fs/btrfs/relocation.c

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