Linux Audio

Check our new training course

Loading...
v6.8
   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#include "space-info.h"
  31#include "fs.h"
  32#include "accessors.h"
  33#include "extent-tree.h"
  34#include "root-tree.h"
  35#include "file-item.h"
  36#include "relocation.h"
  37#include "super.h"
  38#include "tree-checker.h"
  39
  40/*
  41 * Relocation overview
  42 *
  43 * [What does relocation do]
  44 *
  45 * The objective of relocation is to relocate all extents of the target block
  46 * group to other block groups.
  47 * This is utilized by resize (shrink only), profile converting, compacting
  48 * space, or balance routine to spread chunks over devices.
  49 *
  50 * 		Before		|		After
  51 * ------------------------------------------------------------------
  52 *  BG A: 10 data extents	| BG A: deleted
  53 *  BG B:  2 data extents	| BG B: 10 data extents (2 old + 8 relocated)
  54 *  BG C:  1 extents		| BG C:  3 data extents (1 old + 2 relocated)
  55 *
  56 * [How does relocation work]
  57 *
  58 * 1.   Mark the target block group read-only
  59 *      New extents won't be allocated from the target block group.
  60 *
  61 * 2.1  Record each extent in the target block group
  62 *      To build a proper map of extents to be relocated.
  63 *
  64 * 2.2  Build data reloc tree and reloc trees
  65 *      Data reloc tree will contain an inode, recording all newly relocated
  66 *      data extents.
  67 *      There will be only one data reloc tree for one data block group.
  68 *
  69 *      Reloc tree will be a special snapshot of its source tree, containing
  70 *      relocated tree blocks.
  71 *      Each tree referring to a tree block in target block group will get its
  72 *      reloc tree built.
  73 *
  74 * 2.3  Swap source tree with its corresponding reloc tree
  75 *      Each involved tree only refers to new extents after swap.
  76 *
  77 * 3.   Cleanup reloc trees and data reloc tree.
  78 *      As old extents in the target block group are still referenced by reloc
  79 *      trees, we need to clean them up before really freeing the target block
  80 *      group.
  81 *
  82 * The main complexity is in steps 2.2 and 2.3.
  83 *
  84 * The entry point of relocation is relocate_block_group() function.
  85 */
  86
  87#define RELOCATION_RESERVED_NODES	256
  88/*
  89 * map address of tree root to tree
  90 */
  91struct mapping_node {
  92	struct {
  93		struct rb_node rb_node;
  94		u64 bytenr;
  95	}; /* Use rb_simle_node for search/insert */
  96	void *data;
  97};
  98
  99struct mapping_tree {
 100	struct rb_root rb_root;
 101	spinlock_t lock;
 102};
 103
 104/*
 105 * present a tree block to process
 106 */
 107struct tree_block {
 108	struct {
 109		struct rb_node rb_node;
 110		u64 bytenr;
 111	}; /* Use rb_simple_node for search/insert */
 112	u64 owner;
 113	struct btrfs_key key;
 114	u8 level;
 115	bool key_ready;
 116};
 117
 118#define MAX_EXTENTS 128
 119
 120struct file_extent_cluster {
 121	u64 start;
 122	u64 end;
 123	u64 boundary[MAX_EXTENTS];
 124	unsigned int nr;
 125	u64 owning_root;
 126};
 127
 128/* Stages of data relocation. */
 129enum reloc_stage {
 130	MOVE_DATA_EXTENTS,
 131	UPDATE_DATA_PTRS
 132};
 133
 134struct reloc_control {
 135	/* block group to relocate */
 136	struct btrfs_block_group *block_group;
 137	/* extent tree */
 138	struct btrfs_root *extent_root;
 139	/* inode for moving data */
 140	struct inode *data_inode;
 141
 142	struct btrfs_block_rsv *block_rsv;
 143
 144	struct btrfs_backref_cache backref_cache;
 145
 146	struct file_extent_cluster cluster;
 147	/* tree blocks have been processed */
 148	struct extent_io_tree processed_blocks;
 149	/* map start of tree root to corresponding reloc tree */
 150	struct mapping_tree reloc_root_tree;
 151	/* list of reloc trees */
 152	struct list_head reloc_roots;
 153	/* list of subvolume trees that get relocated */
 154	struct list_head dirty_subvol_roots;
 155	/* size of metadata reservation for merging reloc trees */
 156	u64 merging_rsv_size;
 157	/* size of relocated tree nodes */
 158	u64 nodes_relocated;
 159	/* reserved size for block group relocation*/
 160	u64 reserved_bytes;
 161
 162	u64 search_start;
 163	u64 extents_found;
 164
 165	enum reloc_stage stage;
 166	bool create_reloc_tree;
 167	bool merge_reloc_tree;
 168	bool found_file_extent;
 169};
 170
 171static void mark_block_processed(struct reloc_control *rc,
 172				 struct btrfs_backref_node *node)
 173{
 174	u32 blocksize;
 175
 176	if (node->level == 0 ||
 177	    in_range(node->bytenr, rc->block_group->start,
 178		     rc->block_group->length)) {
 179		blocksize = rc->extent_root->fs_info->nodesize;
 180		set_extent_bit(&rc->processed_blocks, node->bytenr,
 181			       node->bytenr + blocksize - 1, EXTENT_DIRTY, NULL);
 182	}
 183	node->processed = 1;
 184}
 185
 186/*
 187 * walk up backref nodes until reach node presents tree root
 188 */
 189static struct btrfs_backref_node *walk_up_backref(
 190		struct btrfs_backref_node *node,
 191		struct btrfs_backref_edge *edges[], int *index)
 192{
 193	struct btrfs_backref_edge *edge;
 194	int idx = *index;
 195
 196	while (!list_empty(&node->upper)) {
 197		edge = list_entry(node->upper.next,
 198				  struct btrfs_backref_edge, list[LOWER]);
 199		edges[idx++] = edge;
 200		node = edge->node[UPPER];
 201	}
 202	BUG_ON(node->detached);
 203	*index = idx;
 204	return node;
 205}
 206
 207/*
 208 * walk down backref nodes to find start of next reference path
 209 */
 210static struct btrfs_backref_node *walk_down_backref(
 211		struct btrfs_backref_edge *edges[], int *index)
 212{
 213	struct btrfs_backref_edge *edge;
 214	struct btrfs_backref_node *lower;
 215	int idx = *index;
 216
 217	while (idx > 0) {
 218		edge = edges[idx - 1];
 219		lower = edge->node[LOWER];
 220		if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 221			idx--;
 222			continue;
 223		}
 224		edge = list_entry(edge->list[LOWER].next,
 225				  struct btrfs_backref_edge, list[LOWER]);
 226		edges[idx - 1] = edge;
 227		*index = idx;
 228		return edge->node[UPPER];
 229	}
 230	*index = 0;
 231	return NULL;
 232}
 233
 234static void update_backref_node(struct btrfs_backref_cache *cache,
 235				struct btrfs_backref_node *node, u64 bytenr)
 236{
 237	struct rb_node *rb_node;
 238	rb_erase(&node->rb_node, &cache->rb_root);
 239	node->bytenr = bytenr;
 240	rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 241	if (rb_node)
 242		btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
 243}
 244
 245/*
 246 * update backref cache after a transaction commit
 247 */
 248static int update_backref_cache(struct btrfs_trans_handle *trans,
 249				struct btrfs_backref_cache *cache)
 250{
 251	struct btrfs_backref_node *node;
 252	int level = 0;
 253
 254	if (cache->last_trans == 0) {
 255		cache->last_trans = trans->transid;
 256		return 0;
 257	}
 258
 259	if (cache->last_trans == trans->transid)
 260		return 0;
 261
 262	/*
 263	 * detached nodes are used to avoid unnecessary backref
 264	 * lookup. transaction commit changes the extent tree.
 265	 * so the detached nodes are no longer useful.
 266	 */
 267	while (!list_empty(&cache->detached)) {
 268		node = list_entry(cache->detached.next,
 269				  struct btrfs_backref_node, list);
 270		btrfs_backref_cleanup_node(cache, node);
 271	}
 272
 273	while (!list_empty(&cache->changed)) {
 274		node = list_entry(cache->changed.next,
 275				  struct btrfs_backref_node, list);
 276		list_del_init(&node->list);
 277		BUG_ON(node->pending);
 278		update_backref_node(cache, node, node->new_bytenr);
 279	}
 280
 281	/*
 282	 * some nodes can be left in the pending list if there were
 283	 * errors during processing the pending nodes.
 284	 */
 285	for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 286		list_for_each_entry(node, &cache->pending[level], list) {
 287			BUG_ON(!node->pending);
 288			if (node->bytenr == node->new_bytenr)
 289				continue;
 290			update_backref_node(cache, node, node->new_bytenr);
 291		}
 292	}
 293
 294	cache->last_trans = 0;
 295	return 1;
 296}
 297
 298static bool reloc_root_is_dead(const struct btrfs_root *root)
 299{
 300	/*
 301	 * Pair with set_bit/clear_bit in clean_dirty_subvols and
 302	 * btrfs_update_reloc_root. We need to see the updated bit before
 303	 * trying to access reloc_root
 304	 */
 305	smp_rmb();
 306	if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
 307		return true;
 308	return false;
 309}
 310
 311/*
 312 * Check if this subvolume tree has valid reloc tree.
 313 *
 314 * Reloc tree after swap is considered dead, thus not considered as valid.
 315 * This is enough for most callers, as they don't distinguish dead reloc root
 316 * from no reloc root.  But btrfs_should_ignore_reloc_root() below is a
 317 * special case.
 318 */
 319static bool have_reloc_root(const struct btrfs_root *root)
 320{
 321	if (reloc_root_is_dead(root))
 322		return false;
 323	if (!root->reloc_root)
 324		return false;
 325	return true;
 326}
 327
 328bool btrfs_should_ignore_reloc_root(const struct btrfs_root *root)
 329{
 330	struct btrfs_root *reloc_root;
 331
 332	if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
 333		return false;
 334
 335	/* This root has been merged with its reloc tree, we can ignore it */
 336	if (reloc_root_is_dead(root))
 337		return true;
 338
 339	reloc_root = root->reloc_root;
 340	if (!reloc_root)
 341		return false;
 342
 343	if (btrfs_header_generation(reloc_root->commit_root) ==
 344	    root->fs_info->running_transaction->transid)
 345		return false;
 346	/*
 347	 * If there is reloc tree and it was created in previous transaction
 348	 * backref lookup can find the reloc tree, so backref node for the fs
 349	 * tree root is useless for relocation.
 350	 */
 351	return true;
 352}
 353
 354/*
 355 * find reloc tree by address of tree root
 356 */
 357struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
 358{
 359	struct reloc_control *rc = fs_info->reloc_ctl;
 360	struct rb_node *rb_node;
 361	struct mapping_node *node;
 362	struct btrfs_root *root = NULL;
 363
 364	ASSERT(rc);
 365	spin_lock(&rc->reloc_root_tree.lock);
 366	rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
 367	if (rb_node) {
 368		node = rb_entry(rb_node, struct mapping_node, rb_node);
 369		root = node->data;
 370	}
 371	spin_unlock(&rc->reloc_root_tree.lock);
 372	return btrfs_grab_root(root);
 373}
 374
 375/*
 376 * For useless nodes, do two major clean ups:
 377 *
 378 * - Cleanup the children edges and nodes
 379 *   If child node is also orphan (no parent) during cleanup, then the child
 380 *   node will also be cleaned up.
 381 *
 382 * - Freeing up leaves (level 0), keeps nodes detached
 383 *   For nodes, the node is still cached as "detached"
 384 *
 385 * Return false if @node is not in the @useless_nodes list.
 386 * Return true if @node is in the @useless_nodes list.
 387 */
 388static bool handle_useless_nodes(struct reloc_control *rc,
 389				 struct btrfs_backref_node *node)
 390{
 391	struct btrfs_backref_cache *cache = &rc->backref_cache;
 392	struct list_head *useless_node = &cache->useless_node;
 393	bool ret = false;
 394
 395	while (!list_empty(useless_node)) {
 396		struct btrfs_backref_node *cur;
 397
 398		cur = list_first_entry(useless_node, struct btrfs_backref_node,
 399				 list);
 400		list_del_init(&cur->list);
 401
 402		/* Only tree root nodes can be added to @useless_nodes */
 403		ASSERT(list_empty(&cur->upper));
 404
 405		if (cur == node)
 406			ret = true;
 407
 408		/* The node is the lowest node */
 409		if (cur->lowest) {
 410			list_del_init(&cur->lower);
 411			cur->lowest = 0;
 412		}
 413
 414		/* Cleanup the lower edges */
 415		while (!list_empty(&cur->lower)) {
 416			struct btrfs_backref_edge *edge;
 417			struct btrfs_backref_node *lower;
 418
 419			edge = list_entry(cur->lower.next,
 420					struct btrfs_backref_edge, list[UPPER]);
 421			list_del(&edge->list[UPPER]);
 422			list_del(&edge->list[LOWER]);
 423			lower = edge->node[LOWER];
 424			btrfs_backref_free_edge(cache, edge);
 425
 426			/* Child node is also orphan, queue for cleanup */
 427			if (list_empty(&lower->upper))
 428				list_add(&lower->list, useless_node);
 429		}
 430		/* Mark this block processed for relocation */
 431		mark_block_processed(rc, cur);
 432
 433		/*
 434		 * Backref nodes for tree leaves are deleted from the cache.
 435		 * Backref nodes for upper level tree blocks are left in the
 436		 * cache to avoid unnecessary backref lookup.
 437		 */
 438		if (cur->level > 0) {
 439			list_add(&cur->list, &cache->detached);
 440			cur->detached = 1;
 441		} else {
 442			rb_erase(&cur->rb_node, &cache->rb_root);
 443			btrfs_backref_free_node(cache, cur);
 444		}
 445	}
 446	return ret;
 447}
 448
 449/*
 450 * Build backref tree for a given tree block. Root of the backref tree
 451 * corresponds the tree block, leaves of the backref tree correspond roots of
 452 * b-trees that reference the tree block.
 453 *
 454 * The basic idea of this function is check backrefs of a given block to find
 455 * upper level blocks that reference the block, and then check backrefs of
 456 * these upper level blocks recursively. The recursion stops when tree root is
 457 * reached or backrefs for the block is cached.
 458 *
 459 * NOTE: if we find that backrefs for a block are cached, we know backrefs for
 460 * all upper level blocks that directly/indirectly reference the block are also
 461 * cached.
 462 */
 463static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
 464			struct btrfs_trans_handle *trans,
 465			struct reloc_control *rc, struct btrfs_key *node_key,
 466			int level, u64 bytenr)
 467{
 468	struct btrfs_backref_iter *iter;
 469	struct btrfs_backref_cache *cache = &rc->backref_cache;
 470	/* For searching parent of TREE_BLOCK_REF */
 471	struct btrfs_path *path;
 472	struct btrfs_backref_node *cur;
 473	struct btrfs_backref_node *node = NULL;
 474	struct btrfs_backref_edge *edge;
 475	int ret;
 476	int err = 0;
 477
 478	iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info);
 479	if (!iter)
 480		return ERR_PTR(-ENOMEM);
 481	path = btrfs_alloc_path();
 482	if (!path) {
 483		err = -ENOMEM;
 484		goto out;
 485	}
 486
 487	node = btrfs_backref_alloc_node(cache, bytenr, level);
 488	if (!node) {
 489		err = -ENOMEM;
 490		goto out;
 491	}
 492
 493	node->lowest = 1;
 494	cur = node;
 495
 496	/* Breadth-first search to build backref cache */
 497	do {
 498		ret = btrfs_backref_add_tree_node(trans, cache, path, iter,
 499						  node_key, cur);
 500		if (ret < 0) {
 501			err = ret;
 502			goto out;
 503		}
 504		edge = list_first_entry_or_null(&cache->pending_edge,
 505				struct btrfs_backref_edge, list[UPPER]);
 506		/*
 507		 * The pending list isn't empty, take the first block to
 508		 * process
 509		 */
 510		if (edge) {
 511			list_del_init(&edge->list[UPPER]);
 512			cur = edge->node[UPPER];
 513		}
 514	} while (edge);
 515
 516	/* Finish the upper linkage of newly added edges/nodes */
 517	ret = btrfs_backref_finish_upper_links(cache, node);
 518	if (ret < 0) {
 519		err = ret;
 520		goto out;
 521	}
 522
 523	if (handle_useless_nodes(rc, node))
 524		node = NULL;
 525out:
 526	btrfs_backref_iter_free(iter);
 
 527	btrfs_free_path(path);
 528	if (err) {
 529		btrfs_backref_error_cleanup(cache, node);
 530		return ERR_PTR(err);
 531	}
 532	ASSERT(!node || !node->detached);
 533	ASSERT(list_empty(&cache->useless_node) &&
 534	       list_empty(&cache->pending_edge));
 535	return node;
 536}
 537
 538/*
 539 * helper to add backref node for the newly created snapshot.
 540 * the backref node is created by cloning backref node that
 541 * corresponds to root of source tree
 542 */
 543static int clone_backref_node(struct btrfs_trans_handle *trans,
 544			      struct reloc_control *rc,
 545			      const struct btrfs_root *src,
 546			      struct btrfs_root *dest)
 547{
 548	struct btrfs_root *reloc_root = src->reloc_root;
 549	struct btrfs_backref_cache *cache = &rc->backref_cache;
 550	struct btrfs_backref_node *node = NULL;
 551	struct btrfs_backref_node *new_node;
 552	struct btrfs_backref_edge *edge;
 553	struct btrfs_backref_edge *new_edge;
 554	struct rb_node *rb_node;
 555
 556	if (cache->last_trans > 0)
 557		update_backref_cache(trans, cache);
 558
 559	rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
 560	if (rb_node) {
 561		node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
 562		if (node->detached)
 563			node = NULL;
 564		else
 565			BUG_ON(node->new_bytenr != reloc_root->node->start);
 566	}
 567
 568	if (!node) {
 569		rb_node = rb_simple_search(&cache->rb_root,
 570					   reloc_root->commit_root->start);
 571		if (rb_node) {
 572			node = rb_entry(rb_node, struct btrfs_backref_node,
 573					rb_node);
 574			BUG_ON(node->detached);
 575		}
 576	}
 577
 578	if (!node)
 579		return 0;
 580
 581	new_node = btrfs_backref_alloc_node(cache, dest->node->start,
 582					    node->level);
 583	if (!new_node)
 584		return -ENOMEM;
 585
 586	new_node->lowest = node->lowest;
 587	new_node->checked = 1;
 588	new_node->root = btrfs_grab_root(dest);
 589	ASSERT(new_node->root);
 590
 591	if (!node->lowest) {
 592		list_for_each_entry(edge, &node->lower, list[UPPER]) {
 593			new_edge = btrfs_backref_alloc_edge(cache);
 594			if (!new_edge)
 595				goto fail;
 596
 597			btrfs_backref_link_edge(new_edge, edge->node[LOWER],
 598						new_node, LINK_UPPER);
 599		}
 600	} else {
 601		list_add_tail(&new_node->lower, &cache->leaves);
 602	}
 603
 604	rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
 605				   &new_node->rb_node);
 606	if (rb_node)
 607		btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
 608
 609	if (!new_node->lowest) {
 610		list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
 611			list_add_tail(&new_edge->list[LOWER],
 612				      &new_edge->node[LOWER]->upper);
 613		}
 614	}
 615	return 0;
 616fail:
 617	while (!list_empty(&new_node->lower)) {
 618		new_edge = list_entry(new_node->lower.next,
 619				      struct btrfs_backref_edge, list[UPPER]);
 620		list_del(&new_edge->list[UPPER]);
 621		btrfs_backref_free_edge(cache, new_edge);
 622	}
 623	btrfs_backref_free_node(cache, new_node);
 624	return -ENOMEM;
 625}
 626
 627/*
 628 * helper to add 'address of tree root -> reloc tree' mapping
 629 */
 630static int __add_reloc_root(struct btrfs_root *root)
 631{
 632	struct btrfs_fs_info *fs_info = root->fs_info;
 633	struct rb_node *rb_node;
 634	struct mapping_node *node;
 635	struct reloc_control *rc = fs_info->reloc_ctl;
 636
 637	node = kmalloc(sizeof(*node), GFP_NOFS);
 638	if (!node)
 639		return -ENOMEM;
 640
 641	node->bytenr = root->commit_root->start;
 642	node->data = root;
 643
 644	spin_lock(&rc->reloc_root_tree.lock);
 645	rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
 646				   node->bytenr, &node->rb_node);
 647	spin_unlock(&rc->reloc_root_tree.lock);
 648	if (rb_node) {
 649		btrfs_err(fs_info,
 650			    "Duplicate root found for start=%llu while inserting into relocation tree",
 651			    node->bytenr);
 652		return -EEXIST;
 653	}
 654
 655	list_add_tail(&root->root_list, &rc->reloc_roots);
 656	return 0;
 657}
 658
 659/*
 660 * helper to delete the 'address of tree root -> reloc tree'
 661 * mapping
 662 */
 663static void __del_reloc_root(struct btrfs_root *root)
 664{
 665	struct btrfs_fs_info *fs_info = root->fs_info;
 666	struct rb_node *rb_node;
 667	struct mapping_node *node = NULL;
 668	struct reloc_control *rc = fs_info->reloc_ctl;
 669	bool put_ref = false;
 670
 671	if (rc && root->node) {
 672		spin_lock(&rc->reloc_root_tree.lock);
 673		rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
 674					   root->commit_root->start);
 675		if (rb_node) {
 676			node = rb_entry(rb_node, struct mapping_node, rb_node);
 677			rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
 678			RB_CLEAR_NODE(&node->rb_node);
 679		}
 680		spin_unlock(&rc->reloc_root_tree.lock);
 681		ASSERT(!node || (struct btrfs_root *)node->data == root);
 682	}
 683
 684	/*
 685	 * We only put the reloc root here if it's on the list.  There's a lot
 686	 * of places where the pattern is to splice the rc->reloc_roots, process
 687	 * the reloc roots, and then add the reloc root back onto
 688	 * rc->reloc_roots.  If we call __del_reloc_root while it's off of the
 689	 * list we don't want the reference being dropped, because the guy
 690	 * messing with the list is in charge of the reference.
 691	 */
 692	spin_lock(&fs_info->trans_lock);
 693	if (!list_empty(&root->root_list)) {
 694		put_ref = true;
 695		list_del_init(&root->root_list);
 696	}
 697	spin_unlock(&fs_info->trans_lock);
 698	if (put_ref)
 699		btrfs_put_root(root);
 700	kfree(node);
 701}
 702
 703/*
 704 * helper to update the 'address of tree root -> reloc tree'
 705 * mapping
 706 */
 707static int __update_reloc_root(struct btrfs_root *root)
 708{
 709	struct btrfs_fs_info *fs_info = root->fs_info;
 710	struct rb_node *rb_node;
 711	struct mapping_node *node = NULL;
 712	struct reloc_control *rc = fs_info->reloc_ctl;
 713
 714	spin_lock(&rc->reloc_root_tree.lock);
 715	rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
 716				   root->commit_root->start);
 717	if (rb_node) {
 718		node = rb_entry(rb_node, struct mapping_node, rb_node);
 719		rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
 720	}
 721	spin_unlock(&rc->reloc_root_tree.lock);
 722
 723	if (!node)
 724		return 0;
 725	BUG_ON((struct btrfs_root *)node->data != root);
 726
 727	spin_lock(&rc->reloc_root_tree.lock);
 728	node->bytenr = root->node->start;
 729	rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
 730				   node->bytenr, &node->rb_node);
 731	spin_unlock(&rc->reloc_root_tree.lock);
 732	if (rb_node)
 733		btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
 734	return 0;
 735}
 736
 737static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
 738					struct btrfs_root *root, u64 objectid)
 739{
 740	struct btrfs_fs_info *fs_info = root->fs_info;
 741	struct btrfs_root *reloc_root;
 742	struct extent_buffer *eb;
 743	struct btrfs_root_item *root_item;
 744	struct btrfs_key root_key;
 745	int ret = 0;
 746	bool must_abort = false;
 747
 748	root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
 749	if (!root_item)
 750		return ERR_PTR(-ENOMEM);
 751
 752	root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
 753	root_key.type = BTRFS_ROOT_ITEM_KEY;
 754	root_key.offset = objectid;
 755
 756	if (root->root_key.objectid == objectid) {
 757		u64 commit_root_gen;
 758
 759		/* called by btrfs_init_reloc_root */
 760		ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
 761				      BTRFS_TREE_RELOC_OBJECTID);
 762		if (ret)
 763			goto fail;
 764
 765		/*
 766		 * Set the last_snapshot field to the generation of the commit
 767		 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
 768		 * correctly (returns true) when the relocation root is created
 769		 * either inside the critical section of a transaction commit
 770		 * (through transaction.c:qgroup_account_snapshot()) and when
 771		 * it's created before the transaction commit is started.
 772		 */
 773		commit_root_gen = btrfs_header_generation(root->commit_root);
 774		btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
 775	} else {
 776		/*
 777		 * called by btrfs_reloc_post_snapshot_hook.
 778		 * the source tree is a reloc tree, all tree blocks
 779		 * modified after it was created have RELOC flag
 780		 * set in their headers. so it's OK to not update
 781		 * the 'last_snapshot'.
 782		 */
 783		ret = btrfs_copy_root(trans, root, root->node, &eb,
 784				      BTRFS_TREE_RELOC_OBJECTID);
 785		if (ret)
 786			goto fail;
 787	}
 788
 789	/*
 790	 * We have changed references at this point, we must abort the
 791	 * transaction if anything fails.
 792	 */
 793	must_abort = true;
 794
 795	memcpy(root_item, &root->root_item, sizeof(*root_item));
 796	btrfs_set_root_bytenr(root_item, eb->start);
 797	btrfs_set_root_level(root_item, btrfs_header_level(eb));
 798	btrfs_set_root_generation(root_item, trans->transid);
 799
 800	if (root->root_key.objectid == objectid) {
 801		btrfs_set_root_refs(root_item, 0);
 802		memset(&root_item->drop_progress, 0,
 803		       sizeof(struct btrfs_disk_key));
 804		btrfs_set_root_drop_level(root_item, 0);
 805	}
 806
 807	btrfs_tree_unlock(eb);
 808	free_extent_buffer(eb);
 809
 810	ret = btrfs_insert_root(trans, fs_info->tree_root,
 811				&root_key, root_item);
 812	if (ret)
 813		goto fail;
 814
 815	kfree(root_item);
 816
 817	reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
 818	if (IS_ERR(reloc_root)) {
 819		ret = PTR_ERR(reloc_root);
 820		goto abort;
 821	}
 822	set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
 823	reloc_root->last_trans = trans->transid;
 824	return reloc_root;
 825fail:
 826	kfree(root_item);
 827abort:
 828	if (must_abort)
 829		btrfs_abort_transaction(trans, ret);
 830	return ERR_PTR(ret);
 831}
 832
 833/*
 834 * create reloc tree for a given fs tree. reloc tree is just a
 835 * snapshot of the fs tree with special root objectid.
 836 *
 837 * The reloc_root comes out of here with two references, one for
 838 * root->reloc_root, and another for being on the rc->reloc_roots list.
 839 */
 840int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
 841			  struct btrfs_root *root)
 842{
 843	struct btrfs_fs_info *fs_info = root->fs_info;
 844	struct btrfs_root *reloc_root;
 845	struct reloc_control *rc = fs_info->reloc_ctl;
 846	struct btrfs_block_rsv *rsv;
 847	int clear_rsv = 0;
 848	int ret;
 849
 850	if (!rc)
 851		return 0;
 852
 853	/*
 854	 * The subvolume has reloc tree but the swap is finished, no need to
 855	 * create/update the dead reloc tree
 856	 */
 857	if (reloc_root_is_dead(root))
 858		return 0;
 859
 860	/*
 861	 * This is subtle but important.  We do not do
 862	 * record_root_in_transaction for reloc roots, instead we record their
 863	 * corresponding fs root, and then here we update the last trans for the
 864	 * reloc root.  This means that we have to do this for the entire life
 865	 * of the reloc root, regardless of which stage of the relocation we are
 866	 * in.
 867	 */
 868	if (root->reloc_root) {
 869		reloc_root = root->reloc_root;
 870		reloc_root->last_trans = trans->transid;
 871		return 0;
 872	}
 873
 874	/*
 875	 * We are merging reloc roots, we do not need new reloc trees.  Also
 876	 * reloc trees never need their own reloc tree.
 877	 */
 878	if (!rc->create_reloc_tree ||
 879	    root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 880		return 0;
 881
 882	if (!trans->reloc_reserved) {
 883		rsv = trans->block_rsv;
 884		trans->block_rsv = rc->block_rsv;
 885		clear_rsv = 1;
 886	}
 887	reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
 888	if (clear_rsv)
 889		trans->block_rsv = rsv;
 890	if (IS_ERR(reloc_root))
 891		return PTR_ERR(reloc_root);
 892
 893	ret = __add_reloc_root(reloc_root);
 894	ASSERT(ret != -EEXIST);
 895	if (ret) {
 896		/* Pairs with create_reloc_root */
 897		btrfs_put_root(reloc_root);
 898		return ret;
 899	}
 900	root->reloc_root = btrfs_grab_root(reloc_root);
 901	return 0;
 902}
 903
 904/*
 905 * update root item of reloc tree
 906 */
 907int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
 908			    struct btrfs_root *root)
 909{
 910	struct btrfs_fs_info *fs_info = root->fs_info;
 911	struct btrfs_root *reloc_root;
 912	struct btrfs_root_item *root_item;
 913	int ret;
 914
 915	if (!have_reloc_root(root))
 916		return 0;
 917
 918	reloc_root = root->reloc_root;
 919	root_item = &reloc_root->root_item;
 920
 921	/*
 922	 * We are probably ok here, but __del_reloc_root() will drop its ref of
 923	 * the root.  We have the ref for root->reloc_root, but just in case
 924	 * hold it while we update the reloc root.
 925	 */
 926	btrfs_grab_root(reloc_root);
 927
 928	/* root->reloc_root will stay until current relocation finished */
 929	if (fs_info->reloc_ctl->merge_reloc_tree &&
 930	    btrfs_root_refs(root_item) == 0) {
 931		set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
 932		/*
 933		 * Mark the tree as dead before we change reloc_root so
 934		 * have_reloc_root will not touch it from now on.
 935		 */
 936		smp_wmb();
 937		__del_reloc_root(reloc_root);
 938	}
 939
 940	if (reloc_root->commit_root != reloc_root->node) {
 941		__update_reloc_root(reloc_root);
 942		btrfs_set_root_node(root_item, reloc_root->node);
 943		free_extent_buffer(reloc_root->commit_root);
 944		reloc_root->commit_root = btrfs_root_node(reloc_root);
 945	}
 946
 947	ret = btrfs_update_root(trans, fs_info->tree_root,
 948				&reloc_root->root_key, root_item);
 949	btrfs_put_root(reloc_root);
 950	return ret;
 951}
 952
 953/*
 954 * helper to find first cached inode with inode number >= objectid
 955 * in a subvolume
 956 */
 957static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
 958{
 959	struct rb_node *node;
 960	struct rb_node *prev;
 961	struct btrfs_inode *entry;
 962	struct inode *inode;
 963
 964	spin_lock(&root->inode_lock);
 965again:
 966	node = root->inode_tree.rb_node;
 967	prev = NULL;
 968	while (node) {
 969		prev = node;
 970		entry = rb_entry(node, struct btrfs_inode, rb_node);
 971
 972		if (objectid < btrfs_ino(entry))
 973			node = node->rb_left;
 974		else if (objectid > btrfs_ino(entry))
 975			node = node->rb_right;
 976		else
 977			break;
 978	}
 979	if (!node) {
 980		while (prev) {
 981			entry = rb_entry(prev, struct btrfs_inode, rb_node);
 982			if (objectid <= btrfs_ino(entry)) {
 983				node = prev;
 984				break;
 985			}
 986			prev = rb_next(prev);
 987		}
 988	}
 989	while (node) {
 990		entry = rb_entry(node, struct btrfs_inode, rb_node);
 991		inode = igrab(&entry->vfs_inode);
 992		if (inode) {
 993			spin_unlock(&root->inode_lock);
 994			return inode;
 995		}
 996
 997		objectid = btrfs_ino(entry) + 1;
 998		if (cond_resched_lock(&root->inode_lock))
 999			goto again;
1000
1001		node = rb_next(node);
1002	}
1003	spin_unlock(&root->inode_lock);
1004	return NULL;
1005}
1006
1007/*
1008 * get new location of data
1009 */
1010static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1011			    u64 bytenr, u64 num_bytes)
1012{
1013	struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1014	struct btrfs_path *path;
1015	struct btrfs_file_extent_item *fi;
1016	struct extent_buffer *leaf;
1017	int ret;
1018
1019	path = btrfs_alloc_path();
1020	if (!path)
1021		return -ENOMEM;
1022
1023	bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1024	ret = btrfs_lookup_file_extent(NULL, root, path,
1025			btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1026	if (ret < 0)
1027		goto out;
1028	if (ret > 0) {
1029		ret = -ENOENT;
1030		goto out;
1031	}
1032
1033	leaf = path->nodes[0];
1034	fi = btrfs_item_ptr(leaf, path->slots[0],
1035			    struct btrfs_file_extent_item);
1036
1037	BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1038	       btrfs_file_extent_compression(leaf, fi) ||
1039	       btrfs_file_extent_encryption(leaf, fi) ||
1040	       btrfs_file_extent_other_encoding(leaf, fi));
1041
1042	if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1043		ret = -EINVAL;
1044		goto out;
1045	}
1046
1047	*new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1048	ret = 0;
1049out:
1050	btrfs_free_path(path);
1051	return ret;
1052}
1053
1054/*
1055 * update file extent items in the tree leaf to point to
1056 * the new locations.
1057 */
1058static noinline_for_stack
1059int replace_file_extents(struct btrfs_trans_handle *trans,
1060			 struct reloc_control *rc,
1061			 struct btrfs_root *root,
1062			 struct extent_buffer *leaf)
1063{
1064	struct btrfs_fs_info *fs_info = root->fs_info;
1065	struct btrfs_key key;
1066	struct btrfs_file_extent_item *fi;
1067	struct inode *inode = NULL;
1068	u64 parent;
1069	u64 bytenr;
1070	u64 new_bytenr = 0;
1071	u64 num_bytes;
1072	u64 end;
1073	u32 nritems;
1074	u32 i;
1075	int ret = 0;
1076	int first = 1;
1077	int dirty = 0;
1078
1079	if (rc->stage != UPDATE_DATA_PTRS)
1080		return 0;
1081
1082	/* reloc trees always use full backref */
1083	if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1084		parent = leaf->start;
1085	else
1086		parent = 0;
1087
1088	nritems = btrfs_header_nritems(leaf);
1089	for (i = 0; i < nritems; i++) {
1090		struct btrfs_ref ref = { 0 };
1091
1092		cond_resched();
1093		btrfs_item_key_to_cpu(leaf, &key, i);
1094		if (key.type != BTRFS_EXTENT_DATA_KEY)
1095			continue;
1096		fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1097		if (btrfs_file_extent_type(leaf, fi) ==
1098		    BTRFS_FILE_EXTENT_INLINE)
1099			continue;
1100		bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1101		num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1102		if (bytenr == 0)
1103			continue;
1104		if (!in_range(bytenr, rc->block_group->start,
1105			      rc->block_group->length))
1106			continue;
1107
1108		/*
1109		 * if we are modifying block in fs tree, wait for read_folio
1110		 * to complete and drop the extent cache
1111		 */
1112		if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1113			if (first) {
1114				inode = find_next_inode(root, key.objectid);
1115				first = 0;
1116			} else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1117				btrfs_add_delayed_iput(BTRFS_I(inode));
1118				inode = find_next_inode(root, key.objectid);
1119			}
1120			if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1121				struct extent_state *cached_state = NULL;
1122
1123				end = key.offset +
1124				      btrfs_file_extent_num_bytes(leaf, fi);
1125				WARN_ON(!IS_ALIGNED(key.offset,
1126						    fs_info->sectorsize));
1127				WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1128				end--;
1129				ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1130						      key.offset, end,
1131						      &cached_state);
1132				if (!ret)
1133					continue;
1134
1135				btrfs_drop_extent_map_range(BTRFS_I(inode),
1136							    key.offset, end, true);
1137				unlock_extent(&BTRFS_I(inode)->io_tree,
1138					      key.offset, end, &cached_state);
1139			}
1140		}
1141
1142		ret = get_new_location(rc->data_inode, &new_bytenr,
1143				       bytenr, num_bytes);
1144		if (ret) {
1145			/*
1146			 * Don't have to abort since we've not changed anything
1147			 * in the file extent yet.
1148			 */
1149			break;
1150		}
1151
1152		btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1153		dirty = 1;
1154
1155		key.offset -= btrfs_file_extent_offset(leaf, fi);
1156		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1157				       num_bytes, parent, root->root_key.objectid);
1158		btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1159				    key.objectid, key.offset,
1160				    root->root_key.objectid, false);
1161		ret = btrfs_inc_extent_ref(trans, &ref);
1162		if (ret) {
1163			btrfs_abort_transaction(trans, ret);
1164			break;
1165		}
1166
1167		btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1168				       num_bytes, parent, root->root_key.objectid);
1169		btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1170				    key.objectid, key.offset,
1171				    root->root_key.objectid, false);
1172		ret = btrfs_free_extent(trans, &ref);
1173		if (ret) {
1174			btrfs_abort_transaction(trans, ret);
1175			break;
1176		}
1177	}
1178	if (dirty)
1179		btrfs_mark_buffer_dirty(trans, leaf);
1180	if (inode)
1181		btrfs_add_delayed_iput(BTRFS_I(inode));
1182	return ret;
1183}
1184
1185static noinline_for_stack int memcmp_node_keys(const struct extent_buffer *eb,
1186					       int slot, const struct btrfs_path *path,
1187					       int level)
1188{
1189	struct btrfs_disk_key key1;
1190	struct btrfs_disk_key key2;
1191	btrfs_node_key(eb, &key1, slot);
1192	btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1193	return memcmp(&key1, &key2, sizeof(key1));
1194}
1195
1196/*
1197 * try to replace tree blocks in fs tree with the new blocks
1198 * in reloc tree. tree blocks haven't been modified since the
1199 * reloc tree was create can be replaced.
1200 *
1201 * if a block was replaced, level of the block + 1 is returned.
1202 * if no block got replaced, 0 is returned. if there are other
1203 * errors, a negative error number is returned.
1204 */
1205static noinline_for_stack
1206int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1207		 struct btrfs_root *dest, struct btrfs_root *src,
1208		 struct btrfs_path *path, struct btrfs_key *next_key,
1209		 int lowest_level, int max_level)
1210{
1211	struct btrfs_fs_info *fs_info = dest->fs_info;
1212	struct extent_buffer *eb;
1213	struct extent_buffer *parent;
1214	struct btrfs_ref ref = { 0 };
1215	struct btrfs_key key;
1216	u64 old_bytenr;
1217	u64 new_bytenr;
1218	u64 old_ptr_gen;
1219	u64 new_ptr_gen;
1220	u64 last_snapshot;
1221	u32 blocksize;
1222	int cow = 0;
1223	int level;
1224	int ret;
1225	int slot;
1226
1227	ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1228	ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1229
1230	last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1231again:
1232	slot = path->slots[lowest_level];
1233	btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1234
1235	eb = btrfs_lock_root_node(dest);
1236	level = btrfs_header_level(eb);
1237
1238	if (level < lowest_level) {
1239		btrfs_tree_unlock(eb);
1240		free_extent_buffer(eb);
1241		return 0;
1242	}
1243
1244	if (cow) {
1245		ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1246				      BTRFS_NESTING_COW);
1247		if (ret) {
1248			btrfs_tree_unlock(eb);
1249			free_extent_buffer(eb);
1250			return ret;
1251		}
1252	}
1253
1254	if (next_key) {
1255		next_key->objectid = (u64)-1;
1256		next_key->type = (u8)-1;
1257		next_key->offset = (u64)-1;
1258	}
1259
1260	parent = eb;
1261	while (1) {
1262		level = btrfs_header_level(parent);
1263		ASSERT(level >= lowest_level);
1264
1265		ret = btrfs_bin_search(parent, 0, &key, &slot);
1266		if (ret < 0)
1267			break;
1268		if (ret && slot > 0)
1269			slot--;
1270
1271		if (next_key && slot + 1 < btrfs_header_nritems(parent))
1272			btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1273
1274		old_bytenr = btrfs_node_blockptr(parent, slot);
1275		blocksize = fs_info->nodesize;
1276		old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1277
1278		if (level <= max_level) {
1279			eb = path->nodes[level];
1280			new_bytenr = btrfs_node_blockptr(eb,
1281							path->slots[level]);
1282			new_ptr_gen = btrfs_node_ptr_generation(eb,
1283							path->slots[level]);
1284		} else {
1285			new_bytenr = 0;
1286			new_ptr_gen = 0;
1287		}
1288
1289		if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1290			ret = level;
1291			break;
1292		}
1293
1294		if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1295		    memcmp_node_keys(parent, slot, path, level)) {
1296			if (level <= lowest_level) {
1297				ret = 0;
1298				break;
1299			}
1300
1301			eb = btrfs_read_node_slot(parent, slot);
1302			if (IS_ERR(eb)) {
1303				ret = PTR_ERR(eb);
1304				break;
1305			}
1306			btrfs_tree_lock(eb);
1307			if (cow) {
1308				ret = btrfs_cow_block(trans, dest, eb, parent,
1309						      slot, &eb,
1310						      BTRFS_NESTING_COW);
1311				if (ret) {
1312					btrfs_tree_unlock(eb);
1313					free_extent_buffer(eb);
1314					break;
1315				}
1316			}
1317
1318			btrfs_tree_unlock(parent);
1319			free_extent_buffer(parent);
1320
1321			parent = eb;
1322			continue;
1323		}
1324
1325		if (!cow) {
1326			btrfs_tree_unlock(parent);
1327			free_extent_buffer(parent);
1328			cow = 1;
1329			goto again;
1330		}
1331
1332		btrfs_node_key_to_cpu(path->nodes[level], &key,
1333				      path->slots[level]);
1334		btrfs_release_path(path);
1335
1336		path->lowest_level = level;
1337		set_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1338		ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1339		clear_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1340		path->lowest_level = 0;
1341		if (ret) {
1342			if (ret > 0)
1343				ret = -ENOENT;
1344			break;
1345		}
1346
1347		/*
1348		 * Info qgroup to trace both subtrees.
1349		 *
1350		 * We must trace both trees.
1351		 * 1) Tree reloc subtree
1352		 *    If not traced, we will leak data numbers
1353		 * 2) Fs subtree
1354		 *    If not traced, we will double count old data
1355		 *
1356		 * We don't scan the subtree right now, but only record
1357		 * the swapped tree blocks.
1358		 * The real subtree rescan is delayed until we have new
1359		 * CoW on the subtree root node before transaction commit.
1360		 */
1361		ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1362				rc->block_group, parent, slot,
1363				path->nodes[level], path->slots[level],
1364				last_snapshot);
1365		if (ret < 0)
1366			break;
1367		/*
1368		 * swap blocks in fs tree and reloc tree.
1369		 */
1370		btrfs_set_node_blockptr(parent, slot, new_bytenr);
1371		btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1372		btrfs_mark_buffer_dirty(trans, parent);
1373
1374		btrfs_set_node_blockptr(path->nodes[level],
1375					path->slots[level], old_bytenr);
1376		btrfs_set_node_ptr_generation(path->nodes[level],
1377					      path->slots[level], old_ptr_gen);
1378		btrfs_mark_buffer_dirty(trans, path->nodes[level]);
1379
1380		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1381				       blocksize, path->nodes[level]->start,
1382				       src->root_key.objectid);
1383		btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1384				    0, true);
1385		ret = btrfs_inc_extent_ref(trans, &ref);
1386		if (ret) {
1387			btrfs_abort_transaction(trans, ret);
1388			break;
1389		}
1390		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1391				       blocksize, 0, dest->root_key.objectid);
1392		btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid, 0,
1393				    true);
1394		ret = btrfs_inc_extent_ref(trans, &ref);
1395		if (ret) {
1396			btrfs_abort_transaction(trans, ret);
1397			break;
1398		}
1399
1400		/* We don't know the real owning_root, use 0. */
1401		btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1402				       blocksize, path->nodes[level]->start, 0);
1403		btrfs_init_tree_ref(&ref, level - 1, src->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		/* We don't know the real owning_root, use 0. */
1412		btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1413				       blocksize, 0, 0);
1414		btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid,
1415				    0, true);
1416		ret = btrfs_free_extent(trans, &ref);
1417		if (ret) {
1418			btrfs_abort_transaction(trans, ret);
1419			break;
1420		}
1421
1422		btrfs_unlock_up_safe(path, 0);
1423
1424		ret = level;
1425		break;
1426	}
1427	btrfs_tree_unlock(parent);
1428	free_extent_buffer(parent);
1429	return ret;
1430}
1431
1432/*
1433 * helper to find next relocated block in reloc tree
1434 */
1435static noinline_for_stack
1436int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1437		       int *level)
1438{
1439	struct extent_buffer *eb;
1440	int i;
1441	u64 last_snapshot;
1442	u32 nritems;
1443
1444	last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1445
1446	for (i = 0; i < *level; i++) {
1447		free_extent_buffer(path->nodes[i]);
1448		path->nodes[i] = NULL;
1449	}
1450
1451	for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1452		eb = path->nodes[i];
1453		nritems = btrfs_header_nritems(eb);
1454		while (path->slots[i] + 1 < nritems) {
1455			path->slots[i]++;
1456			if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1457			    last_snapshot)
1458				continue;
1459
1460			*level = i;
1461			return 0;
1462		}
1463		free_extent_buffer(path->nodes[i]);
1464		path->nodes[i] = NULL;
1465	}
1466	return 1;
1467}
1468
1469/*
1470 * walk down reloc tree to find relocated block of lowest level
1471 */
1472static noinline_for_stack
1473int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1474			 int *level)
1475{
1476	struct extent_buffer *eb = NULL;
1477	int i;
1478	u64 ptr_gen = 0;
1479	u64 last_snapshot;
1480	u32 nritems;
1481
1482	last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1483
1484	for (i = *level; i > 0; i--) {
1485		eb = path->nodes[i];
1486		nritems = btrfs_header_nritems(eb);
1487		while (path->slots[i] < nritems) {
1488			ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1489			if (ptr_gen > last_snapshot)
1490				break;
1491			path->slots[i]++;
1492		}
1493		if (path->slots[i] >= nritems) {
1494			if (i == *level)
1495				break;
1496			*level = i + 1;
1497			return 0;
1498		}
1499		if (i == 1) {
1500			*level = i;
1501			return 0;
1502		}
1503
1504		eb = btrfs_read_node_slot(eb, path->slots[i]);
1505		if (IS_ERR(eb))
1506			return PTR_ERR(eb);
1507		BUG_ON(btrfs_header_level(eb) != i - 1);
1508		path->nodes[i - 1] = eb;
1509		path->slots[i - 1] = 0;
1510	}
1511	return 1;
1512}
1513
1514/*
1515 * invalidate extent cache for file extents whose key in range of
1516 * [min_key, max_key)
1517 */
1518static int invalidate_extent_cache(struct btrfs_root *root,
1519				   const struct btrfs_key *min_key,
1520				   const struct btrfs_key *max_key)
1521{
1522	struct btrfs_fs_info *fs_info = root->fs_info;
1523	struct inode *inode = NULL;
1524	u64 objectid;
1525	u64 start, end;
1526	u64 ino;
1527
1528	objectid = min_key->objectid;
1529	while (1) {
1530		struct extent_state *cached_state = NULL;
1531
1532		cond_resched();
1533		iput(inode);
1534
1535		if (objectid > max_key->objectid)
1536			break;
1537
1538		inode = find_next_inode(root, objectid);
1539		if (!inode)
1540			break;
1541		ino = btrfs_ino(BTRFS_I(inode));
1542
1543		if (ino > max_key->objectid) {
1544			iput(inode);
1545			break;
1546		}
1547
1548		objectid = ino + 1;
1549		if (!S_ISREG(inode->i_mode))
1550			continue;
1551
1552		if (unlikely(min_key->objectid == ino)) {
1553			if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1554				continue;
1555			if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1556				start = 0;
1557			else {
1558				start = min_key->offset;
1559				WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1560			}
1561		} else {
1562			start = 0;
1563		}
1564
1565		if (unlikely(max_key->objectid == ino)) {
1566			if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1567				continue;
1568			if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1569				end = (u64)-1;
1570			} else {
1571				if (max_key->offset == 0)
1572					continue;
1573				end = max_key->offset;
1574				WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1575				end--;
1576			}
1577		} else {
1578			end = (u64)-1;
1579		}
1580
1581		/* the lock_extent waits for read_folio to complete */
1582		lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1583		btrfs_drop_extent_map_range(BTRFS_I(inode), start, end, true);
1584		unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1585	}
1586	return 0;
1587}
1588
1589static int find_next_key(struct btrfs_path *path, int level,
1590			 struct btrfs_key *key)
1591
1592{
1593	while (level < BTRFS_MAX_LEVEL) {
1594		if (!path->nodes[level])
1595			break;
1596		if (path->slots[level] + 1 <
1597		    btrfs_header_nritems(path->nodes[level])) {
1598			btrfs_node_key_to_cpu(path->nodes[level], key,
1599					      path->slots[level] + 1);
1600			return 0;
1601		}
1602		level++;
1603	}
1604	return 1;
1605}
1606
1607/*
1608 * Insert current subvolume into reloc_control::dirty_subvol_roots
1609 */
1610static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1611			       struct reloc_control *rc,
1612			       struct btrfs_root *root)
1613{
1614	struct btrfs_root *reloc_root = root->reloc_root;
1615	struct btrfs_root_item *reloc_root_item;
1616	int ret;
1617
1618	/* @root must be a subvolume tree root with a valid reloc tree */
1619	ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1620	ASSERT(reloc_root);
1621
1622	reloc_root_item = &reloc_root->root_item;
1623	memset(&reloc_root_item->drop_progress, 0,
1624		sizeof(reloc_root_item->drop_progress));
1625	btrfs_set_root_drop_level(reloc_root_item, 0);
1626	btrfs_set_root_refs(reloc_root_item, 0);
1627	ret = btrfs_update_reloc_root(trans, root);
1628	if (ret)
1629		return ret;
1630
1631	if (list_empty(&root->reloc_dirty_list)) {
1632		btrfs_grab_root(root);
1633		list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1634	}
1635
1636	return 0;
1637}
1638
1639static int clean_dirty_subvols(struct reloc_control *rc)
1640{
1641	struct btrfs_root *root;
1642	struct btrfs_root *next;
1643	int ret = 0;
1644	int ret2;
1645
1646	list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1647				 reloc_dirty_list) {
1648		if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1649			/* Merged subvolume, cleanup its reloc root */
1650			struct btrfs_root *reloc_root = root->reloc_root;
1651
1652			list_del_init(&root->reloc_dirty_list);
1653			root->reloc_root = NULL;
1654			/*
1655			 * Need barrier to ensure clear_bit() only happens after
1656			 * root->reloc_root = NULL. Pairs with have_reloc_root.
1657			 */
1658			smp_wmb();
1659			clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1660			if (reloc_root) {
1661				/*
1662				 * btrfs_drop_snapshot drops our ref we hold for
1663				 * ->reloc_root.  If it fails however we must
1664				 * drop the ref ourselves.
1665				 */
1666				ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1667				if (ret2 < 0) {
1668					btrfs_put_root(reloc_root);
1669					if (!ret)
1670						ret = ret2;
1671				}
1672			}
1673			btrfs_put_root(root);
1674		} else {
1675			/* Orphan reloc tree, just clean it up */
1676			ret2 = btrfs_drop_snapshot(root, 0, 1);
1677			if (ret2 < 0) {
1678				btrfs_put_root(root);
1679				if (!ret)
1680					ret = ret2;
1681			}
1682		}
1683	}
1684	return ret;
1685}
1686
1687/*
1688 * merge the relocated tree blocks in reloc tree with corresponding
1689 * fs tree.
1690 */
1691static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1692					       struct btrfs_root *root)
1693{
1694	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1695	struct btrfs_key key;
1696	struct btrfs_key next_key;
1697	struct btrfs_trans_handle *trans = NULL;
1698	struct btrfs_root *reloc_root;
1699	struct btrfs_root_item *root_item;
1700	struct btrfs_path *path;
1701	struct extent_buffer *leaf;
1702	int reserve_level;
1703	int level;
1704	int max_level;
1705	int replaced = 0;
1706	int ret = 0;
1707	u32 min_reserved;
1708
1709	path = btrfs_alloc_path();
1710	if (!path)
1711		return -ENOMEM;
1712	path->reada = READA_FORWARD;
1713
1714	reloc_root = root->reloc_root;
1715	root_item = &reloc_root->root_item;
1716
1717	if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1718		level = btrfs_root_level(root_item);
1719		atomic_inc(&reloc_root->node->refs);
1720		path->nodes[level] = reloc_root->node;
1721		path->slots[level] = 0;
1722	} else {
1723		btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1724
1725		level = btrfs_root_drop_level(root_item);
1726		BUG_ON(level == 0);
1727		path->lowest_level = level;
1728		ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1729		path->lowest_level = 0;
1730		if (ret < 0) {
1731			btrfs_free_path(path);
1732			return ret;
1733		}
1734
1735		btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1736				      path->slots[level]);
1737		WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1738
1739		btrfs_unlock_up_safe(path, 0);
1740	}
1741
1742	/*
1743	 * In merge_reloc_root(), we modify the upper level pointer to swap the
1744	 * tree blocks between reloc tree and subvolume tree.  Thus for tree
1745	 * block COW, we COW at most from level 1 to root level for each tree.
1746	 *
1747	 * Thus the needed metadata size is at most root_level * nodesize,
1748	 * and * 2 since we have two trees to COW.
1749	 */
1750	reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1751	min_reserved = fs_info->nodesize * reserve_level * 2;
1752	memset(&next_key, 0, sizeof(next_key));
1753
1754	while (1) {
1755		ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
1756					     min_reserved,
1757					     BTRFS_RESERVE_FLUSH_LIMIT);
1758		if (ret)
1759			goto out;
1760		trans = btrfs_start_transaction(root, 0);
1761		if (IS_ERR(trans)) {
1762			ret = PTR_ERR(trans);
1763			trans = NULL;
1764			goto out;
1765		}
1766
1767		/*
1768		 * At this point we no longer have a reloc_control, so we can't
1769		 * depend on btrfs_init_reloc_root to update our last_trans.
1770		 *
1771		 * But that's ok, we started the trans handle on our
1772		 * corresponding fs_root, which means it's been added to the
1773		 * dirty list.  At commit time we'll still call
1774		 * btrfs_update_reloc_root() and update our root item
1775		 * appropriately.
1776		 */
1777		reloc_root->last_trans = trans->transid;
1778		trans->block_rsv = rc->block_rsv;
1779
1780		replaced = 0;
1781		max_level = level;
1782
1783		ret = walk_down_reloc_tree(reloc_root, path, &level);
1784		if (ret < 0)
1785			goto out;
1786		if (ret > 0)
1787			break;
1788
1789		if (!find_next_key(path, level, &key) &&
1790		    btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1791			ret = 0;
1792		} else {
1793			ret = replace_path(trans, rc, root, reloc_root, path,
1794					   &next_key, level, max_level);
1795		}
1796		if (ret < 0)
1797			goto out;
1798		if (ret > 0) {
1799			level = ret;
1800			btrfs_node_key_to_cpu(path->nodes[level], &key,
1801					      path->slots[level]);
1802			replaced = 1;
1803		}
1804
1805		ret = walk_up_reloc_tree(reloc_root, path, &level);
1806		if (ret > 0)
1807			break;
1808
1809		BUG_ON(level == 0);
1810		/*
1811		 * save the merging progress in the drop_progress.
1812		 * this is OK since root refs == 1 in this case.
1813		 */
1814		btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1815			       path->slots[level]);
1816		btrfs_set_root_drop_level(root_item, level);
1817
1818		btrfs_end_transaction_throttle(trans);
1819		trans = NULL;
1820
1821		btrfs_btree_balance_dirty(fs_info);
1822
1823		if (replaced && rc->stage == UPDATE_DATA_PTRS)
1824			invalidate_extent_cache(root, &key, &next_key);
1825	}
1826
1827	/*
1828	 * handle the case only one block in the fs tree need to be
1829	 * relocated and the block is tree root.
1830	 */
1831	leaf = btrfs_lock_root_node(root);
1832	ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1833			      BTRFS_NESTING_COW);
1834	btrfs_tree_unlock(leaf);
1835	free_extent_buffer(leaf);
1836out:
1837	btrfs_free_path(path);
1838
1839	if (ret == 0) {
1840		ret = insert_dirty_subvol(trans, rc, root);
1841		if (ret)
1842			btrfs_abort_transaction(trans, ret);
1843	}
1844
1845	if (trans)
1846		btrfs_end_transaction_throttle(trans);
1847
1848	btrfs_btree_balance_dirty(fs_info);
1849
1850	if (replaced && rc->stage == UPDATE_DATA_PTRS)
1851		invalidate_extent_cache(root, &key, &next_key);
1852
1853	return ret;
1854}
1855
1856static noinline_for_stack
1857int prepare_to_merge(struct reloc_control *rc, int err)
1858{
1859	struct btrfs_root *root = rc->extent_root;
1860	struct btrfs_fs_info *fs_info = root->fs_info;
1861	struct btrfs_root *reloc_root;
1862	struct btrfs_trans_handle *trans;
1863	LIST_HEAD(reloc_roots);
1864	u64 num_bytes = 0;
1865	int ret;
1866
1867	mutex_lock(&fs_info->reloc_mutex);
1868	rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1869	rc->merging_rsv_size += rc->nodes_relocated * 2;
1870	mutex_unlock(&fs_info->reloc_mutex);
1871
1872again:
1873	if (!err) {
1874		num_bytes = rc->merging_rsv_size;
1875		ret = btrfs_block_rsv_add(fs_info, rc->block_rsv, num_bytes,
1876					  BTRFS_RESERVE_FLUSH_ALL);
1877		if (ret)
1878			err = ret;
1879	}
1880
1881	trans = btrfs_join_transaction(rc->extent_root);
1882	if (IS_ERR(trans)) {
1883		if (!err)
1884			btrfs_block_rsv_release(fs_info, rc->block_rsv,
1885						num_bytes, NULL);
1886		return PTR_ERR(trans);
1887	}
1888
1889	if (!err) {
1890		if (num_bytes != rc->merging_rsv_size) {
1891			btrfs_end_transaction(trans);
1892			btrfs_block_rsv_release(fs_info, rc->block_rsv,
1893						num_bytes, NULL);
1894			goto again;
1895		}
1896	}
1897
1898	rc->merge_reloc_tree = true;
1899
1900	while (!list_empty(&rc->reloc_roots)) {
1901		reloc_root = list_entry(rc->reloc_roots.next,
1902					struct btrfs_root, root_list);
1903		list_del_init(&reloc_root->root_list);
1904
1905		root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1906				false);
1907		if (IS_ERR(root)) {
1908			/*
1909			 * Even if we have an error we need this reloc root
1910			 * back on our list so we can clean up properly.
1911			 */
1912			list_add(&reloc_root->root_list, &reloc_roots);
1913			btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1914			if (!err)
1915				err = PTR_ERR(root);
1916			break;
1917		}
1918
1919		if (unlikely(root->reloc_root != reloc_root)) {
1920			if (root->reloc_root) {
1921				btrfs_err(fs_info,
1922"reloc tree mismatch, root %lld has reloc root key (%lld %u %llu) gen %llu, expect reloc root key (%lld %u %llu) gen %llu",
1923					  root->root_key.objectid,
1924					  root->reloc_root->root_key.objectid,
1925					  root->reloc_root->root_key.type,
1926					  root->reloc_root->root_key.offset,
1927					  btrfs_root_generation(
1928						  &root->reloc_root->root_item),
1929					  reloc_root->root_key.objectid,
1930					  reloc_root->root_key.type,
1931					  reloc_root->root_key.offset,
1932					  btrfs_root_generation(
1933						  &reloc_root->root_item));
1934			} else {
1935				btrfs_err(fs_info,
1936"reloc tree mismatch, root %lld has no reloc root, expect reloc root key (%lld %u %llu) gen %llu",
1937					  root->root_key.objectid,
1938					  reloc_root->root_key.objectid,
1939					  reloc_root->root_key.type,
1940					  reloc_root->root_key.offset,
1941					  btrfs_root_generation(
1942						  &reloc_root->root_item));
1943			}
1944			list_add(&reloc_root->root_list, &reloc_roots);
1945			btrfs_put_root(root);
1946			btrfs_abort_transaction(trans, -EUCLEAN);
1947			if (!err)
1948				err = -EUCLEAN;
1949			break;
1950		}
1951
1952		/*
1953		 * set reference count to 1, so btrfs_recover_relocation
1954		 * knows it should resumes merging
1955		 */
1956		if (!err)
1957			btrfs_set_root_refs(&reloc_root->root_item, 1);
1958		ret = btrfs_update_reloc_root(trans, root);
1959
1960		/*
1961		 * Even if we have an error we need this reloc root back on our
1962		 * list so we can clean up properly.
1963		 */
1964		list_add(&reloc_root->root_list, &reloc_roots);
1965		btrfs_put_root(root);
1966
1967		if (ret) {
1968			btrfs_abort_transaction(trans, ret);
1969			if (!err)
1970				err = ret;
1971			break;
1972		}
1973	}
1974
1975	list_splice(&reloc_roots, &rc->reloc_roots);
1976
1977	if (!err)
1978		err = btrfs_commit_transaction(trans);
1979	else
1980		btrfs_end_transaction(trans);
1981	return err;
1982}
1983
1984static noinline_for_stack
1985void free_reloc_roots(struct list_head *list)
1986{
1987	struct btrfs_root *reloc_root, *tmp;
1988
1989	list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1990		__del_reloc_root(reloc_root);
1991}
1992
1993static noinline_for_stack
1994void merge_reloc_roots(struct reloc_control *rc)
1995{
1996	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1997	struct btrfs_root *root;
1998	struct btrfs_root *reloc_root;
1999	LIST_HEAD(reloc_roots);
2000	int found = 0;
2001	int ret = 0;
2002again:
2003	root = rc->extent_root;
2004
2005	/*
2006	 * this serializes us with btrfs_record_root_in_transaction,
2007	 * we have to make sure nobody is in the middle of
2008	 * adding their roots to the list while we are
2009	 * doing this splice
2010	 */
2011	mutex_lock(&fs_info->reloc_mutex);
2012	list_splice_init(&rc->reloc_roots, &reloc_roots);
2013	mutex_unlock(&fs_info->reloc_mutex);
2014
2015	while (!list_empty(&reloc_roots)) {
2016		found = 1;
2017		reloc_root = list_entry(reloc_roots.next,
2018					struct btrfs_root, root_list);
2019
2020		root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
2021					 false);
2022		if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2023			if (WARN_ON(IS_ERR(root))) {
2024				/*
2025				 * For recovery we read the fs roots on mount,
2026				 * and if we didn't find the root then we marked
2027				 * the reloc root as a garbage root.  For normal
2028				 * relocation obviously the root should exist in
2029				 * memory.  However there's no reason we can't
2030				 * handle the error properly here just in case.
2031				 */
2032				ret = PTR_ERR(root);
2033				goto out;
2034			}
2035			if (WARN_ON(root->reloc_root != reloc_root)) {
2036				/*
2037				 * This can happen if on-disk metadata has some
2038				 * corruption, e.g. bad reloc tree key offset.
2039				 */
2040				ret = -EINVAL;
2041				goto out;
2042			}
2043			ret = merge_reloc_root(rc, root);
2044			btrfs_put_root(root);
2045			if (ret) {
2046				if (list_empty(&reloc_root->root_list))
2047					list_add_tail(&reloc_root->root_list,
2048						      &reloc_roots);
2049				goto out;
2050			}
2051		} else {
2052			if (!IS_ERR(root)) {
2053				if (root->reloc_root == reloc_root) {
2054					root->reloc_root = NULL;
2055					btrfs_put_root(reloc_root);
2056				}
2057				clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
2058					  &root->state);
2059				btrfs_put_root(root);
2060			}
2061
2062			list_del_init(&reloc_root->root_list);
2063			/* Don't forget to queue this reloc root for cleanup */
2064			list_add_tail(&reloc_root->reloc_dirty_list,
2065				      &rc->dirty_subvol_roots);
2066		}
2067	}
2068
2069	if (found) {
2070		found = 0;
2071		goto again;
2072	}
2073out:
2074	if (ret) {
2075		btrfs_handle_fs_error(fs_info, ret, NULL);
2076		free_reloc_roots(&reloc_roots);
2077
2078		/* new reloc root may be added */
2079		mutex_lock(&fs_info->reloc_mutex);
2080		list_splice_init(&rc->reloc_roots, &reloc_roots);
2081		mutex_unlock(&fs_info->reloc_mutex);
2082		free_reloc_roots(&reloc_roots);
2083	}
2084
2085	/*
2086	 * We used to have
2087	 *
2088	 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2089	 *
2090	 * here, but it's wrong.  If we fail to start the transaction in
2091	 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2092	 * have actually been removed from the reloc_root_tree rb tree.  This is
2093	 * fine because we're bailing here, and we hold a reference on the root
2094	 * for the list that holds it, so these roots will be cleaned up when we
2095	 * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
2096	 * will be cleaned up on unmount.
2097	 *
2098	 * The remaining nodes will be cleaned up by free_reloc_control.
2099	 */
2100}
2101
2102static void free_block_list(struct rb_root *blocks)
2103{
2104	struct tree_block *block;
2105	struct rb_node *rb_node;
2106	while ((rb_node = rb_first(blocks))) {
2107		block = rb_entry(rb_node, struct tree_block, rb_node);
2108		rb_erase(rb_node, blocks);
2109		kfree(block);
2110	}
2111}
2112
2113static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2114				      struct btrfs_root *reloc_root)
2115{
2116	struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2117	struct btrfs_root *root;
2118	int ret;
2119
2120	if (reloc_root->last_trans == trans->transid)
2121		return 0;
2122
2123	root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2124
2125	/*
2126	 * This should succeed, since we can't have a reloc root without having
2127	 * already looked up the actual root and created the reloc root for this
2128	 * root.
2129	 *
2130	 * However if there's some sort of corruption where we have a ref to a
2131	 * reloc root without a corresponding root this could return ENOENT.
2132	 */
2133	if (IS_ERR(root)) {
2134		ASSERT(0);
2135		return PTR_ERR(root);
2136	}
2137	if (root->reloc_root != reloc_root) {
2138		ASSERT(0);
2139		btrfs_err(fs_info,
2140			  "root %llu has two reloc roots associated with it",
2141			  reloc_root->root_key.offset);
2142		btrfs_put_root(root);
2143		return -EUCLEAN;
2144	}
2145	ret = btrfs_record_root_in_trans(trans, root);
2146	btrfs_put_root(root);
2147
2148	return ret;
2149}
2150
2151static noinline_for_stack
2152struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2153				     struct reloc_control *rc,
2154				     struct btrfs_backref_node *node,
2155				     struct btrfs_backref_edge *edges[])
2156{
2157	struct btrfs_backref_node *next;
2158	struct btrfs_root *root;
2159	int index = 0;
2160	int ret;
2161
2162	next = node;
2163	while (1) {
2164		cond_resched();
2165		next = walk_up_backref(next, edges, &index);
2166		root = next->root;
2167
2168		/*
2169		 * If there is no root, then our references for this block are
2170		 * incomplete, as we should be able to walk all the way up to a
2171		 * block that is owned by a root.
2172		 *
2173		 * This path is only for SHAREABLE roots, so if we come upon a
2174		 * non-SHAREABLE root then we have backrefs that resolve
2175		 * improperly.
2176		 *
2177		 * Both of these cases indicate file system corruption, or a bug
2178		 * in the backref walking code.
2179		 */
2180		if (!root) {
2181			ASSERT(0);
2182			btrfs_err(trans->fs_info,
2183		"bytenr %llu doesn't have a backref path ending in a root",
2184				  node->bytenr);
2185			return ERR_PTR(-EUCLEAN);
2186		}
2187		if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2188			ASSERT(0);
2189			btrfs_err(trans->fs_info,
2190	"bytenr %llu has multiple refs with one ending in a non-shareable root",
2191				  node->bytenr);
2192			return ERR_PTR(-EUCLEAN);
2193		}
2194
2195		if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2196			ret = record_reloc_root_in_trans(trans, root);
2197			if (ret)
2198				return ERR_PTR(ret);
2199			break;
2200		}
2201
2202		ret = btrfs_record_root_in_trans(trans, root);
2203		if (ret)
2204			return ERR_PTR(ret);
2205		root = root->reloc_root;
2206
2207		/*
2208		 * We could have raced with another thread which failed, so
2209		 * root->reloc_root may not be set, return ENOENT in this case.
2210		 */
2211		if (!root)
2212			return ERR_PTR(-ENOENT);
2213
2214		if (next->new_bytenr != root->node->start) {
2215			/*
2216			 * We just created the reloc root, so we shouldn't have
2217			 * ->new_bytenr set and this shouldn't be in the changed
2218			 *  list.  If it is then we have multiple roots pointing
2219			 *  at the same bytenr which indicates corruption, or
2220			 *  we've made a mistake in the backref walking code.
2221			 */
2222			ASSERT(next->new_bytenr == 0);
2223			ASSERT(list_empty(&next->list));
2224			if (next->new_bytenr || !list_empty(&next->list)) {
2225				btrfs_err(trans->fs_info,
2226	"bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2227					  node->bytenr, next->bytenr);
2228				return ERR_PTR(-EUCLEAN);
2229			}
2230
2231			next->new_bytenr = root->node->start;
2232			btrfs_put_root(next->root);
2233			next->root = btrfs_grab_root(root);
2234			ASSERT(next->root);
2235			list_add_tail(&next->list,
2236				      &rc->backref_cache.changed);
2237			mark_block_processed(rc, next);
2238			break;
2239		}
2240
2241		WARN_ON(1);
2242		root = NULL;
2243		next = walk_down_backref(edges, &index);
2244		if (!next || next->level <= node->level)
2245			break;
2246	}
2247	if (!root) {
2248		/*
2249		 * This can happen if there's fs corruption or if there's a bug
2250		 * in the backref lookup code.
2251		 */
2252		ASSERT(0);
2253		return ERR_PTR(-ENOENT);
2254	}
2255
2256	next = node;
2257	/* setup backref node path for btrfs_reloc_cow_block */
2258	while (1) {
2259		rc->backref_cache.path[next->level] = next;
2260		if (--index < 0)
2261			break;
2262		next = edges[index]->node[UPPER];
2263	}
2264	return root;
2265}
2266
2267/*
2268 * Select a tree root for relocation.
2269 *
2270 * Return NULL if the block is not shareable. We should use do_relocation() in
2271 * this case.
2272 *
2273 * Return a tree root pointer if the block is shareable.
2274 * Return -ENOENT if the block is root of reloc tree.
2275 */
2276static noinline_for_stack
2277struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2278{
2279	struct btrfs_backref_node *next;
2280	struct btrfs_root *root;
2281	struct btrfs_root *fs_root = NULL;
2282	struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2283	int index = 0;
2284
2285	next = node;
2286	while (1) {
2287		cond_resched();
2288		next = walk_up_backref(next, edges, &index);
2289		root = next->root;
2290
2291		/*
2292		 * This can occur if we have incomplete extent refs leading all
2293		 * the way up a particular path, in this case return -EUCLEAN.
2294		 */
2295		if (!root)
2296			return ERR_PTR(-EUCLEAN);
2297
2298		/* No other choice for non-shareable tree */
2299		if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2300			return root;
2301
2302		if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2303			fs_root = root;
2304
2305		if (next != node)
2306			return NULL;
2307
2308		next = walk_down_backref(edges, &index);
2309		if (!next || next->level <= node->level)
2310			break;
2311	}
2312
2313	if (!fs_root)
2314		return ERR_PTR(-ENOENT);
2315	return fs_root;
2316}
2317
2318static noinline_for_stack
2319u64 calcu_metadata_size(struct reloc_control *rc,
2320			struct btrfs_backref_node *node, int reserve)
2321{
2322	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2323	struct btrfs_backref_node *next = node;
2324	struct btrfs_backref_edge *edge;
2325	struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2326	u64 num_bytes = 0;
2327	int index = 0;
2328
2329	BUG_ON(reserve && node->processed);
2330
2331	while (next) {
2332		cond_resched();
2333		while (1) {
2334			if (next->processed && (reserve || next != node))
2335				break;
2336
2337			num_bytes += fs_info->nodesize;
2338
2339			if (list_empty(&next->upper))
2340				break;
2341
2342			edge = list_entry(next->upper.next,
2343					struct btrfs_backref_edge, list[LOWER]);
2344			edges[index++] = edge;
2345			next = edge->node[UPPER];
2346		}
2347		next = walk_down_backref(edges, &index);
2348	}
2349	return num_bytes;
2350}
2351
2352static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2353				  struct reloc_control *rc,
2354				  struct btrfs_backref_node *node)
2355{
2356	struct btrfs_root *root = rc->extent_root;
2357	struct btrfs_fs_info *fs_info = root->fs_info;
2358	u64 num_bytes;
2359	int ret;
2360	u64 tmp;
2361
2362	num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2363
2364	trans->block_rsv = rc->block_rsv;
2365	rc->reserved_bytes += num_bytes;
2366
2367	/*
2368	 * We are under a transaction here so we can only do limited flushing.
2369	 * If we get an enospc just kick back -EAGAIN so we know to drop the
2370	 * transaction and try to refill when we can flush all the things.
2371	 */
2372	ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv, num_bytes,
2373				     BTRFS_RESERVE_FLUSH_LIMIT);
2374	if (ret) {
2375		tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2376		while (tmp <= rc->reserved_bytes)
2377			tmp <<= 1;
2378		/*
2379		 * only one thread can access block_rsv at this point,
2380		 * so we don't need hold lock to protect block_rsv.
2381		 * we expand more reservation size here to allow enough
2382		 * space for relocation and we will return earlier in
2383		 * enospc case.
2384		 */
2385		rc->block_rsv->size = tmp + fs_info->nodesize *
2386				      RELOCATION_RESERVED_NODES;
2387		return -EAGAIN;
2388	}
2389
2390	return 0;
2391}
2392
2393/*
2394 * relocate a block tree, and then update pointers in upper level
2395 * blocks that reference the block to point to the new location.
2396 *
2397 * if called by link_to_upper, the block has already been relocated.
2398 * in that case this function just updates pointers.
2399 */
2400static int do_relocation(struct btrfs_trans_handle *trans,
2401			 struct reloc_control *rc,
2402			 struct btrfs_backref_node *node,
2403			 struct btrfs_key *key,
2404			 struct btrfs_path *path, int lowest)
2405{
2406	struct btrfs_backref_node *upper;
2407	struct btrfs_backref_edge *edge;
2408	struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2409	struct btrfs_root *root;
2410	struct extent_buffer *eb;
2411	u32 blocksize;
2412	u64 bytenr;
2413	int slot;
2414	int ret = 0;
2415
2416	/*
2417	 * If we are lowest then this is the first time we're processing this
2418	 * block, and thus shouldn't have an eb associated with it yet.
2419	 */
2420	ASSERT(!lowest || !node->eb);
2421
2422	path->lowest_level = node->level + 1;
2423	rc->backref_cache.path[node->level] = node;
2424	list_for_each_entry(edge, &node->upper, list[LOWER]) {
2425		struct btrfs_ref ref = { 0 };
2426
2427		cond_resched();
2428
2429		upper = edge->node[UPPER];
2430		root = select_reloc_root(trans, rc, upper, edges);
2431		if (IS_ERR(root)) {
2432			ret = PTR_ERR(root);
2433			goto next;
2434		}
2435
2436		if (upper->eb && !upper->locked) {
2437			if (!lowest) {
2438				ret = btrfs_bin_search(upper->eb, 0, key, &slot);
2439				if (ret < 0)
2440					goto next;
2441				BUG_ON(ret);
2442				bytenr = btrfs_node_blockptr(upper->eb, slot);
2443				if (node->eb->start == bytenr)
2444					goto next;
2445			}
2446			btrfs_backref_drop_node_buffer(upper);
2447		}
2448
2449		if (!upper->eb) {
2450			ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2451			if (ret) {
2452				if (ret > 0)
2453					ret = -ENOENT;
2454
2455				btrfs_release_path(path);
2456				break;
2457			}
2458
2459			if (!upper->eb) {
2460				upper->eb = path->nodes[upper->level];
2461				path->nodes[upper->level] = NULL;
2462			} else {
2463				BUG_ON(upper->eb != path->nodes[upper->level]);
2464			}
2465
2466			upper->locked = 1;
2467			path->locks[upper->level] = 0;
2468
2469			slot = path->slots[upper->level];
2470			btrfs_release_path(path);
2471		} else {
2472			ret = btrfs_bin_search(upper->eb, 0, key, &slot);
2473			if (ret < 0)
2474				goto next;
2475			BUG_ON(ret);
2476		}
2477
2478		bytenr = btrfs_node_blockptr(upper->eb, slot);
2479		if (lowest) {
2480			if (bytenr != node->bytenr) {
2481				btrfs_err(root->fs_info,
2482		"lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2483					  bytenr, node->bytenr, slot,
2484					  upper->eb->start);
2485				ret = -EIO;
2486				goto next;
2487			}
2488		} else {
2489			if (node->eb->start == bytenr)
2490				goto next;
2491		}
2492
2493		blocksize = root->fs_info->nodesize;
2494		eb = btrfs_read_node_slot(upper->eb, slot);
2495		if (IS_ERR(eb)) {
2496			ret = PTR_ERR(eb);
2497			goto next;
2498		}
2499		btrfs_tree_lock(eb);
2500
2501		if (!node->eb) {
2502			ret = btrfs_cow_block(trans, root, eb, upper->eb,
2503					      slot, &eb, BTRFS_NESTING_COW);
2504			btrfs_tree_unlock(eb);
2505			free_extent_buffer(eb);
2506			if (ret < 0)
2507				goto next;
2508			/*
2509			 * We've just COWed this block, it should have updated
2510			 * the correct backref node entry.
2511			 */
2512			ASSERT(node->eb == eb);
2513		} else {
2514			btrfs_set_node_blockptr(upper->eb, slot,
2515						node->eb->start);
2516			btrfs_set_node_ptr_generation(upper->eb, slot,
2517						      trans->transid);
2518			btrfs_mark_buffer_dirty(trans, upper->eb);
2519
2520			btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2521					       node->eb->start, blocksize,
2522					       upper->eb->start,
2523					       btrfs_header_owner(upper->eb));
2524			btrfs_init_tree_ref(&ref, node->level,
2525					    btrfs_header_owner(upper->eb),
2526					    root->root_key.objectid, false);
2527			ret = btrfs_inc_extent_ref(trans, &ref);
2528			if (!ret)
2529				ret = btrfs_drop_subtree(trans, root, eb,
2530							 upper->eb);
2531			if (ret)
2532				btrfs_abort_transaction(trans, ret);
2533		}
2534next:
2535		if (!upper->pending)
2536			btrfs_backref_drop_node_buffer(upper);
2537		else
2538			btrfs_backref_unlock_node_buffer(upper);
2539		if (ret)
2540			break;
2541	}
2542
2543	if (!ret && node->pending) {
2544		btrfs_backref_drop_node_buffer(node);
2545		list_move_tail(&node->list, &rc->backref_cache.changed);
2546		node->pending = 0;
2547	}
2548
2549	path->lowest_level = 0;
2550
2551	/*
2552	 * We should have allocated all of our space in the block rsv and thus
2553	 * shouldn't ENOSPC.
2554	 */
2555	ASSERT(ret != -ENOSPC);
2556	return ret;
2557}
2558
2559static int link_to_upper(struct btrfs_trans_handle *trans,
2560			 struct reloc_control *rc,
2561			 struct btrfs_backref_node *node,
2562			 struct btrfs_path *path)
2563{
2564	struct btrfs_key key;
2565
2566	btrfs_node_key_to_cpu(node->eb, &key, 0);
2567	return do_relocation(trans, rc, node, &key, path, 0);
2568}
2569
2570static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2571				struct reloc_control *rc,
2572				struct btrfs_path *path, int err)
2573{
2574	LIST_HEAD(list);
2575	struct btrfs_backref_cache *cache = &rc->backref_cache;
2576	struct btrfs_backref_node *node;
2577	int level;
2578	int ret;
2579
2580	for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2581		while (!list_empty(&cache->pending[level])) {
2582			node = list_entry(cache->pending[level].next,
2583					  struct btrfs_backref_node, list);
2584			list_move_tail(&node->list, &list);
2585			BUG_ON(!node->pending);
2586
2587			if (!err) {
2588				ret = link_to_upper(trans, rc, node, path);
2589				if (ret < 0)
2590					err = ret;
2591			}
2592		}
2593		list_splice_init(&list, &cache->pending[level]);
2594	}
2595	return err;
2596}
2597
2598/*
2599 * mark a block and all blocks directly/indirectly reference the block
2600 * as processed.
2601 */
2602static void update_processed_blocks(struct reloc_control *rc,
2603				    struct btrfs_backref_node *node)
2604{
2605	struct btrfs_backref_node *next = node;
2606	struct btrfs_backref_edge *edge;
2607	struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2608	int index = 0;
2609
2610	while (next) {
2611		cond_resched();
2612		while (1) {
2613			if (next->processed)
2614				break;
2615
2616			mark_block_processed(rc, next);
2617
2618			if (list_empty(&next->upper))
2619				break;
2620
2621			edge = list_entry(next->upper.next,
2622					struct btrfs_backref_edge, list[LOWER]);
2623			edges[index++] = edge;
2624			next = edge->node[UPPER];
2625		}
2626		next = walk_down_backref(edges, &index);
2627	}
2628}
2629
2630static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2631{
2632	u32 blocksize = rc->extent_root->fs_info->nodesize;
2633
2634	if (test_range_bit(&rc->processed_blocks, bytenr,
2635			   bytenr + blocksize - 1, EXTENT_DIRTY, NULL))
2636		return 1;
2637	return 0;
2638}
2639
2640static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2641			      struct tree_block *block)
2642{
2643	struct btrfs_tree_parent_check check = {
2644		.level = block->level,
2645		.owner_root = block->owner,
2646		.transid = block->key.offset
2647	};
2648	struct extent_buffer *eb;
2649
2650	eb = read_tree_block(fs_info, block->bytenr, &check);
2651	if (IS_ERR(eb))
2652		return PTR_ERR(eb);
2653	if (!extent_buffer_uptodate(eb)) {
2654		free_extent_buffer(eb);
2655		return -EIO;
2656	}
2657	if (block->level == 0)
2658		btrfs_item_key_to_cpu(eb, &block->key, 0);
2659	else
2660		btrfs_node_key_to_cpu(eb, &block->key, 0);
2661	free_extent_buffer(eb);
2662	block->key_ready = true;
2663	return 0;
2664}
2665
2666/*
2667 * helper function to relocate a tree block
2668 */
2669static int relocate_tree_block(struct btrfs_trans_handle *trans,
2670				struct reloc_control *rc,
2671				struct btrfs_backref_node *node,
2672				struct btrfs_key *key,
2673				struct btrfs_path *path)
2674{
2675	struct btrfs_root *root;
2676	int ret = 0;
2677
2678	if (!node)
2679		return 0;
2680
2681	/*
2682	 * If we fail here we want to drop our backref_node because we are going
2683	 * to start over and regenerate the tree for it.
2684	 */
2685	ret = reserve_metadata_space(trans, rc, node);
2686	if (ret)
2687		goto out;
2688
2689	BUG_ON(node->processed);
2690	root = select_one_root(node);
2691	if (IS_ERR(root)) {
2692		ret = PTR_ERR(root);
2693
2694		/* See explanation in select_one_root for the -EUCLEAN case. */
2695		ASSERT(ret == -ENOENT);
2696		if (ret == -ENOENT) {
2697			ret = 0;
2698			update_processed_blocks(rc, node);
2699		}
2700		goto out;
2701	}
2702
2703	if (root) {
2704		if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2705			/*
2706			 * This block was the root block of a root, and this is
2707			 * the first time we're processing the block and thus it
2708			 * should not have had the ->new_bytenr modified and
2709			 * should have not been included on the changed list.
2710			 *
2711			 * However in the case of corruption we could have
2712			 * multiple refs pointing to the same block improperly,
2713			 * and thus we would trip over these checks.  ASSERT()
2714			 * for the developer case, because it could indicate a
2715			 * bug in the backref code, however error out for a
2716			 * normal user in the case of corruption.
2717			 */
2718			ASSERT(node->new_bytenr == 0);
2719			ASSERT(list_empty(&node->list));
2720			if (node->new_bytenr || !list_empty(&node->list)) {
2721				btrfs_err(root->fs_info,
2722				  "bytenr %llu has improper references to it",
2723					  node->bytenr);
2724				ret = -EUCLEAN;
2725				goto out;
2726			}
2727			ret = btrfs_record_root_in_trans(trans, root);
2728			if (ret)
2729				goto out;
2730			/*
2731			 * Another thread could have failed, need to check if we
2732			 * have reloc_root actually set.
2733			 */
2734			if (!root->reloc_root) {
2735				ret = -ENOENT;
2736				goto out;
2737			}
2738			root = root->reloc_root;
2739			node->new_bytenr = root->node->start;
2740			btrfs_put_root(node->root);
2741			node->root = btrfs_grab_root(root);
2742			ASSERT(node->root);
2743			list_add_tail(&node->list, &rc->backref_cache.changed);
2744		} else {
2745			path->lowest_level = node->level;
2746			if (root == root->fs_info->chunk_root)
2747				btrfs_reserve_chunk_metadata(trans, false);
2748			ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2749			btrfs_release_path(path);
2750			if (root == root->fs_info->chunk_root)
2751				btrfs_trans_release_chunk_metadata(trans);
2752			if (ret > 0)
2753				ret = 0;
2754		}
2755		if (!ret)
2756			update_processed_blocks(rc, node);
2757	} else {
2758		ret = do_relocation(trans, rc, node, key, path, 1);
2759	}
2760out:
2761	if (ret || node->level == 0 || node->cowonly)
2762		btrfs_backref_cleanup_node(&rc->backref_cache, node);
2763	return ret;
2764}
2765
2766/*
2767 * relocate a list of blocks
2768 */
2769static noinline_for_stack
2770int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2771			 struct reloc_control *rc, struct rb_root *blocks)
2772{
2773	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2774	struct btrfs_backref_node *node;
2775	struct btrfs_path *path;
2776	struct tree_block *block;
2777	struct tree_block *next;
2778	int ret;
2779	int err = 0;
2780
2781	path = btrfs_alloc_path();
2782	if (!path) {
2783		err = -ENOMEM;
2784		goto out_free_blocks;
2785	}
2786
2787	/* Kick in readahead for tree blocks with missing keys */
2788	rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2789		if (!block->key_ready)
2790			btrfs_readahead_tree_block(fs_info, block->bytenr,
2791						   block->owner, 0,
2792						   block->level);
2793	}
2794
2795	/* Get first keys */
2796	rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2797		if (!block->key_ready) {
2798			err = get_tree_block_key(fs_info, block);
2799			if (err)
2800				goto out_free_path;
2801		}
2802	}
2803
2804	/* Do tree relocation */
2805	rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2806		node = build_backref_tree(trans, rc, &block->key,
2807					  block->level, block->bytenr);
2808		if (IS_ERR(node)) {
2809			err = PTR_ERR(node);
2810			goto out;
2811		}
2812
2813		ret = relocate_tree_block(trans, rc, node, &block->key,
2814					  path);
2815		if (ret < 0) {
2816			err = ret;
2817			break;
2818		}
2819	}
2820out:
2821	err = finish_pending_nodes(trans, rc, path, err);
2822
2823out_free_path:
2824	btrfs_free_path(path);
2825out_free_blocks:
2826	free_block_list(blocks);
2827	return err;
2828}
2829
2830static noinline_for_stack int prealloc_file_extent_cluster(
2831				struct btrfs_inode *inode,
2832				const struct file_extent_cluster *cluster)
2833{
2834	u64 alloc_hint = 0;
2835	u64 start;
2836	u64 end;
2837	u64 offset = inode->index_cnt;
2838	u64 num_bytes;
2839	int nr;
2840	int ret = 0;
2841	u64 i_size = i_size_read(&inode->vfs_inode);
2842	u64 prealloc_start = cluster->start - offset;
2843	u64 prealloc_end = cluster->end - offset;
2844	u64 cur_offset = prealloc_start;
2845
2846	/*
2847	 * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2848	 * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2849	 * btrfs_do_readpage() call of previously relocated file cluster.
2850	 *
2851	 * If the current cluster starts in the above range, btrfs_do_readpage()
2852	 * will skip the read, and relocate_one_page() will later writeback
2853	 * the padding zeros as new data, causing data corruption.
2854	 *
2855	 * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2856	 */
2857	if (!PAGE_ALIGNED(i_size)) {
2858		struct address_space *mapping = inode->vfs_inode.i_mapping;
2859		struct btrfs_fs_info *fs_info = inode->root->fs_info;
2860		const u32 sectorsize = fs_info->sectorsize;
2861		struct page *page;
2862
2863		ASSERT(sectorsize < PAGE_SIZE);
2864		ASSERT(IS_ALIGNED(i_size, sectorsize));
2865
2866		/*
2867		 * Subpage can't handle page with DIRTY but without UPTODATE
2868		 * bit as it can lead to the following deadlock:
2869		 *
2870		 * btrfs_read_folio()
2871		 * | Page already *locked*
2872		 * |- btrfs_lock_and_flush_ordered_range()
2873		 *    |- btrfs_start_ordered_extent()
2874		 *       |- extent_write_cache_pages()
2875		 *          |- lock_page()
2876		 *             We try to lock the page we already hold.
2877		 *
2878		 * Here we just writeback the whole data reloc inode, so that
2879		 * we will be ensured to have no dirty range in the page, and
2880		 * are safe to clear the uptodate bits.
2881		 *
2882		 * This shouldn't cause too much overhead, as we need to write
2883		 * the data back anyway.
2884		 */
2885		ret = filemap_write_and_wait(mapping);
2886		if (ret < 0)
2887			return ret;
2888
2889		clear_extent_bits(&inode->io_tree, i_size,
2890				  round_up(i_size, PAGE_SIZE) - 1,
2891				  EXTENT_UPTODATE);
2892		page = find_lock_page(mapping, i_size >> PAGE_SHIFT);
2893		/*
2894		 * If page is freed we don't need to do anything then, as we
2895		 * will re-read the whole page anyway.
2896		 */
2897		if (page) {
2898			btrfs_subpage_clear_uptodate(fs_info, page_folio(page), i_size,
2899					round_up(i_size, PAGE_SIZE) - i_size);
2900			unlock_page(page);
2901			put_page(page);
2902		}
2903	}
2904
2905	BUG_ON(cluster->start != cluster->boundary[0]);
2906	ret = btrfs_alloc_data_chunk_ondemand(inode,
2907					      prealloc_end + 1 - prealloc_start);
2908	if (ret)
2909		return ret;
2910
2911	btrfs_inode_lock(inode, 0);
2912	for (nr = 0; nr < cluster->nr; nr++) {
2913		struct extent_state *cached_state = NULL;
2914
2915		start = cluster->boundary[nr] - offset;
2916		if (nr + 1 < cluster->nr)
2917			end = cluster->boundary[nr + 1] - 1 - offset;
2918		else
2919			end = cluster->end - offset;
2920
2921		lock_extent(&inode->io_tree, start, end, &cached_state);
2922		num_bytes = end + 1 - start;
2923		ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2924						num_bytes, num_bytes,
2925						end + 1, &alloc_hint);
2926		cur_offset = end + 1;
2927		unlock_extent(&inode->io_tree, start, end, &cached_state);
2928		if (ret)
2929			break;
2930	}
2931	btrfs_inode_unlock(inode, 0);
2932
2933	if (cur_offset < prealloc_end)
2934		btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2935					       prealloc_end + 1 - cur_offset);
2936	return ret;
2937}
2938
2939static noinline_for_stack int setup_relocation_extent_mapping(struct inode *inode,
2940				u64 start, u64 end, u64 block_start)
2941{
2942	struct extent_map *em;
2943	struct extent_state *cached_state = NULL;
2944	int ret = 0;
2945
2946	em = alloc_extent_map();
2947	if (!em)
2948		return -ENOMEM;
2949
2950	em->start = start;
2951	em->len = end + 1 - start;
2952	em->block_len = em->len;
2953	em->block_start = block_start;
2954	em->flags |= EXTENT_FLAG_PINNED;
2955
2956	lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2957	ret = btrfs_replace_extent_map_range(BTRFS_I(inode), em, false);
2958	unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2959	free_extent_map(em);
2960
2961	return ret;
2962}
2963
2964/*
2965 * Allow error injection to test balance/relocation cancellation
2966 */
2967noinline int btrfs_should_cancel_balance(const struct btrfs_fs_info *fs_info)
2968{
2969	return atomic_read(&fs_info->balance_cancel_req) ||
2970		atomic_read(&fs_info->reloc_cancel_req) ||
2971		fatal_signal_pending(current);
2972}
2973ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2974
2975static u64 get_cluster_boundary_end(const struct file_extent_cluster *cluster,
2976				    int cluster_nr)
2977{
2978	/* Last extent, use cluster end directly */
2979	if (cluster_nr >= cluster->nr - 1)
2980		return cluster->end;
2981
2982	/* Use next boundary start*/
2983	return cluster->boundary[cluster_nr + 1] - 1;
2984}
2985
2986static int relocate_one_page(struct inode *inode, struct file_ra_state *ra,
2987			     const struct file_extent_cluster *cluster,
2988			     int *cluster_nr, unsigned long page_index)
2989{
2990	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2991	u64 offset = BTRFS_I(inode)->index_cnt;
2992	const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2993	gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2994	struct page *page;
2995	u64 page_start;
2996	u64 page_end;
2997	u64 cur;
2998	int ret;
2999
3000	ASSERT(page_index <= last_index);
3001	page = find_lock_page(inode->i_mapping, page_index);
3002	if (!page) {
3003		page_cache_sync_readahead(inode->i_mapping, ra, NULL,
3004				page_index, last_index + 1 - page_index);
3005		page = find_or_create_page(inode->i_mapping, page_index, mask);
3006		if (!page)
3007			return -ENOMEM;
3008	}
3009
3010	if (PageReadahead(page))
3011		page_cache_async_readahead(inode->i_mapping, ra, NULL,
3012				page_folio(page), page_index,
3013				last_index + 1 - page_index);
3014
3015	if (!PageUptodate(page)) {
3016		btrfs_read_folio(NULL, page_folio(page));
3017		lock_page(page);
3018		if (!PageUptodate(page)) {
3019			ret = -EIO;
3020			goto release_page;
3021		}
3022	}
3023
3024	/*
3025	 * We could have lost page private when we dropped the lock to read the
3026	 * page above, make sure we set_page_extent_mapped here so we have any
3027	 * of the subpage blocksize stuff we need in place.
3028	 */
3029	ret = set_page_extent_mapped(page);
3030	if (ret < 0)
3031		goto release_page;
3032
3033	page_start = page_offset(page);
3034	page_end = page_start + PAGE_SIZE - 1;
3035
3036	/*
3037	 * Start from the cluster, as for subpage case, the cluster can start
3038	 * inside the page.
3039	 */
3040	cur = max(page_start, cluster->boundary[*cluster_nr] - offset);
3041	while (cur <= page_end) {
3042		struct extent_state *cached_state = NULL;
3043		u64 extent_start = cluster->boundary[*cluster_nr] - offset;
3044		u64 extent_end = get_cluster_boundary_end(cluster,
3045						*cluster_nr) - offset;
3046		u64 clamped_start = max(page_start, extent_start);
3047		u64 clamped_end = min(page_end, extent_end);
3048		u32 clamped_len = clamped_end + 1 - clamped_start;
3049
3050		/* Reserve metadata for this range */
3051		ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3052						      clamped_len, clamped_len,
3053						      false);
3054		if (ret)
3055			goto release_page;
3056
3057		/* Mark the range delalloc and dirty for later writeback */
3058		lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3059			    &cached_state);
3060		ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
3061						clamped_end, 0, &cached_state);
3062		if (ret) {
3063			clear_extent_bit(&BTRFS_I(inode)->io_tree,
3064					 clamped_start, clamped_end,
3065					 EXTENT_LOCKED | EXTENT_BOUNDARY,
3066					 &cached_state);
3067			btrfs_delalloc_release_metadata(BTRFS_I(inode),
3068							clamped_len, true);
3069			btrfs_delalloc_release_extents(BTRFS_I(inode),
3070						       clamped_len);
3071			goto release_page;
3072		}
3073		btrfs_folio_set_dirty(fs_info, page_folio(page),
3074				      clamped_start, clamped_len);
3075
3076		/*
3077		 * Set the boundary if it's inside the page.
3078		 * Data relocation requires the destination extents to have the
3079		 * same size as the source.
3080		 * EXTENT_BOUNDARY bit prevents current extent from being merged
3081		 * with previous extent.
3082		 */
3083		if (in_range(cluster->boundary[*cluster_nr] - offset,
3084			     page_start, PAGE_SIZE)) {
3085			u64 boundary_start = cluster->boundary[*cluster_nr] -
3086						offset;
3087			u64 boundary_end = boundary_start +
3088					   fs_info->sectorsize - 1;
3089
3090			set_extent_bit(&BTRFS_I(inode)->io_tree,
3091				       boundary_start, boundary_end,
3092				       EXTENT_BOUNDARY, NULL);
3093		}
3094		unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3095			      &cached_state);
3096		btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3097		cur += clamped_len;
3098
3099		/* Crossed extent end, go to next extent */
3100		if (cur >= extent_end) {
3101			(*cluster_nr)++;
3102			/* Just finished the last extent of the cluster, exit. */
3103			if (*cluster_nr >= cluster->nr)
3104				break;
3105		}
3106	}
3107	unlock_page(page);
3108	put_page(page);
3109
3110	balance_dirty_pages_ratelimited(inode->i_mapping);
3111	btrfs_throttle(fs_info);
3112	if (btrfs_should_cancel_balance(fs_info))
3113		ret = -ECANCELED;
3114	return ret;
3115
3116release_page:
3117	unlock_page(page);
3118	put_page(page);
3119	return ret;
3120}
3121
3122static int relocate_file_extent_cluster(struct inode *inode,
3123					const struct file_extent_cluster *cluster)
3124{
3125	u64 offset = BTRFS_I(inode)->index_cnt;
3126	unsigned long index;
3127	unsigned long last_index;
3128	struct file_ra_state *ra;
3129	int cluster_nr = 0;
3130	int ret = 0;
3131
3132	if (!cluster->nr)
3133		return 0;
3134
3135	ra = kzalloc(sizeof(*ra), GFP_NOFS);
3136	if (!ra)
3137		return -ENOMEM;
3138
3139	ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
3140	if (ret)
3141		goto out;
3142
3143	file_ra_state_init(ra, inode->i_mapping);
3144
3145	ret = setup_relocation_extent_mapping(inode, cluster->start - offset,
3146				   cluster->end - offset, cluster->start);
3147	if (ret)
3148		goto out;
3149
3150	last_index = (cluster->end - offset) >> PAGE_SHIFT;
3151	for (index = (cluster->start - offset) >> PAGE_SHIFT;
3152	     index <= last_index && !ret; index++)
3153		ret = relocate_one_page(inode, ra, cluster, &cluster_nr, index);
3154	if (ret == 0)
3155		WARN_ON(cluster_nr != cluster->nr);
3156out:
3157	kfree(ra);
3158	return ret;
3159}
3160
3161static noinline_for_stack int relocate_data_extent(struct inode *inode,
3162				const struct btrfs_key *extent_key,
3163				struct file_extent_cluster *cluster)
3164{
3165	int ret;
3166	struct btrfs_root *root = BTRFS_I(inode)->root;
3167
3168	if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3169		ret = relocate_file_extent_cluster(inode, cluster);
3170		if (ret)
3171			return ret;
3172		cluster->nr = 0;
3173	}
3174
3175	/*
3176	 * Under simple quotas, we set root->relocation_src_root when we find
3177	 * the extent. If adjacent extents have different owners, we can't merge
3178	 * them while relocating. Handle this by storing the owning root that
3179	 * started a cluster and if we see an extent from a different root break
3180	 * cluster formation (just like the above case of non-adjacent extents).
3181	 *
3182	 * Without simple quotas, relocation_src_root is always 0, so we should
3183	 * never see a mismatch, and it should have no effect on relocation
3184	 * clusters.
3185	 */
3186	if (cluster->nr > 0 && cluster->owning_root != root->relocation_src_root) {
3187		u64 tmp = root->relocation_src_root;
3188
3189		/*
3190		 * root->relocation_src_root is the state that actually affects
3191		 * the preallocation we do here, so set it to the root owning
3192		 * the cluster we need to relocate.
3193		 */
3194		root->relocation_src_root = cluster->owning_root;
3195		ret = relocate_file_extent_cluster(inode, cluster);
3196		if (ret)
3197			return ret;
3198		cluster->nr = 0;
3199		/* And reset it back for the current extent's owning root. */
3200		root->relocation_src_root = tmp;
3201	}
3202
3203	if (!cluster->nr) {
3204		cluster->start = extent_key->objectid;
3205		cluster->owning_root = root->relocation_src_root;
3206	}
3207	else
3208		BUG_ON(cluster->nr >= MAX_EXTENTS);
3209	cluster->end = extent_key->objectid + extent_key->offset - 1;
3210	cluster->boundary[cluster->nr] = extent_key->objectid;
3211	cluster->nr++;
3212
3213	if (cluster->nr >= MAX_EXTENTS) {
3214		ret = relocate_file_extent_cluster(inode, cluster);
3215		if (ret)
3216			return ret;
3217		cluster->nr = 0;
3218	}
3219	return 0;
3220}
3221
3222/*
3223 * helper to add a tree block to the list.
3224 * the major work is getting the generation and level of the block
3225 */
3226static int add_tree_block(struct reloc_control *rc,
3227			  const struct btrfs_key *extent_key,
3228			  struct btrfs_path *path,
3229			  struct rb_root *blocks)
3230{
3231	struct extent_buffer *eb;
3232	struct btrfs_extent_item *ei;
3233	struct btrfs_tree_block_info *bi;
3234	struct tree_block *block;
3235	struct rb_node *rb_node;
3236	u32 item_size;
3237	int level = -1;
3238	u64 generation;
3239	u64 owner = 0;
3240
3241	eb =  path->nodes[0];
3242	item_size = btrfs_item_size(eb, path->slots[0]);
3243
3244	if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3245	    item_size >= sizeof(*ei) + sizeof(*bi)) {
3246		unsigned long ptr = 0, end;
3247
3248		ei = btrfs_item_ptr(eb, path->slots[0],
3249				struct btrfs_extent_item);
3250		end = (unsigned long)ei + item_size;
3251		if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3252			bi = (struct btrfs_tree_block_info *)(ei + 1);
3253			level = btrfs_tree_block_level(eb, bi);
3254			ptr = (unsigned long)(bi + 1);
3255		} else {
3256			level = (int)extent_key->offset;
3257			ptr = (unsigned long)(ei + 1);
3258		}
3259		generation = btrfs_extent_generation(eb, ei);
3260
3261		/*
3262		 * We're reading random blocks without knowing their owner ahead
3263		 * of time.  This is ok most of the time, as all reloc roots and
3264		 * fs roots have the same lock type.  However normal trees do
3265		 * not, and the only way to know ahead of time is to read the
3266		 * inline ref offset.  We know it's an fs root if
3267		 *
3268		 * 1. There's more than one ref.
3269		 * 2. There's a SHARED_DATA_REF_KEY set.
3270		 * 3. FULL_BACKREF is set on the flags.
3271		 *
3272		 * Otherwise it's safe to assume that the ref offset == the
3273		 * owner of this block, so we can use that when calling
3274		 * read_tree_block.
3275		 */
3276		if (btrfs_extent_refs(eb, ei) == 1 &&
3277		    !(btrfs_extent_flags(eb, ei) &
3278		      BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3279		    ptr < end) {
3280			struct btrfs_extent_inline_ref *iref;
3281			int type;
3282
3283			iref = (struct btrfs_extent_inline_ref *)ptr;
3284			type = btrfs_get_extent_inline_ref_type(eb, iref,
3285							BTRFS_REF_TYPE_BLOCK);
3286			if (type == BTRFS_REF_TYPE_INVALID)
3287				return -EINVAL;
3288			if (type == BTRFS_TREE_BLOCK_REF_KEY)
3289				owner = btrfs_extent_inline_ref_offset(eb, iref);
3290		}
3291	} else {
3292		btrfs_print_leaf(eb);
3293		btrfs_err(rc->block_group->fs_info,
3294			  "unrecognized tree backref at tree block %llu slot %u",
3295			  eb->start, path->slots[0]);
3296		btrfs_release_path(path);
3297		return -EUCLEAN;
3298	}
3299
3300	btrfs_release_path(path);
3301
3302	BUG_ON(level == -1);
3303
3304	block = kmalloc(sizeof(*block), GFP_NOFS);
3305	if (!block)
3306		return -ENOMEM;
3307
3308	block->bytenr = extent_key->objectid;
3309	block->key.objectid = rc->extent_root->fs_info->nodesize;
3310	block->key.offset = generation;
3311	block->level = level;
3312	block->key_ready = false;
3313	block->owner = owner;
3314
3315	rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3316	if (rb_node)
3317		btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3318				    -EEXIST);
3319
3320	return 0;
3321}
3322
3323/*
3324 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3325 */
3326static int __add_tree_block(struct reloc_control *rc,
3327			    u64 bytenr, u32 blocksize,
3328			    struct rb_root *blocks)
3329{
3330	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3331	struct btrfs_path *path;
3332	struct btrfs_key key;
3333	int ret;
3334	bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3335
3336	if (tree_block_processed(bytenr, rc))
3337		return 0;
3338
3339	if (rb_simple_search(blocks, bytenr))
3340		return 0;
3341
3342	path = btrfs_alloc_path();
3343	if (!path)
3344		return -ENOMEM;
3345again:
3346	key.objectid = bytenr;
3347	if (skinny) {
3348		key.type = BTRFS_METADATA_ITEM_KEY;
3349		key.offset = (u64)-1;
3350	} else {
3351		key.type = BTRFS_EXTENT_ITEM_KEY;
3352		key.offset = blocksize;
3353	}
3354
3355	path->search_commit_root = 1;
3356	path->skip_locking = 1;
3357	ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3358	if (ret < 0)
3359		goto out;
3360
3361	if (ret > 0 && skinny) {
3362		if (path->slots[0]) {
3363			path->slots[0]--;
3364			btrfs_item_key_to_cpu(path->nodes[0], &key,
3365					      path->slots[0]);
3366			if (key.objectid == bytenr &&
3367			    (key.type == BTRFS_METADATA_ITEM_KEY ||
3368			     (key.type == BTRFS_EXTENT_ITEM_KEY &&
3369			      key.offset == blocksize)))
3370				ret = 0;
3371		}
3372
3373		if (ret) {
3374			skinny = false;
3375			btrfs_release_path(path);
3376			goto again;
3377		}
3378	}
3379	if (ret) {
3380		ASSERT(ret == 1);
3381		btrfs_print_leaf(path->nodes[0]);
3382		btrfs_err(fs_info,
3383	     "tree block extent item (%llu) is not found in extent tree",
3384		     bytenr);
3385		WARN_ON(1);
3386		ret = -EINVAL;
3387		goto out;
3388	}
3389
3390	ret = add_tree_block(rc, &key, path, blocks);
3391out:
3392	btrfs_free_path(path);
3393	return ret;
3394}
3395
3396static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3397				    struct btrfs_block_group *block_group,
3398				    struct inode *inode,
3399				    u64 ino)
3400{
3401	struct btrfs_root *root = fs_info->tree_root;
3402	struct btrfs_trans_handle *trans;
3403	int ret = 0;
3404
3405	if (inode)
3406		goto truncate;
3407
3408	inode = btrfs_iget(fs_info->sb, ino, root);
3409	if (IS_ERR(inode))
3410		return -ENOENT;
3411
3412truncate:
3413	ret = btrfs_check_trunc_cache_free_space(fs_info,
3414						 &fs_info->global_block_rsv);
3415	if (ret)
3416		goto out;
3417
3418	trans = btrfs_join_transaction(root);
3419	if (IS_ERR(trans)) {
3420		ret = PTR_ERR(trans);
3421		goto out;
3422	}
3423
3424	ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3425
3426	btrfs_end_transaction(trans);
3427	btrfs_btree_balance_dirty(fs_info);
3428out:
3429	iput(inode);
3430	return ret;
3431}
3432
3433/*
3434 * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3435 * cache inode, to avoid free space cache data extent blocking data relocation.
3436 */
3437static int delete_v1_space_cache(struct extent_buffer *leaf,
3438				 struct btrfs_block_group *block_group,
3439				 u64 data_bytenr)
3440{
3441	u64 space_cache_ino;
3442	struct btrfs_file_extent_item *ei;
3443	struct btrfs_key key;
3444	bool found = false;
3445	int i;
3446	int ret;
3447
3448	if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3449		return 0;
3450
3451	for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3452		u8 type;
3453
3454		btrfs_item_key_to_cpu(leaf, &key, i);
3455		if (key.type != BTRFS_EXTENT_DATA_KEY)
3456			continue;
3457		ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3458		type = btrfs_file_extent_type(leaf, ei);
3459
3460		if ((type == BTRFS_FILE_EXTENT_REG ||
3461		     type == BTRFS_FILE_EXTENT_PREALLOC) &&
3462		    btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3463			found = true;
3464			space_cache_ino = key.objectid;
3465			break;
3466		}
3467	}
3468	if (!found)
3469		return -ENOENT;
3470	ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3471					space_cache_ino);
3472	return ret;
3473}
3474
3475/*
3476 * helper to find all tree blocks that reference a given data extent
3477 */
3478static noinline_for_stack int add_data_references(struct reloc_control *rc,
3479						  const struct btrfs_key *extent_key,
3480						  struct btrfs_path *path,
3481						  struct rb_root *blocks)
3482{
3483	struct btrfs_backref_walk_ctx ctx = { 0 };
3484	struct ulist_iterator leaf_uiter;
3485	struct ulist_node *ref_node = NULL;
3486	const u32 blocksize = rc->extent_root->fs_info->nodesize;
3487	int ret = 0;
3488
3489	btrfs_release_path(path);
3490
3491	ctx.bytenr = extent_key->objectid;
3492	ctx.skip_inode_ref_list = true;
3493	ctx.fs_info = rc->extent_root->fs_info;
3494
3495	ret = btrfs_find_all_leafs(&ctx);
3496	if (ret < 0)
3497		return ret;
3498
3499	ULIST_ITER_INIT(&leaf_uiter);
3500	while ((ref_node = ulist_next(ctx.refs, &leaf_uiter))) {
3501		struct btrfs_tree_parent_check check = { 0 };
3502		struct extent_buffer *eb;
3503
3504		eb = read_tree_block(ctx.fs_info, ref_node->val, &check);
3505		if (IS_ERR(eb)) {
3506			ret = PTR_ERR(eb);
3507			break;
3508		}
3509		ret = delete_v1_space_cache(eb, rc->block_group,
3510					    extent_key->objectid);
3511		free_extent_buffer(eb);
3512		if (ret < 0)
3513			break;
3514		ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3515		if (ret < 0)
3516			break;
3517	}
3518	if (ret < 0)
3519		free_block_list(blocks);
3520	ulist_free(ctx.refs);
3521	return ret;
3522}
3523
3524/*
3525 * helper to find next unprocessed extent
3526 */
3527static noinline_for_stack
3528int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3529		     struct btrfs_key *extent_key)
3530{
3531	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3532	struct btrfs_key key;
3533	struct extent_buffer *leaf;
3534	u64 start, end, last;
3535	int ret;
3536
3537	last = rc->block_group->start + rc->block_group->length;
3538	while (1) {
3539		bool block_found;
3540
3541		cond_resched();
3542		if (rc->search_start >= last) {
3543			ret = 1;
3544			break;
3545		}
3546
3547		key.objectid = rc->search_start;
3548		key.type = BTRFS_EXTENT_ITEM_KEY;
3549		key.offset = 0;
3550
3551		path->search_commit_root = 1;
3552		path->skip_locking = 1;
3553		ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3554					0, 0);
3555		if (ret < 0)
3556			break;
3557next:
3558		leaf = path->nodes[0];
3559		if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3560			ret = btrfs_next_leaf(rc->extent_root, path);
3561			if (ret != 0)
3562				break;
3563			leaf = path->nodes[0];
3564		}
3565
3566		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3567		if (key.objectid >= last) {
3568			ret = 1;
3569			break;
3570		}
3571
3572		if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3573		    key.type != BTRFS_METADATA_ITEM_KEY) {
3574			path->slots[0]++;
3575			goto next;
3576		}
3577
3578		if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3579		    key.objectid + key.offset <= rc->search_start) {
3580			path->slots[0]++;
3581			goto next;
3582		}
3583
3584		if (key.type == BTRFS_METADATA_ITEM_KEY &&
3585		    key.objectid + fs_info->nodesize <=
3586		    rc->search_start) {
3587			path->slots[0]++;
3588			goto next;
3589		}
3590
3591		block_found = find_first_extent_bit(&rc->processed_blocks,
3592						    key.objectid, &start, &end,
3593						    EXTENT_DIRTY, NULL);
3594
3595		if (block_found && start <= key.objectid) {
3596			btrfs_release_path(path);
3597			rc->search_start = end + 1;
3598		} else {
3599			if (key.type == BTRFS_EXTENT_ITEM_KEY)
3600				rc->search_start = key.objectid + key.offset;
3601			else
3602				rc->search_start = key.objectid +
3603					fs_info->nodesize;
3604			memcpy(extent_key, &key, sizeof(key));
3605			return 0;
3606		}
3607	}
3608	btrfs_release_path(path);
3609	return ret;
3610}
3611
3612static void set_reloc_control(struct reloc_control *rc)
3613{
3614	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3615
3616	mutex_lock(&fs_info->reloc_mutex);
3617	fs_info->reloc_ctl = rc;
3618	mutex_unlock(&fs_info->reloc_mutex);
3619}
3620
3621static void unset_reloc_control(struct reloc_control *rc)
3622{
3623	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3624
3625	mutex_lock(&fs_info->reloc_mutex);
3626	fs_info->reloc_ctl = NULL;
3627	mutex_unlock(&fs_info->reloc_mutex);
3628}
3629
3630static noinline_for_stack
3631int prepare_to_relocate(struct reloc_control *rc)
3632{
3633	struct btrfs_trans_handle *trans;
3634	int ret;
3635
3636	rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3637					      BTRFS_BLOCK_RSV_TEMP);
3638	if (!rc->block_rsv)
3639		return -ENOMEM;
3640
3641	memset(&rc->cluster, 0, sizeof(rc->cluster));
3642	rc->search_start = rc->block_group->start;
3643	rc->extents_found = 0;
3644	rc->nodes_relocated = 0;
3645	rc->merging_rsv_size = 0;
3646	rc->reserved_bytes = 0;
3647	rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3648			      RELOCATION_RESERVED_NODES;
3649	ret = btrfs_block_rsv_refill(rc->extent_root->fs_info,
3650				     rc->block_rsv, rc->block_rsv->size,
3651				     BTRFS_RESERVE_FLUSH_ALL);
3652	if (ret)
3653		return ret;
3654
3655	rc->create_reloc_tree = true;
3656	set_reloc_control(rc);
3657
3658	trans = btrfs_join_transaction(rc->extent_root);
3659	if (IS_ERR(trans)) {
3660		unset_reloc_control(rc);
3661		/*
3662		 * extent tree is not a ref_cow tree and has no reloc_root to
3663		 * cleanup.  And callers are responsible to free the above
3664		 * block rsv.
3665		 */
3666		return PTR_ERR(trans);
3667	}
3668
3669	ret = btrfs_commit_transaction(trans);
3670	if (ret)
3671		unset_reloc_control(rc);
3672
3673	return ret;
3674}
3675
3676static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3677{
3678	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3679	struct rb_root blocks = RB_ROOT;
3680	struct btrfs_key key;
3681	struct btrfs_trans_handle *trans = NULL;
3682	struct btrfs_path *path;
3683	struct btrfs_extent_item *ei;
3684	u64 flags;
3685	int ret;
3686	int err = 0;
3687	int progress = 0;
3688
3689	path = btrfs_alloc_path();
3690	if (!path)
3691		return -ENOMEM;
3692	path->reada = READA_FORWARD;
3693
3694	ret = prepare_to_relocate(rc);
3695	if (ret) {
3696		err = ret;
3697		goto out_free;
3698	}
3699
3700	while (1) {
3701		rc->reserved_bytes = 0;
3702		ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
3703					     rc->block_rsv->size,
3704					     BTRFS_RESERVE_FLUSH_ALL);
3705		if (ret) {
3706			err = ret;
3707			break;
3708		}
3709		progress++;
3710		trans = btrfs_start_transaction(rc->extent_root, 0);
3711		if (IS_ERR(trans)) {
3712			err = PTR_ERR(trans);
3713			trans = NULL;
3714			break;
3715		}
3716restart:
3717		if (update_backref_cache(trans, &rc->backref_cache)) {
3718			btrfs_end_transaction(trans);
3719			trans = NULL;
3720			continue;
3721		}
3722
3723		ret = find_next_extent(rc, path, &key);
3724		if (ret < 0)
3725			err = ret;
3726		if (ret != 0)
3727			break;
3728
3729		rc->extents_found++;
3730
3731		ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3732				    struct btrfs_extent_item);
3733		flags = btrfs_extent_flags(path->nodes[0], ei);
3734
3735		/*
3736		 * If we are relocating a simple quota owned extent item, we
3737		 * need to note the owner on the reloc data root so that when
3738		 * we allocate the replacement item, we can attribute it to the
3739		 * correct eventual owner (rather than the reloc data root).
3740		 */
3741		if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) {
3742			struct btrfs_root *root = BTRFS_I(rc->data_inode)->root;
3743			u64 owning_root_id = btrfs_get_extent_owner_root(fs_info,
3744								 path->nodes[0],
3745								 path->slots[0]);
3746
3747			root->relocation_src_root = owning_root_id;
3748		}
3749
3750		if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3751			ret = add_tree_block(rc, &key, path, &blocks);
3752		} else if (rc->stage == UPDATE_DATA_PTRS &&
3753			   (flags & BTRFS_EXTENT_FLAG_DATA)) {
3754			ret = add_data_references(rc, &key, path, &blocks);
3755		} else {
3756			btrfs_release_path(path);
3757			ret = 0;
3758		}
3759		if (ret < 0) {
3760			err = ret;
3761			break;
3762		}
3763
3764		if (!RB_EMPTY_ROOT(&blocks)) {
3765			ret = relocate_tree_blocks(trans, rc, &blocks);
3766			if (ret < 0) {
3767				if (ret != -EAGAIN) {
3768					err = ret;
3769					break;
3770				}
3771				rc->extents_found--;
3772				rc->search_start = key.objectid;
3773			}
3774		}
3775
3776		btrfs_end_transaction_throttle(trans);
3777		btrfs_btree_balance_dirty(fs_info);
3778		trans = NULL;
3779
3780		if (rc->stage == MOVE_DATA_EXTENTS &&
3781		    (flags & BTRFS_EXTENT_FLAG_DATA)) {
3782			rc->found_file_extent = true;
3783			ret = relocate_data_extent(rc->data_inode,
3784						   &key, &rc->cluster);
3785			if (ret < 0) {
3786				err = ret;
3787				break;
3788			}
3789		}
3790		if (btrfs_should_cancel_balance(fs_info)) {
3791			err = -ECANCELED;
3792			break;
3793		}
3794	}
3795	if (trans && progress && err == -ENOSPC) {
3796		ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3797		if (ret == 1) {
3798			err = 0;
3799			progress = 0;
3800			goto restart;
3801		}
3802	}
3803
3804	btrfs_release_path(path);
3805	clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3806
3807	if (trans) {
3808		btrfs_end_transaction_throttle(trans);
3809		btrfs_btree_balance_dirty(fs_info);
3810	}
3811
3812	if (!err) {
3813		ret = relocate_file_extent_cluster(rc->data_inode,
3814						   &rc->cluster);
3815		if (ret < 0)
3816			err = ret;
3817	}
3818
3819	rc->create_reloc_tree = false;
3820	set_reloc_control(rc);
3821
3822	btrfs_backref_release_cache(&rc->backref_cache);
3823	btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3824
3825	/*
3826	 * Even in the case when the relocation is cancelled, we should all go
3827	 * through prepare_to_merge() and merge_reloc_roots().
3828	 *
3829	 * For error (including cancelled balance), prepare_to_merge() will
3830	 * mark all reloc trees orphan, then queue them for cleanup in
3831	 * merge_reloc_roots()
3832	 */
3833	err = prepare_to_merge(rc, err);
3834
3835	merge_reloc_roots(rc);
3836
3837	rc->merge_reloc_tree = false;
3838	unset_reloc_control(rc);
3839	btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3840
3841	/* get rid of pinned extents */
3842	trans = btrfs_join_transaction(rc->extent_root);
3843	if (IS_ERR(trans)) {
3844		err = PTR_ERR(trans);
3845		goto out_free;
3846	}
3847	ret = btrfs_commit_transaction(trans);
3848	if (ret && !err)
3849		err = ret;
3850out_free:
3851	ret = clean_dirty_subvols(rc);
3852	if (ret < 0 && !err)
3853		err = ret;
3854	btrfs_free_block_rsv(fs_info, rc->block_rsv);
3855	btrfs_free_path(path);
3856	return err;
3857}
3858
3859static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3860				 struct btrfs_root *root, u64 objectid)
3861{
3862	struct btrfs_path *path;
3863	struct btrfs_inode_item *item;
3864	struct extent_buffer *leaf;
3865	int ret;
3866
3867	path = btrfs_alloc_path();
3868	if (!path)
3869		return -ENOMEM;
3870
3871	ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3872	if (ret)
3873		goto out;
3874
3875	leaf = path->nodes[0];
3876	item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3877	memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3878	btrfs_set_inode_generation(leaf, item, 1);
3879	btrfs_set_inode_size(leaf, item, 0);
3880	btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3881	btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3882					  BTRFS_INODE_PREALLOC);
3883	btrfs_mark_buffer_dirty(trans, leaf);
3884out:
3885	btrfs_free_path(path);
3886	return ret;
3887}
3888
3889static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3890				struct btrfs_root *root, u64 objectid)
3891{
3892	struct btrfs_path *path;
3893	struct btrfs_key key;
3894	int ret = 0;
3895
3896	path = btrfs_alloc_path();
3897	if (!path) {
3898		ret = -ENOMEM;
3899		goto out;
3900	}
3901
3902	key.objectid = objectid;
3903	key.type = BTRFS_INODE_ITEM_KEY;
3904	key.offset = 0;
3905	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3906	if (ret) {
3907		if (ret > 0)
3908			ret = -ENOENT;
3909		goto out;
3910	}
3911	ret = btrfs_del_item(trans, root, path);
3912out:
3913	if (ret)
3914		btrfs_abort_transaction(trans, ret);
3915	btrfs_free_path(path);
3916}
3917
3918/*
3919 * helper to create inode for data relocation.
3920 * the inode is in data relocation tree and its link count is 0
3921 */
3922static noinline_for_stack struct inode *create_reloc_inode(
3923					struct btrfs_fs_info *fs_info,
3924					const struct btrfs_block_group *group)
3925{
3926	struct inode *inode = NULL;
3927	struct btrfs_trans_handle *trans;
3928	struct btrfs_root *root;
3929	u64 objectid;
3930	int err = 0;
3931
3932	root = btrfs_grab_root(fs_info->data_reloc_root);
3933	trans = btrfs_start_transaction(root, 6);
3934	if (IS_ERR(trans)) {
3935		btrfs_put_root(root);
3936		return ERR_CAST(trans);
3937	}
3938
3939	err = btrfs_get_free_objectid(root, &objectid);
3940	if (err)
3941		goto out;
3942
3943	err = __insert_orphan_inode(trans, root, objectid);
3944	if (err)
3945		goto out;
3946
3947	inode = btrfs_iget(fs_info->sb, objectid, root);
3948	if (IS_ERR(inode)) {
3949		delete_orphan_inode(trans, root, objectid);
3950		err = PTR_ERR(inode);
3951		inode = NULL;
3952		goto out;
3953	}
3954	BTRFS_I(inode)->index_cnt = group->start;
3955
3956	err = btrfs_orphan_add(trans, BTRFS_I(inode));
3957out:
3958	btrfs_put_root(root);
3959	btrfs_end_transaction(trans);
3960	btrfs_btree_balance_dirty(fs_info);
3961	if (err) {
3962		iput(inode);
3963		inode = ERR_PTR(err);
3964	}
3965	return inode;
3966}
3967
3968/*
3969 * Mark start of chunk relocation that is cancellable. Check if the cancellation
3970 * has been requested meanwhile and don't start in that case.
3971 *
3972 * Return:
3973 *   0             success
3974 *   -EINPROGRESS  operation is already in progress, that's probably a bug
3975 *   -ECANCELED    cancellation request was set before the operation started
3976 */
3977static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3978{
3979	if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3980		/* This should not happen */
3981		btrfs_err(fs_info, "reloc already running, cannot start");
3982		return -EINPROGRESS;
3983	}
3984
3985	if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3986		btrfs_info(fs_info, "chunk relocation canceled on start");
3987		/*
3988		 * On cancel, clear all requests but let the caller mark
3989		 * the end after cleanup operations.
3990		 */
3991		atomic_set(&fs_info->reloc_cancel_req, 0);
3992		return -ECANCELED;
3993	}
3994	return 0;
3995}
3996
3997/*
3998 * Mark end of chunk relocation that is cancellable and wake any waiters.
3999 */
4000static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
4001{
4002	/* Requested after start, clear bit first so any waiters can continue */
4003	if (atomic_read(&fs_info->reloc_cancel_req) > 0)
4004		btrfs_info(fs_info, "chunk relocation canceled during operation");
4005	clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
4006	atomic_set(&fs_info->reloc_cancel_req, 0);
4007}
4008
4009static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4010{
4011	struct reloc_control *rc;
4012
4013	rc = kzalloc(sizeof(*rc), GFP_NOFS);
4014	if (!rc)
4015		return NULL;
4016
4017	INIT_LIST_HEAD(&rc->reloc_roots);
4018	INIT_LIST_HEAD(&rc->dirty_subvol_roots);
4019	btrfs_backref_init_cache(fs_info, &rc->backref_cache, true);
4020	rc->reloc_root_tree.rb_root = RB_ROOT;
4021	spin_lock_init(&rc->reloc_root_tree.lock);
4022	extent_io_tree_init(fs_info, &rc->processed_blocks, IO_TREE_RELOC_BLOCKS);
4023	return rc;
4024}
4025
4026static void free_reloc_control(struct reloc_control *rc)
4027{
4028	struct mapping_node *node, *tmp;
4029
4030	free_reloc_roots(&rc->reloc_roots);
4031	rbtree_postorder_for_each_entry_safe(node, tmp,
4032			&rc->reloc_root_tree.rb_root, rb_node)
4033		kfree(node);
4034
4035	kfree(rc);
4036}
4037
4038/*
4039 * Print the block group being relocated
4040 */
4041static void describe_relocation(struct btrfs_fs_info *fs_info,
4042				struct btrfs_block_group *block_group)
4043{
4044	char buf[128] = {'\0'};
4045
4046	btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
4047
4048	btrfs_info(fs_info,
4049		   "relocating block group %llu flags %s",
4050		   block_group->start, buf);
4051}
4052
4053static const char *stage_to_string(enum reloc_stage stage)
4054{
4055	if (stage == MOVE_DATA_EXTENTS)
4056		return "move data extents";
4057	if (stage == UPDATE_DATA_PTRS)
4058		return "update data pointers";
4059	return "unknown";
4060}
4061
4062/*
4063 * function to relocate all extents in a block group.
4064 */
4065int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4066{
4067	struct btrfs_block_group *bg;
4068	struct btrfs_root *extent_root = btrfs_extent_root(fs_info, group_start);
4069	struct reloc_control *rc;
4070	struct inode *inode;
4071	struct btrfs_path *path;
4072	int ret;
4073	int rw = 0;
4074	int err = 0;
4075
4076	/*
4077	 * This only gets set if we had a half-deleted snapshot on mount.  We
4078	 * cannot allow relocation to start while we're still trying to clean up
4079	 * these pending deletions.
4080	 */
4081	ret = wait_on_bit(&fs_info->flags, BTRFS_FS_UNFINISHED_DROPS, TASK_INTERRUPTIBLE);
4082	if (ret)
4083		return ret;
4084
4085	/* We may have been woken up by close_ctree, so bail if we're closing. */
4086	if (btrfs_fs_closing(fs_info))
4087		return -EINTR;
4088
4089	bg = btrfs_lookup_block_group(fs_info, group_start);
4090	if (!bg)
4091		return -ENOENT;
4092
4093	/*
4094	 * Relocation of a data block group creates ordered extents.  Without
4095	 * sb_start_write(), we can freeze the filesystem while unfinished
4096	 * ordered extents are left. Such ordered extents can cause a deadlock
4097	 * e.g. when syncfs() is waiting for their completion but they can't
4098	 * finish because they block when joining a transaction, due to the
4099	 * fact that the freeze locks are being held in write mode.
4100	 */
4101	if (bg->flags & BTRFS_BLOCK_GROUP_DATA)
4102		ASSERT(sb_write_started(fs_info->sb));
4103
4104	if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4105		btrfs_put_block_group(bg);
4106		return -ETXTBSY;
4107	}
4108
4109	rc = alloc_reloc_control(fs_info);
4110	if (!rc) {
4111		btrfs_put_block_group(bg);
4112		return -ENOMEM;
4113	}
4114
4115	ret = reloc_chunk_start(fs_info);
4116	if (ret < 0) {
4117		err = ret;
4118		goto out_put_bg;
4119	}
4120
4121	rc->extent_root = extent_root;
4122	rc->block_group = bg;
4123
4124	ret = btrfs_inc_block_group_ro(rc->block_group, true);
4125	if (ret) {
4126		err = ret;
4127		goto out;
4128	}
4129	rw = 1;
4130
4131	path = btrfs_alloc_path();
4132	if (!path) {
4133		err = -ENOMEM;
4134		goto out;
4135	}
4136
4137	inode = lookup_free_space_inode(rc->block_group, path);
4138	btrfs_free_path(path);
4139
4140	if (!IS_ERR(inode))
4141		ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4142	else
4143		ret = PTR_ERR(inode);
4144
4145	if (ret && ret != -ENOENT) {
4146		err = ret;
4147		goto out;
4148	}
4149
4150	rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4151	if (IS_ERR(rc->data_inode)) {
4152		err = PTR_ERR(rc->data_inode);
4153		rc->data_inode = NULL;
4154		goto out;
4155	}
4156
4157	describe_relocation(fs_info, rc->block_group);
4158
4159	btrfs_wait_block_group_reservations(rc->block_group);
4160	btrfs_wait_nocow_writers(rc->block_group);
4161	btrfs_wait_ordered_roots(fs_info, U64_MAX,
4162				 rc->block_group->start,
4163				 rc->block_group->length);
4164
4165	ret = btrfs_zone_finish(rc->block_group);
4166	WARN_ON(ret && ret != -EAGAIN);
4167
4168	while (1) {
4169		enum reloc_stage finishes_stage;
4170
4171		mutex_lock(&fs_info->cleaner_mutex);
4172		ret = relocate_block_group(rc);
4173		mutex_unlock(&fs_info->cleaner_mutex);
4174		if (ret < 0)
4175			err = ret;
4176
4177		finishes_stage = rc->stage;
4178		/*
4179		 * We may have gotten ENOSPC after we already dirtied some
4180		 * extents.  If writeout happens while we're relocating a
4181		 * different block group we could end up hitting the
4182		 * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4183		 * btrfs_reloc_cow_block.  Make sure we write everything out
4184		 * properly so we don't trip over this problem, and then break
4185		 * out of the loop if we hit an error.
4186		 */
4187		if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4188			ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4189						       (u64)-1);
4190			if (ret)
4191				err = ret;
4192			invalidate_mapping_pages(rc->data_inode->i_mapping,
4193						 0, -1);
4194			rc->stage = UPDATE_DATA_PTRS;
4195		}
4196
4197		if (err < 0)
4198			goto out;
4199
4200		if (rc->extents_found == 0)
4201			break;
4202
4203		btrfs_info(fs_info, "found %llu extents, stage: %s",
4204			   rc->extents_found, stage_to_string(finishes_stage));
4205	}
4206
4207	WARN_ON(rc->block_group->pinned > 0);
4208	WARN_ON(rc->block_group->reserved > 0);
4209	WARN_ON(rc->block_group->used > 0);
4210out:
4211	if (err && rw)
4212		btrfs_dec_block_group_ro(rc->block_group);
4213	iput(rc->data_inode);
4214out_put_bg:
4215	btrfs_put_block_group(bg);
4216	reloc_chunk_end(fs_info);
4217	free_reloc_control(rc);
4218	return err;
4219}
4220
4221static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4222{
4223	struct btrfs_fs_info *fs_info = root->fs_info;
4224	struct btrfs_trans_handle *trans;
4225	int ret, err;
4226
4227	trans = btrfs_start_transaction(fs_info->tree_root, 0);
4228	if (IS_ERR(trans))
4229		return PTR_ERR(trans);
4230
4231	memset(&root->root_item.drop_progress, 0,
4232		sizeof(root->root_item.drop_progress));
4233	btrfs_set_root_drop_level(&root->root_item, 0);
4234	btrfs_set_root_refs(&root->root_item, 0);
4235	ret = btrfs_update_root(trans, fs_info->tree_root,
4236				&root->root_key, &root->root_item);
4237
4238	err = btrfs_end_transaction(trans);
4239	if (err)
4240		return err;
4241	return ret;
4242}
4243
4244/*
4245 * recover relocation interrupted by system crash.
4246 *
4247 * this function resumes merging reloc trees with corresponding fs trees.
4248 * this is important for keeping the sharing of tree blocks
4249 */
4250int btrfs_recover_relocation(struct btrfs_fs_info *fs_info)
4251{
4252	LIST_HEAD(reloc_roots);
4253	struct btrfs_key key;
4254	struct btrfs_root *fs_root;
4255	struct btrfs_root *reloc_root;
4256	struct btrfs_path *path;
4257	struct extent_buffer *leaf;
4258	struct reloc_control *rc = NULL;
4259	struct btrfs_trans_handle *trans;
4260	int ret;
4261	int err = 0;
4262
4263	path = btrfs_alloc_path();
4264	if (!path)
4265		return -ENOMEM;
4266	path->reada = READA_BACK;
4267
4268	key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4269	key.type = BTRFS_ROOT_ITEM_KEY;
4270	key.offset = (u64)-1;
4271
4272	while (1) {
4273		ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4274					path, 0, 0);
4275		if (ret < 0) {
4276			err = ret;
4277			goto out;
4278		}
4279		if (ret > 0) {
4280			if (path->slots[0] == 0)
4281				break;
4282			path->slots[0]--;
4283		}
4284		leaf = path->nodes[0];
4285		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4286		btrfs_release_path(path);
4287
4288		if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4289		    key.type != BTRFS_ROOT_ITEM_KEY)
4290			break;
4291
4292		reloc_root = btrfs_read_tree_root(fs_info->tree_root, &key);
4293		if (IS_ERR(reloc_root)) {
4294			err = PTR_ERR(reloc_root);
4295			goto out;
4296		}
4297
4298		set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4299		list_add(&reloc_root->root_list, &reloc_roots);
4300
4301		if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4302			fs_root = btrfs_get_fs_root(fs_info,
4303					reloc_root->root_key.offset, false);
4304			if (IS_ERR(fs_root)) {
4305				ret = PTR_ERR(fs_root);
4306				if (ret != -ENOENT) {
4307					err = ret;
4308					goto out;
4309				}
4310				ret = mark_garbage_root(reloc_root);
4311				if (ret < 0) {
4312					err = ret;
4313					goto out;
4314				}
4315			} else {
4316				btrfs_put_root(fs_root);
4317			}
4318		}
4319
4320		if (key.offset == 0)
4321			break;
4322
4323		key.offset--;
4324	}
4325	btrfs_release_path(path);
4326
4327	if (list_empty(&reloc_roots))
4328		goto out;
4329
4330	rc = alloc_reloc_control(fs_info);
4331	if (!rc) {
4332		err = -ENOMEM;
4333		goto out;
4334	}
4335
4336	ret = reloc_chunk_start(fs_info);
4337	if (ret < 0) {
4338		err = ret;
4339		goto out_end;
4340	}
4341
4342	rc->extent_root = btrfs_extent_root(fs_info, 0);
4343
4344	set_reloc_control(rc);
4345
4346	trans = btrfs_join_transaction(rc->extent_root);
4347	if (IS_ERR(trans)) {
4348		err = PTR_ERR(trans);
4349		goto out_unset;
4350	}
4351
4352	rc->merge_reloc_tree = true;
4353
4354	while (!list_empty(&reloc_roots)) {
4355		reloc_root = list_entry(reloc_roots.next,
4356					struct btrfs_root, root_list);
4357		list_del(&reloc_root->root_list);
4358
4359		if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4360			list_add_tail(&reloc_root->root_list,
4361				      &rc->reloc_roots);
4362			continue;
4363		}
4364
4365		fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4366					    false);
4367		if (IS_ERR(fs_root)) {
4368			err = PTR_ERR(fs_root);
4369			list_add_tail(&reloc_root->root_list, &reloc_roots);
4370			btrfs_end_transaction(trans);
4371			goto out_unset;
4372		}
4373
4374		err = __add_reloc_root(reloc_root);
4375		ASSERT(err != -EEXIST);
4376		if (err) {
4377			list_add_tail(&reloc_root->root_list, &reloc_roots);
4378			btrfs_put_root(fs_root);
4379			btrfs_end_transaction(trans);
4380			goto out_unset;
4381		}
4382		fs_root->reloc_root = btrfs_grab_root(reloc_root);
4383		btrfs_put_root(fs_root);
4384	}
4385
4386	err = btrfs_commit_transaction(trans);
4387	if (err)
4388		goto out_unset;
4389
4390	merge_reloc_roots(rc);
4391
4392	unset_reloc_control(rc);
4393
4394	trans = btrfs_join_transaction(rc->extent_root);
4395	if (IS_ERR(trans)) {
4396		err = PTR_ERR(trans);
4397		goto out_clean;
4398	}
4399	err = btrfs_commit_transaction(trans);
4400out_clean:
4401	ret = clean_dirty_subvols(rc);
4402	if (ret < 0 && !err)
4403		err = ret;
4404out_unset:
4405	unset_reloc_control(rc);
4406out_end:
4407	reloc_chunk_end(fs_info);
4408	free_reloc_control(rc);
4409out:
4410	free_reloc_roots(&reloc_roots);
4411
4412	btrfs_free_path(path);
4413
4414	if (err == 0) {
4415		/* cleanup orphan inode in data relocation tree */
4416		fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4417		ASSERT(fs_root);
4418		err = btrfs_orphan_cleanup(fs_root);
4419		btrfs_put_root(fs_root);
4420	}
4421	return err;
4422}
4423
4424/*
4425 * helper to add ordered checksum for data relocation.
4426 *
4427 * cloning checksum properly handles the nodatasum extents.
4428 * it also saves CPU time to re-calculate the checksum.
4429 */
4430int btrfs_reloc_clone_csums(struct btrfs_ordered_extent *ordered)
4431{
4432	struct btrfs_inode *inode = BTRFS_I(ordered->inode);
4433	struct btrfs_fs_info *fs_info = inode->root->fs_info;
4434	u64 disk_bytenr = ordered->file_offset + inode->index_cnt;
4435	struct btrfs_root *csum_root = btrfs_csum_root(fs_info, disk_bytenr);
4436	LIST_HEAD(list);
4437	int ret;
4438
4439	ret = btrfs_lookup_csums_list(csum_root, disk_bytenr,
4440				      disk_bytenr + ordered->num_bytes - 1,
4441				      &list, 0, false);
4442	if (ret)
4443		return ret;
4444
4445	while (!list_empty(&list)) {
4446		struct btrfs_ordered_sum *sums =
4447			list_entry(list.next, struct btrfs_ordered_sum, list);
4448
4449		list_del_init(&sums->list);
4450
4451		/*
4452		 * We need to offset the new_bytenr based on where the csum is.
4453		 * We need to do this because we will read in entire prealloc
4454		 * extents but we may have written to say the middle of the
4455		 * prealloc extent, so we need to make sure the csum goes with
4456		 * the right disk offset.
4457		 *
4458		 * We can do this because the data reloc inode refers strictly
4459		 * to the on disk bytes, so we don't have to worry about
4460		 * disk_len vs real len like with real inodes since it's all
4461		 * disk length.
4462		 */
4463		sums->logical = ordered->disk_bytenr + sums->logical - disk_bytenr;
4464		btrfs_add_ordered_sum(ordered, sums);
4465	}
4466
4467	return 0;
4468}
4469
4470int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4471			  struct btrfs_root *root,
4472			  const struct extent_buffer *buf,
4473			  struct extent_buffer *cow)
4474{
4475	struct btrfs_fs_info *fs_info = root->fs_info;
4476	struct reloc_control *rc;
4477	struct btrfs_backref_node *node;
4478	int first_cow = 0;
4479	int level;
4480	int ret = 0;
4481
4482	rc = fs_info->reloc_ctl;
4483	if (!rc)
4484		return 0;
4485
4486	BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4487
4488	level = btrfs_header_level(buf);
4489	if (btrfs_header_generation(buf) <=
4490	    btrfs_root_last_snapshot(&root->root_item))
4491		first_cow = 1;
4492
4493	if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4494	    rc->create_reloc_tree) {
4495		WARN_ON(!first_cow && level == 0);
4496
4497		node = rc->backref_cache.path[level];
4498		BUG_ON(node->bytenr != buf->start &&
4499		       node->new_bytenr != buf->start);
4500
4501		btrfs_backref_drop_node_buffer(node);
4502		atomic_inc(&cow->refs);
4503		node->eb = cow;
4504		node->new_bytenr = cow->start;
4505
4506		if (!node->pending) {
4507			list_move_tail(&node->list,
4508				       &rc->backref_cache.pending[level]);
4509			node->pending = 1;
4510		}
4511
4512		if (first_cow)
4513			mark_block_processed(rc, node);
4514
4515		if (first_cow && level > 0)
4516			rc->nodes_relocated += buf->len;
4517	}
4518
4519	if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4520		ret = replace_file_extents(trans, rc, root, cow);
4521	return ret;
4522}
4523
4524/*
4525 * called before creating snapshot. it calculates metadata reservation
4526 * required for relocating tree blocks in the snapshot
4527 */
4528void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4529			      u64 *bytes_to_reserve)
4530{
4531	struct btrfs_root *root = pending->root;
4532	struct reloc_control *rc = root->fs_info->reloc_ctl;
4533
4534	if (!rc || !have_reloc_root(root))
4535		return;
4536
4537	if (!rc->merge_reloc_tree)
4538		return;
4539
4540	root = root->reloc_root;
4541	BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4542	/*
4543	 * relocation is in the stage of merging trees. the space
4544	 * used by merging a reloc tree is twice the size of
4545	 * relocated tree nodes in the worst case. half for cowing
4546	 * the reloc tree, half for cowing the fs tree. the space
4547	 * used by cowing the reloc tree will be freed after the
4548	 * tree is dropped. if we create snapshot, cowing the fs
4549	 * tree may use more space than it frees. so we need
4550	 * reserve extra space.
4551	 */
4552	*bytes_to_reserve += rc->nodes_relocated;
4553}
4554
4555/*
4556 * called after snapshot is created. migrate block reservation
4557 * and create reloc root for the newly created snapshot
4558 *
4559 * This is similar to btrfs_init_reloc_root(), we come out of here with two
4560 * references held on the reloc_root, one for root->reloc_root and one for
4561 * rc->reloc_roots.
4562 */
4563int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4564			       struct btrfs_pending_snapshot *pending)
4565{
4566	struct btrfs_root *root = pending->root;
4567	struct btrfs_root *reloc_root;
4568	struct btrfs_root *new_root;
4569	struct reloc_control *rc = root->fs_info->reloc_ctl;
4570	int ret;
4571
4572	if (!rc || !have_reloc_root(root))
4573		return 0;
4574
4575	rc = root->fs_info->reloc_ctl;
4576	rc->merging_rsv_size += rc->nodes_relocated;
4577
4578	if (rc->merge_reloc_tree) {
4579		ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4580					      rc->block_rsv,
4581					      rc->nodes_relocated, true);
4582		if (ret)
4583			return ret;
4584	}
4585
4586	new_root = pending->snap;
4587	reloc_root = create_reloc_root(trans, root->reloc_root,
4588				       new_root->root_key.objectid);
4589	if (IS_ERR(reloc_root))
4590		return PTR_ERR(reloc_root);
4591
4592	ret = __add_reloc_root(reloc_root);
4593	ASSERT(ret != -EEXIST);
4594	if (ret) {
4595		/* Pairs with create_reloc_root */
4596		btrfs_put_root(reloc_root);
4597		return ret;
4598	}
4599	new_root->reloc_root = btrfs_grab_root(reloc_root);
4600
4601	if (rc->create_reloc_tree)
4602		ret = clone_backref_node(trans, rc, root, reloc_root);
4603	return ret;
4604}
4605
4606/*
4607 * Get the current bytenr for the block group which is being relocated.
4608 *
4609 * Return U64_MAX if no running relocation.
4610 */
4611u64 btrfs_get_reloc_bg_bytenr(const struct btrfs_fs_info *fs_info)
4612{
4613	u64 logical = U64_MAX;
4614
4615	lockdep_assert_held(&fs_info->reloc_mutex);
4616
4617	if (fs_info->reloc_ctl && fs_info->reloc_ctl->block_group)
4618		logical = fs_info->reloc_ctl->block_group->start;
4619	return logical;
4620}
v6.9.4
   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#include "space-info.h"
  31#include "fs.h"
  32#include "accessors.h"
  33#include "extent-tree.h"
  34#include "root-tree.h"
  35#include "file-item.h"
  36#include "relocation.h"
  37#include "super.h"
  38#include "tree-checker.h"
  39
  40/*
  41 * Relocation overview
  42 *
  43 * [What does relocation do]
  44 *
  45 * The objective of relocation is to relocate all extents of the target block
  46 * group to other block groups.
  47 * This is utilized by resize (shrink only), profile converting, compacting
  48 * space, or balance routine to spread chunks over devices.
  49 *
  50 * 		Before		|		After
  51 * ------------------------------------------------------------------
  52 *  BG A: 10 data extents	| BG A: deleted
  53 *  BG B:  2 data extents	| BG B: 10 data extents (2 old + 8 relocated)
  54 *  BG C:  1 extents		| BG C:  3 data extents (1 old + 2 relocated)
  55 *
  56 * [How does relocation work]
  57 *
  58 * 1.   Mark the target block group read-only
  59 *      New extents won't be allocated from the target block group.
  60 *
  61 * 2.1  Record each extent in the target block group
  62 *      To build a proper map of extents to be relocated.
  63 *
  64 * 2.2  Build data reloc tree and reloc trees
  65 *      Data reloc tree will contain an inode, recording all newly relocated
  66 *      data extents.
  67 *      There will be only one data reloc tree for one data block group.
  68 *
  69 *      Reloc tree will be a special snapshot of its source tree, containing
  70 *      relocated tree blocks.
  71 *      Each tree referring to a tree block in target block group will get its
  72 *      reloc tree built.
  73 *
  74 * 2.3  Swap source tree with its corresponding reloc tree
  75 *      Each involved tree only refers to new extents after swap.
  76 *
  77 * 3.   Cleanup reloc trees and data reloc tree.
  78 *      As old extents in the target block group are still referenced by reloc
  79 *      trees, we need to clean them up before really freeing the target block
  80 *      group.
  81 *
  82 * The main complexity is in steps 2.2 and 2.3.
  83 *
  84 * The entry point of relocation is relocate_block_group() function.
  85 */
  86
  87#define RELOCATION_RESERVED_NODES	256
  88/*
  89 * map address of tree root to tree
  90 */
  91struct mapping_node {
  92	struct {
  93		struct rb_node rb_node;
  94		u64 bytenr;
  95	}; /* Use rb_simle_node for search/insert */
  96	void *data;
  97};
  98
  99struct mapping_tree {
 100	struct rb_root rb_root;
 101	spinlock_t lock;
 102};
 103
 104/*
 105 * present a tree block to process
 106 */
 107struct tree_block {
 108	struct {
 109		struct rb_node rb_node;
 110		u64 bytenr;
 111	}; /* Use rb_simple_node for search/insert */
 112	u64 owner;
 113	struct btrfs_key key;
 114	u8 level;
 115	bool key_ready;
 116};
 117
 118#define MAX_EXTENTS 128
 119
 120struct file_extent_cluster {
 121	u64 start;
 122	u64 end;
 123	u64 boundary[MAX_EXTENTS];
 124	unsigned int nr;
 125	u64 owning_root;
 126};
 127
 128/* Stages of data relocation. */
 129enum reloc_stage {
 130	MOVE_DATA_EXTENTS,
 131	UPDATE_DATA_PTRS
 132};
 133
 134struct reloc_control {
 135	/* block group to relocate */
 136	struct btrfs_block_group *block_group;
 137	/* extent tree */
 138	struct btrfs_root *extent_root;
 139	/* inode for moving data */
 140	struct inode *data_inode;
 141
 142	struct btrfs_block_rsv *block_rsv;
 143
 144	struct btrfs_backref_cache backref_cache;
 145
 146	struct file_extent_cluster cluster;
 147	/* tree blocks have been processed */
 148	struct extent_io_tree processed_blocks;
 149	/* map start of tree root to corresponding reloc tree */
 150	struct mapping_tree reloc_root_tree;
 151	/* list of reloc trees */
 152	struct list_head reloc_roots;
 153	/* list of subvolume trees that get relocated */
 154	struct list_head dirty_subvol_roots;
 155	/* size of metadata reservation for merging reloc trees */
 156	u64 merging_rsv_size;
 157	/* size of relocated tree nodes */
 158	u64 nodes_relocated;
 159	/* reserved size for block group relocation*/
 160	u64 reserved_bytes;
 161
 162	u64 search_start;
 163	u64 extents_found;
 164
 165	enum reloc_stage stage;
 166	bool create_reloc_tree;
 167	bool merge_reloc_tree;
 168	bool found_file_extent;
 169};
 170
 171static void mark_block_processed(struct reloc_control *rc,
 172				 struct btrfs_backref_node *node)
 173{
 174	u32 blocksize;
 175
 176	if (node->level == 0 ||
 177	    in_range(node->bytenr, rc->block_group->start,
 178		     rc->block_group->length)) {
 179		blocksize = rc->extent_root->fs_info->nodesize;
 180		set_extent_bit(&rc->processed_blocks, node->bytenr,
 181			       node->bytenr + blocksize - 1, EXTENT_DIRTY, NULL);
 182	}
 183	node->processed = 1;
 184}
 185
 186/*
 187 * walk up backref nodes until reach node presents tree root
 188 */
 189static struct btrfs_backref_node *walk_up_backref(
 190		struct btrfs_backref_node *node,
 191		struct btrfs_backref_edge *edges[], int *index)
 192{
 193	struct btrfs_backref_edge *edge;
 194	int idx = *index;
 195
 196	while (!list_empty(&node->upper)) {
 197		edge = list_entry(node->upper.next,
 198				  struct btrfs_backref_edge, list[LOWER]);
 199		edges[idx++] = edge;
 200		node = edge->node[UPPER];
 201	}
 202	BUG_ON(node->detached);
 203	*index = idx;
 204	return node;
 205}
 206
 207/*
 208 * walk down backref nodes to find start of next reference path
 209 */
 210static struct btrfs_backref_node *walk_down_backref(
 211		struct btrfs_backref_edge *edges[], int *index)
 212{
 213	struct btrfs_backref_edge *edge;
 214	struct btrfs_backref_node *lower;
 215	int idx = *index;
 216
 217	while (idx > 0) {
 218		edge = edges[idx - 1];
 219		lower = edge->node[LOWER];
 220		if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 221			idx--;
 222			continue;
 223		}
 224		edge = list_entry(edge->list[LOWER].next,
 225				  struct btrfs_backref_edge, list[LOWER]);
 226		edges[idx - 1] = edge;
 227		*index = idx;
 228		return edge->node[UPPER];
 229	}
 230	*index = 0;
 231	return NULL;
 232}
 233
 234static void update_backref_node(struct btrfs_backref_cache *cache,
 235				struct btrfs_backref_node *node, u64 bytenr)
 236{
 237	struct rb_node *rb_node;
 238	rb_erase(&node->rb_node, &cache->rb_root);
 239	node->bytenr = bytenr;
 240	rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 241	if (rb_node)
 242		btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
 243}
 244
 245/*
 246 * update backref cache after a transaction commit
 247 */
 248static int update_backref_cache(struct btrfs_trans_handle *trans,
 249				struct btrfs_backref_cache *cache)
 250{
 251	struct btrfs_backref_node *node;
 252	int level = 0;
 253
 254	if (cache->last_trans == 0) {
 255		cache->last_trans = trans->transid;
 256		return 0;
 257	}
 258
 259	if (cache->last_trans == trans->transid)
 260		return 0;
 261
 262	/*
 263	 * detached nodes are used to avoid unnecessary backref
 264	 * lookup. transaction commit changes the extent tree.
 265	 * so the detached nodes are no longer useful.
 266	 */
 267	while (!list_empty(&cache->detached)) {
 268		node = list_entry(cache->detached.next,
 269				  struct btrfs_backref_node, list);
 270		btrfs_backref_cleanup_node(cache, node);
 271	}
 272
 273	while (!list_empty(&cache->changed)) {
 274		node = list_entry(cache->changed.next,
 275				  struct btrfs_backref_node, list);
 276		list_del_init(&node->list);
 277		BUG_ON(node->pending);
 278		update_backref_node(cache, node, node->new_bytenr);
 279	}
 280
 281	/*
 282	 * some nodes can be left in the pending list if there were
 283	 * errors during processing the pending nodes.
 284	 */
 285	for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 286		list_for_each_entry(node, &cache->pending[level], list) {
 287			BUG_ON(!node->pending);
 288			if (node->bytenr == node->new_bytenr)
 289				continue;
 290			update_backref_node(cache, node, node->new_bytenr);
 291		}
 292	}
 293
 294	cache->last_trans = 0;
 295	return 1;
 296}
 297
 298static bool reloc_root_is_dead(const struct btrfs_root *root)
 299{
 300	/*
 301	 * Pair with set_bit/clear_bit in clean_dirty_subvols and
 302	 * btrfs_update_reloc_root. We need to see the updated bit before
 303	 * trying to access reloc_root
 304	 */
 305	smp_rmb();
 306	if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
 307		return true;
 308	return false;
 309}
 310
 311/*
 312 * Check if this subvolume tree has valid reloc tree.
 313 *
 314 * Reloc tree after swap is considered dead, thus not considered as valid.
 315 * This is enough for most callers, as they don't distinguish dead reloc root
 316 * from no reloc root.  But btrfs_should_ignore_reloc_root() below is a
 317 * special case.
 318 */
 319static bool have_reloc_root(const struct btrfs_root *root)
 320{
 321	if (reloc_root_is_dead(root))
 322		return false;
 323	if (!root->reloc_root)
 324		return false;
 325	return true;
 326}
 327
 328bool btrfs_should_ignore_reloc_root(const struct btrfs_root *root)
 329{
 330	struct btrfs_root *reloc_root;
 331
 332	if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
 333		return false;
 334
 335	/* This root has been merged with its reloc tree, we can ignore it */
 336	if (reloc_root_is_dead(root))
 337		return true;
 338
 339	reloc_root = root->reloc_root;
 340	if (!reloc_root)
 341		return false;
 342
 343	if (btrfs_header_generation(reloc_root->commit_root) ==
 344	    root->fs_info->running_transaction->transid)
 345		return false;
 346	/*
 347	 * If there is reloc tree and it was created in previous transaction
 348	 * backref lookup can find the reloc tree, so backref node for the fs
 349	 * tree root is useless for relocation.
 350	 */
 351	return true;
 352}
 353
 354/*
 355 * find reloc tree by address of tree root
 356 */
 357struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
 358{
 359	struct reloc_control *rc = fs_info->reloc_ctl;
 360	struct rb_node *rb_node;
 361	struct mapping_node *node;
 362	struct btrfs_root *root = NULL;
 363
 364	ASSERT(rc);
 365	spin_lock(&rc->reloc_root_tree.lock);
 366	rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
 367	if (rb_node) {
 368		node = rb_entry(rb_node, struct mapping_node, rb_node);
 369		root = node->data;
 370	}
 371	spin_unlock(&rc->reloc_root_tree.lock);
 372	return btrfs_grab_root(root);
 373}
 374
 375/*
 376 * For useless nodes, do two major clean ups:
 377 *
 378 * - Cleanup the children edges and nodes
 379 *   If child node is also orphan (no parent) during cleanup, then the child
 380 *   node will also be cleaned up.
 381 *
 382 * - Freeing up leaves (level 0), keeps nodes detached
 383 *   For nodes, the node is still cached as "detached"
 384 *
 385 * Return false if @node is not in the @useless_nodes list.
 386 * Return true if @node is in the @useless_nodes list.
 387 */
 388static bool handle_useless_nodes(struct reloc_control *rc,
 389				 struct btrfs_backref_node *node)
 390{
 391	struct btrfs_backref_cache *cache = &rc->backref_cache;
 392	struct list_head *useless_node = &cache->useless_node;
 393	bool ret = false;
 394
 395	while (!list_empty(useless_node)) {
 396		struct btrfs_backref_node *cur;
 397
 398		cur = list_first_entry(useless_node, struct btrfs_backref_node,
 399				 list);
 400		list_del_init(&cur->list);
 401
 402		/* Only tree root nodes can be added to @useless_nodes */
 403		ASSERT(list_empty(&cur->upper));
 404
 405		if (cur == node)
 406			ret = true;
 407
 408		/* The node is the lowest node */
 409		if (cur->lowest) {
 410			list_del_init(&cur->lower);
 411			cur->lowest = 0;
 412		}
 413
 414		/* Cleanup the lower edges */
 415		while (!list_empty(&cur->lower)) {
 416			struct btrfs_backref_edge *edge;
 417			struct btrfs_backref_node *lower;
 418
 419			edge = list_entry(cur->lower.next,
 420					struct btrfs_backref_edge, list[UPPER]);
 421			list_del(&edge->list[UPPER]);
 422			list_del(&edge->list[LOWER]);
 423			lower = edge->node[LOWER];
 424			btrfs_backref_free_edge(cache, edge);
 425
 426			/* Child node is also orphan, queue for cleanup */
 427			if (list_empty(&lower->upper))
 428				list_add(&lower->list, useless_node);
 429		}
 430		/* Mark this block processed for relocation */
 431		mark_block_processed(rc, cur);
 432
 433		/*
 434		 * Backref nodes for tree leaves are deleted from the cache.
 435		 * Backref nodes for upper level tree blocks are left in the
 436		 * cache to avoid unnecessary backref lookup.
 437		 */
 438		if (cur->level > 0) {
 439			list_add(&cur->list, &cache->detached);
 440			cur->detached = 1;
 441		} else {
 442			rb_erase(&cur->rb_node, &cache->rb_root);
 443			btrfs_backref_free_node(cache, cur);
 444		}
 445	}
 446	return ret;
 447}
 448
 449/*
 450 * Build backref tree for a given tree block. Root of the backref tree
 451 * corresponds the tree block, leaves of the backref tree correspond roots of
 452 * b-trees that reference the tree block.
 453 *
 454 * The basic idea of this function is check backrefs of a given block to find
 455 * upper level blocks that reference the block, and then check backrefs of
 456 * these upper level blocks recursively. The recursion stops when tree root is
 457 * reached or backrefs for the block is cached.
 458 *
 459 * NOTE: if we find that backrefs for a block are cached, we know backrefs for
 460 * all upper level blocks that directly/indirectly reference the block are also
 461 * cached.
 462 */
 463static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
 464			struct btrfs_trans_handle *trans,
 465			struct reloc_control *rc, struct btrfs_key *node_key,
 466			int level, u64 bytenr)
 467{
 468	struct btrfs_backref_iter *iter;
 469	struct btrfs_backref_cache *cache = &rc->backref_cache;
 470	/* For searching parent of TREE_BLOCK_REF */
 471	struct btrfs_path *path;
 472	struct btrfs_backref_node *cur;
 473	struct btrfs_backref_node *node = NULL;
 474	struct btrfs_backref_edge *edge;
 475	int ret;
 476	int err = 0;
 477
 478	iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info);
 479	if (!iter)
 480		return ERR_PTR(-ENOMEM);
 481	path = btrfs_alloc_path();
 482	if (!path) {
 483		err = -ENOMEM;
 484		goto out;
 485	}
 486
 487	node = btrfs_backref_alloc_node(cache, bytenr, level);
 488	if (!node) {
 489		err = -ENOMEM;
 490		goto out;
 491	}
 492
 493	node->lowest = 1;
 494	cur = node;
 495
 496	/* Breadth-first search to build backref cache */
 497	do {
 498		ret = btrfs_backref_add_tree_node(trans, cache, path, iter,
 499						  node_key, cur);
 500		if (ret < 0) {
 501			err = ret;
 502			goto out;
 503		}
 504		edge = list_first_entry_or_null(&cache->pending_edge,
 505				struct btrfs_backref_edge, list[UPPER]);
 506		/*
 507		 * The pending list isn't empty, take the first block to
 508		 * process
 509		 */
 510		if (edge) {
 511			list_del_init(&edge->list[UPPER]);
 512			cur = edge->node[UPPER];
 513		}
 514	} while (edge);
 515
 516	/* Finish the upper linkage of newly added edges/nodes */
 517	ret = btrfs_backref_finish_upper_links(cache, node);
 518	if (ret < 0) {
 519		err = ret;
 520		goto out;
 521	}
 522
 523	if (handle_useless_nodes(rc, node))
 524		node = NULL;
 525out:
 526	btrfs_free_path(iter->path);
 527	kfree(iter);
 528	btrfs_free_path(path);
 529	if (err) {
 530		btrfs_backref_error_cleanup(cache, node);
 531		return ERR_PTR(err);
 532	}
 533	ASSERT(!node || !node->detached);
 534	ASSERT(list_empty(&cache->useless_node) &&
 535	       list_empty(&cache->pending_edge));
 536	return node;
 537}
 538
 539/*
 540 * helper to add backref node for the newly created snapshot.
 541 * the backref node is created by cloning backref node that
 542 * corresponds to root of source tree
 543 */
 544static int clone_backref_node(struct btrfs_trans_handle *trans,
 545			      struct reloc_control *rc,
 546			      const struct btrfs_root *src,
 547			      struct btrfs_root *dest)
 548{
 549	struct btrfs_root *reloc_root = src->reloc_root;
 550	struct btrfs_backref_cache *cache = &rc->backref_cache;
 551	struct btrfs_backref_node *node = NULL;
 552	struct btrfs_backref_node *new_node;
 553	struct btrfs_backref_edge *edge;
 554	struct btrfs_backref_edge *new_edge;
 555	struct rb_node *rb_node;
 556
 557	if (cache->last_trans > 0)
 558		update_backref_cache(trans, cache);
 559
 560	rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
 561	if (rb_node) {
 562		node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
 563		if (node->detached)
 564			node = NULL;
 565		else
 566			BUG_ON(node->new_bytenr != reloc_root->node->start);
 567	}
 568
 569	if (!node) {
 570		rb_node = rb_simple_search(&cache->rb_root,
 571					   reloc_root->commit_root->start);
 572		if (rb_node) {
 573			node = rb_entry(rb_node, struct btrfs_backref_node,
 574					rb_node);
 575			BUG_ON(node->detached);
 576		}
 577	}
 578
 579	if (!node)
 580		return 0;
 581
 582	new_node = btrfs_backref_alloc_node(cache, dest->node->start,
 583					    node->level);
 584	if (!new_node)
 585		return -ENOMEM;
 586
 587	new_node->lowest = node->lowest;
 588	new_node->checked = 1;
 589	new_node->root = btrfs_grab_root(dest);
 590	ASSERT(new_node->root);
 591
 592	if (!node->lowest) {
 593		list_for_each_entry(edge, &node->lower, list[UPPER]) {
 594			new_edge = btrfs_backref_alloc_edge(cache);
 595			if (!new_edge)
 596				goto fail;
 597
 598			btrfs_backref_link_edge(new_edge, edge->node[LOWER],
 599						new_node, LINK_UPPER);
 600		}
 601	} else {
 602		list_add_tail(&new_node->lower, &cache->leaves);
 603	}
 604
 605	rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
 606				   &new_node->rb_node);
 607	if (rb_node)
 608		btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
 609
 610	if (!new_node->lowest) {
 611		list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
 612			list_add_tail(&new_edge->list[LOWER],
 613				      &new_edge->node[LOWER]->upper);
 614		}
 615	}
 616	return 0;
 617fail:
 618	while (!list_empty(&new_node->lower)) {
 619		new_edge = list_entry(new_node->lower.next,
 620				      struct btrfs_backref_edge, list[UPPER]);
 621		list_del(&new_edge->list[UPPER]);
 622		btrfs_backref_free_edge(cache, new_edge);
 623	}
 624	btrfs_backref_free_node(cache, new_node);
 625	return -ENOMEM;
 626}
 627
 628/*
 629 * helper to add 'address of tree root -> reloc tree' mapping
 630 */
 631static int __add_reloc_root(struct btrfs_root *root)
 632{
 633	struct btrfs_fs_info *fs_info = root->fs_info;
 634	struct rb_node *rb_node;
 635	struct mapping_node *node;
 636	struct reloc_control *rc = fs_info->reloc_ctl;
 637
 638	node = kmalloc(sizeof(*node), GFP_NOFS);
 639	if (!node)
 640		return -ENOMEM;
 641
 642	node->bytenr = root->commit_root->start;
 643	node->data = root;
 644
 645	spin_lock(&rc->reloc_root_tree.lock);
 646	rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
 647				   node->bytenr, &node->rb_node);
 648	spin_unlock(&rc->reloc_root_tree.lock);
 649	if (rb_node) {
 650		btrfs_err(fs_info,
 651			    "Duplicate root found for start=%llu while inserting into relocation tree",
 652			    node->bytenr);
 653		return -EEXIST;
 654	}
 655
 656	list_add_tail(&root->root_list, &rc->reloc_roots);
 657	return 0;
 658}
 659
 660/*
 661 * helper to delete the 'address of tree root -> reloc tree'
 662 * mapping
 663 */
 664static void __del_reloc_root(struct btrfs_root *root)
 665{
 666	struct btrfs_fs_info *fs_info = root->fs_info;
 667	struct rb_node *rb_node;
 668	struct mapping_node *node = NULL;
 669	struct reloc_control *rc = fs_info->reloc_ctl;
 670	bool put_ref = false;
 671
 672	if (rc && root->node) {
 673		spin_lock(&rc->reloc_root_tree.lock);
 674		rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
 675					   root->commit_root->start);
 676		if (rb_node) {
 677			node = rb_entry(rb_node, struct mapping_node, rb_node);
 678			rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
 679			RB_CLEAR_NODE(&node->rb_node);
 680		}
 681		spin_unlock(&rc->reloc_root_tree.lock);
 682		ASSERT(!node || (struct btrfs_root *)node->data == root);
 683	}
 684
 685	/*
 686	 * We only put the reloc root here if it's on the list.  There's a lot
 687	 * of places where the pattern is to splice the rc->reloc_roots, process
 688	 * the reloc roots, and then add the reloc root back onto
 689	 * rc->reloc_roots.  If we call __del_reloc_root while it's off of the
 690	 * list we don't want the reference being dropped, because the guy
 691	 * messing with the list is in charge of the reference.
 692	 */
 693	spin_lock(&fs_info->trans_lock);
 694	if (!list_empty(&root->root_list)) {
 695		put_ref = true;
 696		list_del_init(&root->root_list);
 697	}
 698	spin_unlock(&fs_info->trans_lock);
 699	if (put_ref)
 700		btrfs_put_root(root);
 701	kfree(node);
 702}
 703
 704/*
 705 * helper to update the 'address of tree root -> reloc tree'
 706 * mapping
 707 */
 708static int __update_reloc_root(struct btrfs_root *root)
 709{
 710	struct btrfs_fs_info *fs_info = root->fs_info;
 711	struct rb_node *rb_node;
 712	struct mapping_node *node = NULL;
 713	struct reloc_control *rc = fs_info->reloc_ctl;
 714
 715	spin_lock(&rc->reloc_root_tree.lock);
 716	rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
 717				   root->commit_root->start);
 718	if (rb_node) {
 719		node = rb_entry(rb_node, struct mapping_node, rb_node);
 720		rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
 721	}
 722	spin_unlock(&rc->reloc_root_tree.lock);
 723
 724	if (!node)
 725		return 0;
 726	BUG_ON((struct btrfs_root *)node->data != root);
 727
 728	spin_lock(&rc->reloc_root_tree.lock);
 729	node->bytenr = root->node->start;
 730	rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
 731				   node->bytenr, &node->rb_node);
 732	spin_unlock(&rc->reloc_root_tree.lock);
 733	if (rb_node)
 734		btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
 735	return 0;
 736}
 737
 738static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
 739					struct btrfs_root *root, u64 objectid)
 740{
 741	struct btrfs_fs_info *fs_info = root->fs_info;
 742	struct btrfs_root *reloc_root;
 743	struct extent_buffer *eb;
 744	struct btrfs_root_item *root_item;
 745	struct btrfs_key root_key;
 746	int ret = 0;
 747	bool must_abort = false;
 748
 749	root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
 750	if (!root_item)
 751		return ERR_PTR(-ENOMEM);
 752
 753	root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
 754	root_key.type = BTRFS_ROOT_ITEM_KEY;
 755	root_key.offset = objectid;
 756
 757	if (root->root_key.objectid == objectid) {
 758		u64 commit_root_gen;
 759
 760		/* called by btrfs_init_reloc_root */
 761		ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
 762				      BTRFS_TREE_RELOC_OBJECTID);
 763		if (ret)
 764			goto fail;
 765
 766		/*
 767		 * Set the last_snapshot field to the generation of the commit
 768		 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
 769		 * correctly (returns true) when the relocation root is created
 770		 * either inside the critical section of a transaction commit
 771		 * (through transaction.c:qgroup_account_snapshot()) and when
 772		 * it's created before the transaction commit is started.
 773		 */
 774		commit_root_gen = btrfs_header_generation(root->commit_root);
 775		btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
 776	} else {
 777		/*
 778		 * called by btrfs_reloc_post_snapshot_hook.
 779		 * the source tree is a reloc tree, all tree blocks
 780		 * modified after it was created have RELOC flag
 781		 * set in their headers. so it's OK to not update
 782		 * the 'last_snapshot'.
 783		 */
 784		ret = btrfs_copy_root(trans, root, root->node, &eb,
 785				      BTRFS_TREE_RELOC_OBJECTID);
 786		if (ret)
 787			goto fail;
 788	}
 789
 790	/*
 791	 * We have changed references at this point, we must abort the
 792	 * transaction if anything fails.
 793	 */
 794	must_abort = true;
 795
 796	memcpy(root_item, &root->root_item, sizeof(*root_item));
 797	btrfs_set_root_bytenr(root_item, eb->start);
 798	btrfs_set_root_level(root_item, btrfs_header_level(eb));
 799	btrfs_set_root_generation(root_item, trans->transid);
 800
 801	if (root->root_key.objectid == objectid) {
 802		btrfs_set_root_refs(root_item, 0);
 803		memset(&root_item->drop_progress, 0,
 804		       sizeof(struct btrfs_disk_key));
 805		btrfs_set_root_drop_level(root_item, 0);
 806	}
 807
 808	btrfs_tree_unlock(eb);
 809	free_extent_buffer(eb);
 810
 811	ret = btrfs_insert_root(trans, fs_info->tree_root,
 812				&root_key, root_item);
 813	if (ret)
 814		goto fail;
 815
 816	kfree(root_item);
 817
 818	reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
 819	if (IS_ERR(reloc_root)) {
 820		ret = PTR_ERR(reloc_root);
 821		goto abort;
 822	}
 823	set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
 824	reloc_root->last_trans = trans->transid;
 825	return reloc_root;
 826fail:
 827	kfree(root_item);
 828abort:
 829	if (must_abort)
 830		btrfs_abort_transaction(trans, ret);
 831	return ERR_PTR(ret);
 832}
 833
 834/*
 835 * create reloc tree for a given fs tree. reloc tree is just a
 836 * snapshot of the fs tree with special root objectid.
 837 *
 838 * The reloc_root comes out of here with two references, one for
 839 * root->reloc_root, and another for being on the rc->reloc_roots list.
 840 */
 841int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
 842			  struct btrfs_root *root)
 843{
 844	struct btrfs_fs_info *fs_info = root->fs_info;
 845	struct btrfs_root *reloc_root;
 846	struct reloc_control *rc = fs_info->reloc_ctl;
 847	struct btrfs_block_rsv *rsv;
 848	int clear_rsv = 0;
 849	int ret;
 850
 851	if (!rc)
 852		return 0;
 853
 854	/*
 855	 * The subvolume has reloc tree but the swap is finished, no need to
 856	 * create/update the dead reloc tree
 857	 */
 858	if (reloc_root_is_dead(root))
 859		return 0;
 860
 861	/*
 862	 * This is subtle but important.  We do not do
 863	 * record_root_in_transaction for reloc roots, instead we record their
 864	 * corresponding fs root, and then here we update the last trans for the
 865	 * reloc root.  This means that we have to do this for the entire life
 866	 * of the reloc root, regardless of which stage of the relocation we are
 867	 * in.
 868	 */
 869	if (root->reloc_root) {
 870		reloc_root = root->reloc_root;
 871		reloc_root->last_trans = trans->transid;
 872		return 0;
 873	}
 874
 875	/*
 876	 * We are merging reloc roots, we do not need new reloc trees.  Also
 877	 * reloc trees never need their own reloc tree.
 878	 */
 879	if (!rc->create_reloc_tree ||
 880	    root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 881		return 0;
 882
 883	if (!trans->reloc_reserved) {
 884		rsv = trans->block_rsv;
 885		trans->block_rsv = rc->block_rsv;
 886		clear_rsv = 1;
 887	}
 888	reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
 889	if (clear_rsv)
 890		trans->block_rsv = rsv;
 891	if (IS_ERR(reloc_root))
 892		return PTR_ERR(reloc_root);
 893
 894	ret = __add_reloc_root(reloc_root);
 895	ASSERT(ret != -EEXIST);
 896	if (ret) {
 897		/* Pairs with create_reloc_root */
 898		btrfs_put_root(reloc_root);
 899		return ret;
 900	}
 901	root->reloc_root = btrfs_grab_root(reloc_root);
 902	return 0;
 903}
 904
 905/*
 906 * update root item of reloc tree
 907 */
 908int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
 909			    struct btrfs_root *root)
 910{
 911	struct btrfs_fs_info *fs_info = root->fs_info;
 912	struct btrfs_root *reloc_root;
 913	struct btrfs_root_item *root_item;
 914	int ret;
 915
 916	if (!have_reloc_root(root))
 917		return 0;
 918
 919	reloc_root = root->reloc_root;
 920	root_item = &reloc_root->root_item;
 921
 922	/*
 923	 * We are probably ok here, but __del_reloc_root() will drop its ref of
 924	 * the root.  We have the ref for root->reloc_root, but just in case
 925	 * hold it while we update the reloc root.
 926	 */
 927	btrfs_grab_root(reloc_root);
 928
 929	/* root->reloc_root will stay until current relocation finished */
 930	if (fs_info->reloc_ctl->merge_reloc_tree &&
 931	    btrfs_root_refs(root_item) == 0) {
 932		set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
 933		/*
 934		 * Mark the tree as dead before we change reloc_root so
 935		 * have_reloc_root will not touch it from now on.
 936		 */
 937		smp_wmb();
 938		__del_reloc_root(reloc_root);
 939	}
 940
 941	if (reloc_root->commit_root != reloc_root->node) {
 942		__update_reloc_root(reloc_root);
 943		btrfs_set_root_node(root_item, reloc_root->node);
 944		free_extent_buffer(reloc_root->commit_root);
 945		reloc_root->commit_root = btrfs_root_node(reloc_root);
 946	}
 947
 948	ret = btrfs_update_root(trans, fs_info->tree_root,
 949				&reloc_root->root_key, root_item);
 950	btrfs_put_root(reloc_root);
 951	return ret;
 952}
 953
 954/*
 955 * helper to find first cached inode with inode number >= objectid
 956 * in a subvolume
 957 */
 958static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
 959{
 960	struct rb_node *node;
 961	struct rb_node *prev;
 962	struct btrfs_inode *entry;
 963	struct inode *inode;
 964
 965	spin_lock(&root->inode_lock);
 966again:
 967	node = root->inode_tree.rb_node;
 968	prev = NULL;
 969	while (node) {
 970		prev = node;
 971		entry = rb_entry(node, struct btrfs_inode, rb_node);
 972
 973		if (objectid < btrfs_ino(entry))
 974			node = node->rb_left;
 975		else if (objectid > btrfs_ino(entry))
 976			node = node->rb_right;
 977		else
 978			break;
 979	}
 980	if (!node) {
 981		while (prev) {
 982			entry = rb_entry(prev, struct btrfs_inode, rb_node);
 983			if (objectid <= btrfs_ino(entry)) {
 984				node = prev;
 985				break;
 986			}
 987			prev = rb_next(prev);
 988		}
 989	}
 990	while (node) {
 991		entry = rb_entry(node, struct btrfs_inode, rb_node);
 992		inode = igrab(&entry->vfs_inode);
 993		if (inode) {
 994			spin_unlock(&root->inode_lock);
 995			return inode;
 996		}
 997
 998		objectid = btrfs_ino(entry) + 1;
 999		if (cond_resched_lock(&root->inode_lock))
1000			goto again;
1001
1002		node = rb_next(node);
1003	}
1004	spin_unlock(&root->inode_lock);
1005	return NULL;
1006}
1007
1008/*
1009 * get new location of data
1010 */
1011static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1012			    u64 bytenr, u64 num_bytes)
1013{
1014	struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1015	struct btrfs_path *path;
1016	struct btrfs_file_extent_item *fi;
1017	struct extent_buffer *leaf;
1018	int ret;
1019
1020	path = btrfs_alloc_path();
1021	if (!path)
1022		return -ENOMEM;
1023
1024	bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1025	ret = btrfs_lookup_file_extent(NULL, root, path,
1026			btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1027	if (ret < 0)
1028		goto out;
1029	if (ret > 0) {
1030		ret = -ENOENT;
1031		goto out;
1032	}
1033
1034	leaf = path->nodes[0];
1035	fi = btrfs_item_ptr(leaf, path->slots[0],
1036			    struct btrfs_file_extent_item);
1037
1038	BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1039	       btrfs_file_extent_compression(leaf, fi) ||
1040	       btrfs_file_extent_encryption(leaf, fi) ||
1041	       btrfs_file_extent_other_encoding(leaf, fi));
1042
1043	if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1044		ret = -EINVAL;
1045		goto out;
1046	}
1047
1048	*new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1049	ret = 0;
1050out:
1051	btrfs_free_path(path);
1052	return ret;
1053}
1054
1055/*
1056 * update file extent items in the tree leaf to point to
1057 * the new locations.
1058 */
1059static noinline_for_stack
1060int replace_file_extents(struct btrfs_trans_handle *trans,
1061			 struct reloc_control *rc,
1062			 struct btrfs_root *root,
1063			 struct extent_buffer *leaf)
1064{
1065	struct btrfs_fs_info *fs_info = root->fs_info;
1066	struct btrfs_key key;
1067	struct btrfs_file_extent_item *fi;
1068	struct inode *inode = NULL;
1069	u64 parent;
1070	u64 bytenr;
1071	u64 new_bytenr = 0;
1072	u64 num_bytes;
1073	u64 end;
1074	u32 nritems;
1075	u32 i;
1076	int ret = 0;
1077	int first = 1;
1078	int dirty = 0;
1079
1080	if (rc->stage != UPDATE_DATA_PTRS)
1081		return 0;
1082
1083	/* reloc trees always use full backref */
1084	if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1085		parent = leaf->start;
1086	else
1087		parent = 0;
1088
1089	nritems = btrfs_header_nritems(leaf);
1090	for (i = 0; i < nritems; i++) {
1091		struct btrfs_ref ref = { 0 };
1092
1093		cond_resched();
1094		btrfs_item_key_to_cpu(leaf, &key, i);
1095		if (key.type != BTRFS_EXTENT_DATA_KEY)
1096			continue;
1097		fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1098		if (btrfs_file_extent_type(leaf, fi) ==
1099		    BTRFS_FILE_EXTENT_INLINE)
1100			continue;
1101		bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1102		num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1103		if (bytenr == 0)
1104			continue;
1105		if (!in_range(bytenr, rc->block_group->start,
1106			      rc->block_group->length))
1107			continue;
1108
1109		/*
1110		 * if we are modifying block in fs tree, wait for read_folio
1111		 * to complete and drop the extent cache
1112		 */
1113		if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1114			if (first) {
1115				inode = find_next_inode(root, key.objectid);
1116				first = 0;
1117			} else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1118				btrfs_add_delayed_iput(BTRFS_I(inode));
1119				inode = find_next_inode(root, key.objectid);
1120			}
1121			if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1122				struct extent_state *cached_state = NULL;
1123
1124				end = key.offset +
1125				      btrfs_file_extent_num_bytes(leaf, fi);
1126				WARN_ON(!IS_ALIGNED(key.offset,
1127						    fs_info->sectorsize));
1128				WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1129				end--;
1130				ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1131						      key.offset, end,
1132						      &cached_state);
1133				if (!ret)
1134					continue;
1135
1136				btrfs_drop_extent_map_range(BTRFS_I(inode),
1137							    key.offset, end, true);
1138				unlock_extent(&BTRFS_I(inode)->io_tree,
1139					      key.offset, end, &cached_state);
1140			}
1141		}
1142
1143		ret = get_new_location(rc->data_inode, &new_bytenr,
1144				       bytenr, num_bytes);
1145		if (ret) {
1146			/*
1147			 * Don't have to abort since we've not changed anything
1148			 * in the file extent yet.
1149			 */
1150			break;
1151		}
1152
1153		btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1154		dirty = 1;
1155
1156		key.offset -= btrfs_file_extent_offset(leaf, fi);
1157		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1158				       num_bytes, parent, root->root_key.objectid);
1159		btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1160				    key.objectid, key.offset,
1161				    root->root_key.objectid, false);
1162		ret = btrfs_inc_extent_ref(trans, &ref);
1163		if (ret) {
1164			btrfs_abort_transaction(trans, ret);
1165			break;
1166		}
1167
1168		btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1169				       num_bytes, parent, root->root_key.objectid);
1170		btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1171				    key.objectid, key.offset,
1172				    root->root_key.objectid, false);
1173		ret = btrfs_free_extent(trans, &ref);
1174		if (ret) {
1175			btrfs_abort_transaction(trans, ret);
1176			break;
1177		}
1178	}
1179	if (dirty)
1180		btrfs_mark_buffer_dirty(trans, leaf);
1181	if (inode)
1182		btrfs_add_delayed_iput(BTRFS_I(inode));
1183	return ret;
1184}
1185
1186static noinline_for_stack int memcmp_node_keys(const struct extent_buffer *eb,
1187					       int slot, const struct btrfs_path *path,
1188					       int level)
1189{
1190	struct btrfs_disk_key key1;
1191	struct btrfs_disk_key key2;
1192	btrfs_node_key(eb, &key1, slot);
1193	btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1194	return memcmp(&key1, &key2, sizeof(key1));
1195}
1196
1197/*
1198 * try to replace tree blocks in fs tree with the new blocks
1199 * in reloc tree. tree blocks haven't been modified since the
1200 * reloc tree was create can be replaced.
1201 *
1202 * if a block was replaced, level of the block + 1 is returned.
1203 * if no block got replaced, 0 is returned. if there are other
1204 * errors, a negative error number is returned.
1205 */
1206static noinline_for_stack
1207int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1208		 struct btrfs_root *dest, struct btrfs_root *src,
1209		 struct btrfs_path *path, struct btrfs_key *next_key,
1210		 int lowest_level, int max_level)
1211{
1212	struct btrfs_fs_info *fs_info = dest->fs_info;
1213	struct extent_buffer *eb;
1214	struct extent_buffer *parent;
1215	struct btrfs_ref ref = { 0 };
1216	struct btrfs_key key;
1217	u64 old_bytenr;
1218	u64 new_bytenr;
1219	u64 old_ptr_gen;
1220	u64 new_ptr_gen;
1221	u64 last_snapshot;
1222	u32 blocksize;
1223	int cow = 0;
1224	int level;
1225	int ret;
1226	int slot;
1227
1228	ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1229	ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1230
1231	last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1232again:
1233	slot = path->slots[lowest_level];
1234	btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1235
1236	eb = btrfs_lock_root_node(dest);
1237	level = btrfs_header_level(eb);
1238
1239	if (level < lowest_level) {
1240		btrfs_tree_unlock(eb);
1241		free_extent_buffer(eb);
1242		return 0;
1243	}
1244
1245	if (cow) {
1246		ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1247				      BTRFS_NESTING_COW);
1248		if (ret) {
1249			btrfs_tree_unlock(eb);
1250			free_extent_buffer(eb);
1251			return ret;
1252		}
1253	}
1254
1255	if (next_key) {
1256		next_key->objectid = (u64)-1;
1257		next_key->type = (u8)-1;
1258		next_key->offset = (u64)-1;
1259	}
1260
1261	parent = eb;
1262	while (1) {
1263		level = btrfs_header_level(parent);
1264		ASSERT(level >= lowest_level);
1265
1266		ret = btrfs_bin_search(parent, 0, &key, &slot);
1267		if (ret < 0)
1268			break;
1269		if (ret && slot > 0)
1270			slot--;
1271
1272		if (next_key && slot + 1 < btrfs_header_nritems(parent))
1273			btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1274
1275		old_bytenr = btrfs_node_blockptr(parent, slot);
1276		blocksize = fs_info->nodesize;
1277		old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1278
1279		if (level <= max_level) {
1280			eb = path->nodes[level];
1281			new_bytenr = btrfs_node_blockptr(eb,
1282							path->slots[level]);
1283			new_ptr_gen = btrfs_node_ptr_generation(eb,
1284							path->slots[level]);
1285		} else {
1286			new_bytenr = 0;
1287			new_ptr_gen = 0;
1288		}
1289
1290		if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1291			ret = level;
1292			break;
1293		}
1294
1295		if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1296		    memcmp_node_keys(parent, slot, path, level)) {
1297			if (level <= lowest_level) {
1298				ret = 0;
1299				break;
1300			}
1301
1302			eb = btrfs_read_node_slot(parent, slot);
1303			if (IS_ERR(eb)) {
1304				ret = PTR_ERR(eb);
1305				break;
1306			}
1307			btrfs_tree_lock(eb);
1308			if (cow) {
1309				ret = btrfs_cow_block(trans, dest, eb, parent,
1310						      slot, &eb,
1311						      BTRFS_NESTING_COW);
1312				if (ret) {
1313					btrfs_tree_unlock(eb);
1314					free_extent_buffer(eb);
1315					break;
1316				}
1317			}
1318
1319			btrfs_tree_unlock(parent);
1320			free_extent_buffer(parent);
1321
1322			parent = eb;
1323			continue;
1324		}
1325
1326		if (!cow) {
1327			btrfs_tree_unlock(parent);
1328			free_extent_buffer(parent);
1329			cow = 1;
1330			goto again;
1331		}
1332
1333		btrfs_node_key_to_cpu(path->nodes[level], &key,
1334				      path->slots[level]);
1335		btrfs_release_path(path);
1336
1337		path->lowest_level = level;
1338		set_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1339		ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1340		clear_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1341		path->lowest_level = 0;
1342		if (ret) {
1343			if (ret > 0)
1344				ret = -ENOENT;
1345			break;
1346		}
1347
1348		/*
1349		 * Info qgroup to trace both subtrees.
1350		 *
1351		 * We must trace both trees.
1352		 * 1) Tree reloc subtree
1353		 *    If not traced, we will leak data numbers
1354		 * 2) Fs subtree
1355		 *    If not traced, we will double count old data
1356		 *
1357		 * We don't scan the subtree right now, but only record
1358		 * the swapped tree blocks.
1359		 * The real subtree rescan is delayed until we have new
1360		 * CoW on the subtree root node before transaction commit.
1361		 */
1362		ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1363				rc->block_group, parent, slot,
1364				path->nodes[level], path->slots[level],
1365				last_snapshot);
1366		if (ret < 0)
1367			break;
1368		/*
1369		 * swap blocks in fs tree and reloc tree.
1370		 */
1371		btrfs_set_node_blockptr(parent, slot, new_bytenr);
1372		btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1373		btrfs_mark_buffer_dirty(trans, parent);
1374
1375		btrfs_set_node_blockptr(path->nodes[level],
1376					path->slots[level], old_bytenr);
1377		btrfs_set_node_ptr_generation(path->nodes[level],
1378					      path->slots[level], old_ptr_gen);
1379		btrfs_mark_buffer_dirty(trans, path->nodes[level]);
1380
1381		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1382				       blocksize, path->nodes[level]->start,
1383				       src->root_key.objectid);
1384		btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1385				    0, true);
1386		ret = btrfs_inc_extent_ref(trans, &ref);
1387		if (ret) {
1388			btrfs_abort_transaction(trans, ret);
1389			break;
1390		}
1391		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1392				       blocksize, 0, dest->root_key.objectid);
1393		btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid, 0,
1394				    true);
1395		ret = btrfs_inc_extent_ref(trans, &ref);
1396		if (ret) {
1397			btrfs_abort_transaction(trans, ret);
1398			break;
1399		}
1400
1401		/* We don't know the real owning_root, use 0. */
1402		btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1403				       blocksize, path->nodes[level]->start, 0);
1404		btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1405				    0, true);
1406		ret = btrfs_free_extent(trans, &ref);
1407		if (ret) {
1408			btrfs_abort_transaction(trans, ret);
1409			break;
1410		}
1411
1412		/* We don't know the real owning_root, use 0. */
1413		btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1414				       blocksize, 0, 0);
1415		btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid,
1416				    0, true);
1417		ret = btrfs_free_extent(trans, &ref);
1418		if (ret) {
1419			btrfs_abort_transaction(trans, ret);
1420			break;
1421		}
1422
1423		btrfs_unlock_up_safe(path, 0);
1424
1425		ret = level;
1426		break;
1427	}
1428	btrfs_tree_unlock(parent);
1429	free_extent_buffer(parent);
1430	return ret;
1431}
1432
1433/*
1434 * helper to find next relocated block in reloc tree
1435 */
1436static noinline_for_stack
1437int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1438		       int *level)
1439{
1440	struct extent_buffer *eb;
1441	int i;
1442	u64 last_snapshot;
1443	u32 nritems;
1444
1445	last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1446
1447	for (i = 0; i < *level; i++) {
1448		free_extent_buffer(path->nodes[i]);
1449		path->nodes[i] = NULL;
1450	}
1451
1452	for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1453		eb = path->nodes[i];
1454		nritems = btrfs_header_nritems(eb);
1455		while (path->slots[i] + 1 < nritems) {
1456			path->slots[i]++;
1457			if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1458			    last_snapshot)
1459				continue;
1460
1461			*level = i;
1462			return 0;
1463		}
1464		free_extent_buffer(path->nodes[i]);
1465		path->nodes[i] = NULL;
1466	}
1467	return 1;
1468}
1469
1470/*
1471 * walk down reloc tree to find relocated block of lowest level
1472 */
1473static noinline_for_stack
1474int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1475			 int *level)
1476{
1477	struct extent_buffer *eb = NULL;
1478	int i;
1479	u64 ptr_gen = 0;
1480	u64 last_snapshot;
1481	u32 nritems;
1482
1483	last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1484
1485	for (i = *level; i > 0; i--) {
1486		eb = path->nodes[i];
1487		nritems = btrfs_header_nritems(eb);
1488		while (path->slots[i] < nritems) {
1489			ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1490			if (ptr_gen > last_snapshot)
1491				break;
1492			path->slots[i]++;
1493		}
1494		if (path->slots[i] >= nritems) {
1495			if (i == *level)
1496				break;
1497			*level = i + 1;
1498			return 0;
1499		}
1500		if (i == 1) {
1501			*level = i;
1502			return 0;
1503		}
1504
1505		eb = btrfs_read_node_slot(eb, path->slots[i]);
1506		if (IS_ERR(eb))
1507			return PTR_ERR(eb);
1508		BUG_ON(btrfs_header_level(eb) != i - 1);
1509		path->nodes[i - 1] = eb;
1510		path->slots[i - 1] = 0;
1511	}
1512	return 1;
1513}
1514
1515/*
1516 * invalidate extent cache for file extents whose key in range of
1517 * [min_key, max_key)
1518 */
1519static int invalidate_extent_cache(struct btrfs_root *root,
1520				   const struct btrfs_key *min_key,
1521				   const struct btrfs_key *max_key)
1522{
1523	struct btrfs_fs_info *fs_info = root->fs_info;
1524	struct inode *inode = NULL;
1525	u64 objectid;
1526	u64 start, end;
1527	u64 ino;
1528
1529	objectid = min_key->objectid;
1530	while (1) {
1531		struct extent_state *cached_state = NULL;
1532
1533		cond_resched();
1534		iput(inode);
1535
1536		if (objectid > max_key->objectid)
1537			break;
1538
1539		inode = find_next_inode(root, objectid);
1540		if (!inode)
1541			break;
1542		ino = btrfs_ino(BTRFS_I(inode));
1543
1544		if (ino > max_key->objectid) {
1545			iput(inode);
1546			break;
1547		}
1548
1549		objectid = ino + 1;
1550		if (!S_ISREG(inode->i_mode))
1551			continue;
1552
1553		if (unlikely(min_key->objectid == ino)) {
1554			if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1555				continue;
1556			if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1557				start = 0;
1558			else {
1559				start = min_key->offset;
1560				WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1561			}
1562		} else {
1563			start = 0;
1564		}
1565
1566		if (unlikely(max_key->objectid == ino)) {
1567			if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1568				continue;
1569			if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1570				end = (u64)-1;
1571			} else {
1572				if (max_key->offset == 0)
1573					continue;
1574				end = max_key->offset;
1575				WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1576				end--;
1577			}
1578		} else {
1579			end = (u64)-1;
1580		}
1581
1582		/* the lock_extent waits for read_folio to complete */
1583		lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1584		btrfs_drop_extent_map_range(BTRFS_I(inode), start, end, true);
1585		unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1586	}
1587	return 0;
1588}
1589
1590static int find_next_key(struct btrfs_path *path, int level,
1591			 struct btrfs_key *key)
1592
1593{
1594	while (level < BTRFS_MAX_LEVEL) {
1595		if (!path->nodes[level])
1596			break;
1597		if (path->slots[level] + 1 <
1598		    btrfs_header_nritems(path->nodes[level])) {
1599			btrfs_node_key_to_cpu(path->nodes[level], key,
1600					      path->slots[level] + 1);
1601			return 0;
1602		}
1603		level++;
1604	}
1605	return 1;
1606}
1607
1608/*
1609 * Insert current subvolume into reloc_control::dirty_subvol_roots
1610 */
1611static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1612			       struct reloc_control *rc,
1613			       struct btrfs_root *root)
1614{
1615	struct btrfs_root *reloc_root = root->reloc_root;
1616	struct btrfs_root_item *reloc_root_item;
1617	int ret;
1618
1619	/* @root must be a subvolume tree root with a valid reloc tree */
1620	ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1621	ASSERT(reloc_root);
1622
1623	reloc_root_item = &reloc_root->root_item;
1624	memset(&reloc_root_item->drop_progress, 0,
1625		sizeof(reloc_root_item->drop_progress));
1626	btrfs_set_root_drop_level(reloc_root_item, 0);
1627	btrfs_set_root_refs(reloc_root_item, 0);
1628	ret = btrfs_update_reloc_root(trans, root);
1629	if (ret)
1630		return ret;
1631
1632	if (list_empty(&root->reloc_dirty_list)) {
1633		btrfs_grab_root(root);
1634		list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1635	}
1636
1637	return 0;
1638}
1639
1640static int clean_dirty_subvols(struct reloc_control *rc)
1641{
1642	struct btrfs_root *root;
1643	struct btrfs_root *next;
1644	int ret = 0;
1645	int ret2;
1646
1647	list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1648				 reloc_dirty_list) {
1649		if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1650			/* Merged subvolume, cleanup its reloc root */
1651			struct btrfs_root *reloc_root = root->reloc_root;
1652
1653			list_del_init(&root->reloc_dirty_list);
1654			root->reloc_root = NULL;
1655			/*
1656			 * Need barrier to ensure clear_bit() only happens after
1657			 * root->reloc_root = NULL. Pairs with have_reloc_root.
1658			 */
1659			smp_wmb();
1660			clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1661			if (reloc_root) {
1662				/*
1663				 * btrfs_drop_snapshot drops our ref we hold for
1664				 * ->reloc_root.  If it fails however we must
1665				 * drop the ref ourselves.
1666				 */
1667				ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1668				if (ret2 < 0) {
1669					btrfs_put_root(reloc_root);
1670					if (!ret)
1671						ret = ret2;
1672				}
1673			}
1674			btrfs_put_root(root);
1675		} else {
1676			/* Orphan reloc tree, just clean it up */
1677			ret2 = btrfs_drop_snapshot(root, 0, 1);
1678			if (ret2 < 0) {
1679				btrfs_put_root(root);
1680				if (!ret)
1681					ret = ret2;
1682			}
1683		}
1684	}
1685	return ret;
1686}
1687
1688/*
1689 * merge the relocated tree blocks in reloc tree with corresponding
1690 * fs tree.
1691 */
1692static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1693					       struct btrfs_root *root)
1694{
1695	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1696	struct btrfs_key key;
1697	struct btrfs_key next_key;
1698	struct btrfs_trans_handle *trans = NULL;
1699	struct btrfs_root *reloc_root;
1700	struct btrfs_root_item *root_item;
1701	struct btrfs_path *path;
1702	struct extent_buffer *leaf;
1703	int reserve_level;
1704	int level;
1705	int max_level;
1706	int replaced = 0;
1707	int ret = 0;
1708	u32 min_reserved;
1709
1710	path = btrfs_alloc_path();
1711	if (!path)
1712		return -ENOMEM;
1713	path->reada = READA_FORWARD;
1714
1715	reloc_root = root->reloc_root;
1716	root_item = &reloc_root->root_item;
1717
1718	if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1719		level = btrfs_root_level(root_item);
1720		atomic_inc(&reloc_root->node->refs);
1721		path->nodes[level] = reloc_root->node;
1722		path->slots[level] = 0;
1723	} else {
1724		btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1725
1726		level = btrfs_root_drop_level(root_item);
1727		BUG_ON(level == 0);
1728		path->lowest_level = level;
1729		ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1730		path->lowest_level = 0;
1731		if (ret < 0) {
1732			btrfs_free_path(path);
1733			return ret;
1734		}
1735
1736		btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1737				      path->slots[level]);
1738		WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1739
1740		btrfs_unlock_up_safe(path, 0);
1741	}
1742
1743	/*
1744	 * In merge_reloc_root(), we modify the upper level pointer to swap the
1745	 * tree blocks between reloc tree and subvolume tree.  Thus for tree
1746	 * block COW, we COW at most from level 1 to root level for each tree.
1747	 *
1748	 * Thus the needed metadata size is at most root_level * nodesize,
1749	 * and * 2 since we have two trees to COW.
1750	 */
1751	reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1752	min_reserved = fs_info->nodesize * reserve_level * 2;
1753	memset(&next_key, 0, sizeof(next_key));
1754
1755	while (1) {
1756		ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
1757					     min_reserved,
1758					     BTRFS_RESERVE_FLUSH_LIMIT);
1759		if (ret)
1760			goto out;
1761		trans = btrfs_start_transaction(root, 0);
1762		if (IS_ERR(trans)) {
1763			ret = PTR_ERR(trans);
1764			trans = NULL;
1765			goto out;
1766		}
1767
1768		/*
1769		 * At this point we no longer have a reloc_control, so we can't
1770		 * depend on btrfs_init_reloc_root to update our last_trans.
1771		 *
1772		 * But that's ok, we started the trans handle on our
1773		 * corresponding fs_root, which means it's been added to the
1774		 * dirty list.  At commit time we'll still call
1775		 * btrfs_update_reloc_root() and update our root item
1776		 * appropriately.
1777		 */
1778		reloc_root->last_trans = trans->transid;
1779		trans->block_rsv = rc->block_rsv;
1780
1781		replaced = 0;
1782		max_level = level;
1783
1784		ret = walk_down_reloc_tree(reloc_root, path, &level);
1785		if (ret < 0)
1786			goto out;
1787		if (ret > 0)
1788			break;
1789
1790		if (!find_next_key(path, level, &key) &&
1791		    btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1792			ret = 0;
1793		} else {
1794			ret = replace_path(trans, rc, root, reloc_root, path,
1795					   &next_key, level, max_level);
1796		}
1797		if (ret < 0)
1798			goto out;
1799		if (ret > 0) {
1800			level = ret;
1801			btrfs_node_key_to_cpu(path->nodes[level], &key,
1802					      path->slots[level]);
1803			replaced = 1;
1804		}
1805
1806		ret = walk_up_reloc_tree(reloc_root, path, &level);
1807		if (ret > 0)
1808			break;
1809
1810		BUG_ON(level == 0);
1811		/*
1812		 * save the merging progress in the drop_progress.
1813		 * this is OK since root refs == 1 in this case.
1814		 */
1815		btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1816			       path->slots[level]);
1817		btrfs_set_root_drop_level(root_item, level);
1818
1819		btrfs_end_transaction_throttle(trans);
1820		trans = NULL;
1821
1822		btrfs_btree_balance_dirty(fs_info);
1823
1824		if (replaced && rc->stage == UPDATE_DATA_PTRS)
1825			invalidate_extent_cache(root, &key, &next_key);
1826	}
1827
1828	/*
1829	 * handle the case only one block in the fs tree need to be
1830	 * relocated and the block is tree root.
1831	 */
1832	leaf = btrfs_lock_root_node(root);
1833	ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1834			      BTRFS_NESTING_COW);
1835	btrfs_tree_unlock(leaf);
1836	free_extent_buffer(leaf);
1837out:
1838	btrfs_free_path(path);
1839
1840	if (ret == 0) {
1841		ret = insert_dirty_subvol(trans, rc, root);
1842		if (ret)
1843			btrfs_abort_transaction(trans, ret);
1844	}
1845
1846	if (trans)
1847		btrfs_end_transaction_throttle(trans);
1848
1849	btrfs_btree_balance_dirty(fs_info);
1850
1851	if (replaced && rc->stage == UPDATE_DATA_PTRS)
1852		invalidate_extent_cache(root, &key, &next_key);
1853
1854	return ret;
1855}
1856
1857static noinline_for_stack
1858int prepare_to_merge(struct reloc_control *rc, int err)
1859{
1860	struct btrfs_root *root = rc->extent_root;
1861	struct btrfs_fs_info *fs_info = root->fs_info;
1862	struct btrfs_root *reloc_root;
1863	struct btrfs_trans_handle *trans;
1864	LIST_HEAD(reloc_roots);
1865	u64 num_bytes = 0;
1866	int ret;
1867
1868	mutex_lock(&fs_info->reloc_mutex);
1869	rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1870	rc->merging_rsv_size += rc->nodes_relocated * 2;
1871	mutex_unlock(&fs_info->reloc_mutex);
1872
1873again:
1874	if (!err) {
1875		num_bytes = rc->merging_rsv_size;
1876		ret = btrfs_block_rsv_add(fs_info, rc->block_rsv, num_bytes,
1877					  BTRFS_RESERVE_FLUSH_ALL);
1878		if (ret)
1879			err = ret;
1880	}
1881
1882	trans = btrfs_join_transaction(rc->extent_root);
1883	if (IS_ERR(trans)) {
1884		if (!err)
1885			btrfs_block_rsv_release(fs_info, rc->block_rsv,
1886						num_bytes, NULL);
1887		return PTR_ERR(trans);
1888	}
1889
1890	if (!err) {
1891		if (num_bytes != rc->merging_rsv_size) {
1892			btrfs_end_transaction(trans);
1893			btrfs_block_rsv_release(fs_info, rc->block_rsv,
1894						num_bytes, NULL);
1895			goto again;
1896		}
1897	}
1898
1899	rc->merge_reloc_tree = true;
1900
1901	while (!list_empty(&rc->reloc_roots)) {
1902		reloc_root = list_entry(rc->reloc_roots.next,
1903					struct btrfs_root, root_list);
1904		list_del_init(&reloc_root->root_list);
1905
1906		root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1907				false);
1908		if (IS_ERR(root)) {
1909			/*
1910			 * Even if we have an error we need this reloc root
1911			 * back on our list so we can clean up properly.
1912			 */
1913			list_add(&reloc_root->root_list, &reloc_roots);
1914			btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1915			if (!err)
1916				err = PTR_ERR(root);
1917			break;
1918		}
1919
1920		if (unlikely(root->reloc_root != reloc_root)) {
1921			if (root->reloc_root) {
1922				btrfs_err(fs_info,
1923"reloc tree mismatch, root %lld has reloc root key (%lld %u %llu) gen %llu, expect reloc root key (%lld %u %llu) gen %llu",
1924					  root->root_key.objectid,
1925					  root->reloc_root->root_key.objectid,
1926					  root->reloc_root->root_key.type,
1927					  root->reloc_root->root_key.offset,
1928					  btrfs_root_generation(
1929						  &root->reloc_root->root_item),
1930					  reloc_root->root_key.objectid,
1931					  reloc_root->root_key.type,
1932					  reloc_root->root_key.offset,
1933					  btrfs_root_generation(
1934						  &reloc_root->root_item));
1935			} else {
1936				btrfs_err(fs_info,
1937"reloc tree mismatch, root %lld has no reloc root, expect reloc root key (%lld %u %llu) gen %llu",
1938					  root->root_key.objectid,
1939					  reloc_root->root_key.objectid,
1940					  reloc_root->root_key.type,
1941					  reloc_root->root_key.offset,
1942					  btrfs_root_generation(
1943						  &reloc_root->root_item));
1944			}
1945			list_add(&reloc_root->root_list, &reloc_roots);
1946			btrfs_put_root(root);
1947			btrfs_abort_transaction(trans, -EUCLEAN);
1948			if (!err)
1949				err = -EUCLEAN;
1950			break;
1951		}
1952
1953		/*
1954		 * set reference count to 1, so btrfs_recover_relocation
1955		 * knows it should resumes merging
1956		 */
1957		if (!err)
1958			btrfs_set_root_refs(&reloc_root->root_item, 1);
1959		ret = btrfs_update_reloc_root(trans, root);
1960
1961		/*
1962		 * Even if we have an error we need this reloc root back on our
1963		 * list so we can clean up properly.
1964		 */
1965		list_add(&reloc_root->root_list, &reloc_roots);
1966		btrfs_put_root(root);
1967
1968		if (ret) {
1969			btrfs_abort_transaction(trans, ret);
1970			if (!err)
1971				err = ret;
1972			break;
1973		}
1974	}
1975
1976	list_splice(&reloc_roots, &rc->reloc_roots);
1977
1978	if (!err)
1979		err = btrfs_commit_transaction(trans);
1980	else
1981		btrfs_end_transaction(trans);
1982	return err;
1983}
1984
1985static noinline_for_stack
1986void free_reloc_roots(struct list_head *list)
1987{
1988	struct btrfs_root *reloc_root, *tmp;
1989
1990	list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1991		__del_reloc_root(reloc_root);
1992}
1993
1994static noinline_for_stack
1995void merge_reloc_roots(struct reloc_control *rc)
1996{
1997	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1998	struct btrfs_root *root;
1999	struct btrfs_root *reloc_root;
2000	LIST_HEAD(reloc_roots);
2001	int found = 0;
2002	int ret = 0;
2003again:
2004	root = rc->extent_root;
2005
2006	/*
2007	 * this serializes us with btrfs_record_root_in_transaction,
2008	 * we have to make sure nobody is in the middle of
2009	 * adding their roots to the list while we are
2010	 * doing this splice
2011	 */
2012	mutex_lock(&fs_info->reloc_mutex);
2013	list_splice_init(&rc->reloc_roots, &reloc_roots);
2014	mutex_unlock(&fs_info->reloc_mutex);
2015
2016	while (!list_empty(&reloc_roots)) {
2017		found = 1;
2018		reloc_root = list_entry(reloc_roots.next,
2019					struct btrfs_root, root_list);
2020
2021		root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
2022					 false);
2023		if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2024			if (WARN_ON(IS_ERR(root))) {
2025				/*
2026				 * For recovery we read the fs roots on mount,
2027				 * and if we didn't find the root then we marked
2028				 * the reloc root as a garbage root.  For normal
2029				 * relocation obviously the root should exist in
2030				 * memory.  However there's no reason we can't
2031				 * handle the error properly here just in case.
2032				 */
2033				ret = PTR_ERR(root);
2034				goto out;
2035			}
2036			if (WARN_ON(root->reloc_root != reloc_root)) {
2037				/*
2038				 * This can happen if on-disk metadata has some
2039				 * corruption, e.g. bad reloc tree key offset.
2040				 */
2041				ret = -EINVAL;
2042				goto out;
2043			}
2044			ret = merge_reloc_root(rc, root);
2045			btrfs_put_root(root);
2046			if (ret) {
2047				if (list_empty(&reloc_root->root_list))
2048					list_add_tail(&reloc_root->root_list,
2049						      &reloc_roots);
2050				goto out;
2051			}
2052		} else {
2053			if (!IS_ERR(root)) {
2054				if (root->reloc_root == reloc_root) {
2055					root->reloc_root = NULL;
2056					btrfs_put_root(reloc_root);
2057				}
2058				clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
2059					  &root->state);
2060				btrfs_put_root(root);
2061			}
2062
2063			list_del_init(&reloc_root->root_list);
2064			/* Don't forget to queue this reloc root for cleanup */
2065			list_add_tail(&reloc_root->reloc_dirty_list,
2066				      &rc->dirty_subvol_roots);
2067		}
2068	}
2069
2070	if (found) {
2071		found = 0;
2072		goto again;
2073	}
2074out:
2075	if (ret) {
2076		btrfs_handle_fs_error(fs_info, ret, NULL);
2077		free_reloc_roots(&reloc_roots);
2078
2079		/* new reloc root may be added */
2080		mutex_lock(&fs_info->reloc_mutex);
2081		list_splice_init(&rc->reloc_roots, &reloc_roots);
2082		mutex_unlock(&fs_info->reloc_mutex);
2083		free_reloc_roots(&reloc_roots);
2084	}
2085
2086	/*
2087	 * We used to have
2088	 *
2089	 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2090	 *
2091	 * here, but it's wrong.  If we fail to start the transaction in
2092	 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2093	 * have actually been removed from the reloc_root_tree rb tree.  This is
2094	 * fine because we're bailing here, and we hold a reference on the root
2095	 * for the list that holds it, so these roots will be cleaned up when we
2096	 * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
2097	 * will be cleaned up on unmount.
2098	 *
2099	 * The remaining nodes will be cleaned up by free_reloc_control.
2100	 */
2101}
2102
2103static void free_block_list(struct rb_root *blocks)
2104{
2105	struct tree_block *block;
2106	struct rb_node *rb_node;
2107	while ((rb_node = rb_first(blocks))) {
2108		block = rb_entry(rb_node, struct tree_block, rb_node);
2109		rb_erase(rb_node, blocks);
2110		kfree(block);
2111	}
2112}
2113
2114static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2115				      struct btrfs_root *reloc_root)
2116{
2117	struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2118	struct btrfs_root *root;
2119	int ret;
2120
2121	if (reloc_root->last_trans == trans->transid)
2122		return 0;
2123
2124	root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2125
2126	/*
2127	 * This should succeed, since we can't have a reloc root without having
2128	 * already looked up the actual root and created the reloc root for this
2129	 * root.
2130	 *
2131	 * However if there's some sort of corruption where we have a ref to a
2132	 * reloc root without a corresponding root this could return ENOENT.
2133	 */
2134	if (IS_ERR(root)) {
2135		ASSERT(0);
2136		return PTR_ERR(root);
2137	}
2138	if (root->reloc_root != reloc_root) {
2139		ASSERT(0);
2140		btrfs_err(fs_info,
2141			  "root %llu has two reloc roots associated with it",
2142			  reloc_root->root_key.offset);
2143		btrfs_put_root(root);
2144		return -EUCLEAN;
2145	}
2146	ret = btrfs_record_root_in_trans(trans, root);
2147	btrfs_put_root(root);
2148
2149	return ret;
2150}
2151
2152static noinline_for_stack
2153struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2154				     struct reloc_control *rc,
2155				     struct btrfs_backref_node *node,
2156				     struct btrfs_backref_edge *edges[])
2157{
2158	struct btrfs_backref_node *next;
2159	struct btrfs_root *root;
2160	int index = 0;
2161	int ret;
2162
2163	next = node;
2164	while (1) {
2165		cond_resched();
2166		next = walk_up_backref(next, edges, &index);
2167		root = next->root;
2168
2169		/*
2170		 * If there is no root, then our references for this block are
2171		 * incomplete, as we should be able to walk all the way up to a
2172		 * block that is owned by a root.
2173		 *
2174		 * This path is only for SHAREABLE roots, so if we come upon a
2175		 * non-SHAREABLE root then we have backrefs that resolve
2176		 * improperly.
2177		 *
2178		 * Both of these cases indicate file system corruption, or a bug
2179		 * in the backref walking code.
2180		 */
2181		if (!root) {
2182			ASSERT(0);
2183			btrfs_err(trans->fs_info,
2184		"bytenr %llu doesn't have a backref path ending in a root",
2185				  node->bytenr);
2186			return ERR_PTR(-EUCLEAN);
2187		}
2188		if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2189			ASSERT(0);
2190			btrfs_err(trans->fs_info,
2191	"bytenr %llu has multiple refs with one ending in a non-shareable root",
2192				  node->bytenr);
2193			return ERR_PTR(-EUCLEAN);
2194		}
2195
2196		if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2197			ret = record_reloc_root_in_trans(trans, root);
2198			if (ret)
2199				return ERR_PTR(ret);
2200			break;
2201		}
2202
2203		ret = btrfs_record_root_in_trans(trans, root);
2204		if (ret)
2205			return ERR_PTR(ret);
2206		root = root->reloc_root;
2207
2208		/*
2209		 * We could have raced with another thread which failed, so
2210		 * root->reloc_root may not be set, return ENOENT in this case.
2211		 */
2212		if (!root)
2213			return ERR_PTR(-ENOENT);
2214
2215		if (next->new_bytenr != root->node->start) {
2216			/*
2217			 * We just created the reloc root, so we shouldn't have
2218			 * ->new_bytenr set and this shouldn't be in the changed
2219			 *  list.  If it is then we have multiple roots pointing
2220			 *  at the same bytenr which indicates corruption, or
2221			 *  we've made a mistake in the backref walking code.
2222			 */
2223			ASSERT(next->new_bytenr == 0);
2224			ASSERT(list_empty(&next->list));
2225			if (next->new_bytenr || !list_empty(&next->list)) {
2226				btrfs_err(trans->fs_info,
2227	"bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2228					  node->bytenr, next->bytenr);
2229				return ERR_PTR(-EUCLEAN);
2230			}
2231
2232			next->new_bytenr = root->node->start;
2233			btrfs_put_root(next->root);
2234			next->root = btrfs_grab_root(root);
2235			ASSERT(next->root);
2236			list_add_tail(&next->list,
2237				      &rc->backref_cache.changed);
2238			mark_block_processed(rc, next);
2239			break;
2240		}
2241
2242		WARN_ON(1);
2243		root = NULL;
2244		next = walk_down_backref(edges, &index);
2245		if (!next || next->level <= node->level)
2246			break;
2247	}
2248	if (!root) {
2249		/*
2250		 * This can happen if there's fs corruption or if there's a bug
2251		 * in the backref lookup code.
2252		 */
2253		ASSERT(0);
2254		return ERR_PTR(-ENOENT);
2255	}
2256
2257	next = node;
2258	/* setup backref node path for btrfs_reloc_cow_block */
2259	while (1) {
2260		rc->backref_cache.path[next->level] = next;
2261		if (--index < 0)
2262			break;
2263		next = edges[index]->node[UPPER];
2264	}
2265	return root;
2266}
2267
2268/*
2269 * Select a tree root for relocation.
2270 *
2271 * Return NULL if the block is not shareable. We should use do_relocation() in
2272 * this case.
2273 *
2274 * Return a tree root pointer if the block is shareable.
2275 * Return -ENOENT if the block is root of reloc tree.
2276 */
2277static noinline_for_stack
2278struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2279{
2280	struct btrfs_backref_node *next;
2281	struct btrfs_root *root;
2282	struct btrfs_root *fs_root = NULL;
2283	struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2284	int index = 0;
2285
2286	next = node;
2287	while (1) {
2288		cond_resched();
2289		next = walk_up_backref(next, edges, &index);
2290		root = next->root;
2291
2292		/*
2293		 * This can occur if we have incomplete extent refs leading all
2294		 * the way up a particular path, in this case return -EUCLEAN.
2295		 */
2296		if (!root)
2297			return ERR_PTR(-EUCLEAN);
2298
2299		/* No other choice for non-shareable tree */
2300		if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2301			return root;
2302
2303		if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2304			fs_root = root;
2305
2306		if (next != node)
2307			return NULL;
2308
2309		next = walk_down_backref(edges, &index);
2310		if (!next || next->level <= node->level)
2311			break;
2312	}
2313
2314	if (!fs_root)
2315		return ERR_PTR(-ENOENT);
2316	return fs_root;
2317}
2318
2319static noinline_for_stack
2320u64 calcu_metadata_size(struct reloc_control *rc,
2321			struct btrfs_backref_node *node, int reserve)
2322{
2323	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2324	struct btrfs_backref_node *next = node;
2325	struct btrfs_backref_edge *edge;
2326	struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2327	u64 num_bytes = 0;
2328	int index = 0;
2329
2330	BUG_ON(reserve && node->processed);
2331
2332	while (next) {
2333		cond_resched();
2334		while (1) {
2335			if (next->processed && (reserve || next != node))
2336				break;
2337
2338			num_bytes += fs_info->nodesize;
2339
2340			if (list_empty(&next->upper))
2341				break;
2342
2343			edge = list_entry(next->upper.next,
2344					struct btrfs_backref_edge, list[LOWER]);
2345			edges[index++] = edge;
2346			next = edge->node[UPPER];
2347		}
2348		next = walk_down_backref(edges, &index);
2349	}
2350	return num_bytes;
2351}
2352
2353static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2354				  struct reloc_control *rc,
2355				  struct btrfs_backref_node *node)
2356{
2357	struct btrfs_root *root = rc->extent_root;
2358	struct btrfs_fs_info *fs_info = root->fs_info;
2359	u64 num_bytes;
2360	int ret;
2361	u64 tmp;
2362
2363	num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2364
2365	trans->block_rsv = rc->block_rsv;
2366	rc->reserved_bytes += num_bytes;
2367
2368	/*
2369	 * We are under a transaction here so we can only do limited flushing.
2370	 * If we get an enospc just kick back -EAGAIN so we know to drop the
2371	 * transaction and try to refill when we can flush all the things.
2372	 */
2373	ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv, num_bytes,
2374				     BTRFS_RESERVE_FLUSH_LIMIT);
2375	if (ret) {
2376		tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2377		while (tmp <= rc->reserved_bytes)
2378			tmp <<= 1;
2379		/*
2380		 * only one thread can access block_rsv at this point,
2381		 * so we don't need hold lock to protect block_rsv.
2382		 * we expand more reservation size here to allow enough
2383		 * space for relocation and we will return earlier in
2384		 * enospc case.
2385		 */
2386		rc->block_rsv->size = tmp + fs_info->nodesize *
2387				      RELOCATION_RESERVED_NODES;
2388		return -EAGAIN;
2389	}
2390
2391	return 0;
2392}
2393
2394/*
2395 * relocate a block tree, and then update pointers in upper level
2396 * blocks that reference the block to point to the new location.
2397 *
2398 * if called by link_to_upper, the block has already been relocated.
2399 * in that case this function just updates pointers.
2400 */
2401static int do_relocation(struct btrfs_trans_handle *trans,
2402			 struct reloc_control *rc,
2403			 struct btrfs_backref_node *node,
2404			 struct btrfs_key *key,
2405			 struct btrfs_path *path, int lowest)
2406{
2407	struct btrfs_backref_node *upper;
2408	struct btrfs_backref_edge *edge;
2409	struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2410	struct btrfs_root *root;
2411	struct extent_buffer *eb;
2412	u32 blocksize;
2413	u64 bytenr;
2414	int slot;
2415	int ret = 0;
2416
2417	/*
2418	 * If we are lowest then this is the first time we're processing this
2419	 * block, and thus shouldn't have an eb associated with it yet.
2420	 */
2421	ASSERT(!lowest || !node->eb);
2422
2423	path->lowest_level = node->level + 1;
2424	rc->backref_cache.path[node->level] = node;
2425	list_for_each_entry(edge, &node->upper, list[LOWER]) {
2426		struct btrfs_ref ref = { 0 };
2427
2428		cond_resched();
2429
2430		upper = edge->node[UPPER];
2431		root = select_reloc_root(trans, rc, upper, edges);
2432		if (IS_ERR(root)) {
2433			ret = PTR_ERR(root);
2434			goto next;
2435		}
2436
2437		if (upper->eb && !upper->locked) {
2438			if (!lowest) {
2439				ret = btrfs_bin_search(upper->eb, 0, key, &slot);
2440				if (ret < 0)
2441					goto next;
2442				BUG_ON(ret);
2443				bytenr = btrfs_node_blockptr(upper->eb, slot);
2444				if (node->eb->start == bytenr)
2445					goto next;
2446			}
2447			btrfs_backref_drop_node_buffer(upper);
2448		}
2449
2450		if (!upper->eb) {
2451			ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2452			if (ret) {
2453				if (ret > 0)
2454					ret = -ENOENT;
2455
2456				btrfs_release_path(path);
2457				break;
2458			}
2459
2460			if (!upper->eb) {
2461				upper->eb = path->nodes[upper->level];
2462				path->nodes[upper->level] = NULL;
2463			} else {
2464				BUG_ON(upper->eb != path->nodes[upper->level]);
2465			}
2466
2467			upper->locked = 1;
2468			path->locks[upper->level] = 0;
2469
2470			slot = path->slots[upper->level];
2471			btrfs_release_path(path);
2472		} else {
2473			ret = btrfs_bin_search(upper->eb, 0, key, &slot);
2474			if (ret < 0)
2475				goto next;
2476			BUG_ON(ret);
2477		}
2478
2479		bytenr = btrfs_node_blockptr(upper->eb, slot);
2480		if (lowest) {
2481			if (bytenr != node->bytenr) {
2482				btrfs_err(root->fs_info,
2483		"lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2484					  bytenr, node->bytenr, slot,
2485					  upper->eb->start);
2486				ret = -EIO;
2487				goto next;
2488			}
2489		} else {
2490			if (node->eb->start == bytenr)
2491				goto next;
2492		}
2493
2494		blocksize = root->fs_info->nodesize;
2495		eb = btrfs_read_node_slot(upper->eb, slot);
2496		if (IS_ERR(eb)) {
2497			ret = PTR_ERR(eb);
2498			goto next;
2499		}
2500		btrfs_tree_lock(eb);
2501
2502		if (!node->eb) {
2503			ret = btrfs_cow_block(trans, root, eb, upper->eb,
2504					      slot, &eb, BTRFS_NESTING_COW);
2505			btrfs_tree_unlock(eb);
2506			free_extent_buffer(eb);
2507			if (ret < 0)
2508				goto next;
2509			/*
2510			 * We've just COWed this block, it should have updated
2511			 * the correct backref node entry.
2512			 */
2513			ASSERT(node->eb == eb);
2514		} else {
2515			btrfs_set_node_blockptr(upper->eb, slot,
2516						node->eb->start);
2517			btrfs_set_node_ptr_generation(upper->eb, slot,
2518						      trans->transid);
2519			btrfs_mark_buffer_dirty(trans, upper->eb);
2520
2521			btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2522					       node->eb->start, blocksize,
2523					       upper->eb->start,
2524					       btrfs_header_owner(upper->eb));
2525			btrfs_init_tree_ref(&ref, node->level,
2526					    btrfs_header_owner(upper->eb),
2527					    root->root_key.objectid, false);
2528			ret = btrfs_inc_extent_ref(trans, &ref);
2529			if (!ret)
2530				ret = btrfs_drop_subtree(trans, root, eb,
2531							 upper->eb);
2532			if (ret)
2533				btrfs_abort_transaction(trans, ret);
2534		}
2535next:
2536		if (!upper->pending)
2537			btrfs_backref_drop_node_buffer(upper);
2538		else
2539			btrfs_backref_unlock_node_buffer(upper);
2540		if (ret)
2541			break;
2542	}
2543
2544	if (!ret && node->pending) {
2545		btrfs_backref_drop_node_buffer(node);
2546		list_move_tail(&node->list, &rc->backref_cache.changed);
2547		node->pending = 0;
2548	}
2549
2550	path->lowest_level = 0;
2551
2552	/*
2553	 * We should have allocated all of our space in the block rsv and thus
2554	 * shouldn't ENOSPC.
2555	 */
2556	ASSERT(ret != -ENOSPC);
2557	return ret;
2558}
2559
2560static int link_to_upper(struct btrfs_trans_handle *trans,
2561			 struct reloc_control *rc,
2562			 struct btrfs_backref_node *node,
2563			 struct btrfs_path *path)
2564{
2565	struct btrfs_key key;
2566
2567	btrfs_node_key_to_cpu(node->eb, &key, 0);
2568	return do_relocation(trans, rc, node, &key, path, 0);
2569}
2570
2571static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2572				struct reloc_control *rc,
2573				struct btrfs_path *path, int err)
2574{
2575	LIST_HEAD(list);
2576	struct btrfs_backref_cache *cache = &rc->backref_cache;
2577	struct btrfs_backref_node *node;
2578	int level;
2579	int ret;
2580
2581	for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2582		while (!list_empty(&cache->pending[level])) {
2583			node = list_entry(cache->pending[level].next,
2584					  struct btrfs_backref_node, list);
2585			list_move_tail(&node->list, &list);
2586			BUG_ON(!node->pending);
2587
2588			if (!err) {
2589				ret = link_to_upper(trans, rc, node, path);
2590				if (ret < 0)
2591					err = ret;
2592			}
2593		}
2594		list_splice_init(&list, &cache->pending[level]);
2595	}
2596	return err;
2597}
2598
2599/*
2600 * mark a block and all blocks directly/indirectly reference the block
2601 * as processed.
2602 */
2603static void update_processed_blocks(struct reloc_control *rc,
2604				    struct btrfs_backref_node *node)
2605{
2606	struct btrfs_backref_node *next = node;
2607	struct btrfs_backref_edge *edge;
2608	struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2609	int index = 0;
2610
2611	while (next) {
2612		cond_resched();
2613		while (1) {
2614			if (next->processed)
2615				break;
2616
2617			mark_block_processed(rc, next);
2618
2619			if (list_empty(&next->upper))
2620				break;
2621
2622			edge = list_entry(next->upper.next,
2623					struct btrfs_backref_edge, list[LOWER]);
2624			edges[index++] = edge;
2625			next = edge->node[UPPER];
2626		}
2627		next = walk_down_backref(edges, &index);
2628	}
2629}
2630
2631static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2632{
2633	u32 blocksize = rc->extent_root->fs_info->nodesize;
2634
2635	if (test_range_bit(&rc->processed_blocks, bytenr,
2636			   bytenr + blocksize - 1, EXTENT_DIRTY, NULL))
2637		return 1;
2638	return 0;
2639}
2640
2641static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2642			      struct tree_block *block)
2643{
2644	struct btrfs_tree_parent_check check = {
2645		.level = block->level,
2646		.owner_root = block->owner,
2647		.transid = block->key.offset
2648	};
2649	struct extent_buffer *eb;
2650
2651	eb = read_tree_block(fs_info, block->bytenr, &check);
2652	if (IS_ERR(eb))
2653		return PTR_ERR(eb);
2654	if (!extent_buffer_uptodate(eb)) {
2655		free_extent_buffer(eb);
2656		return -EIO;
2657	}
2658	if (block->level == 0)
2659		btrfs_item_key_to_cpu(eb, &block->key, 0);
2660	else
2661		btrfs_node_key_to_cpu(eb, &block->key, 0);
2662	free_extent_buffer(eb);
2663	block->key_ready = true;
2664	return 0;
2665}
2666
2667/*
2668 * helper function to relocate a tree block
2669 */
2670static int relocate_tree_block(struct btrfs_trans_handle *trans,
2671				struct reloc_control *rc,
2672				struct btrfs_backref_node *node,
2673				struct btrfs_key *key,
2674				struct btrfs_path *path)
2675{
2676	struct btrfs_root *root;
2677	int ret = 0;
2678
2679	if (!node)
2680		return 0;
2681
2682	/*
2683	 * If we fail here we want to drop our backref_node because we are going
2684	 * to start over and regenerate the tree for it.
2685	 */
2686	ret = reserve_metadata_space(trans, rc, node);
2687	if (ret)
2688		goto out;
2689
2690	BUG_ON(node->processed);
2691	root = select_one_root(node);
2692	if (IS_ERR(root)) {
2693		ret = PTR_ERR(root);
2694
2695		/* See explanation in select_one_root for the -EUCLEAN case. */
2696		ASSERT(ret == -ENOENT);
2697		if (ret == -ENOENT) {
2698			ret = 0;
2699			update_processed_blocks(rc, node);
2700		}
2701		goto out;
2702	}
2703
2704	if (root) {
2705		if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2706			/*
2707			 * This block was the root block of a root, and this is
2708			 * the first time we're processing the block and thus it
2709			 * should not have had the ->new_bytenr modified and
2710			 * should have not been included on the changed list.
2711			 *
2712			 * However in the case of corruption we could have
2713			 * multiple refs pointing to the same block improperly,
2714			 * and thus we would trip over these checks.  ASSERT()
2715			 * for the developer case, because it could indicate a
2716			 * bug in the backref code, however error out for a
2717			 * normal user in the case of corruption.
2718			 */
2719			ASSERT(node->new_bytenr == 0);
2720			ASSERT(list_empty(&node->list));
2721			if (node->new_bytenr || !list_empty(&node->list)) {
2722				btrfs_err(root->fs_info,
2723				  "bytenr %llu has improper references to it",
2724					  node->bytenr);
2725				ret = -EUCLEAN;
2726				goto out;
2727			}
2728			ret = btrfs_record_root_in_trans(trans, root);
2729			if (ret)
2730				goto out;
2731			/*
2732			 * Another thread could have failed, need to check if we
2733			 * have reloc_root actually set.
2734			 */
2735			if (!root->reloc_root) {
2736				ret = -ENOENT;
2737				goto out;
2738			}
2739			root = root->reloc_root;
2740			node->new_bytenr = root->node->start;
2741			btrfs_put_root(node->root);
2742			node->root = btrfs_grab_root(root);
2743			ASSERT(node->root);
2744			list_add_tail(&node->list, &rc->backref_cache.changed);
2745		} else {
2746			path->lowest_level = node->level;
2747			if (root == root->fs_info->chunk_root)
2748				btrfs_reserve_chunk_metadata(trans, false);
2749			ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2750			btrfs_release_path(path);
2751			if (root == root->fs_info->chunk_root)
2752				btrfs_trans_release_chunk_metadata(trans);
2753			if (ret > 0)
2754				ret = 0;
2755		}
2756		if (!ret)
2757			update_processed_blocks(rc, node);
2758	} else {
2759		ret = do_relocation(trans, rc, node, key, path, 1);
2760	}
2761out:
2762	if (ret || node->level == 0 || node->cowonly)
2763		btrfs_backref_cleanup_node(&rc->backref_cache, node);
2764	return ret;
2765}
2766
2767/*
2768 * relocate a list of blocks
2769 */
2770static noinline_for_stack
2771int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2772			 struct reloc_control *rc, struct rb_root *blocks)
2773{
2774	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2775	struct btrfs_backref_node *node;
2776	struct btrfs_path *path;
2777	struct tree_block *block;
2778	struct tree_block *next;
2779	int ret;
2780	int err = 0;
2781
2782	path = btrfs_alloc_path();
2783	if (!path) {
2784		err = -ENOMEM;
2785		goto out_free_blocks;
2786	}
2787
2788	/* Kick in readahead for tree blocks with missing keys */
2789	rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2790		if (!block->key_ready)
2791			btrfs_readahead_tree_block(fs_info, block->bytenr,
2792						   block->owner, 0,
2793						   block->level);
2794	}
2795
2796	/* Get first keys */
2797	rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2798		if (!block->key_ready) {
2799			err = get_tree_block_key(fs_info, block);
2800			if (err)
2801				goto out_free_path;
2802		}
2803	}
2804
2805	/* Do tree relocation */
2806	rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2807		node = build_backref_tree(trans, rc, &block->key,
2808					  block->level, block->bytenr);
2809		if (IS_ERR(node)) {
2810			err = PTR_ERR(node);
2811			goto out;
2812		}
2813
2814		ret = relocate_tree_block(trans, rc, node, &block->key,
2815					  path);
2816		if (ret < 0) {
2817			err = ret;
2818			break;
2819		}
2820	}
2821out:
2822	err = finish_pending_nodes(trans, rc, path, err);
2823
2824out_free_path:
2825	btrfs_free_path(path);
2826out_free_blocks:
2827	free_block_list(blocks);
2828	return err;
2829}
2830
2831static noinline_for_stack int prealloc_file_extent_cluster(
2832				struct btrfs_inode *inode,
2833				const struct file_extent_cluster *cluster)
2834{
2835	u64 alloc_hint = 0;
2836	u64 start;
2837	u64 end;
2838	u64 offset = inode->index_cnt;
2839	u64 num_bytes;
2840	int nr;
2841	int ret = 0;
2842	u64 i_size = i_size_read(&inode->vfs_inode);
2843	u64 prealloc_start = cluster->start - offset;
2844	u64 prealloc_end = cluster->end - offset;
2845	u64 cur_offset = prealloc_start;
2846
2847	/*
2848	 * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2849	 * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2850	 * btrfs_do_readpage() call of previously relocated file cluster.
2851	 *
2852	 * If the current cluster starts in the above range, btrfs_do_readpage()
2853	 * will skip the read, and relocate_one_page() will later writeback
2854	 * the padding zeros as new data, causing data corruption.
2855	 *
2856	 * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2857	 */
2858	if (!PAGE_ALIGNED(i_size)) {
2859		struct address_space *mapping = inode->vfs_inode.i_mapping;
2860		struct btrfs_fs_info *fs_info = inode->root->fs_info;
2861		const u32 sectorsize = fs_info->sectorsize;
2862		struct page *page;
2863
2864		ASSERT(sectorsize < PAGE_SIZE);
2865		ASSERT(IS_ALIGNED(i_size, sectorsize));
2866
2867		/*
2868		 * Subpage can't handle page with DIRTY but without UPTODATE
2869		 * bit as it can lead to the following deadlock:
2870		 *
2871		 * btrfs_read_folio()
2872		 * | Page already *locked*
2873		 * |- btrfs_lock_and_flush_ordered_range()
2874		 *    |- btrfs_start_ordered_extent()
2875		 *       |- extent_write_cache_pages()
2876		 *          |- lock_page()
2877		 *             We try to lock the page we already hold.
2878		 *
2879		 * Here we just writeback the whole data reloc inode, so that
2880		 * we will be ensured to have no dirty range in the page, and
2881		 * are safe to clear the uptodate bits.
2882		 *
2883		 * This shouldn't cause too much overhead, as we need to write
2884		 * the data back anyway.
2885		 */
2886		ret = filemap_write_and_wait(mapping);
2887		if (ret < 0)
2888			return ret;
2889
2890		clear_extent_bits(&inode->io_tree, i_size,
2891				  round_up(i_size, PAGE_SIZE) - 1,
2892				  EXTENT_UPTODATE);
2893		page = find_lock_page(mapping, i_size >> PAGE_SHIFT);
2894		/*
2895		 * If page is freed we don't need to do anything then, as we
2896		 * will re-read the whole page anyway.
2897		 */
2898		if (page) {
2899			btrfs_subpage_clear_uptodate(fs_info, page_folio(page), i_size,
2900					round_up(i_size, PAGE_SIZE) - i_size);
2901			unlock_page(page);
2902			put_page(page);
2903		}
2904	}
2905
2906	BUG_ON(cluster->start != cluster->boundary[0]);
2907	ret = btrfs_alloc_data_chunk_ondemand(inode,
2908					      prealloc_end + 1 - prealloc_start);
2909	if (ret)
2910		return ret;
2911
2912	btrfs_inode_lock(inode, 0);
2913	for (nr = 0; nr < cluster->nr; nr++) {
2914		struct extent_state *cached_state = NULL;
2915
2916		start = cluster->boundary[nr] - offset;
2917		if (nr + 1 < cluster->nr)
2918			end = cluster->boundary[nr + 1] - 1 - offset;
2919		else
2920			end = cluster->end - offset;
2921
2922		lock_extent(&inode->io_tree, start, end, &cached_state);
2923		num_bytes = end + 1 - start;
2924		ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2925						num_bytes, num_bytes,
2926						end + 1, &alloc_hint);
2927		cur_offset = end + 1;
2928		unlock_extent(&inode->io_tree, start, end, &cached_state);
2929		if (ret)
2930			break;
2931	}
2932	btrfs_inode_unlock(inode, 0);
2933
2934	if (cur_offset < prealloc_end)
2935		btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2936					       prealloc_end + 1 - cur_offset);
2937	return ret;
2938}
2939
2940static noinline_for_stack int setup_relocation_extent_mapping(struct inode *inode,
2941				u64 start, u64 end, u64 block_start)
2942{
2943	struct extent_map *em;
2944	struct extent_state *cached_state = NULL;
2945	int ret = 0;
2946
2947	em = alloc_extent_map();
2948	if (!em)
2949		return -ENOMEM;
2950
2951	em->start = start;
2952	em->len = end + 1 - start;
2953	em->block_len = em->len;
2954	em->block_start = block_start;
2955	em->flags |= EXTENT_FLAG_PINNED;
2956
2957	lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2958	ret = btrfs_replace_extent_map_range(BTRFS_I(inode), em, false);
2959	unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2960	free_extent_map(em);
2961
2962	return ret;
2963}
2964
2965/*
2966 * Allow error injection to test balance/relocation cancellation
2967 */
2968noinline int btrfs_should_cancel_balance(const struct btrfs_fs_info *fs_info)
2969{
2970	return atomic_read(&fs_info->balance_cancel_req) ||
2971		atomic_read(&fs_info->reloc_cancel_req) ||
2972		fatal_signal_pending(current);
2973}
2974ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2975
2976static u64 get_cluster_boundary_end(const struct file_extent_cluster *cluster,
2977				    int cluster_nr)
2978{
2979	/* Last extent, use cluster end directly */
2980	if (cluster_nr >= cluster->nr - 1)
2981		return cluster->end;
2982
2983	/* Use next boundary start*/
2984	return cluster->boundary[cluster_nr + 1] - 1;
2985}
2986
2987static int relocate_one_page(struct inode *inode, struct file_ra_state *ra,
2988			     const struct file_extent_cluster *cluster,
2989			     int *cluster_nr, unsigned long page_index)
2990{
2991	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2992	u64 offset = BTRFS_I(inode)->index_cnt;
2993	const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2994	gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2995	struct page *page;
2996	u64 page_start;
2997	u64 page_end;
2998	u64 cur;
2999	int ret;
3000
3001	ASSERT(page_index <= last_index);
3002	page = find_lock_page(inode->i_mapping, page_index);
3003	if (!page) {
3004		page_cache_sync_readahead(inode->i_mapping, ra, NULL,
3005				page_index, last_index + 1 - page_index);
3006		page = find_or_create_page(inode->i_mapping, page_index, mask);
3007		if (!page)
3008			return -ENOMEM;
3009	}
3010
3011	if (PageReadahead(page))
3012		page_cache_async_readahead(inode->i_mapping, ra, NULL,
3013				page_folio(page), page_index,
3014				last_index + 1 - page_index);
3015
3016	if (!PageUptodate(page)) {
3017		btrfs_read_folio(NULL, page_folio(page));
3018		lock_page(page);
3019		if (!PageUptodate(page)) {
3020			ret = -EIO;
3021			goto release_page;
3022		}
3023	}
3024
3025	/*
3026	 * We could have lost page private when we dropped the lock to read the
3027	 * page above, make sure we set_page_extent_mapped here so we have any
3028	 * of the subpage blocksize stuff we need in place.
3029	 */
3030	ret = set_page_extent_mapped(page);
3031	if (ret < 0)
3032		goto release_page;
3033
3034	page_start = page_offset(page);
3035	page_end = page_start + PAGE_SIZE - 1;
3036
3037	/*
3038	 * Start from the cluster, as for subpage case, the cluster can start
3039	 * inside the page.
3040	 */
3041	cur = max(page_start, cluster->boundary[*cluster_nr] - offset);
3042	while (cur <= page_end) {
3043		struct extent_state *cached_state = NULL;
3044		u64 extent_start = cluster->boundary[*cluster_nr] - offset;
3045		u64 extent_end = get_cluster_boundary_end(cluster,
3046						*cluster_nr) - offset;
3047		u64 clamped_start = max(page_start, extent_start);
3048		u64 clamped_end = min(page_end, extent_end);
3049		u32 clamped_len = clamped_end + 1 - clamped_start;
3050
3051		/* Reserve metadata for this range */
3052		ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3053						      clamped_len, clamped_len,
3054						      false);
3055		if (ret)
3056			goto release_page;
3057
3058		/* Mark the range delalloc and dirty for later writeback */
3059		lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3060			    &cached_state);
3061		ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
3062						clamped_end, 0, &cached_state);
3063		if (ret) {
3064			clear_extent_bit(&BTRFS_I(inode)->io_tree,
3065					 clamped_start, clamped_end,
3066					 EXTENT_LOCKED | EXTENT_BOUNDARY,
3067					 &cached_state);
3068			btrfs_delalloc_release_metadata(BTRFS_I(inode),
3069							clamped_len, true);
3070			btrfs_delalloc_release_extents(BTRFS_I(inode),
3071						       clamped_len);
3072			goto release_page;
3073		}
3074		btrfs_folio_set_dirty(fs_info, page_folio(page),
3075				      clamped_start, clamped_len);
3076
3077		/*
3078		 * Set the boundary if it's inside the page.
3079		 * Data relocation requires the destination extents to have the
3080		 * same size as the source.
3081		 * EXTENT_BOUNDARY bit prevents current extent from being merged
3082		 * with previous extent.
3083		 */
3084		if (in_range(cluster->boundary[*cluster_nr] - offset,
3085			     page_start, PAGE_SIZE)) {
3086			u64 boundary_start = cluster->boundary[*cluster_nr] -
3087						offset;
3088			u64 boundary_end = boundary_start +
3089					   fs_info->sectorsize - 1;
3090
3091			set_extent_bit(&BTRFS_I(inode)->io_tree,
3092				       boundary_start, boundary_end,
3093				       EXTENT_BOUNDARY, NULL);
3094		}
3095		unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3096			      &cached_state);
3097		btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3098		cur += clamped_len;
3099
3100		/* Crossed extent end, go to next extent */
3101		if (cur >= extent_end) {
3102			(*cluster_nr)++;
3103			/* Just finished the last extent of the cluster, exit. */
3104			if (*cluster_nr >= cluster->nr)
3105				break;
3106		}
3107	}
3108	unlock_page(page);
3109	put_page(page);
3110
3111	balance_dirty_pages_ratelimited(inode->i_mapping);
3112	btrfs_throttle(fs_info);
3113	if (btrfs_should_cancel_balance(fs_info))
3114		ret = -ECANCELED;
3115	return ret;
3116
3117release_page:
3118	unlock_page(page);
3119	put_page(page);
3120	return ret;
3121}
3122
3123static int relocate_file_extent_cluster(struct inode *inode,
3124					const struct file_extent_cluster *cluster)
3125{
3126	u64 offset = BTRFS_I(inode)->index_cnt;
3127	unsigned long index;
3128	unsigned long last_index;
3129	struct file_ra_state *ra;
3130	int cluster_nr = 0;
3131	int ret = 0;
3132
3133	if (!cluster->nr)
3134		return 0;
3135
3136	ra = kzalloc(sizeof(*ra), GFP_NOFS);
3137	if (!ra)
3138		return -ENOMEM;
3139
3140	ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
3141	if (ret)
3142		goto out;
3143
3144	file_ra_state_init(ra, inode->i_mapping);
3145
3146	ret = setup_relocation_extent_mapping(inode, cluster->start - offset,
3147				   cluster->end - offset, cluster->start);
3148	if (ret)
3149		goto out;
3150
3151	last_index = (cluster->end - offset) >> PAGE_SHIFT;
3152	for (index = (cluster->start - offset) >> PAGE_SHIFT;
3153	     index <= last_index && !ret; index++)
3154		ret = relocate_one_page(inode, ra, cluster, &cluster_nr, index);
3155	if (ret == 0)
3156		WARN_ON(cluster_nr != cluster->nr);
3157out:
3158	kfree(ra);
3159	return ret;
3160}
3161
3162static noinline_for_stack int relocate_data_extent(struct inode *inode,
3163				const struct btrfs_key *extent_key,
3164				struct file_extent_cluster *cluster)
3165{
3166	int ret;
3167	struct btrfs_root *root = BTRFS_I(inode)->root;
3168
3169	if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3170		ret = relocate_file_extent_cluster(inode, cluster);
3171		if (ret)
3172			return ret;
3173		cluster->nr = 0;
3174	}
3175
3176	/*
3177	 * Under simple quotas, we set root->relocation_src_root when we find
3178	 * the extent. If adjacent extents have different owners, we can't merge
3179	 * them while relocating. Handle this by storing the owning root that
3180	 * started a cluster and if we see an extent from a different root break
3181	 * cluster formation (just like the above case of non-adjacent extents).
3182	 *
3183	 * Without simple quotas, relocation_src_root is always 0, so we should
3184	 * never see a mismatch, and it should have no effect on relocation
3185	 * clusters.
3186	 */
3187	if (cluster->nr > 0 && cluster->owning_root != root->relocation_src_root) {
3188		u64 tmp = root->relocation_src_root;
3189
3190		/*
3191		 * root->relocation_src_root is the state that actually affects
3192		 * the preallocation we do here, so set it to the root owning
3193		 * the cluster we need to relocate.
3194		 */
3195		root->relocation_src_root = cluster->owning_root;
3196		ret = relocate_file_extent_cluster(inode, cluster);
3197		if (ret)
3198			return ret;
3199		cluster->nr = 0;
3200		/* And reset it back for the current extent's owning root. */
3201		root->relocation_src_root = tmp;
3202	}
3203
3204	if (!cluster->nr) {
3205		cluster->start = extent_key->objectid;
3206		cluster->owning_root = root->relocation_src_root;
3207	}
3208	else
3209		BUG_ON(cluster->nr >= MAX_EXTENTS);
3210	cluster->end = extent_key->objectid + extent_key->offset - 1;
3211	cluster->boundary[cluster->nr] = extent_key->objectid;
3212	cluster->nr++;
3213
3214	if (cluster->nr >= MAX_EXTENTS) {
3215		ret = relocate_file_extent_cluster(inode, cluster);
3216		if (ret)
3217			return ret;
3218		cluster->nr = 0;
3219	}
3220	return 0;
3221}
3222
3223/*
3224 * helper to add a tree block to the list.
3225 * the major work is getting the generation and level of the block
3226 */
3227static int add_tree_block(struct reloc_control *rc,
3228			  const struct btrfs_key *extent_key,
3229			  struct btrfs_path *path,
3230			  struct rb_root *blocks)
3231{
3232	struct extent_buffer *eb;
3233	struct btrfs_extent_item *ei;
3234	struct btrfs_tree_block_info *bi;
3235	struct tree_block *block;
3236	struct rb_node *rb_node;
3237	u32 item_size;
3238	int level = -1;
3239	u64 generation;
3240	u64 owner = 0;
3241
3242	eb =  path->nodes[0];
3243	item_size = btrfs_item_size(eb, path->slots[0]);
3244
3245	if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3246	    item_size >= sizeof(*ei) + sizeof(*bi)) {
3247		unsigned long ptr = 0, end;
3248
3249		ei = btrfs_item_ptr(eb, path->slots[0],
3250				struct btrfs_extent_item);
3251		end = (unsigned long)ei + item_size;
3252		if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3253			bi = (struct btrfs_tree_block_info *)(ei + 1);
3254			level = btrfs_tree_block_level(eb, bi);
3255			ptr = (unsigned long)(bi + 1);
3256		} else {
3257			level = (int)extent_key->offset;
3258			ptr = (unsigned long)(ei + 1);
3259		}
3260		generation = btrfs_extent_generation(eb, ei);
3261
3262		/*
3263		 * We're reading random blocks without knowing their owner ahead
3264		 * of time.  This is ok most of the time, as all reloc roots and
3265		 * fs roots have the same lock type.  However normal trees do
3266		 * not, and the only way to know ahead of time is to read the
3267		 * inline ref offset.  We know it's an fs root if
3268		 *
3269		 * 1. There's more than one ref.
3270		 * 2. There's a SHARED_DATA_REF_KEY set.
3271		 * 3. FULL_BACKREF is set on the flags.
3272		 *
3273		 * Otherwise it's safe to assume that the ref offset == the
3274		 * owner of this block, so we can use that when calling
3275		 * read_tree_block.
3276		 */
3277		if (btrfs_extent_refs(eb, ei) == 1 &&
3278		    !(btrfs_extent_flags(eb, ei) &
3279		      BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3280		    ptr < end) {
3281			struct btrfs_extent_inline_ref *iref;
3282			int type;
3283
3284			iref = (struct btrfs_extent_inline_ref *)ptr;
3285			type = btrfs_get_extent_inline_ref_type(eb, iref,
3286							BTRFS_REF_TYPE_BLOCK);
3287			if (type == BTRFS_REF_TYPE_INVALID)
3288				return -EINVAL;
3289			if (type == BTRFS_TREE_BLOCK_REF_KEY)
3290				owner = btrfs_extent_inline_ref_offset(eb, iref);
3291		}
3292	} else {
3293		btrfs_print_leaf(eb);
3294		btrfs_err(rc->block_group->fs_info,
3295			  "unrecognized tree backref at tree block %llu slot %u",
3296			  eb->start, path->slots[0]);
3297		btrfs_release_path(path);
3298		return -EUCLEAN;
3299	}
3300
3301	btrfs_release_path(path);
3302
3303	BUG_ON(level == -1);
3304
3305	block = kmalloc(sizeof(*block), GFP_NOFS);
3306	if (!block)
3307		return -ENOMEM;
3308
3309	block->bytenr = extent_key->objectid;
3310	block->key.objectid = rc->extent_root->fs_info->nodesize;
3311	block->key.offset = generation;
3312	block->level = level;
3313	block->key_ready = false;
3314	block->owner = owner;
3315
3316	rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3317	if (rb_node)
3318		btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3319				    -EEXIST);
3320
3321	return 0;
3322}
3323
3324/*
3325 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3326 */
3327static int __add_tree_block(struct reloc_control *rc,
3328			    u64 bytenr, u32 blocksize,
3329			    struct rb_root *blocks)
3330{
3331	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3332	struct btrfs_path *path;
3333	struct btrfs_key key;
3334	int ret;
3335	bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3336
3337	if (tree_block_processed(bytenr, rc))
3338		return 0;
3339
3340	if (rb_simple_search(blocks, bytenr))
3341		return 0;
3342
3343	path = btrfs_alloc_path();
3344	if (!path)
3345		return -ENOMEM;
3346again:
3347	key.objectid = bytenr;
3348	if (skinny) {
3349		key.type = BTRFS_METADATA_ITEM_KEY;
3350		key.offset = (u64)-1;
3351	} else {
3352		key.type = BTRFS_EXTENT_ITEM_KEY;
3353		key.offset = blocksize;
3354	}
3355
3356	path->search_commit_root = 1;
3357	path->skip_locking = 1;
3358	ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3359	if (ret < 0)
3360		goto out;
3361
3362	if (ret > 0 && skinny) {
3363		if (path->slots[0]) {
3364			path->slots[0]--;
3365			btrfs_item_key_to_cpu(path->nodes[0], &key,
3366					      path->slots[0]);
3367			if (key.objectid == bytenr &&
3368			    (key.type == BTRFS_METADATA_ITEM_KEY ||
3369			     (key.type == BTRFS_EXTENT_ITEM_KEY &&
3370			      key.offset == blocksize)))
3371				ret = 0;
3372		}
3373
3374		if (ret) {
3375			skinny = false;
3376			btrfs_release_path(path);
3377			goto again;
3378		}
3379	}
3380	if (ret) {
3381		ASSERT(ret == 1);
3382		btrfs_print_leaf(path->nodes[0]);
3383		btrfs_err(fs_info,
3384	     "tree block extent item (%llu) is not found in extent tree",
3385		     bytenr);
3386		WARN_ON(1);
3387		ret = -EINVAL;
3388		goto out;
3389	}
3390
3391	ret = add_tree_block(rc, &key, path, blocks);
3392out:
3393	btrfs_free_path(path);
3394	return ret;
3395}
3396
3397static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3398				    struct btrfs_block_group *block_group,
3399				    struct inode *inode,
3400				    u64 ino)
3401{
3402	struct btrfs_root *root = fs_info->tree_root;
3403	struct btrfs_trans_handle *trans;
3404	int ret = 0;
3405
3406	if (inode)
3407		goto truncate;
3408
3409	inode = btrfs_iget(fs_info->sb, ino, root);
3410	if (IS_ERR(inode))
3411		return -ENOENT;
3412
3413truncate:
3414	ret = btrfs_check_trunc_cache_free_space(fs_info,
3415						 &fs_info->global_block_rsv);
3416	if (ret)
3417		goto out;
3418
3419	trans = btrfs_join_transaction(root);
3420	if (IS_ERR(trans)) {
3421		ret = PTR_ERR(trans);
3422		goto out;
3423	}
3424
3425	ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3426
3427	btrfs_end_transaction(trans);
3428	btrfs_btree_balance_dirty(fs_info);
3429out:
3430	iput(inode);
3431	return ret;
3432}
3433
3434/*
3435 * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3436 * cache inode, to avoid free space cache data extent blocking data relocation.
3437 */
3438static int delete_v1_space_cache(struct extent_buffer *leaf,
3439				 struct btrfs_block_group *block_group,
3440				 u64 data_bytenr)
3441{
3442	u64 space_cache_ino;
3443	struct btrfs_file_extent_item *ei;
3444	struct btrfs_key key;
3445	bool found = false;
3446	int i;
3447	int ret;
3448
3449	if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3450		return 0;
3451
3452	for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3453		u8 type;
3454
3455		btrfs_item_key_to_cpu(leaf, &key, i);
3456		if (key.type != BTRFS_EXTENT_DATA_KEY)
3457			continue;
3458		ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3459		type = btrfs_file_extent_type(leaf, ei);
3460
3461		if ((type == BTRFS_FILE_EXTENT_REG ||
3462		     type == BTRFS_FILE_EXTENT_PREALLOC) &&
3463		    btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3464			found = true;
3465			space_cache_ino = key.objectid;
3466			break;
3467		}
3468	}
3469	if (!found)
3470		return -ENOENT;
3471	ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3472					space_cache_ino);
3473	return ret;
3474}
3475
3476/*
3477 * helper to find all tree blocks that reference a given data extent
3478 */
3479static noinline_for_stack int add_data_references(struct reloc_control *rc,
3480						  const struct btrfs_key *extent_key,
3481						  struct btrfs_path *path,
3482						  struct rb_root *blocks)
3483{
3484	struct btrfs_backref_walk_ctx ctx = { 0 };
3485	struct ulist_iterator leaf_uiter;
3486	struct ulist_node *ref_node = NULL;
3487	const u32 blocksize = rc->extent_root->fs_info->nodesize;
3488	int ret = 0;
3489
3490	btrfs_release_path(path);
3491
3492	ctx.bytenr = extent_key->objectid;
3493	ctx.skip_inode_ref_list = true;
3494	ctx.fs_info = rc->extent_root->fs_info;
3495
3496	ret = btrfs_find_all_leafs(&ctx);
3497	if (ret < 0)
3498		return ret;
3499
3500	ULIST_ITER_INIT(&leaf_uiter);
3501	while ((ref_node = ulist_next(ctx.refs, &leaf_uiter))) {
3502		struct btrfs_tree_parent_check check = { 0 };
3503		struct extent_buffer *eb;
3504
3505		eb = read_tree_block(ctx.fs_info, ref_node->val, &check);
3506		if (IS_ERR(eb)) {
3507			ret = PTR_ERR(eb);
3508			break;
3509		}
3510		ret = delete_v1_space_cache(eb, rc->block_group,
3511					    extent_key->objectid);
3512		free_extent_buffer(eb);
3513		if (ret < 0)
3514			break;
3515		ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3516		if (ret < 0)
3517			break;
3518	}
3519	if (ret < 0)
3520		free_block_list(blocks);
3521	ulist_free(ctx.refs);
3522	return ret;
3523}
3524
3525/*
3526 * helper to find next unprocessed extent
3527 */
3528static noinline_for_stack
3529int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3530		     struct btrfs_key *extent_key)
3531{
3532	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3533	struct btrfs_key key;
3534	struct extent_buffer *leaf;
3535	u64 start, end, last;
3536	int ret;
3537
3538	last = rc->block_group->start + rc->block_group->length;
3539	while (1) {
3540		bool block_found;
3541
3542		cond_resched();
3543		if (rc->search_start >= last) {
3544			ret = 1;
3545			break;
3546		}
3547
3548		key.objectid = rc->search_start;
3549		key.type = BTRFS_EXTENT_ITEM_KEY;
3550		key.offset = 0;
3551
3552		path->search_commit_root = 1;
3553		path->skip_locking = 1;
3554		ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3555					0, 0);
3556		if (ret < 0)
3557			break;
3558next:
3559		leaf = path->nodes[0];
3560		if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3561			ret = btrfs_next_leaf(rc->extent_root, path);
3562			if (ret != 0)
3563				break;
3564			leaf = path->nodes[0];
3565		}
3566
3567		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3568		if (key.objectid >= last) {
3569			ret = 1;
3570			break;
3571		}
3572
3573		if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3574		    key.type != BTRFS_METADATA_ITEM_KEY) {
3575			path->slots[0]++;
3576			goto next;
3577		}
3578
3579		if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3580		    key.objectid + key.offset <= rc->search_start) {
3581			path->slots[0]++;
3582			goto next;
3583		}
3584
3585		if (key.type == BTRFS_METADATA_ITEM_KEY &&
3586		    key.objectid + fs_info->nodesize <=
3587		    rc->search_start) {
3588			path->slots[0]++;
3589			goto next;
3590		}
3591
3592		block_found = find_first_extent_bit(&rc->processed_blocks,
3593						    key.objectid, &start, &end,
3594						    EXTENT_DIRTY, NULL);
3595
3596		if (block_found && start <= key.objectid) {
3597			btrfs_release_path(path);
3598			rc->search_start = end + 1;
3599		} else {
3600			if (key.type == BTRFS_EXTENT_ITEM_KEY)
3601				rc->search_start = key.objectid + key.offset;
3602			else
3603				rc->search_start = key.objectid +
3604					fs_info->nodesize;
3605			memcpy(extent_key, &key, sizeof(key));
3606			return 0;
3607		}
3608	}
3609	btrfs_release_path(path);
3610	return ret;
3611}
3612
3613static void set_reloc_control(struct reloc_control *rc)
3614{
3615	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3616
3617	mutex_lock(&fs_info->reloc_mutex);
3618	fs_info->reloc_ctl = rc;
3619	mutex_unlock(&fs_info->reloc_mutex);
3620}
3621
3622static void unset_reloc_control(struct reloc_control *rc)
3623{
3624	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3625
3626	mutex_lock(&fs_info->reloc_mutex);
3627	fs_info->reloc_ctl = NULL;
3628	mutex_unlock(&fs_info->reloc_mutex);
3629}
3630
3631static noinline_for_stack
3632int prepare_to_relocate(struct reloc_control *rc)
3633{
3634	struct btrfs_trans_handle *trans;
3635	int ret;
3636
3637	rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3638					      BTRFS_BLOCK_RSV_TEMP);
3639	if (!rc->block_rsv)
3640		return -ENOMEM;
3641
3642	memset(&rc->cluster, 0, sizeof(rc->cluster));
3643	rc->search_start = rc->block_group->start;
3644	rc->extents_found = 0;
3645	rc->nodes_relocated = 0;
3646	rc->merging_rsv_size = 0;
3647	rc->reserved_bytes = 0;
3648	rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3649			      RELOCATION_RESERVED_NODES;
3650	ret = btrfs_block_rsv_refill(rc->extent_root->fs_info,
3651				     rc->block_rsv, rc->block_rsv->size,
3652				     BTRFS_RESERVE_FLUSH_ALL);
3653	if (ret)
3654		return ret;
3655
3656	rc->create_reloc_tree = true;
3657	set_reloc_control(rc);
3658
3659	trans = btrfs_join_transaction(rc->extent_root);
3660	if (IS_ERR(trans)) {
3661		unset_reloc_control(rc);
3662		/*
3663		 * extent tree is not a ref_cow tree and has no reloc_root to
3664		 * cleanup.  And callers are responsible to free the above
3665		 * block rsv.
3666		 */
3667		return PTR_ERR(trans);
3668	}
3669
3670	ret = btrfs_commit_transaction(trans);
3671	if (ret)
3672		unset_reloc_control(rc);
3673
3674	return ret;
3675}
3676
3677static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3678{
3679	struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3680	struct rb_root blocks = RB_ROOT;
3681	struct btrfs_key key;
3682	struct btrfs_trans_handle *trans = NULL;
3683	struct btrfs_path *path;
3684	struct btrfs_extent_item *ei;
3685	u64 flags;
3686	int ret;
3687	int err = 0;
3688	int progress = 0;
3689
3690	path = btrfs_alloc_path();
3691	if (!path)
3692		return -ENOMEM;
3693	path->reada = READA_FORWARD;
3694
3695	ret = prepare_to_relocate(rc);
3696	if (ret) {
3697		err = ret;
3698		goto out_free;
3699	}
3700
3701	while (1) {
3702		rc->reserved_bytes = 0;
3703		ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
3704					     rc->block_rsv->size,
3705					     BTRFS_RESERVE_FLUSH_ALL);
3706		if (ret) {
3707			err = ret;
3708			break;
3709		}
3710		progress++;
3711		trans = btrfs_start_transaction(rc->extent_root, 0);
3712		if (IS_ERR(trans)) {
3713			err = PTR_ERR(trans);
3714			trans = NULL;
3715			break;
3716		}
3717restart:
3718		if (update_backref_cache(trans, &rc->backref_cache)) {
3719			btrfs_end_transaction(trans);
3720			trans = NULL;
3721			continue;
3722		}
3723
3724		ret = find_next_extent(rc, path, &key);
3725		if (ret < 0)
3726			err = ret;
3727		if (ret != 0)
3728			break;
3729
3730		rc->extents_found++;
3731
3732		ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3733				    struct btrfs_extent_item);
3734		flags = btrfs_extent_flags(path->nodes[0], ei);
3735
3736		/*
3737		 * If we are relocating a simple quota owned extent item, we
3738		 * need to note the owner on the reloc data root so that when
3739		 * we allocate the replacement item, we can attribute it to the
3740		 * correct eventual owner (rather than the reloc data root).
3741		 */
3742		if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) {
3743			struct btrfs_root *root = BTRFS_I(rc->data_inode)->root;
3744			u64 owning_root_id = btrfs_get_extent_owner_root(fs_info,
3745								 path->nodes[0],
3746								 path->slots[0]);
3747
3748			root->relocation_src_root = owning_root_id;
3749		}
3750
3751		if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3752			ret = add_tree_block(rc, &key, path, &blocks);
3753		} else if (rc->stage == UPDATE_DATA_PTRS &&
3754			   (flags & BTRFS_EXTENT_FLAG_DATA)) {
3755			ret = add_data_references(rc, &key, path, &blocks);
3756		} else {
3757			btrfs_release_path(path);
3758			ret = 0;
3759		}
3760		if (ret < 0) {
3761			err = ret;
3762			break;
3763		}
3764
3765		if (!RB_EMPTY_ROOT(&blocks)) {
3766			ret = relocate_tree_blocks(trans, rc, &blocks);
3767			if (ret < 0) {
3768				if (ret != -EAGAIN) {
3769					err = ret;
3770					break;
3771				}
3772				rc->extents_found--;
3773				rc->search_start = key.objectid;
3774			}
3775		}
3776
3777		btrfs_end_transaction_throttle(trans);
3778		btrfs_btree_balance_dirty(fs_info);
3779		trans = NULL;
3780
3781		if (rc->stage == MOVE_DATA_EXTENTS &&
3782		    (flags & BTRFS_EXTENT_FLAG_DATA)) {
3783			rc->found_file_extent = true;
3784			ret = relocate_data_extent(rc->data_inode,
3785						   &key, &rc->cluster);
3786			if (ret < 0) {
3787				err = ret;
3788				break;
3789			}
3790		}
3791		if (btrfs_should_cancel_balance(fs_info)) {
3792			err = -ECANCELED;
3793			break;
3794		}
3795	}
3796	if (trans && progress && err == -ENOSPC) {
3797		ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3798		if (ret == 1) {
3799			err = 0;
3800			progress = 0;
3801			goto restart;
3802		}
3803	}
3804
3805	btrfs_release_path(path);
3806	clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3807
3808	if (trans) {
3809		btrfs_end_transaction_throttle(trans);
3810		btrfs_btree_balance_dirty(fs_info);
3811	}
3812
3813	if (!err) {
3814		ret = relocate_file_extent_cluster(rc->data_inode,
3815						   &rc->cluster);
3816		if (ret < 0)
3817			err = ret;
3818	}
3819
3820	rc->create_reloc_tree = false;
3821	set_reloc_control(rc);
3822
3823	btrfs_backref_release_cache(&rc->backref_cache);
3824	btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3825
3826	/*
3827	 * Even in the case when the relocation is cancelled, we should all go
3828	 * through prepare_to_merge() and merge_reloc_roots().
3829	 *
3830	 * For error (including cancelled balance), prepare_to_merge() will
3831	 * mark all reloc trees orphan, then queue them for cleanup in
3832	 * merge_reloc_roots()
3833	 */
3834	err = prepare_to_merge(rc, err);
3835
3836	merge_reloc_roots(rc);
3837
3838	rc->merge_reloc_tree = false;
3839	unset_reloc_control(rc);
3840	btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3841
3842	/* get rid of pinned extents */
3843	trans = btrfs_join_transaction(rc->extent_root);
3844	if (IS_ERR(trans)) {
3845		err = PTR_ERR(trans);
3846		goto out_free;
3847	}
3848	ret = btrfs_commit_transaction(trans);
3849	if (ret && !err)
3850		err = ret;
3851out_free:
3852	ret = clean_dirty_subvols(rc);
3853	if (ret < 0 && !err)
3854		err = ret;
3855	btrfs_free_block_rsv(fs_info, rc->block_rsv);
3856	btrfs_free_path(path);
3857	return err;
3858}
3859
3860static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3861				 struct btrfs_root *root, u64 objectid)
3862{
3863	struct btrfs_path *path;
3864	struct btrfs_inode_item *item;
3865	struct extent_buffer *leaf;
3866	int ret;
3867
3868	path = btrfs_alloc_path();
3869	if (!path)
3870		return -ENOMEM;
3871
3872	ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3873	if (ret)
3874		goto out;
3875
3876	leaf = path->nodes[0];
3877	item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3878	memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3879	btrfs_set_inode_generation(leaf, item, 1);
3880	btrfs_set_inode_size(leaf, item, 0);
3881	btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3882	btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3883					  BTRFS_INODE_PREALLOC);
3884	btrfs_mark_buffer_dirty(trans, leaf);
3885out:
3886	btrfs_free_path(path);
3887	return ret;
3888}
3889
3890static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3891				struct btrfs_root *root, u64 objectid)
3892{
3893	struct btrfs_path *path;
3894	struct btrfs_key key;
3895	int ret = 0;
3896
3897	path = btrfs_alloc_path();
3898	if (!path) {
3899		ret = -ENOMEM;
3900		goto out;
3901	}
3902
3903	key.objectid = objectid;
3904	key.type = BTRFS_INODE_ITEM_KEY;
3905	key.offset = 0;
3906	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3907	if (ret) {
3908		if (ret > 0)
3909			ret = -ENOENT;
3910		goto out;
3911	}
3912	ret = btrfs_del_item(trans, root, path);
3913out:
3914	if (ret)
3915		btrfs_abort_transaction(trans, ret);
3916	btrfs_free_path(path);
3917}
3918
3919/*
3920 * helper to create inode for data relocation.
3921 * the inode is in data relocation tree and its link count is 0
3922 */
3923static noinline_for_stack struct inode *create_reloc_inode(
3924					struct btrfs_fs_info *fs_info,
3925					const struct btrfs_block_group *group)
3926{
3927	struct inode *inode = NULL;
3928	struct btrfs_trans_handle *trans;
3929	struct btrfs_root *root;
3930	u64 objectid;
3931	int err = 0;
3932
3933	root = btrfs_grab_root(fs_info->data_reloc_root);
3934	trans = btrfs_start_transaction(root, 6);
3935	if (IS_ERR(trans)) {
3936		btrfs_put_root(root);
3937		return ERR_CAST(trans);
3938	}
3939
3940	err = btrfs_get_free_objectid(root, &objectid);
3941	if (err)
3942		goto out;
3943
3944	err = __insert_orphan_inode(trans, root, objectid);
3945	if (err)
3946		goto out;
3947
3948	inode = btrfs_iget(fs_info->sb, objectid, root);
3949	if (IS_ERR(inode)) {
3950		delete_orphan_inode(trans, root, objectid);
3951		err = PTR_ERR(inode);
3952		inode = NULL;
3953		goto out;
3954	}
3955	BTRFS_I(inode)->index_cnt = group->start;
3956
3957	err = btrfs_orphan_add(trans, BTRFS_I(inode));
3958out:
3959	btrfs_put_root(root);
3960	btrfs_end_transaction(trans);
3961	btrfs_btree_balance_dirty(fs_info);
3962	if (err) {
3963		iput(inode);
3964		inode = ERR_PTR(err);
3965	}
3966	return inode;
3967}
3968
3969/*
3970 * Mark start of chunk relocation that is cancellable. Check if the cancellation
3971 * has been requested meanwhile and don't start in that case.
3972 *
3973 * Return:
3974 *   0             success
3975 *   -EINPROGRESS  operation is already in progress, that's probably a bug
3976 *   -ECANCELED    cancellation request was set before the operation started
3977 */
3978static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3979{
3980	if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3981		/* This should not happen */
3982		btrfs_err(fs_info, "reloc already running, cannot start");
3983		return -EINPROGRESS;
3984	}
3985
3986	if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3987		btrfs_info(fs_info, "chunk relocation canceled on start");
3988		/*
3989		 * On cancel, clear all requests but let the caller mark
3990		 * the end after cleanup operations.
3991		 */
3992		atomic_set(&fs_info->reloc_cancel_req, 0);
3993		return -ECANCELED;
3994	}
3995	return 0;
3996}
3997
3998/*
3999 * Mark end of chunk relocation that is cancellable and wake any waiters.
4000 */
4001static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
4002{
4003	/* Requested after start, clear bit first so any waiters can continue */
4004	if (atomic_read(&fs_info->reloc_cancel_req) > 0)
4005		btrfs_info(fs_info, "chunk relocation canceled during operation");
4006	clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
4007	atomic_set(&fs_info->reloc_cancel_req, 0);
4008}
4009
4010static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4011{
4012	struct reloc_control *rc;
4013
4014	rc = kzalloc(sizeof(*rc), GFP_NOFS);
4015	if (!rc)
4016		return NULL;
4017
4018	INIT_LIST_HEAD(&rc->reloc_roots);
4019	INIT_LIST_HEAD(&rc->dirty_subvol_roots);
4020	btrfs_backref_init_cache(fs_info, &rc->backref_cache, true);
4021	rc->reloc_root_tree.rb_root = RB_ROOT;
4022	spin_lock_init(&rc->reloc_root_tree.lock);
4023	extent_io_tree_init(fs_info, &rc->processed_blocks, IO_TREE_RELOC_BLOCKS);
4024	return rc;
4025}
4026
4027static void free_reloc_control(struct reloc_control *rc)
4028{
4029	struct mapping_node *node, *tmp;
4030
4031	free_reloc_roots(&rc->reloc_roots);
4032	rbtree_postorder_for_each_entry_safe(node, tmp,
4033			&rc->reloc_root_tree.rb_root, rb_node)
4034		kfree(node);
4035
4036	kfree(rc);
4037}
4038
4039/*
4040 * Print the block group being relocated
4041 */
4042static void describe_relocation(struct btrfs_fs_info *fs_info,
4043				struct btrfs_block_group *block_group)
4044{
4045	char buf[128] = {'\0'};
4046
4047	btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
4048
4049	btrfs_info(fs_info,
4050		   "relocating block group %llu flags %s",
4051		   block_group->start, buf);
4052}
4053
4054static const char *stage_to_string(enum reloc_stage stage)
4055{
4056	if (stage == MOVE_DATA_EXTENTS)
4057		return "move data extents";
4058	if (stage == UPDATE_DATA_PTRS)
4059		return "update data pointers";
4060	return "unknown";
4061}
4062
4063/*
4064 * function to relocate all extents in a block group.
4065 */
4066int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4067{
4068	struct btrfs_block_group *bg;
4069	struct btrfs_root *extent_root = btrfs_extent_root(fs_info, group_start);
4070	struct reloc_control *rc;
4071	struct inode *inode;
4072	struct btrfs_path *path;
4073	int ret;
4074	int rw = 0;
4075	int err = 0;
4076
4077	/*
4078	 * This only gets set if we had a half-deleted snapshot on mount.  We
4079	 * cannot allow relocation to start while we're still trying to clean up
4080	 * these pending deletions.
4081	 */
4082	ret = wait_on_bit(&fs_info->flags, BTRFS_FS_UNFINISHED_DROPS, TASK_INTERRUPTIBLE);
4083	if (ret)
4084		return ret;
4085
4086	/* We may have been woken up by close_ctree, so bail if we're closing. */
4087	if (btrfs_fs_closing(fs_info))
4088		return -EINTR;
4089
4090	bg = btrfs_lookup_block_group(fs_info, group_start);
4091	if (!bg)
4092		return -ENOENT;
4093
4094	/*
4095	 * Relocation of a data block group creates ordered extents.  Without
4096	 * sb_start_write(), we can freeze the filesystem while unfinished
4097	 * ordered extents are left. Such ordered extents can cause a deadlock
4098	 * e.g. when syncfs() is waiting for their completion but they can't
4099	 * finish because they block when joining a transaction, due to the
4100	 * fact that the freeze locks are being held in write mode.
4101	 */
4102	if (bg->flags & BTRFS_BLOCK_GROUP_DATA)
4103		ASSERT(sb_write_started(fs_info->sb));
4104
4105	if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4106		btrfs_put_block_group(bg);
4107		return -ETXTBSY;
4108	}
4109
4110	rc = alloc_reloc_control(fs_info);
4111	if (!rc) {
4112		btrfs_put_block_group(bg);
4113		return -ENOMEM;
4114	}
4115
4116	ret = reloc_chunk_start(fs_info);
4117	if (ret < 0) {
4118		err = ret;
4119		goto out_put_bg;
4120	}
4121
4122	rc->extent_root = extent_root;
4123	rc->block_group = bg;
4124
4125	ret = btrfs_inc_block_group_ro(rc->block_group, true);
4126	if (ret) {
4127		err = ret;
4128		goto out;
4129	}
4130	rw = 1;
4131
4132	path = btrfs_alloc_path();
4133	if (!path) {
4134		err = -ENOMEM;
4135		goto out;
4136	}
4137
4138	inode = lookup_free_space_inode(rc->block_group, path);
4139	btrfs_free_path(path);
4140
4141	if (!IS_ERR(inode))
4142		ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4143	else
4144		ret = PTR_ERR(inode);
4145
4146	if (ret && ret != -ENOENT) {
4147		err = ret;
4148		goto out;
4149	}
4150
4151	rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4152	if (IS_ERR(rc->data_inode)) {
4153		err = PTR_ERR(rc->data_inode);
4154		rc->data_inode = NULL;
4155		goto out;
4156	}
4157
4158	describe_relocation(fs_info, rc->block_group);
4159
4160	btrfs_wait_block_group_reservations(rc->block_group);
4161	btrfs_wait_nocow_writers(rc->block_group);
4162	btrfs_wait_ordered_roots(fs_info, U64_MAX,
4163				 rc->block_group->start,
4164				 rc->block_group->length);
4165
4166	ret = btrfs_zone_finish(rc->block_group);
4167	WARN_ON(ret && ret != -EAGAIN);
4168
4169	while (1) {
4170		enum reloc_stage finishes_stage;
4171
4172		mutex_lock(&fs_info->cleaner_mutex);
4173		ret = relocate_block_group(rc);
4174		mutex_unlock(&fs_info->cleaner_mutex);
4175		if (ret < 0)
4176			err = ret;
4177
4178		finishes_stage = rc->stage;
4179		/*
4180		 * We may have gotten ENOSPC after we already dirtied some
4181		 * extents.  If writeout happens while we're relocating a
4182		 * different block group we could end up hitting the
4183		 * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4184		 * btrfs_reloc_cow_block.  Make sure we write everything out
4185		 * properly so we don't trip over this problem, and then break
4186		 * out of the loop if we hit an error.
4187		 */
4188		if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4189			ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4190						       (u64)-1);
4191			if (ret)
4192				err = ret;
4193			invalidate_mapping_pages(rc->data_inode->i_mapping,
4194						 0, -1);
4195			rc->stage = UPDATE_DATA_PTRS;
4196		}
4197
4198		if (err < 0)
4199			goto out;
4200
4201		if (rc->extents_found == 0)
4202			break;
4203
4204		btrfs_info(fs_info, "found %llu extents, stage: %s",
4205			   rc->extents_found, stage_to_string(finishes_stage));
4206	}
4207
4208	WARN_ON(rc->block_group->pinned > 0);
4209	WARN_ON(rc->block_group->reserved > 0);
4210	WARN_ON(rc->block_group->used > 0);
4211out:
4212	if (err && rw)
4213		btrfs_dec_block_group_ro(rc->block_group);
4214	iput(rc->data_inode);
4215out_put_bg:
4216	btrfs_put_block_group(bg);
4217	reloc_chunk_end(fs_info);
4218	free_reloc_control(rc);
4219	return err;
4220}
4221
4222static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4223{
4224	struct btrfs_fs_info *fs_info = root->fs_info;
4225	struct btrfs_trans_handle *trans;
4226	int ret, err;
4227
4228	trans = btrfs_start_transaction(fs_info->tree_root, 0);
4229	if (IS_ERR(trans))
4230		return PTR_ERR(trans);
4231
4232	memset(&root->root_item.drop_progress, 0,
4233		sizeof(root->root_item.drop_progress));
4234	btrfs_set_root_drop_level(&root->root_item, 0);
4235	btrfs_set_root_refs(&root->root_item, 0);
4236	ret = btrfs_update_root(trans, fs_info->tree_root,
4237				&root->root_key, &root->root_item);
4238
4239	err = btrfs_end_transaction(trans);
4240	if (err)
4241		return err;
4242	return ret;
4243}
4244
4245/*
4246 * recover relocation interrupted by system crash.
4247 *
4248 * this function resumes merging reloc trees with corresponding fs trees.
4249 * this is important for keeping the sharing of tree blocks
4250 */
4251int btrfs_recover_relocation(struct btrfs_fs_info *fs_info)
4252{
4253	LIST_HEAD(reloc_roots);
4254	struct btrfs_key key;
4255	struct btrfs_root *fs_root;
4256	struct btrfs_root *reloc_root;
4257	struct btrfs_path *path;
4258	struct extent_buffer *leaf;
4259	struct reloc_control *rc = NULL;
4260	struct btrfs_trans_handle *trans;
4261	int ret;
4262	int err = 0;
4263
4264	path = btrfs_alloc_path();
4265	if (!path)
4266		return -ENOMEM;
4267	path->reada = READA_BACK;
4268
4269	key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4270	key.type = BTRFS_ROOT_ITEM_KEY;
4271	key.offset = (u64)-1;
4272
4273	while (1) {
4274		ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4275					path, 0, 0);
4276		if (ret < 0) {
4277			err = ret;
4278			goto out;
4279		}
4280		if (ret > 0) {
4281			if (path->slots[0] == 0)
4282				break;
4283			path->slots[0]--;
4284		}
4285		leaf = path->nodes[0];
4286		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4287		btrfs_release_path(path);
4288
4289		if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4290		    key.type != BTRFS_ROOT_ITEM_KEY)
4291			break;
4292
4293		reloc_root = btrfs_read_tree_root(fs_info->tree_root, &key);
4294		if (IS_ERR(reloc_root)) {
4295			err = PTR_ERR(reloc_root);
4296			goto out;
4297		}
4298
4299		set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4300		list_add(&reloc_root->root_list, &reloc_roots);
4301
4302		if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4303			fs_root = btrfs_get_fs_root(fs_info,
4304					reloc_root->root_key.offset, false);
4305			if (IS_ERR(fs_root)) {
4306				ret = PTR_ERR(fs_root);
4307				if (ret != -ENOENT) {
4308					err = ret;
4309					goto out;
4310				}
4311				ret = mark_garbage_root(reloc_root);
4312				if (ret < 0) {
4313					err = ret;
4314					goto out;
4315				}
4316			} else {
4317				btrfs_put_root(fs_root);
4318			}
4319		}
4320
4321		if (key.offset == 0)
4322			break;
4323
4324		key.offset--;
4325	}
4326	btrfs_release_path(path);
4327
4328	if (list_empty(&reloc_roots))
4329		goto out;
4330
4331	rc = alloc_reloc_control(fs_info);
4332	if (!rc) {
4333		err = -ENOMEM;
4334		goto out;
4335	}
4336
4337	ret = reloc_chunk_start(fs_info);
4338	if (ret < 0) {
4339		err = ret;
4340		goto out_end;
4341	}
4342
4343	rc->extent_root = btrfs_extent_root(fs_info, 0);
4344
4345	set_reloc_control(rc);
4346
4347	trans = btrfs_join_transaction(rc->extent_root);
4348	if (IS_ERR(trans)) {
4349		err = PTR_ERR(trans);
4350		goto out_unset;
4351	}
4352
4353	rc->merge_reloc_tree = true;
4354
4355	while (!list_empty(&reloc_roots)) {
4356		reloc_root = list_entry(reloc_roots.next,
4357					struct btrfs_root, root_list);
4358		list_del(&reloc_root->root_list);
4359
4360		if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4361			list_add_tail(&reloc_root->root_list,
4362				      &rc->reloc_roots);
4363			continue;
4364		}
4365
4366		fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4367					    false);
4368		if (IS_ERR(fs_root)) {
4369			err = PTR_ERR(fs_root);
4370			list_add_tail(&reloc_root->root_list, &reloc_roots);
4371			btrfs_end_transaction(trans);
4372			goto out_unset;
4373		}
4374
4375		err = __add_reloc_root(reloc_root);
4376		ASSERT(err != -EEXIST);
4377		if (err) {
4378			list_add_tail(&reloc_root->root_list, &reloc_roots);
4379			btrfs_put_root(fs_root);
4380			btrfs_end_transaction(trans);
4381			goto out_unset;
4382		}
4383		fs_root->reloc_root = btrfs_grab_root(reloc_root);
4384		btrfs_put_root(fs_root);
4385	}
4386
4387	err = btrfs_commit_transaction(trans);
4388	if (err)
4389		goto out_unset;
4390
4391	merge_reloc_roots(rc);
4392
4393	unset_reloc_control(rc);
4394
4395	trans = btrfs_join_transaction(rc->extent_root);
4396	if (IS_ERR(trans)) {
4397		err = PTR_ERR(trans);
4398		goto out_clean;
4399	}
4400	err = btrfs_commit_transaction(trans);
4401out_clean:
4402	ret = clean_dirty_subvols(rc);
4403	if (ret < 0 && !err)
4404		err = ret;
4405out_unset:
4406	unset_reloc_control(rc);
4407out_end:
4408	reloc_chunk_end(fs_info);
4409	free_reloc_control(rc);
4410out:
4411	free_reloc_roots(&reloc_roots);
4412
4413	btrfs_free_path(path);
4414
4415	if (err == 0) {
4416		/* cleanup orphan inode in data relocation tree */
4417		fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4418		ASSERT(fs_root);
4419		err = btrfs_orphan_cleanup(fs_root);
4420		btrfs_put_root(fs_root);
4421	}
4422	return err;
4423}
4424
4425/*
4426 * helper to add ordered checksum for data relocation.
4427 *
4428 * cloning checksum properly handles the nodatasum extents.
4429 * it also saves CPU time to re-calculate the checksum.
4430 */
4431int btrfs_reloc_clone_csums(struct btrfs_ordered_extent *ordered)
4432{
4433	struct btrfs_inode *inode = BTRFS_I(ordered->inode);
4434	struct btrfs_fs_info *fs_info = inode->root->fs_info;
4435	u64 disk_bytenr = ordered->file_offset + inode->index_cnt;
4436	struct btrfs_root *csum_root = btrfs_csum_root(fs_info, disk_bytenr);
4437	LIST_HEAD(list);
4438	int ret;
4439
4440	ret = btrfs_lookup_csums_list(csum_root, disk_bytenr,
4441				      disk_bytenr + ordered->num_bytes - 1,
4442				      &list, 0, false);
4443	if (ret)
4444		return ret;
4445
4446	while (!list_empty(&list)) {
4447		struct btrfs_ordered_sum *sums =
4448			list_entry(list.next, struct btrfs_ordered_sum, list);
4449
4450		list_del_init(&sums->list);
4451
4452		/*
4453		 * We need to offset the new_bytenr based on where the csum is.
4454		 * We need to do this because we will read in entire prealloc
4455		 * extents but we may have written to say the middle of the
4456		 * prealloc extent, so we need to make sure the csum goes with
4457		 * the right disk offset.
4458		 *
4459		 * We can do this because the data reloc inode refers strictly
4460		 * to the on disk bytes, so we don't have to worry about
4461		 * disk_len vs real len like with real inodes since it's all
4462		 * disk length.
4463		 */
4464		sums->logical = ordered->disk_bytenr + sums->logical - disk_bytenr;
4465		btrfs_add_ordered_sum(ordered, sums);
4466	}
4467
4468	return 0;
4469}
4470
4471int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4472			  struct btrfs_root *root,
4473			  const struct extent_buffer *buf,
4474			  struct extent_buffer *cow)
4475{
4476	struct btrfs_fs_info *fs_info = root->fs_info;
4477	struct reloc_control *rc;
4478	struct btrfs_backref_node *node;
4479	int first_cow = 0;
4480	int level;
4481	int ret = 0;
4482
4483	rc = fs_info->reloc_ctl;
4484	if (!rc)
4485		return 0;
4486
4487	BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4488
4489	level = btrfs_header_level(buf);
4490	if (btrfs_header_generation(buf) <=
4491	    btrfs_root_last_snapshot(&root->root_item))
4492		first_cow = 1;
4493
4494	if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4495	    rc->create_reloc_tree) {
4496		WARN_ON(!first_cow && level == 0);
4497
4498		node = rc->backref_cache.path[level];
4499		BUG_ON(node->bytenr != buf->start &&
4500		       node->new_bytenr != buf->start);
4501
4502		btrfs_backref_drop_node_buffer(node);
4503		atomic_inc(&cow->refs);
4504		node->eb = cow;
4505		node->new_bytenr = cow->start;
4506
4507		if (!node->pending) {
4508			list_move_tail(&node->list,
4509				       &rc->backref_cache.pending[level]);
4510			node->pending = 1;
4511		}
4512
4513		if (first_cow)
4514			mark_block_processed(rc, node);
4515
4516		if (first_cow && level > 0)
4517			rc->nodes_relocated += buf->len;
4518	}
4519
4520	if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4521		ret = replace_file_extents(trans, rc, root, cow);
4522	return ret;
4523}
4524
4525/*
4526 * called before creating snapshot. it calculates metadata reservation
4527 * required for relocating tree blocks in the snapshot
4528 */
4529void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4530			      u64 *bytes_to_reserve)
4531{
4532	struct btrfs_root *root = pending->root;
4533	struct reloc_control *rc = root->fs_info->reloc_ctl;
4534
4535	if (!rc || !have_reloc_root(root))
4536		return;
4537
4538	if (!rc->merge_reloc_tree)
4539		return;
4540
4541	root = root->reloc_root;
4542	BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4543	/*
4544	 * relocation is in the stage of merging trees. the space
4545	 * used by merging a reloc tree is twice the size of
4546	 * relocated tree nodes in the worst case. half for cowing
4547	 * the reloc tree, half for cowing the fs tree. the space
4548	 * used by cowing the reloc tree will be freed after the
4549	 * tree is dropped. if we create snapshot, cowing the fs
4550	 * tree may use more space than it frees. so we need
4551	 * reserve extra space.
4552	 */
4553	*bytes_to_reserve += rc->nodes_relocated;
4554}
4555
4556/*
4557 * called after snapshot is created. migrate block reservation
4558 * and create reloc root for the newly created snapshot
4559 *
4560 * This is similar to btrfs_init_reloc_root(), we come out of here with two
4561 * references held on the reloc_root, one for root->reloc_root and one for
4562 * rc->reloc_roots.
4563 */
4564int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4565			       struct btrfs_pending_snapshot *pending)
4566{
4567	struct btrfs_root *root = pending->root;
4568	struct btrfs_root *reloc_root;
4569	struct btrfs_root *new_root;
4570	struct reloc_control *rc = root->fs_info->reloc_ctl;
4571	int ret;
4572
4573	if (!rc || !have_reloc_root(root))
4574		return 0;
4575
4576	rc = root->fs_info->reloc_ctl;
4577	rc->merging_rsv_size += rc->nodes_relocated;
4578
4579	if (rc->merge_reloc_tree) {
4580		ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4581					      rc->block_rsv,
4582					      rc->nodes_relocated, true);
4583		if (ret)
4584			return ret;
4585	}
4586
4587	new_root = pending->snap;
4588	reloc_root = create_reloc_root(trans, root->reloc_root,
4589				       new_root->root_key.objectid);
4590	if (IS_ERR(reloc_root))
4591		return PTR_ERR(reloc_root);
4592
4593	ret = __add_reloc_root(reloc_root);
4594	ASSERT(ret != -EEXIST);
4595	if (ret) {
4596		/* Pairs with create_reloc_root */
4597		btrfs_put_root(reloc_root);
4598		return ret;
4599	}
4600	new_root->reloc_root = btrfs_grab_root(reloc_root);
4601
4602	if (rc->create_reloc_tree)
4603		ret = clone_backref_node(trans, rc, root, reloc_root);
4604	return ret;
4605}
4606
4607/*
4608 * Get the current bytenr for the block group which is being relocated.
4609 *
4610 * Return U64_MAX if no running relocation.
4611 */
4612u64 btrfs_get_reloc_bg_bytenr(const struct btrfs_fs_info *fs_info)
4613{
4614	u64 logical = U64_MAX;
4615
4616	lockdep_assert_held(&fs_info->reloc_mutex);
4617
4618	if (fs_info->reloc_ctl && fs_info->reloc_ctl->block_group)
4619		logical = fs_info->reloc_ctl->block_group->start;
4620	return logical;
4621}