fs/btrfs/relocation.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2009 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include <linux/error-injection.h>
  13 #include "ctree.h"
  14 #include "disk-io.h"
  15 #include "transaction.h"
  16 #include "volumes.h"
  17 #include "locking.h"
  18 #include "btrfs_inode.h"
  19 #include "async-thread.h"
  20 #include "free-space-cache.h"
  21 #include "qgroup.h"
  22 #include "print-tree.h"
  23 #include "delalloc-space.h"
  24 #include "block-group.h"
  25 #include "backref.h"
  26 #include "misc.h"
  27 #include "subpage.h"
  28 #include "zoned.h"
  29 #include "inode-item.h"
  30
  31 /*
  32  * Relocation overview
  33  *
  34  * [What does relocation do]
  35  *
  36  * The objective of relocation is to relocate all extents of the target block
  37  * group to other block groups.
  38  * This is utilized by resize (shrink only), profile converting, compacting
  39  * space, or balance routine to spread chunks over devices.
  40  *
  41  *              Before          |               After
  42  * ------------------------------------------------------------------
  43  *  BG A: 10 data extents       | BG A: deleted
  44  *  BG B:  2 data extents       | BG B: 10 data extents (2 old + 8 relocated)
  45  *  BG C:  1 extents            | BG C:  3 data extents (1 old + 2 relocated)
  46  *
  47  * [How does relocation work]
  48  *
  49  * 1.   Mark the target block group read-only
  50  *      New extents won't be allocated from the target block group.
  51  *
  52  * 2.1  Record each extent in the target block group
  53  *      To build a proper map of extents to be relocated.
  54  *
  55  * 2.2  Build data reloc tree and reloc trees
  56  *      Data reloc tree will contain an inode, recording all newly relocated
  57  *      data extents.
  58  *      There will be only one data reloc tree for one data block group.
  59  *
  60  *      Reloc tree will be a special snapshot of its source tree, containing
  61  *      relocated tree blocks.
  62  *      Each tree referring to a tree block in target block group will get its
  63  *      reloc tree built.
  64  *
  65  * 2.3  Swap source tree with its corresponding reloc tree
  66  *      Each involved tree only refers to new extents after swap.
  67  *
  68  * 3.   Cleanup reloc trees and data reloc tree.
  69  *      As old extents in the target block group are still referenced by reloc
  70  *      trees, we need to clean them up before really freeing the target block
  71  *      group.
  72  *
  73  * The main complexity is in steps 2.2 and 2.3.
  74  *
  75  * The entry point of relocation is relocate_block_group() function.
  76  */
  77
  78 #define RELOCATION_RESERVED_NODES       256
  79 /*
  80  * map address of tree root to tree
  81  */
  82 struct mapping_node {
  83         struct {
  84                 struct rb_node rb_node;
  85                 u64 bytenr;
  86         }; /* Use rb_simle_node for search/insert */
  87         void *data;
  88 };
  89
  90 struct mapping_tree {
  91         struct rb_root rb_root;
  92         spinlock_t lock;
  93 };
  94
  95 /*
  96  * present a tree block to process
  97  */
  98 struct tree_block {
  99         struct {
 100                 struct rb_node rb_node;
 101                 u64 bytenr;
 102         }; /* Use rb_simple_node for search/insert */
 103         u64 owner;
 104         struct btrfs_key key;
 105         unsigned int level:8;
 106         unsigned int key_ready:1;
 107 };
 108
 109 #define MAX_EXTENTS 128
 110
 111 struct file_extent_cluster {
 112         u64 start;
 113         u64 end;
 114         u64 boundary[MAX_EXTENTS];
 115         unsigned int nr;
 116 };
 117
 118 struct reloc_control {
 119         /* block group to relocate */
 120         struct btrfs_block_group *block_group;
 121         /* extent tree */
 122         struct btrfs_root *extent_root;
 123         /* inode for moving data */
 124         struct inode *data_inode;
 125
 126         struct btrfs_block_rsv *block_rsv;
 127
 128         struct btrfs_backref_cache backref_cache;
 129
 130         struct file_extent_cluster cluster;
 131         /* tree blocks have been processed */
 132         struct extent_io_tree processed_blocks;
 133         /* map start of tree root to corresponding reloc tree */
 134         struct mapping_tree reloc_root_tree;
 135         /* list of reloc trees */
 136         struct list_head reloc_roots;
 137         /* list of subvolume trees that get relocated */
 138         struct list_head dirty_subvol_roots;
 139         /* size of metadata reservation for merging reloc trees */
 140         u64 merging_rsv_size;
 141         /* size of relocated tree nodes */
 142         u64 nodes_relocated;
 143         /* reserved size for block group relocation*/
 144         u64 reserved_bytes;
 145
 146         u64 search_start;
 147         u64 extents_found;
 148
 149         unsigned int stage:8;
 150         unsigned int create_reloc_tree:1;
 151         unsigned int merge_reloc_tree:1;
 152         unsigned int found_file_extent:1;
 153 };
 154
 155 /* stages of data relocation */
 156 #define MOVE_DATA_EXTENTS       0
 157 #define UPDATE_DATA_PTRS        1
 158
 159 static void mark_block_processed(struct reloc_control *rc,
 160                                  struct btrfs_backref_node *node)
 161 {
 162         u32 blocksize;
 163
 164         if (node->level == 0 ||
 165             in_range(node->bytenr, rc->block_group->start,
 166                      rc->block_group->length)) {
 167                 blocksize = rc->extent_root->fs_info->nodesize;
 168                 set_extent_bits(&rc->processed_blocks, node->bytenr,
 169                                 node->bytenr + blocksize - 1, EXTENT_DIRTY);
 170         }
 171         node->processed = 1;
 172 }
 173
 174
 175 static void mapping_tree_init(struct mapping_tree *tree)
 176 {
 177         tree->rb_root = RB_ROOT;
 178         spin_lock_init(&tree->lock);
 179 }
 180
 181 /*
 182  * walk up backref nodes until reach node presents tree root
 183  */
 184 static struct btrfs_backref_node *walk_up_backref(
 185                 struct btrfs_backref_node *node,
 186                 struct btrfs_backref_edge *edges[], int *index)
 187 {
 188         struct btrfs_backref_edge *edge;
 189         int idx = *index;
 190
 191         while (!list_empty(&node->upper)) {
 192                 edge = list_entry(node->upper.next,
 193                                   struct btrfs_backref_edge, list[LOWER]);
 194                 edges[idx++] = edge;
 195                 node = edge->node[UPPER];
 196         }
 197         BUG_ON(node->detached);
 198         *index = idx;
 199         return node;
 200 }
 201
 202 /*
 203  * walk down backref nodes to find start of next reference path
 204  */
 205 static struct btrfs_backref_node *walk_down_backref(
 206                 struct btrfs_backref_edge *edges[], int *index)
 207 {
 208         struct btrfs_backref_edge *edge;
 209         struct btrfs_backref_node *lower;
 210         int idx = *index;
 211
 212         while (idx > 0) {
 213                 edge = edges[idx - 1];
 214                 lower = edge->node[LOWER];
 215                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 216                         idx--;
 217                         continue;
 218                 }
 219                 edge = list_entry(edge->list[LOWER].next,
 220                                   struct btrfs_backref_edge, list[LOWER]);
 221                 edges[idx - 1] = edge;
 222                 *index = idx;
 223                 return edge->node[UPPER];
 224         }
 225         *index = 0;
 226         return NULL;
 227 }
 228
 229 static void update_backref_node(struct btrfs_backref_cache *cache,
 230                                 struct btrfs_backref_node *node, u64 bytenr)
 231 {
 232         struct rb_node *rb_node;
 233         rb_erase(&node->rb_node, &cache->rb_root);
 234         node->bytenr = bytenr;
 235         rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 236         if (rb_node)
 237                 btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
 238 }
 239
 240 /*
 241  * update backref cache after a transaction commit
 242  */
 243 static int update_backref_cache(struct btrfs_trans_handle *trans,
 244                                 struct btrfs_backref_cache *cache)
 245 {
 246         struct btrfs_backref_node *node;
 247         int level = 0;
 248
 249         if (cache->last_trans == 0) {
 250                 cache->last_trans = trans->transid;
 251                 return 0;
 252         }
 253
 254         if (cache->last_trans == trans->transid)
 255                 return 0;
 256
 257         /*
 258          * detached nodes are used to avoid unnecessary backref
 259          * lookup. transaction commit changes the extent tree.
 260          * so the detached nodes are no longer useful.
 261          */
 262         while (!list_empty(&cache->detached)) {
 263                 node = list_entry(cache->detached.next,
 264                                   struct btrfs_backref_node, list);
 265                 btrfs_backref_cleanup_node(cache, node);
 266         }
 267
 268         while (!list_empty(&cache->changed)) {
 269                 node = list_entry(cache->changed.next,
 270                                   struct btrfs_backref_node, list);
 271                 list_del_init(&node->list);
 272                 BUG_ON(node->pending);
 273                 update_backref_node(cache, node, node->new_bytenr);
 274         }
 275
 276         /*
 277          * some nodes can be left in the pending list if there were
 278          * errors during processing the pending nodes.
 279          */
 280         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 281                 list_for_each_entry(node, &cache->pending[level], list) {
 282                         BUG_ON(!node->pending);
 283                         if (node->bytenr == node->new_bytenr)
 284                                 continue;
 285                         update_backref_node(cache, node, node->new_bytenr);
 286                 }
 287         }
 288
 289         cache->last_trans = 0;
 290         return 1;
 291 }
 292
 293 static bool reloc_root_is_dead(struct btrfs_root *root)
 294 {
 295         /*
 296          * Pair with set_bit/clear_bit in clean_dirty_subvols and
 297          * btrfs_update_reloc_root. We need to see the updated bit before
 298          * trying to access reloc_root
 299          */
 300         smp_rmb();
 301         if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
 302                 return true;
 303         return false;
 304 }
 305
 306 /*
 307  * Check if this subvolume tree has valid reloc tree.
 308  *
 309  * Reloc tree after swap is considered dead, thus not considered as valid.
 310  * This is enough for most callers, as they don't distinguish dead reloc root
 311  * from no reloc root.  But btrfs_should_ignore_reloc_root() below is a
 312  * special case.
 313  */
 314 static bool have_reloc_root(struct btrfs_root *root)
 315 {
 316         if (reloc_root_is_dead(root))
 317                 return false;
 318         if (!root->reloc_root)
 319                 return false;
 320         return true;
 321 }
 322
 323 int btrfs_should_ignore_reloc_root(struct btrfs_root *root)
 324 {
 325         struct btrfs_root *reloc_root;
 326
 327         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
 328                 return 0;
 329
 330         /* This root has been merged with its reloc tree, we can ignore it */
 331         if (reloc_root_is_dead(root))
 332                 return 1;
 333
 334         reloc_root = root->reloc_root;
 335         if (!reloc_root)
 336                 return 0;
 337
 338         if (btrfs_header_generation(reloc_root->commit_root) ==
 339             root->fs_info->running_transaction->transid)
 340                 return 0;
 341         /*
 342          * if there is reloc tree and it was created in previous
 343          * transaction backref lookup can find the reloc tree,
 344          * so backref node for the fs tree root is useless for
 345          * relocation.
 346          */
 347         return 1;
 348 }
 349
 350 /*
 351  * find reloc tree by address of tree root
 352  */
 353 struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
 354 {
 355         struct reloc_control *rc = fs_info->reloc_ctl;
 356         struct rb_node *rb_node;
 357         struct mapping_node *node;
 358         struct btrfs_root *root = NULL;
 359
 360         ASSERT(rc);
 361         spin_lock(&rc->reloc_root_tree.lock);
 362         rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
 363         if (rb_node) {
 364                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 365                 root = node->data;
 366         }
 367         spin_unlock(&rc->reloc_root_tree.lock);
 368         return btrfs_grab_root(root);
 369 }
 370
 371 /*
 372  * For useless nodes, do two major clean ups:
 373  *
 374  * - Cleanup the children edges and nodes
 375  *   If child node is also orphan (no parent) during cleanup, then the child
 376  *   node will also be cleaned up.
 377  *
 378  * - Freeing up leaves (level 0), keeps nodes detached
 379  *   For nodes, the node is still cached as "detached"
 380  *
 381  * Return false if @node is not in the @useless_nodes list.
 382  * Return true if @node is in the @useless_nodes list.
 383  */
 384 static bool handle_useless_nodes(struct reloc_control *rc,
 385                                  struct btrfs_backref_node *node)
 386 {
 387         struct btrfs_backref_cache *cache = &rc->backref_cache;
 388         struct list_head *useless_node = &cache->useless_node;
 389         bool ret = false;
 390
 391         while (!list_empty(useless_node)) {
 392                 struct btrfs_backref_node *cur;
 393
 394                 cur = list_first_entry(useless_node, struct btrfs_backref_node,
 395                                  list);
 396                 list_del_init(&cur->list);
 397
 398                 /* Only tree root nodes can be added to @useless_nodes */
 399                 ASSERT(list_empty(&cur->upper));
 400
 401                 if (cur == node)
 402                         ret = true;
 403
 404                 /* The node is the lowest node */
 405                 if (cur->lowest) {
 406                         list_del_init(&cur->lower);
 407                         cur->lowest = 0;
 408                 }
 409
 410                 /* Cleanup the lower edges */
 411                 while (!list_empty(&cur->lower)) {
 412                         struct btrfs_backref_edge *edge;
 413                         struct btrfs_backref_node *lower;
 414
 415                         edge = list_entry(cur->lower.next,
 416                                         struct btrfs_backref_edge, list[UPPER]);
 417                         list_del(&edge->list[UPPER]);
 418                         list_del(&edge->list[LOWER]);
 419                         lower = edge->node[LOWER];
 420                         btrfs_backref_free_edge(cache, edge);
 421
 422                         /* Child node is also orphan, queue for cleanup */
 423                         if (list_empty(&lower->upper))
 424                                 list_add(&lower->list, useless_node);
 425                 }
 426                 /* Mark this block processed for relocation */
 427                 mark_block_processed(rc, cur);
 428
 429                 /*
 430                  * Backref nodes for tree leaves are deleted from the cache.
 431                  * Backref nodes for upper level tree blocks are left in the
 432                  * cache to avoid unnecessary backref lookup.
 433                  */
 434                 if (cur->level > 0) {
 435                         list_add(&cur->list, &cache->detached);
 436                         cur->detached = 1;
 437                 } else {
 438                         rb_erase(&cur->rb_node, &cache->rb_root);
 439                         btrfs_backref_free_node(cache, cur);
 440                 }
 441         }
 442         return ret;
 443 }
 444
 445 /*
 446  * Build backref tree for a given tree block. Root of the backref tree
 447  * corresponds the tree block, leaves of the backref tree correspond roots of
 448  * b-trees that reference the tree block.
 449  *
 450  * The basic idea of this function is check backrefs of a given block to find
 451  * upper level blocks that reference the block, and then check backrefs of
 452  * these upper level blocks recursively. The recursion stops when tree root is
 453  * reached or backrefs for the block is cached.
 454  *
 455  * NOTE: if we find that backrefs for a block are cached, we know backrefs for
 456  * all upper level blocks that directly/indirectly reference the block are also
 457  * cached.
 458  */
 459 static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
 460                         struct reloc_control *rc, struct btrfs_key *node_key,
 461                         int level, u64 bytenr)
 462 {
 463         struct btrfs_backref_iter *iter;
 464         struct btrfs_backref_cache *cache = &rc->backref_cache;
 465         /* For searching parent of TREE_BLOCK_REF */
 466         struct btrfs_path *path;
 467         struct btrfs_backref_node *cur;
 468         struct btrfs_backref_node *node = NULL;
 469         struct btrfs_backref_edge *edge;
 470         int ret;
 471         int err = 0;
 472
 473         iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS);
 474         if (!iter)
 475                 return ERR_PTR(-ENOMEM);
 476         path = btrfs_alloc_path();
 477         if (!path) {
 478                 err = -ENOMEM;
 479                 goto out;
 480         }
 481
 482         node = btrfs_backref_alloc_node(cache, bytenr, level);
 483         if (!node) {
 484                 err = -ENOMEM;
 485                 goto out;
 486         }
 487
 488         node->lowest = 1;
 489         cur = node;
 490
 491         /* Breadth-first search to build backref cache */
 492         do {
 493                 ret = btrfs_backref_add_tree_node(cache, path, iter, node_key,
 494                                                   cur);
 495                 if (ret < 0) {
 496                         err = ret;
 497                         goto out;
 498                 }
 499                 edge = list_first_entry_or_null(&cache->pending_edge,
 500                                 struct btrfs_backref_edge, list[UPPER]);
 501                 /*
 502                  * The pending list isn't empty, take the first block to
 503                  * process
 504                  */
 505                 if (edge) {
 506                         list_del_init(&edge->list[UPPER]);
 507                         cur = edge->node[UPPER];
 508                 }
 509         } while (edge);
 510
 511         /* Finish the upper linkage of newly added edges/nodes */
 512         ret = btrfs_backref_finish_upper_links(cache, node);
 513         if (ret < 0) {
 514                 err = ret;
 515                 goto out;
 516         }
 517
 518         if (handle_useless_nodes(rc, node))
 519                 node = NULL;
 520 out:
 521         btrfs_backref_iter_free(iter);
 522         btrfs_free_path(path);
 523         if (err) {
 524                 btrfs_backref_error_cleanup(cache, node);
 525                 return ERR_PTR(err);
 526         }
 527         ASSERT(!node || !node->detached);
 528         ASSERT(list_empty(&cache->useless_node) &&
 529                list_empty(&cache->pending_edge));
 530         return node;
 531 }
 532
 533 /*
 534  * helper to add backref node for the newly created snapshot.
 535  * the backref node is created by cloning backref node that
 536  * corresponds to root of source tree
 537  */
 538 static int clone_backref_node(struct btrfs_trans_handle *trans,
 539                               struct reloc_control *rc,
 540                               struct btrfs_root *src,
 541                               struct btrfs_root *dest)
 542 {
 543         struct btrfs_root *reloc_root = src->reloc_root;
 544         struct btrfs_backref_cache *cache = &rc->backref_cache;
 545         struct btrfs_backref_node *node = NULL;
 546         struct btrfs_backref_node *new_node;
 547         struct btrfs_backref_edge *edge;
 548         struct btrfs_backref_edge *new_edge;
 549         struct rb_node *rb_node;
 550
 551         if (cache->last_trans > 0)
 552                 update_backref_cache(trans, cache);
 553
 554         rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
 555         if (rb_node) {
 556                 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
 557                 if (node->detached)
 558                         node = NULL;
 559                 else
 560                         BUG_ON(node->new_bytenr != reloc_root->node->start);
 561         }
 562
 563         if (!node) {
 564                 rb_node = rb_simple_search(&cache->rb_root,
 565                                            reloc_root->commit_root->start);
 566                 if (rb_node) {
 567                         node = rb_entry(rb_node, struct btrfs_backref_node,
 568                                         rb_node);
 569                         BUG_ON(node->detached);
 570                 }
 571         }
 572
 573         if (!node)
 574                 return 0;
 575
 576         new_node = btrfs_backref_alloc_node(cache, dest->node->start,
 577                                             node->level);
 578         if (!new_node)
 579                 return -ENOMEM;
 580
 581         new_node->lowest = node->lowest;
 582         new_node->checked = 1;
 583         new_node->root = btrfs_grab_root(dest);
 584         ASSERT(new_node->root);
 585
 586         if (!node->lowest) {
 587                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
 588                         new_edge = btrfs_backref_alloc_edge(cache);
 589                         if (!new_edge)
 590                                 goto fail;
 591
 592                         btrfs_backref_link_edge(new_edge, edge->node[LOWER],
 593                                                 new_node, LINK_UPPER);
 594                 }
 595         } else {
 596                 list_add_tail(&new_node->lower, &cache->leaves);
 597         }
 598
 599         rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
 600                                    &new_node->rb_node);
 601         if (rb_node)
 602                 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
 603
 604         if (!new_node->lowest) {
 605                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
 606                         list_add_tail(&new_edge->list[LOWER],
 607                                       &new_edge->node[LOWER]->upper);
 608                 }
 609         }
 610         return 0;
 611 fail:
 612         while (!list_empty(&new_node->lower)) {
 613                 new_edge = list_entry(new_node->lower.next,
 614                                       struct btrfs_backref_edge, list[UPPER]);
 615                 list_del(&new_edge->list[UPPER]);
 616                 btrfs_backref_free_edge(cache, new_edge);
 617         }
 618         btrfs_backref_free_node(cache, new_node);
 619         return -ENOMEM;
 620 }
 621
 622 /*
 623  * helper to add 'address of tree root -> reloc tree' mapping
 624  */
 625 static int __must_check __add_reloc_root(struct btrfs_root *root)
 626 {
 627         struct btrfs_fs_info *fs_info = root->fs_info;
 628         struct rb_node *rb_node;
 629         struct mapping_node *node;
 630         struct reloc_control *rc = fs_info->reloc_ctl;
 631
 632         node = kmalloc(sizeof(*node), GFP_NOFS);
 633         if (!node)
 634                 return -ENOMEM;
 635
 636         node->bytenr = root->commit_root->start;
 637         node->data = root;
 638
 639         spin_lock(&rc->reloc_root_tree.lock);
 640         rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
 641                                    node->bytenr, &node->rb_node);
 642         spin_unlock(&rc->reloc_root_tree.lock);
 643         if (rb_node) {
 644                 btrfs_err(fs_info,
 645                             "Duplicate root found for start=%llu while inserting into relocation tree",
 646                             node->bytenr);
 647                 return -EEXIST;
 648         }
 649
 650         list_add_tail(&root->root_list, &rc->reloc_roots);
 651         return 0;
 652 }
 653
 654 /*
 655  * helper to delete the 'address of tree root -> reloc tree'
 656  * mapping
 657  */
 658 static void __del_reloc_root(struct btrfs_root *root)
 659 {
 660         struct btrfs_fs_info *fs_info = root->fs_info;
 661         struct rb_node *rb_node;
 662         struct mapping_node *node = NULL;
 663         struct reloc_control *rc = fs_info->reloc_ctl;
 664         bool put_ref = false;
 665
 666         if (rc && root->node) {
 667                 spin_lock(&rc->reloc_root_tree.lock);
 668                 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
 669                                            root->commit_root->start);
 670                 if (rb_node) {
 671                         node = rb_entry(rb_node, struct mapping_node, rb_node);
 672                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
 673                         RB_CLEAR_NODE(&node->rb_node);
 674                 }
 675                 spin_unlock(&rc->reloc_root_tree.lock);
 676                 ASSERT(!node || (struct btrfs_root *)node->data == root);
 677         }
 678
 679         /*
 680          * We only put the reloc root here if it's on the list.  There's a lot
 681          * of places where the pattern is to splice the rc->reloc_roots, process
 682          * the reloc roots, and then add the reloc root back onto
 683          * rc->reloc_roots.  If we call __del_reloc_root while it's off of the
 684          * list we don't want the reference being dropped, because the guy
 685          * messing with the list is in charge of the reference.
 686          */
 687         spin_lock(&fs_info->trans_lock);
 688         if (!list_empty(&root->root_list)) {
 689                 put_ref = true;
 690                 list_del_init(&root->root_list);
 691         }
 692         spin_unlock(&fs_info->trans_lock);
 693         if (put_ref)
 694                 btrfs_put_root(root);
 695         kfree(node);
 696 }
 697
 698 /*
 699  * helper to update the 'address of tree root -> reloc tree'
 700  * mapping
 701  */
 702 static int __update_reloc_root(struct btrfs_root *root)
 703 {
 704         struct btrfs_fs_info *fs_info = root->fs_info;
 705         struct rb_node *rb_node;
 706         struct mapping_node *node = NULL;
 707         struct reloc_control *rc = fs_info->reloc_ctl;
 708
 709         spin_lock(&rc->reloc_root_tree.lock);
 710         rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
 711                                    root->commit_root->start);
 712         if (rb_node) {
 713                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 714                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
 715         }
 716         spin_unlock(&rc->reloc_root_tree.lock);
 717
 718         if (!node)
 719                 return 0;
 720         BUG_ON((struct btrfs_root *)node->data != root);
 721
 722         spin_lock(&rc->reloc_root_tree.lock);
 723         node->bytenr = root->node->start;
 724         rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
 725                                    node->bytenr, &node->rb_node);
 726         spin_unlock(&rc->reloc_root_tree.lock);
 727         if (rb_node)
 728                 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
 729         return 0;
 730 }
 731
 732 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
 733                                         struct btrfs_root *root, u64 objectid)
 734 {
 735         struct btrfs_fs_info *fs_info = root->fs_info;
 736         struct btrfs_root *reloc_root;
 737         struct extent_buffer *eb;
 738         struct btrfs_root_item *root_item;
 739         struct btrfs_key root_key;
 740         int ret = 0;
 741         bool must_abort = false;
 742
 743         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
 744         if (!root_item)
 745                 return ERR_PTR(-ENOMEM);
 746
 747         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
 748         root_key.type = BTRFS_ROOT_ITEM_KEY;
 749         root_key.offset = objectid;
 750
 751         if (root->root_key.objectid == objectid) {
 752                 u64 commit_root_gen;
 753
 754                 /* called by btrfs_init_reloc_root */
 755                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
 756                                       BTRFS_TREE_RELOC_OBJECTID);
 757                 if (ret)
 758                         goto fail;
 759
 760                 /*
 761                  * Set the last_snapshot field to the generation of the commit
 762                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
 763                  * correctly (returns true) when the relocation root is created
 764                  * either inside the critical section of a transaction commit
 765                  * (through transaction.c:qgroup_account_snapshot()) and when
 766                  * it's created before the transaction commit is started.
 767                  */
 768                 commit_root_gen = btrfs_header_generation(root->commit_root);
 769                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
 770         } else {
 771                 /*
 772                  * called by btrfs_reloc_post_snapshot_hook.
 773                  * the source tree is a reloc tree, all tree blocks
 774                  * modified after it was created have RELOC flag
 775                  * set in their headers. so it's OK to not update
 776                  * the 'last_snapshot'.
 777                  */
 778                 ret = btrfs_copy_root(trans, root, root->node, &eb,
 779                                       BTRFS_TREE_RELOC_OBJECTID);
 780                 if (ret)
 781                         goto fail;
 782         }
 783
 784         /*
 785          * We have changed references at this point, we must abort the
 786          * transaction if anything fails.
 787          */
 788         must_abort = true;
 789
 790         memcpy(root_item, &root->root_item, sizeof(*root_item));
 791         btrfs_set_root_bytenr(root_item, eb->start);
 792         btrfs_set_root_level(root_item, btrfs_header_level(eb));
 793         btrfs_set_root_generation(root_item, trans->transid);
 794
 795         if (root->root_key.objectid == objectid) {
 796                 btrfs_set_root_refs(root_item, 0);
 797                 memset(&root_item->drop_progress, 0,
 798                        sizeof(struct btrfs_disk_key));
 799                 btrfs_set_root_drop_level(root_item, 0);
 800         }
 801
 802         btrfs_tree_unlock(eb);
 803         free_extent_buffer(eb);
 804
 805         ret = btrfs_insert_root(trans, fs_info->tree_root,
 806                                 &root_key, root_item);
 807         if (ret)
 808                 goto fail;
 809
 810         kfree(root_item);
 811
 812         reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
 813         if (IS_ERR(reloc_root)) {
 814                 ret = PTR_ERR(reloc_root);
 815                 goto abort;
 816         }
 817         set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
 818         reloc_root->last_trans = trans->transid;
 819         return reloc_root;
 820 fail:
 821         kfree(root_item);
 822 abort:
 823         if (must_abort)
 824                 btrfs_abort_transaction(trans, ret);
 825         return ERR_PTR(ret);
 826 }
 827
 828 /*
 829  * create reloc tree for a given fs tree. reloc tree is just a
 830  * snapshot of the fs tree with special root objectid.
 831  *
 832  * The reloc_root comes out of here with two references, one for
 833  * root->reloc_root, and another for being on the rc->reloc_roots list.
 834  */
 835 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
 836                           struct btrfs_root *root)
 837 {
 838         struct btrfs_fs_info *fs_info = root->fs_info;
 839         struct btrfs_root *reloc_root;
 840         struct reloc_control *rc = fs_info->reloc_ctl;
 841         struct btrfs_block_rsv *rsv;
 842         int clear_rsv = 0;
 843         int ret;
 844
 845         if (!rc)
 846                 return 0;
 847
 848         /*
 849          * The subvolume has reloc tree but the swap is finished, no need to
 850          * create/update the dead reloc tree
 851          */
 852         if (reloc_root_is_dead(root))
 853                 return 0;
 854
 855         /*
 856          * This is subtle but important.  We do not do
 857          * record_root_in_transaction for reloc roots, instead we record their
 858          * corresponding fs root, and then here we update the last trans for the
 859          * reloc root.  This means that we have to do this for the entire life
 860          * of the reloc root, regardless of which stage of the relocation we are
 861          * in.
 862          */
 863         if (root->reloc_root) {
 864                 reloc_root = root->reloc_root;
 865                 reloc_root->last_trans = trans->transid;
 866                 return 0;
 867         }
 868
 869         /*
 870          * We are merging reloc roots, we do not need new reloc trees.  Also
 871          * reloc trees never need their own reloc tree.
 872          */
 873         if (!rc->create_reloc_tree ||
 874             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 875                 return 0;
 876
 877         if (!trans->reloc_reserved) {
 878                 rsv = trans->block_rsv;
 879                 trans->block_rsv = rc->block_rsv;
 880                 clear_rsv = 1;
 881         }
 882         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
 883         if (clear_rsv)
 884                 trans->block_rsv = rsv;
 885         if (IS_ERR(reloc_root))
 886                 return PTR_ERR(reloc_root);
 887
 888         ret = __add_reloc_root(reloc_root);
 889         ASSERT(ret != -EEXIST);
 890         if (ret) {
 891                 /* Pairs with create_reloc_root */
 892                 btrfs_put_root(reloc_root);
 893                 return ret;
 894         }
 895         root->reloc_root = btrfs_grab_root(reloc_root);
 896         return 0;
 897 }
 898
 899 /*
 900  * update root item of reloc tree
 901  */
 902 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
 903                             struct btrfs_root *root)
 904 {
 905         struct btrfs_fs_info *fs_info = root->fs_info;
 906         struct btrfs_root *reloc_root;
 907         struct btrfs_root_item *root_item;
 908         int ret;
 909
 910         if (!have_reloc_root(root))
 911                 return 0;
 912
 913         reloc_root = root->reloc_root;
 914         root_item = &reloc_root->root_item;
 915
 916         /*
 917          * We are probably ok here, but __del_reloc_root() will drop its ref of
 918          * the root.  We have the ref for root->reloc_root, but just in case
 919          * hold it while we update the reloc root.
 920          */
 921         btrfs_grab_root(reloc_root);
 922
 923         /* root->reloc_root will stay until current relocation finished */
 924         if (fs_info->reloc_ctl->merge_reloc_tree &&
 925             btrfs_root_refs(root_item) == 0) {
 926                 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
 927                 /*
 928                  * Mark the tree as dead before we change reloc_root so
 929                  * have_reloc_root will not touch it from now on.
 930                  */
 931                 smp_wmb();
 932                 __del_reloc_root(reloc_root);
 933         }
 934
 935         if (reloc_root->commit_root != reloc_root->node) {
 936                 __update_reloc_root(reloc_root);
 937                 btrfs_set_root_node(root_item, reloc_root->node);
 938                 free_extent_buffer(reloc_root->commit_root);
 939                 reloc_root->commit_root = btrfs_root_node(reloc_root);
 940         }
 941
 942         ret = btrfs_update_root(trans, fs_info->tree_root,
 943                                 &reloc_root->root_key, root_item);
 944         btrfs_put_root(reloc_root);
 945         return ret;
 946 }
 947
 948 /*
 949  * helper to find first cached inode with inode number >= objectid
 950  * in a subvolume
 951  */
 952 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
 953 {
 954         struct rb_node *node;
 955         struct rb_node *prev;
 956         struct btrfs_inode *entry;
 957         struct inode *inode;
 958
 959         spin_lock(&root->inode_lock);
 960 again:
 961         node = root->inode_tree.rb_node;
 962         prev = NULL;
 963         while (node) {
 964                 prev = node;
 965                 entry = rb_entry(node, struct btrfs_inode, rb_node);
 966
 967                 if (objectid < btrfs_ino(entry))
 968                         node = node->rb_left;
 969                 else if (objectid > btrfs_ino(entry))
 970                         node = node->rb_right;
 971                 else
 972                         break;
 973         }
 974         if (!node) {
 975                 while (prev) {
 976                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
 977                         if (objectid <= btrfs_ino(entry)) {
 978                                 node = prev;
 979                                 break;
 980                         }
 981                         prev = rb_next(prev);
 982                 }
 983         }
 984         while (node) {
 985                 entry = rb_entry(node, struct btrfs_inode, rb_node);
 986                 inode = igrab(&entry->vfs_inode);
 987                 if (inode) {
 988                         spin_unlock(&root->inode_lock);
 989                         return inode;
 990                 }
 991
 992                 objectid = btrfs_ino(entry) + 1;
 993                 if (cond_resched_lock(&root->inode_lock))
 994                         goto again;
 995
 996                 node = rb_next(node);
 997         }
 998         spin_unlock(&root->inode_lock);
 999         return NULL;
1000 }
1001
1002 /*
1003  * get new location of data
1004  */
1005 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1006                             u64 bytenr, u64 num_bytes)
1007 {
1008         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1009         struct btrfs_path *path;
1010         struct btrfs_file_extent_item *fi;
1011         struct extent_buffer *leaf;
1012         int ret;
1013
1014         path = btrfs_alloc_path();
1015         if (!path)
1016                 return -ENOMEM;
1017
1018         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1019         ret = btrfs_lookup_file_extent(NULL, root, path,
1020                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1021         if (ret < 0)
1022                 goto out;
1023         if (ret > 0) {
1024                 ret = -ENOENT;
1025                 goto out;
1026         }
1027
1028         leaf = path->nodes[0];
1029         fi = btrfs_item_ptr(leaf, path->slots[0],
1030                             struct btrfs_file_extent_item);
1031
1032         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1033                btrfs_file_extent_compression(leaf, fi) ||
1034                btrfs_file_extent_encryption(leaf, fi) ||
1035                btrfs_file_extent_other_encoding(leaf, fi));
1036
1037         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1038                 ret = -EINVAL;
1039                 goto out;
1040         }
1041
1042         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1043         ret = 0;
1044 out:
1045         btrfs_free_path(path);
1046         return ret;
1047 }
1048
1049 /*
1050  * update file extent items in the tree leaf to point to
1051  * the new locations.
1052  */
1053 static noinline_for_stack
1054 int replace_file_extents(struct btrfs_trans_handle *trans,
1055                          struct reloc_control *rc,
1056                          struct btrfs_root *root,
1057                          struct extent_buffer *leaf)
1058 {
1059         struct btrfs_fs_info *fs_info = root->fs_info;
1060         struct btrfs_key key;
1061         struct btrfs_file_extent_item *fi;
1062         struct inode *inode = NULL;
1063         u64 parent;
1064         u64 bytenr;
1065         u64 new_bytenr = 0;
1066         u64 num_bytes;
1067         u64 end;
1068         u32 nritems;
1069         u32 i;
1070         int ret = 0;
1071         int first = 1;
1072         int dirty = 0;
1073
1074         if (rc->stage != UPDATE_DATA_PTRS)
1075                 return 0;
1076
1077         /* reloc trees always use full backref */
1078         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1079                 parent = leaf->start;
1080         else
1081                 parent = 0;
1082
1083         nritems = btrfs_header_nritems(leaf);
1084         for (i = 0; i < nritems; i++) {
1085                 struct btrfs_ref ref = { 0 };
1086
1087                 cond_resched();
1088                 btrfs_item_key_to_cpu(leaf, &key, i);
1089                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1090                         continue;
1091                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1092                 if (btrfs_file_extent_type(leaf, fi) ==
1093                     BTRFS_FILE_EXTENT_INLINE)
1094                         continue;
1095                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1096                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1097                 if (bytenr == 0)
1098                         continue;
1099                 if (!in_range(bytenr, rc->block_group->start,
1100                               rc->block_group->length))
1101                         continue;
1102
1103                 /*
1104                  * if we are modifying block in fs tree, wait for read_folio
1105                  * to complete and drop the extent cache
1106                  */
1107                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1108                         if (first) {
1109                                 inode = find_next_inode(root, key.objectid);
1110                                 first = 0;
1111                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1112                                 btrfs_add_delayed_iput(inode);
1113                                 inode = find_next_inode(root, key.objectid);
1114                         }
1115                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1116                                 end = key.offset +
1117                                       btrfs_file_extent_num_bytes(leaf, fi);
1118                                 WARN_ON(!IS_ALIGNED(key.offset,
1119                                                     fs_info->sectorsize));
1120                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1121                                 end--;
1122                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1123                                                       key.offset, end);
1124                                 if (!ret)
1125                                         continue;
1126
1127                                 btrfs_drop_extent_map_range(BTRFS_I(inode),
1128                                                             key.offset, end, true);
1129                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1130                                               key.offset, end, NULL);
1131                         }
1132                 }
1133
1134                 ret = get_new_location(rc->data_inode, &new_bytenr,
1135                                        bytenr, num_bytes);
1136                 if (ret) {
1137                         /*
1138                          * Don't have to abort since we've not changed anything
1139                          * in the file extent yet.
1140                          */
1141                         break;
1142                 }
1143
1144                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1145                 dirty = 1;
1146
1147                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1148                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1149                                        num_bytes, parent);
1150                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1151                                     key.objectid, key.offset,
1152                                     root->root_key.objectid, false);
1153                 ret = btrfs_inc_extent_ref(trans, &ref);
1154                 if (ret) {
1155                         btrfs_abort_transaction(trans, ret);
1156                         break;
1157                 }
1158
1159                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1160                                        num_bytes, parent);
1161                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1162                                     key.objectid, key.offset,
1163                                     root->root_key.objectid, false);
1164                 ret = btrfs_free_extent(trans, &ref);
1165                 if (ret) {
1166                         btrfs_abort_transaction(trans, ret);
1167                         break;
1168                 }
1169         }
1170         if (dirty)
1171                 btrfs_mark_buffer_dirty(leaf);
1172         if (inode)
1173                 btrfs_add_delayed_iput(inode);
1174         return ret;
1175 }
1176
1177 static noinline_for_stack
1178 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1179                      struct btrfs_path *path, int level)
1180 {
1181         struct btrfs_disk_key key1;
1182         struct btrfs_disk_key key2;
1183         btrfs_node_key(eb, &key1, slot);
1184         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1185         return memcmp(&key1, &key2, sizeof(key1));
1186 }
1187
1188 /*
1189  * try to replace tree blocks in fs tree with the new blocks
1190  * in reloc tree. tree blocks haven't been modified since the
1191  * reloc tree was create can be replaced.
1192  *
1193  * if a block was replaced, level of the block + 1 is returned.
1194  * if no block got replaced, 0 is returned. if there are other
1195  * errors, a negative error number is returned.
1196  */
1197 static noinline_for_stack
1198 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1199                  struct btrfs_root *dest, struct btrfs_root *src,
1200                  struct btrfs_path *path, struct btrfs_key *next_key,
1201                  int lowest_level, int max_level)
1202 {
1203         struct btrfs_fs_info *fs_info = dest->fs_info;
1204         struct extent_buffer *eb;
1205         struct extent_buffer *parent;
1206         struct btrfs_ref ref = { 0 };
1207         struct btrfs_key key;
1208         u64 old_bytenr;
1209         u64 new_bytenr;
1210         u64 old_ptr_gen;
1211         u64 new_ptr_gen;
1212         u64 last_snapshot;
1213         u32 blocksize;
1214         int cow = 0;
1215         int level;
1216         int ret;
1217         int slot;
1218
1219         ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1220         ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1221
1222         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1223 again:
1224         slot = path->slots[lowest_level];
1225         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1226
1227         eb = btrfs_lock_root_node(dest);
1228         level = btrfs_header_level(eb);
1229
1230         if (level < lowest_level) {
1231                 btrfs_tree_unlock(eb);
1232                 free_extent_buffer(eb);
1233                 return 0;
1234         }
1235
1236         if (cow) {
1237                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1238                                       BTRFS_NESTING_COW);
1239                 if (ret) {
1240                         btrfs_tree_unlock(eb);
1241                         free_extent_buffer(eb);
1242                         return ret;
1243                 }
1244         }
1245
1246         if (next_key) {
1247                 next_key->objectid = (u64)-1;
1248                 next_key->type = (u8)-1;
1249                 next_key->offset = (u64)-1;
1250         }
1251
1252         parent = eb;
1253         while (1) {
1254                 level = btrfs_header_level(parent);
1255                 ASSERT(level >= lowest_level);
1256
1257                 ret = btrfs_bin_search(parent, &key, &slot);
1258                 if (ret < 0)
1259                         break;
1260                 if (ret && slot > 0)
1261                         slot--;
1262
1263                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1264                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1265
1266                 old_bytenr = btrfs_node_blockptr(parent, slot);
1267                 blocksize = fs_info->nodesize;
1268                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1269
1270                 if (level <= max_level) {
1271                         eb = path->nodes[level];
1272                         new_bytenr = btrfs_node_blockptr(eb,
1273                                                         path->slots[level]);
1274                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1275                                                         path->slots[level]);
1276                 } else {
1277                         new_bytenr = 0;
1278                         new_ptr_gen = 0;
1279                 }
1280
1281                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1282                         ret = level;
1283                         break;
1284                 }
1285
1286                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1287                     memcmp_node_keys(parent, slot, path, level)) {
1288                         if (level <= lowest_level) {
1289                                 ret = 0;
1290                                 break;
1291                         }
1292
1293                         eb = btrfs_read_node_slot(parent, slot);
1294                         if (IS_ERR(eb)) {
1295                                 ret = PTR_ERR(eb);
1296                                 break;
1297                         }
1298                         btrfs_tree_lock(eb);
1299                         if (cow) {
1300                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1301                                                       slot, &eb,
1302                                                       BTRFS_NESTING_COW);
1303                                 if (ret) {
1304                                         btrfs_tree_unlock(eb);
1305                                         free_extent_buffer(eb);
1306                                         break;
1307                                 }
1308                         }
1309
1310                         btrfs_tree_unlock(parent);
1311                         free_extent_buffer(parent);
1312
1313                         parent = eb;
1314                         continue;
1315                 }
1316
1317                 if (!cow) {
1318                         btrfs_tree_unlock(parent);
1319                         free_extent_buffer(parent);
1320                         cow = 1;
1321                         goto again;
1322                 }
1323
1324                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1325                                       path->slots[level]);
1326                 btrfs_release_path(path);
1327
1328                 path->lowest_level = level;
1329                 set_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1330                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1331                 clear_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1332                 path->lowest_level = 0;
1333                 if (ret) {
1334                         if (ret > 0)
1335                                 ret = -ENOENT;
1336                         break;
1337                 }
1338
1339                 /*
1340                  * Info qgroup to trace both subtrees.
1341                  *
1342                  * We must trace both trees.
1343                  * 1) Tree reloc subtree
1344                  *    If not traced, we will leak data numbers
1345                  * 2) Fs subtree
1346                  *    If not traced, we will double count old data
1347                  *
1348                  * We don't scan the subtree right now, but only record
1349                  * the swapped tree blocks.
1350                  * The real subtree rescan is delayed until we have new
1351                  * CoW on the subtree root node before transaction commit.
1352                  */
1353                 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1354                                 rc->block_group, parent, slot,
1355                                 path->nodes[level], path->slots[level],
1356                                 last_snapshot);
1357                 if (ret < 0)
1358                         break;
1359                 /*
1360                  * swap blocks in fs tree and reloc tree.
1361                  */
1362                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1363                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1364                 btrfs_mark_buffer_dirty(parent);
1365
1366                 btrfs_set_node_blockptr(path->nodes[level],
1367                                         path->slots[level], old_bytenr);
1368                 btrfs_set_node_ptr_generation(path->nodes[level],
1369                                               path->slots[level], old_ptr_gen);
1370                 btrfs_mark_buffer_dirty(path->nodes[level]);
1371
1372                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1373                                        blocksize, path->nodes[level]->start);
1374                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1375                                     0, true);
1376                 ret = btrfs_inc_extent_ref(trans, &ref);
1377                 if (ret) {
1378                         btrfs_abort_transaction(trans, ret);
1379                         break;
1380                 }
1381                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1382                                        blocksize, 0);
1383                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid, 0,
1384                                     true);
1385                 ret = btrfs_inc_extent_ref(trans, &ref);
1386                 if (ret) {
1387                         btrfs_abort_transaction(trans, ret);
1388                         break;
1389                 }
1390
1391                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1392                                        blocksize, path->nodes[level]->start);
1393                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1394                                     0, true);
1395                 ret = btrfs_free_extent(trans, &ref);
1396                 if (ret) {
1397                         btrfs_abort_transaction(trans, ret);
1398                         break;
1399                 }
1400
1401                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1402                                        blocksize, 0);
1403                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid,
1404                                     0, true);
1405                 ret = btrfs_free_extent(trans, &ref);
1406                 if (ret) {
1407                         btrfs_abort_transaction(trans, ret);
1408                         break;
1409                 }
1410
1411                 btrfs_unlock_up_safe(path, 0);
1412
1413                 ret = level;
1414                 break;
1415         }
1416         btrfs_tree_unlock(parent);
1417         free_extent_buffer(parent);
1418         return ret;
1419 }
1420
1421 /*
1422  * helper to find next relocated block in reloc tree
1423  */
1424 static noinline_for_stack
1425 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1426                        int *level)
1427 {
1428         struct extent_buffer *eb;
1429         int i;
1430         u64 last_snapshot;
1431         u32 nritems;
1432
1433         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1434
1435         for (i = 0; i < *level; i++) {
1436                 free_extent_buffer(path->nodes[i]);
1437                 path->nodes[i] = NULL;
1438         }
1439
1440         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1441                 eb = path->nodes[i];
1442                 nritems = btrfs_header_nritems(eb);
1443                 while (path->slots[i] + 1 < nritems) {
1444                         path->slots[i]++;
1445                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1446                             last_snapshot)
1447                                 continue;
1448
1449                         *level = i;
1450                         return 0;
1451                 }
1452                 free_extent_buffer(path->nodes[i]);
1453                 path->nodes[i] = NULL;
1454         }
1455         return 1;
1456 }
1457
1458 /*
1459  * walk down reloc tree to find relocated block of lowest level
1460  */
1461 static noinline_for_stack
1462 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1463                          int *level)
1464 {
1465         struct extent_buffer *eb = NULL;
1466         int i;
1467         u64 ptr_gen = 0;
1468         u64 last_snapshot;
1469         u32 nritems;
1470
1471         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1472
1473         for (i = *level; i > 0; i--) {
1474                 eb = path->nodes[i];
1475                 nritems = btrfs_header_nritems(eb);
1476                 while (path->slots[i] < nritems) {
1477                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1478                         if (ptr_gen > last_snapshot)
1479                                 break;
1480                         path->slots[i]++;
1481                 }
1482                 if (path->slots[i] >= nritems) {
1483                         if (i == *level)
1484                                 break;
1485                         *level = i + 1;
1486                         return 0;
1487                 }
1488                 if (i == 1) {
1489                         *level = i;
1490                         return 0;
1491                 }
1492
1493                 eb = btrfs_read_node_slot(eb, path->slots[i]);
1494                 if (IS_ERR(eb))
1495                         return PTR_ERR(eb);
1496                 BUG_ON(btrfs_header_level(eb) != i - 1);
1497                 path->nodes[i - 1] = eb;
1498                 path->slots[i - 1] = 0;
1499         }
1500         return 1;
1501 }
1502
1503 /*
1504  * invalidate extent cache for file extents whose key in range of
1505  * [min_key, max_key)
1506  */
1507 static int invalidate_extent_cache(struct btrfs_root *root,
1508                                    struct btrfs_key *min_key,
1509                                    struct btrfs_key *max_key)
1510 {
1511         struct btrfs_fs_info *fs_info = root->fs_info;
1512         struct inode *inode = NULL;
1513         u64 objectid;
1514         u64 start, end;
1515         u64 ino;
1516
1517         objectid = min_key->objectid;
1518         while (1) {
1519                 cond_resched();
1520                 iput(inode);
1521
1522                 if (objectid > max_key->objectid)
1523                         break;
1524
1525                 inode = find_next_inode(root, objectid);
1526                 if (!inode)
1527                         break;
1528                 ino = btrfs_ino(BTRFS_I(inode));
1529
1530                 if (ino > max_key->objectid) {
1531                         iput(inode);
1532                         break;
1533                 }
1534
1535                 objectid = ino + 1;
1536                 if (!S_ISREG(inode->i_mode))
1537                         continue;
1538
1539                 if (unlikely(min_key->objectid == ino)) {
1540                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1541                                 continue;
1542                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1543                                 start = 0;
1544                         else {
1545                                 start = min_key->offset;
1546                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1547                         }
1548                 } else {
1549                         start = 0;
1550                 }
1551
1552                 if (unlikely(max_key->objectid == ino)) {
1553                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1554                                 continue;
1555                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1556                                 end = (u64)-1;
1557                         } else {
1558                                 if (max_key->offset == 0)
1559                                         continue;
1560                                 end = max_key->offset;
1561                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1562                                 end--;
1563                         }
1564                 } else {
1565                         end = (u64)-1;
1566                 }
1567
1568                 /* the lock_extent waits for read_folio to complete */
1569                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, NULL);
1570                 btrfs_drop_extent_map_range(BTRFS_I(inode), start, end, true);
1571                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, NULL);
1572         }
1573         return 0;
1574 }
1575
1576 static int find_next_key(struct btrfs_path *path, int level,
1577                          struct btrfs_key *key)
1578
1579 {
1580         while (level < BTRFS_MAX_LEVEL) {
1581                 if (!path->nodes[level])
1582                         break;
1583                 if (path->slots[level] + 1 <
1584                     btrfs_header_nritems(path->nodes[level])) {
1585                         btrfs_node_key_to_cpu(path->nodes[level], key,
1586                                               path->slots[level] + 1);
1587                         return 0;
1588                 }
1589                 level++;
1590         }
1591         return 1;
1592 }
1593
1594 /*
1595  * Insert current subvolume into reloc_control::dirty_subvol_roots
1596  */
1597 static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1598                                struct reloc_control *rc,
1599                                struct btrfs_root *root)
1600 {
1601         struct btrfs_root *reloc_root = root->reloc_root;
1602         struct btrfs_root_item *reloc_root_item;
1603         int ret;
1604
1605         /* @root must be a subvolume tree root with a valid reloc tree */
1606         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1607         ASSERT(reloc_root);
1608
1609         reloc_root_item = &reloc_root->root_item;
1610         memset(&reloc_root_item->drop_progress, 0,
1611                 sizeof(reloc_root_item->drop_progress));
1612         btrfs_set_root_drop_level(reloc_root_item, 0);
1613         btrfs_set_root_refs(reloc_root_item, 0);
1614         ret = btrfs_update_reloc_root(trans, root);
1615         if (ret)
1616                 return ret;
1617
1618         if (list_empty(&root->reloc_dirty_list)) {
1619                 btrfs_grab_root(root);
1620                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1621         }
1622
1623         return 0;
1624 }
1625
1626 static int clean_dirty_subvols(struct reloc_control *rc)
1627 {
1628         struct btrfs_root *root;
1629         struct btrfs_root *next;
1630         int ret = 0;
1631         int ret2;
1632
1633         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1634                                  reloc_dirty_list) {
1635                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1636                         /* Merged subvolume, cleanup its reloc root */
1637                         struct btrfs_root *reloc_root = root->reloc_root;
1638
1639                         list_del_init(&root->reloc_dirty_list);
1640                         root->reloc_root = NULL;
1641                         /*
1642                          * Need barrier to ensure clear_bit() only happens after
1643                          * root->reloc_root = NULL. Pairs with have_reloc_root.
1644                          */
1645                         smp_wmb();
1646                         clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1647                         if (reloc_root) {
1648                                 /*
1649                                  * btrfs_drop_snapshot drops our ref we hold for
1650                                  * ->reloc_root.  If it fails however we must
1651                                  * drop the ref ourselves.
1652                                  */
1653                                 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1654                                 if (ret2 < 0) {
1655                                         btrfs_put_root(reloc_root);
1656                                         if (!ret)
1657                                                 ret = ret2;
1658                                 }
1659                         }
1660                         btrfs_put_root(root);
1661                 } else {
1662                         /* Orphan reloc tree, just clean it up */
1663                         ret2 = btrfs_drop_snapshot(root, 0, 1);
1664                         if (ret2 < 0) {
1665                                 btrfs_put_root(root);
1666                                 if (!ret)
1667                                         ret = ret2;
1668                         }
1669                 }
1670         }
1671         return ret;
1672 }
1673
1674 /*
1675  * merge the relocated tree blocks in reloc tree with corresponding
1676  * fs tree.
1677  */
1678 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1679                                                struct btrfs_root *root)
1680 {
1681         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1682         struct btrfs_key key;
1683         struct btrfs_key next_key;
1684         struct btrfs_trans_handle *trans = NULL;
1685         struct btrfs_root *reloc_root;
1686         struct btrfs_root_item *root_item;
1687         struct btrfs_path *path;
1688         struct extent_buffer *leaf;
1689         int reserve_level;
1690         int level;
1691         int max_level;
1692         int replaced = 0;
1693         int ret = 0;
1694         u32 min_reserved;
1695
1696         path = btrfs_alloc_path();
1697         if (!path)
1698                 return -ENOMEM;
1699         path->reada = READA_FORWARD;
1700
1701         reloc_root = root->reloc_root;
1702         root_item = &reloc_root->root_item;
1703
1704         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1705                 level = btrfs_root_level(root_item);
1706                 atomic_inc(&reloc_root->node->refs);
1707                 path->nodes[level] = reloc_root->node;
1708                 path->slots[level] = 0;
1709         } else {
1710                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1711
1712                 level = btrfs_root_drop_level(root_item);
1713                 BUG_ON(level == 0);
1714                 path->lowest_level = level;
1715                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1716                 path->lowest_level = 0;
1717                 if (ret < 0) {
1718                         btrfs_free_path(path);
1719                         return ret;
1720                 }
1721
1722                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1723                                       path->slots[level]);
1724                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1725
1726                 btrfs_unlock_up_safe(path, 0);
1727         }
1728
1729         /*
1730          * In merge_reloc_root(), we modify the upper level pointer to swap the
1731          * tree blocks between reloc tree and subvolume tree.  Thus for tree
1732          * block COW, we COW at most from level 1 to root level for each tree.
1733          *
1734          * Thus the needed metadata size is at most root_level * nodesize,
1735          * and * 2 since we have two trees to COW.
1736          */
1737         reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1738         min_reserved = fs_info->nodesize * reserve_level * 2;
1739         memset(&next_key, 0, sizeof(next_key));
1740
1741         while (1) {
1742                 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
1743                                              min_reserved,
1744                                              BTRFS_RESERVE_FLUSH_LIMIT);
1745                 if (ret)
1746                         goto out;
1747                 trans = btrfs_start_transaction(root, 0);
1748                 if (IS_ERR(trans)) {
1749                         ret = PTR_ERR(trans);
1750                         trans = NULL;
1751                         goto out;
1752                 }
1753
1754                 /*
1755                  * At this point we no longer have a reloc_control, so we can't
1756                  * depend on btrfs_init_reloc_root to update our last_trans.
1757                  *
1758                  * But that's ok, we started the trans handle on our
1759                  * corresponding fs_root, which means it's been added to the
1760                  * dirty list.  At commit time we'll still call
1761                  * btrfs_update_reloc_root() and update our root item
1762                  * appropriately.
1763                  */
1764                 reloc_root->last_trans = trans->transid;
1765                 trans->block_rsv = rc->block_rsv;
1766
1767                 replaced = 0;
1768                 max_level = level;
1769
1770                 ret = walk_down_reloc_tree(reloc_root, path, &level);
1771                 if (ret < 0)
1772                         goto out;
1773                 if (ret > 0)
1774                         break;
1775
1776                 if (!find_next_key(path, level, &key) &&
1777                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1778                         ret = 0;
1779                 } else {
1780                         ret = replace_path(trans, rc, root, reloc_root, path,
1781                                            &next_key, level, max_level);
1782                 }
1783                 if (ret < 0)
1784                         goto out;
1785                 if (ret > 0) {
1786                         level = ret;
1787                         btrfs_node_key_to_cpu(path->nodes[level], &key,
1788                                               path->slots[level]);
1789                         replaced = 1;
1790                 }
1791
1792                 ret = walk_up_reloc_tree(reloc_root, path, &level);
1793                 if (ret > 0)
1794                         break;
1795
1796                 BUG_ON(level == 0);
1797                 /*
1798                  * save the merging progress in the drop_progress.
1799                  * this is OK since root refs == 1 in this case.
1800                  */
1801                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1802                                path->slots[level]);
1803                 btrfs_set_root_drop_level(root_item, level);
1804
1805                 btrfs_end_transaction_throttle(trans);
1806                 trans = NULL;
1807
1808                 btrfs_btree_balance_dirty(fs_info);
1809
1810                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1811                         invalidate_extent_cache(root, &key, &next_key);
1812         }
1813
1814         /*
1815          * handle the case only one block in the fs tree need to be
1816          * relocated and the block is tree root.
1817          */
1818         leaf = btrfs_lock_root_node(root);
1819         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1820                               BTRFS_NESTING_COW);
1821         btrfs_tree_unlock(leaf);
1822         free_extent_buffer(leaf);
1823 out:
1824         btrfs_free_path(path);
1825
1826         if (ret == 0) {
1827                 ret = insert_dirty_subvol(trans, rc, root);
1828                 if (ret)
1829                         btrfs_abort_transaction(trans, ret);
1830         }
1831
1832         if (trans)
1833                 btrfs_end_transaction_throttle(trans);
1834
1835         btrfs_btree_balance_dirty(fs_info);
1836
1837         if (replaced && rc->stage == UPDATE_DATA_PTRS)
1838                 invalidate_extent_cache(root, &key, &next_key);
1839
1840         return ret;
1841 }
1842
1843 static noinline_for_stack
1844 int prepare_to_merge(struct reloc_control *rc, int err)
1845 {
1846         struct btrfs_root *root = rc->extent_root;
1847         struct btrfs_fs_info *fs_info = root->fs_info;
1848         struct btrfs_root *reloc_root;
1849         struct btrfs_trans_handle *trans;
1850         LIST_HEAD(reloc_roots);
1851         u64 num_bytes = 0;
1852         int ret;
1853
1854         mutex_lock(&fs_info->reloc_mutex);
1855         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1856         rc->merging_rsv_size += rc->nodes_relocated * 2;
1857         mutex_unlock(&fs_info->reloc_mutex);
1858
1859 again:
1860         if (!err) {
1861                 num_bytes = rc->merging_rsv_size;
1862                 ret = btrfs_block_rsv_add(fs_info, rc->block_rsv, num_bytes,
1863                                           BTRFS_RESERVE_FLUSH_ALL);
1864                 if (ret)
1865                         err = ret;
1866         }
1867
1868         trans = btrfs_join_transaction(rc->extent_root);
1869         if (IS_ERR(trans)) {
1870                 if (!err)
1871                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1872                                                 num_bytes, NULL);
1873                 return PTR_ERR(trans);
1874         }
1875
1876         if (!err) {
1877                 if (num_bytes != rc->merging_rsv_size) {
1878                         btrfs_end_transaction(trans);
1879                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1880                                                 num_bytes, NULL);
1881                         goto again;
1882                 }
1883         }
1884
1885         rc->merge_reloc_tree = 1;
1886
1887         while (!list_empty(&rc->reloc_roots)) {
1888                 reloc_root = list_entry(rc->reloc_roots.next,
1889                                         struct btrfs_root, root_list);
1890                 list_del_init(&reloc_root->root_list);
1891
1892                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1893                                 false);
1894                 if (IS_ERR(root)) {
1895                         /*
1896                          * Even if we have an error we need this reloc root
1897                          * back on our list so we can clean up properly.
1898                          */
1899                         list_add(&reloc_root->root_list, &reloc_roots);
1900                         btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1901                         if (!err)
1902                                 err = PTR_ERR(root);
1903                         break;
1904                 }
1905
1906                 if (unlikely(root->reloc_root != reloc_root)) {
1907                         if (root->reloc_root) {
1908                                 btrfs_err(fs_info,
1909 "reloc tree mismatch, root %lld has reloc root key (%lld %u %llu) gen %llu, expect reloc root key (%lld %u %llu) gen %llu",
1910                                           root->root_key.objectid,
1911                                           root->reloc_root->root_key.objectid,
1912                                           root->reloc_root->root_key.type,
1913                                           root->reloc_root->root_key.offset,
1914                                           btrfs_root_generation(
1915                                                   &root->reloc_root->root_item),
1916                                           reloc_root->root_key.objectid,
1917                                           reloc_root->root_key.type,
1918                                           reloc_root->root_key.offset,
1919                                           btrfs_root_generation(
1920                                                   &reloc_root->root_item));
1921                         } else {
1922                                 btrfs_err(fs_info,
1923 "reloc tree mismatch, root %lld has no reloc root, expect reloc root key (%lld %u %llu) gen %llu",
1924                                           root->root_key.objectid,
1925                                           reloc_root->root_key.objectid,
1926                                           reloc_root->root_key.type,
1927                                           reloc_root->root_key.offset,
1928                                           btrfs_root_generation(
1929                                                   &reloc_root->root_item));
1930                         }
1931                         list_add(&reloc_root->root_list, &reloc_roots);
1932                         btrfs_put_root(root);
1933                         btrfs_abort_transaction(trans, -EUCLEAN);
1934                         if (!err)
1935                                 err = -EUCLEAN;
1936                         break;
1937                 }
1938
1939                 /*
1940                  * set reference count to 1, so btrfs_recover_relocation
1941                  * knows it should resumes merging
1942                  */
1943                 if (!err)
1944                         btrfs_set_root_refs(&reloc_root->root_item, 1);
1945                 ret = btrfs_update_reloc_root(trans, root);
1946
1947                 /*
1948                  * Even if we have an error we need this reloc root back on our
1949                  * list so we can clean up properly.
1950                  */
1951                 list_add(&reloc_root->root_list, &reloc_roots);
1952                 btrfs_put_root(root);
1953
1954                 if (ret) {
1955                         btrfs_abort_transaction(trans, ret);
1956                         if (!err)
1957                                 err = ret;
1958                         break;
1959                 }
1960         }
1961
1962         list_splice(&reloc_roots, &rc->reloc_roots);
1963
1964         if (!err)
1965                 err = btrfs_commit_transaction(trans);
1966         else
1967                 btrfs_end_transaction(trans);
1968         return err;
1969 }
1970
1971 static noinline_for_stack
1972 void free_reloc_roots(struct list_head *list)
1973 {
1974         struct btrfs_root *reloc_root, *tmp;
1975
1976         list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1977                 __del_reloc_root(reloc_root);
1978 }
1979
1980 static noinline_for_stack
1981 void merge_reloc_roots(struct reloc_control *rc)
1982 {
1983         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1984         struct btrfs_root *root;
1985         struct btrfs_root *reloc_root;
1986         LIST_HEAD(reloc_roots);
1987         int found = 0;
1988         int ret = 0;
1989 again:
1990         root = rc->extent_root;
1991
1992         /*
1993          * this serializes us with btrfs_record_root_in_transaction,
1994          * we have to make sure nobody is in the middle of
1995          * adding their roots to the list while we are
1996          * doing this splice
1997          */
1998         mutex_lock(&fs_info->reloc_mutex);
1999         list_splice_init(&rc->reloc_roots, &reloc_roots);
2000         mutex_unlock(&fs_info->reloc_mutex);
2001
2002         while (!list_empty(&reloc_roots)) {
2003                 found = 1;
2004                 reloc_root = list_entry(reloc_roots.next,
2005                                         struct btrfs_root, root_list);
2006
2007                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
2008                                          false);
2009                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2010                         if (WARN_ON(IS_ERR(root))) {
2011                                 /*
2012                                  * For recovery we read the fs roots on mount,
2013                                  * and if we didn't find the root then we marked
2014                                  * the reloc root as a garbage root.  For normal
2015                                  * relocation obviously the root should exist in
2016                                  * memory.  However there's no reason we can't
2017                                  * handle the error properly here just in case.
2018                                  */
2019                                 ret = PTR_ERR(root);
2020                                 goto out;
2021                         }
2022                         if (WARN_ON(root->reloc_root != reloc_root)) {
2023                                 /*
2024                                  * This can happen if on-disk metadata has some
2025                                  * corruption, e.g. bad reloc tree key offset.
2026                                  */
2027                                 ret = -EINVAL;
2028                                 goto out;
2029                         }
2030                         ret = merge_reloc_root(rc, root);
2031                         btrfs_put_root(root);
2032                         if (ret) {
2033                                 if (list_empty(&reloc_root->root_list))
2034                                         list_add_tail(&reloc_root->root_list,
2035                                                       &reloc_roots);
2036                                 goto out;
2037                         }
2038                 } else {
2039                         if (!IS_ERR(root)) {
2040                                 if (root->reloc_root == reloc_root) {
2041                                         root->reloc_root = NULL;
2042                                         btrfs_put_root(reloc_root);
2043                                 }
2044                                 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
2045                                           &root->state);
2046                                 btrfs_put_root(root);
2047                         }
2048
2049                         list_del_init(&reloc_root->root_list);
2050                         /* Don't forget to queue this reloc root for cleanup */
2051                         list_add_tail(&reloc_root->reloc_dirty_list,
2052                                       &rc->dirty_subvol_roots);
2053                 }
2054         }
2055
2056         if (found) {
2057                 found = 0;
2058                 goto again;
2059         }
2060 out:
2061         if (ret) {
2062                 btrfs_handle_fs_error(fs_info, ret, NULL);
2063                 free_reloc_roots(&reloc_roots);
2064
2065                 /* new reloc root may be added */
2066                 mutex_lock(&fs_info->reloc_mutex);
2067                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2068                 mutex_unlock(&fs_info->reloc_mutex);
2069                 free_reloc_roots(&reloc_roots);
2070         }
2071
2072         /*
2073          * We used to have
2074          *
2075          * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2076          *
2077          * here, but it's wrong.  If we fail to start the transaction in
2078          * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2079          * have actually been removed from the reloc_root_tree rb tree.  This is
2080          * fine because we're bailing here, and we hold a reference on the root
2081          * for the list that holds it, so these roots will be cleaned up when we
2082          * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
2083          * will be cleaned up on unmount.
2084          *
2085          * The remaining nodes will be cleaned up by free_reloc_control.
2086          */
2087 }
2088
2089 static void free_block_list(struct rb_root *blocks)
2090 {
2091         struct tree_block *block;
2092         struct rb_node *rb_node;
2093         while ((rb_node = rb_first(blocks))) {
2094                 block = rb_entry(rb_node, struct tree_block, rb_node);
2095                 rb_erase(rb_node, blocks);
2096                 kfree(block);
2097         }
2098 }
2099
2100 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2101                                       struct btrfs_root *reloc_root)
2102 {
2103         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2104         struct btrfs_root *root;
2105         int ret;
2106
2107         if (reloc_root->last_trans == trans->transid)
2108                 return 0;
2109
2110         root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2111
2112         /*
2113          * This should succeed, since we can't have a reloc root without having
2114          * already looked up the actual root and created the reloc root for this
2115          * root.
2116          *
2117          * However if there's some sort of corruption where we have a ref to a
2118          * reloc root without a corresponding root this could return ENOENT.
2119          */
2120         if (IS_ERR(root)) {
2121                 ASSERT(0);
2122                 return PTR_ERR(root);
2123         }
2124         if (root->reloc_root != reloc_root) {
2125                 ASSERT(0);
2126                 btrfs_err(fs_info,
2127                           "root %llu has two reloc roots associated with it",
2128                           reloc_root->root_key.offset);
2129                 btrfs_put_root(root);
2130                 return -EUCLEAN;
2131         }
2132         ret = btrfs_record_root_in_trans(trans, root);
2133         btrfs_put_root(root);
2134
2135         return ret;
2136 }
2137
2138 static noinline_for_stack
2139 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2140                                      struct reloc_control *rc,
2141                                      struct btrfs_backref_node *node,
2142                                      struct btrfs_backref_edge *edges[])
2143 {
2144         struct btrfs_backref_node *next;
2145         struct btrfs_root *root;
2146         int index = 0;
2147         int ret;
2148
2149         next = node;
2150         while (1) {
2151                 cond_resched();
2152                 next = walk_up_backref(next, edges, &index);
2153                 root = next->root;
2154
2155                 /*
2156                  * If there is no root, then our references for this block are
2157                  * incomplete, as we should be able to walk all the way up to a
2158                  * block that is owned by a root.
2159                  *
2160                  * This path is only for SHAREABLE roots, so if we come upon a
2161                  * non-SHAREABLE root then we have backrefs that resolve
2162                  * improperly.
2163                  *
2164                  * Both of these cases indicate file system corruption, or a bug
2165                  * in the backref walking code.
2166                  */
2167                 if (!root) {
2168                         ASSERT(0);
2169                         btrfs_err(trans->fs_info,
2170                 "bytenr %llu doesn't have a backref path ending in a root",
2171                                   node->bytenr);
2172                         return ERR_PTR(-EUCLEAN);
2173                 }
2174                 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2175                         ASSERT(0);
2176                         btrfs_err(trans->fs_info,
2177         "bytenr %llu has multiple refs with one ending in a non-shareable root",
2178                                   node->bytenr);
2179                         return ERR_PTR(-EUCLEAN);
2180                 }
2181
2182                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2183                         ret = record_reloc_root_in_trans(trans, root);
2184                         if (ret)
2185                                 return ERR_PTR(ret);
2186                         break;
2187                 }
2188
2189                 ret = btrfs_record_root_in_trans(trans, root);
2190                 if (ret)
2191                         return ERR_PTR(ret);
2192                 root = root->reloc_root;
2193
2194                 /*
2195                  * We could have raced with another thread which failed, so
2196                  * root->reloc_root may not be set, return ENOENT in this case.
2197                  */
2198                 if (!root)
2199                         return ERR_PTR(-ENOENT);
2200
2201                 if (next->new_bytenr != root->node->start) {
2202                         /*
2203                          * We just created the reloc root, so we shouldn't have
2204                          * ->new_bytenr set and this shouldn't be in the changed
2205                          *  list.  If it is then we have multiple roots pointing
2206                          *  at the same bytenr which indicates corruption, or
2207                          *  we've made a mistake in the backref walking code.
2208                          */
2209                         ASSERT(next->new_bytenr == 0);
2210                         ASSERT(list_empty(&next->list));
2211                         if (next->new_bytenr || !list_empty(&next->list)) {
2212                                 btrfs_err(trans->fs_info,
2213         "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2214                                           node->bytenr, next->bytenr);
2215                                 return ERR_PTR(-EUCLEAN);
2216                         }
2217
2218                         next->new_bytenr = root->node->start;
2219                         btrfs_put_root(next->root);
2220                         next->root = btrfs_grab_root(root);
2221                         ASSERT(next->root);
2222                         list_add_tail(&next->list,
2223                                       &rc->backref_cache.changed);
2224                         mark_block_processed(rc, next);
2225                         break;
2226                 }
2227
2228                 WARN_ON(1);
2229                 root = NULL;
2230                 next = walk_down_backref(edges, &index);
2231                 if (!next || next->level <= node->level)
2232                         break;
2233         }
2234         if (!root) {
2235                 /*
2236                  * This can happen if there's fs corruption or if there's a bug
2237                  * in the backref lookup code.
2238                  */
2239                 ASSERT(0);
2240                 return ERR_PTR(-ENOENT);
2241         }
2242
2243         next = node;
2244         /* setup backref node path for btrfs_reloc_cow_block */
2245         while (1) {
2246                 rc->backref_cache.path[next->level] = next;
2247                 if (--index < 0)
2248                         break;
2249                 next = edges[index]->node[UPPER];
2250         }
2251         return root;
2252 }
2253
2254 /*
2255  * Select a tree root for relocation.
2256  *
2257  * Return NULL if the block is not shareable. We should use do_relocation() in
2258  * this case.
2259  *
2260  * Return a tree root pointer if the block is shareable.
2261  * Return -ENOENT if the block is root of reloc tree.
2262  */
2263 static noinline_for_stack
2264 struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2265 {
2266         struct btrfs_backref_node *next;
2267         struct btrfs_root *root;
2268         struct btrfs_root *fs_root = NULL;
2269         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2270         int index = 0;
2271
2272         next = node;
2273         while (1) {
2274                 cond_resched();
2275                 next = walk_up_backref(next, edges, &index);
2276                 root = next->root;
2277
2278                 /*
2279                  * This can occur if we have incomplete extent refs leading all
2280                  * the way up a particular path, in this case return -EUCLEAN.
2281                  */
2282                 if (!root)
2283                         return ERR_PTR(-EUCLEAN);
2284
2285                 /* No other choice for non-shareable tree */
2286                 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2287                         return root;
2288
2289                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2290                         fs_root = root;
2291
2292                 if (next != node)
2293                         return NULL;
2294
2295                 next = walk_down_backref(edges, &index);
2296                 if (!next || next->level <= node->level)
2297                         break;
2298         }
2299
2300         if (!fs_root)
2301                 return ERR_PTR(-ENOENT);
2302         return fs_root;
2303 }
2304
2305 static noinline_for_stack
2306 u64 calcu_metadata_size(struct reloc_control *rc,
2307                         struct btrfs_backref_node *node, int reserve)
2308 {
2309         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2310         struct btrfs_backref_node *next = node;
2311         struct btrfs_backref_edge *edge;
2312         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2313         u64 num_bytes = 0;
2314         int index = 0;
2315
2316         BUG_ON(reserve && node->processed);
2317
2318         while (next) {
2319                 cond_resched();
2320                 while (1) {
2321                         if (next->processed && (reserve || next != node))
2322                                 break;
2323
2324                         num_bytes += fs_info->nodesize;
2325
2326                         if (list_empty(&next->upper))
2327                                 break;
2328
2329                         edge = list_entry(next->upper.next,
2330                                         struct btrfs_backref_edge, list[LOWER]);
2331                         edges[index++] = edge;
2332                         next = edge->node[UPPER];
2333                 }
2334                 next = walk_down_backref(edges, &index);
2335         }
2336         return num_bytes;
2337 }
2338
2339 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2340                                   struct reloc_control *rc,
2341                                   struct btrfs_backref_node *node)
2342 {
2343         struct btrfs_root *root = rc->extent_root;
2344         struct btrfs_fs_info *fs_info = root->fs_info;
2345         u64 num_bytes;
2346         int ret;
2347         u64 tmp;
2348
2349         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2350
2351         trans->block_rsv = rc->block_rsv;
2352         rc->reserved_bytes += num_bytes;
2353
2354         /*
2355          * We are under a transaction here so we can only do limited flushing.
2356          * If we get an enospc just kick back -EAGAIN so we know to drop the
2357          * transaction and try to refill when we can flush all the things.
2358          */
2359         ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv, num_bytes,
2360                                      BTRFS_RESERVE_FLUSH_LIMIT);
2361         if (ret) {
2362                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2363                 while (tmp <= rc->reserved_bytes)
2364                         tmp <<= 1;
2365                 /*
2366                  * only one thread can access block_rsv at this point,
2367                  * so we don't need hold lock to protect block_rsv.
2368                  * we expand more reservation size here to allow enough
2369                  * space for relocation and we will return earlier in
2370                  * enospc case.
2371                  */
2372                 rc->block_rsv->size = tmp + fs_info->nodesize *
2373                                       RELOCATION_RESERVED_NODES;
2374                 return -EAGAIN;
2375         }
2376
2377         return 0;
2378 }
2379
2380 /*
2381  * relocate a block tree, and then update pointers in upper level
2382  * blocks that reference the block to point to the new location.
2383  *
2384  * if called by link_to_upper, the block has already been relocated.
2385  * in that case this function just updates pointers.
2386  */
2387 static int do_relocation(struct btrfs_trans_handle *trans,
2388                          struct reloc_control *rc,
2389                          struct btrfs_backref_node *node,
2390                          struct btrfs_key *key,
2391                          struct btrfs_path *path, int lowest)
2392 {
2393         struct btrfs_backref_node *upper;
2394         struct btrfs_backref_edge *edge;
2395         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2396         struct btrfs_root *root;
2397         struct extent_buffer *eb;
2398         u32 blocksize;
2399         u64 bytenr;
2400         int slot;
2401         int ret = 0;
2402
2403         /*
2404          * If we are lowest then this is the first time we're processing this
2405          * block, and thus shouldn't have an eb associated with it yet.
2406          */
2407         ASSERT(!lowest || !node->eb);
2408
2409         path->lowest_level = node->level + 1;
2410         rc->backref_cache.path[node->level] = node;
2411         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2412                 struct btrfs_ref ref = { 0 };
2413
2414                 cond_resched();
2415
2416                 upper = edge->node[UPPER];
2417                 root = select_reloc_root(trans, rc, upper, edges);
2418                 if (IS_ERR(root)) {
2419                         ret = PTR_ERR(root);
2420                         goto next;
2421                 }
2422
2423                 if (upper->eb && !upper->locked) {
2424                         if (!lowest) {
2425                                 ret = btrfs_bin_search(upper->eb, key, &slot);
2426                                 if (ret < 0)
2427                                         goto next;
2428                                 BUG_ON(ret);
2429                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2430                                 if (node->eb->start == bytenr)
2431                                         goto next;
2432                         }
2433                         btrfs_backref_drop_node_buffer(upper);
2434                 }
2435
2436                 if (!upper->eb) {
2437                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2438                         if (ret) {
2439                                 if (ret > 0)
2440                                         ret = -ENOENT;
2441
2442                                 btrfs_release_path(path);
2443                                 break;
2444                         }
2445
2446                         if (!upper->eb) {
2447                                 upper->eb = path->nodes[upper->level];
2448                                 path->nodes[upper->level] = NULL;
2449                         } else {
2450                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2451                         }
2452
2453                         upper->locked = 1;
2454                         path->locks[upper->level] = 0;
2455
2456                         slot = path->slots[upper->level];
2457                         btrfs_release_path(path);
2458                 } else {
2459                         ret = btrfs_bin_search(upper->eb, key, &slot);
2460                         if (ret < 0)
2461                                 goto next;
2462                         BUG_ON(ret);
2463                 }
2464
2465                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2466                 if (lowest) {
2467                         if (bytenr != node->bytenr) {
2468                                 btrfs_err(root->fs_info,
2469                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2470                                           bytenr, node->bytenr, slot,
2471                                           upper->eb->start);
2472                                 ret = -EIO;
2473                                 goto next;
2474                         }
2475                 } else {
2476                         if (node->eb->start == bytenr)
2477                                 goto next;
2478                 }
2479
2480                 blocksize = root->fs_info->nodesize;
2481                 eb = btrfs_read_node_slot(upper->eb, slot);
2482                 if (IS_ERR(eb)) {
2483                         ret = PTR_ERR(eb);
2484                         goto next;
2485                 }
2486                 btrfs_tree_lock(eb);
2487
2488                 if (!node->eb) {
2489                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2490                                               slot, &eb, BTRFS_NESTING_COW);
2491                         btrfs_tree_unlock(eb);
2492                         free_extent_buffer(eb);
2493                         if (ret < 0)
2494                                 goto next;
2495                         /*
2496                          * We've just COWed this block, it should have updated
2497                          * the correct backref node entry.
2498                          */
2499                         ASSERT(node->eb == eb);
2500                 } else {
2501                         btrfs_set_node_blockptr(upper->eb, slot,
2502                                                 node->eb->start);
2503                         btrfs_set_node_ptr_generation(upper->eb, slot,
2504                                                       trans->transid);
2505                         btrfs_mark_buffer_dirty(upper->eb);
2506
2507                         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2508                                                node->eb->start, blocksize,
2509                                                upper->eb->start);
2510                         btrfs_init_tree_ref(&ref, node->level,
2511                                             btrfs_header_owner(upper->eb),
2512                                             root->root_key.objectid, false);
2513                         ret = btrfs_inc_extent_ref(trans, &ref);
2514                         if (!ret)
2515                                 ret = btrfs_drop_subtree(trans, root, eb,
2516                                                          upper->eb);
2517                         if (ret)
2518                                 btrfs_abort_transaction(trans, ret);
2519                 }
2520 next:
2521                 if (!upper->pending)
2522                         btrfs_backref_drop_node_buffer(upper);
2523                 else
2524                         btrfs_backref_unlock_node_buffer(upper);
2525                 if (ret)
2526                         break;
2527         }
2528
2529         if (!ret && node->pending) {
2530                 btrfs_backref_drop_node_buffer(node);
2531                 list_move_tail(&node->list, &rc->backref_cache.changed);
2532                 node->pending = 0;
2533         }
2534
2535         path->lowest_level = 0;
2536
2537         /*
2538          * We should have allocated all of our space in the block rsv and thus
2539          * shouldn't ENOSPC.
2540          */
2541         ASSERT(ret != -ENOSPC);
2542         return ret;
2543 }
2544
2545 static int link_to_upper(struct btrfs_trans_handle *trans,
2546                          struct reloc_control *rc,
2547                          struct btrfs_backref_node *node,
2548                          struct btrfs_path *path)
2549 {
2550         struct btrfs_key key;
2551
2552         btrfs_node_key_to_cpu(node->eb, &key, 0);
2553         return do_relocation(trans, rc, node, &key, path, 0);
2554 }
2555
2556 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2557                                 struct reloc_control *rc,
2558                                 struct btrfs_path *path, int err)
2559 {
2560         LIST_HEAD(list);
2561         struct btrfs_backref_cache *cache = &rc->backref_cache;
2562         struct btrfs_backref_node *node;
2563         int level;
2564         int ret;
2565
2566         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2567                 while (!list_empty(&cache->pending[level])) {
2568                         node = list_entry(cache->pending[level].next,
2569                                           struct btrfs_backref_node, list);
2570                         list_move_tail(&node->list, &list);
2571                         BUG_ON(!node->pending);
2572
2573                         if (!err) {
2574                                 ret = link_to_upper(trans, rc, node, path);
2575                                 if (ret < 0)
2576                                         err = ret;
2577                         }
2578                 }
2579                 list_splice_init(&list, &cache->pending[level]);
2580         }
2581         return err;
2582 }
2583
2584 /*
2585  * mark a block and all blocks directly/indirectly reference the block
2586  * as processed.
2587  */
2588 static void update_processed_blocks(struct reloc_control *rc,
2589                                     struct btrfs_backref_node *node)
2590 {
2591         struct btrfs_backref_node *next = node;
2592         struct btrfs_backref_edge *edge;
2593         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2594         int index = 0;
2595
2596         while (next) {
2597                 cond_resched();
2598                 while (1) {
2599                         if (next->processed)
2600                                 break;
2601
2602                         mark_block_processed(rc, next);
2603
2604                         if (list_empty(&next->upper))
2605                                 break;
2606
2607                         edge = list_entry(next->upper.next,
2608                                         struct btrfs_backref_edge, list[LOWER]);
2609                         edges[index++] = edge;
2610                         next = edge->node[UPPER];
2611                 }
2612                 next = walk_down_backref(edges, &index);
2613         }
2614 }
2615
2616 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2617 {
2618         u32 blocksize = rc->extent_root->fs_info->nodesize;
2619
2620         if (test_range_bit(&rc->processed_blocks, bytenr,
2621                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2622                 return 1;
2623         return 0;
2624 }
2625
2626 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2627                               struct tree_block *block)
2628 {
2629         struct extent_buffer *eb;
2630
2631         eb = read_tree_block(fs_info, block->bytenr, block->owner,
2632                              block->key.offset, block->level, NULL);
2633         if (IS_ERR(eb))
2634                 return PTR_ERR(eb);
2635         if (!extent_buffer_uptodate(eb)) {
2636                 free_extent_buffer(eb);
2637                 return -EIO;
2638         }
2639         if (block->level == 0)
2640                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2641         else
2642                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2643         free_extent_buffer(eb);
2644         block->key_ready = 1;
2645         return 0;
2646 }
2647
2648 /*
2649  * helper function to relocate a tree block
2650  */
2651 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2652                                 struct reloc_control *rc,
2653                                 struct btrfs_backref_node *node,
2654                                 struct btrfs_key *key,
2655                                 struct btrfs_path *path)
2656 {
2657         struct btrfs_root *root;
2658         int ret = 0;
2659
2660         if (!node)
2661                 return 0;
2662
2663         /*
2664          * If we fail here we want to drop our backref_node because we are going
2665          * to start over and regenerate the tree for it.
2666          */
2667         ret = reserve_metadata_space(trans, rc, node);
2668         if (ret)
2669                 goto out;
2670
2671         BUG_ON(node->processed);
2672         root = select_one_root(node);
2673         if (IS_ERR(root)) {
2674                 ret = PTR_ERR(root);
2675
2676                 /* See explanation in select_one_root for the -EUCLEAN case. */
2677                 ASSERT(ret == -ENOENT);
2678                 if (ret == -ENOENT) {
2679                         ret = 0;
2680                         update_processed_blocks(rc, node);
2681                 }
2682                 goto out;
2683         }
2684
2685         if (root) {
2686                 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2687                         /*
2688                          * This block was the root block of a root, and this is
2689                          * the first time we're processing the block and thus it
2690                          * should not have had the ->new_bytenr modified and
2691                          * should have not been included on the changed list.
2692                          *
2693                          * However in the case of corruption we could have
2694                          * multiple refs pointing to the same block improperly,
2695                          * and thus we would trip over these checks.  ASSERT()
2696                          * for the developer case, because it could indicate a
2697                          * bug in the backref code, however error out for a
2698                          * normal user in the case of corruption.
2699                          */
2700                         ASSERT(node->new_bytenr == 0);
2701                         ASSERT(list_empty(&node->list));
2702                         if (node->new_bytenr || !list_empty(&node->list)) {
2703                                 btrfs_err(root->fs_info,
2704                                   "bytenr %llu has improper references to it",
2705                                           node->bytenr);
2706                                 ret = -EUCLEAN;
2707                                 goto out;
2708                         }
2709                         ret = btrfs_record_root_in_trans(trans, root);
2710                         if (ret)
2711                                 goto out;
2712                         /*
2713                          * Another thread could have failed, need to check if we
2714                          * have reloc_root actually set.
2715                          */
2716                         if (!root->reloc_root) {
2717                                 ret = -ENOENT;
2718                                 goto out;
2719                         }
2720                         root = root->reloc_root;
2721                         node->new_bytenr = root->node->start;
2722                         btrfs_put_root(node->root);
2723                         node->root = btrfs_grab_root(root);
2724                         ASSERT(node->root);
2725                         list_add_tail(&node->list, &rc->backref_cache.changed);
2726                 } else {
2727                         path->lowest_level = node->level;
2728                         if (root == root->fs_info->chunk_root)
2729                                 btrfs_reserve_chunk_metadata(trans, false);
2730                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2731                         btrfs_release_path(path);
2732                         if (root == root->fs_info->chunk_root)
2733                                 btrfs_trans_release_chunk_metadata(trans);
2734                         if (ret > 0)
2735                                 ret = 0;
2736                 }
2737                 if (!ret)
2738                         update_processed_blocks(rc, node);
2739         } else {
2740                 ret = do_relocation(trans, rc, node, key, path, 1);
2741         }
2742 out:
2743         if (ret || node->level == 0 || node->cowonly)
2744                 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2745         return ret;
2746 }
2747
2748 /*
2749  * relocate a list of blocks
2750  */
2751 static noinline_for_stack
2752 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2753                          struct reloc_control *rc, struct rb_root *blocks)
2754 {
2755         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2756         struct btrfs_backref_node *node;
2757         struct btrfs_path *path;
2758         struct tree_block *block;
2759         struct tree_block *next;
2760         int ret;
2761         int err = 0;
2762
2763         path = btrfs_alloc_path();
2764         if (!path) {
2765                 err = -ENOMEM;
2766                 goto out_free_blocks;
2767         }
2768
2769         /* Kick in readahead for tree blocks with missing keys */
2770         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2771                 if (!block->key_ready)
2772                         btrfs_readahead_tree_block(fs_info, block->bytenr,
2773                                                    block->owner, 0,
2774                                                    block->level);
2775         }
2776
2777         /* Get first keys */
2778         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2779                 if (!block->key_ready) {
2780                         err = get_tree_block_key(fs_info, block);
2781                         if (err)
2782                                 goto out_free_path;
2783                 }
2784         }
2785
2786         /* Do tree relocation */
2787         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2788                 node = build_backref_tree(rc, &block->key,
2789                                           block->level, block->bytenr);
2790                 if (IS_ERR(node)) {
2791                         err = PTR_ERR(node);
2792                         goto out;
2793                 }
2794
2795                 ret = relocate_tree_block(trans, rc, node, &block->key,
2796                                           path);
2797                 if (ret < 0) {
2798                         err = ret;
2799                         break;
2800                 }
2801         }
2802 out:
2803         err = finish_pending_nodes(trans, rc, path, err);
2804
2805 out_free_path:
2806         btrfs_free_path(path);
2807 out_free_blocks:
2808         free_block_list(blocks);
2809         return err;
2810 }
2811
2812 static noinline_for_stack int prealloc_file_extent_cluster(
2813                                 struct btrfs_inode *inode,
2814                                 struct file_extent_cluster *cluster)
2815 {
2816         u64 alloc_hint = 0;
2817         u64 start;
2818         u64 end;
2819         u64 offset = inode->index_cnt;
2820         u64 num_bytes;
2821         int nr;
2822         int ret = 0;
2823         u64 i_size = i_size_read(&inode->vfs_inode);
2824         u64 prealloc_start = cluster->start - offset;
2825         u64 prealloc_end = cluster->end - offset;
2826         u64 cur_offset = prealloc_start;
2827
2828         /*
2829          * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2830          * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2831          * btrfs_do_readpage() call of previously relocated file cluster.
2832          *
2833          * If the current cluster starts in the above range, btrfs_do_readpage()
2834          * will skip the read, and relocate_one_page() will later writeback
2835          * the padding zeros as new data, causing data corruption.
2836          *
2837          * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2838          */
2839         if (!IS_ALIGNED(i_size, PAGE_SIZE)) {
2840                 struct address_space *mapping = inode->vfs_inode.i_mapping;
2841                 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2842                 const u32 sectorsize = fs_info->sectorsize;
2843                 struct page *page;
2844
2845                 ASSERT(sectorsize < PAGE_SIZE);
2846                 ASSERT(IS_ALIGNED(i_size, sectorsize));
2847
2848                 /*
2849                  * Subpage can't handle page with DIRTY but without UPTODATE
2850                  * bit as it can lead to the following deadlock:
2851                  *
2852                  * btrfs_read_folio()
2853                  * | Page already *locked*
2854                  * |- btrfs_lock_and_flush_ordered_range()
2855                  *    |- btrfs_start_ordered_extent()
2856                  *       |- extent_write_cache_pages()
2857                  *          |- lock_page()
2858                  *             We try to lock the page we already hold.
2859                  *
2860                  * Here we just writeback the whole data reloc inode, so that
2861                  * we will be ensured to have no dirty range in the page, and
2862                  * are safe to clear the uptodate bits.
2863                  *
2864                  * This shouldn't cause too much overhead, as we need to write
2865                  * the data back anyway.
2866                  */
2867                 ret = filemap_write_and_wait(mapping);
2868                 if (ret < 0)
2869                         return ret;
2870
2871                 clear_extent_bits(&inode->io_tree, i_size,
2872                                   round_up(i_size, PAGE_SIZE) - 1,
2873                                   EXTENT_UPTODATE);
2874                 page = find_lock_page(mapping, i_size >> PAGE_SHIFT);
2875                 /*
2876                  * If page is freed we don't need to do anything then, as we
2877                  * will re-read the whole page anyway.
2878                  */
2879                 if (page) {
2880                         btrfs_subpage_clear_uptodate(fs_info, page, i_size,
2881                                         round_up(i_size, PAGE_SIZE) - i_size);
2882                         unlock_page(page);
2883                         put_page(page);
2884                 }
2885         }
2886
2887         BUG_ON(cluster->start != cluster->boundary[0]);
2888         ret = btrfs_alloc_data_chunk_ondemand(inode,
2889                                               prealloc_end + 1 - prealloc_start);
2890         if (ret)
2891                 return ret;
2892
2893         btrfs_inode_lock(&inode->vfs_inode, 0);
2894         for (nr = 0; nr < cluster->nr; nr++) {
2895                 start = cluster->boundary[nr] - offset;
2896                 if (nr + 1 < cluster->nr)
2897                         end = cluster->boundary[nr + 1] - 1 - offset;
2898                 else
2899                         end = cluster->end - offset;
2900
2901                 lock_extent(&inode->io_tree, start, end, NULL);
2902                 num_bytes = end + 1 - start;
2903                 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2904                                                 num_bytes, num_bytes,
2905                                                 end + 1, &alloc_hint);
2906                 cur_offset = end + 1;
2907                 unlock_extent(&inode->io_tree, start, end, NULL);
2908                 if (ret)
2909                         break;
2910         }
2911         btrfs_inode_unlock(&inode->vfs_inode, 0);
2912
2913         if (cur_offset < prealloc_end)
2914                 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2915                                                prealloc_end + 1 - cur_offset);
2916         return ret;
2917 }
2918
2919 static noinline_for_stack int setup_relocation_extent_mapping(struct inode *inode,
2920                                 u64 start, u64 end, u64 block_start)
2921 {
2922         struct extent_map *em;
2923         int ret = 0;
2924
2925         em = alloc_extent_map();
2926         if (!em)
2927                 return -ENOMEM;
2928
2929         em->start = start;
2930         em->len = end + 1 - start;
2931         em->block_len = em->len;
2932         em->block_start = block_start;
2933         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2934
2935         lock_extent(&BTRFS_I(inode)->io_tree, start, end, NULL);
2936         ret = btrfs_replace_extent_map_range(BTRFS_I(inode), em, false);
2937         unlock_extent(&BTRFS_I(inode)->io_tree, start, end, NULL);
2938         free_extent_map(em);
2939
2940         return ret;
2941 }
2942
2943 /*
2944  * Allow error injection to test balance/relocation cancellation
2945  */
2946 noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2947 {
2948         return atomic_read(&fs_info->balance_cancel_req) ||
2949                 atomic_read(&fs_info->reloc_cancel_req) ||
2950                 fatal_signal_pending(current);
2951 }
2952 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2953
2954 static u64 get_cluster_boundary_end(struct file_extent_cluster *cluster,
2955                                     int cluster_nr)
2956 {
2957         /* Last extent, use cluster end directly */
2958         if (cluster_nr >= cluster->nr - 1)
2959                 return cluster->end;
2960
2961         /* Use next boundary start*/
2962         return cluster->boundary[cluster_nr + 1] - 1;
2963 }
2964
2965 static int relocate_one_page(struct inode *inode, struct file_ra_state *ra,
2966                              struct file_extent_cluster *cluster,
2967                              int *cluster_nr, unsigned long page_index)
2968 {
2969         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2970         u64 offset = BTRFS_I(inode)->index_cnt;
2971         const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2972         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2973         struct page *page;
2974         u64 page_start;
2975         u64 page_end;
2976         u64 cur;
2977         int ret;
2978
2979         ASSERT(page_index <= last_index);
2980         page = find_lock_page(inode->i_mapping, page_index);
2981         if (!page) {
2982                 page_cache_sync_readahead(inode->i_mapping, ra, NULL,
2983                                 page_index, last_index + 1 - page_index);
2984                 page = find_or_create_page(inode->i_mapping, page_index, mask);
2985                 if (!page)
2986                         return -ENOMEM;
2987         }
2988         ret = set_page_extent_mapped(page);
2989         if (ret < 0)
2990                 goto release_page;
2991
2992         if (PageReadahead(page))
2993                 page_cache_async_readahead(inode->i_mapping, ra, NULL,
2994                                 page_folio(page), page_index,
2995                                 last_index + 1 - page_index);
2996
2997         if (!PageUptodate(page)) {
2998                 btrfs_read_folio(NULL, page_folio(page));
2999                 lock_page(page);
3000                 if (!PageUptodate(page)) {
3001                         ret = -EIO;
3002                         goto release_page;
3003                 }
3004         }
3005
3006         page_start = page_offset(page);
3007         page_end = page_start + PAGE_SIZE - 1;
3008
3009         /*
3010          * Start from the cluster, as for subpage case, the cluster can start
3011          * inside the page.
3012          */
3013         cur = max(page_start, cluster->boundary[*cluster_nr] - offset);
3014         while (cur <= page_end) {
3015                 u64 extent_start = cluster->boundary[*cluster_nr] - offset;
3016                 u64 extent_end = get_cluster_boundary_end(cluster,
3017                                                 *cluster_nr) - offset;
3018                 u64 clamped_start = max(page_start, extent_start);
3019                 u64 clamped_end = min(page_end, extent_end);
3020                 u32 clamped_len = clamped_end + 1 - clamped_start;
3021
3022                 /* Reserve metadata for this range */
3023                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3024                                                       clamped_len, clamped_len,
3025                                                       false);
3026                 if (ret)
3027                         goto release_page;
3028
3029                 /* Mark the range delalloc and dirty for later writeback */
3030                 lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end, NULL);
3031                 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
3032                                                 clamped_end, 0, NULL);
3033                 if (ret) {
3034                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
3035                                         clamped_start, clamped_end,
3036                                         EXTENT_LOCKED | EXTENT_BOUNDARY);
3037                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3038                                                         clamped_len, true);
3039                         btrfs_delalloc_release_extents(BTRFS_I(inode),
3040                                                        clamped_len);
3041                         goto release_page;
3042                 }
3043                 btrfs_page_set_dirty(fs_info, page, clamped_start, clamped_len);
3044
3045                 /*
3046                  * Set the boundary if it's inside the page.
3047                  * Data relocation requires the destination extents to have the
3048                  * same size as the source.
3049                  * EXTENT_BOUNDARY bit prevents current extent from being merged
3050                  * with previous extent.
3051                  */
3052                 if (in_range(cluster->boundary[*cluster_nr] - offset,
3053                              page_start, PAGE_SIZE)) {
3054                         u64 boundary_start = cluster->boundary[*cluster_nr] -
3055                                                 offset;
3056                         u64 boundary_end = boundary_start +
3057                                            fs_info->sectorsize - 1;
3058
3059                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3060                                         boundary_start, boundary_end,
3061                                         EXTENT_BOUNDARY);
3062                 }
3063                 unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end, NULL);
3064                 btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3065                 cur += clamped_len;
3066
3067                 /* Crossed extent end, go to next extent */
3068                 if (cur >= extent_end) {
3069                         (*cluster_nr)++;
3070                         /* Just finished the last extent of the cluster, exit. */
3071                         if (*cluster_nr >= cluster->nr)
3072                                 break;
3073                 }
3074         }
3075         unlock_page(page);
3076         put_page(page);
3077
3078         balance_dirty_pages_ratelimited(inode->i_mapping);
3079         btrfs_throttle(fs_info);
3080         if (btrfs_should_cancel_balance(fs_info))
3081                 ret = -ECANCELED;
3082         return ret;
3083
3084 release_page:
3085         unlock_page(page);
3086         put_page(page);
3087         return ret;
3088 }
3089
3090 static int relocate_file_extent_cluster(struct inode *inode,
3091                                         struct file_extent_cluster *cluster)
3092 {
3093         u64 offset = BTRFS_I(inode)->index_cnt;
3094         unsigned long index;
3095         unsigned long last_index;
3096         struct file_ra_state *ra;
3097         int cluster_nr = 0;
3098         int ret = 0;
3099
3100         if (!cluster->nr)
3101                 return 0;
3102
3103         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3104         if (!ra)
3105                 return -ENOMEM;
3106
3107         ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
3108         if (ret)
3109                 goto out;
3110
3111         file_ra_state_init(ra, inode->i_mapping);
3112
3113         ret = setup_relocation_extent_mapping(inode, cluster->start - offset,
3114                                    cluster->end - offset, cluster->start);
3115         if (ret)
3116                 goto out;
3117
3118         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3119         for (index = (cluster->start - offset) >> PAGE_SHIFT;
3120              index <= last_index && !ret; index++)
3121                 ret = relocate_one_page(inode, ra, cluster, &cluster_nr, index);
3122         if (ret == 0)
3123                 WARN_ON(cluster_nr != cluster->nr);
3124 out:
3125         kfree(ra);
3126         return ret;
3127 }
3128
3129 static noinline_for_stack
3130 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3131                          struct file_extent_cluster *cluster)
3132 {
3133         int ret;
3134
3135         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3136                 ret = relocate_file_extent_cluster(inode, cluster);
3137                 if (ret)
3138                         return ret;
3139                 cluster->nr = 0;
3140         }
3141
3142         if (!cluster->nr)
3143                 cluster->start = extent_key->objectid;
3144         else
3145                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3146         cluster->end = extent_key->objectid + extent_key->offset - 1;
3147         cluster->boundary[cluster->nr] = extent_key->objectid;
3148         cluster->nr++;
3149
3150         if (cluster->nr >= MAX_EXTENTS) {
3151                 ret = relocate_file_extent_cluster(inode, cluster);
3152                 if (ret)
3153                         return ret;
3154                 cluster->nr = 0;
3155         }
3156         return 0;
3157 }
3158
3159 /*
3160  * helper to add a tree block to the list.
3161  * the major work is getting the generation and level of the block
3162  */
3163 static int add_tree_block(struct reloc_control *rc,
3164                           struct btrfs_key *extent_key,
3165                           struct btrfs_path *path,
3166                           struct rb_root *blocks)
3167 {
3168         struct extent_buffer *eb;
3169         struct btrfs_extent_item *ei;
3170         struct btrfs_tree_block_info *bi;
3171         struct tree_block *block;
3172         struct rb_node *rb_node;
3173         u32 item_size;
3174         int level = -1;
3175         u64 generation;
3176         u64 owner = 0;
3177
3178         eb =  path->nodes[0];
3179         item_size = btrfs_item_size(eb, path->slots[0]);
3180
3181         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3182             item_size >= sizeof(*ei) + sizeof(*bi)) {
3183                 unsigned long ptr = 0, end;
3184
3185                 ei = btrfs_item_ptr(eb, path->slots[0],
3186                                 struct btrfs_extent_item);
3187                 end = (unsigned long)ei + item_size;
3188                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3189                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3190                         level = btrfs_tree_block_level(eb, bi);
3191                         ptr = (unsigned long)(bi + 1);
3192                 } else {
3193                         level = (int)extent_key->offset;
3194                         ptr = (unsigned long)(ei + 1);
3195                 }
3196                 generation = btrfs_extent_generation(eb, ei);
3197
3198                 /*
3199                  * We're reading random blocks without knowing their owner ahead
3200                  * of time.  This is ok most of the time, as all reloc roots and
3201                  * fs roots have the same lock type.  However normal trees do
3202                  * not, and the only way to know ahead of time is to read the
3203                  * inline ref offset.  We know it's an fs root if
3204                  *
3205                  * 1. There's more than one ref.
3206                  * 2. There's a SHARED_DATA_REF_KEY set.
3207                  * 3. FULL_BACKREF is set on the flags.
3208                  *
3209                  * Otherwise it's safe to assume that the ref offset == the
3210                  * owner of this block, so we can use that when calling
3211                  * read_tree_block.
3212                  */
3213                 if (btrfs_extent_refs(eb, ei) == 1 &&
3214                     !(btrfs_extent_flags(eb, ei) &
3215                       BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3216                     ptr < end) {
3217                         struct btrfs_extent_inline_ref *iref;
3218                         int type;
3219
3220                         iref = (struct btrfs_extent_inline_ref *)ptr;
3221                         type = btrfs_get_extent_inline_ref_type(eb, iref,
3222                                                         BTRFS_REF_TYPE_BLOCK);
3223                         if (type == BTRFS_REF_TYPE_INVALID)
3224                                 return -EINVAL;
3225                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
3226                                 owner = btrfs_extent_inline_ref_offset(eb, iref);
3227                 }
3228         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3229                 btrfs_print_v0_err(eb->fs_info);
3230                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3231                 return -EINVAL;
3232         } else {
3233                 BUG();
3234         }
3235
3236         btrfs_release_path(path);
3237
3238         BUG_ON(level == -1);
3239
3240         block = kmalloc(sizeof(*block), GFP_NOFS);
3241         if (!block)
3242                 return -ENOMEM;
3243
3244         block->bytenr = extent_key->objectid;
3245         block->key.objectid = rc->extent_root->fs_info->nodesize;
3246         block->key.offset = generation;
3247         block->level = level;
3248         block->key_ready = 0;
3249         block->owner = owner;
3250
3251         rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3252         if (rb_node)
3253                 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3254                                     -EEXIST);
3255
3256         return 0;
3257 }
3258
3259 /*
3260  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3261  */
3262 static int __add_tree_block(struct reloc_control *rc,
3263                             u64 bytenr, u32 blocksize,
3264                             struct rb_root *blocks)
3265 {
3266         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3267         struct btrfs_path *path;
3268         struct btrfs_key key;
3269         int ret;
3270         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3271
3272         if (tree_block_processed(bytenr, rc))
3273                 return 0;
3274
3275         if (rb_simple_search(blocks, bytenr))
3276                 return 0;
3277
3278         path = btrfs_alloc_path();
3279         if (!path)
3280                 return -ENOMEM;
3281 again:
3282         key.objectid = bytenr;
3283         if (skinny) {
3284                 key.type = BTRFS_METADATA_ITEM_KEY;
3285                 key.offset = (u64)-1;
3286         } else {
3287                 key.type = BTRFS_EXTENT_ITEM_KEY;
3288                 key.offset = blocksize;
3289         }
3290
3291         path->search_commit_root = 1;
3292         path->skip_locking = 1;
3293         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3294         if (ret < 0)
3295                 goto out;
3296
3297         if (ret > 0 && skinny) {
3298                 if (path->slots[0]) {
3299                         path->slots[0]--;
3300                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3301                                               path->slots[0]);
3302                         if (key.objectid == bytenr &&
3303                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3304                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3305                               key.offset == blocksize)))
3306                                 ret = 0;
3307                 }
3308
3309                 if (ret) {
3310                         skinny = false;
3311                         btrfs_release_path(path);
3312                         goto again;
3313                 }
3314         }
3315         if (ret) {
3316                 ASSERT(ret == 1);
3317                 btrfs_print_leaf(path->nodes[0]);
3318                 btrfs_err(fs_info,
3319              "tree block extent item (%llu) is not found in extent tree",
3320                      bytenr);
3321                 WARN_ON(1);
3322                 ret = -EINVAL;
3323                 goto out;
3324         }
3325
3326         ret = add_tree_block(rc, &key, path, blocks);
3327 out:
3328         btrfs_free_path(path);
3329         return ret;
3330 }
3331
3332 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3333                                     struct btrfs_block_group *block_group,
3334                                     struct inode *inode,
3335                                     u64 ino)
3336 {
3337         struct btrfs_root *root = fs_info->tree_root;
3338         struct btrfs_trans_handle *trans;
3339         int ret = 0;
3340
3341         if (inode)
3342                 goto truncate;
3343
3344         inode = btrfs_iget(fs_info->sb, ino, root);
3345         if (IS_ERR(inode))
3346                 return -ENOENT;
3347
3348 truncate:
3349         ret = btrfs_check_trunc_cache_free_space(fs_info,
3350                                                  &fs_info->global_block_rsv);
3351         if (ret)
3352                 goto out;
3353
3354         trans = btrfs_join_transaction(root);
3355         if (IS_ERR(trans)) {
3356                 ret = PTR_ERR(trans);
3357                 goto out;
3358         }
3359
3360         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3361
3362         btrfs_end_transaction(trans);
3363         btrfs_btree_balance_dirty(fs_info);
3364 out:
3365         iput(inode);
3366         return ret;
3367 }
3368
3369 /*
3370  * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3371  * cache inode, to avoid free space cache data extent blocking data relocation.
3372  */
3373 static int delete_v1_space_cache(struct extent_buffer *leaf,
3374                                  struct btrfs_block_group *block_group,
3375                                  u64 data_bytenr)
3376 {
3377         u64 space_cache_ino;
3378         struct btrfs_file_extent_item *ei;
3379         struct btrfs_key key;
3380         bool found = false;
3381         int i;
3382         int ret;
3383
3384         if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3385                 return 0;
3386
3387         for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3388                 u8 type;
3389
3390                 btrfs_item_key_to_cpu(leaf, &key, i);
3391                 if (key.type != BTRFS_EXTENT_DATA_KEY)
3392                         continue;
3393                 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3394                 type = btrfs_file_extent_type(leaf, ei);
3395
3396                 if ((type == BTRFS_FILE_EXTENT_REG ||
3397                      type == BTRFS_FILE_EXTENT_PREALLOC) &&
3398                     btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3399                         found = true;
3400                         space_cache_ino = key.objectid;
3401                         break;
3402                 }
3403         }
3404         if (!found)
3405                 return -ENOENT;
3406         ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3407                                         space_cache_ino);
3408         return ret;
3409 }
3410
3411 /*
3412  * helper to find all tree blocks that reference a given data extent
3413  */
3414 static noinline_for_stack
3415 int add_data_references(struct reloc_control *rc,
3416                         struct btrfs_key *extent_key,
3417                         struct btrfs_path *path,
3418                         struct rb_root *blocks)
3419 {
3420         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3421         struct ulist *leaves = NULL;
3422         struct ulist_iterator leaf_uiter;
3423         struct ulist_node *ref_node = NULL;
3424         const u32 blocksize = fs_info->nodesize;
3425         int ret = 0;
3426
3427         btrfs_release_path(path);
3428         ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid,
3429                                    0, &leaves, NULL, true);
3430         if (ret < 0)
3431                 return ret;
3432
3433         ULIST_ITER_INIT(&leaf_uiter);
3434         while ((ref_node = ulist_next(leaves, &leaf_uiter))) {
3435                 struct extent_buffer *eb;
3436
3437                 eb = read_tree_block(fs_info, ref_node->val, 0, 0, 0, NULL);
3438                 if (IS_ERR(eb)) {
3439                         ret = PTR_ERR(eb);
3440                         break;
3441                 }
3442                 ret = delete_v1_space_cache(eb, rc->block_group,
3443                                             extent_key->objectid);
3444                 free_extent_buffer(eb);
3445                 if (ret < 0)
3446                         break;
3447                 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3448                 if (ret < 0)
3449                         break;
3450         }
3451         if (ret < 0)
3452                 free_block_list(blocks);
3453         ulist_free(leaves);
3454         return ret;
3455 }
3456
3457 /*
3458  * helper to find next unprocessed extent
3459  */
3460 static noinline_for_stack
3461 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3462                      struct btrfs_key *extent_key)
3463 {
3464         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3465         struct btrfs_key key;
3466         struct extent_buffer *leaf;
3467         u64 start, end, last;
3468         int ret;
3469
3470         last = rc->block_group->start + rc->block_group->length;
3471         while (1) {
3472                 cond_resched();
3473                 if (rc->search_start >= last) {
3474                         ret = 1;
3475                         break;
3476                 }
3477
3478                 key.objectid = rc->search_start;
3479                 key.type = BTRFS_EXTENT_ITEM_KEY;
3480                 key.offset = 0;
3481
3482                 path->search_commit_root = 1;
3483                 path->skip_locking = 1;
3484                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3485                                         0, 0);
3486                 if (ret < 0)
3487                         break;
3488 next:
3489                 leaf = path->nodes[0];
3490                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3491                         ret = btrfs_next_leaf(rc->extent_root, path);
3492                         if (ret != 0)
3493                                 break;
3494                         leaf = path->nodes[0];
3495                 }
3496
3497                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3498                 if (key.objectid >= last) {
3499                         ret = 1;
3500                         break;
3501                 }
3502
3503                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3504                     key.type != BTRFS_METADATA_ITEM_KEY) {
3505                         path->slots[0]++;
3506                         goto next;
3507                 }
3508
3509                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3510                     key.objectid + key.offset <= rc->search_start) {
3511                         path->slots[0]++;
3512                         goto next;
3513                 }
3514
3515                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3516                     key.objectid + fs_info->nodesize <=
3517                     rc->search_start) {
3518                         path->slots[0]++;
3519                         goto next;
3520                 }
3521
3522                 ret = find_first_extent_bit(&rc->processed_blocks,
3523                                             key.objectid, &start, &end,
3524                                             EXTENT_DIRTY, NULL);
3525
3526                 if (ret == 0 && start <= key.objectid) {
3527                         btrfs_release_path(path);
3528                         rc->search_start = end + 1;
3529                 } else {
3530                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3531                                 rc->search_start = key.objectid + key.offset;
3532                         else
3533                                 rc->search_start = key.objectid +
3534                                         fs_info->nodesize;
3535                         memcpy(extent_key, &key, sizeof(key));
3536                         return 0;
3537                 }
3538         }
3539         btrfs_release_path(path);
3540         return ret;
3541 }
3542
3543 static void set_reloc_control(struct reloc_control *rc)
3544 {
3545         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3546
3547         mutex_lock(&fs_info->reloc_mutex);
3548         fs_info->reloc_ctl = rc;
3549         mutex_unlock(&fs_info->reloc_mutex);
3550 }
3551
3552 static void unset_reloc_control(struct reloc_control *rc)
3553 {
3554         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3555
3556         mutex_lock(&fs_info->reloc_mutex);
3557         fs_info->reloc_ctl = NULL;
3558         mutex_unlock(&fs_info->reloc_mutex);
3559 }
3560
3561 static noinline_for_stack
3562 int prepare_to_relocate(struct reloc_control *rc)
3563 {
3564         struct btrfs_trans_handle *trans;
3565         int ret;
3566
3567         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3568                                               BTRFS_BLOCK_RSV_TEMP);
3569         if (!rc->block_rsv)
3570                 return -ENOMEM;
3571
3572         memset(&rc->cluster, 0, sizeof(rc->cluster));
3573         rc->search_start = rc->block_group->start;
3574         rc->extents_found = 0;
3575         rc->nodes_relocated = 0;
3576         rc->merging_rsv_size = 0;
3577         rc->reserved_bytes = 0;
3578         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3579                               RELOCATION_RESERVED_NODES;
3580         ret = btrfs_block_rsv_refill(rc->extent_root->fs_info,
3581                                      rc->block_rsv, rc->block_rsv->size,
3582                                      BTRFS_RESERVE_FLUSH_ALL);
3583         if (ret)
3584                 return ret;
3585
3586         rc->create_reloc_tree = 1;
3587         set_reloc_control(rc);
3588
3589         trans = btrfs_join_transaction(rc->extent_root);
3590         if (IS_ERR(trans)) {
3591                 unset_reloc_control(rc);
3592                 /*
3593                  * extent tree is not a ref_cow tree and has no reloc_root to
3594                  * cleanup.  And callers are responsible to free the above
3595                  * block rsv.
3596                  */
3597                 return PTR_ERR(trans);
3598         }
3599
3600         ret = btrfs_commit_transaction(trans);
3601         if (ret)
3602                 unset_reloc_control(rc);
3603
3604         return ret;
3605 }
3606
3607 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3608 {
3609         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3610         struct rb_root blocks = RB_ROOT;
3611         struct btrfs_key key;
3612         struct btrfs_trans_handle *trans = NULL;
3613         struct btrfs_path *path;
3614         struct btrfs_extent_item *ei;
3615         u64 flags;
3616         int ret;
3617         int err = 0;
3618         int progress = 0;
3619
3620         path = btrfs_alloc_path();
3621         if (!path)
3622                 return -ENOMEM;
3623         path->reada = READA_FORWARD;
3624
3625         ret = prepare_to_relocate(rc);
3626         if (ret) {
3627                 err = ret;
3628                 goto out_free;
3629         }
3630
3631         while (1) {
3632                 rc->reserved_bytes = 0;
3633                 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
3634                                              rc->block_rsv->size,
3635                                              BTRFS_RESERVE_FLUSH_ALL);
3636                 if (ret) {
3637                         err = ret;
3638                         break;
3639                 }
3640                 progress++;
3641                 trans = btrfs_start_transaction(rc->extent_root, 0);
3642                 if (IS_ERR(trans)) {
3643                         err = PTR_ERR(trans);
3644                         trans = NULL;
3645                         break;
3646                 }
3647 restart:
3648                 if (update_backref_cache(trans, &rc->backref_cache)) {
3649                         btrfs_end_transaction(trans);
3650                         trans = NULL;
3651                         continue;
3652                 }
3653
3654                 ret = find_next_extent(rc, path, &key);
3655                 if (ret < 0)
3656                         err = ret;
3657                 if (ret != 0)
3658                         break;
3659
3660                 rc->extents_found++;
3661
3662                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3663                                     struct btrfs_extent_item);
3664                 flags = btrfs_extent_flags(path->nodes[0], ei);
3665
3666                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3667                         ret = add_tree_block(rc, &key, path, &blocks);
3668                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3669                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3670                         ret = add_data_references(rc, &key, path, &blocks);
3671                 } else {
3672                         btrfs_release_path(path);
3673                         ret = 0;
3674                 }
3675                 if (ret < 0) {
3676                         err = ret;
3677                         break;
3678                 }
3679
3680                 if (!RB_EMPTY_ROOT(&blocks)) {
3681                         ret = relocate_tree_blocks(trans, rc, &blocks);
3682                         if (ret < 0) {
3683                                 if (ret != -EAGAIN) {
3684                                         err = ret;
3685                                         break;
3686                                 }
3687                                 rc->extents_found--;
3688                                 rc->search_start = key.objectid;
3689                         }
3690                 }
3691
3692                 btrfs_end_transaction_throttle(trans);
3693                 btrfs_btree_balance_dirty(fs_info);
3694                 trans = NULL;
3695
3696                 if (rc->stage == MOVE_DATA_EXTENTS &&
3697                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3698                         rc->found_file_extent = 1;
3699                         ret = relocate_data_extent(rc->data_inode,
3700                                                    &key, &rc->cluster);
3701                         if (ret < 0) {
3702                                 err = ret;
3703                                 break;
3704                         }
3705                 }
3706                 if (btrfs_should_cancel_balance(fs_info)) {
3707                         err = -ECANCELED;
3708                         break;
3709                 }
3710         }
3711         if (trans && progress && err == -ENOSPC) {
3712                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3713                 if (ret == 1) {
3714                         err = 0;
3715                         progress = 0;
3716                         goto restart;
3717                 }
3718         }
3719
3720         btrfs_release_path(path);
3721         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3722
3723         if (trans) {
3724                 btrfs_end_transaction_throttle(trans);
3725                 btrfs_btree_balance_dirty(fs_info);
3726         }
3727
3728         if (!err) {
3729                 ret = relocate_file_extent_cluster(rc->data_inode,
3730                                                    &rc->cluster);
3731                 if (ret < 0)
3732                         err = ret;
3733         }
3734
3735         rc->create_reloc_tree = 0;
3736         set_reloc_control(rc);
3737
3738         btrfs_backref_release_cache(&rc->backref_cache);
3739         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3740
3741         /*
3742          * Even in the case when the relocation is cancelled, we should all go
3743          * through prepare_to_merge() and merge_reloc_roots().
3744          *
3745          * For error (including cancelled balance), prepare_to_merge() will
3746          * mark all reloc trees orphan, then queue them for cleanup in
3747          * merge_reloc_roots()
3748          */
3749         err = prepare_to_merge(rc, err);
3750
3751         merge_reloc_roots(rc);
3752
3753         rc->merge_reloc_tree = 0;
3754         unset_reloc_control(rc);
3755         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3756
3757         /* get rid of pinned extents */
3758         trans = btrfs_join_transaction(rc->extent_root);
3759         if (IS_ERR(trans)) {
3760                 err = PTR_ERR(trans);
3761                 goto out_free;
3762         }
3763         ret = btrfs_commit_transaction(trans);
3764         if (ret && !err)
3765                 err = ret;
3766 out_free:
3767         ret = clean_dirty_subvols(rc);
3768         if (ret < 0 && !err)
3769                 err = ret;
3770         btrfs_free_block_rsv(fs_info, rc->block_rsv);
3771         btrfs_free_path(path);
3772         return err;
3773 }
3774
3775 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3776                                  struct btrfs_root *root, u64 objectid)
3777 {
3778         struct btrfs_path *path;
3779         struct btrfs_inode_item *item;
3780         struct extent_buffer *leaf;
3781         int ret;
3782
3783         path = btrfs_alloc_path();
3784         if (!path)
3785                 return -ENOMEM;
3786
3787         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3788         if (ret)
3789                 goto out;
3790
3791         leaf = path->nodes[0];
3792         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3793         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3794         btrfs_set_inode_generation(leaf, item, 1);
3795         btrfs_set_inode_size(leaf, item, 0);
3796         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3797         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3798                                           BTRFS_INODE_PREALLOC);
3799         btrfs_mark_buffer_dirty(leaf);
3800 out:
3801         btrfs_free_path(path);
3802         return ret;
3803 }
3804
3805 static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3806                                 struct btrfs_root *root, u64 objectid)
3807 {
3808         struct btrfs_path *path;
3809         struct btrfs_key key;
3810         int ret = 0;
3811
3812         path = btrfs_alloc_path();
3813         if (!path) {
3814                 ret = -ENOMEM;
3815                 goto out;
3816         }
3817
3818         key.objectid = objectid;
3819         key.type = BTRFS_INODE_ITEM_KEY;
3820         key.offset = 0;
3821         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3822         if (ret) {
3823                 if (ret > 0)
3824                         ret = -ENOENT;
3825                 goto out;
3826         }
3827         ret = btrfs_del_item(trans, root, path);
3828 out:
3829         if (ret)
3830                 btrfs_abort_transaction(trans, ret);
3831         btrfs_free_path(path);
3832 }
3833
3834 /*
3835  * helper to create inode for data relocation.
3836  * the inode is in data relocation tree and its link count is 0
3837  */
3838 static noinline_for_stack
3839 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3840                                  struct btrfs_block_group *group)
3841 {
3842         struct inode *inode = NULL;
3843         struct btrfs_trans_handle *trans;
3844         struct btrfs_root *root;
3845         u64 objectid;
3846         int err = 0;
3847
3848         root = btrfs_grab_root(fs_info->data_reloc_root);
3849         trans = btrfs_start_transaction(root, 6);
3850         if (IS_ERR(trans)) {
3851                 btrfs_put_root(root);
3852                 return ERR_CAST(trans);
3853         }
3854
3855         err = btrfs_get_free_objectid(root, &objectid);
3856         if (err)
3857                 goto out;
3858
3859         err = __insert_orphan_inode(trans, root, objectid);
3860         if (err)
3861                 goto out;
3862
3863         inode = btrfs_iget(fs_info->sb, objectid, root);
3864         if (IS_ERR(inode)) {
3865                 delete_orphan_inode(trans, root, objectid);
3866                 err = PTR_ERR(inode);
3867                 inode = NULL;
3868                 goto out;
3869         }
3870         BTRFS_I(inode)->index_cnt = group->start;
3871
3872         err = btrfs_orphan_add(trans, BTRFS_I(inode));
3873 out:
3874         btrfs_put_root(root);
3875         btrfs_end_transaction(trans);
3876         btrfs_btree_balance_dirty(fs_info);
3877         if (err) {
3878                 iput(inode);
3879                 inode = ERR_PTR(err);
3880         }
3881         return inode;
3882 }
3883
3884 /*
3885  * Mark start of chunk relocation that is cancellable. Check if the cancellation
3886  * has been requested meanwhile and don't start in that case.
3887  *
3888  * Return:
3889  *   0             success
3890  *   -EINPROGRESS  operation is already in progress, that's probably a bug
3891  *   -ECANCELED    cancellation request was set before the operation started
3892  */
3893 static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3894 {
3895         if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3896                 /* This should not happen */
3897                 btrfs_err(fs_info, "reloc already running, cannot start");
3898                 return -EINPROGRESS;
3899         }
3900
3901         if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3902                 btrfs_info(fs_info, "chunk relocation canceled on start");
3903                 /*
3904                  * On cancel, clear all requests but let the caller mark
3905                  * the end after cleanup operations.
3906                  */
3907                 atomic_set(&fs_info->reloc_cancel_req, 0);
3908                 return -ECANCELED;
3909         }
3910         return 0;
3911 }
3912
3913 /*
3914  * Mark end of chunk relocation that is cancellable and wake any waiters.
3915  */
3916 static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
3917 {
3918         /* Requested after start, clear bit first so any waiters can continue */
3919         if (atomic_read(&fs_info->reloc_cancel_req) > 0)
3920                 btrfs_info(fs_info, "chunk relocation canceled during operation");
3921         clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
3922         atomic_set(&fs_info->reloc_cancel_req, 0);
3923 }
3924
3925 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3926 {
3927         struct reloc_control *rc;
3928
3929         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3930         if (!rc)
3931                 return NULL;
3932
3933         INIT_LIST_HEAD(&rc->reloc_roots);
3934         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3935         btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3936         mapping_tree_init(&rc->reloc_root_tree);
3937         extent_io_tree_init(fs_info, &rc->processed_blocks,
3938                             IO_TREE_RELOC_BLOCKS, NULL);
3939         return rc;
3940 }
3941
3942 static void free_reloc_control(struct reloc_control *rc)
3943 {
3944         struct mapping_node *node, *tmp;
3945
3946         free_reloc_roots(&rc->reloc_roots);
3947         rbtree_postorder_for_each_entry_safe(node, tmp,
3948                         &rc->reloc_root_tree.rb_root, rb_node)
3949                 kfree(node);
3950
3951         kfree(rc);
3952 }
3953
3954 /*
3955  * Print the block group being relocated
3956  */
3957 static void describe_relocation(struct btrfs_fs_info *fs_info,
3958                                 struct btrfs_block_group *block_group)
3959 {
3960         char buf[128] = {'\0'};
3961
3962         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3963
3964         btrfs_info(fs_info,
3965                    "relocating block group %llu flags %s",
3966                    block_group->start, buf);
3967 }
3968
3969 static const char *stage_to_string(int stage)
3970 {
3971         if (stage == MOVE_DATA_EXTENTS)
3972                 return "move data extents";
3973         if (stage == UPDATE_DATA_PTRS)
3974                 return "update data pointers";
3975         return "unknown";
3976 }
3977
3978 /*
3979  * function to relocate all extents in a block group.
3980  */
3981 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3982 {
3983         struct btrfs_block_group *bg;
3984         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, group_start);
3985         struct reloc_control *rc;
3986         struct inode *inode;
3987         struct btrfs_path *path;
3988         int ret;
3989         int rw = 0;
3990         int err = 0;
3991
3992         /*
3993          * This only gets set if we had a half-deleted snapshot on mount.  We
3994          * cannot allow relocation to start while we're still trying to clean up
3995          * these pending deletions.
3996          */
3997         ret = wait_on_bit(&fs_info->flags, BTRFS_FS_UNFINISHED_DROPS, TASK_INTERRUPTIBLE);
3998         if (ret)
3999                 return ret;
4000
4001         /* We may have been woken up by close_ctree, so bail if we're closing. */
4002         if (btrfs_fs_closing(fs_info))
4003                 return -EINTR;
4004
4005         bg = btrfs_lookup_block_group(fs_info, group_start);
4006         if (!bg)
4007                 return -ENOENT;
4008
4009         /*
4010          * Relocation of a data block group creates ordered extents.  Without
4011          * sb_start_write(), we can freeze the filesystem while unfinished
4012          * ordered extents are left. Such ordered extents can cause a deadlock
4013          * e.g. when syncfs() is waiting for their completion but they can't
4014          * finish because they block when joining a transaction, due to the
4015          * fact that the freeze locks are being held in write mode.
4016          */
4017         if (bg->flags & BTRFS_BLOCK_GROUP_DATA)
4018                 ASSERT(sb_write_started(fs_info->sb));
4019
4020         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4021                 btrfs_put_block_group(bg);
4022                 return -ETXTBSY;
4023         }
4024
4025         rc = alloc_reloc_control(fs_info);
4026         if (!rc) {
4027                 btrfs_put_block_group(bg);
4028                 return -ENOMEM;
4029         }
4030
4031         ret = reloc_chunk_start(fs_info);
4032         if (ret < 0) {
4033                 err = ret;
4034                 goto out_put_bg;
4035         }
4036
4037         rc->extent_root = extent_root;
4038         rc->block_group = bg;
4039
4040         ret = btrfs_inc_block_group_ro(rc->block_group, true);
4041         if (ret) {
4042                 err = ret;
4043                 goto out;
4044         }
4045         rw = 1;
4046
4047         path = btrfs_alloc_path();
4048         if (!path) {
4049                 err = -ENOMEM;
4050                 goto out;
4051         }
4052
4053         inode = lookup_free_space_inode(rc->block_group, path);
4054         btrfs_free_path(path);
4055
4056         if (!IS_ERR(inode))
4057                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4058         else
4059                 ret = PTR_ERR(inode);
4060
4061         if (ret && ret != -ENOENT) {
4062                 err = ret;
4063                 goto out;
4064         }
4065
4066         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4067         if (IS_ERR(rc->data_inode)) {
4068                 err = PTR_ERR(rc->data_inode);
4069                 rc->data_inode = NULL;
4070                 goto out;
4071         }
4072
4073         describe_relocation(fs_info, rc->block_group);
4074
4075         btrfs_wait_block_group_reservations(rc->block_group);
4076         btrfs_wait_nocow_writers(rc->block_group);
4077         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4078                                  rc->block_group->start,
4079                                  rc->block_group->length);
4080
4081         ret = btrfs_zone_finish(rc->block_group);
4082         WARN_ON(ret && ret != -EAGAIN);
4083
4084         while (1) {
4085                 int finishes_stage;
4086
4087                 mutex_lock(&fs_info->cleaner_mutex);
4088                 ret = relocate_block_group(rc);
4089                 mutex_unlock(&fs_info->cleaner_mutex);
4090                 if (ret < 0)
4091                         err = ret;
4092
4093                 finishes_stage = rc->stage;
4094                 /*
4095                  * We may have gotten ENOSPC after we already dirtied some
4096                  * extents.  If writeout happens while we're relocating a
4097                  * different block group we could end up hitting the
4098                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4099                  * btrfs_reloc_cow_block.  Make sure we write everything out
4100                  * properly so we don't trip over this problem, and then break
4101                  * out of the loop if we hit an error.
4102                  */
4103                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4104                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4105                                                        (u64)-1);
4106                         if (ret)
4107                                 err = ret;
4108                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4109                                                  0, -1);
4110                         rc->stage = UPDATE_DATA_PTRS;
4111                 }
4112
4113                 if (err < 0)
4114                         goto out;
4115
4116                 if (rc->extents_found == 0)
4117                         break;
4118
4119                 btrfs_info(fs_info, "found %llu extents, stage: %s",
4120                            rc->extents_found, stage_to_string(finishes_stage));
4121         }
4122
4123         WARN_ON(rc->block_group->pinned > 0);
4124         WARN_ON(rc->block_group->reserved > 0);
4125         WARN_ON(rc->block_group->used > 0);
4126 out:
4127         if (err && rw)
4128                 btrfs_dec_block_group_ro(rc->block_group);
4129         iput(rc->data_inode);
4130 out_put_bg:
4131         btrfs_put_block_group(bg);
4132         reloc_chunk_end(fs_info);
4133         free_reloc_control(rc);
4134         return err;
4135 }
4136
4137 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4138 {
4139         struct btrfs_fs_info *fs_info = root->fs_info;
4140         struct btrfs_trans_handle *trans;
4141         int ret, err;
4142
4143         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4144         if (IS_ERR(trans))
4145                 return PTR_ERR(trans);
4146
4147         memset(&root->root_item.drop_progress, 0,
4148                 sizeof(root->root_item.drop_progress));
4149         btrfs_set_root_drop_level(&root->root_item, 0);
4150         btrfs_set_root_refs(&root->root_item, 0);
4151         ret = btrfs_update_root(trans, fs_info->tree_root,
4152                                 &root->root_key, &root->root_item);
4153
4154         err = btrfs_end_transaction(trans);
4155         if (err)
4156                 return err;
4157         return ret;
4158 }
4159
4160 /*
4161  * recover relocation interrupted by system crash.
4162  *
4163  * this function resumes merging reloc trees with corresponding fs trees.
4164  * this is important for keeping the sharing of tree blocks
4165  */
4166 int btrfs_recover_relocation(struct btrfs_fs_info *fs_info)
4167 {
4168         LIST_HEAD(reloc_roots);
4169         struct btrfs_key key;
4170         struct btrfs_root *fs_root;
4171         struct btrfs_root *reloc_root;
4172         struct btrfs_path *path;
4173         struct extent_buffer *leaf;
4174         struct reloc_control *rc = NULL;
4175         struct btrfs_trans_handle *trans;
4176         int ret;
4177         int err = 0;
4178
4179         path = btrfs_alloc_path();
4180         if (!path)
4181                 return -ENOMEM;
4182         path->reada = READA_BACK;
4183
4184         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4185         key.type = BTRFS_ROOT_ITEM_KEY;
4186         key.offset = (u64)-1;
4187
4188         while (1) {
4189                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4190                                         path, 0, 0);
4191                 if (ret < 0) {
4192                         err = ret;
4193                         goto out;
4194                 }
4195                 if (ret > 0) {
4196                         if (path->slots[0] == 0)
4197                                 break;
4198                         path->slots[0]--;
4199                 }
4200                 leaf = path->nodes[0];
4201                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4202                 btrfs_release_path(path);
4203
4204                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4205                     key.type != BTRFS_ROOT_ITEM_KEY)
4206                         break;
4207
4208                 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &key);
4209                 if (IS_ERR(reloc_root)) {
4210                         err = PTR_ERR(reloc_root);
4211                         goto out;
4212                 }
4213
4214                 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4215                 list_add(&reloc_root->root_list, &reloc_roots);
4216
4217                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4218                         fs_root = btrfs_get_fs_root(fs_info,
4219                                         reloc_root->root_key.offset, false);
4220                         if (IS_ERR(fs_root)) {
4221                                 ret = PTR_ERR(fs_root);
4222                                 if (ret != -ENOENT) {
4223                                         err = ret;
4224                                         goto out;
4225                                 }
4226                                 ret = mark_garbage_root(reloc_root);
4227                                 if (ret < 0) {
4228                                         err = ret;
4229                                         goto out;
4230                                 }
4231                         } else {
4232                                 btrfs_put_root(fs_root);
4233                         }
4234                 }
4235
4236                 if (key.offset == 0)
4237                         break;
4238
4239                 key.offset--;
4240         }
4241         btrfs_release_path(path);
4242
4243         if (list_empty(&reloc_roots))
4244                 goto out;
4245
4246         rc = alloc_reloc_control(fs_info);
4247         if (!rc) {
4248                 err = -ENOMEM;
4249                 goto out;
4250         }
4251
4252         ret = reloc_chunk_start(fs_info);
4253         if (ret < 0) {
4254                 err = ret;
4255                 goto out_end;
4256         }
4257
4258         rc->extent_root = btrfs_extent_root(fs_info, 0);
4259
4260         set_reloc_control(rc);
4261
4262         trans = btrfs_join_transaction(rc->extent_root);
4263         if (IS_ERR(trans)) {
4264                 err = PTR_ERR(trans);
4265                 goto out_unset;
4266         }
4267
4268         rc->merge_reloc_tree = 1;
4269
4270         while (!list_empty(&reloc_roots)) {
4271                 reloc_root = list_entry(reloc_roots.next,
4272                                         struct btrfs_root, root_list);
4273                 list_del(&reloc_root->root_list);
4274
4275                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4276                         list_add_tail(&reloc_root->root_list,
4277                                       &rc->reloc_roots);
4278                         continue;
4279                 }
4280
4281                 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4282                                             false);
4283                 if (IS_ERR(fs_root)) {
4284                         err = PTR_ERR(fs_root);
4285                         list_add_tail(&reloc_root->root_list, &reloc_roots);
4286                         btrfs_end_transaction(trans);
4287                         goto out_unset;
4288                 }
4289
4290                 err = __add_reloc_root(reloc_root);
4291                 ASSERT(err != -EEXIST);
4292                 if (err) {
4293                         list_add_tail(&reloc_root->root_list, &reloc_roots);
4294                         btrfs_put_root(fs_root);
4295                         btrfs_end_transaction(trans);
4296                         goto out_unset;
4297                 }
4298                 fs_root->reloc_root = btrfs_grab_root(reloc_root);
4299                 btrfs_put_root(fs_root);
4300         }
4301
4302         err = btrfs_commit_transaction(trans);
4303         if (err)
4304                 goto out_unset;
4305
4306         merge_reloc_roots(rc);
4307
4308         unset_reloc_control(rc);
4309
4310         trans = btrfs_join_transaction(rc->extent_root);
4311         if (IS_ERR(trans)) {
4312                 err = PTR_ERR(trans);
4313                 goto out_clean;
4314         }
4315         err = btrfs_commit_transaction(trans);
4316 out_clean:
4317         ret = clean_dirty_subvols(rc);
4318         if (ret < 0 && !err)
4319                 err = ret;
4320 out_unset:
4321         unset_reloc_control(rc);
4322 out_end:
4323         reloc_chunk_end(fs_info);
4324         free_reloc_control(rc);
4325 out:
4326         free_reloc_roots(&reloc_roots);
4327
4328         btrfs_free_path(path);
4329
4330         if (err == 0) {
4331                 /* cleanup orphan inode in data relocation tree */
4332                 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4333                 ASSERT(fs_root);
4334                 err = btrfs_orphan_cleanup(fs_root);
4335                 btrfs_put_root(fs_root);
4336         }
4337         return err;
4338 }
4339
4340 /*
4341  * helper to add ordered checksum for data relocation.
4342  *
4343  * cloning checksum properly handles the nodatasum extents.
4344  * it also saves CPU time to re-calculate the checksum.
4345  */
4346 int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
4347 {
4348         struct btrfs_fs_info *fs_info = inode->root->fs_info;
4349         struct btrfs_root *csum_root;
4350         struct btrfs_ordered_sum *sums;
4351         struct btrfs_ordered_extent *ordered;
4352         int ret;
4353         u64 disk_bytenr;
4354         u64 new_bytenr;
4355         LIST_HEAD(list);
4356
4357         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4358         BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
4359
4360         disk_bytenr = file_pos + inode->index_cnt;
4361         csum_root = btrfs_csum_root(fs_info, disk_bytenr);
4362         ret = btrfs_lookup_csums_range(csum_root, disk_bytenr,
4363                                        disk_bytenr + len - 1, &list, 0, false);
4364         if (ret)
4365                 goto out;
4366
4367         while (!list_empty(&list)) {
4368                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4369                 list_del_init(&sums->list);
4370
4371                 /*
4372                  * We need to offset the new_bytenr based on where the csum is.
4373                  * We need to do this because we will read in entire prealloc
4374                  * extents but we may have written to say the middle of the
4375                  * prealloc extent, so we need to make sure the csum goes with
4376                  * the right disk offset.
4377                  *
4378                  * We can do this because the data reloc inode refers strictly
4379                  * to the on disk bytes, so we don't have to worry about
4380                  * disk_len vs real len like with real inodes since it's all
4381                  * disk length.
4382                  */
4383                 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
4384                 sums->bytenr = new_bytenr;
4385
4386                 btrfs_add_ordered_sum(ordered, sums);
4387         }
4388 out:
4389         btrfs_put_ordered_extent(ordered);
4390         return ret;
4391 }
4392
4393 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4394                           struct btrfs_root *root, struct extent_buffer *buf,
4395                           struct extent_buffer *cow)
4396 {
4397         struct btrfs_fs_info *fs_info = root->fs_info;
4398         struct reloc_control *rc;
4399         struct btrfs_backref_node *node;
4400         int first_cow = 0;
4401         int level;
4402         int ret = 0;
4403
4404         rc = fs_info->reloc_ctl;
4405         if (!rc)
4406                 return 0;
4407
4408         BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4409
4410         level = btrfs_header_level(buf);
4411         if (btrfs_header_generation(buf) <=
4412             btrfs_root_last_snapshot(&root->root_item))
4413                 first_cow = 1;
4414
4415         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4416             rc->create_reloc_tree) {
4417                 WARN_ON(!first_cow && level == 0);
4418
4419                 node = rc->backref_cache.path[level];
4420                 BUG_ON(node->bytenr != buf->start &&
4421                        node->new_bytenr != buf->start);
4422
4423                 btrfs_backref_drop_node_buffer(node);
4424                 atomic_inc(&cow->refs);
4425                 node->eb = cow;
4426                 node->new_bytenr = cow->start;
4427
4428                 if (!node->pending) {
4429                         list_move_tail(&node->list,
4430                                        &rc->backref_cache.pending[level]);
4431                         node->pending = 1;
4432                 }
4433
4434                 if (first_cow)
4435                         mark_block_processed(rc, node);
4436
4437                 if (first_cow && level > 0)
4438                         rc->nodes_relocated += buf->len;
4439         }
4440
4441         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4442                 ret = replace_file_extents(trans, rc, root, cow);
4443         return ret;
4444 }
4445
4446 /*
4447  * called before creating snapshot. it calculates metadata reservation
4448  * required for relocating tree blocks in the snapshot
4449  */
4450 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4451                               u64 *bytes_to_reserve)
4452 {
4453         struct btrfs_root *root = pending->root;
4454         struct reloc_control *rc = root->fs_info->reloc_ctl;
4455
4456         if (!rc || !have_reloc_root(root))
4457                 return;
4458
4459         if (!rc->merge_reloc_tree)
4460                 return;
4461
4462         root = root->reloc_root;
4463         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4464         /*
4465          * relocation is in the stage of merging trees. the space
4466          * used by merging a reloc tree is twice the size of
4467          * relocated tree nodes in the worst case. half for cowing
4468          * the reloc tree, half for cowing the fs tree. the space
4469          * used by cowing the reloc tree will be freed after the
4470          * tree is dropped. if we create snapshot, cowing the fs
4471          * tree may use more space than it frees. so we need
4472          * reserve extra space.
4473          */
4474         *bytes_to_reserve += rc->nodes_relocated;
4475 }
4476
4477 /*
4478  * called after snapshot is created. migrate block reservation
4479  * and create reloc root for the newly created snapshot
4480  *
4481  * This is similar to btrfs_init_reloc_root(), we come out of here with two
4482  * references held on the reloc_root, one for root->reloc_root and one for
4483  * rc->reloc_roots.
4484  */
4485 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4486                                struct btrfs_pending_snapshot *pending)
4487 {
4488         struct btrfs_root *root = pending->root;
4489         struct btrfs_root *reloc_root;
4490         struct btrfs_root *new_root;
4491         struct reloc_control *rc = root->fs_info->reloc_ctl;
4492         int ret;
4493
4494         if (!rc || !have_reloc_root(root))
4495                 return 0;
4496
4497         rc = root->fs_info->reloc_ctl;
4498         rc->merging_rsv_size += rc->nodes_relocated;
4499
4500         if (rc->merge_reloc_tree) {
4501                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4502                                               rc->block_rsv,
4503                                               rc->nodes_relocated, true);
4504                 if (ret)
4505                         return ret;
4506         }
4507
4508         new_root = pending->snap;
4509         reloc_root = create_reloc_root(trans, root->reloc_root,
4510                                        new_root->root_key.objectid);
4511         if (IS_ERR(reloc_root))
4512                 return PTR_ERR(reloc_root);
4513
4514         ret = __add_reloc_root(reloc_root);
4515         ASSERT(ret != -EEXIST);
4516         if (ret) {
4517                 /* Pairs with create_reloc_root */
4518                 btrfs_put_root(reloc_root);
4519                 return ret;
4520         }
4521         new_root->reloc_root = btrfs_grab_root(reloc_root);
4522
4523         if (rc->create_reloc_tree)
4524                 ret = clone_backref_node(trans, rc, root, reloc_root);
4525         return ret;
4526 }