                        if (ret > 0)
3812                                break;
3813                        eb = path->nodes[0];
3814                }
3815
3816                btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3817                if (key.objectid != extent_key->objectid)
3818                        break;
3819
3820#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3821                if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3822                    key.type == BTRFS_EXTENT_REF_V0_KEY) {
3823#else
3824                BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3825                if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3826#endif
3827                        ret = __add_tree_block(rc, key.offset, blocksize,
3828                                               blocks);
3829                } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3830                        dref = btrfs_item_ptr(eb, path->slots[0],
3831                                              struct btrfs_extent_data_ref);
3832                        ret = find_data_references(rc, extent_key,
3833                                                   eb, dref, blocks);
3834                } else {
3835                        ret = 0;
3836                }
3837                if (ret) {
3838                        err = ret;
3839                        break;
3840                }
3841                path->slots[0]++;
3842        }
3843out:
3844        btrfs_release_path(path);
3845        if (err)
3846                free_block_list(blocks);
3847        return err;
3848}
3849
3850/*
3851 * helper to find next unprocessed extent
3852 */
3853static noinline_for_stack
3854int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3855                     struct btrfs_key *extent_key)
3856{
3857        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3858        struct btrfs_key key;
3859        struct extent_buffer *leaf;
3860        u64 start, end, last;
3861        int ret;
3862
3863        last = rc->block_group->key.objectid + rc->block_group->key.offset;
3864        while (1) {
3865                cond_resched();
3866                if (rc->search_start >= last) {
3867                        ret = 1;
3868                        break;
3869                }
3870
3871                key.objectid = rc->search_start;
3872                key.type = BTRFS_EXTENT_ITEM_KEY;
3873                key.offset = 0;
3874
3875                path->search_commit_root = 1;
3876                path->skip_locking = 1;
3877                ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3878                                        0, 0);
3879                if (ret < 0)
3880                        break;
3881next:
3882                leaf = path->nodes[0];
3883                if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3884                        ret = btrfs_next_leaf(rc->extent_root, path);
3885                        if (ret != 0)
3886                                break;
3887                        leaf = path->nodes[0];
3888                }
3889
3890                btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3891                if (key.objectid >= last) {
3892                        ret = 1;
3893                        break;
3894                }
3895
3896                if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3897                    key.type != BTRFS_METADATA_ITEM_KEY) {
3898                        path->slots[0]++;
3899                        goto next;
3900                }
3901
3902                if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3903                    key.objectid + key.offset <= rc->search_start) {
3904                        path->slots[0]++;
3905                        goto next;
3906                }
3907
3908                if (key.type == BTRFS_METADATA_ITEM_KEY &&
3909                    key.objectid + fs_info->nodesize <=
3910                    rc->search_start) {
3911                        path->slots[0]++;
3912                        goto next;
3913                }
3914
3915                ret = find_first_extent_bit(&rc->processed_blocks,
3916                                            key.objectid, &start, &end,
3917                                            EXTENT_DIRTY, NULL);
3918
3919                if (ret == 0 && start <= key.objectid) {
3920                        btrfs_release_path(path);
3921                        rc->search_start = end + 1;
3922                } else {
3923                        if (key.type == BTRFS_EXTENT_ITEM_KEY)
3924                                rc->search_start = key.objectid + key.offset;
3925                        else
3926                                rc->search_start = key.objectid +
3927                                        fs_info->nodesize;
3928                        memcpy(extent_key, &key, sizeof(key));
3929                        return 0;
3930                }
3931        }
3932        btrfs_release_path(path);
3933        return ret;
3934}
3935
3936static void set_reloc_control(struct reloc_control *rc)
3937{
3938        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3939
3940        mutex_lock(&fs_info->reloc_mutex);
3941        fs_info->reloc_ctl = rc;
3942        mutex_unlock(&fs_info->reloc_mutex);
3943}
3944
3945static void unset_reloc_control(struct reloc_control *rc)
3946{
3947        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3948
3949        mutex_lock(&fs_info->reloc_mutex);
3950        fs_info->reloc_ctl = NULL;
3951        mutex_unlock(&fs_info->reloc_mutex);
3952}
3953
3954static int check_extent_flags(u64 flags)
3955{
3956        if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3957            (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3958                return 1;
3959        if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3960            !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3961                return 1;
3962        if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3963            (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3964                return 1;
3965        return 0;
3966}
3967
3968static noinline_for_stack
3969int prepare_to_relocate(struct reloc_control *rc)
3970{
3971        struct btrfs_trans_handle *trans;
3972        int ret;
3973
3974        rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3975                                              BTRFS_BLOCK_RSV_TEMP);
3976        if (!rc->block_rsv)
3977                return -ENOMEM;
3978
3979        memset(&rc->cluster, 0, sizeof(rc->cluster));
3980        rc->search_start = rc->block_group->key.objectid;
3981        rc->extents_found = 0;
3982        rc->nodes_relocated = 0;
3983        rc->merging_rsv_size = 0;
3984        rc->reserved_bytes = 0;
3985        rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3986                              RELOCATION_RESERVED_NODES;
3987        ret = btrfs_block_rsv_refill(rc->extent_root,
3988                                     rc->block_rsv, rc->block_rsv->size,
3989                                     BTRFS_RESERVE_FLUSH_ALL);
3990        if (ret)
3991                return ret;
3992
3993        rc->create_reloc_tree = 1;
3994        set_reloc_control(rc);
3995
3996        trans = btrfs_join_transaction(rc->extent_root);
3997        if (IS_ERR(trans)) {
3998                unset_reloc_control(rc);
3999                /*
4000                 * extent tree is not a ref_cow tree and has no reloc_root to
4001                 * cleanup.  And callers are responsible to free the above
4002                 * block rsv.
4003                 */
4004                return PTR_ERR(trans);
4005        }
4006        btrfs_commit_transaction(trans);
4007        return 0;
4008}
4009
4010static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
4011{
4012        struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
4013        struct rb_root blocks = RB_ROOT;
4014        struct btrfs_key key;
4015        struct btrfs_trans_handle *trans = NULL;
4016        struct btrfs_path *path;
4017        struct btrfs_extent_item *ei;
4018        u64 flags;
4019        u32 item_size;
4020        int ret;
4021        int err = 0;
4022        int progress = 0;
4023
4024        path = btrfs_alloc_path();
4025        if (!path)
4026                return -ENOMEM;
4027        path->reada = READA_FORWARD;
4028
4029        ret = prepare_to_relocate(rc);
4030        if (ret) {
4031                err = ret;
4032                goto out_free;
4033        }
4034
4035        while (1) {
4036                rc->reserved_bytes = 0;
4037                ret = btrfs_block_rsv_refill(rc->extent_root,
4038                                        rc->block_rsv, rc->block_rsv->size,
4039                                        BTRFS_RESERVE_FLUSH_ALL);
4040                if (ret) {
4041                        err = ret;
4042                        break;
4043                }
4044                progress++;
4045                trans = btrfs_start_transaction(rc->extent_root, 0);
4046                if (IS_ERR(trans)) {
4047                        err = PTR_ERR(trans);
4048                        trans = NULL;
4049                        break;
4050                }
4051restart:
4052                if (update_backref_cache(trans, &rc->backref_cache)) {
4053                        btrfs_end_transaction(trans);
4054                        continue;
4055                }
4056
4057                ret = find_next_extent(rc, path, &key);
4058                if (ret < 0)
4059                        err = ret;
4060                if (ret != 0)
4061                        break;
4062
4063                rc->extents_found++;
4064
4065                ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4066                                    struct btrfs_extent_item);
4067                item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4068                if (item_size >= sizeof(*ei)) {
4069                        flags = btrfs_extent_flags(path->nodes[0], ei);
4070                        ret = check_extent_flags(flags);
4071                        BUG_ON(ret);
4072
4073                } else {
4074#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4075                        u64 ref_owner;
4076                        int path_change = 0;
4077
4078                        BUG_ON(item_size !=
4079                               sizeof(struct btrfs_extent_item_v0));
4080                        ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
4081                                                  &path_change);
4082                        if (ret < 0) {
4083                                err = ret;
4084                                break;
4085                        }
4086                        if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
4087                                flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
4088                        else
4089                                flags = BTRFS_EXTENT_FLAG_DATA;
4090
4091                        if (path_change) {
4092                                btrfs_release_path(path);
4093
4094                                path->search_commit_root = 1;
4095                                path->skip_locking = 1;
4096                                ret = btrfs_search_slot(NULL, rc->extent_root,
4097                                                        &key, path, 0, 0);
4098                                if (ret < 0) {
4099                                        err = ret;
4100                                        break;
4101                                }
4102                                BUG_ON(ret > 0);
4103                        }
4104#else
4105                        BUG();
4106#endif
4107                }
4108
4109                if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4110                        ret = add_tree_block(rc, &key, path, &blocks);
4111                } else if (rc->stage == UPDATE_DATA_PTRS &&
4112                           (flags & BTRFS_EXTENT_FLAG_DATA)) {
4113                        ret = add_data_references(rc, &key, path, &blocks);
4114                } else {
4115                        btrfs_release_path(path);
4116                        ret = 0;
4117                }
4118                if (ret < 0) {
4119                        err = ret;
4120                        break;
4121                }
4122
4123                if (!RB_EMPTY_ROOT(&blocks)) {
4124                        ret = relocate_tree_blocks(trans, rc, &blocks);
4125                        if (ret < 0) {
4126                                /*
4127                                 * if we fail to relocate tree blocks, force to update
4128                                 * backref cache when committing transaction.
4129                                 */
4130                                rc->backref_cache.last_trans = trans->transid - 1;
4131
4132                                if (ret != -EAGAIN) {
4133                                        err = ret;
4134                                        break;
4135                                }
4136                                rc->extents_found--;
4137                                rc->search_start = key.objectid;
4138                        }
4139                }
4140
4141                btrfs_end_transaction_throttle(trans);
4142                btrfs_btree_balance_dirty(fs_info);
4143                trans = NULL;
4144
4145                if (rc->stage == MOVE_DATA_EXTENTS &&
4146                    (flags & BTRFS_EXTENT_FLAG_DATA)) {
4147                        rc->found_file_extent = 1;
4148                        ret = relocate_data_extent(rc->data_inode,
4149                                                   &key, &rc->cluster);
4150                        if (ret < 0) {
4151                                err = ret;
4152                                break;
4153                        }
4154                }
4155        }
4156        if (trans && progress && err == -ENOSPC) {
4157                ret = btrfs_force_chunk_alloc(trans, fs_info,
4158                                              rc->block_group->flags);
4159                if (ret == 1) {
4160                        err = 0;
4161                        progress = 0;
4162                        goto restart;
4163                }
4164        }
4165
4166        btrfs_release_path(path);
4167        clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4168
4169        if (trans) {
4170                btrfs_end_transaction_throttle(trans);
4171                btrfs_btree_balance_dirty(fs_info);
4172        }
4173
4174        if (!err) {
4175                ret = relocate_file_extent_cluster(rc->data_inode,
4176                                                   &rc->cluster);
4177                if (ret < 0)
4178                        err = ret;
4179        }
4180
4181        rc->create_reloc_tree = 0;
4182        set_reloc_control(rc);
4183
4184        backref_cache_cleanup(&rc->backref_cache);
4185        btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4186
4187        err = prepare_to_merge(rc, err);
4188
4189        merge_reloc_roots(rc);
4190
4191        rc->merge_reloc_tree = 0;
4192        unset_reloc_control(rc);
4193        btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1);
4194
4195        /* get rid of pinned extents */
4196        trans = btrfs_join_transaction(rc->extent_root);
4197        if (IS_ERR(trans)) {
4198                err = PTR_ERR(trans);
4199                goto out_free;
4200        }
4201        btrfs_commit_transaction(trans);
4202out_free:
4203        btrfs_free_block_rsv(fs_info, rc->block_rsv);
4204        btrfs_free_path(path);
4205        return err;
4206}
4207
4208static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4209                                 struct btrfs_root *root, u64 objectid)
4210{
4211        struct btrfs_path *path;
4212        struct btrfs_inode_item *item;
4213        struct extent_buffer *leaf;
4214        int ret;
4215
4216        path = btrfs_alloc_path();
4217        if (!path)
4218                return -ENOMEM;
4219
4220        ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4221        if (ret)
4222                goto out;
4223
4224        leaf = path->nodes[0];
4225        item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4226        memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
4227        btrfs_set_inode_generation(leaf, item, 1);
4228        btrfs_set_inode_size(leaf, item, 0);
4229        btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4230        btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4231                                          BTRFS_INODE_PREALLOC);
4232        btrfs_mark_buffer_dirty(leaf);
4233out:
4234        btrfs_free_path(path);
4235        return ret;
4236}
4237
4238/*
4239 * helper to create inode for data relocation.
4240 * the inode is in data relocation tree and its link count is 0
4241 */
4242static noinline_for_stack
4243struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4244                                 struct btrfs_block_group_cache *group)
4245{
4246        struct inode *inode = NULL;
4247        struct btrfs_trans_handle *trans;
4248        struct btrfs_root *root;
4249        struct btrfs_key key;
4250        u64 objectid;
4251        int err = 0;
4252
4253        root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4254        if (IS_ERR(root))
4255                return ERR_CAST(root);
4256
4257        trans = btrfs_start_transaction(root, 6);
4258        if (IS_ERR(trans))
4259                return ERR_CAST(trans);
4260
4261        err = btrfs_find_free_objectid(root, &objectid);
4262        if (err)
4263                goto out;
4264
4265        err = __insert_orphan_inode(trans, root, objectid);
4266        BUG_ON(err);
4267
4268        key.objectid = objectid;
4269        key.type = BTRFS_INODE_ITEM_KEY;
4270        key.offset = 0;
4271        inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4272        BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4273        BTRFS_I(inode)->index_cnt = group->key.objectid;
4274
4275        err = btrfs_orphan_add(trans, BTRFS_I(inode));
4276out:
4277        btrfs_end_transaction(trans);
4278        btrfs_btree_balance_dirty(fs_info);
4279        if (err) {
4280                if (inode)
4281                        iput(inode);
4282                inode = ERR_PTR(err);
4283        }
4284        return inode;
4285}
4286
4287static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4288{
4289        struct reloc_control *rc;
4290
4291        rc = kzalloc(sizeof(*rc), GFP_NOFS);
4292        if (!rc)
4293                return NULL;
4294
4295        INIT_LIST_HEAD(&rc->reloc_roots);
4296        backref_cache_init(&rc->backref_cache);
4297        mapping_tree_init(&rc->reloc_root_tree);
4298        extent_io_tree_init(&rc->processed_blocks, NULL);
4299        return rc;
4300}
4301
4302/*
4303 * Print the block group being relocated
4304 */
4305static void describe_relocation(struct btrfs_fs_info *fs_info,
4306                                struct btrfs_block_group_cache *block_group)
4307{
4308        char buf[128];          /* prefixed by a '|' that'll be dropped */
4309        u64 flags = block_group->flags;
4310
4311        /* Shouldn't happen */
4312        if (!flags) {
4313                strcpy(buf, "|NONE");
4314        } else {
4315                char *bp = buf;
4316
4317#define DESCRIBE_FLAG(f, d) \
4318                if (flags & BTRFS_BLOCK_GROUP_##f) { \
4319                        bp += snprintf(bp, buf - bp + sizeof(buf), "|%s", d); \
4320                        flags &= ~BTRFS_BLOCK_GROUP_##f; \
4321                }
4322                DESCRIBE_FLAG(DATA,     "data");
4323                DESCRIBE_FLAG(SYSTEM,   "system");
4324                DESCRIBE_FLAG(METADATA, "metadata");
4325                DESCRIBE_FLAG(RAID0,    "raid0");
4326                DESCRIBE_FLAG(RAID1,    "raid1");
4327                DESCRIBE_FLAG(DUP,      "dup");
4328                DESCRIBE_FLAG(RAID10,   "raid10");
4329                DESCRIBE_FLAG(RAID5,    "raid5");
4330                DESCRIBE_FLAG(RAID6,    "raid6");
4331                if (flags)
4332                        snprintf(buf, buf - bp + sizeof(buf), "|0x%llx", flags);
4333#undef DESCRIBE_FLAG
4334        }
4335
4336        btrfs_info(fs_info,
4337                   "relocating block group %llu flags %s",
4338                   block_group->key.objectid, buf + 1);
4339}
4340
4341/*
4342 * function to relocate all extents in a block group.
4343 */
4344int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4345{
4346        struct btrfs_root *extent_root = fs_info->extent_root;
4347        struct reloc_control *rc;
4348        struct inode *inode;
4349        struct btrfs_path *path;
4350        int ret;
4351        int rw = 0;
4352        int err = 0;
4353
4354        rc = alloc_reloc_control(fs_info);
4355        if (!rc)
4356                return -ENOMEM;
4357
4358        rc->extent_root = extent_root;
4359
4360        rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4361        BUG_ON(!rc->block_group);
4362
4363        ret = btrfs_inc_block_group_ro(fs_info, rc->block_group);
4364        if (ret) {
4365                err = ret;
4366                goto out;
4367        }
4368        rw = 1;
4369
4370        path = btrfs_alloc_path();
4371        if (!path) {
4372                err = -ENOMEM;
4373                goto out;
4374        }
4375
4376        inode = lookup_free_space_inode(fs_info, rc->block_group, path);
4377        btrfs_free_path(path);
4378
4379        if (!IS_ERR(inode))
4380                ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4381        else
4382                ret = PTR_ERR(inode);
4383
4384        if (ret && ret != -ENOENT) {
4385                err = ret;
4386                goto out;
4387        }
4388
4389        rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4390        if (IS_ERR(rc->data_inode)) {
4391                err = PTR_ERR(rc->data_inode);
4392                rc->data_inode = NULL;
4393                goto out;
4394        }
4395
4396        describe_relocation(fs_info, rc->block_group);
4397
4398        btrfs_wait_block_group_reservations(rc->block_group);
4399        btrfs_wait_nocow_writers(rc->block_group);
4400        btrfs_wait_ordered_roots(fs_info, U64_MAX,
4401                                 rc->block_group->key.objectid,
4402                                 rc->block_group->key.offset);
4403
4404        while (1) {
4405                mutex_lock(&fs_info->cleaner_mutex);
4406                ret = relocate_block_group(rc);
4407                mutex_unlock(&fs_info->cleaner_mutex);
4408                if (ret < 0) {
4409                        err = ret;
4410                        goto out;
4411                }
4412
4413                if (rc->extents_found == 0)
4414                        break;
4415
4416                btrfs_info(fs_info, "found %llu extents", rc->extents_found);
4417
4418                if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4419                        ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4420                                                       (u64)-1);
4421                        if (ret) {
4422                                err = ret;
4423                                goto out;
4424                        }
4425                        invalidate_mapping_pages(rc->data_inode->i_mapping,
4426                                                 0, -1);
4427                        rc->stage = UPDATE_DATA_PTRS;
4428                }
4429        }
4430
4431        WARN_ON(rc->block_group->pinned > 0);
4432        WARN_ON(rc->block_group->reserved > 0);
4433        WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4434out:
4435        if (err && rw)
4436                btrfs_dec_block_group_ro(rc->block_group);
4437        iput(rc->data_inode);
4438        btrfs_put_block_group(rc->block_group);
4439        kfree(rc);
4440        return err;
4441}
4442
4443static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4444{
4445        struct btrfs_fs_info *fs_info = root->fs_info;
4446        struct btrfs_trans_handle *trans;
4447        int ret, err;
4448
4449        trans = btrfs_start_transaction(fs_info->tree_root, 0);
4450        if (IS_ERR(trans))
4451                return PTR_ERR(trans);
4452
4453        memset(&root->root_item.drop_progress, 0,
4454                sizeof(root->root_item.drop_progress));
4455        root->root_item.drop_level = 0;
4456        btrfs_set_root_refs(&root->root_item, 0);
4457        ret = btrfs_update_root(trans, fs_info->tree_root,
4458                                &root->root_key, &root->root_item);
4459
4460        err = btrfs_end_transaction(trans);
4461        if (err)
4462                return err;
4463        return ret;
4464}
4465
4466/*
4467 * recover relocation interrupted by system crash.
4468 *
4469 * this function resumes merging reloc trees with corresponding fs trees.
4470 * this is important for keeping the sharing of tree blocks
4471 */
4472int btrfs_recover_relocation(struct btrfs_root *root)
4473{
4474        struct btrfs_fs_info *fs_info = root->fs_info;
4475        LIST_HEAD(reloc_roots);
4476        struct btrfs_key key;
4477        struct btrfs_root *fs_root;
4478        struct btrfs_root *reloc_root;
4479        struct btrfs_path *path;
4480        struct extent_buffer *leaf;
4481        struct reloc_control *rc = NULL;
4482        struct btrfs_trans_handle *trans;
4483        int ret;
4484        int err = 0;
4485
4486        path = btrfs_alloc_path();
4487        if (!path)
4488                return -ENOMEM;
4489        path->reada = READA_BACK;
4490
4491        key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4492        key.type = BTRFS_ROOT_ITEM_KEY;
4493        key.offset = (u64)-1;
4494
4495        while (1) {
4496                ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4497                                        path, 0, 0);
4498                if (ret < 0) {
4499                        err = ret;
4500                        goto out;
4501                }
4502                if (ret > 0) {
4503                        if (path->slots[0] == 0)
4504                                break;
4505                        path->slots[0]--;
4506                }
4507                leaf = path->nodes[0];
4508                btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4509                btrfs_release_path(path);
4510
4511                if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4512                    key.type != BTRFS_ROOT_ITEM_KEY)
4513                        break;
4514
4515                reloc_root = btrfs_read_fs_root(root, &key);
4516                if (IS_ERR(reloc_root)) {
4517                        err = PTR_ERR(reloc_root);
4518                        goto out;
4519                }
4520
4521                list_add(&reloc_root->root_list, &reloc_roots);
4522
4523                if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4524                        fs_root = read_fs_root(fs_info,
4525                                               reloc_root->root_key.offset);
4526                        if (IS_ERR(fs_root)) {
4527                                ret = PTR_ERR(fs_root);
4528                                if (ret != -ENOENT) {
4529                                        err = ret;
4530                                        goto out;
4531                                }
4532                                ret = mark_garbage_root(reloc_root);
4533                                if (ret < 0) {
4534                                        err = ret;
4535                                        goto out;
4536                                }
4537                        }
4538                }
4539
4540                if (key.offset == 0)
4541                        break;
4542
4543                key.offset--;
4544        }
4545        btrfs_release_path(path);
4546
4547        if (list_empty(&reloc_roots))
4548                goto out;
4549
4550        rc = alloc_reloc_control(fs_info);
4551        if (!rc) {
4552                err = -ENOMEM;
4553                goto out;
4554        }
4555
4556        rc->extent_root = fs_info->extent_root;
4557
4558        set_reloc_control(rc);
4559
4560        trans = btrfs_join_transaction(rc->extent_root);
4561        if (IS_ERR(trans)) {
4562                unset_reloc_control(rc);
4563                err = PTR_ERR(trans);
4564                goto out_free;
4565        }
4566
4567        rc->merge_reloc_tree = 1;
4568
4569        while (!list_empty(&reloc_roots)) {
4570                reloc_root = list_entry(reloc_roots.next,
4571                                        struct btrfs_root, root_list);
4572                list_del(&reloc_root->root_list);
4573
4574                if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4575                        list_add_tail(&reloc_root->root_list,
4576                                      &rc->reloc_roots);
4577                        continue;
4578                }
4579
4580                fs_root = read_fs_root(fs_info, reloc_root->root_key.offset);
4581                if (IS_ERR(fs_root)) {
4582                        err = PTR_ERR(fs_root);
4583                        goto out_free;
4584                }
4585
4586                err = __add_reloc_root(reloc_root);
4587                BUG_ON(err < 0); /* -ENOMEM or logic error */
4588                fs_root->reloc_root = reloc_root;
4589        }
4590
4591        err = btrfs_commit_transaction(trans);
4592        if (err)
4593                goto out_free;
4594
4595        merge_reloc_roots(rc);
4596
4597        unset_reloc_control(rc);
4598
4599        trans = btrfs_join_transaction(rc->extent_root);
4600        if (IS_ERR(trans)) {
4601                err = PTR_ERR(trans);
4602                goto out_free;
4603        }
4604        err = btrfs_commit_transaction(trans);
4605out_free:
4606        kfree(rc);
4607out:
4608        if (!list_empty(&reloc_roots))
4609                free_reloc_roots(&reloc_roots);
4610
4611        btrfs_free_path(path);
4612
4613        if (err == 0) {
4614                /* cleanup orphan inode in data relocation tree */
4615                fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4616                if (IS_ERR(fs_root))
4617                        err = PTR_ERR(fs_root);
4618                else
4619                        err = btrfs_orphan_cleanup(fs_root);
4620        }
4621        return err;
4622}
4623
4624/*
4625 * helper to add ordered checksum for data relocation.
4626 *
4627 * cloning checksum properly handles the nodatasum extents.
4628 * it also saves CPU time to re-calculate the checksum.
4629 */
4630int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4631{
4632        struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4633        struct btrfs_ordered_sum *sums;
4634        struct btrfs_ordered_extent *ordered;
4635        int ret;
4636        u64 disk_bytenr;
4637        u64 new_bytenr;
4638        LIST_HEAD(list);
4639
4640        ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4641        BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4642
4643        disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4644        ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4645                                       disk_bytenr + len - 1, &list, 0);
4646        if (ret)
4647                goto out;
4648
4649        while (!list_empty(&list)) {
4650                sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4651                list_del_init(&sums->list);
4652
4653                /*
4654                 * We need to offset the new_bytenr based on where the csum is.
4655                 * We need to do this because we will read in entire prealloc
4656                 * extents but we may have written to say the middle of the
4657                 * prealloc extent, so we need to make sure the csum goes with
4658                 * the right disk offset.
4659                 *
4660                 * We can do this because the data reloc inode refers strictly
4661                 * to the on disk bytes, so we don't have to worry about
4662                 * disk_len vs real len like with real inodes since it's all
4663                 * disk length.
4664                 */
4665                new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4666                sums->bytenr = new_bytenr;
4667
4668                btrfs_add_ordered_sum(inode, ordered, sums);
4669        }
4670out:
4671        btrfs_put_ordered_extent(ordered);
4672        return ret;
4673}
4674
4675int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4676                          struct btrfs_root *root, struct extent_buffer *buf,
4677                          struct extent_buffer *cow)
4678{
4679        struct btrfs_fs_info *fs_info = root->fs_info;
4680        struct reloc_control *rc;
4681        struct backref_node *node;
4682        int first_cow = 0;
4683        int level;
4684        int ret = 0;
4685
4686        rc = fs_info->reloc_ctl;
4687        if (!rc)
4688                return 0;
4689
4690        BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4691               root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4692
4693        if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4694                if (buf == root->node)
4695                        __update_reloc_root(root, cow->start);
4696        }
4697
4698        level = btrfs_header_level(buf);
4699        if (btrfs_header_generation(buf) <=
4700            btrfs_root_last_snapshot(&root->root_item))
4701                first_cow = 1;
4702
4703        if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4704            rc->create_reloc_tree) {
4705                WARN_ON(!first_cow && level == 0);
4706
4707                node = rc->backref_cache.path[level];
4708                BUG_ON(node->bytenr != buf->start &&
4709                       node->new_bytenr != buf->start);
4710
4711                drop_node_buffer(node);
4712                extent_buffer_get(cow);
4713                node->eb = cow;
4714                node->new_bytenr = cow->start;
4715
4716                if (!node->pending) {
4717                        list_move_tail(&node->list,
4718                                       &rc->backref_cache.pending[level]);
4719                        node->pending = 1;
4720                }
4721
4722                if (first_cow)
4723                        __mark_block_processed(rc, node);
4724
4725                if (first_cow && level > 0)
4726                        rc->nodes_relocated += buf->len;
4727        }
4728
4729        if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4730                ret = replace_file_extents(trans, rc, root, cow);
4731        return ret;
4732}
4733
4734/*
4735 * called before creating snapshot. it calculates metadata reservation
4736 * required for relocating tree blocks in the snapshot
4737 */
4738void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4739                              u64 *bytes_to_reserve)
4740{
4741        struct btrfs_root *root;
4742        struct reloc_control *rc;
4743
4744        root = pending->root;
4745        if (!root->reloc_root)
4746                return;
4747
4748        rc = root->fs_info->reloc_ctl;
4749        if (!rc->merge_reloc_tree)
4750                return;
4751
4752        root = root->reloc_root;
4753        BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4754        /*
4755         * relocation is in the stage of merging trees. the space
4756         * used by merging a reloc tree is twice the size of
4757         * relocated tree nodes in the worst case. half for cowing
4758         * the reloc tree, half for cowing the fs tree. the space
4759         * used by cowing the reloc tree will be freed after the
4760         * tree is dropped. if we create snapshot, cowing the fs
4761         * tree may use more space than it frees. so we need
4762         * reserve extra space.
4763         */
4764        *bytes_to_reserve += rc->nodes_relocated;
4765}
4766
4767/*
4768 * called after snapshot is created. migrate block reservation
4769 * and create reloc root for the newly created snapshot
4770 */
4771int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4772                               struct btrfs_pending_snapshot *pending)
4773{
4774        struct btrfs_root *root = pending->root;
4775        struct btrfs_root *reloc_root;
4776        struct btrfs_root *new_root;
4777        struct reloc_control *rc;
4778        int ret;
4779
4780        if (!root->reloc_root)
4781                return 0;
4782
4783        rc = root->fs_info->reloc_ctl;
4784        rc->merging_rsv_size += rc->nodes_relocated;
4785
4786        if (rc->merge_reloc_tree) {
4787                ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4788                                              rc->block_rsv,
4789                                              rc->nodes_relocated, 1);
4790                if (ret)
4791                        return ret;
4792        }
4793
4794        new_root = pending->snap;
4795        reloc_root = create_reloc_root(trans, root->reloc_root,
4796                                       new_root->root_key.objectid);
4797        if (IS_ERR(reloc_root))
4798                return PTR_ERR(reloc_root);
4799
4800        ret = __add_reloc_root(reloc_root);
4801        BUG_ON(ret < 0);
4802        new_root->reloc_root = reloc_root;
4803
4804        if (rc->create_reloc_tree)
4805                ret = clone_backref_node(trans, rc, root, reloc_root);
4806        return ret;
4807}
4808