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_cache(BTRFS_I(inode),
1128                                                 key.offset,     end, 1);
1129                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1130                                               key.offset, end);
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                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1330                 path->lowest_level = 0;
1331                 if (ret) {
1332                         if (ret > 0)
1333                                 ret = -ENOENT;
1334                         break;
1335                 }
1336
1337                 /*
1338                  * Info qgroup to trace both subtrees.
1339                  *
1340                  * We must trace both trees.
1341                  * 1) Tree reloc subtree
1342                  *    If not traced, we will leak data numbers
1343                  * 2) Fs subtree
1344                  *    If not traced, we will double count old data
1345                  *
1346                  * We don't scan the subtree right now, but only record
1347                  * the swapped tree blocks.
1348                  * The real subtree rescan is delayed until we have new
1349                  * CoW on the subtree root node before transaction commit.
1350                  */
1351                 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1352                                 rc->block_group, parent, slot,
1353                                 path->nodes[level], path->slots[level],
1354                                 last_snapshot);
1355                 if (ret < 0)
1356                         break;
1357                 /*
1358                  * swap blocks in fs tree and reloc tree.
1359                  */
1360                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1361                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1362                 btrfs_mark_buffer_dirty(parent);
1363
1364                 btrfs_set_node_blockptr(path->nodes[level],
1365                                         path->slots[level], old_bytenr);
1366                 btrfs_set_node_ptr_generation(path->nodes[level],
1367                                               path->slots[level], old_ptr_gen);
1368                 btrfs_mark_buffer_dirty(path->nodes[level]);
1369
1370                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1371                                        blocksize, path->nodes[level]->start);
1372                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1373                                     0, true);
1374                 ret = btrfs_inc_extent_ref(trans, &ref);
1375                 if (ret) {
1376                         btrfs_abort_transaction(trans, ret);
1377                         break;
1378                 }
1379                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1380                                        blocksize, 0);
1381                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid, 0,
1382                                     true);
1383                 ret = btrfs_inc_extent_ref(trans, &ref);
1384                 if (ret) {
1385                         btrfs_abort_transaction(trans, ret);
1386                         break;
1387                 }
1388
1389                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1390                                        blocksize, path->nodes[level]->start);
1391                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1392                                     0, true);
1393                 ret = btrfs_free_extent(trans, &ref);
1394                 if (ret) {
1395                         btrfs_abort_transaction(trans, ret);
1396                         break;
1397                 }
1398
1399                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1400                                        blocksize, 0);
1401                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid,
1402                                     0, true);
1403                 ret = btrfs_free_extent(trans, &ref);
1404                 if (ret) {
1405                         btrfs_abort_transaction(trans, ret);
1406                         break;
1407                 }
1408
1409                 btrfs_unlock_up_safe(path, 0);
1410
1411                 ret = level;
1412                 break;
1413         }
1414         btrfs_tree_unlock(parent);
1415         free_extent_buffer(parent);
1416         return ret;
1417 }
1418
1419 /*
1420  * helper to find next relocated block in reloc tree
1421  */
1422 static noinline_for_stack
1423 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1424                        int *level)
1425 {
1426         struct extent_buffer *eb;
1427         int i;
1428         u64 last_snapshot;
1429         u32 nritems;
1430
1431         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1432
1433         for (i = 0; i < *level; i++) {
1434                 free_extent_buffer(path->nodes[i]);
1435                 path->nodes[i] = NULL;
1436         }
1437
1438         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1439                 eb = path->nodes[i];
1440                 nritems = btrfs_header_nritems(eb);
1441                 while (path->slots[i] + 1 < nritems) {
1442                         path->slots[i]++;
1443                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1444                             last_snapshot)
1445                                 continue;
1446
1447                         *level = i;
1448                         return 0;
1449                 }
1450                 free_extent_buffer(path->nodes[i]);
1451                 path->nodes[i] = NULL;
1452         }
1453         return 1;
1454 }
1455
1456 /*
1457  * walk down reloc tree to find relocated block of lowest level
1458  */
1459 static noinline_for_stack
1460 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1461                          int *level)
1462 {
1463         struct extent_buffer *eb = NULL;
1464         int i;
1465         u64 ptr_gen = 0;
1466         u64 last_snapshot;
1467         u32 nritems;
1468
1469         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1470
1471         for (i = *level; i > 0; i--) {
1472                 eb = path->nodes[i];
1473                 nritems = btrfs_header_nritems(eb);
1474                 while (path->slots[i] < nritems) {
1475                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1476                         if (ptr_gen > last_snapshot)
1477                                 break;
1478                         path->slots[i]++;
1479                 }
1480                 if (path->slots[i] >= nritems) {
1481                         if (i == *level)
1482                                 break;
1483                         *level = i + 1;
1484                         return 0;
1485                 }
1486                 if (i == 1) {
1487                         *level = i;
1488                         return 0;
1489                 }
1490
1491                 eb = btrfs_read_node_slot(eb, path->slots[i]);
1492                 if (IS_ERR(eb))
1493                         return PTR_ERR(eb);
1494                 BUG_ON(btrfs_header_level(eb) != i - 1);
1495                 path->nodes[i - 1] = eb;
1496                 path->slots[i - 1] = 0;
1497         }
1498         return 1;
1499 }
1500
1501 /*
1502  * invalidate extent cache for file extents whose key in range of
1503  * [min_key, max_key)
1504  */
1505 static int invalidate_extent_cache(struct btrfs_root *root,
1506                                    struct btrfs_key *min_key,
1507                                    struct btrfs_key *max_key)
1508 {
1509         struct btrfs_fs_info *fs_info = root->fs_info;
1510         struct inode *inode = NULL;
1511         u64 objectid;
1512         u64 start, end;
1513         u64 ino;
1514
1515         objectid = min_key->objectid;
1516         while (1) {
1517                 cond_resched();
1518                 iput(inode);
1519
1520                 if (objectid > max_key->objectid)
1521                         break;
1522
1523                 inode = find_next_inode(root, objectid);
1524                 if (!inode)
1525                         break;
1526                 ino = btrfs_ino(BTRFS_I(inode));
1527
1528                 if (ino > max_key->objectid) {
1529                         iput(inode);
1530                         break;
1531                 }
1532
1533                 objectid = ino + 1;
1534                 if (!S_ISREG(inode->i_mode))
1535                         continue;
1536
1537                 if (unlikely(min_key->objectid == ino)) {
1538                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1539                                 continue;
1540                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1541                                 start = 0;
1542                         else {
1543                                 start = min_key->offset;
1544                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1545                         }
1546                 } else {
1547                         start = 0;
1548                 }
1549
1550                 if (unlikely(max_key->objectid == ino)) {
1551                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1552                                 continue;
1553                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1554                                 end = (u64)-1;
1555                         } else {
1556                                 if (max_key->offset == 0)
1557                                         continue;
1558                                 end = max_key->offset;
1559                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1560                                 end--;
1561                         }
1562                 } else {
1563                         end = (u64)-1;
1564                 }
1565
1566                 /* the lock_extent waits for read_folio to complete */
1567                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
1568                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
1569                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1570         }
1571         return 0;
1572 }
1573
1574 static int find_next_key(struct btrfs_path *path, int level,
1575                          struct btrfs_key *key)
1576
1577 {
1578         while (level < BTRFS_MAX_LEVEL) {
1579                 if (!path->nodes[level])
1580                         break;
1581                 if (path->slots[level] + 1 <
1582                     btrfs_header_nritems(path->nodes[level])) {
1583                         btrfs_node_key_to_cpu(path->nodes[level], key,
1584                                               path->slots[level] + 1);
1585                         return 0;
1586                 }
1587                 level++;
1588         }
1589         return 1;
1590 }
1591
1592 /*
1593  * Insert current subvolume into reloc_control::dirty_subvol_roots
1594  */
1595 static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1596                                struct reloc_control *rc,
1597                                struct btrfs_root *root)
1598 {
1599         struct btrfs_root *reloc_root = root->reloc_root;
1600         struct btrfs_root_item *reloc_root_item;
1601         int ret;
1602
1603         /* @root must be a subvolume tree root with a valid reloc tree */
1604         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1605         ASSERT(reloc_root);
1606
1607         reloc_root_item = &reloc_root->root_item;
1608         memset(&reloc_root_item->drop_progress, 0,
1609                 sizeof(reloc_root_item->drop_progress));
1610         btrfs_set_root_drop_level(reloc_root_item, 0);
1611         btrfs_set_root_refs(reloc_root_item, 0);
1612         ret = btrfs_update_reloc_root(trans, root);
1613         if (ret)
1614                 return ret;
1615
1616         if (list_empty(&root->reloc_dirty_list)) {
1617                 btrfs_grab_root(root);
1618                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1619         }
1620
1621         return 0;
1622 }
1623
1624 static int clean_dirty_subvols(struct reloc_control *rc)
1625 {
1626         struct btrfs_root *root;
1627         struct btrfs_root *next;
1628         int ret = 0;
1629         int ret2;
1630
1631         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1632                                  reloc_dirty_list) {
1633                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1634                         /* Merged subvolume, cleanup its reloc root */
1635                         struct btrfs_root *reloc_root = root->reloc_root;
1636
1637                         list_del_init(&root->reloc_dirty_list);
1638                         root->reloc_root = NULL;
1639                         /*
1640                          * Need barrier to ensure clear_bit() only happens after
1641                          * root->reloc_root = NULL. Pairs with have_reloc_root.
1642                          */
1643                         smp_wmb();
1644                         clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1645                         if (reloc_root) {
1646                                 /*
1647                                  * btrfs_drop_snapshot drops our ref we hold for
1648                                  * ->reloc_root.  If it fails however we must
1649                                  * drop the ref ourselves.
1650                                  */
1651                                 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1652                                 if (ret2 < 0) {
1653                                         btrfs_put_root(reloc_root);
1654                                         if (!ret)
1655                                                 ret = ret2;
1656                                 }
1657                         }
1658                         btrfs_put_root(root);
1659                 } else {
1660                         /* Orphan reloc tree, just clean it up */
1661                         ret2 = btrfs_drop_snapshot(root, 0, 1);
1662                         if (ret2 < 0) {
1663                                 btrfs_put_root(root);
1664                                 if (!ret)
1665                                         ret = ret2;
1666                         }
1667                 }
1668         }
1669         return ret;
1670 }
1671
1672 /*
1673  * merge the relocated tree blocks in reloc tree with corresponding
1674  * fs tree.
1675  */
1676 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1677                                                struct btrfs_root *root)
1678 {
1679         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1680         struct btrfs_key key;
1681         struct btrfs_key next_key;
1682         struct btrfs_trans_handle *trans = NULL;
1683         struct btrfs_root *reloc_root;
1684         struct btrfs_root_item *root_item;
1685         struct btrfs_path *path;
1686         struct extent_buffer *leaf;
1687         int reserve_level;
1688         int level;
1689         int max_level;
1690         int replaced = 0;
1691         int ret = 0;
1692         u32 min_reserved;
1693
1694         path = btrfs_alloc_path();
1695         if (!path)
1696                 return -ENOMEM;
1697         path->reada = READA_FORWARD;
1698
1699         reloc_root = root->reloc_root;
1700         root_item = &reloc_root->root_item;
1701
1702         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1703                 level = btrfs_root_level(root_item);
1704                 atomic_inc(&reloc_root->node->refs);
1705                 path->nodes[level] = reloc_root->node;
1706                 path->slots[level] = 0;
1707         } else {
1708                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1709
1710                 level = btrfs_root_drop_level(root_item);
1711                 BUG_ON(level == 0);
1712                 path->lowest_level = level;
1713                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1714                 path->lowest_level = 0;
1715                 if (ret < 0) {
1716                         btrfs_free_path(path);
1717                         return ret;
1718                 }
1719
1720                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1721                                       path->slots[level]);
1722                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1723
1724                 btrfs_unlock_up_safe(path, 0);
1725         }
1726
1727         /*
1728          * In merge_reloc_root(), we modify the upper level pointer to swap the
1729          * tree blocks between reloc tree and subvolume tree.  Thus for tree
1730          * block COW, we COW at most from level 1 to root level for each tree.
1731          *
1732          * Thus the needed metadata size is at most root_level * nodesize,
1733          * and * 2 since we have two trees to COW.
1734          */
1735         reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1736         min_reserved = fs_info->nodesize * reserve_level * 2;
1737         memset(&next_key, 0, sizeof(next_key));
1738
1739         while (1) {
1740                 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
1741                                              min_reserved,
1742                                              BTRFS_RESERVE_FLUSH_LIMIT);
1743                 if (ret)
1744                         goto out;
1745                 trans = btrfs_start_transaction(root, 0);
1746                 if (IS_ERR(trans)) {
1747                         ret = PTR_ERR(trans);
1748                         trans = NULL;
1749                         goto out;
1750                 }
1751
1752                 /*
1753                  * At this point we no longer have a reloc_control, so we can't
1754                  * depend on btrfs_init_reloc_root to update our last_trans.
1755                  *
1756                  * But that's ok, we started the trans handle on our
1757                  * corresponding fs_root, which means it's been added to the
1758                  * dirty list.  At commit time we'll still call
1759                  * btrfs_update_reloc_root() and update our root item
1760                  * appropriately.
1761                  */
1762                 reloc_root->last_trans = trans->transid;
1763                 trans->block_rsv = rc->block_rsv;
1764
1765                 replaced = 0;
1766                 max_level = level;
1767
1768                 ret = walk_down_reloc_tree(reloc_root, path, &level);
1769                 if (ret < 0)
1770                         goto out;
1771                 if (ret > 0)
1772                         break;
1773
1774                 if (!find_next_key(path, level, &key) &&
1775                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1776                         ret = 0;
1777                 } else {
1778                         ret = replace_path(trans, rc, root, reloc_root, path,
1779                                            &next_key, level, max_level);
1780                 }
1781                 if (ret < 0)
1782                         goto out;
1783                 if (ret > 0) {
1784                         level = ret;
1785                         btrfs_node_key_to_cpu(path->nodes[level], &key,
1786                                               path->slots[level]);
1787                         replaced = 1;
1788                 }
1789
1790                 ret = walk_up_reloc_tree(reloc_root, path, &level);
1791                 if (ret > 0)
1792                         break;
1793
1794                 BUG_ON(level == 0);
1795                 /*
1796                  * save the merging progress in the drop_progress.
1797                  * this is OK since root refs == 1 in this case.
1798                  */
1799                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1800                                path->slots[level]);
1801                 btrfs_set_root_drop_level(root_item, level);
1802
1803                 btrfs_end_transaction_throttle(trans);
1804                 trans = NULL;
1805
1806                 btrfs_btree_balance_dirty(fs_info);
1807
1808                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1809                         invalidate_extent_cache(root, &key, &next_key);
1810         }
1811
1812         /*
1813          * handle the case only one block in the fs tree need to be
1814          * relocated and the block is tree root.
1815          */
1816         leaf = btrfs_lock_root_node(root);
1817         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1818                               BTRFS_NESTING_COW);
1819         btrfs_tree_unlock(leaf);
1820         free_extent_buffer(leaf);
1821 out:
1822         btrfs_free_path(path);
1823
1824         if (ret == 0) {
1825                 ret = insert_dirty_subvol(trans, rc, root);
1826                 if (ret)
1827                         btrfs_abort_transaction(trans, ret);
1828         }
1829
1830         if (trans)
1831                 btrfs_end_transaction_throttle(trans);
1832
1833         btrfs_btree_balance_dirty(fs_info);
1834
1835         if (replaced && rc->stage == UPDATE_DATA_PTRS)
1836                 invalidate_extent_cache(root, &key, &next_key);
1837
1838         return ret;
1839 }
1840
1841 static noinline_for_stack
1842 int prepare_to_merge(struct reloc_control *rc, int err)
1843 {
1844         struct btrfs_root *root = rc->extent_root;
1845         struct btrfs_fs_info *fs_info = root->fs_info;
1846         struct btrfs_root *reloc_root;
1847         struct btrfs_trans_handle *trans;
1848         LIST_HEAD(reloc_roots);
1849         u64 num_bytes = 0;
1850         int ret;
1851
1852         mutex_lock(&fs_info->reloc_mutex);
1853         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1854         rc->merging_rsv_size += rc->nodes_relocated * 2;
1855         mutex_unlock(&fs_info->reloc_mutex);
1856
1857 again:
1858         if (!err) {
1859                 num_bytes = rc->merging_rsv_size;
1860                 ret = btrfs_block_rsv_add(fs_info, rc->block_rsv, num_bytes,
1861                                           BTRFS_RESERVE_FLUSH_ALL);
1862                 if (ret)
1863                         err = ret;
1864         }
1865
1866         trans = btrfs_join_transaction(rc->extent_root);
1867         if (IS_ERR(trans)) {
1868                 if (!err)
1869                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1870                                                 num_bytes, NULL);
1871                 return PTR_ERR(trans);
1872         }
1873
1874         if (!err) {
1875                 if (num_bytes != rc->merging_rsv_size) {
1876                         btrfs_end_transaction(trans);
1877                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1878                                                 num_bytes, NULL);
1879                         goto again;
1880                 }
1881         }
1882
1883         rc->merge_reloc_tree = 1;
1884
1885         while (!list_empty(&rc->reloc_roots)) {
1886                 reloc_root = list_entry(rc->reloc_roots.next,
1887                                         struct btrfs_root, root_list);
1888                 list_del_init(&reloc_root->root_list);
1889
1890                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1891                                 false);
1892                 if (IS_ERR(root)) {
1893                         /*
1894                          * Even if we have an error we need this reloc root
1895                          * back on our list so we can clean up properly.
1896                          */
1897                         list_add(&reloc_root->root_list, &reloc_roots);
1898                         btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1899                         if (!err)
1900                                 err = PTR_ERR(root);
1901                         break;
1902                 }
1903                 ASSERT(root->reloc_root == reloc_root);
1904
1905                 /*
1906                  * set reference count to 1, so btrfs_recover_relocation
1907                  * knows it should resumes merging
1908                  */
1909                 if (!err)
1910                         btrfs_set_root_refs(&reloc_root->root_item, 1);
1911                 ret = btrfs_update_reloc_root(trans, root);
1912
1913                 /*
1914                  * Even if we have an error we need this reloc root back on our
1915                  * list so we can clean up properly.
1916                  */
1917                 list_add(&reloc_root->root_list, &reloc_roots);
1918                 btrfs_put_root(root);
1919
1920                 if (ret) {
1921                         btrfs_abort_transaction(trans, ret);
1922                         if (!err)
1923                                 err = ret;
1924                         break;
1925                 }
1926         }
1927
1928         list_splice(&reloc_roots, &rc->reloc_roots);
1929
1930         if (!err)
1931                 err = btrfs_commit_transaction(trans);
1932         else
1933                 btrfs_end_transaction(trans);
1934         return err;
1935 }
1936
1937 static noinline_for_stack
1938 void free_reloc_roots(struct list_head *list)
1939 {
1940         struct btrfs_root *reloc_root, *tmp;
1941
1942         list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1943                 __del_reloc_root(reloc_root);
1944 }
1945
1946 static noinline_for_stack
1947 void merge_reloc_roots(struct reloc_control *rc)
1948 {
1949         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1950         struct btrfs_root *root;
1951         struct btrfs_root *reloc_root;
1952         LIST_HEAD(reloc_roots);
1953         int found = 0;
1954         int ret = 0;
1955 again:
1956         root = rc->extent_root;
1957
1958         /*
1959          * this serializes us with btrfs_record_root_in_transaction,
1960          * we have to make sure nobody is in the middle of
1961          * adding their roots to the list while we are
1962          * doing this splice
1963          */
1964         mutex_lock(&fs_info->reloc_mutex);
1965         list_splice_init(&rc->reloc_roots, &reloc_roots);
1966         mutex_unlock(&fs_info->reloc_mutex);
1967
1968         while (!list_empty(&reloc_roots)) {
1969                 found = 1;
1970                 reloc_root = list_entry(reloc_roots.next,
1971                                         struct btrfs_root, root_list);
1972
1973                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1974                                          false);
1975                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1976                         if (IS_ERR(root)) {
1977                                 /*
1978                                  * For recovery we read the fs roots on mount,
1979                                  * and if we didn't find the root then we marked
1980                                  * the reloc root as a garbage root.  For normal
1981                                  * relocation obviously the root should exist in
1982                                  * memory.  However there's no reason we can't
1983                                  * handle the error properly here just in case.
1984                                  */
1985                                 ASSERT(0);
1986                                 ret = PTR_ERR(root);
1987                                 goto out;
1988                         }
1989                         if (root->reloc_root != reloc_root) {
1990                                 /*
1991                                  * This is actually impossible without something
1992                                  * going really wrong (like weird race condition
1993                                  * or cosmic rays).
1994                                  */
1995                                 ASSERT(0);
1996                                 ret = -EINVAL;
1997                                 goto out;
1998                         }
1999                         ret = merge_reloc_root(rc, root);
2000                         btrfs_put_root(root);
2001                         if (ret) {
2002                                 if (list_empty(&reloc_root->root_list))
2003                                         list_add_tail(&reloc_root->root_list,
2004                                                       &reloc_roots);
2005                                 goto out;
2006                         }
2007                 } else {
2008                         if (!IS_ERR(root)) {
2009                                 if (root->reloc_root == reloc_root) {
2010                                         root->reloc_root = NULL;
2011                                         btrfs_put_root(reloc_root);
2012                                 }
2013                                 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
2014                                           &root->state);
2015                                 btrfs_put_root(root);
2016                         }
2017
2018                         list_del_init(&reloc_root->root_list);
2019                         /* Don't forget to queue this reloc root for cleanup */
2020                         list_add_tail(&reloc_root->reloc_dirty_list,
2021                                       &rc->dirty_subvol_roots);
2022                 }
2023         }
2024
2025         if (found) {
2026                 found = 0;
2027                 goto again;
2028         }
2029 out:
2030         if (ret) {
2031                 btrfs_handle_fs_error(fs_info, ret, NULL);
2032                 free_reloc_roots(&reloc_roots);
2033
2034                 /* new reloc root may be added */
2035                 mutex_lock(&fs_info->reloc_mutex);
2036                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2037                 mutex_unlock(&fs_info->reloc_mutex);
2038                 free_reloc_roots(&reloc_roots);
2039         }
2040
2041         /*
2042          * We used to have
2043          *
2044          * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2045          *
2046          * here, but it's wrong.  If we fail to start the transaction in
2047          * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2048          * have actually been removed from the reloc_root_tree rb tree.  This is
2049          * fine because we're bailing here, and we hold a reference on the root
2050          * for the list that holds it, so these roots will be cleaned up when we
2051          * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
2052          * will be cleaned up on unmount.
2053          *
2054          * The remaining nodes will be cleaned up by free_reloc_control.
2055          */
2056 }
2057
2058 static void free_block_list(struct rb_root *blocks)
2059 {
2060         struct tree_block *block;
2061         struct rb_node *rb_node;
2062         while ((rb_node = rb_first(blocks))) {
2063                 block = rb_entry(rb_node, struct tree_block, rb_node);
2064                 rb_erase(rb_node, blocks);
2065                 kfree(block);
2066         }
2067 }
2068
2069 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2070                                       struct btrfs_root *reloc_root)
2071 {
2072         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2073         struct btrfs_root *root;
2074         int ret;
2075
2076         if (reloc_root->last_trans == trans->transid)
2077                 return 0;
2078
2079         root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2080
2081         /*
2082          * This should succeed, since we can't have a reloc root without having
2083          * already looked up the actual root and created the reloc root for this
2084          * root.
2085          *
2086          * However if there's some sort of corruption where we have a ref to a
2087          * reloc root without a corresponding root this could return ENOENT.
2088          */
2089         if (IS_ERR(root)) {
2090                 ASSERT(0);
2091                 return PTR_ERR(root);
2092         }
2093         if (root->reloc_root != reloc_root) {
2094                 ASSERT(0);
2095                 btrfs_err(fs_info,
2096                           "root %llu has two reloc roots associated with it",
2097                           reloc_root->root_key.offset);
2098                 btrfs_put_root(root);
2099                 return -EUCLEAN;
2100         }
2101         ret = btrfs_record_root_in_trans(trans, root);
2102         btrfs_put_root(root);
2103
2104         return ret;
2105 }
2106
2107 static noinline_for_stack
2108 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2109                                      struct reloc_control *rc,
2110                                      struct btrfs_backref_node *node,
2111                                      struct btrfs_backref_edge *edges[])
2112 {
2113         struct btrfs_backref_node *next;
2114         struct btrfs_root *root;
2115         int index = 0;
2116         int ret;
2117
2118         next = node;
2119         while (1) {
2120                 cond_resched();
2121                 next = walk_up_backref(next, edges, &index);
2122                 root = next->root;
2123
2124                 /*
2125                  * If there is no root, then our references for this block are
2126                  * incomplete, as we should be able to walk all the way up to a
2127                  * block that is owned by a root.
2128                  *
2129                  * This path is only for SHAREABLE roots, so if we come upon a
2130                  * non-SHAREABLE root then we have backrefs that resolve
2131                  * improperly.
2132                  *
2133                  * Both of these cases indicate file system corruption, or a bug
2134                  * in the backref walking code.
2135                  */
2136                 if (!root) {
2137                         ASSERT(0);
2138                         btrfs_err(trans->fs_info,
2139                 "bytenr %llu doesn't have a backref path ending in a root",
2140                                   node->bytenr);
2141                         return ERR_PTR(-EUCLEAN);
2142                 }
2143                 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2144                         ASSERT(0);
2145                         btrfs_err(trans->fs_info,
2146         "bytenr %llu has multiple refs with one ending in a non-shareable root",
2147                                   node->bytenr);
2148                         return ERR_PTR(-EUCLEAN);
2149                 }
2150
2151                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2152                         ret = record_reloc_root_in_trans(trans, root);
2153                         if (ret)
2154                                 return ERR_PTR(ret);
2155                         break;
2156                 }
2157
2158                 ret = btrfs_record_root_in_trans(trans, root);
2159                 if (ret)
2160                         return ERR_PTR(ret);
2161                 root = root->reloc_root;
2162
2163                 /*
2164                  * We could have raced with another thread which failed, so
2165                  * root->reloc_root may not be set, return ENOENT in this case.
2166                  */
2167                 if (!root)
2168                         return ERR_PTR(-ENOENT);
2169
2170                 if (next->new_bytenr != root->node->start) {
2171                         /*
2172                          * We just created the reloc root, so we shouldn't have
2173                          * ->new_bytenr set and this shouldn't be in the changed
2174                          *  list.  If it is then we have multiple roots pointing
2175                          *  at the same bytenr which indicates corruption, or
2176                          *  we've made a mistake in the backref walking code.
2177                          */
2178                         ASSERT(next->new_bytenr == 0);
2179                         ASSERT(list_empty(&next->list));
2180                         if (next->new_bytenr || !list_empty(&next->list)) {
2181                                 btrfs_err(trans->fs_info,
2182         "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2183                                           node->bytenr, next->bytenr);
2184                                 return ERR_PTR(-EUCLEAN);
2185                         }
2186
2187                         next->new_bytenr = root->node->start;
2188                         btrfs_put_root(next->root);
2189                         next->root = btrfs_grab_root(root);
2190                         ASSERT(next->root);
2191                         list_add_tail(&next->list,
2192                                       &rc->backref_cache.changed);
2193                         mark_block_processed(rc, next);
2194                         break;
2195                 }
2196
2197                 WARN_ON(1);
2198                 root = NULL;
2199                 next = walk_down_backref(edges, &index);
2200                 if (!next || next->level <= node->level)
2201                         break;
2202         }
2203         if (!root) {
2204                 /*
2205                  * This can happen if there's fs corruption or if there's a bug
2206                  * in the backref lookup code.
2207                  */
2208                 ASSERT(0);
2209                 return ERR_PTR(-ENOENT);
2210         }
2211
2212         next = node;
2213         /* setup backref node path for btrfs_reloc_cow_block */
2214         while (1) {
2215                 rc->backref_cache.path[next->level] = next;
2216                 if (--index < 0)
2217                         break;
2218                 next = edges[index]->node[UPPER];
2219         }
2220         return root;
2221 }
2222
2223 /*
2224  * Select a tree root for relocation.
2225  *
2226  * Return NULL if the block is not shareable. We should use do_relocation() in
2227  * this case.
2228  *
2229  * Return a tree root pointer if the block is shareable.
2230  * Return -ENOENT if the block is root of reloc tree.
2231  */
2232 static noinline_for_stack
2233 struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2234 {
2235         struct btrfs_backref_node *next;
2236         struct btrfs_root *root;
2237         struct btrfs_root *fs_root = NULL;
2238         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2239         int index = 0;
2240
2241         next = node;
2242         while (1) {
2243                 cond_resched();
2244                 next = walk_up_backref(next, edges, &index);
2245                 root = next->root;
2246
2247                 /*
2248                  * This can occur if we have incomplete extent refs leading all
2249                  * the way up a particular path, in this case return -EUCLEAN.
2250                  */
2251                 if (!root)
2252                         return ERR_PTR(-EUCLEAN);
2253
2254                 /* No other choice for non-shareable tree */
2255                 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2256                         return root;
2257
2258                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2259                         fs_root = root;
2260
2261                 if (next != node)
2262                         return NULL;
2263
2264                 next = walk_down_backref(edges, &index);
2265                 if (!next || next->level <= node->level)
2266                         break;
2267         }
2268
2269         if (!fs_root)
2270                 return ERR_PTR(-ENOENT);
2271         return fs_root;
2272 }
2273
2274 static noinline_for_stack
2275 u64 calcu_metadata_size(struct reloc_control *rc,
2276                         struct btrfs_backref_node *node, int reserve)
2277 {
2278         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2279         struct btrfs_backref_node *next = node;
2280         struct btrfs_backref_edge *edge;
2281         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2282         u64 num_bytes = 0;
2283         int index = 0;
2284
2285         BUG_ON(reserve && node->processed);
2286
2287         while (next) {
2288                 cond_resched();
2289                 while (1) {
2290                         if (next->processed && (reserve || next != node))
2291                                 break;
2292
2293                         num_bytes += fs_info->nodesize;
2294
2295                         if (list_empty(&next->upper))
2296                                 break;
2297
2298                         edge = list_entry(next->upper.next,
2299                                         struct btrfs_backref_edge, list[LOWER]);
2300                         edges[index++] = edge;
2301                         next = edge->node[UPPER];
2302                 }
2303                 next = walk_down_backref(edges, &index);
2304         }
2305         return num_bytes;
2306 }
2307
2308 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2309                                   struct reloc_control *rc,
2310                                   struct btrfs_backref_node *node)
2311 {
2312         struct btrfs_root *root = rc->extent_root;
2313         struct btrfs_fs_info *fs_info = root->fs_info;
2314         u64 num_bytes;
2315         int ret;
2316         u64 tmp;
2317
2318         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2319
2320         trans->block_rsv = rc->block_rsv;
2321         rc->reserved_bytes += num_bytes;
2322
2323         /*
2324          * We are under a transaction here so we can only do limited flushing.
2325          * If we get an enospc just kick back -EAGAIN so we know to drop the
2326          * transaction and try to refill when we can flush all the things.
2327          */
2328         ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv, num_bytes,
2329                                      BTRFS_RESERVE_FLUSH_LIMIT);
2330         if (ret) {
2331                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2332                 while (tmp <= rc->reserved_bytes)
2333                         tmp <<= 1;
2334                 /*
2335                  * only one thread can access block_rsv at this point,
2336                  * so we don't need hold lock to protect block_rsv.
2337                  * we expand more reservation size here to allow enough
2338                  * space for relocation and we will return earlier in
2339                  * enospc case.
2340                  */
2341                 rc->block_rsv->size = tmp + fs_info->nodesize *
2342                                       RELOCATION_RESERVED_NODES;
2343                 return -EAGAIN;
2344         }
2345
2346         return 0;
2347 }
2348
2349 /*
2350  * relocate a block tree, and then update pointers in upper level
2351  * blocks that reference the block to point to the new location.
2352  *
2353  * if called by link_to_upper, the block has already been relocated.
2354  * in that case this function just updates pointers.
2355  */
2356 static int do_relocation(struct btrfs_trans_handle *trans,
2357                          struct reloc_control *rc,
2358                          struct btrfs_backref_node *node,
2359                          struct btrfs_key *key,
2360                          struct btrfs_path *path, int lowest)
2361 {
2362         struct btrfs_backref_node *upper;
2363         struct btrfs_backref_edge *edge;
2364         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2365         struct btrfs_root *root;
2366         struct extent_buffer *eb;
2367         u32 blocksize;
2368         u64 bytenr;
2369         int slot;
2370         int ret = 0;
2371
2372         /*
2373          * If we are lowest then this is the first time we're processing this
2374          * block, and thus shouldn't have an eb associated with it yet.
2375          */
2376         ASSERT(!lowest || !node->eb);
2377
2378         path->lowest_level = node->level + 1;
2379         rc->backref_cache.path[node->level] = node;
2380         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2381                 struct btrfs_ref ref = { 0 };
2382
2383                 cond_resched();
2384
2385                 upper = edge->node[UPPER];
2386                 root = select_reloc_root(trans, rc, upper, edges);
2387                 if (IS_ERR(root)) {
2388                         ret = PTR_ERR(root);
2389                         goto next;
2390                 }
2391
2392                 if (upper->eb && !upper->locked) {
2393                         if (!lowest) {
2394                                 ret = btrfs_bin_search(upper->eb, key, &slot);
2395                                 if (ret < 0)
2396                                         goto next;
2397                                 BUG_ON(ret);
2398                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2399                                 if (node->eb->start == bytenr)
2400                                         goto next;
2401                         }
2402                         btrfs_backref_drop_node_buffer(upper);
2403                 }
2404
2405                 if (!upper->eb) {
2406                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2407                         if (ret) {
2408                                 if (ret > 0)
2409                                         ret = -ENOENT;
2410
2411                                 btrfs_release_path(path);
2412                                 break;
2413                         }
2414
2415                         if (!upper->eb) {
2416                                 upper->eb = path->nodes[upper->level];
2417                                 path->nodes[upper->level] = NULL;
2418                         } else {
2419                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2420                         }
2421
2422                         upper->locked = 1;
2423                         path->locks[upper->level] = 0;
2424
2425                         slot = path->slots[upper->level];
2426                         btrfs_release_path(path);
2427                 } else {
2428                         ret = btrfs_bin_search(upper->eb, key, &slot);
2429                         if (ret < 0)
2430                                 goto next;
2431                         BUG_ON(ret);
2432                 }
2433
2434                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2435                 if (lowest) {
2436                         if (bytenr != node->bytenr) {
2437                                 btrfs_err(root->fs_info,
2438                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2439                                           bytenr, node->bytenr, slot,
2440                                           upper->eb->start);
2441                                 ret = -EIO;
2442                                 goto next;
2443                         }
2444                 } else {
2445                         if (node->eb->start == bytenr)
2446                                 goto next;
2447                 }
2448
2449                 blocksize = root->fs_info->nodesize;
2450                 eb = btrfs_read_node_slot(upper->eb, slot);
2451                 if (IS_ERR(eb)) {
2452                         ret = PTR_ERR(eb);
2453                         goto next;
2454                 }
2455                 btrfs_tree_lock(eb);
2456
2457                 if (!node->eb) {
2458                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2459                                               slot, &eb, BTRFS_NESTING_COW);
2460                         btrfs_tree_unlock(eb);
2461                         free_extent_buffer(eb);
2462                         if (ret < 0)
2463                                 goto next;
2464                         /*
2465                          * We've just COWed this block, it should have updated
2466                          * the correct backref node entry.
2467                          */
2468                         ASSERT(node->eb == eb);
2469                 } else {
2470                         btrfs_set_node_blockptr(upper->eb, slot,
2471                                                 node->eb->start);
2472                         btrfs_set_node_ptr_generation(upper->eb, slot,
2473                                                       trans->transid);
2474                         btrfs_mark_buffer_dirty(upper->eb);
2475
2476                         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2477                                                node->eb->start, blocksize,
2478                                                upper->eb->start);
2479                         btrfs_init_tree_ref(&ref, node->level,
2480                                             btrfs_header_owner(upper->eb),
2481                                             root->root_key.objectid, false);
2482                         ret = btrfs_inc_extent_ref(trans, &ref);
2483                         if (!ret)
2484                                 ret = btrfs_drop_subtree(trans, root, eb,
2485                                                          upper->eb);
2486                         if (ret)
2487                                 btrfs_abort_transaction(trans, ret);
2488                 }
2489 next:
2490                 if (!upper->pending)
2491                         btrfs_backref_drop_node_buffer(upper);
2492                 else
2493                         btrfs_backref_unlock_node_buffer(upper);
2494                 if (ret)
2495                         break;
2496         }
2497
2498         if (!ret && node->pending) {
2499                 btrfs_backref_drop_node_buffer(node);
2500                 list_move_tail(&node->list, &rc->backref_cache.changed);
2501                 node->pending = 0;
2502         }
2503
2504         path->lowest_level = 0;
2505
2506         /*
2507          * We should have allocated all of our space in the block rsv and thus
2508          * shouldn't ENOSPC.
2509          */
2510         ASSERT(ret != -ENOSPC);
2511         return ret;
2512 }
2513
2514 static int link_to_upper(struct btrfs_trans_handle *trans,
2515                          struct reloc_control *rc,
2516                          struct btrfs_backref_node *node,
2517                          struct btrfs_path *path)
2518 {
2519         struct btrfs_key key;
2520
2521         btrfs_node_key_to_cpu(node->eb, &key, 0);
2522         return do_relocation(trans, rc, node, &key, path, 0);
2523 }
2524
2525 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2526                                 struct reloc_control *rc,
2527                                 struct btrfs_path *path, int err)
2528 {
2529         LIST_HEAD(list);
2530         struct btrfs_backref_cache *cache = &rc->backref_cache;
2531         struct btrfs_backref_node *node;
2532         int level;
2533         int ret;
2534
2535         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2536                 while (!list_empty(&cache->pending[level])) {
2537                         node = list_entry(cache->pending[level].next,
2538                                           struct btrfs_backref_node, list);
2539                         list_move_tail(&node->list, &list);
2540                         BUG_ON(!node->pending);
2541
2542                         if (!err) {
2543                                 ret = link_to_upper(trans, rc, node, path);
2544                                 if (ret < 0)
2545                                         err = ret;
2546                         }
2547                 }
2548                 list_splice_init(&list, &cache->pending[level]);
2549         }
2550         return err;
2551 }
2552
2553 /*
2554  * mark a block and all blocks directly/indirectly reference the block
2555  * as processed.
2556  */
2557 static void update_processed_blocks(struct reloc_control *rc,
2558                                     struct btrfs_backref_node *node)
2559 {
2560         struct btrfs_backref_node *next = node;
2561         struct btrfs_backref_edge *edge;
2562         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2563         int index = 0;
2564
2565         while (next) {
2566                 cond_resched();
2567                 while (1) {
2568                         if (next->processed)
2569                                 break;
2570
2571                         mark_block_processed(rc, next);
2572
2573                         if (list_empty(&next->upper))
2574                                 break;
2575
2576                         edge = list_entry(next->upper.next,
2577                                         struct btrfs_backref_edge, list[LOWER]);
2578                         edges[index++] = edge;
2579                         next = edge->node[UPPER];
2580                 }
2581                 next = walk_down_backref(edges, &index);
2582         }
2583 }
2584
2585 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2586 {
2587         u32 blocksize = rc->extent_root->fs_info->nodesize;
2588
2589         if (test_range_bit(&rc->processed_blocks, bytenr,
2590                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2591                 return 1;
2592         return 0;
2593 }
2594
2595 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2596                               struct tree_block *block)
2597 {
2598         struct extent_buffer *eb;
2599
2600         eb = read_tree_block(fs_info, block->bytenr, block->owner,
2601                              block->key.offset, block->level, NULL);
2602         if (IS_ERR(eb))
2603                 return PTR_ERR(eb);
2604         if (!extent_buffer_uptodate(eb)) {
2605                 free_extent_buffer(eb);
2606                 return -EIO;
2607         }
2608         if (block->level == 0)
2609                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2610         else
2611                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2612         free_extent_buffer(eb);
2613         block->key_ready = 1;
2614         return 0;
2615 }
2616
2617 /*
2618  * helper function to relocate a tree block
2619  */
2620 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2621                                 struct reloc_control *rc,
2622                                 struct btrfs_backref_node *node,
2623                                 struct btrfs_key *key,
2624                                 struct btrfs_path *path)
2625 {
2626         struct btrfs_root *root;
2627         int ret = 0;
2628
2629         if (!node)
2630                 return 0;
2631
2632         /*
2633          * If we fail here we want to drop our backref_node because we are going
2634          * to start over and regenerate the tree for it.
2635          */
2636         ret = reserve_metadata_space(trans, rc, node);
2637         if (ret)
2638                 goto out;
2639
2640         BUG_ON(node->processed);
2641         root = select_one_root(node);
2642         if (IS_ERR(root)) {
2643                 ret = PTR_ERR(root);
2644
2645                 /* See explanation in select_one_root for the -EUCLEAN case. */
2646                 ASSERT(ret == -ENOENT);
2647                 if (ret == -ENOENT) {
2648                         ret = 0;
2649                         update_processed_blocks(rc, node);
2650                 }
2651                 goto out;
2652         }
2653
2654         if (root) {
2655                 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2656                         /*
2657                          * This block was the root block of a root, and this is
2658                          * the first time we're processing the block and thus it
2659                          * should not have had the ->new_bytenr modified and
2660                          * should have not been included on the changed list.
2661                          *
2662                          * However in the case of corruption we could have
2663                          * multiple refs pointing to the same block improperly,
2664                          * and thus we would trip over these checks.  ASSERT()
2665                          * for the developer case, because it could indicate a
2666                          * bug in the backref code, however error out for a
2667                          * normal user in the case of corruption.
2668                          */
2669                         ASSERT(node->new_bytenr == 0);
2670                         ASSERT(list_empty(&node->list));
2671                         if (node->new_bytenr || !list_empty(&node->list)) {
2672                                 btrfs_err(root->fs_info,
2673                                   "bytenr %llu has improper references to it",
2674                                           node->bytenr);
2675                                 ret = -EUCLEAN;
2676                                 goto out;
2677                         }
2678                         ret = btrfs_record_root_in_trans(trans, root);
2679                         if (ret)
2680                                 goto out;
2681                         /*
2682                          * Another thread could have failed, need to check if we
2683                          * have reloc_root actually set.
2684                          */
2685                         if (!root->reloc_root) {
2686                                 ret = -ENOENT;
2687                                 goto out;
2688                         }
2689                         root = root->reloc_root;
2690                         node->new_bytenr = root->node->start;
2691                         btrfs_put_root(node->root);
2692                         node->root = btrfs_grab_root(root);
2693                         ASSERT(node->root);
2694                         list_add_tail(&node->list, &rc->backref_cache.changed);
2695                 } else {
2696                         path->lowest_level = node->level;
2697                         if (root == root->fs_info->chunk_root)
2698                                 btrfs_reserve_chunk_metadata(trans, false);
2699                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2700                         btrfs_release_path(path);
2701                         if (root == root->fs_info->chunk_root)
2702                                 btrfs_trans_release_chunk_metadata(trans);
2703                         if (ret > 0)
2704                                 ret = 0;
2705                 }
2706                 if (!ret)
2707                         update_processed_blocks(rc, node);
2708         } else {
2709                 ret = do_relocation(trans, rc, node, key, path, 1);
2710         }
2711 out:
2712         if (ret || node->level == 0 || node->cowonly)
2713                 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2714         return ret;
2715 }
2716
2717 /*
2718  * relocate a list of blocks
2719  */
2720 static noinline_for_stack
2721 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2722                          struct reloc_control *rc, struct rb_root *blocks)
2723 {
2724         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2725         struct btrfs_backref_node *node;
2726         struct btrfs_path *path;
2727         struct tree_block *block;
2728         struct tree_block *next;
2729         int ret;
2730         int err = 0;
2731
2732         path = btrfs_alloc_path();
2733         if (!path) {
2734                 err = -ENOMEM;
2735                 goto out_free_blocks;
2736         }
2737
2738         /* Kick in readahead for tree blocks with missing keys */
2739         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2740                 if (!block->key_ready)
2741                         btrfs_readahead_tree_block(fs_info, block->bytenr,
2742                                                    block->owner, 0,
2743                                                    block->level);
2744         }
2745
2746         /* Get first keys */
2747         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2748                 if (!block->key_ready) {
2749                         err = get_tree_block_key(fs_info, block);
2750                         if (err)
2751                                 goto out_free_path;
2752                 }
2753         }
2754
2755         /* Do tree relocation */
2756         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2757                 node = build_backref_tree(rc, &block->key,
2758                                           block->level, block->bytenr);
2759                 if (IS_ERR(node)) {
2760                         err = PTR_ERR(node);
2761                         goto out;
2762                 }
2763
2764                 ret = relocate_tree_block(trans, rc, node, &block->key,
2765                                           path);
2766                 if (ret < 0) {
2767                         err = ret;
2768                         break;
2769                 }
2770         }
2771 out:
2772         err = finish_pending_nodes(trans, rc, path, err);
2773
2774 out_free_path:
2775         btrfs_free_path(path);
2776 out_free_blocks:
2777         free_block_list(blocks);
2778         return err;
2779 }
2780
2781 static noinline_for_stack int prealloc_file_extent_cluster(
2782                                 struct btrfs_inode *inode,
2783                                 struct file_extent_cluster *cluster)
2784 {
2785         u64 alloc_hint = 0;
2786         u64 start;
2787         u64 end;
2788         u64 offset = inode->index_cnt;
2789         u64 num_bytes;
2790         int nr;
2791         int ret = 0;
2792         u64 i_size = i_size_read(&inode->vfs_inode);
2793         u64 prealloc_start = cluster->start - offset;
2794         u64 prealloc_end = cluster->end - offset;
2795         u64 cur_offset = prealloc_start;
2796
2797         /*
2798          * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2799          * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2800          * btrfs_do_readpage() call of previously relocated file cluster.
2801          *
2802          * If the current cluster starts in the above range, btrfs_do_readpage()
2803          * will skip the read, and relocate_one_page() will later writeback
2804          * the padding zeros as new data, causing data corruption.
2805          *
2806          * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2807          */
2808         if (!IS_ALIGNED(i_size, PAGE_SIZE)) {
2809                 struct address_space *mapping = inode->vfs_inode.i_mapping;
2810                 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2811                 const u32 sectorsize = fs_info->sectorsize;
2812                 struct page *page;
2813
2814                 ASSERT(sectorsize < PAGE_SIZE);
2815                 ASSERT(IS_ALIGNED(i_size, sectorsize));
2816
2817                 /*
2818                  * Subpage can't handle page with DIRTY but without UPTODATE
2819                  * bit as it can lead to the following deadlock:
2820                  *
2821                  * btrfs_read_folio()
2822                  * | Page already *locked*
2823                  * |- btrfs_lock_and_flush_ordered_range()
2824                  *    |- btrfs_start_ordered_extent()
2825                  *       |- extent_write_cache_pages()
2826                  *          |- lock_page()
2827                  *             We try to lock the page we already hold.
2828                  *
2829                  * Here we just writeback the whole data reloc inode, so that
2830                  * we will be ensured to have no dirty range in the page, and
2831                  * are safe to clear the uptodate bits.
2832                  *
2833                  * This shouldn't cause too much overhead, as we need to write
2834                  * the data back anyway.
2835                  */
2836                 ret = filemap_write_and_wait(mapping);
2837                 if (ret < 0)
2838                         return ret;
2839
2840                 clear_extent_bits(&inode->io_tree, i_size,
2841                                   round_up(i_size, PAGE_SIZE) - 1,
2842                                   EXTENT_UPTODATE);
2843                 page = find_lock_page(mapping, i_size >> PAGE_SHIFT);
2844                 /*
2845                  * If page is freed we don't need to do anything then, as we
2846                  * will re-read the whole page anyway.
2847                  */
2848                 if (page) {
2849                         btrfs_subpage_clear_uptodate(fs_info, page, i_size,
2850                                         round_up(i_size, PAGE_SIZE) - i_size);
2851                         unlock_page(page);
2852                         put_page(page);
2853                 }
2854         }
2855
2856         BUG_ON(cluster->start != cluster->boundary[0]);
2857         ret = btrfs_alloc_data_chunk_ondemand(inode,
2858                                               prealloc_end + 1 - prealloc_start);
2859         if (ret)
2860                 return ret;
2861
2862         btrfs_inode_lock(&inode->vfs_inode, 0);
2863         for (nr = 0; nr < cluster->nr; nr++) {
2864                 start = cluster->boundary[nr] - offset;
2865                 if (nr + 1 < cluster->nr)
2866                         end = cluster->boundary[nr + 1] - 1 - offset;
2867                 else
2868                         end = cluster->end - offset;
2869
2870                 lock_extent(&inode->io_tree, start, end);
2871                 num_bytes = end + 1 - start;
2872                 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2873                                                 num_bytes, num_bytes,
2874                                                 end + 1, &alloc_hint);
2875                 cur_offset = end + 1;
2876                 unlock_extent(&inode->io_tree, start, end);
2877                 if (ret)
2878                         break;
2879         }
2880         btrfs_inode_unlock(&inode->vfs_inode, 0);
2881
2882         if (cur_offset < prealloc_end)
2883                 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2884                                                prealloc_end + 1 - cur_offset);
2885         return ret;
2886 }
2887
2888 static noinline_for_stack int setup_relocation_extent_mapping(struct inode *inode,
2889                                 u64 start, u64 end, u64 block_start)
2890 {
2891         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2892         struct extent_map *em;
2893         int ret = 0;
2894
2895         em = alloc_extent_map();
2896         if (!em)
2897                 return -ENOMEM;
2898
2899         em->start = start;
2900         em->len = end + 1 - start;
2901         em->block_len = em->len;
2902         em->block_start = block_start;
2903         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2904
2905         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2906         while (1) {
2907                 write_lock(&em_tree->lock);
2908                 ret = add_extent_mapping(em_tree, em, 0);
2909                 write_unlock(&em_tree->lock);
2910                 if (ret != -EEXIST) {
2911                         free_extent_map(em);
2912                         break;
2913                 }
2914                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
2915         }
2916         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2917         return ret;
2918 }
2919
2920 /*
2921  * Allow error injection to test balance/relocation cancellation
2922  */
2923 noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2924 {
2925         return atomic_read(&fs_info->balance_cancel_req) ||
2926                 atomic_read(&fs_info->reloc_cancel_req) ||
2927                 fatal_signal_pending(current);
2928 }
2929 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2930
2931 static u64 get_cluster_boundary_end(struct file_extent_cluster *cluster,
2932                                     int cluster_nr)
2933 {
2934         /* Last extent, use cluster end directly */
2935         if (cluster_nr >= cluster->nr - 1)
2936                 return cluster->end;
2937
2938         /* Use next boundary start*/
2939         return cluster->boundary[cluster_nr + 1] - 1;
2940 }
2941
2942 static int relocate_one_page(struct inode *inode, struct file_ra_state *ra,
2943                              struct file_extent_cluster *cluster,
2944                              int *cluster_nr, unsigned long page_index)
2945 {
2946         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2947         u64 offset = BTRFS_I(inode)->index_cnt;
2948         const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2949         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2950         struct page *page;
2951         u64 page_start;
2952         u64 page_end;
2953         u64 cur;
2954         int ret;
2955
2956         ASSERT(page_index <= last_index);
2957         page = find_lock_page(inode->i_mapping, page_index);
2958         if (!page) {
2959                 page_cache_sync_readahead(inode->i_mapping, ra, NULL,
2960                                 page_index, last_index + 1 - page_index);
2961                 page = find_or_create_page(inode->i_mapping, page_index, mask);
2962                 if (!page)
2963                         return -ENOMEM;
2964         }
2965         ret = set_page_extent_mapped(page);
2966         if (ret < 0)
2967                 goto release_page;
2968
2969         if (PageReadahead(page))
2970                 page_cache_async_readahead(inode->i_mapping, ra, NULL,
2971                                 page_folio(page), page_index,
2972                                 last_index + 1 - page_index);
2973
2974         if (!PageUptodate(page)) {
2975                 btrfs_read_folio(NULL, page_folio(page));
2976                 lock_page(page);
2977                 if (!PageUptodate(page)) {
2978                         ret = -EIO;
2979                         goto release_page;
2980                 }
2981         }
2982
2983         page_start = page_offset(page);
2984         page_end = page_start + PAGE_SIZE - 1;
2985
2986         /*
2987          * Start from the cluster, as for subpage case, the cluster can start
2988          * inside the page.
2989          */
2990         cur = max(page_start, cluster->boundary[*cluster_nr] - offset);
2991         while (cur <= page_end) {
2992                 u64 extent_start = cluster->boundary[*cluster_nr] - offset;
2993                 u64 extent_end = get_cluster_boundary_end(cluster,
2994                                                 *cluster_nr) - offset;
2995                 u64 clamped_start = max(page_start, extent_start);
2996                 u64 clamped_end = min(page_end, extent_end);
2997                 u32 clamped_len = clamped_end + 1 - clamped_start;
2998
2999                 /* Reserve metadata for this range */
3000                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3001                                                       clamped_len, clamped_len,
3002                                                       false);
3003                 if (ret)
3004                         goto release_page;
3005
3006                 /* Mark the range delalloc and dirty for later writeback */
3007                 lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end);
3008                 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
3009                                                 clamped_end, 0, NULL);
3010                 if (ret) {
3011                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
3012                                         clamped_start, clamped_end,
3013                                         EXTENT_LOCKED | EXTENT_BOUNDARY);
3014                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3015                                                         clamped_len, true);
3016                         btrfs_delalloc_release_extents(BTRFS_I(inode),
3017                                                        clamped_len);
3018                         goto release_page;
3019                 }
3020                 btrfs_page_set_dirty(fs_info, page, clamped_start, clamped_len);
3021
3022                 /*
3023                  * Set the boundary if it's inside the page.
3024                  * Data relocation requires the destination extents to have the
3025                  * same size as the source.
3026                  * EXTENT_BOUNDARY bit prevents current extent from being merged
3027                  * with previous extent.
3028                  */
3029                 if (in_range(cluster->boundary[*cluster_nr] - offset,
3030                              page_start, PAGE_SIZE)) {
3031                         u64 boundary_start = cluster->boundary[*cluster_nr] -
3032                                                 offset;
3033                         u64 boundary_end = boundary_start +
3034                                            fs_info->sectorsize - 1;
3035
3036                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3037                                         boundary_start, boundary_end,
3038                                         EXTENT_BOUNDARY);
3039                 }
3040                 unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end);
3041                 btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3042                 cur += clamped_len;
3043
3044                 /* Crossed extent end, go to next extent */
3045                 if (cur >= extent_end) {
3046                         (*cluster_nr)++;
3047                         /* Just finished the last extent of the cluster, exit. */
3048                         if (*cluster_nr >= cluster->nr)
3049                                 break;
3050                 }
3051         }
3052         unlock_page(page);
3053         put_page(page);
3054
3055         balance_dirty_pages_ratelimited(inode->i_mapping);
3056         btrfs_throttle(fs_info);
3057         if (btrfs_should_cancel_balance(fs_info))
3058                 ret = -ECANCELED;
3059         return ret;
3060
3061 release_page:
3062         unlock_page(page);
3063         put_page(page);
3064         return ret;
3065 }
3066
3067 static int relocate_file_extent_cluster(struct inode *inode,
3068                                         struct file_extent_cluster *cluster)
3069 {
3070         u64 offset = BTRFS_I(inode)->index_cnt;
3071         unsigned long index;
3072         unsigned long last_index;
3073         struct file_ra_state *ra;
3074         int cluster_nr = 0;
3075         int ret = 0;
3076
3077         if (!cluster->nr)
3078                 return 0;
3079
3080         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3081         if (!ra)
3082                 return -ENOMEM;
3083
3084         ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
3085         if (ret)
3086                 goto out;
3087
3088         file_ra_state_init(ra, inode->i_mapping);
3089
3090         ret = setup_relocation_extent_mapping(inode, cluster->start - offset,
3091                                    cluster->end - offset, cluster->start);
3092         if (ret)
3093                 goto out;
3094
3095         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3096         for (index = (cluster->start - offset) >> PAGE_SHIFT;
3097              index <= last_index && !ret; index++)
3098                 ret = relocate_one_page(inode, ra, cluster, &cluster_nr, index);
3099         if (ret == 0)
3100                 WARN_ON(cluster_nr != cluster->nr);
3101 out:
3102         kfree(ra);
3103         return ret;
3104 }
3105
3106 static noinline_for_stack
3107 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3108                          struct file_extent_cluster *cluster)
3109 {
3110         int ret;
3111
3112         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3113                 ret = relocate_file_extent_cluster(inode, cluster);
3114                 if (ret)
3115                         return ret;
3116                 cluster->nr = 0;
3117         }
3118
3119         if (!cluster->nr)
3120                 cluster->start = extent_key->objectid;
3121         else
3122                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3123         cluster->end = extent_key->objectid + extent_key->offset - 1;
3124         cluster->boundary[cluster->nr] = extent_key->objectid;
3125         cluster->nr++;
3126
3127         if (cluster->nr >= MAX_EXTENTS) {
3128                 ret = relocate_file_extent_cluster(inode, cluster);
3129                 if (ret)
3130                         return ret;
3131                 cluster->nr = 0;
3132         }
3133         return 0;
3134 }
3135
3136 /*
3137  * helper to add a tree block to the list.
3138  * the major work is getting the generation and level of the block
3139  */
3140 static int add_tree_block(struct reloc_control *rc,
3141                           struct btrfs_key *extent_key,
3142                           struct btrfs_path *path,
3143                           struct rb_root *blocks)
3144 {
3145         struct extent_buffer *eb;
3146         struct btrfs_extent_item *ei;
3147         struct btrfs_tree_block_info *bi;
3148         struct tree_block *block;
3149         struct rb_node *rb_node;
3150         u32 item_size;
3151         int level = -1;
3152         u64 generation;
3153         u64 owner = 0;
3154
3155         eb =  path->nodes[0];
3156         item_size = btrfs_item_size(eb, path->slots[0]);
3157
3158         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3159             item_size >= sizeof(*ei) + sizeof(*bi)) {
3160                 unsigned long ptr = 0, end;
3161
3162                 ei = btrfs_item_ptr(eb, path->slots[0],
3163                                 struct btrfs_extent_item);
3164                 end = (unsigned long)ei + item_size;
3165                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3166                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3167                         level = btrfs_tree_block_level(eb, bi);
3168                         ptr = (unsigned long)(bi + 1);
3169                 } else {
3170                         level = (int)extent_key->offset;
3171                         ptr = (unsigned long)(ei + 1);
3172                 }
3173                 generation = btrfs_extent_generation(eb, ei);
3174
3175                 /*
3176                  * We're reading random blocks without knowing their owner ahead
3177                  * of time.  This is ok most of the time, as all reloc roots and
3178                  * fs roots have the same lock type.  However normal trees do
3179                  * not, and the only way to know ahead of time is to read the
3180                  * inline ref offset.  We know it's an fs root if
3181                  *
3182                  * 1. There's more than one ref.
3183                  * 2. There's a SHARED_DATA_REF_KEY set.
3184                  * 3. FULL_BACKREF is set on the flags.
3185                  *
3186                  * Otherwise it's safe to assume that the ref offset == the
3187                  * owner of this block, so we can use that when calling
3188                  * read_tree_block.
3189                  */
3190                 if (btrfs_extent_refs(eb, ei) == 1 &&
3191                     !(btrfs_extent_flags(eb, ei) &
3192                       BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3193                     ptr < end) {
3194                         struct btrfs_extent_inline_ref *iref;
3195                         int type;
3196
3197                         iref = (struct btrfs_extent_inline_ref *)ptr;
3198                         type = btrfs_get_extent_inline_ref_type(eb, iref,
3199                                                         BTRFS_REF_TYPE_BLOCK);
3200                         if (type == BTRFS_REF_TYPE_INVALID)
3201                                 return -EINVAL;
3202                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
3203                                 owner = btrfs_extent_inline_ref_offset(eb, iref);
3204                 }
3205         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3206                 btrfs_print_v0_err(eb->fs_info);
3207                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3208                 return -EINVAL;
3209         } else {
3210                 BUG();
3211         }
3212
3213         btrfs_release_path(path);
3214
3215         BUG_ON(level == -1);
3216
3217         block = kmalloc(sizeof(*block), GFP_NOFS);
3218         if (!block)
3219                 return -ENOMEM;
3220
3221         block->bytenr = extent_key->objectid;
3222         block->key.objectid = rc->extent_root->fs_info->nodesize;
3223         block->key.offset = generation;
3224         block->level = level;
3225         block->key_ready = 0;
3226         block->owner = owner;
3227
3228         rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3229         if (rb_node)
3230                 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3231                                     -EEXIST);
3232
3233         return 0;
3234 }
3235
3236 /*
3237  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3238  */
3239 static int __add_tree_block(struct reloc_control *rc,
3240                             u64 bytenr, u32 blocksize,
3241                             struct rb_root *blocks)
3242 {
3243         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3244         struct btrfs_path *path;
3245         struct btrfs_key key;
3246         int ret;
3247         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3248
3249         if (tree_block_processed(bytenr, rc))
3250                 return 0;
3251
3252         if (rb_simple_search(blocks, bytenr))
3253                 return 0;
3254
3255         path = btrfs_alloc_path();
3256         if (!path)
3257                 return -ENOMEM;
3258 again:
3259         key.objectid = bytenr;
3260         if (skinny) {
3261                 key.type = BTRFS_METADATA_ITEM_KEY;
3262                 key.offset = (u64)-1;
3263         } else {
3264                 key.type = BTRFS_EXTENT_ITEM_KEY;
3265                 key.offset = blocksize;
3266         }
3267
3268         path->search_commit_root = 1;
3269         path->skip_locking = 1;
3270         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3271         if (ret < 0)
3272                 goto out;
3273
3274         if (ret > 0 && skinny) {
3275                 if (path->slots[0]) {
3276                         path->slots[0]--;
3277                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3278                                               path->slots[0]);
3279                         if (key.objectid == bytenr &&
3280                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3281                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3282                               key.offset == blocksize)))
3283                                 ret = 0;
3284                 }
3285
3286                 if (ret) {
3287                         skinny = false;
3288                         btrfs_release_path(path);
3289                         goto again;
3290                 }
3291         }
3292         if (ret) {
3293                 ASSERT(ret == 1);
3294                 btrfs_print_leaf(path->nodes[0]);
3295                 btrfs_err(fs_info,
3296              "tree block extent item (%llu) is not found in extent tree",
3297                      bytenr);
3298                 WARN_ON(1);
3299                 ret = -EINVAL;
3300                 goto out;
3301         }
3302
3303         ret = add_tree_block(rc, &key, path, blocks);
3304 out:
3305         btrfs_free_path(path);
3306         return ret;
3307 }
3308
3309 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3310                                     struct btrfs_block_group *block_group,
3311                                     struct inode *inode,
3312                                     u64 ino)
3313 {
3314         struct btrfs_root *root = fs_info->tree_root;
3315         struct btrfs_trans_handle *trans;
3316         int ret = 0;
3317
3318         if (inode)
3319                 goto truncate;
3320
3321         inode = btrfs_iget(fs_info->sb, ino, root);
3322         if (IS_ERR(inode))
3323                 return -ENOENT;
3324
3325 truncate:
3326         ret = btrfs_check_trunc_cache_free_space(fs_info,
3327                                                  &fs_info->global_block_rsv);
3328         if (ret)
3329                 goto out;
3330
3331         trans = btrfs_join_transaction(root);
3332         if (IS_ERR(trans)) {
3333                 ret = PTR_ERR(trans);
3334                 goto out;
3335         }
3336
3337         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3338
3339         btrfs_end_transaction(trans);
3340         btrfs_btree_balance_dirty(fs_info);
3341 out:
3342         iput(inode);
3343         return ret;
3344 }
3345
3346 /*
3347  * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3348  * cache inode, to avoid free space cache data extent blocking data relocation.
3349  */
3350 static int delete_v1_space_cache(struct extent_buffer *leaf,
3351                                  struct btrfs_block_group *block_group,
3352                                  u64 data_bytenr)
3353 {
3354         u64 space_cache_ino;
3355         struct btrfs_file_extent_item *ei;
3356         struct btrfs_key key;
3357         bool found = false;
3358         int i;
3359         int ret;
3360
3361         if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3362                 return 0;
3363
3364         for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3365                 u8 type;
3366
3367                 btrfs_item_key_to_cpu(leaf, &key, i);
3368                 if (key.type != BTRFS_EXTENT_DATA_KEY)
3369                         continue;
3370                 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3371                 type = btrfs_file_extent_type(leaf, ei);
3372
3373                 if ((type == BTRFS_FILE_EXTENT_REG ||
3374                      type == BTRFS_FILE_EXTENT_PREALLOC) &&
3375                     btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3376                         found = true;
3377                         space_cache_ino = key.objectid;
3378                         break;
3379                 }
3380         }
3381         if (!found)
3382                 return -ENOENT;
3383         ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3384                                         space_cache_ino);
3385         return ret;
3386 }
3387
3388 /*
3389  * helper to find all tree blocks that reference a given data extent
3390  */
3391 static noinline_for_stack
3392 int add_data_references(struct reloc_control *rc,
3393                         struct btrfs_key *extent_key,
3394                         struct btrfs_path *path,
3395                         struct rb_root *blocks)
3396 {
3397         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3398         struct ulist *leaves = NULL;
3399         struct ulist_iterator leaf_uiter;
3400         struct ulist_node *ref_node = NULL;
3401         const u32 blocksize = fs_info->nodesize;
3402         int ret = 0;
3403
3404         btrfs_release_path(path);
3405         ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid,
3406                                    0, &leaves, NULL, true);
3407         if (ret < 0)
3408                 return ret;
3409
3410         ULIST_ITER_INIT(&leaf_uiter);
3411         while ((ref_node = ulist_next(leaves, &leaf_uiter))) {
3412                 struct extent_buffer *eb;
3413
3414                 eb = read_tree_block(fs_info, ref_node->val, 0, 0, 0, NULL);
3415                 if (IS_ERR(eb)) {
3416                         ret = PTR_ERR(eb);
3417                         break;
3418                 }
3419                 ret = delete_v1_space_cache(eb, rc->block_group,
3420                                             extent_key->objectid);
3421                 free_extent_buffer(eb);
3422                 if (ret < 0)
3423                         break;
3424                 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3425                 if (ret < 0)
3426                         break;
3427         }
3428         if (ret < 0)
3429                 free_block_list(blocks);
3430         ulist_free(leaves);
3431         return ret;
3432 }
3433
3434 /*
3435  * helper to find next unprocessed extent
3436  */
3437 static noinline_for_stack
3438 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3439                      struct btrfs_key *extent_key)
3440 {
3441         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3442         struct btrfs_key key;
3443         struct extent_buffer *leaf;
3444         u64 start, end, last;
3445         int ret;
3446
3447         last = rc->block_group->start + rc->block_group->length;
3448         while (1) {
3449                 cond_resched();
3450                 if (rc->search_start >= last) {
3451                         ret = 1;
3452                         break;
3453                 }
3454
3455                 key.objectid = rc->search_start;
3456                 key.type = BTRFS_EXTENT_ITEM_KEY;
3457                 key.offset = 0;
3458
3459                 path->search_commit_root = 1;
3460                 path->skip_locking = 1;
3461                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3462                                         0, 0);
3463                 if (ret < 0)
3464                         break;
3465 next:
3466                 leaf = path->nodes[0];
3467                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3468                         ret = btrfs_next_leaf(rc->extent_root, path);
3469                         if (ret != 0)
3470                                 break;
3471                         leaf = path->nodes[0];
3472                 }
3473
3474                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3475                 if (key.objectid >= last) {
3476                         ret = 1;
3477                         break;
3478                 }
3479
3480                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3481                     key.type != BTRFS_METADATA_ITEM_KEY) {
3482                         path->slots[0]++;
3483                         goto next;
3484                 }
3485
3486                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3487                     key.objectid + key.offset <= rc->search_start) {
3488                         path->slots[0]++;
3489                         goto next;
3490                 }
3491
3492                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3493                     key.objectid + fs_info->nodesize <=
3494                     rc->search_start) {
3495                         path->slots[0]++;
3496                         goto next;
3497                 }
3498
3499                 ret = find_first_extent_bit(&rc->processed_blocks,
3500                                             key.objectid, &start, &end,
3501                                             EXTENT_DIRTY, NULL);
3502
3503                 if (ret == 0 && start <= key.objectid) {
3504                         btrfs_release_path(path);
3505                         rc->search_start = end + 1;
3506                 } else {
3507                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3508                                 rc->search_start = key.objectid + key.offset;
3509                         else
3510                                 rc->search_start = key.objectid +
3511                                         fs_info->nodesize;
3512                         memcpy(extent_key, &key, sizeof(key));
3513                         return 0;
3514                 }
3515         }
3516         btrfs_release_path(path);
3517         return ret;
3518 }
3519
3520 static void set_reloc_control(struct reloc_control *rc)
3521 {
3522         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3523
3524         mutex_lock(&fs_info->reloc_mutex);
3525         fs_info->reloc_ctl = rc;
3526         mutex_unlock(&fs_info->reloc_mutex);
3527 }
3528
3529 static void unset_reloc_control(struct reloc_control *rc)
3530 {
3531         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3532
3533         mutex_lock(&fs_info->reloc_mutex);
3534         fs_info->reloc_ctl = NULL;
3535         mutex_unlock(&fs_info->reloc_mutex);
3536 }
3537
3538 static noinline_for_stack
3539 int prepare_to_relocate(struct reloc_control *rc)
3540 {
3541         struct btrfs_trans_handle *trans;
3542         int ret;
3543
3544         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3545                                               BTRFS_BLOCK_RSV_TEMP);
3546         if (!rc->block_rsv)
3547                 return -ENOMEM;
3548
3549         memset(&rc->cluster, 0, sizeof(rc->cluster));
3550         rc->search_start = rc->block_group->start;
3551         rc->extents_found = 0;
3552         rc->nodes_relocated = 0;
3553         rc->merging_rsv_size = 0;
3554         rc->reserved_bytes = 0;
3555         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3556                               RELOCATION_RESERVED_NODES;
3557         ret = btrfs_block_rsv_refill(rc->extent_root->fs_info,
3558                                      rc->block_rsv, rc->block_rsv->size,
3559                                      BTRFS_RESERVE_FLUSH_ALL);
3560         if (ret)
3561                 return ret;
3562
3563         rc->create_reloc_tree = 1;
3564         set_reloc_control(rc);
3565
3566         trans = btrfs_join_transaction(rc->extent_root);
3567         if (IS_ERR(trans)) {
3568                 unset_reloc_control(rc);
3569                 /*
3570                  * extent tree is not a ref_cow tree and has no reloc_root to
3571                  * cleanup.  And callers are responsible to free the above
3572                  * block rsv.
3573                  */
3574                 return PTR_ERR(trans);
3575         }
3576         return btrfs_commit_transaction(trans);
3577 }
3578
3579 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3580 {
3581         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3582         struct rb_root blocks = RB_ROOT;
3583         struct btrfs_key key;
3584         struct btrfs_trans_handle *trans = NULL;
3585         struct btrfs_path *path;
3586         struct btrfs_extent_item *ei;
3587         u64 flags;
3588         int ret;
3589         int err = 0;
3590         int progress = 0;
3591
3592         path = btrfs_alloc_path();
3593         if (!path)
3594                 return -ENOMEM;
3595         path->reada = READA_FORWARD;
3596
3597         ret = prepare_to_relocate(rc);
3598         if (ret) {
3599                 err = ret;
3600                 goto out_free;
3601         }
3602
3603         while (1) {
3604                 rc->reserved_bytes = 0;
3605                 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
3606                                              rc->block_rsv->size,
3607                                              BTRFS_RESERVE_FLUSH_ALL);
3608                 if (ret) {
3609                         err = ret;
3610                         break;
3611                 }
3612                 progress++;
3613                 trans = btrfs_start_transaction(rc->extent_root, 0);
3614                 if (IS_ERR(trans)) {
3615                         err = PTR_ERR(trans);
3616                         trans = NULL;
3617                         break;
3618                 }
3619 restart:
3620                 if (update_backref_cache(trans, &rc->backref_cache)) {
3621                         btrfs_end_transaction(trans);
3622                         trans = NULL;
3623                         continue;
3624                 }
3625
3626                 ret = find_next_extent(rc, path, &key);
3627                 if (ret < 0)
3628                         err = ret;
3629                 if (ret != 0)
3630                         break;
3631
3632                 rc->extents_found++;
3633
3634                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3635                                     struct btrfs_extent_item);
3636                 flags = btrfs_extent_flags(path->nodes[0], ei);
3637
3638                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3639                         ret = add_tree_block(rc, &key, path, &blocks);
3640                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3641                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3642                         ret = add_data_references(rc, &key, path, &blocks);
3643                 } else {
3644                         btrfs_release_path(path);
3645                         ret = 0;
3646                 }
3647                 if (ret < 0) {
3648                         err = ret;
3649                         break;
3650                 }
3651
3652                 if (!RB_EMPTY_ROOT(&blocks)) {
3653                         ret = relocate_tree_blocks(trans, rc, &blocks);
3654                         if (ret < 0) {
3655                                 if (ret != -EAGAIN) {
3656                                         err = ret;
3657                                         break;
3658                                 }
3659                                 rc->extents_found--;
3660                                 rc->search_start = key.objectid;
3661                         }
3662                 }
3663
3664                 btrfs_end_transaction_throttle(trans);
3665                 btrfs_btree_balance_dirty(fs_info);
3666                 trans = NULL;
3667
3668                 if (rc->stage == MOVE_DATA_EXTENTS &&
3669                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3670                         rc->found_file_extent = 1;
3671                         ret = relocate_data_extent(rc->data_inode,
3672                                                    &key, &rc->cluster);
3673                         if (ret < 0) {
3674                                 err = ret;
3675                                 break;
3676                         }
3677                 }
3678                 if (btrfs_should_cancel_balance(fs_info)) {
3679                         err = -ECANCELED;
3680                         break;
3681                 }
3682         }
3683         if (trans && progress && err == -ENOSPC) {
3684                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3685                 if (ret == 1) {
3686                         err = 0;
3687                         progress = 0;
3688                         goto restart;
3689                 }
3690         }
3691
3692         btrfs_release_path(path);
3693         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3694
3695         if (trans) {
3696                 btrfs_end_transaction_throttle(trans);
3697                 btrfs_btree_balance_dirty(fs_info);
3698         }
3699
3700         if (!err) {
3701                 ret = relocate_file_extent_cluster(rc->data_inode,
3702                                                    &rc->cluster);
3703                 if (ret < 0)
3704                         err = ret;
3705         }
3706
3707         rc->create_reloc_tree = 0;
3708         set_reloc_control(rc);
3709
3710         btrfs_backref_release_cache(&rc->backref_cache);
3711         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3712
3713         /*
3714          * Even in the case when the relocation is cancelled, we should all go
3715          * through prepare_to_merge() and merge_reloc_roots().
3716          *
3717          * For error (including cancelled balance), prepare_to_merge() will
3718          * mark all reloc trees orphan, then queue them for cleanup in
3719          * merge_reloc_roots()
3720          */
3721         err = prepare_to_merge(rc, err);
3722
3723         merge_reloc_roots(rc);
3724
3725         rc->merge_reloc_tree = 0;
3726         unset_reloc_control(rc);
3727         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3728
3729         /* get rid of pinned extents */
3730         trans = btrfs_join_transaction(rc->extent_root);
3731         if (IS_ERR(trans)) {
3732                 err = PTR_ERR(trans);
3733                 goto out_free;
3734         }
3735         ret = btrfs_commit_transaction(trans);
3736         if (ret && !err)
3737                 err = ret;
3738 out_free:
3739         ret = clean_dirty_subvols(rc);
3740         if (ret < 0 && !err)
3741                 err = ret;
3742         btrfs_free_block_rsv(fs_info, rc->block_rsv);
3743         btrfs_free_path(path);
3744         return err;
3745 }
3746
3747 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3748                                  struct btrfs_root *root, u64 objectid)
3749 {
3750         struct btrfs_path *path;
3751         struct btrfs_inode_item *item;
3752         struct extent_buffer *leaf;
3753         int ret;
3754
3755         path = btrfs_alloc_path();
3756         if (!path)
3757                 return -ENOMEM;
3758
3759         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3760         if (ret)
3761                 goto out;
3762
3763         leaf = path->nodes[0];
3764         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3765         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3766         btrfs_set_inode_generation(leaf, item, 1);
3767         btrfs_set_inode_size(leaf, item, 0);
3768         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3769         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3770                                           BTRFS_INODE_PREALLOC);
3771         btrfs_mark_buffer_dirty(leaf);
3772 out:
3773         btrfs_free_path(path);
3774         return ret;
3775 }
3776
3777 static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3778                                 struct btrfs_root *root, u64 objectid)
3779 {
3780         struct btrfs_path *path;
3781         struct btrfs_key key;
3782         int ret = 0;
3783
3784         path = btrfs_alloc_path();
3785         if (!path) {
3786                 ret = -ENOMEM;
3787                 goto out;
3788         }
3789
3790         key.objectid = objectid;
3791         key.type = BTRFS_INODE_ITEM_KEY;
3792         key.offset = 0;
3793         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3794         if (ret) {
3795                 if (ret > 0)
3796                         ret = -ENOENT;
3797                 goto out;
3798         }
3799         ret = btrfs_del_item(trans, root, path);
3800 out:
3801         if (ret)
3802                 btrfs_abort_transaction(trans, ret);
3803         btrfs_free_path(path);
3804 }
3805
3806 /*
3807  * helper to create inode for data relocation.
3808  * the inode is in data relocation tree and its link count is 0
3809  */
3810 static noinline_for_stack
3811 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3812                                  struct btrfs_block_group *group)
3813 {
3814         struct inode *inode = NULL;
3815         struct btrfs_trans_handle *trans;
3816         struct btrfs_root *root;
3817         u64 objectid;
3818         int err = 0;
3819
3820         root = btrfs_grab_root(fs_info->data_reloc_root);
3821         trans = btrfs_start_transaction(root, 6);
3822         if (IS_ERR(trans)) {
3823                 btrfs_put_root(root);
3824                 return ERR_CAST(trans);
3825         }
3826
3827         err = btrfs_get_free_objectid(root, &objectid);
3828         if (err)
3829                 goto out;
3830
3831         err = __insert_orphan_inode(trans, root, objectid);
3832         if (err)
3833                 goto out;
3834
3835         inode = btrfs_iget(fs_info->sb, objectid, root);
3836         if (IS_ERR(inode)) {
3837                 delete_orphan_inode(trans, root, objectid);
3838                 err = PTR_ERR(inode);
3839                 inode = NULL;
3840                 goto out;
3841         }
3842         BTRFS_I(inode)->index_cnt = group->start;
3843
3844         err = btrfs_orphan_add(trans, BTRFS_I(inode));
3845 out:
3846         btrfs_put_root(root);
3847         btrfs_end_transaction(trans);
3848         btrfs_btree_balance_dirty(fs_info);
3849         if (err) {
3850                 iput(inode);
3851                 inode = ERR_PTR(err);
3852         }
3853         return inode;
3854 }
3855
3856 /*
3857  * Mark start of chunk relocation that is cancellable. Check if the cancellation
3858  * has been requested meanwhile and don't start in that case.
3859  *
3860  * Return:
3861  *   0             success
3862  *   -EINPROGRESS  operation is already in progress, that's probably a bug
3863  *   -ECANCELED    cancellation request was set before the operation started
3864  */
3865 static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3866 {
3867         if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3868                 /* This should not happen */
3869                 btrfs_err(fs_info, "reloc already running, cannot start");
3870                 return -EINPROGRESS;
3871         }
3872
3873         if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3874                 btrfs_info(fs_info, "chunk relocation canceled on start");
3875                 /*
3876                  * On cancel, clear all requests but let the caller mark
3877                  * the end after cleanup operations.
3878                  */
3879                 atomic_set(&fs_info->reloc_cancel_req, 0);
3880                 return -ECANCELED;
3881         }
3882         return 0;
3883 }
3884
3885 /*
3886  * Mark end of chunk relocation that is cancellable and wake any waiters.
3887  */
3888 static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
3889 {
3890         /* Requested after start, clear bit first so any waiters can continue */
3891         if (atomic_read(&fs_info->reloc_cancel_req) > 0)
3892                 btrfs_info(fs_info, "chunk relocation canceled during operation");
3893         clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
3894         atomic_set(&fs_info->reloc_cancel_req, 0);
3895 }
3896
3897 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3898 {
3899         struct reloc_control *rc;
3900
3901         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3902         if (!rc)
3903                 return NULL;
3904
3905         INIT_LIST_HEAD(&rc->reloc_roots);
3906         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3907         btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3908         mapping_tree_init(&rc->reloc_root_tree);
3909         extent_io_tree_init(fs_info, &rc->processed_blocks,
3910                             IO_TREE_RELOC_BLOCKS, NULL);
3911         return rc;
3912 }
3913
3914 static void free_reloc_control(struct reloc_control *rc)
3915 {
3916         struct mapping_node *node, *tmp;
3917
3918         free_reloc_roots(&rc->reloc_roots);
3919         rbtree_postorder_for_each_entry_safe(node, tmp,
3920                         &rc->reloc_root_tree.rb_root, rb_node)
3921                 kfree(node);
3922
3923         kfree(rc);
3924 }
3925
3926 /*
3927  * Print the block group being relocated
3928  */
3929 static void describe_relocation(struct btrfs_fs_info *fs_info,
3930                                 struct btrfs_block_group *block_group)
3931 {
3932         char buf[128] = {'\0'};
3933
3934         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3935
3936         btrfs_info(fs_info,
3937                    "relocating block group %llu flags %s",
3938                    block_group->start, buf);
3939 }
3940
3941 static const char *stage_to_string(int stage)
3942 {
3943         if (stage == MOVE_DATA_EXTENTS)
3944                 return "move data extents";
3945         if (stage == UPDATE_DATA_PTRS)
3946                 return "update data pointers";
3947         return "unknown";
3948 }
3949
3950 /*
3951  * function to relocate all extents in a block group.
3952  */
3953 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3954 {
3955         struct btrfs_block_group *bg;
3956         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, group_start);
3957         struct reloc_control *rc;
3958         struct inode *inode;
3959         struct btrfs_path *path;
3960         int ret;
3961         int rw = 0;
3962         int err = 0;
3963
3964         /*
3965          * This only gets set if we had a half-deleted snapshot on mount.  We
3966          * cannot allow relocation to start while we're still trying to clean up
3967          * these pending deletions.
3968          */
3969         ret = wait_on_bit(&fs_info->flags, BTRFS_FS_UNFINISHED_DROPS, TASK_INTERRUPTIBLE);
3970         if (ret)
3971                 return ret;
3972
3973         /* We may have been woken up by close_ctree, so bail if we're closing. */
3974         if (btrfs_fs_closing(fs_info))
3975                 return -EINTR;
3976
3977         bg = btrfs_lookup_block_group(fs_info, group_start);
3978         if (!bg)
3979                 return -ENOENT;
3980
3981         /*
3982          * Relocation of a data block group creates ordered extents.  Without
3983          * sb_start_write(), we can freeze the filesystem while unfinished
3984          * ordered extents are left. Such ordered extents can cause a deadlock
3985          * e.g. when syncfs() is waiting for their completion but they can't
3986          * finish because they block when joining a transaction, due to the
3987          * fact that the freeze locks are being held in write mode.
3988          */
3989         if (bg->flags & BTRFS_BLOCK_GROUP_DATA)
3990                 ASSERT(sb_write_started(fs_info->sb));
3991
3992         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
3993                 btrfs_put_block_group(bg);
3994                 return -ETXTBSY;
3995         }
3996
3997         rc = alloc_reloc_control(fs_info);
3998         if (!rc) {
3999                 btrfs_put_block_group(bg);
4000                 return -ENOMEM;
4001         }
4002
4003         ret = reloc_chunk_start(fs_info);
4004         if (ret < 0) {
4005                 err = ret;
4006                 goto out_put_bg;
4007         }
4008
4009         rc->extent_root = extent_root;
4010         rc->block_group = bg;
4011
4012         ret = btrfs_inc_block_group_ro(rc->block_group, true);
4013         if (ret) {
4014                 err = ret;
4015                 goto out;
4016         }
4017         rw = 1;
4018
4019         path = btrfs_alloc_path();
4020         if (!path) {
4021                 err = -ENOMEM;
4022                 goto out;
4023         }
4024
4025         inode = lookup_free_space_inode(rc->block_group, path);
4026         btrfs_free_path(path);
4027
4028         if (!IS_ERR(inode))
4029                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4030         else
4031                 ret = PTR_ERR(inode);
4032
4033         if (ret && ret != -ENOENT) {
4034                 err = ret;
4035                 goto out;
4036         }
4037
4038         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4039         if (IS_ERR(rc->data_inode)) {
4040                 err = PTR_ERR(rc->data_inode);
4041                 rc->data_inode = NULL;
4042                 goto out;
4043         }
4044
4045         describe_relocation(fs_info, rc->block_group);
4046
4047         btrfs_wait_block_group_reservations(rc->block_group);
4048         btrfs_wait_nocow_writers(rc->block_group);
4049         btrfs_wait_ordered_roots(fs_info, U64_MAX,
4050                                  rc->block_group->start,
4051                                  rc->block_group->length);
4052
4053         ret = btrfs_zone_finish(rc->block_group);
4054         WARN_ON(ret && ret != -EAGAIN);
4055
4056         while (1) {
4057                 int finishes_stage;
4058
4059                 mutex_lock(&fs_info->cleaner_mutex);
4060                 ret = relocate_block_group(rc);
4061                 mutex_unlock(&fs_info->cleaner_mutex);
4062                 if (ret < 0)
4063                         err = ret;
4064
4065                 finishes_stage = rc->stage;
4066                 /*
4067                  * We may have gotten ENOSPC after we already dirtied some
4068                  * extents.  If writeout happens while we're relocating a
4069                  * different block group we could end up hitting the
4070                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4071                  * btrfs_reloc_cow_block.  Make sure we write everything out
4072                  * properly so we don't trip over this problem, and then break
4073                  * out of the loop if we hit an error.
4074                  */
4075                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4076                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4077                                                        (u64)-1);
4078                         if (ret)
4079                                 err = ret;
4080                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4081                                                  0, -1);
4082                         rc->stage = UPDATE_DATA_PTRS;
4083                 }
4084
4085                 if (err < 0)
4086                         goto out;
4087
4088                 if (rc->extents_found == 0)
4089                         break;
4090
4091                 btrfs_info(fs_info, "found %llu extents, stage: %s",
4092                            rc->extents_found, stage_to_string(finishes_stage));
4093         }
4094
4095         WARN_ON(rc->block_group->pinned > 0);
4096         WARN_ON(rc->block_group->reserved > 0);
4097         WARN_ON(rc->block_group->used > 0);
4098 out:
4099         if (err && rw)
4100                 btrfs_dec_block_group_ro(rc->block_group);
4101         iput(rc->data_inode);
4102 out_put_bg:
4103         btrfs_put_block_group(bg);
4104         reloc_chunk_end(fs_info);
4105         free_reloc_control(rc);
4106         return err;
4107 }
4108
4109 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4110 {
4111         struct btrfs_fs_info *fs_info = root->fs_info;
4112         struct btrfs_trans_handle *trans;
4113         int ret, err;
4114
4115         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4116         if (IS_ERR(trans))
4117                 return PTR_ERR(trans);
4118
4119         memset(&root->root_item.drop_progress, 0,
4120                 sizeof(root->root_item.drop_progress));
4121         btrfs_set_root_drop_level(&root->root_item, 0);
4122         btrfs_set_root_refs(&root->root_item, 0);
4123         ret = btrfs_update_root(trans, fs_info->tree_root,
4124                                 &root->root_key, &root->root_item);
4125
4126         err = btrfs_end_transaction(trans);
4127         if (err)
4128                 return err;
4129         return ret;
4130 }
4131
4132 /*
4133  * recover relocation interrupted by system crash.
4134  *
4135  * this function resumes merging reloc trees with corresponding fs trees.
4136  * this is important for keeping the sharing of tree blocks
4137  */
4138 int btrfs_recover_relocation(struct btrfs_fs_info *fs_info)
4139 {
4140         LIST_HEAD(reloc_roots);
4141         struct btrfs_key key;
4142         struct btrfs_root *fs_root;
4143         struct btrfs_root *reloc_root;
4144         struct btrfs_path *path;
4145         struct extent_buffer *leaf;
4146         struct reloc_control *rc = NULL;
4147         struct btrfs_trans_handle *trans;
4148         int ret;
4149         int err = 0;
4150
4151         path = btrfs_alloc_path();
4152         if (!path)
4153                 return -ENOMEM;
4154         path->reada = READA_BACK;
4155
4156         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4157         key.type = BTRFS_ROOT_ITEM_KEY;
4158         key.offset = (u64)-1;
4159
4160         while (1) {
4161                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4162                                         path, 0, 0);
4163                 if (ret < 0) {
4164                         err = ret;
4165                         goto out;
4166                 }
4167                 if (ret > 0) {
4168                         if (path->slots[0] == 0)
4169                                 break;
4170                         path->slots[0]--;
4171                 }
4172                 leaf = path->nodes[0];
4173                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4174                 btrfs_release_path(path);
4175
4176                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4177                     key.type != BTRFS_ROOT_ITEM_KEY)
4178                         break;
4179
4180                 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &key);
4181                 if (IS_ERR(reloc_root)) {
4182                         err = PTR_ERR(reloc_root);
4183                         goto out;
4184                 }
4185
4186                 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4187                 list_add(&reloc_root->root_list, &reloc_roots);
4188
4189                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4190                         fs_root = btrfs_get_fs_root(fs_info,
4191                                         reloc_root->root_key.offset, false);
4192                         if (IS_ERR(fs_root)) {
4193                                 ret = PTR_ERR(fs_root);
4194                                 if (ret != -ENOENT) {
4195                                         err = ret;
4196                                         goto out;
4197                                 }
4198                                 ret = mark_garbage_root(reloc_root);
4199                                 if (ret < 0) {
4200                                         err = ret;
4201                                         goto out;
4202                                 }
4203                         } else {
4204                                 btrfs_put_root(fs_root);
4205                         }
4206                 }
4207
4208                 if (key.offset == 0)
4209                         break;
4210
4211                 key.offset--;
4212         }
4213         btrfs_release_path(path);
4214
4215         if (list_empty(&reloc_roots))
4216                 goto out;
4217
4218         rc = alloc_reloc_control(fs_info);
4219         if (!rc) {
4220                 err = -ENOMEM;
4221                 goto out;
4222         }
4223
4224         ret = reloc_chunk_start(fs_info);
4225         if (ret < 0) {
4226                 err = ret;
4227                 goto out_end;
4228         }
4229
4230         rc->extent_root = btrfs_extent_root(fs_info, 0);
4231
4232         set_reloc_control(rc);
4233
4234         trans = btrfs_join_transaction(rc->extent_root);
4235         if (IS_ERR(trans)) {
4236                 err = PTR_ERR(trans);
4237                 goto out_unset;
4238         }
4239
4240         rc->merge_reloc_tree = 1;
4241
4242         while (!list_empty(&reloc_roots)) {
4243                 reloc_root = list_entry(reloc_roots.next,
4244                                         struct btrfs_root, root_list);
4245                 list_del(&reloc_root->root_list);
4246
4247                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4248                         list_add_tail(&reloc_root->root_list,
4249                                       &rc->reloc_roots);
4250                         continue;
4251                 }
4252
4253                 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4254                                             false);
4255                 if (IS_ERR(fs_root)) {
4256                         err = PTR_ERR(fs_root);
4257                         list_add_tail(&reloc_root->root_list, &reloc_roots);
4258                         btrfs_end_transaction(trans);
4259                         goto out_unset;
4260                 }
4261
4262                 err = __add_reloc_root(reloc_root);
4263                 ASSERT(err != -EEXIST);
4264                 if (err) {
4265                         list_add_tail(&reloc_root->root_list, &reloc_roots);
4266                         btrfs_put_root(fs_root);
4267                         btrfs_end_transaction(trans);
4268                         goto out_unset;
4269                 }
4270                 fs_root->reloc_root = btrfs_grab_root(reloc_root);
4271                 btrfs_put_root(fs_root);
4272         }
4273
4274         err = btrfs_commit_transaction(trans);
4275         if (err)
4276                 goto out_unset;
4277
4278         merge_reloc_roots(rc);
4279
4280         unset_reloc_control(rc);
4281
4282         trans = btrfs_join_transaction(rc->extent_root);
4283         if (IS_ERR(trans)) {
4284                 err = PTR_ERR(trans);
4285                 goto out_clean;
4286         }
4287         err = btrfs_commit_transaction(trans);
4288 out_clean:
4289         ret = clean_dirty_subvols(rc);
4290         if (ret < 0 && !err)
4291                 err = ret;
4292 out_unset:
4293         unset_reloc_control(rc);
4294 out_end:
4295         reloc_chunk_end(fs_info);
4296         free_reloc_control(rc);
4297 out:
4298         free_reloc_roots(&reloc_roots);
4299
4300         btrfs_free_path(path);
4301
4302         if (err == 0) {
4303                 /* cleanup orphan inode in data relocation tree */
4304                 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4305                 ASSERT(fs_root);
4306                 err = btrfs_orphan_cleanup(fs_root);
4307                 btrfs_put_root(fs_root);
4308         }
4309         return err;
4310 }
4311
4312 /*
4313  * helper to add ordered checksum for data relocation.
4314  *
4315  * cloning checksum properly handles the nodatasum extents.
4316  * it also saves CPU time to re-calculate the checksum.
4317  */
4318 int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
4319 {
4320         struct btrfs_fs_info *fs_info = inode->root->fs_info;
4321         struct btrfs_root *csum_root;
4322         struct btrfs_ordered_sum *sums;
4323         struct btrfs_ordered_extent *ordered;
4324         int ret;
4325         u64 disk_bytenr;
4326         u64 new_bytenr;
4327         LIST_HEAD(list);
4328
4329         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4330         BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
4331
4332         disk_bytenr = file_pos + inode->index_cnt;
4333         csum_root = btrfs_csum_root(fs_info, disk_bytenr);
4334         ret = btrfs_lookup_csums_range(csum_root, disk_bytenr,
4335                                        disk_bytenr + len - 1, &list, 0);
4336         if (ret)
4337                 goto out;
4338
4339         while (!list_empty(&list)) {
4340                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4341                 list_del_init(&sums->list);
4342
4343                 /*
4344                  * We need to offset the new_bytenr based on where the csum is.
4345                  * We need to do this because we will read in entire prealloc
4346                  * extents but we may have written to say the middle of the
4347                  * prealloc extent, so we need to make sure the csum goes with
4348                  * the right disk offset.
4349                  *
4350                  * We can do this because the data reloc inode refers strictly
4351                  * to the on disk bytes, so we don't have to worry about
4352                  * disk_len vs real len like with real inodes since it's all
4353                  * disk length.
4354                  */
4355                 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
4356                 sums->bytenr = new_bytenr;
4357
4358                 btrfs_add_ordered_sum(ordered, sums);
4359         }
4360 out:
4361         btrfs_put_ordered_extent(ordered);
4362         return ret;
4363 }
4364
4365 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4366                           struct btrfs_root *root, struct extent_buffer *buf,
4367                           struct extent_buffer *cow)
4368 {
4369         struct btrfs_fs_info *fs_info = root->fs_info;
4370         struct reloc_control *rc;
4371         struct btrfs_backref_node *node;
4372         int first_cow = 0;
4373         int level;
4374         int ret = 0;
4375
4376         rc = fs_info->reloc_ctl;
4377         if (!rc)
4378                 return 0;
4379
4380         BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4381
4382         level = btrfs_header_level(buf);
4383         if (btrfs_header_generation(buf) <=
4384             btrfs_root_last_snapshot(&root->root_item))
4385                 first_cow = 1;
4386
4387         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4388             rc->create_reloc_tree) {
4389                 WARN_ON(!first_cow && level == 0);
4390
4391                 node = rc->backref_cache.path[level];
4392                 BUG_ON(node->bytenr != buf->start &&
4393                        node->new_bytenr != buf->start);
4394
4395                 btrfs_backref_drop_node_buffer(node);
4396                 atomic_inc(&cow->refs);
4397                 node->eb = cow;
4398                 node->new_bytenr = cow->start;
4399
4400                 if (!node->pending) {
4401                         list_move_tail(&node->list,
4402                                        &rc->backref_cache.pending[level]);
4403                         node->pending = 1;
4404                 }
4405
4406                 if (first_cow)
4407                         mark_block_processed(rc, node);
4408
4409                 if (first_cow && level > 0)
4410                         rc->nodes_relocated += buf->len;
4411         }
4412
4413         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4414                 ret = replace_file_extents(trans, rc, root, cow);
4415         return ret;
4416 }
4417
4418 /*
4419  * called before creating snapshot. it calculates metadata reservation
4420  * required for relocating tree blocks in the snapshot
4421  */
4422 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4423                               u64 *bytes_to_reserve)
4424 {
4425         struct btrfs_root *root = pending->root;
4426         struct reloc_control *rc = root->fs_info->reloc_ctl;
4427
4428         if (!rc || !have_reloc_root(root))
4429                 return;
4430
4431         if (!rc->merge_reloc_tree)
4432                 return;
4433
4434         root = root->reloc_root;
4435         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4436         /*
4437          * relocation is in the stage of merging trees. the space
4438          * used by merging a reloc tree is twice the size of
4439          * relocated tree nodes in the worst case. half for cowing
4440          * the reloc tree, half for cowing the fs tree. the space
4441          * used by cowing the reloc tree will be freed after the
4442          * tree is dropped. if we create snapshot, cowing the fs
4443          * tree may use more space than it frees. so we need
4444          * reserve extra space.
4445          */
4446         *bytes_to_reserve += rc->nodes_relocated;
4447 }
4448
4449 /*
4450  * called after snapshot is created. migrate block reservation
4451  * and create reloc root for the newly created snapshot
4452  *
4453  * This is similar to btrfs_init_reloc_root(), we come out of here with two
4454  * references held on the reloc_root, one for root->reloc_root and one for
4455  * rc->reloc_roots.
4456  */
4457 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4458                                struct btrfs_pending_snapshot *pending)
4459 {
4460         struct btrfs_root *root = pending->root;
4461         struct btrfs_root *reloc_root;
4462         struct btrfs_root *new_root;
4463         struct reloc_control *rc = root->fs_info->reloc_ctl;
4464         int ret;
4465
4466         if (!rc || !have_reloc_root(root))
4467                 return 0;
4468
4469         rc = root->fs_info->reloc_ctl;
4470         rc->merging_rsv_size += rc->nodes_relocated;
4471
4472         if (rc->merge_reloc_tree) {
4473                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4474                                               rc->block_rsv,
4475                                               rc->nodes_relocated, true);
4476                 if (ret)
4477                         return ret;
4478         }
4479
4480         new_root = pending->snap;
4481         reloc_root = create_reloc_root(trans, root->reloc_root,
4482                                        new_root->root_key.objectid);
4483         if (IS_ERR(reloc_root))
4484                 return PTR_ERR(reloc_root);
4485
4486         ret = __add_reloc_root(reloc_root);
4487         ASSERT(ret != -EEXIST);
4488         if (ret) {
4489                 /* Pairs with create_reloc_root */
4490                 btrfs_put_root(reloc_root);
4491                 return ret;
4492         }
4493         new_root->reloc_root = btrfs_grab_root(reloc_root);
4494
4495         if (rc->create_reloc_tree)
4496                 ret = clone_backref_node(trans, rc, root, reloc_root);
4497         return ret;
4498 }