1c1d7c514SDavid Sterba // SPDX-License-Identifier: GPL-2.0 25d4f98a2SYan Zheng /* 35d4f98a2SYan Zheng * Copyright (C) 2009 Oracle. All rights reserved. 45d4f98a2SYan Zheng */ 55d4f98a2SYan Zheng 65d4f98a2SYan Zheng #include <linux/sched.h> 75d4f98a2SYan Zheng #include <linux/pagemap.h> 85d4f98a2SYan Zheng #include <linux/writeback.h> 95d4f98a2SYan Zheng #include <linux/blkdev.h> 105d4f98a2SYan Zheng #include <linux/rbtree.h> 115a0e3ad6STejun Heo #include <linux/slab.h> 12726a3421SQu Wenruo #include <linux/error-injection.h> 135d4f98a2SYan Zheng #include "ctree.h" 145d4f98a2SYan Zheng #include "disk-io.h" 155d4f98a2SYan Zheng #include "transaction.h" 165d4f98a2SYan Zheng #include "volumes.h" 175d4f98a2SYan Zheng #include "locking.h" 185d4f98a2SYan Zheng #include "btrfs_inode.h" 195d4f98a2SYan Zheng #include "async-thread.h" 200af3d00bSJosef Bacik #include "free-space-cache.h" 2162b99540SQu Wenruo #include "qgroup.h" 22cdccee99SLiu Bo #include "print-tree.h" 2386736342SJosef Bacik #include "delalloc-space.h" 24aac0023cSJosef Bacik #include "block-group.h" 2519b546d7SQu Wenruo #include "backref.h" 26e9a28dc5SQu Wenruo #include "misc.h" 275d4f98a2SYan Zheng 285d4f98a2SYan Zheng /* 290c891389SQu Wenruo * Relocation overview 300c891389SQu Wenruo * 310c891389SQu Wenruo * [What does relocation do] 320c891389SQu Wenruo * 330c891389SQu Wenruo * The objective of relocation is to relocate all extents of the target block 340c891389SQu Wenruo * group to other block groups. 350c891389SQu Wenruo * This is utilized by resize (shrink only), profile converting, compacting 360c891389SQu Wenruo * space, or balance routine to spread chunks over devices. 370c891389SQu Wenruo * 380c891389SQu Wenruo * Before | After 390c891389SQu Wenruo * ------------------------------------------------------------------ 400c891389SQu Wenruo * BG A: 10 data extents | BG A: deleted 410c891389SQu Wenruo * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated) 420c891389SQu Wenruo * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated) 430c891389SQu Wenruo * 440c891389SQu Wenruo * [How does relocation work] 450c891389SQu Wenruo * 460c891389SQu Wenruo * 1. Mark the target block group read-only 470c891389SQu Wenruo * New extents won't be allocated from the target block group. 480c891389SQu Wenruo * 490c891389SQu Wenruo * 2.1 Record each extent in the target block group 500c891389SQu Wenruo * To build a proper map of extents to be relocated. 510c891389SQu Wenruo * 520c891389SQu Wenruo * 2.2 Build data reloc tree and reloc trees 530c891389SQu Wenruo * Data reloc tree will contain an inode, recording all newly relocated 540c891389SQu Wenruo * data extents. 550c891389SQu Wenruo * There will be only one data reloc tree for one data block group. 560c891389SQu Wenruo * 570c891389SQu Wenruo * Reloc tree will be a special snapshot of its source tree, containing 580c891389SQu Wenruo * relocated tree blocks. 590c891389SQu Wenruo * Each tree referring to a tree block in target block group will get its 600c891389SQu Wenruo * reloc tree built. 610c891389SQu Wenruo * 620c891389SQu Wenruo * 2.3 Swap source tree with its corresponding reloc tree 630c891389SQu Wenruo * Each involved tree only refers to new extents after swap. 640c891389SQu Wenruo * 650c891389SQu Wenruo * 3. Cleanup reloc trees and data reloc tree. 660c891389SQu Wenruo * As old extents in the target block group are still referenced by reloc 670c891389SQu Wenruo * trees, we need to clean them up before really freeing the target block 680c891389SQu Wenruo * group. 690c891389SQu Wenruo * 700c891389SQu Wenruo * The main complexity is in steps 2.2 and 2.3. 710c891389SQu Wenruo * 720c891389SQu Wenruo * The entry point of relocation is relocate_block_group() function. 730c891389SQu Wenruo */ 740c891389SQu Wenruo 750647bf56SWang Shilong #define RELOCATION_RESERVED_NODES 256 762a979612SQu Wenruo /* 775d4f98a2SYan Zheng * map address of tree root to tree 785d4f98a2SYan Zheng */ 795d4f98a2SYan Zheng struct mapping_node { 80e9a28dc5SQu Wenruo struct { 815d4f98a2SYan Zheng struct rb_node rb_node; 825d4f98a2SYan Zheng u64 bytenr; 83e9a28dc5SQu Wenruo }; /* Use rb_simle_node for search/insert */ 845d4f98a2SYan Zheng void *data; 855d4f98a2SYan Zheng }; 865d4f98a2SYan Zheng 875d4f98a2SYan Zheng struct mapping_tree { 885d4f98a2SYan Zheng struct rb_root rb_root; 895d4f98a2SYan Zheng spinlock_t lock; 905d4f98a2SYan Zheng }; 915d4f98a2SYan Zheng 925d4f98a2SYan Zheng /* 935d4f98a2SYan Zheng * present a tree block to process 945d4f98a2SYan Zheng */ 955d4f98a2SYan Zheng struct tree_block { 96e9a28dc5SQu Wenruo struct { 975d4f98a2SYan Zheng struct rb_node rb_node; 985d4f98a2SYan Zheng u64 bytenr; 99e9a28dc5SQu Wenruo }; /* Use rb_simple_node for search/insert */ 100f7ba2d37SJosef Bacik u64 owner; 1015d4f98a2SYan Zheng struct btrfs_key key; 1025d4f98a2SYan Zheng unsigned int level:8; 1035d4f98a2SYan Zheng unsigned int key_ready:1; 1045d4f98a2SYan Zheng }; 1055d4f98a2SYan Zheng 1060257bb82SYan, Zheng #define MAX_EXTENTS 128 1070257bb82SYan, Zheng 1080257bb82SYan, Zheng struct file_extent_cluster { 1090257bb82SYan, Zheng u64 start; 1100257bb82SYan, Zheng u64 end; 1110257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 1120257bb82SYan, Zheng unsigned int nr; 1130257bb82SYan, Zheng }; 1140257bb82SYan, Zheng 1155d4f98a2SYan Zheng struct reloc_control { 1165d4f98a2SYan Zheng /* block group to relocate */ 11732da5386SDavid Sterba struct btrfs_block_group *block_group; 1185d4f98a2SYan Zheng /* extent tree */ 1195d4f98a2SYan Zheng struct btrfs_root *extent_root; 1205d4f98a2SYan Zheng /* inode for moving data */ 1215d4f98a2SYan Zheng struct inode *data_inode; 1223fd0a558SYan, Zheng 1233fd0a558SYan, Zheng struct btrfs_block_rsv *block_rsv; 1243fd0a558SYan, Zheng 125a26195a5SQu Wenruo struct btrfs_backref_cache backref_cache; 1263fd0a558SYan, Zheng 1273fd0a558SYan, Zheng struct file_extent_cluster cluster; 1285d4f98a2SYan Zheng /* tree blocks have been processed */ 1295d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 1305d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 1315d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 1325d4f98a2SYan Zheng /* list of reloc trees */ 1335d4f98a2SYan Zheng struct list_head reloc_roots; 134d2311e69SQu Wenruo /* list of subvolume trees that get relocated */ 135d2311e69SQu Wenruo struct list_head dirty_subvol_roots; 1363fd0a558SYan, Zheng /* size of metadata reservation for merging reloc trees */ 1373fd0a558SYan, Zheng u64 merging_rsv_size; 1383fd0a558SYan, Zheng /* size of relocated tree nodes */ 1393fd0a558SYan, Zheng u64 nodes_relocated; 1400647bf56SWang Shilong /* reserved size for block group relocation*/ 1410647bf56SWang Shilong u64 reserved_bytes; 1423fd0a558SYan, Zheng 1435d4f98a2SYan Zheng u64 search_start; 1445d4f98a2SYan Zheng u64 extents_found; 1453fd0a558SYan, Zheng 1463fd0a558SYan, Zheng unsigned int stage:8; 1473fd0a558SYan, Zheng unsigned int create_reloc_tree:1; 1483fd0a558SYan, Zheng unsigned int merge_reloc_tree:1; 1495d4f98a2SYan Zheng unsigned int found_file_extent:1; 1505d4f98a2SYan Zheng }; 1515d4f98a2SYan Zheng 1525d4f98a2SYan Zheng /* stages of data relocation */ 1535d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 1545d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 1555d4f98a2SYan Zheng 1569569cc20SQu Wenruo static void mark_block_processed(struct reloc_control *rc, 157a26195a5SQu Wenruo struct btrfs_backref_node *node) 1589569cc20SQu Wenruo { 1599569cc20SQu Wenruo u32 blocksize; 1609569cc20SQu Wenruo 1619569cc20SQu Wenruo if (node->level == 0 || 1629569cc20SQu Wenruo in_range(node->bytenr, rc->block_group->start, 1639569cc20SQu Wenruo rc->block_group->length)) { 1649569cc20SQu Wenruo blocksize = rc->extent_root->fs_info->nodesize; 1659569cc20SQu Wenruo set_extent_bits(&rc->processed_blocks, node->bytenr, 1669569cc20SQu Wenruo node->bytenr + blocksize - 1, EXTENT_DIRTY); 1679569cc20SQu Wenruo } 1689569cc20SQu Wenruo node->processed = 1; 1699569cc20SQu Wenruo } 1709569cc20SQu Wenruo 1715d4f98a2SYan Zheng 1725d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 1735d4f98a2SYan Zheng { 1746bef4d31SEric Paris tree->rb_root = RB_ROOT; 1755d4f98a2SYan Zheng spin_lock_init(&tree->lock); 1765d4f98a2SYan Zheng } 1775d4f98a2SYan Zheng 1785d4f98a2SYan Zheng /* 1795d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 1805d4f98a2SYan Zheng */ 181a26195a5SQu Wenruo static struct btrfs_backref_node *walk_up_backref( 182a26195a5SQu Wenruo struct btrfs_backref_node *node, 183a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 1845d4f98a2SYan Zheng { 185a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1865d4f98a2SYan Zheng int idx = *index; 1875d4f98a2SYan Zheng 1885d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 1895d4f98a2SYan Zheng edge = list_entry(node->upper.next, 190a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 1915d4f98a2SYan Zheng edges[idx++] = edge; 1925d4f98a2SYan Zheng node = edge->node[UPPER]; 1935d4f98a2SYan Zheng } 1943fd0a558SYan, Zheng BUG_ON(node->detached); 1955d4f98a2SYan Zheng *index = idx; 1965d4f98a2SYan Zheng return node; 1975d4f98a2SYan Zheng } 1985d4f98a2SYan Zheng 1995d4f98a2SYan Zheng /* 2005d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 2015d4f98a2SYan Zheng */ 202a26195a5SQu Wenruo static struct btrfs_backref_node *walk_down_backref( 203a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 2045d4f98a2SYan Zheng { 205a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 206a26195a5SQu Wenruo struct btrfs_backref_node *lower; 2075d4f98a2SYan Zheng int idx = *index; 2085d4f98a2SYan Zheng 2095d4f98a2SYan Zheng while (idx > 0) { 2105d4f98a2SYan Zheng edge = edges[idx - 1]; 2115d4f98a2SYan Zheng lower = edge->node[LOWER]; 2125d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 2135d4f98a2SYan Zheng idx--; 2145d4f98a2SYan Zheng continue; 2155d4f98a2SYan Zheng } 2165d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 217a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 2185d4f98a2SYan Zheng edges[idx - 1] = edge; 2195d4f98a2SYan Zheng *index = idx; 2205d4f98a2SYan Zheng return edge->node[UPPER]; 2215d4f98a2SYan Zheng } 2225d4f98a2SYan Zheng *index = 0; 2235d4f98a2SYan Zheng return NULL; 2245d4f98a2SYan Zheng } 2255d4f98a2SYan Zheng 226a26195a5SQu Wenruo static void update_backref_node(struct btrfs_backref_cache *cache, 227a26195a5SQu Wenruo struct btrfs_backref_node *node, u64 bytenr) 2283fd0a558SYan, Zheng { 2293fd0a558SYan, Zheng struct rb_node *rb_node; 2303fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 2313fd0a558SYan, Zheng node->bytenr = bytenr; 232e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node); 23343c04fb1SJeff Mahoney if (rb_node) 234982c92cbSQu Wenruo btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST); 2353fd0a558SYan, Zheng } 2363fd0a558SYan, Zheng 2373fd0a558SYan, Zheng /* 2383fd0a558SYan, Zheng * update backref cache after a transaction commit 2393fd0a558SYan, Zheng */ 2403fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 241a26195a5SQu Wenruo struct btrfs_backref_cache *cache) 2423fd0a558SYan, Zheng { 243a26195a5SQu Wenruo struct btrfs_backref_node *node; 2443fd0a558SYan, Zheng int level = 0; 2453fd0a558SYan, Zheng 2463fd0a558SYan, Zheng if (cache->last_trans == 0) { 2473fd0a558SYan, Zheng cache->last_trans = trans->transid; 2483fd0a558SYan, Zheng return 0; 2493fd0a558SYan, Zheng } 2503fd0a558SYan, Zheng 2513fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 2523fd0a558SYan, Zheng return 0; 2533fd0a558SYan, Zheng 2543fd0a558SYan, Zheng /* 2553fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 2563fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 2573fd0a558SYan, Zheng * so the detached nodes are no longer useful. 2583fd0a558SYan, Zheng */ 2593fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 2603fd0a558SYan, Zheng node = list_entry(cache->detached.next, 261a26195a5SQu Wenruo struct btrfs_backref_node, list); 262023acb07SQu Wenruo btrfs_backref_cleanup_node(cache, node); 2633fd0a558SYan, Zheng } 2643fd0a558SYan, Zheng 2653fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 2663fd0a558SYan, Zheng node = list_entry(cache->changed.next, 267a26195a5SQu Wenruo struct btrfs_backref_node, list); 2683fd0a558SYan, Zheng list_del_init(&node->list); 2693fd0a558SYan, Zheng BUG_ON(node->pending); 2703fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 2713fd0a558SYan, Zheng } 2723fd0a558SYan, Zheng 2733fd0a558SYan, Zheng /* 2743fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 2753fd0a558SYan, Zheng * errors during processing the pending nodes. 2763fd0a558SYan, Zheng */ 2773fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 2783fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 2793fd0a558SYan, Zheng BUG_ON(!node->pending); 2803fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 2813fd0a558SYan, Zheng continue; 2823fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 2833fd0a558SYan, Zheng } 2843fd0a558SYan, Zheng } 2853fd0a558SYan, Zheng 2863fd0a558SYan, Zheng cache->last_trans = 0; 2873fd0a558SYan, Zheng return 1; 2883fd0a558SYan, Zheng } 2893fd0a558SYan, Zheng 2906282675eSQu Wenruo static bool reloc_root_is_dead(struct btrfs_root *root) 2916282675eSQu Wenruo { 2926282675eSQu Wenruo /* 2936282675eSQu Wenruo * Pair with set_bit/clear_bit in clean_dirty_subvols and 2946282675eSQu Wenruo * btrfs_update_reloc_root. We need to see the updated bit before 2956282675eSQu Wenruo * trying to access reloc_root 2966282675eSQu Wenruo */ 2976282675eSQu Wenruo smp_rmb(); 2986282675eSQu Wenruo if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state)) 2996282675eSQu Wenruo return true; 3006282675eSQu Wenruo return false; 3016282675eSQu Wenruo } 3026282675eSQu Wenruo 3036282675eSQu Wenruo /* 3046282675eSQu Wenruo * Check if this subvolume tree has valid reloc tree. 3056282675eSQu Wenruo * 3066282675eSQu Wenruo * Reloc tree after swap is considered dead, thus not considered as valid. 3076282675eSQu Wenruo * This is enough for most callers, as they don't distinguish dead reloc root 30855465730SQu Wenruo * from no reloc root. But btrfs_should_ignore_reloc_root() below is a 30955465730SQu Wenruo * special case. 3106282675eSQu Wenruo */ 3116282675eSQu Wenruo static bool have_reloc_root(struct btrfs_root *root) 3126282675eSQu Wenruo { 3136282675eSQu Wenruo if (reloc_root_is_dead(root)) 3146282675eSQu Wenruo return false; 3156282675eSQu Wenruo if (!root->reloc_root) 3166282675eSQu Wenruo return false; 3176282675eSQu Wenruo return true; 3186282675eSQu Wenruo } 319f2a97a9dSDavid Sterba 32055465730SQu Wenruo int btrfs_should_ignore_reloc_root(struct btrfs_root *root) 3213fd0a558SYan, Zheng { 3223fd0a558SYan, Zheng struct btrfs_root *reloc_root; 3233fd0a558SYan, Zheng 32492a7cc42SQu Wenruo if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 3253fd0a558SYan, Zheng return 0; 3263fd0a558SYan, Zheng 3276282675eSQu Wenruo /* This root has been merged with its reloc tree, we can ignore it */ 3286282675eSQu Wenruo if (reloc_root_is_dead(root)) 3296282675eSQu Wenruo return 1; 3306282675eSQu Wenruo 3313fd0a558SYan, Zheng reloc_root = root->reloc_root; 3323fd0a558SYan, Zheng if (!reloc_root) 3333fd0a558SYan, Zheng return 0; 3343fd0a558SYan, Zheng 3354d4225fcSJosef Bacik if (btrfs_header_generation(reloc_root->commit_root) == 3364d4225fcSJosef Bacik root->fs_info->running_transaction->transid) 3373fd0a558SYan, Zheng return 0; 3383fd0a558SYan, Zheng /* 3393fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 3403fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 3413fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 3423fd0a558SYan, Zheng * relocation. 3433fd0a558SYan, Zheng */ 3443fd0a558SYan, Zheng return 1; 3453fd0a558SYan, Zheng } 34655465730SQu Wenruo 3475d4f98a2SYan Zheng /* 3485d4f98a2SYan Zheng * find reloc tree by address of tree root 3495d4f98a2SYan Zheng */ 3502433bea5SQu Wenruo struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr) 3515d4f98a2SYan Zheng { 3522433bea5SQu Wenruo struct reloc_control *rc = fs_info->reloc_ctl; 3535d4f98a2SYan Zheng struct rb_node *rb_node; 3545d4f98a2SYan Zheng struct mapping_node *node; 3555d4f98a2SYan Zheng struct btrfs_root *root = NULL; 3565d4f98a2SYan Zheng 3572433bea5SQu Wenruo ASSERT(rc); 3585d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 359e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr); 3605d4f98a2SYan Zheng if (rb_node) { 3615d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 3625d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 3635d4f98a2SYan Zheng } 3645d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 36500246528SJosef Bacik return btrfs_grab_root(root); 3665d4f98a2SYan Zheng } 3675d4f98a2SYan Zheng 3685d4f98a2SYan Zheng /* 36929db137bSQu Wenruo * For useless nodes, do two major clean ups: 37029db137bSQu Wenruo * 37129db137bSQu Wenruo * - Cleanup the children edges and nodes 37229db137bSQu Wenruo * If child node is also orphan (no parent) during cleanup, then the child 37329db137bSQu Wenruo * node will also be cleaned up. 37429db137bSQu Wenruo * 37529db137bSQu Wenruo * - Freeing up leaves (level 0), keeps nodes detached 37629db137bSQu Wenruo * For nodes, the node is still cached as "detached" 37729db137bSQu Wenruo * 37829db137bSQu Wenruo * Return false if @node is not in the @useless_nodes list. 37929db137bSQu Wenruo * Return true if @node is in the @useless_nodes list. 38029db137bSQu Wenruo */ 38129db137bSQu Wenruo static bool handle_useless_nodes(struct reloc_control *rc, 382a26195a5SQu Wenruo struct btrfs_backref_node *node) 38329db137bSQu Wenruo { 384a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 38529db137bSQu Wenruo struct list_head *useless_node = &cache->useless_node; 38629db137bSQu Wenruo bool ret = false; 38729db137bSQu Wenruo 38829db137bSQu Wenruo while (!list_empty(useless_node)) { 389a26195a5SQu Wenruo struct btrfs_backref_node *cur; 39029db137bSQu Wenruo 391a26195a5SQu Wenruo cur = list_first_entry(useless_node, struct btrfs_backref_node, 39229db137bSQu Wenruo list); 39329db137bSQu Wenruo list_del_init(&cur->list); 39429db137bSQu Wenruo 39529db137bSQu Wenruo /* Only tree root nodes can be added to @useless_nodes */ 39629db137bSQu Wenruo ASSERT(list_empty(&cur->upper)); 39729db137bSQu Wenruo 39829db137bSQu Wenruo if (cur == node) 39929db137bSQu Wenruo ret = true; 40029db137bSQu Wenruo 40129db137bSQu Wenruo /* The node is the lowest node */ 40229db137bSQu Wenruo if (cur->lowest) { 40329db137bSQu Wenruo list_del_init(&cur->lower); 40429db137bSQu Wenruo cur->lowest = 0; 40529db137bSQu Wenruo } 40629db137bSQu Wenruo 40729db137bSQu Wenruo /* Cleanup the lower edges */ 40829db137bSQu Wenruo while (!list_empty(&cur->lower)) { 409a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 410a26195a5SQu Wenruo struct btrfs_backref_node *lower; 41129db137bSQu Wenruo 41229db137bSQu Wenruo edge = list_entry(cur->lower.next, 413a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 41429db137bSQu Wenruo list_del(&edge->list[UPPER]); 41529db137bSQu Wenruo list_del(&edge->list[LOWER]); 41629db137bSQu Wenruo lower = edge->node[LOWER]; 417741188d3SQu Wenruo btrfs_backref_free_edge(cache, edge); 41829db137bSQu Wenruo 41929db137bSQu Wenruo /* Child node is also orphan, queue for cleanup */ 42029db137bSQu Wenruo if (list_empty(&lower->upper)) 42129db137bSQu Wenruo list_add(&lower->list, useless_node); 42229db137bSQu Wenruo } 42329db137bSQu Wenruo /* Mark this block processed for relocation */ 42429db137bSQu Wenruo mark_block_processed(rc, cur); 42529db137bSQu Wenruo 42629db137bSQu Wenruo /* 42729db137bSQu Wenruo * Backref nodes for tree leaves are deleted from the cache. 42829db137bSQu Wenruo * Backref nodes for upper level tree blocks are left in the 42929db137bSQu Wenruo * cache to avoid unnecessary backref lookup. 43029db137bSQu Wenruo */ 43129db137bSQu Wenruo if (cur->level > 0) { 43229db137bSQu Wenruo list_add(&cur->list, &cache->detached); 43329db137bSQu Wenruo cur->detached = 1; 43429db137bSQu Wenruo } else { 43529db137bSQu Wenruo rb_erase(&cur->rb_node, &cache->rb_root); 436741188d3SQu Wenruo btrfs_backref_free_node(cache, cur); 43729db137bSQu Wenruo } 43829db137bSQu Wenruo } 43929db137bSQu Wenruo return ret; 44029db137bSQu Wenruo } 44129db137bSQu Wenruo 44229db137bSQu Wenruo /* 443e7d571c7SQu Wenruo * Build backref tree for a given tree block. Root of the backref tree 444e7d571c7SQu Wenruo * corresponds the tree block, leaves of the backref tree correspond roots of 445e7d571c7SQu Wenruo * b-trees that reference the tree block. 446e7d571c7SQu Wenruo * 447e7d571c7SQu Wenruo * The basic idea of this function is check backrefs of a given block to find 448e7d571c7SQu Wenruo * upper level blocks that reference the block, and then check backrefs of 449e7d571c7SQu Wenruo * these upper level blocks recursively. The recursion stops when tree root is 450e7d571c7SQu Wenruo * reached or backrefs for the block is cached. 451e7d571c7SQu Wenruo * 452e7d571c7SQu Wenruo * NOTE: if we find that backrefs for a block are cached, we know backrefs for 453e7d571c7SQu Wenruo * all upper level blocks that directly/indirectly reference the block are also 454e7d571c7SQu Wenruo * cached. 455e7d571c7SQu Wenruo */ 456a26195a5SQu Wenruo static noinline_for_stack struct btrfs_backref_node *build_backref_tree( 457e7d571c7SQu Wenruo struct reloc_control *rc, struct btrfs_key *node_key, 458e7d571c7SQu Wenruo int level, u64 bytenr) 459e7d571c7SQu Wenruo { 460e7d571c7SQu Wenruo struct btrfs_backref_iter *iter; 461a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 462e7d571c7SQu Wenruo /* For searching parent of TREE_BLOCK_REF */ 463e7d571c7SQu Wenruo struct btrfs_path *path; 464a26195a5SQu Wenruo struct btrfs_backref_node *cur; 465a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 466a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 467e7d571c7SQu Wenruo int ret; 468e7d571c7SQu Wenruo int err = 0; 469e7d571c7SQu Wenruo 470e7d571c7SQu Wenruo iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS); 471e7d571c7SQu Wenruo if (!iter) 472e7d571c7SQu Wenruo return ERR_PTR(-ENOMEM); 473e7d571c7SQu Wenruo path = btrfs_alloc_path(); 474e7d571c7SQu Wenruo if (!path) { 475e7d571c7SQu Wenruo err = -ENOMEM; 476e7d571c7SQu Wenruo goto out; 477e7d571c7SQu Wenruo } 478e7d571c7SQu Wenruo 479b1818dabSQu Wenruo node = btrfs_backref_alloc_node(cache, bytenr, level); 480e7d571c7SQu Wenruo if (!node) { 481e7d571c7SQu Wenruo err = -ENOMEM; 482e7d571c7SQu Wenruo goto out; 483e7d571c7SQu Wenruo } 484e7d571c7SQu Wenruo 485e7d571c7SQu Wenruo node->lowest = 1; 486e7d571c7SQu Wenruo cur = node; 487e7d571c7SQu Wenruo 488e7d571c7SQu Wenruo /* Breadth-first search to build backref cache */ 489e7d571c7SQu Wenruo do { 4901b60d2ecSQu Wenruo ret = btrfs_backref_add_tree_node(cache, path, iter, node_key, 4911b60d2ecSQu Wenruo cur); 492e7d571c7SQu Wenruo if (ret < 0) { 493e7d571c7SQu Wenruo err = ret; 494e7d571c7SQu Wenruo goto out; 495e7d571c7SQu Wenruo } 496e7d571c7SQu Wenruo edge = list_first_entry_or_null(&cache->pending_edge, 497a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 498e7d571c7SQu Wenruo /* 499e7d571c7SQu Wenruo * The pending list isn't empty, take the first block to 500e7d571c7SQu Wenruo * process 501e7d571c7SQu Wenruo */ 502e7d571c7SQu Wenruo if (edge) { 5035d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 5045d4f98a2SYan Zheng cur = edge->node[UPPER]; 5055d4f98a2SYan Zheng } 506e7d571c7SQu Wenruo } while (edge); 5075d4f98a2SYan Zheng 5081f872924SQu Wenruo /* Finish the upper linkage of newly added edges/nodes */ 509fc997ed0SQu Wenruo ret = btrfs_backref_finish_upper_links(cache, node); 5101f872924SQu Wenruo if (ret < 0) { 5111f872924SQu Wenruo err = ret; 51275bfb9afSJosef Bacik goto out; 51375bfb9afSJosef Bacik } 51475bfb9afSJosef Bacik 51529db137bSQu Wenruo if (handle_useless_nodes(rc, node)) 5163fd0a558SYan, Zheng node = NULL; 5175d4f98a2SYan Zheng out: 51871f572a9SQu Wenruo btrfs_backref_iter_free(iter); 51971f572a9SQu Wenruo btrfs_free_path(path); 5205d4f98a2SYan Zheng if (err) { 5211b23ea18SQu Wenruo btrfs_backref_error_cleanup(cache, node); 5225d4f98a2SYan Zheng return ERR_PTR(err); 5235d4f98a2SYan Zheng } 52475bfb9afSJosef Bacik ASSERT(!node || !node->detached); 52584780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 52684780289SQu Wenruo list_empty(&cache->pending_edge)); 5275d4f98a2SYan Zheng return node; 5285d4f98a2SYan Zheng } 5295d4f98a2SYan Zheng 5305d4f98a2SYan Zheng /* 5313fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 5323fd0a558SYan, Zheng * the backref node is created by cloning backref node that 5333fd0a558SYan, Zheng * corresponds to root of source tree 5343fd0a558SYan, Zheng */ 5353fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 5363fd0a558SYan, Zheng struct reloc_control *rc, 5373fd0a558SYan, Zheng struct btrfs_root *src, 5383fd0a558SYan, Zheng struct btrfs_root *dest) 5393fd0a558SYan, Zheng { 5403fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 541a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 542a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 543a26195a5SQu Wenruo struct btrfs_backref_node *new_node; 544a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 545a26195a5SQu Wenruo struct btrfs_backref_edge *new_edge; 5463fd0a558SYan, Zheng struct rb_node *rb_node; 5473fd0a558SYan, Zheng 5483fd0a558SYan, Zheng if (cache->last_trans > 0) 5493fd0a558SYan, Zheng update_backref_cache(trans, cache); 5503fd0a558SYan, Zheng 551e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start); 5523fd0a558SYan, Zheng if (rb_node) { 553a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, rb_node); 5543fd0a558SYan, Zheng if (node->detached) 5553fd0a558SYan, Zheng node = NULL; 5563fd0a558SYan, Zheng else 5573fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 5583fd0a558SYan, Zheng } 5593fd0a558SYan, Zheng 5603fd0a558SYan, Zheng if (!node) { 561e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, 5623fd0a558SYan, Zheng reloc_root->commit_root->start); 5633fd0a558SYan, Zheng if (rb_node) { 564a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, 5653fd0a558SYan, Zheng rb_node); 5663fd0a558SYan, Zheng BUG_ON(node->detached); 5673fd0a558SYan, Zheng } 5683fd0a558SYan, Zheng } 5693fd0a558SYan, Zheng 5703fd0a558SYan, Zheng if (!node) 5713fd0a558SYan, Zheng return 0; 5723fd0a558SYan, Zheng 573b1818dabSQu Wenruo new_node = btrfs_backref_alloc_node(cache, dest->node->start, 574b1818dabSQu Wenruo node->level); 5753fd0a558SYan, Zheng if (!new_node) 5763fd0a558SYan, Zheng return -ENOMEM; 5773fd0a558SYan, Zheng 5783fd0a558SYan, Zheng new_node->lowest = node->lowest; 5796848ad64SYan, Zheng new_node->checked = 1; 58000246528SJosef Bacik new_node->root = btrfs_grab_root(dest); 5810b530bc5SJosef Bacik ASSERT(new_node->root); 5823fd0a558SYan, Zheng 5833fd0a558SYan, Zheng if (!node->lowest) { 5843fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 58547254d07SQu Wenruo new_edge = btrfs_backref_alloc_edge(cache); 5863fd0a558SYan, Zheng if (!new_edge) 5873fd0a558SYan, Zheng goto fail; 5883fd0a558SYan, Zheng 589f39911e5SQu Wenruo btrfs_backref_link_edge(new_edge, edge->node[LOWER], 590f39911e5SQu Wenruo new_node, LINK_UPPER); 5913fd0a558SYan, Zheng } 59276b9e23dSMiao Xie } else { 59376b9e23dSMiao Xie list_add_tail(&new_node->lower, &cache->leaves); 5943fd0a558SYan, Zheng } 5953fd0a558SYan, Zheng 596e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr, 5973fd0a558SYan, Zheng &new_node->rb_node); 59843c04fb1SJeff Mahoney if (rb_node) 599982c92cbSQu Wenruo btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST); 6003fd0a558SYan, Zheng 6013fd0a558SYan, Zheng if (!new_node->lowest) { 6023fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 6033fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 6043fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 6053fd0a558SYan, Zheng } 6063fd0a558SYan, Zheng } 6073fd0a558SYan, Zheng return 0; 6083fd0a558SYan, Zheng fail: 6093fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 6103fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 611a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 6123fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 613741188d3SQu Wenruo btrfs_backref_free_edge(cache, new_edge); 6143fd0a558SYan, Zheng } 615741188d3SQu Wenruo btrfs_backref_free_node(cache, new_node); 6163fd0a558SYan, Zheng return -ENOMEM; 6173fd0a558SYan, Zheng } 6183fd0a558SYan, Zheng 6193fd0a558SYan, Zheng /* 6205d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 6215d4f98a2SYan Zheng */ 622ffd7b339SJeff Mahoney static int __must_check __add_reloc_root(struct btrfs_root *root) 6235d4f98a2SYan Zheng { 6240b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 6255d4f98a2SYan Zheng struct rb_node *rb_node; 6265d4f98a2SYan Zheng struct mapping_node *node; 6270b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 6285d4f98a2SYan Zheng 6295d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 630ffd7b339SJeff Mahoney if (!node) 631ffd7b339SJeff Mahoney return -ENOMEM; 6325d4f98a2SYan Zheng 633ea287ab1SJosef Bacik node->bytenr = root->commit_root->start; 6345d4f98a2SYan Zheng node->data = root; 6355d4f98a2SYan Zheng 6365d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 637e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 6385d4f98a2SYan Zheng node->bytenr, &node->rb_node); 6395d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 640ffd7b339SJeff Mahoney if (rb_node) { 6410b246afaSJeff Mahoney btrfs_panic(fs_info, -EEXIST, 6425d163e0eSJeff Mahoney "Duplicate root found for start=%llu while inserting into relocation tree", 6435d163e0eSJeff Mahoney node->bytenr); 644ffd7b339SJeff Mahoney } 6455d4f98a2SYan Zheng 6465d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 6475d4f98a2SYan Zheng return 0; 6485d4f98a2SYan Zheng } 6495d4f98a2SYan Zheng 6505d4f98a2SYan Zheng /* 651c974c464SWang Shilong * helper to delete the 'address of tree root -> reloc tree' 6525d4f98a2SYan Zheng * mapping 6535d4f98a2SYan Zheng */ 654c974c464SWang Shilong static void __del_reloc_root(struct btrfs_root *root) 6555d4f98a2SYan Zheng { 6560b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 6575d4f98a2SYan Zheng struct rb_node *rb_node; 6585d4f98a2SYan Zheng struct mapping_node *node = NULL; 6590b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 660f44deb74SJosef Bacik bool put_ref = false; 6615d4f98a2SYan Zheng 66265c6e82bSQu Wenruo if (rc && root->node) { 6635d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 664e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 665ea287ab1SJosef Bacik root->commit_root->start); 666c974c464SWang Shilong if (rb_node) { 667c974c464SWang Shilong node = rb_entry(rb_node, struct mapping_node, rb_node); 668c974c464SWang Shilong rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 669ea287ab1SJosef Bacik RB_CLEAR_NODE(&node->rb_node); 670c974c464SWang Shilong } 671c974c464SWang Shilong spin_unlock(&rc->reloc_root_tree.lock); 672c78a10aeSJosef Bacik ASSERT(!node || (struct btrfs_root *)node->data == root); 673389305b2SQu Wenruo } 674c974c464SWang Shilong 675f44deb74SJosef Bacik /* 676f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 677f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 678f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 679f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 680f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 681f44deb74SJosef Bacik * messing with the list is in charge of the reference. 682f44deb74SJosef Bacik */ 6830b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 684f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 685f44deb74SJosef Bacik put_ref = true; 686c974c464SWang Shilong list_del_init(&root->root_list); 687f44deb74SJosef Bacik } 6880b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 689f44deb74SJosef Bacik if (put_ref) 690f44deb74SJosef Bacik btrfs_put_root(root); 691c974c464SWang Shilong kfree(node); 692c974c464SWang Shilong } 693c974c464SWang Shilong 694c974c464SWang Shilong /* 695c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 696c974c464SWang Shilong * mapping 697c974c464SWang Shilong */ 698ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 699c974c464SWang Shilong { 7000b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 701c974c464SWang Shilong struct rb_node *rb_node; 702c974c464SWang Shilong struct mapping_node *node = NULL; 7030b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 704c974c464SWang Shilong 705c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 706e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 707ea287ab1SJosef Bacik root->commit_root->start); 7085d4f98a2SYan Zheng if (rb_node) { 7095d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 7105d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 7115d4f98a2SYan Zheng } 7125d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 7135d4f98a2SYan Zheng 7148f71f3e0SLiu Bo if (!node) 7158f71f3e0SLiu Bo return 0; 7165d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 7175d4f98a2SYan Zheng 7185d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 719ea287ab1SJosef Bacik node->bytenr = root->node->start; 720e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 7215d4f98a2SYan Zheng node->bytenr, &node->rb_node); 7225d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 72343c04fb1SJeff Mahoney if (rb_node) 724982c92cbSQu Wenruo btrfs_backref_panic(fs_info, node->bytenr, -EEXIST); 7255d4f98a2SYan Zheng return 0; 7265d4f98a2SYan Zheng } 7275d4f98a2SYan Zheng 7283fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 7293fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 7305d4f98a2SYan Zheng { 7310b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 7325d4f98a2SYan Zheng struct btrfs_root *reloc_root; 7335d4f98a2SYan Zheng struct extent_buffer *eb; 7345d4f98a2SYan Zheng struct btrfs_root_item *root_item; 7355d4f98a2SYan Zheng struct btrfs_key root_key; 73684c50ba5SJosef Bacik int ret = 0; 73784c50ba5SJosef Bacik bool must_abort = false; 7385d4f98a2SYan Zheng 7395d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 74084c50ba5SJosef Bacik if (!root_item) 74184c50ba5SJosef Bacik return ERR_PTR(-ENOMEM); 7425d4f98a2SYan Zheng 7435d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 7445d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 7453fd0a558SYan, Zheng root_key.offset = objectid; 7465d4f98a2SYan Zheng 7473fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 748054570a1SFilipe Manana u64 commit_root_gen; 749054570a1SFilipe Manana 7503fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 7515d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 7525d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 75384c50ba5SJosef Bacik if (ret) 75484c50ba5SJosef Bacik goto fail; 75584c50ba5SJosef Bacik 756054570a1SFilipe Manana /* 757054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 758054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 759054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 760054570a1SFilipe Manana * either inside the critical section of a transaction commit 761054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 762054570a1SFilipe Manana * it's created before the transaction commit is started. 763054570a1SFilipe Manana */ 764054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 765054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 7663fd0a558SYan, Zheng } else { 7673fd0a558SYan, Zheng /* 7683fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 7693fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 7703fd0a558SYan, Zheng * modified after it was created have RELOC flag 7713fd0a558SYan, Zheng * set in their headers. so it's OK to not update 7723fd0a558SYan, Zheng * the 'last_snapshot'. 7733fd0a558SYan, Zheng */ 7743fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 7753fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 77684c50ba5SJosef Bacik if (ret) 77784c50ba5SJosef Bacik goto fail; 7783fd0a558SYan, Zheng } 7793fd0a558SYan, Zheng 78084c50ba5SJosef Bacik /* 78184c50ba5SJosef Bacik * We have changed references at this point, we must abort the 78284c50ba5SJosef Bacik * transaction if anything fails. 78384c50ba5SJosef Bacik */ 78484c50ba5SJosef Bacik must_abort = true; 78584c50ba5SJosef Bacik 7865d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 7875d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 7885d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 7895d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 7903fd0a558SYan, Zheng 7913fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 7923fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 7933fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 7943fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 795c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, 0); 7963fd0a558SYan, Zheng } 7975d4f98a2SYan Zheng 7985d4f98a2SYan Zheng btrfs_tree_unlock(eb); 7995d4f98a2SYan Zheng free_extent_buffer(eb); 8005d4f98a2SYan Zheng 8010b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 8025d4f98a2SYan Zheng &root_key, root_item); 80384c50ba5SJosef Bacik if (ret) 80484c50ba5SJosef Bacik goto fail; 80584c50ba5SJosef Bacik 8065d4f98a2SYan Zheng kfree(root_item); 8075d4f98a2SYan Zheng 8083dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 80984c50ba5SJosef Bacik if (IS_ERR(reloc_root)) { 81084c50ba5SJosef Bacik ret = PTR_ERR(reloc_root); 81184c50ba5SJosef Bacik goto abort; 81284c50ba5SJosef Bacik } 81392a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 8145d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 8153fd0a558SYan, Zheng return reloc_root; 81684c50ba5SJosef Bacik fail: 81784c50ba5SJosef Bacik kfree(root_item); 81884c50ba5SJosef Bacik abort: 81984c50ba5SJosef Bacik if (must_abort) 82084c50ba5SJosef Bacik btrfs_abort_transaction(trans, ret); 82184c50ba5SJosef Bacik return ERR_PTR(ret); 8223fd0a558SYan, Zheng } 8233fd0a558SYan, Zheng 8243fd0a558SYan, Zheng /* 8253fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 8263fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 827f44deb74SJosef Bacik * 828f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 829f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 8303fd0a558SYan, Zheng */ 8313fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 8323fd0a558SYan, Zheng struct btrfs_root *root) 8333fd0a558SYan, Zheng { 8340b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 8353fd0a558SYan, Zheng struct btrfs_root *reloc_root; 8360b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 83720dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 8383fd0a558SYan, Zheng int clear_rsv = 0; 839ffd7b339SJeff Mahoney int ret; 8403fd0a558SYan, Zheng 841aec7db3bSJosef Bacik if (!rc) 8422abc726aSJosef Bacik return 0; 8432abc726aSJosef Bacik 8441fac4a54SQu Wenruo /* 8451fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 8461fac4a54SQu Wenruo * create/update the dead reloc tree 8471fac4a54SQu Wenruo */ 8486282675eSQu Wenruo if (reloc_root_is_dead(root)) 8491fac4a54SQu Wenruo return 0; 8501fac4a54SQu Wenruo 851aec7db3bSJosef Bacik /* 852aec7db3bSJosef Bacik * This is subtle but important. We do not do 853aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 854aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 855aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 856aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 857aec7db3bSJosef Bacik * in. 858aec7db3bSJosef Bacik */ 8593fd0a558SYan, Zheng if (root->reloc_root) { 8603fd0a558SYan, Zheng reloc_root = root->reloc_root; 8613fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 8623fd0a558SYan, Zheng return 0; 8633fd0a558SYan, Zheng } 8643fd0a558SYan, Zheng 865aec7db3bSJosef Bacik /* 866aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 867aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 868aec7db3bSJosef Bacik */ 869aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 870aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 871aec7db3bSJosef Bacik return 0; 872aec7db3bSJosef Bacik 87320dd2cbfSMiao Xie if (!trans->reloc_reserved) { 87420dd2cbfSMiao Xie rsv = trans->block_rsv; 8753fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 8763fd0a558SYan, Zheng clear_rsv = 1; 8773fd0a558SYan, Zheng } 8783fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 8793fd0a558SYan, Zheng if (clear_rsv) 88020dd2cbfSMiao Xie trans->block_rsv = rsv; 88100bb36a0SJosef Bacik if (IS_ERR(reloc_root)) 88200bb36a0SJosef Bacik return PTR_ERR(reloc_root); 8835d4f98a2SYan Zheng 884ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 88500bb36a0SJosef Bacik if (ret) { 88600bb36a0SJosef Bacik /* Pairs with create_reloc_root */ 88700bb36a0SJosef Bacik btrfs_put_root(reloc_root); 88800bb36a0SJosef Bacik return ret; 88900bb36a0SJosef Bacik } 890f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 8915d4f98a2SYan Zheng return 0; 8925d4f98a2SYan Zheng } 8935d4f98a2SYan Zheng 8945d4f98a2SYan Zheng /* 8955d4f98a2SYan Zheng * update root item of reloc tree 8965d4f98a2SYan Zheng */ 8975d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 8985d4f98a2SYan Zheng struct btrfs_root *root) 8995d4f98a2SYan Zheng { 9000b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 9015d4f98a2SYan Zheng struct btrfs_root *reloc_root; 9025d4f98a2SYan Zheng struct btrfs_root_item *root_item; 9035d4f98a2SYan Zheng int ret; 9045d4f98a2SYan Zheng 9056282675eSQu Wenruo if (!have_reloc_root(root)) 9067585717fSChris Mason goto out; 9075d4f98a2SYan Zheng 9085d4f98a2SYan Zheng reloc_root = root->reloc_root; 9095d4f98a2SYan Zheng root_item = &reloc_root->root_item; 9105d4f98a2SYan Zheng 911f44deb74SJosef Bacik /* 912f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 913f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 914f44deb74SJosef Bacik * hold it while we update the reloc root. 915f44deb74SJosef Bacik */ 916f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 917f44deb74SJosef Bacik 918d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 9190b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 9203fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 921d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 9226282675eSQu Wenruo /* 9236282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 9246282675eSQu Wenruo * have_reloc_root will not touch it from now on. 9256282675eSQu Wenruo */ 9266282675eSQu Wenruo smp_wmb(); 927c974c464SWang Shilong __del_reloc_root(reloc_root); 9285d4f98a2SYan Zheng } 9295d4f98a2SYan Zheng 9305d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 931ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 9325d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 9335d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 9345d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 9355d4f98a2SYan Zheng } 9365d4f98a2SYan Zheng 9370b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 9385d4f98a2SYan Zheng &reloc_root->root_key, root_item); 9395d4f98a2SYan Zheng BUG_ON(ret); 940f44deb74SJosef Bacik btrfs_put_root(reloc_root); 9417585717fSChris Mason out: 9425d4f98a2SYan Zheng return 0; 9435d4f98a2SYan Zheng } 9445d4f98a2SYan Zheng 9455d4f98a2SYan Zheng /* 9465d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 9475d4f98a2SYan Zheng * in a subvolume 9485d4f98a2SYan Zheng */ 9495d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 9505d4f98a2SYan Zheng { 9515d4f98a2SYan Zheng struct rb_node *node; 9525d4f98a2SYan Zheng struct rb_node *prev; 9535d4f98a2SYan Zheng struct btrfs_inode *entry; 9545d4f98a2SYan Zheng struct inode *inode; 9555d4f98a2SYan Zheng 9565d4f98a2SYan Zheng spin_lock(&root->inode_lock); 9575d4f98a2SYan Zheng again: 9585d4f98a2SYan Zheng node = root->inode_tree.rb_node; 9595d4f98a2SYan Zheng prev = NULL; 9605d4f98a2SYan Zheng while (node) { 9615d4f98a2SYan Zheng prev = node; 9625d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9635d4f98a2SYan Zheng 9644a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 9655d4f98a2SYan Zheng node = node->rb_left; 9664a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 9675d4f98a2SYan Zheng node = node->rb_right; 9685d4f98a2SYan Zheng else 9695d4f98a2SYan Zheng break; 9705d4f98a2SYan Zheng } 9715d4f98a2SYan Zheng if (!node) { 9725d4f98a2SYan Zheng while (prev) { 9735d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 9744a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 9755d4f98a2SYan Zheng node = prev; 9765d4f98a2SYan Zheng break; 9775d4f98a2SYan Zheng } 9785d4f98a2SYan Zheng prev = rb_next(prev); 9795d4f98a2SYan Zheng } 9805d4f98a2SYan Zheng } 9815d4f98a2SYan Zheng while (node) { 9825d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9835d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 9845d4f98a2SYan Zheng if (inode) { 9855d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9865d4f98a2SYan Zheng return inode; 9875d4f98a2SYan Zheng } 9885d4f98a2SYan Zheng 9894a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 9905d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 9915d4f98a2SYan Zheng goto again; 9925d4f98a2SYan Zheng 9935d4f98a2SYan Zheng node = rb_next(node); 9945d4f98a2SYan Zheng } 9955d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9965d4f98a2SYan Zheng return NULL; 9975d4f98a2SYan Zheng } 9985d4f98a2SYan Zheng 9995d4f98a2SYan Zheng /* 10005d4f98a2SYan Zheng * get new location of data 10015d4f98a2SYan Zheng */ 10025d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 10035d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 10045d4f98a2SYan Zheng { 10055d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 10065d4f98a2SYan Zheng struct btrfs_path *path; 10075d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10085d4f98a2SYan Zheng struct extent_buffer *leaf; 10095d4f98a2SYan Zheng int ret; 10105d4f98a2SYan Zheng 10115d4f98a2SYan Zheng path = btrfs_alloc_path(); 10125d4f98a2SYan Zheng if (!path) 10135d4f98a2SYan Zheng return -ENOMEM; 10145d4f98a2SYan Zheng 10155d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 1016f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 1017f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 10185d4f98a2SYan Zheng if (ret < 0) 10195d4f98a2SYan Zheng goto out; 10205d4f98a2SYan Zheng if (ret > 0) { 10215d4f98a2SYan Zheng ret = -ENOENT; 10225d4f98a2SYan Zheng goto out; 10235d4f98a2SYan Zheng } 10245d4f98a2SYan Zheng 10255d4f98a2SYan Zheng leaf = path->nodes[0]; 10265d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 10275d4f98a2SYan Zheng struct btrfs_file_extent_item); 10285d4f98a2SYan Zheng 10295d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 10305d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 10315d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 10325d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 10335d4f98a2SYan Zheng 10345d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 103583d4cfd4SJosef Bacik ret = -EINVAL; 10365d4f98a2SYan Zheng goto out; 10375d4f98a2SYan Zheng } 10385d4f98a2SYan Zheng 10395d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10405d4f98a2SYan Zheng ret = 0; 10415d4f98a2SYan Zheng out: 10425d4f98a2SYan Zheng btrfs_free_path(path); 10435d4f98a2SYan Zheng return ret; 10445d4f98a2SYan Zheng } 10455d4f98a2SYan Zheng 10465d4f98a2SYan Zheng /* 10475d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 10485d4f98a2SYan Zheng * the new locations. 10495d4f98a2SYan Zheng */ 10503fd0a558SYan, Zheng static noinline_for_stack 10513fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 10525d4f98a2SYan Zheng struct reloc_control *rc, 10535d4f98a2SYan Zheng struct btrfs_root *root, 10543fd0a558SYan, Zheng struct extent_buffer *leaf) 10555d4f98a2SYan Zheng { 10560b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 10575d4f98a2SYan Zheng struct btrfs_key key; 10585d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10595d4f98a2SYan Zheng struct inode *inode = NULL; 10605d4f98a2SYan Zheng u64 parent; 10615d4f98a2SYan Zheng u64 bytenr; 10623fd0a558SYan, Zheng u64 new_bytenr = 0; 10635d4f98a2SYan Zheng u64 num_bytes; 10645d4f98a2SYan Zheng u64 end; 10655d4f98a2SYan Zheng u32 nritems; 10665d4f98a2SYan Zheng u32 i; 106783d4cfd4SJosef Bacik int ret = 0; 10685d4f98a2SYan Zheng int first = 1; 10695d4f98a2SYan Zheng int dirty = 0; 10705d4f98a2SYan Zheng 10715d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 10725d4f98a2SYan Zheng return 0; 10735d4f98a2SYan Zheng 10745d4f98a2SYan Zheng /* reloc trees always use full backref */ 10755d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 10765d4f98a2SYan Zheng parent = leaf->start; 10775d4f98a2SYan Zheng else 10785d4f98a2SYan Zheng parent = 0; 10795d4f98a2SYan Zheng 10805d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 10815d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 108282fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 108382fa113fSQu Wenruo 10845d4f98a2SYan Zheng cond_resched(); 10855d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 10865d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 10875d4f98a2SYan Zheng continue; 10885d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 10895d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 10905d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 10915d4f98a2SYan Zheng continue; 10925d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10935d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 10945d4f98a2SYan Zheng if (bytenr == 0) 10955d4f98a2SYan Zheng continue; 10969569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 10979569cc20SQu Wenruo rc->block_group->length)) 10985d4f98a2SYan Zheng continue; 10995d4f98a2SYan Zheng 11005d4f98a2SYan Zheng /* 11015d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 11025d4f98a2SYan Zheng * to complete and drop the extent cache 11035d4f98a2SYan Zheng */ 11045d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 11055d4f98a2SYan Zheng if (first) { 11065d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11075d4f98a2SYan Zheng first = 0; 11084a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 11093fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 11105d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11115d4f98a2SYan Zheng } 11124a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 11135d4f98a2SYan Zheng end = key.offset + 11145d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 11155d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 11160b246afaSJeff Mahoney fs_info->sectorsize)); 11170b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 11185d4f98a2SYan Zheng end--; 11195d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1120d0082371SJeff Mahoney key.offset, end); 11215d4f98a2SYan Zheng if (!ret) 11225d4f98a2SYan Zheng continue; 11235d4f98a2SYan Zheng 1124dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1125dcdbc059SNikolay Borisov key.offset, end, 1); 11265d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1127d0082371SJeff Mahoney key.offset, end); 11285d4f98a2SYan Zheng } 11295d4f98a2SYan Zheng } 11305d4f98a2SYan Zheng 11315d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 11325d4f98a2SYan Zheng bytenr, num_bytes); 113383d4cfd4SJosef Bacik if (ret) { 113483d4cfd4SJosef Bacik /* 113583d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 113683d4cfd4SJosef Bacik * in the file extent yet. 113783d4cfd4SJosef Bacik */ 113883d4cfd4SJosef Bacik break; 11393fd0a558SYan, Zheng } 11405d4f98a2SYan Zheng 11415d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 11425d4f98a2SYan Zheng dirty = 1; 11435d4f98a2SYan Zheng 11445d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 114582fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 114682fa113fSQu Wenruo num_bytes, parent); 114782fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 114882fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1149b06c4bf5SFilipe Manana key.objectid, key.offset); 115082fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 115183d4cfd4SJosef Bacik if (ret) { 115266642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 115383d4cfd4SJosef Bacik break; 115483d4cfd4SJosef Bacik } 11555d4f98a2SYan Zheng 1156ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1157ffd4bb2aSQu Wenruo num_bytes, parent); 1158ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1159ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1160b06c4bf5SFilipe Manana key.objectid, key.offset); 1161ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 116283d4cfd4SJosef Bacik if (ret) { 116366642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 116483d4cfd4SJosef Bacik break; 116583d4cfd4SJosef Bacik } 11665d4f98a2SYan Zheng } 11675d4f98a2SYan Zheng if (dirty) 11685d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 11693fd0a558SYan, Zheng if (inode) 11703fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 117183d4cfd4SJosef Bacik return ret; 11725d4f98a2SYan Zheng } 11735d4f98a2SYan Zheng 11745d4f98a2SYan Zheng static noinline_for_stack 11755d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 11765d4f98a2SYan Zheng struct btrfs_path *path, int level) 11775d4f98a2SYan Zheng { 11785d4f98a2SYan Zheng struct btrfs_disk_key key1; 11795d4f98a2SYan Zheng struct btrfs_disk_key key2; 11805d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 11815d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 11825d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 11835d4f98a2SYan Zheng } 11845d4f98a2SYan Zheng 11855d4f98a2SYan Zheng /* 11865d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 11875d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 11885d4f98a2SYan Zheng * reloc tree was create can be replaced. 11895d4f98a2SYan Zheng * 11905d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 11915d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 11925d4f98a2SYan Zheng * errors, a negative error number is returned. 11935d4f98a2SYan Zheng */ 11943fd0a558SYan, Zheng static noinline_for_stack 11953d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 11965d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 11975d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 11985d4f98a2SYan Zheng int lowest_level, int max_level) 11995d4f98a2SYan Zheng { 12000b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 12015d4f98a2SYan Zheng struct extent_buffer *eb; 12025d4f98a2SYan Zheng struct extent_buffer *parent; 120382fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 12045d4f98a2SYan Zheng struct btrfs_key key; 12055d4f98a2SYan Zheng u64 old_bytenr; 12065d4f98a2SYan Zheng u64 new_bytenr; 12075d4f98a2SYan Zheng u64 old_ptr_gen; 12085d4f98a2SYan Zheng u64 new_ptr_gen; 12095d4f98a2SYan Zheng u64 last_snapshot; 12105d4f98a2SYan Zheng u32 blocksize; 12113fd0a558SYan, Zheng int cow = 0; 12125d4f98a2SYan Zheng int level; 12135d4f98a2SYan Zheng int ret; 12145d4f98a2SYan Zheng int slot; 12155d4f98a2SYan Zheng 12165d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 12175d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 12185d4f98a2SYan Zheng 12195d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 12203fd0a558SYan, Zheng again: 12215d4f98a2SYan Zheng slot = path->slots[lowest_level]; 12225d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 12235d4f98a2SYan Zheng 12245d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 12255d4f98a2SYan Zheng level = btrfs_header_level(eb); 12265d4f98a2SYan Zheng 12275d4f98a2SYan Zheng if (level < lowest_level) { 12285d4f98a2SYan Zheng btrfs_tree_unlock(eb); 12295d4f98a2SYan Zheng free_extent_buffer(eb); 12305d4f98a2SYan Zheng return 0; 12315d4f98a2SYan Zheng } 12325d4f98a2SYan Zheng 12333fd0a558SYan, Zheng if (cow) { 12349631e4ccSJosef Bacik ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb, 12359631e4ccSJosef Bacik BTRFS_NESTING_COW); 12365d4f98a2SYan Zheng BUG_ON(ret); 12373fd0a558SYan, Zheng } 12385d4f98a2SYan Zheng 12395d4f98a2SYan Zheng if (next_key) { 12405d4f98a2SYan Zheng next_key->objectid = (u64)-1; 12415d4f98a2SYan Zheng next_key->type = (u8)-1; 12425d4f98a2SYan Zheng next_key->offset = (u64)-1; 12435d4f98a2SYan Zheng } 12445d4f98a2SYan Zheng 12455d4f98a2SYan Zheng parent = eb; 12465d4f98a2SYan Zheng while (1) { 12475d4f98a2SYan Zheng level = btrfs_header_level(parent); 12485d4f98a2SYan Zheng BUG_ON(level < lowest_level); 12495d4f98a2SYan Zheng 1250e3b83361SQu Wenruo ret = btrfs_bin_search(parent, &key, &slot); 1251cbca7d59SFilipe Manana if (ret < 0) 1252cbca7d59SFilipe Manana break; 12535d4f98a2SYan Zheng if (ret && slot > 0) 12545d4f98a2SYan Zheng slot--; 12555d4f98a2SYan Zheng 12565d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 12575d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 12585d4f98a2SYan Zheng 12595d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 12600b246afaSJeff Mahoney blocksize = fs_info->nodesize; 12615d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 12625d4f98a2SYan Zheng 12635d4f98a2SYan Zheng if (level <= max_level) { 12645d4f98a2SYan Zheng eb = path->nodes[level]; 12655d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 12665d4f98a2SYan Zheng path->slots[level]); 12675d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 12685d4f98a2SYan Zheng path->slots[level]); 12695d4f98a2SYan Zheng } else { 12705d4f98a2SYan Zheng new_bytenr = 0; 12715d4f98a2SYan Zheng new_ptr_gen = 0; 12725d4f98a2SYan Zheng } 12735d4f98a2SYan Zheng 1274fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 12755d4f98a2SYan Zheng ret = level; 12765d4f98a2SYan Zheng break; 12775d4f98a2SYan Zheng } 12785d4f98a2SYan Zheng 12795d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 12805d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 12813fd0a558SYan, Zheng if (level <= lowest_level) { 12825d4f98a2SYan Zheng ret = 0; 12835d4f98a2SYan Zheng break; 12845d4f98a2SYan Zheng } 12855d4f98a2SYan Zheng 12866b3426beSJosef Bacik eb = btrfs_read_node_slot(parent, slot); 128764c043deSLiu Bo if (IS_ERR(eb)) { 128864c043deSLiu Bo ret = PTR_ERR(eb); 1289264813acSLiu Bo break; 1290416bc658SJosef Bacik } 12915d4f98a2SYan Zheng btrfs_tree_lock(eb); 12923fd0a558SYan, Zheng if (cow) { 12935d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 12949631e4ccSJosef Bacik slot, &eb, 12959631e4ccSJosef Bacik BTRFS_NESTING_COW); 12965d4f98a2SYan Zheng BUG_ON(ret); 12975d4f98a2SYan Zheng } 12985d4f98a2SYan Zheng 12995d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13005d4f98a2SYan Zheng free_extent_buffer(parent); 13015d4f98a2SYan Zheng 13025d4f98a2SYan Zheng parent = eb; 13035d4f98a2SYan Zheng continue; 13045d4f98a2SYan Zheng } 13055d4f98a2SYan Zheng 13063fd0a558SYan, Zheng if (!cow) { 13073fd0a558SYan, Zheng btrfs_tree_unlock(parent); 13083fd0a558SYan, Zheng free_extent_buffer(parent); 13093fd0a558SYan, Zheng cow = 1; 13103fd0a558SYan, Zheng goto again; 13113fd0a558SYan, Zheng } 13123fd0a558SYan, Zheng 13135d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 13145d4f98a2SYan Zheng path->slots[level]); 1315b3b4aa74SDavid Sterba btrfs_release_path(path); 13165d4f98a2SYan Zheng 13175d4f98a2SYan Zheng path->lowest_level = level; 13185d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 13195d4f98a2SYan Zheng path->lowest_level = 0; 13205d4f98a2SYan Zheng BUG_ON(ret); 13215d4f98a2SYan Zheng 13225d4f98a2SYan Zheng /* 1323824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 1324824d8dffSQu Wenruo * 1325824d8dffSQu Wenruo * We must trace both trees. 1326824d8dffSQu Wenruo * 1) Tree reloc subtree 1327824d8dffSQu Wenruo * If not traced, we will leak data numbers 1328824d8dffSQu Wenruo * 2) Fs subtree 1329824d8dffSQu Wenruo * If not traced, we will double count old data 1330f616f5cdSQu Wenruo * 1331f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 1332f616f5cdSQu Wenruo * the swapped tree blocks. 1333f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 1334f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 1335824d8dffSQu Wenruo */ 1336370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 1337370a11b8SQu Wenruo rc->block_group, parent, slot, 1338370a11b8SQu Wenruo path->nodes[level], path->slots[level], 1339370a11b8SQu Wenruo last_snapshot); 1340370a11b8SQu Wenruo if (ret < 0) 1341370a11b8SQu Wenruo break; 1342824d8dffSQu Wenruo /* 13435d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 13445d4f98a2SYan Zheng */ 13455d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 13465d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 13475d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 13485d4f98a2SYan Zheng 13495d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 13505d4f98a2SYan Zheng path->slots[level], old_bytenr); 13515d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 13525d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 13535d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 13545d4f98a2SYan Zheng 135582fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 135682fa113fSQu Wenruo blocksize, path->nodes[level]->start); 135782fa113fSQu Wenruo ref.skip_qgroup = true; 135882fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 135982fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13605d4f98a2SYan Zheng BUG_ON(ret); 136182fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 136282fa113fSQu Wenruo blocksize, 0); 136382fa113fSQu Wenruo ref.skip_qgroup = true; 136482fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 136582fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13665d4f98a2SYan Zheng BUG_ON(ret); 13675d4f98a2SYan Zheng 1368ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 1369ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 1370ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 1371ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1372ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13735d4f98a2SYan Zheng BUG_ON(ret); 13745d4f98a2SYan Zheng 1375ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 1376ffd4bb2aSQu Wenruo blocksize, 0); 1377ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 1378ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1379ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13805d4f98a2SYan Zheng BUG_ON(ret); 13815d4f98a2SYan Zheng 13825d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 13835d4f98a2SYan Zheng 13845d4f98a2SYan Zheng ret = level; 13855d4f98a2SYan Zheng break; 13865d4f98a2SYan Zheng } 13875d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13885d4f98a2SYan Zheng free_extent_buffer(parent); 13895d4f98a2SYan Zheng return ret; 13905d4f98a2SYan Zheng } 13915d4f98a2SYan Zheng 13925d4f98a2SYan Zheng /* 13935d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 13945d4f98a2SYan Zheng */ 13955d4f98a2SYan Zheng static noinline_for_stack 13965d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 13975d4f98a2SYan Zheng int *level) 13985d4f98a2SYan Zheng { 13995d4f98a2SYan Zheng struct extent_buffer *eb; 14005d4f98a2SYan Zheng int i; 14015d4f98a2SYan Zheng u64 last_snapshot; 14025d4f98a2SYan Zheng u32 nritems; 14035d4f98a2SYan Zheng 14045d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14055d4f98a2SYan Zheng 14065d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 14075d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14085d4f98a2SYan Zheng path->nodes[i] = NULL; 14095d4f98a2SYan Zheng } 14105d4f98a2SYan Zheng 14115d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 14125d4f98a2SYan Zheng eb = path->nodes[i]; 14135d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14145d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 14155d4f98a2SYan Zheng path->slots[i]++; 14165d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 14175d4f98a2SYan Zheng last_snapshot) 14185d4f98a2SYan Zheng continue; 14195d4f98a2SYan Zheng 14205d4f98a2SYan Zheng *level = i; 14215d4f98a2SYan Zheng return 0; 14225d4f98a2SYan Zheng } 14235d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14245d4f98a2SYan Zheng path->nodes[i] = NULL; 14255d4f98a2SYan Zheng } 14265d4f98a2SYan Zheng return 1; 14275d4f98a2SYan Zheng } 14285d4f98a2SYan Zheng 14295d4f98a2SYan Zheng /* 14305d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 14315d4f98a2SYan Zheng */ 14325d4f98a2SYan Zheng static noinline_for_stack 14335d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14345d4f98a2SYan Zheng int *level) 14355d4f98a2SYan Zheng { 14365d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 14375d4f98a2SYan Zheng int i; 14385d4f98a2SYan Zheng u64 ptr_gen = 0; 14395d4f98a2SYan Zheng u64 last_snapshot; 14405d4f98a2SYan Zheng u32 nritems; 14415d4f98a2SYan Zheng 14425d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14435d4f98a2SYan Zheng 14445d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 14455d4f98a2SYan Zheng eb = path->nodes[i]; 14465d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14475d4f98a2SYan Zheng while (path->slots[i] < nritems) { 14485d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 14495d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 14505d4f98a2SYan Zheng break; 14515d4f98a2SYan Zheng path->slots[i]++; 14525d4f98a2SYan Zheng } 14535d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 14545d4f98a2SYan Zheng if (i == *level) 14555d4f98a2SYan Zheng break; 14565d4f98a2SYan Zheng *level = i + 1; 14575d4f98a2SYan Zheng return 0; 14585d4f98a2SYan Zheng } 14595d4f98a2SYan Zheng if (i == 1) { 14605d4f98a2SYan Zheng *level = i; 14615d4f98a2SYan Zheng return 0; 14625d4f98a2SYan Zheng } 14635d4f98a2SYan Zheng 14648ef385bbSJosef Bacik eb = btrfs_read_node_slot(eb, path->slots[i]); 14658ef385bbSJosef Bacik if (IS_ERR(eb)) 146664c043deSLiu Bo return PTR_ERR(eb); 14675d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 14685d4f98a2SYan Zheng path->nodes[i - 1] = eb; 14695d4f98a2SYan Zheng path->slots[i - 1] = 0; 14705d4f98a2SYan Zheng } 14715d4f98a2SYan Zheng return 1; 14725d4f98a2SYan Zheng } 14735d4f98a2SYan Zheng 14745d4f98a2SYan Zheng /* 14755d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 14765d4f98a2SYan Zheng * [min_key, max_key) 14775d4f98a2SYan Zheng */ 14785d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 14795d4f98a2SYan Zheng struct btrfs_key *min_key, 14805d4f98a2SYan Zheng struct btrfs_key *max_key) 14815d4f98a2SYan Zheng { 14820b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14835d4f98a2SYan Zheng struct inode *inode = NULL; 14845d4f98a2SYan Zheng u64 objectid; 14855d4f98a2SYan Zheng u64 start, end; 148633345d01SLi Zefan u64 ino; 14875d4f98a2SYan Zheng 14885d4f98a2SYan Zheng objectid = min_key->objectid; 14895d4f98a2SYan Zheng while (1) { 14905d4f98a2SYan Zheng cond_resched(); 14915d4f98a2SYan Zheng iput(inode); 14925d4f98a2SYan Zheng 14935d4f98a2SYan Zheng if (objectid > max_key->objectid) 14945d4f98a2SYan Zheng break; 14955d4f98a2SYan Zheng 14965d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 14975d4f98a2SYan Zheng if (!inode) 14985d4f98a2SYan Zheng break; 14994a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 15005d4f98a2SYan Zheng 150133345d01SLi Zefan if (ino > max_key->objectid) { 15025d4f98a2SYan Zheng iput(inode); 15035d4f98a2SYan Zheng break; 15045d4f98a2SYan Zheng } 15055d4f98a2SYan Zheng 150633345d01SLi Zefan objectid = ino + 1; 15075d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 15085d4f98a2SYan Zheng continue; 15095d4f98a2SYan Zheng 151033345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 15115d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 15125d4f98a2SYan Zheng continue; 15135d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 15145d4f98a2SYan Zheng start = 0; 15155d4f98a2SYan Zheng else { 15165d4f98a2SYan Zheng start = min_key->offset; 15170b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 15185d4f98a2SYan Zheng } 15195d4f98a2SYan Zheng } else { 15205d4f98a2SYan Zheng start = 0; 15215d4f98a2SYan Zheng } 15225d4f98a2SYan Zheng 152333345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 15245d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 15255d4f98a2SYan Zheng continue; 15265d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 15275d4f98a2SYan Zheng end = (u64)-1; 15285d4f98a2SYan Zheng } else { 15295d4f98a2SYan Zheng if (max_key->offset == 0) 15305d4f98a2SYan Zheng continue; 15315d4f98a2SYan Zheng end = max_key->offset; 15320b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 15335d4f98a2SYan Zheng end--; 15345d4f98a2SYan Zheng } 15355d4f98a2SYan Zheng } else { 15365d4f98a2SYan Zheng end = (u64)-1; 15375d4f98a2SYan Zheng } 15385d4f98a2SYan Zheng 15395d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 1540d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 1541dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 1542d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 15435d4f98a2SYan Zheng } 15445d4f98a2SYan Zheng return 0; 15455d4f98a2SYan Zheng } 15465d4f98a2SYan Zheng 15475d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 15485d4f98a2SYan Zheng struct btrfs_key *key) 15495d4f98a2SYan Zheng 15505d4f98a2SYan Zheng { 15515d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 15525d4f98a2SYan Zheng if (!path->nodes[level]) 15535d4f98a2SYan Zheng break; 15545d4f98a2SYan Zheng if (path->slots[level] + 1 < 15555d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 15565d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 15575d4f98a2SYan Zheng path->slots[level] + 1); 15585d4f98a2SYan Zheng return 0; 15595d4f98a2SYan Zheng } 15605d4f98a2SYan Zheng level++; 15615d4f98a2SYan Zheng } 15625d4f98a2SYan Zheng return 1; 15635d4f98a2SYan Zheng } 15645d4f98a2SYan Zheng 15655d4f98a2SYan Zheng /* 1566d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 1567d2311e69SQu Wenruo */ 1568ac54da6cSJosef Bacik static int insert_dirty_subvol(struct btrfs_trans_handle *trans, 1569d2311e69SQu Wenruo struct reloc_control *rc, 1570d2311e69SQu Wenruo struct btrfs_root *root) 1571d2311e69SQu Wenruo { 1572d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1573d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 1574*7934133fSJosef Bacik int ret; 1575d2311e69SQu Wenruo 1576d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 1577d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 1578d2311e69SQu Wenruo ASSERT(reloc_root); 1579d2311e69SQu Wenruo 1580d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 1581d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 1582d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 1583c8422684SDavid Sterba btrfs_set_root_drop_level(reloc_root_item, 0); 1584d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 1585*7934133fSJosef Bacik ret = btrfs_update_reloc_root(trans, root); 1586*7934133fSJosef Bacik if (ret) 1587*7934133fSJosef Bacik return ret; 1588d2311e69SQu Wenruo 1589d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 159000246528SJosef Bacik btrfs_grab_root(root); 1591d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 1592d2311e69SQu Wenruo } 1593ac54da6cSJosef Bacik 1594ac54da6cSJosef Bacik return 0; 1595d2311e69SQu Wenruo } 1596d2311e69SQu Wenruo 1597d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 1598d2311e69SQu Wenruo { 1599d2311e69SQu Wenruo struct btrfs_root *root; 1600d2311e69SQu Wenruo struct btrfs_root *next; 1601d2311e69SQu Wenruo int ret = 0; 160230d40577SQu Wenruo int ret2; 1603d2311e69SQu Wenruo 1604d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 1605d2311e69SQu Wenruo reloc_dirty_list) { 160630d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 160730d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 1608d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1609d2311e69SQu Wenruo 1610d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 1611d2311e69SQu Wenruo root->reloc_root = NULL; 16126282675eSQu Wenruo /* 16136282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 16146282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 16156282675eSQu Wenruo */ 16166282675eSQu Wenruo smp_wmb(); 16171fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 1618f28de8d8SJosef Bacik if (reloc_root) { 1619f44deb74SJosef Bacik /* 1620f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 1621f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 1622f44deb74SJosef Bacik * drop the ref ourselves. 1623f44deb74SJosef Bacik */ 1624f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 1625f44deb74SJosef Bacik if (ret2 < 0) { 1626f44deb74SJosef Bacik btrfs_put_root(reloc_root); 1627f44deb74SJosef Bacik if (!ret) 1628f28de8d8SJosef Bacik ret = ret2; 1629f28de8d8SJosef Bacik } 1630f44deb74SJosef Bacik } 163100246528SJosef Bacik btrfs_put_root(root); 163230d40577SQu Wenruo } else { 163330d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 16340078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 1635f44deb74SJosef Bacik if (ret2 < 0) { 1636f44deb74SJosef Bacik btrfs_put_root(root); 1637f44deb74SJosef Bacik if (!ret) 163830d40577SQu Wenruo ret = ret2; 163930d40577SQu Wenruo } 1640d2311e69SQu Wenruo } 1641f44deb74SJosef Bacik } 1642d2311e69SQu Wenruo return ret; 1643d2311e69SQu Wenruo } 1644d2311e69SQu Wenruo 1645d2311e69SQu Wenruo /* 16465d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 16475d4f98a2SYan Zheng * fs tree. 16485d4f98a2SYan Zheng */ 16495d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 16505d4f98a2SYan Zheng struct btrfs_root *root) 16515d4f98a2SYan Zheng { 16520b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 16535d4f98a2SYan Zheng struct btrfs_key key; 16545d4f98a2SYan Zheng struct btrfs_key next_key; 16559e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 16565d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16575d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16585d4f98a2SYan Zheng struct btrfs_path *path; 16593fd0a558SYan, Zheng struct extent_buffer *leaf; 1660fca3a45dSJosef Bacik int reserve_level; 16615d4f98a2SYan Zheng int level; 16625d4f98a2SYan Zheng int max_level; 16635d4f98a2SYan Zheng int replaced = 0; 1664c6a592f2SNikolay Borisov int ret = 0; 16653fd0a558SYan, Zheng u32 min_reserved; 16665d4f98a2SYan Zheng 16675d4f98a2SYan Zheng path = btrfs_alloc_path(); 16685d4f98a2SYan Zheng if (!path) 16695d4f98a2SYan Zheng return -ENOMEM; 1670e4058b54SDavid Sterba path->reada = READA_FORWARD; 16715d4f98a2SYan Zheng 16725d4f98a2SYan Zheng reloc_root = root->reloc_root; 16735d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16745d4f98a2SYan Zheng 16755d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 16765d4f98a2SYan Zheng level = btrfs_root_level(root_item); 167767439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 16785d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 16795d4f98a2SYan Zheng path->slots[level] = 0; 16805d4f98a2SYan Zheng } else { 16815d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 16825d4f98a2SYan Zheng 1683c8422684SDavid Sterba level = btrfs_root_drop_level(root_item); 16845d4f98a2SYan Zheng BUG_ON(level == 0); 16855d4f98a2SYan Zheng path->lowest_level = level; 16865d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 168733c66f43SYan Zheng path->lowest_level = 0; 16885d4f98a2SYan Zheng if (ret < 0) { 16895d4f98a2SYan Zheng btrfs_free_path(path); 16905d4f98a2SYan Zheng return ret; 16915d4f98a2SYan Zheng } 16925d4f98a2SYan Zheng 16935d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 16945d4f98a2SYan Zheng path->slots[level]); 16955d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 16965d4f98a2SYan Zheng 16975d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 16985d4f98a2SYan Zheng } 16995d4f98a2SYan Zheng 170044d354abSQu Wenruo /* 170144d354abSQu Wenruo * In merge_reloc_root(), we modify the upper level pointer to swap the 170244d354abSQu Wenruo * tree blocks between reloc tree and subvolume tree. Thus for tree 170344d354abSQu Wenruo * block COW, we COW at most from level 1 to root level for each tree. 170444d354abSQu Wenruo * 170544d354abSQu Wenruo * Thus the needed metadata size is at most root_level * nodesize, 170644d354abSQu Wenruo * and * 2 since we have two trees to COW. 170744d354abSQu Wenruo */ 1708fca3a45dSJosef Bacik reserve_level = max_t(int, 1, btrfs_root_level(root_item)); 1709fca3a45dSJosef Bacik min_reserved = fs_info->nodesize * reserve_level * 2; 17105d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 17115d4f98a2SYan Zheng 17125d4f98a2SYan Zheng while (1) { 171308e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 171444d354abSQu Wenruo BTRFS_RESERVE_FLUSH_LIMIT); 1715c6a592f2SNikolay Borisov if (ret) 17169e6a0c52SJosef Bacik goto out; 17179e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 17189e6a0c52SJosef Bacik if (IS_ERR(trans)) { 1719c6a592f2SNikolay Borisov ret = PTR_ERR(trans); 17209e6a0c52SJosef Bacik trans = NULL; 17219e6a0c52SJosef Bacik goto out; 17229e6a0c52SJosef Bacik } 17232abc726aSJosef Bacik 17242abc726aSJosef Bacik /* 17252abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 17262abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 17272abc726aSJosef Bacik * 17282abc726aSJosef Bacik * But that's ok, we started the trans handle on our 17292abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 17302abc726aSJosef Bacik * dirty list. At commit time we'll still call 17312abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 17322abc726aSJosef Bacik * appropriately. 17332abc726aSJosef Bacik */ 17342abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 17359e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 17363fd0a558SYan, Zheng 17373fd0a558SYan, Zheng replaced = 0; 17385d4f98a2SYan Zheng max_level = level; 17395d4f98a2SYan Zheng 17405d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 1741c6a592f2SNikolay Borisov if (ret < 0) 17425d4f98a2SYan Zheng goto out; 17435d4f98a2SYan Zheng if (ret > 0) 17445d4f98a2SYan Zheng break; 17455d4f98a2SYan Zheng 17465d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 17475d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 17485d4f98a2SYan Zheng ret = 0; 17495d4f98a2SYan Zheng } else { 17503d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 17513fd0a558SYan, Zheng &next_key, level, max_level); 17525d4f98a2SYan Zheng } 1753c6a592f2SNikolay Borisov if (ret < 0) 17545d4f98a2SYan Zheng goto out; 17555d4f98a2SYan Zheng if (ret > 0) { 17565d4f98a2SYan Zheng level = ret; 17575d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17585d4f98a2SYan Zheng path->slots[level]); 17595d4f98a2SYan Zheng replaced = 1; 17605d4f98a2SYan Zheng } 17615d4f98a2SYan Zheng 17625d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 17635d4f98a2SYan Zheng if (ret > 0) 17645d4f98a2SYan Zheng break; 17655d4f98a2SYan Zheng 17665d4f98a2SYan Zheng BUG_ON(level == 0); 17675d4f98a2SYan Zheng /* 17685d4f98a2SYan Zheng * save the merging progress in the drop_progress. 17695d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 17705d4f98a2SYan Zheng */ 17715d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 17725d4f98a2SYan Zheng path->slots[level]); 1773c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, level); 17745d4f98a2SYan Zheng 17753a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 17769e6a0c52SJosef Bacik trans = NULL; 17775d4f98a2SYan Zheng 17782ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 17795d4f98a2SYan Zheng 17805d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17815d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17825d4f98a2SYan Zheng } 17835d4f98a2SYan Zheng 17845d4f98a2SYan Zheng /* 17855d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 17865d4f98a2SYan Zheng * relocated and the block is tree root. 17875d4f98a2SYan Zheng */ 17885d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 17899631e4ccSJosef Bacik ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf, 17909631e4ccSJosef Bacik BTRFS_NESTING_COW); 17915d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 17925d4f98a2SYan Zheng free_extent_buffer(leaf); 17935d4f98a2SYan Zheng out: 17945d4f98a2SYan Zheng btrfs_free_path(path); 17955d4f98a2SYan Zheng 1796ac54da6cSJosef Bacik if (ret == 0) { 1797ac54da6cSJosef Bacik ret = insert_dirty_subvol(trans, rc, root); 1798ac54da6cSJosef Bacik if (ret) 1799ac54da6cSJosef Bacik btrfs_abort_transaction(trans, ret); 1800ac54da6cSJosef Bacik } 18015d4f98a2SYan Zheng 18029e6a0c52SJosef Bacik if (trans) 18033a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 18045d4f98a2SYan Zheng 18052ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 18065d4f98a2SYan Zheng 18075d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 18085d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 18095d4f98a2SYan Zheng 1810c6a592f2SNikolay Borisov return ret; 18115d4f98a2SYan Zheng } 18125d4f98a2SYan Zheng 18133fd0a558SYan, Zheng static noinline_for_stack 18143fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 18155d4f98a2SYan Zheng { 18163fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 18170b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 18183fd0a558SYan, Zheng struct btrfs_root *reloc_root; 18195d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 18203fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 18213fd0a558SYan, Zheng u64 num_bytes = 0; 18223fd0a558SYan, Zheng int ret; 18233fd0a558SYan, Zheng 18240b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 18250b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 18263fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 18270b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 18287585717fSChris Mason 18293fd0a558SYan, Zheng again: 18303fd0a558SYan, Zheng if (!err) { 18313fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 183208e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 183308e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 18343fd0a558SYan, Zheng if (ret) 18353fd0a558SYan, Zheng err = ret; 18363fd0a558SYan, Zheng } 18373fd0a558SYan, Zheng 18387a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 18393612b495STsutomu Itoh if (IS_ERR(trans)) { 18403612b495STsutomu Itoh if (!err) 18412ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 184263f018beSNikolay Borisov num_bytes, NULL); 18433612b495STsutomu Itoh return PTR_ERR(trans); 18443612b495STsutomu Itoh } 18453fd0a558SYan, Zheng 18463fd0a558SYan, Zheng if (!err) { 18473fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 18483a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18492ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 185063f018beSNikolay Borisov num_bytes, NULL); 18513fd0a558SYan, Zheng goto again; 18523fd0a558SYan, Zheng } 18533fd0a558SYan, Zheng } 18543fd0a558SYan, Zheng 18553fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 18563fd0a558SYan, Zheng 18573fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 18583fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 18593fd0a558SYan, Zheng struct btrfs_root, root_list); 18603fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 18613fd0a558SYan, Zheng 1862a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1863a820feb5SDavid Sterba false); 18643fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 18653fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 18663fd0a558SYan, Zheng 18673fd0a558SYan, Zheng /* 18683fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 18693fd0a558SYan, Zheng * knows it should resumes merging 18703fd0a558SYan, Zheng */ 18713fd0a558SYan, Zheng if (!err) 18723fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 18733fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 18743fd0a558SYan, Zheng 18753fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 187600246528SJosef Bacik btrfs_put_root(root); 18773fd0a558SYan, Zheng } 18783fd0a558SYan, Zheng 18793fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 18803fd0a558SYan, Zheng 18813fd0a558SYan, Zheng if (!err) 18823a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 18833fd0a558SYan, Zheng else 18843a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18853fd0a558SYan, Zheng return err; 18863fd0a558SYan, Zheng } 18873fd0a558SYan, Zheng 18883fd0a558SYan, Zheng static noinline_for_stack 1889aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 1890aca1bba6SLiu Bo { 1891a7571232SNikolay Borisov struct btrfs_root *reloc_root, *tmp; 1892aca1bba6SLiu Bo 1893a7571232SNikolay Borisov list_for_each_entry_safe(reloc_root, tmp, list, root_list) 1894bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 1895aca1bba6SLiu Bo } 1896aca1bba6SLiu Bo 1897aca1bba6SLiu Bo static noinline_for_stack 189894404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 18993fd0a558SYan, Zheng { 19000b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 19015d4f98a2SYan Zheng struct btrfs_root *root; 19025d4f98a2SYan Zheng struct btrfs_root *reloc_root; 19033fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 19043fd0a558SYan, Zheng int found = 0; 1905aca1bba6SLiu Bo int ret = 0; 19063fd0a558SYan, Zheng again: 19073fd0a558SYan, Zheng root = rc->extent_root; 19087585717fSChris Mason 19097585717fSChris Mason /* 19107585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 19117585717fSChris Mason * we have to make sure nobody is in the middle of 19127585717fSChris Mason * adding their roots to the list while we are 19137585717fSChris Mason * doing this splice 19147585717fSChris Mason */ 19150b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 19163fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 19170b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 19185d4f98a2SYan Zheng 19193fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 19203fd0a558SYan, Zheng found = 1; 19213fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 19223fd0a558SYan, Zheng struct btrfs_root, root_list); 19235d4f98a2SYan Zheng 1924a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1925a820feb5SDavid Sterba false); 19265d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 19275d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 19285d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 19293fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 193000246528SJosef Bacik btrfs_put_root(root); 1931b37b39cdSJosef Bacik if (ret) { 193225e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 193325e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 193425e293c2SWang Shilong &reloc_roots); 1935aca1bba6SLiu Bo goto out; 1936b37b39cdSJosef Bacik } 19373fd0a558SYan, Zheng } else { 193851415b6cSQu Wenruo if (!IS_ERR(root)) { 193951415b6cSQu Wenruo if (root->reloc_root == reloc_root) { 194051415b6cSQu Wenruo root->reloc_root = NULL; 194151415b6cSQu Wenruo btrfs_put_root(reloc_root); 194251415b6cSQu Wenruo } 19431dae7e0eSQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, 19441dae7e0eSQu Wenruo &root->state); 194551415b6cSQu Wenruo btrfs_put_root(root); 194651415b6cSQu Wenruo } 194751415b6cSQu Wenruo 19483fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 194930d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 195030d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 195130d40577SQu Wenruo &rc->dirty_subvol_roots); 19523fd0a558SYan, Zheng } 19535d4f98a2SYan Zheng } 19545d4f98a2SYan Zheng 19553fd0a558SYan, Zheng if (found) { 19563fd0a558SYan, Zheng found = 0; 19573fd0a558SYan, Zheng goto again; 19585d4f98a2SYan Zheng } 1959aca1bba6SLiu Bo out: 1960aca1bba6SLiu Bo if (ret) { 19610b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 1962aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 1963467bb1d2SWang Shilong 1964467bb1d2SWang Shilong /* new reloc root may be added */ 19650b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 1966467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 19670b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 1968467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 1969aca1bba6SLiu Bo } 1970aca1bba6SLiu Bo 19717b7b7431SJosef Bacik /* 19727b7b7431SJosef Bacik * We used to have 19737b7b7431SJosef Bacik * 19747b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 19757b7b7431SJosef Bacik * 19767b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 19777b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 19787b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 19797b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 19807b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 19817b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 19827b7b7431SJosef Bacik * will be cleaned up on unmount. 19837b7b7431SJosef Bacik * 19847b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 19857b7b7431SJosef Bacik */ 19865d4f98a2SYan Zheng } 19875d4f98a2SYan Zheng 19885d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 19895d4f98a2SYan Zheng { 19905d4f98a2SYan Zheng struct tree_block *block; 19915d4f98a2SYan Zheng struct rb_node *rb_node; 19925d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 19935d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 19945d4f98a2SYan Zheng rb_erase(rb_node, blocks); 19955d4f98a2SYan Zheng kfree(block); 19965d4f98a2SYan Zheng } 19975d4f98a2SYan Zheng } 19985d4f98a2SYan Zheng 19995d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 20005d4f98a2SYan Zheng struct btrfs_root *reloc_root) 20015d4f98a2SYan Zheng { 20020b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 20035d4f98a2SYan Zheng struct btrfs_root *root; 2004442b1ac5SJosef Bacik int ret; 20055d4f98a2SYan Zheng 20065d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 20075d4f98a2SYan Zheng return 0; 20085d4f98a2SYan Zheng 2009a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false); 2010404bccbcSJosef Bacik 2011404bccbcSJosef Bacik /* 2012404bccbcSJosef Bacik * This should succeed, since we can't have a reloc root without having 2013404bccbcSJosef Bacik * already looked up the actual root and created the reloc root for this 2014404bccbcSJosef Bacik * root. 2015404bccbcSJosef Bacik * 2016404bccbcSJosef Bacik * However if there's some sort of corruption where we have a ref to a 2017404bccbcSJosef Bacik * reloc root without a corresponding root this could return ENOENT. 2018404bccbcSJosef Bacik */ 2019404bccbcSJosef Bacik if (IS_ERR(root)) { 2020404bccbcSJosef Bacik ASSERT(0); 2021404bccbcSJosef Bacik return PTR_ERR(root); 2022404bccbcSJosef Bacik } 2023404bccbcSJosef Bacik if (root->reloc_root != reloc_root) { 2024404bccbcSJosef Bacik ASSERT(0); 2025404bccbcSJosef Bacik btrfs_err(fs_info, 2026404bccbcSJosef Bacik "root %llu has two reloc roots associated with it", 2027404bccbcSJosef Bacik reloc_root->root_key.offset); 2028404bccbcSJosef Bacik btrfs_put_root(root); 2029404bccbcSJosef Bacik return -EUCLEAN; 2030404bccbcSJosef Bacik } 2031442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 203200246528SJosef Bacik btrfs_put_root(root); 20335d4f98a2SYan Zheng 2034442b1ac5SJosef Bacik return ret; 20355d4f98a2SYan Zheng } 20365d4f98a2SYan Zheng 20373fd0a558SYan, Zheng static noinline_for_stack 20383fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 20393fd0a558SYan, Zheng struct reloc_control *rc, 2040a26195a5SQu Wenruo struct btrfs_backref_node *node, 2041a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 20425d4f98a2SYan Zheng { 2043a26195a5SQu Wenruo struct btrfs_backref_node *next; 20445d4f98a2SYan Zheng struct btrfs_root *root; 20453fd0a558SYan, Zheng int index = 0; 204692de551bSJosef Bacik int ret; 20473fd0a558SYan, Zheng 20485d4f98a2SYan Zheng next = node; 20495d4f98a2SYan Zheng while (1) { 20505d4f98a2SYan Zheng cond_resched(); 20515d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 20525d4f98a2SYan Zheng root = next->root; 20538ee66afeSJosef Bacik 20548ee66afeSJosef Bacik /* 20558ee66afeSJosef Bacik * If there is no root, then our references for this block are 20568ee66afeSJosef Bacik * incomplete, as we should be able to walk all the way up to a 20578ee66afeSJosef Bacik * block that is owned by a root. 20588ee66afeSJosef Bacik * 20598ee66afeSJosef Bacik * This path is only for SHAREABLE roots, so if we come upon a 20608ee66afeSJosef Bacik * non-SHAREABLE root then we have backrefs that resolve 20618ee66afeSJosef Bacik * improperly. 20628ee66afeSJosef Bacik * 20638ee66afeSJosef Bacik * Both of these cases indicate file system corruption, or a bug 20648ee66afeSJosef Bacik * in the backref walking code. 20658ee66afeSJosef Bacik */ 20668ee66afeSJosef Bacik if (!root) { 20678ee66afeSJosef Bacik ASSERT(0); 20688ee66afeSJosef Bacik btrfs_err(trans->fs_info, 20698ee66afeSJosef Bacik "bytenr %llu doesn't have a backref path ending in a root", 20708ee66afeSJosef Bacik node->bytenr); 20718ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 20728ee66afeSJosef Bacik } 20738ee66afeSJosef Bacik if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 20748ee66afeSJosef Bacik ASSERT(0); 20758ee66afeSJosef Bacik btrfs_err(trans->fs_info, 20768ee66afeSJosef Bacik "bytenr %llu has multiple refs with one ending in a non-shareable root", 20778ee66afeSJosef Bacik node->bytenr); 20788ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 20798ee66afeSJosef Bacik } 20805d4f98a2SYan Zheng 20815d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 208292de551bSJosef Bacik ret = record_reloc_root_in_trans(trans, root); 208392de551bSJosef Bacik if (ret) 208492de551bSJosef Bacik return ERR_PTR(ret); 20855d4f98a2SYan Zheng break; 20865d4f98a2SYan Zheng } 20875d4f98a2SYan Zheng 208892de551bSJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 208992de551bSJosef Bacik if (ret) 209092de551bSJosef Bacik return ERR_PTR(ret); 20913fd0a558SYan, Zheng root = root->reloc_root; 20923fd0a558SYan, Zheng 209339200e59SJosef Bacik /* 209439200e59SJosef Bacik * We could have raced with another thread which failed, so 209539200e59SJosef Bacik * root->reloc_root may not be set, return ENOENT in this case. 209639200e59SJosef Bacik */ 209739200e59SJosef Bacik if (!root) 209839200e59SJosef Bacik return ERR_PTR(-ENOENT); 209939200e59SJosef Bacik 21003fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 21018ee66afeSJosef Bacik /* 21028ee66afeSJosef Bacik * We just created the reloc root, so we shouldn't have 21038ee66afeSJosef Bacik * ->new_bytenr set and this shouldn't be in the changed 21048ee66afeSJosef Bacik * list. If it is then we have multiple roots pointing 21058ee66afeSJosef Bacik * at the same bytenr which indicates corruption, or 21068ee66afeSJosef Bacik * we've made a mistake in the backref walking code. 21078ee66afeSJosef Bacik */ 21088ee66afeSJosef Bacik ASSERT(next->new_bytenr == 0); 21098ee66afeSJosef Bacik ASSERT(list_empty(&next->list)); 21108ee66afeSJosef Bacik if (next->new_bytenr || !list_empty(&next->list)) { 21118ee66afeSJosef Bacik btrfs_err(trans->fs_info, 21128ee66afeSJosef Bacik "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu", 21138ee66afeSJosef Bacik node->bytenr, next->bytenr); 21148ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 21158ee66afeSJosef Bacik } 21168ee66afeSJosef Bacik 21173fd0a558SYan, Zheng next->new_bytenr = root->node->start; 211800246528SJosef Bacik btrfs_put_root(next->root); 211900246528SJosef Bacik next->root = btrfs_grab_root(root); 21200b530bc5SJosef Bacik ASSERT(next->root); 21213fd0a558SYan, Zheng list_add_tail(&next->list, 21223fd0a558SYan, Zheng &rc->backref_cache.changed); 21239569cc20SQu Wenruo mark_block_processed(rc, next); 21245d4f98a2SYan Zheng break; 21255d4f98a2SYan Zheng } 21265d4f98a2SYan Zheng 21273fd0a558SYan, Zheng WARN_ON(1); 21285d4f98a2SYan Zheng root = NULL; 21295d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 21305d4f98a2SYan Zheng if (!next || next->level <= node->level) 21315d4f98a2SYan Zheng break; 21325d4f98a2SYan Zheng } 2133cbdc2ebcSJosef Bacik if (!root) { 2134cbdc2ebcSJosef Bacik /* 2135cbdc2ebcSJosef Bacik * This can happen if there's fs corruption or if there's a bug 2136cbdc2ebcSJosef Bacik * in the backref lookup code. 2137cbdc2ebcSJosef Bacik */ 2138cbdc2ebcSJosef Bacik ASSERT(0); 2139cbdc2ebcSJosef Bacik return ERR_PTR(-ENOENT); 2140cbdc2ebcSJosef Bacik } 21415d4f98a2SYan Zheng 21423fd0a558SYan, Zheng next = node; 21433fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 21443fd0a558SYan, Zheng while (1) { 21453fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 21463fd0a558SYan, Zheng if (--index < 0) 21473fd0a558SYan, Zheng break; 21483fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 21493fd0a558SYan, Zheng } 21505d4f98a2SYan Zheng return root; 21515d4f98a2SYan Zheng } 21525d4f98a2SYan Zheng 21533fd0a558SYan, Zheng /* 215492a7cc42SQu Wenruo * Select a tree root for relocation. 215592a7cc42SQu Wenruo * 215692a7cc42SQu Wenruo * Return NULL if the block is not shareable. We should use do_relocation() in 215792a7cc42SQu Wenruo * this case. 215892a7cc42SQu Wenruo * 215992a7cc42SQu Wenruo * Return a tree root pointer if the block is shareable. 216092a7cc42SQu Wenruo * Return -ENOENT if the block is root of reloc tree. 21613fd0a558SYan, Zheng */ 21625d4f98a2SYan Zheng static noinline_for_stack 2163a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 21645d4f98a2SYan Zheng { 2165a26195a5SQu Wenruo struct btrfs_backref_node *next; 21663fd0a558SYan, Zheng struct btrfs_root *root; 21673fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2168a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 21693fd0a558SYan, Zheng int index = 0; 21703fd0a558SYan, Zheng 21713fd0a558SYan, Zheng next = node; 21723fd0a558SYan, Zheng while (1) { 21733fd0a558SYan, Zheng cond_resched(); 21743fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 21753fd0a558SYan, Zheng root = next->root; 21763fd0a558SYan, Zheng BUG_ON(!root); 21773fd0a558SYan, Zheng 217892a7cc42SQu Wenruo /* No other choice for non-shareable tree */ 217992a7cc42SQu Wenruo if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 21803fd0a558SYan, Zheng return root; 21813fd0a558SYan, Zheng 21823fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 21833fd0a558SYan, Zheng fs_root = root; 21843fd0a558SYan, Zheng 21853fd0a558SYan, Zheng if (next != node) 21863fd0a558SYan, Zheng return NULL; 21873fd0a558SYan, Zheng 21883fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 21893fd0a558SYan, Zheng if (!next || next->level <= node->level) 21903fd0a558SYan, Zheng break; 21913fd0a558SYan, Zheng } 21923fd0a558SYan, Zheng 21933fd0a558SYan, Zheng if (!fs_root) 21943fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 21953fd0a558SYan, Zheng return fs_root; 21965d4f98a2SYan Zheng } 21975d4f98a2SYan Zheng 21985d4f98a2SYan Zheng static noinline_for_stack 21993fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2200a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 22015d4f98a2SYan Zheng { 22020b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2203a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2204a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2205a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 22063fd0a558SYan, Zheng u64 num_bytes = 0; 22073fd0a558SYan, Zheng int index = 0; 22085d4f98a2SYan Zheng 22093fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 22103fd0a558SYan, Zheng 22113fd0a558SYan, Zheng while (next) { 22123fd0a558SYan, Zheng cond_resched(); 22135d4f98a2SYan Zheng while (1) { 22143fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 22155d4f98a2SYan Zheng break; 22165d4f98a2SYan Zheng 22170b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 22183fd0a558SYan, Zheng 22193fd0a558SYan, Zheng if (list_empty(&next->upper)) 22203fd0a558SYan, Zheng break; 22213fd0a558SYan, Zheng 22223fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2223a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 22243fd0a558SYan, Zheng edges[index++] = edge; 22253fd0a558SYan, Zheng next = edge->node[UPPER]; 22265d4f98a2SYan Zheng } 22273fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 22283fd0a558SYan, Zheng } 22293fd0a558SYan, Zheng return num_bytes; 22303fd0a558SYan, Zheng } 22313fd0a558SYan, Zheng 22323fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 22333fd0a558SYan, Zheng struct reloc_control *rc, 2234a26195a5SQu Wenruo struct btrfs_backref_node *node) 22353fd0a558SYan, Zheng { 22363fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2237da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 22383fd0a558SYan, Zheng u64 num_bytes; 22393fd0a558SYan, Zheng int ret; 22400647bf56SWang Shilong u64 tmp; 22413fd0a558SYan, Zheng 22423fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 22433fd0a558SYan, Zheng 22443fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 22450647bf56SWang Shilong rc->reserved_bytes += num_bytes; 22468ca17f0fSJosef Bacik 22478ca17f0fSJosef Bacik /* 22488ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 22498ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 22508ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 22518ca17f0fSJosef Bacik */ 22520647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 22538ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 22543fd0a558SYan, Zheng if (ret) { 2255da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 22560647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 22570647bf56SWang Shilong tmp <<= 1; 22580647bf56SWang Shilong /* 22590647bf56SWang Shilong * only one thread can access block_rsv at this point, 22600647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 22610647bf56SWang Shilong * we expand more reservation size here to allow enough 226252042d8eSAndrea Gelmini * space for relocation and we will return earlier in 22630647bf56SWang Shilong * enospc case. 22640647bf56SWang Shilong */ 2265da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 22660647bf56SWang Shilong RELOCATION_RESERVED_NODES; 22678ca17f0fSJosef Bacik return -EAGAIN; 22683fd0a558SYan, Zheng } 22693fd0a558SYan, Zheng 22703fd0a558SYan, Zheng return 0; 22713fd0a558SYan, Zheng } 22723fd0a558SYan, Zheng 22735d4f98a2SYan Zheng /* 22745d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 22755d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 22765d4f98a2SYan Zheng * 22775d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 22785d4f98a2SYan Zheng * in that case this function just updates pointers. 22795d4f98a2SYan Zheng */ 22805d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 22813fd0a558SYan, Zheng struct reloc_control *rc, 2282a26195a5SQu Wenruo struct btrfs_backref_node *node, 22835d4f98a2SYan Zheng struct btrfs_key *key, 22845d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 22855d4f98a2SYan Zheng { 2286a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2287a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2288a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 22895d4f98a2SYan Zheng struct btrfs_root *root; 22905d4f98a2SYan Zheng struct extent_buffer *eb; 22915d4f98a2SYan Zheng u32 blocksize; 22925d4f98a2SYan Zheng u64 bytenr; 22935d4f98a2SYan Zheng int slot; 22948df01fddSNikolay Borisov int ret = 0; 22955d4f98a2SYan Zheng 2296ffe30dd8SJosef Bacik /* 2297ffe30dd8SJosef Bacik * If we are lowest then this is the first time we're processing this 2298ffe30dd8SJosef Bacik * block, and thus shouldn't have an eb associated with it yet. 2299ffe30dd8SJosef Bacik */ 2300ffe30dd8SJosef Bacik ASSERT(!lowest || !node->eb); 23015d4f98a2SYan Zheng 23025d4f98a2SYan Zheng path->lowest_level = node->level + 1; 23033fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 23045d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 230582fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2306581c1760SQu Wenruo 23075d4f98a2SYan Zheng cond_resched(); 23085d4f98a2SYan Zheng 23095d4f98a2SYan Zheng upper = edge->node[UPPER]; 2310dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 2311cbdc2ebcSJosef Bacik if (IS_ERR(root)) { 2312cbdc2ebcSJosef Bacik ret = PTR_ERR(root); 2313cbdc2ebcSJosef Bacik goto next; 2314cbdc2ebcSJosef Bacik } 23155d4f98a2SYan Zheng 23163fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 23173fd0a558SYan, Zheng if (!lowest) { 2318e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 23198df01fddSNikolay Borisov if (ret < 0) 2320cbca7d59SFilipe Manana goto next; 23213fd0a558SYan, Zheng BUG_ON(ret); 23223fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 23233fd0a558SYan, Zheng if (node->eb->start == bytenr) 23243fd0a558SYan, Zheng goto next; 23253fd0a558SYan, Zheng } 2326b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 23273fd0a558SYan, Zheng } 23285d4f98a2SYan Zheng 23295d4f98a2SYan Zheng if (!upper->eb) { 23305d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 23313561b9dbSLiu Bo if (ret) { 23328df01fddSNikolay Borisov if (ret > 0) 23338df01fddSNikolay Borisov ret = -ENOENT; 23343561b9dbSLiu Bo 23353561b9dbSLiu Bo btrfs_release_path(path); 23365d4f98a2SYan Zheng break; 23375d4f98a2SYan Zheng } 23385d4f98a2SYan Zheng 23393fd0a558SYan, Zheng if (!upper->eb) { 23403fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 23413fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 23423fd0a558SYan, Zheng } else { 23433fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 23443fd0a558SYan, Zheng } 23453fd0a558SYan, Zheng 23463fd0a558SYan, Zheng upper->locked = 1; 23473fd0a558SYan, Zheng path->locks[upper->level] = 0; 23483fd0a558SYan, Zheng 23495d4f98a2SYan Zheng slot = path->slots[upper->level]; 2350b3b4aa74SDavid Sterba btrfs_release_path(path); 23515d4f98a2SYan Zheng } else { 2352e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 23538df01fddSNikolay Borisov if (ret < 0) 2354cbca7d59SFilipe Manana goto next; 23555d4f98a2SYan Zheng BUG_ON(ret); 23565d4f98a2SYan Zheng } 23575d4f98a2SYan Zheng 23585d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 23593fd0a558SYan, Zheng if (lowest) { 23604547f4d8SLiu Bo if (bytenr != node->bytenr) { 23614547f4d8SLiu Bo btrfs_err(root->fs_info, 23624547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 23634547f4d8SLiu Bo bytenr, node->bytenr, slot, 23644547f4d8SLiu Bo upper->eb->start); 23658df01fddSNikolay Borisov ret = -EIO; 23664547f4d8SLiu Bo goto next; 23674547f4d8SLiu Bo } 23685d4f98a2SYan Zheng } else { 23693fd0a558SYan, Zheng if (node->eb->start == bytenr) 23703fd0a558SYan, Zheng goto next; 23715d4f98a2SYan Zheng } 23725d4f98a2SYan Zheng 2373da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 2374c9752536SJosef Bacik eb = btrfs_read_node_slot(upper->eb, slot); 237564c043deSLiu Bo if (IS_ERR(eb)) { 23768df01fddSNikolay Borisov ret = PTR_ERR(eb); 237764c043deSLiu Bo goto next; 237897d9a8a4STsutomu Itoh } 23795d4f98a2SYan Zheng btrfs_tree_lock(eb); 23805d4f98a2SYan Zheng 23815d4f98a2SYan Zheng if (!node->eb) { 23825d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 23839631e4ccSJosef Bacik slot, &eb, BTRFS_NESTING_COW); 23843fd0a558SYan, Zheng btrfs_tree_unlock(eb); 23853fd0a558SYan, Zheng free_extent_buffer(eb); 23868df01fddSNikolay Borisov if (ret < 0) 23873fd0a558SYan, Zheng goto next; 2388ffe30dd8SJosef Bacik /* 2389ffe30dd8SJosef Bacik * We've just COWed this block, it should have updated 2390ffe30dd8SJosef Bacik * the correct backref node entry. 2391ffe30dd8SJosef Bacik */ 2392ffe30dd8SJosef Bacik ASSERT(node->eb == eb); 23935d4f98a2SYan Zheng } else { 23945d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 23955d4f98a2SYan Zheng node->eb->start); 23965d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 23975d4f98a2SYan Zheng trans->transid); 23985d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 23995d4f98a2SYan Zheng 240082fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 24015d4f98a2SYan Zheng node->eb->start, blocksize, 240282fa113fSQu Wenruo upper->eb->start); 240382fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 240482fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 240582fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 240682fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 24075d4f98a2SYan Zheng BUG_ON(ret); 24085d4f98a2SYan Zheng 24095d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 24105d4f98a2SYan Zheng BUG_ON(ret); 24115d4f98a2SYan Zheng } 24123fd0a558SYan, Zheng next: 24133fd0a558SYan, Zheng if (!upper->pending) 2414b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 24153fd0a558SYan, Zheng else 2416b0fe7078SQu Wenruo btrfs_backref_unlock_node_buffer(upper); 24178df01fddSNikolay Borisov if (ret) 24183fd0a558SYan, Zheng break; 24195d4f98a2SYan Zheng } 24203fd0a558SYan, Zheng 24218df01fddSNikolay Borisov if (!ret && node->pending) { 2422b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 24233fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 24243fd0a558SYan, Zheng node->pending = 0; 24255d4f98a2SYan Zheng } 24263fd0a558SYan, Zheng 24275d4f98a2SYan Zheng path->lowest_level = 0; 2428ffe30dd8SJosef Bacik 2429ffe30dd8SJosef Bacik /* 2430ffe30dd8SJosef Bacik * We should have allocated all of our space in the block rsv and thus 2431ffe30dd8SJosef Bacik * shouldn't ENOSPC. 2432ffe30dd8SJosef Bacik */ 2433ffe30dd8SJosef Bacik ASSERT(ret != -ENOSPC); 24348df01fddSNikolay Borisov return ret; 24355d4f98a2SYan Zheng } 24365d4f98a2SYan Zheng 24375d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 24383fd0a558SYan, Zheng struct reloc_control *rc, 2439a26195a5SQu Wenruo struct btrfs_backref_node *node, 24405d4f98a2SYan Zheng struct btrfs_path *path) 24415d4f98a2SYan Zheng { 24425d4f98a2SYan Zheng struct btrfs_key key; 24435d4f98a2SYan Zheng 24445d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 24453fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 24465d4f98a2SYan Zheng } 24475d4f98a2SYan Zheng 24485d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 24493fd0a558SYan, Zheng struct reloc_control *rc, 24503fd0a558SYan, Zheng struct btrfs_path *path, int err) 24515d4f98a2SYan Zheng { 24523fd0a558SYan, Zheng LIST_HEAD(list); 2453a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 2454a26195a5SQu Wenruo struct btrfs_backref_node *node; 24555d4f98a2SYan Zheng int level; 24565d4f98a2SYan Zheng int ret; 24575d4f98a2SYan Zheng 24585d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 24595d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 24605d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 2461a26195a5SQu Wenruo struct btrfs_backref_node, list); 24623fd0a558SYan, Zheng list_move_tail(&node->list, &list); 24633fd0a558SYan, Zheng BUG_ON(!node->pending); 24645d4f98a2SYan Zheng 24653fd0a558SYan, Zheng if (!err) { 24663fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 24675d4f98a2SYan Zheng if (ret < 0) 24685d4f98a2SYan Zheng err = ret; 24695d4f98a2SYan Zheng } 24705d4f98a2SYan Zheng } 24713fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 24723fd0a558SYan, Zheng } 24735d4f98a2SYan Zheng return err; 24745d4f98a2SYan Zheng } 24755d4f98a2SYan Zheng 24765d4f98a2SYan Zheng /* 24775d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 24785d4f98a2SYan Zheng * as processed. 24795d4f98a2SYan Zheng */ 24805d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 2481a26195a5SQu Wenruo struct btrfs_backref_node *node) 24825d4f98a2SYan Zheng { 2483a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2484a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2485a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24865d4f98a2SYan Zheng int index = 0; 24875d4f98a2SYan Zheng 24885d4f98a2SYan Zheng while (next) { 24895d4f98a2SYan Zheng cond_resched(); 24905d4f98a2SYan Zheng while (1) { 24915d4f98a2SYan Zheng if (next->processed) 24925d4f98a2SYan Zheng break; 24935d4f98a2SYan Zheng 24949569cc20SQu Wenruo mark_block_processed(rc, next); 24955d4f98a2SYan Zheng 24965d4f98a2SYan Zheng if (list_empty(&next->upper)) 24975d4f98a2SYan Zheng break; 24985d4f98a2SYan Zheng 24995d4f98a2SYan Zheng edge = list_entry(next->upper.next, 2500a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 25015d4f98a2SYan Zheng edges[index++] = edge; 25025d4f98a2SYan Zheng next = edge->node[UPPER]; 25035d4f98a2SYan Zheng } 25045d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 25055d4f98a2SYan Zheng } 25065d4f98a2SYan Zheng } 25075d4f98a2SYan Zheng 25087476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 25095d4f98a2SYan Zheng { 2510da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 25117476dfdaSDavid Sterba 25125d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 25139655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 25145d4f98a2SYan Zheng return 1; 25155d4f98a2SYan Zheng return 0; 25165d4f98a2SYan Zheng } 25175d4f98a2SYan Zheng 25182ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 25195d4f98a2SYan Zheng struct tree_block *block) 25205d4f98a2SYan Zheng { 25215d4f98a2SYan Zheng struct extent_buffer *eb; 25225d4f98a2SYan Zheng 2523f7ba2d37SJosef Bacik eb = read_tree_block(fs_info, block->bytenr, block->owner, 2524f7ba2d37SJosef Bacik block->key.offset, block->level, NULL); 252564c043deSLiu Bo if (IS_ERR(eb)) { 252664c043deSLiu Bo return PTR_ERR(eb); 252764c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2528416bc658SJosef Bacik free_extent_buffer(eb); 2529416bc658SJosef Bacik return -EIO; 2530416bc658SJosef Bacik } 25315d4f98a2SYan Zheng if (block->level == 0) 25325d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 25335d4f98a2SYan Zheng else 25345d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 25355d4f98a2SYan Zheng free_extent_buffer(eb); 25365d4f98a2SYan Zheng block->key_ready = 1; 25375d4f98a2SYan Zheng return 0; 25385d4f98a2SYan Zheng } 25395d4f98a2SYan Zheng 25405d4f98a2SYan Zheng /* 25415d4f98a2SYan Zheng * helper function to relocate a tree block 25425d4f98a2SYan Zheng */ 25435d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 25445d4f98a2SYan Zheng struct reloc_control *rc, 2545a26195a5SQu Wenruo struct btrfs_backref_node *node, 25465d4f98a2SYan Zheng struct btrfs_key *key, 25475d4f98a2SYan Zheng struct btrfs_path *path) 25485d4f98a2SYan Zheng { 25495d4f98a2SYan Zheng struct btrfs_root *root; 25503fd0a558SYan, Zheng int ret = 0; 25515d4f98a2SYan Zheng 25523fd0a558SYan, Zheng if (!node) 25535d4f98a2SYan Zheng return 0; 25543fd0a558SYan, Zheng 25555f6b2e5cSJosef Bacik /* 25565f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 25575f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 25585f6b2e5cSJosef Bacik */ 25595f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 25605f6b2e5cSJosef Bacik if (ret) 25615f6b2e5cSJosef Bacik goto out; 25625f6b2e5cSJosef Bacik 25633fd0a558SYan, Zheng BUG_ON(node->processed); 2564147d256eSZhaolei root = select_one_root(node); 25653fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 25663fd0a558SYan, Zheng update_processed_blocks(rc, node); 25673fd0a558SYan, Zheng goto out; 25685d4f98a2SYan Zheng } 25695d4f98a2SYan Zheng 25703fd0a558SYan, Zheng if (root) { 257192a7cc42SQu Wenruo if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 25721c7bfa15SJosef Bacik /* 25731c7bfa15SJosef Bacik * This block was the root block of a root, and this is 25741c7bfa15SJosef Bacik * the first time we're processing the block and thus it 25751c7bfa15SJosef Bacik * should not have had the ->new_bytenr modified and 25761c7bfa15SJosef Bacik * should have not been included on the changed list. 25771c7bfa15SJosef Bacik * 25781c7bfa15SJosef Bacik * However in the case of corruption we could have 25791c7bfa15SJosef Bacik * multiple refs pointing to the same block improperly, 25801c7bfa15SJosef Bacik * and thus we would trip over these checks. ASSERT() 25811c7bfa15SJosef Bacik * for the developer case, because it could indicate a 25821c7bfa15SJosef Bacik * bug in the backref code, however error out for a 25831c7bfa15SJosef Bacik * normal user in the case of corruption. 25841c7bfa15SJosef Bacik */ 25851c7bfa15SJosef Bacik ASSERT(node->new_bytenr == 0); 25861c7bfa15SJosef Bacik ASSERT(list_empty(&node->list)); 25871c7bfa15SJosef Bacik if (node->new_bytenr || !list_empty(&node->list)) { 25881c7bfa15SJosef Bacik btrfs_err(root->fs_info, 25891c7bfa15SJosef Bacik "bytenr %llu has improper references to it", 25901c7bfa15SJosef Bacik node->bytenr); 25911c7bfa15SJosef Bacik ret = -EUCLEAN; 25921c7bfa15SJosef Bacik goto out; 25931c7bfa15SJosef Bacik } 2594d18c7bd9SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 2595d18c7bd9SJosef Bacik if (ret) 2596d18c7bd9SJosef Bacik goto out; 259739200e59SJosef Bacik /* 259839200e59SJosef Bacik * Another thread could have failed, need to check if we 259939200e59SJosef Bacik * have reloc_root actually set. 260039200e59SJosef Bacik */ 260139200e59SJosef Bacik if (!root->reloc_root) { 260239200e59SJosef Bacik ret = -ENOENT; 260339200e59SJosef Bacik goto out; 260439200e59SJosef Bacik } 26053fd0a558SYan, Zheng root = root->reloc_root; 26063fd0a558SYan, Zheng node->new_bytenr = root->node->start; 260700246528SJosef Bacik btrfs_put_root(node->root); 260800246528SJosef Bacik node->root = btrfs_grab_root(root); 26090b530bc5SJosef Bacik ASSERT(node->root); 26103fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 26113fd0a558SYan, Zheng } else { 26125d4f98a2SYan Zheng path->lowest_level = node->level; 26135d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2614b3b4aa74SDavid Sterba btrfs_release_path(path); 26153fd0a558SYan, Zheng if (ret > 0) 26165d4f98a2SYan Zheng ret = 0; 26173fd0a558SYan, Zheng } 26183fd0a558SYan, Zheng if (!ret) 26193fd0a558SYan, Zheng update_processed_blocks(rc, node); 26203fd0a558SYan, Zheng } else { 26213fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 26223fd0a558SYan, Zheng } 26235d4f98a2SYan Zheng out: 26240647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 2625023acb07SQu Wenruo btrfs_backref_cleanup_node(&rc->backref_cache, node); 26265d4f98a2SYan Zheng return ret; 26275d4f98a2SYan Zheng } 26285d4f98a2SYan Zheng 26295d4f98a2SYan Zheng /* 26305d4f98a2SYan Zheng * relocate a list of blocks 26315d4f98a2SYan Zheng */ 26325d4f98a2SYan Zheng static noinline_for_stack 26335d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 26345d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 26355d4f98a2SYan Zheng { 26362ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2637a26195a5SQu Wenruo struct btrfs_backref_node *node; 26385d4f98a2SYan Zheng struct btrfs_path *path; 26395d4f98a2SYan Zheng struct tree_block *block; 264098ff7b94SQu Wenruo struct tree_block *next; 26415d4f98a2SYan Zheng int ret; 26425d4f98a2SYan Zheng int err = 0; 26435d4f98a2SYan Zheng 26445d4f98a2SYan Zheng path = btrfs_alloc_path(); 2645e1a12670SLiu Bo if (!path) { 2646e1a12670SLiu Bo err = -ENOMEM; 264734c2b290SDavid Sterba goto out_free_blocks; 2648e1a12670SLiu Bo } 26495d4f98a2SYan Zheng 265098ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 265198ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 26525d4f98a2SYan Zheng if (!block->key_ready) 2653f7ba2d37SJosef Bacik btrfs_readahead_tree_block(fs_info, block->bytenr, 2654f7ba2d37SJosef Bacik block->owner, 0, 26553fbaf258SJosef Bacik block->level); 26565d4f98a2SYan Zheng } 26575d4f98a2SYan Zheng 265898ff7b94SQu Wenruo /* Get first keys */ 265998ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 266034c2b290SDavid Sterba if (!block->key_ready) { 26612ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 266234c2b290SDavid Sterba if (err) 266334c2b290SDavid Sterba goto out_free_path; 266434c2b290SDavid Sterba } 26655d4f98a2SYan Zheng } 26665d4f98a2SYan Zheng 266798ff7b94SQu Wenruo /* Do tree relocation */ 266898ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 26693fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 26705d4f98a2SYan Zheng block->level, block->bytenr); 26715d4f98a2SYan Zheng if (IS_ERR(node)) { 26725d4f98a2SYan Zheng err = PTR_ERR(node); 26735d4f98a2SYan Zheng goto out; 26745d4f98a2SYan Zheng } 26755d4f98a2SYan Zheng 26765d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 26775d4f98a2SYan Zheng path); 26785d4f98a2SYan Zheng if (ret < 0) { 26795d4f98a2SYan Zheng err = ret; 268050dbbb71SJosef Bacik break; 26815d4f98a2SYan Zheng } 26825d4f98a2SYan Zheng } 26835d4f98a2SYan Zheng out: 26843fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 26855d4f98a2SYan Zheng 268634c2b290SDavid Sterba out_free_path: 26875d4f98a2SYan Zheng btrfs_free_path(path); 268834c2b290SDavid Sterba out_free_blocks: 2689e1a12670SLiu Bo free_block_list(blocks); 26905d4f98a2SYan Zheng return err; 26915d4f98a2SYan Zheng } 26925d4f98a2SYan Zheng 2693056d9becSNikolay Borisov static noinline_for_stack int prealloc_file_extent_cluster( 2694056d9becSNikolay Borisov struct btrfs_inode *inode, 2695efa56464SYan, Zheng struct file_extent_cluster *cluster) 2696efa56464SYan, Zheng { 2697efa56464SYan, Zheng u64 alloc_hint = 0; 2698efa56464SYan, Zheng u64 start; 2699efa56464SYan, Zheng u64 end; 2700056d9becSNikolay Borisov u64 offset = inode->index_cnt; 2701efa56464SYan, Zheng u64 num_bytes; 27024e9d0d01SNikolay Borisov int nr; 2703efa56464SYan, Zheng int ret = 0; 2704dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 2705dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 2706214e61d0SNikolay Borisov u64 cur_offset = prealloc_start; 2707efa56464SYan, Zheng 2708efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2709056d9becSNikolay Borisov ret = btrfs_alloc_data_chunk_ondemand(inode, 2710dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 2711efa56464SYan, Zheng if (ret) 2712214e61d0SNikolay Borisov return ret; 2713efa56464SYan, Zheng 271432430c61SNaohiro Aota /* 271532430c61SNaohiro Aota * On a zoned filesystem, we cannot preallocate the file region. 271632430c61SNaohiro Aota * Instead, we dirty and fiemap_write the region. 271732430c61SNaohiro Aota */ 271832430c61SNaohiro Aota if (btrfs_is_zoned(inode->root->fs_info)) { 271932430c61SNaohiro Aota struct btrfs_root *root = inode->root; 272032430c61SNaohiro Aota struct btrfs_trans_handle *trans; 272132430c61SNaohiro Aota 272232430c61SNaohiro Aota end = cluster->end - offset + 1; 272332430c61SNaohiro Aota trans = btrfs_start_transaction(root, 1); 272432430c61SNaohiro Aota if (IS_ERR(trans)) 272532430c61SNaohiro Aota return PTR_ERR(trans); 272632430c61SNaohiro Aota 272732430c61SNaohiro Aota inode->vfs_inode.i_ctime = current_time(&inode->vfs_inode); 272832430c61SNaohiro Aota i_size_write(&inode->vfs_inode, end); 272932430c61SNaohiro Aota ret = btrfs_update_inode(trans, root, inode); 273032430c61SNaohiro Aota if (ret) { 273132430c61SNaohiro Aota btrfs_abort_transaction(trans, ret); 273232430c61SNaohiro Aota btrfs_end_transaction(trans); 273332430c61SNaohiro Aota return ret; 273432430c61SNaohiro Aota } 273532430c61SNaohiro Aota 273632430c61SNaohiro Aota return btrfs_end_transaction(trans); 273732430c61SNaohiro Aota } 273832430c61SNaohiro Aota 273964708539SJosef Bacik btrfs_inode_lock(&inode->vfs_inode, 0); 27404e9d0d01SNikolay Borisov for (nr = 0; nr < cluster->nr; nr++) { 2741efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2742efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2743efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2744efa56464SYan, Zheng else 2745efa56464SYan, Zheng end = cluster->end - offset; 2746efa56464SYan, Zheng 2747056d9becSNikolay Borisov lock_extent(&inode->io_tree, start, end); 2748efa56464SYan, Zheng num_bytes = end + 1 - start; 2749056d9becSNikolay Borisov ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start, 2750efa56464SYan, Zheng num_bytes, num_bytes, 2751efa56464SYan, Zheng end + 1, &alloc_hint); 275218513091SWang Xiaoguang cur_offset = end + 1; 2753056d9becSNikolay Borisov unlock_extent(&inode->io_tree, start, end); 2754efa56464SYan, Zheng if (ret) 2755efa56464SYan, Zheng break; 2756efa56464SYan, Zheng } 275764708539SJosef Bacik btrfs_inode_unlock(&inode->vfs_inode, 0); 2758214e61d0SNikolay Borisov 275918513091SWang Xiaoguang if (cur_offset < prealloc_end) 2760056d9becSNikolay Borisov btrfs_free_reserved_data_space_noquota(inode->root->fs_info, 2761a89ef455SFilipe Manana prealloc_end + 1 - cur_offset); 2762efa56464SYan, Zheng return ret; 2763efa56464SYan, Zheng } 2764efa56464SYan, Zheng 2765efa56464SYan, Zheng static noinline_for_stack 27660257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 27670257bb82SYan, Zheng u64 block_start) 27685d4f98a2SYan Zheng { 27695d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 27705d4f98a2SYan Zheng struct extent_map *em; 27710257bb82SYan, Zheng int ret = 0; 27725d4f98a2SYan Zheng 2773172ddd60SDavid Sterba em = alloc_extent_map(); 27740257bb82SYan, Zheng if (!em) 27750257bb82SYan, Zheng return -ENOMEM; 27760257bb82SYan, Zheng 27775d4f98a2SYan Zheng em->start = start; 27780257bb82SYan, Zheng em->len = end + 1 - start; 27790257bb82SYan, Zheng em->block_len = em->len; 27800257bb82SYan, Zheng em->block_start = block_start; 27815d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 27825d4f98a2SYan Zheng 2783d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 27845d4f98a2SYan Zheng while (1) { 2785890871beSChris Mason write_lock(&em_tree->lock); 278609a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 2787890871beSChris Mason write_unlock(&em_tree->lock); 27885d4f98a2SYan Zheng if (ret != -EEXIST) { 27895d4f98a2SYan Zheng free_extent_map(em); 27905d4f98a2SYan Zheng break; 27915d4f98a2SYan Zheng } 2792dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 27935d4f98a2SYan Zheng } 2794d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 27950257bb82SYan, Zheng return ret; 27960257bb82SYan, Zheng } 27975d4f98a2SYan Zheng 2798726a3421SQu Wenruo /* 2799726a3421SQu Wenruo * Allow error injection to test balance cancellation 2800726a3421SQu Wenruo */ 28011fec12a5SJosef Bacik noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 2802726a3421SQu Wenruo { 28035cb502f4SQu Wenruo return atomic_read(&fs_info->balance_cancel_req) || 28045cb502f4SQu Wenruo fatal_signal_pending(current); 2805726a3421SQu Wenruo } 2806726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 2807726a3421SQu Wenruo 28080257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 28090257bb82SYan, Zheng struct file_extent_cluster *cluster) 28100257bb82SYan, Zheng { 28112ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 28120257bb82SYan, Zheng u64 page_start; 28130257bb82SYan, Zheng u64 page_end; 28140257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 28150257bb82SYan, Zheng unsigned long index; 28160257bb82SYan, Zheng unsigned long last_index; 28170257bb82SYan, Zheng struct page *page; 28180257bb82SYan, Zheng struct file_ra_state *ra; 28193b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 28200257bb82SYan, Zheng int nr = 0; 28210257bb82SYan, Zheng int ret = 0; 28220257bb82SYan, Zheng 28230257bb82SYan, Zheng if (!cluster->nr) 28240257bb82SYan, Zheng return 0; 28250257bb82SYan, Zheng 28260257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 28270257bb82SYan, Zheng if (!ra) 28280257bb82SYan, Zheng return -ENOMEM; 28290257bb82SYan, Zheng 2830056d9becSNikolay Borisov ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster); 28310257bb82SYan, Zheng if (ret) 2832efa56464SYan, Zheng goto out; 28330257bb82SYan, Zheng 28340257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 28350257bb82SYan, Zheng 2836efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2837efa56464SYan, Zheng cluster->end - offset, cluster->start); 2838efa56464SYan, Zheng if (ret) 2839efa56464SYan, Zheng goto out; 2840efa56464SYan, Zheng 284109cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 284209cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 28430257bb82SYan, Zheng while (index <= last_index) { 28449f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 28459f3db423SNikolay Borisov PAGE_SIZE); 2846efa56464SYan, Zheng if (ret) 2847efa56464SYan, Zheng goto out; 2848efa56464SYan, Zheng 28490257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 28500257bb82SYan, Zheng if (!page) { 28510257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 28520257bb82SYan, Zheng ra, NULL, index, 28530257bb82SYan, Zheng last_index + 1 - index); 2854a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 28553b16a4e3SJosef Bacik mask); 28560257bb82SYan, Zheng if (!page) { 2857691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 285843b18595SQu Wenruo PAGE_SIZE, true); 285944db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 28608702ba93SQu Wenruo PAGE_SIZE); 28610257bb82SYan, Zheng ret = -ENOMEM; 2862efa56464SYan, Zheng goto out; 28630257bb82SYan, Zheng } 28640257bb82SYan, Zheng } 286532443de3SQu Wenruo ret = set_page_extent_mapped(page); 286632443de3SQu Wenruo if (ret < 0) { 286732443de3SQu Wenruo btrfs_delalloc_release_metadata(BTRFS_I(inode), 286832443de3SQu Wenruo PAGE_SIZE, true); 286932443de3SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 287032443de3SQu Wenruo unlock_page(page); 287132443de3SQu Wenruo put_page(page); 287232443de3SQu Wenruo goto out; 287332443de3SQu Wenruo } 28740257bb82SYan, Zheng 28750257bb82SYan, Zheng if (PageReadahead(page)) { 28760257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 28770257bb82SYan, Zheng ra, NULL, page, index, 28780257bb82SYan, Zheng last_index + 1 - index); 28790257bb82SYan, Zheng } 28800257bb82SYan, Zheng 28810257bb82SYan, Zheng if (!PageUptodate(page)) { 28820257bb82SYan, Zheng btrfs_readpage(NULL, page); 28830257bb82SYan, Zheng lock_page(page); 28840257bb82SYan, Zheng if (!PageUptodate(page)) { 28850257bb82SYan, Zheng unlock_page(page); 288609cbfeafSKirill A. Shutemov put_page(page); 2887691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 288843b18595SQu Wenruo PAGE_SIZE, true); 28898b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 28908702ba93SQu Wenruo PAGE_SIZE); 28910257bb82SYan, Zheng ret = -EIO; 2892efa56464SYan, Zheng goto out; 28930257bb82SYan, Zheng } 28940257bb82SYan, Zheng } 28950257bb82SYan, Zheng 28964eee4fa4SMiao Xie page_start = page_offset(page); 289709cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 28980257bb82SYan, Zheng 2899d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 29000257bb82SYan, Zheng 29010257bb82SYan, Zheng if (nr < cluster->nr && 29020257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 29030257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 29040257bb82SYan, Zheng page_start, page_end, 2905ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 29060257bb82SYan, Zheng nr++; 29070257bb82SYan, Zheng } 29080257bb82SYan, Zheng 2909c2566f22SNikolay Borisov ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, 2910c2566f22SNikolay Borisov page_end, 0, NULL); 2911765f3cebSNikolay Borisov if (ret) { 2912765f3cebSNikolay Borisov unlock_page(page); 2913765f3cebSNikolay Borisov put_page(page); 2914765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 291543b18595SQu Wenruo PAGE_SIZE, true); 2916765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 29178702ba93SQu Wenruo PAGE_SIZE); 2918765f3cebSNikolay Borisov 2919765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 2920765f3cebSNikolay Borisov page_start, page_end, 2921765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 2922765f3cebSNikolay Borisov goto out; 2923765f3cebSNikolay Borisov 2924765f3cebSNikolay Borisov } 29250257bb82SYan, Zheng set_page_dirty(page); 29260257bb82SYan, Zheng 29270257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 2928d0082371SJeff Mahoney page_start, page_end); 29290257bb82SYan, Zheng unlock_page(page); 293009cbfeafSKirill A. Shutemov put_page(page); 29310257bb82SYan, Zheng 29320257bb82SYan, Zheng index++; 29338702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 2934efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 29352ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 29367f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 29377f913c7cSQu Wenruo ret = -ECANCELED; 29387f913c7cSQu Wenruo goto out; 29397f913c7cSQu Wenruo } 29400257bb82SYan, Zheng } 29410257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 294232430c61SNaohiro Aota if (btrfs_is_zoned(fs_info) && !ret) 294332430c61SNaohiro Aota ret = btrfs_wait_ordered_range(inode, 0, (u64)-1); 2944efa56464SYan, Zheng out: 29450257bb82SYan, Zheng kfree(ra); 29460257bb82SYan, Zheng return ret; 29470257bb82SYan, Zheng } 29480257bb82SYan, Zheng 29490257bb82SYan, Zheng static noinline_for_stack 29500257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 29510257bb82SYan, Zheng struct file_extent_cluster *cluster) 29520257bb82SYan, Zheng { 29530257bb82SYan, Zheng int ret; 29540257bb82SYan, Zheng 29550257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 29560257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 29570257bb82SYan, Zheng if (ret) 29580257bb82SYan, Zheng return ret; 29590257bb82SYan, Zheng cluster->nr = 0; 29600257bb82SYan, Zheng } 29610257bb82SYan, Zheng 29620257bb82SYan, Zheng if (!cluster->nr) 29630257bb82SYan, Zheng cluster->start = extent_key->objectid; 29640257bb82SYan, Zheng else 29650257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 29660257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 29670257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 29680257bb82SYan, Zheng cluster->nr++; 29690257bb82SYan, Zheng 29700257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 29710257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 29720257bb82SYan, Zheng if (ret) 29730257bb82SYan, Zheng return ret; 29740257bb82SYan, Zheng cluster->nr = 0; 29750257bb82SYan, Zheng } 29760257bb82SYan, Zheng return 0; 29775d4f98a2SYan Zheng } 29785d4f98a2SYan Zheng 29795d4f98a2SYan Zheng /* 29805d4f98a2SYan Zheng * helper to add a tree block to the list. 29815d4f98a2SYan Zheng * the major work is getting the generation and level of the block 29825d4f98a2SYan Zheng */ 29835d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 29845d4f98a2SYan Zheng struct btrfs_key *extent_key, 29855d4f98a2SYan Zheng struct btrfs_path *path, 29865d4f98a2SYan Zheng struct rb_root *blocks) 29875d4f98a2SYan Zheng { 29885d4f98a2SYan Zheng struct extent_buffer *eb; 29895d4f98a2SYan Zheng struct btrfs_extent_item *ei; 29905d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 29915d4f98a2SYan Zheng struct tree_block *block; 29925d4f98a2SYan Zheng struct rb_node *rb_node; 29935d4f98a2SYan Zheng u32 item_size; 29945d4f98a2SYan Zheng int level = -1; 29957fdf4b60SWang Shilong u64 generation; 2996f7ba2d37SJosef Bacik u64 owner = 0; 29975d4f98a2SYan Zheng 29985d4f98a2SYan Zheng eb = path->nodes[0]; 29995d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 30005d4f98a2SYan Zheng 30013173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 30023173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 3003f7ba2d37SJosef Bacik unsigned long ptr = 0, end; 3004f7ba2d37SJosef Bacik 30055d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 30065d4f98a2SYan Zheng struct btrfs_extent_item); 3007f7ba2d37SJosef Bacik end = (unsigned long)ei + item_size; 30083173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 30095d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 30105d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 3011f7ba2d37SJosef Bacik ptr = (unsigned long)(bi + 1); 30125d4f98a2SYan Zheng } else { 30133173a18fSJosef Bacik level = (int)extent_key->offset; 3014f7ba2d37SJosef Bacik ptr = (unsigned long)(ei + 1); 30153173a18fSJosef Bacik } 30163173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 3017f7ba2d37SJosef Bacik 3018f7ba2d37SJosef Bacik /* 3019f7ba2d37SJosef Bacik * We're reading random blocks without knowing their owner ahead 3020f7ba2d37SJosef Bacik * of time. This is ok most of the time, as all reloc roots and 3021f7ba2d37SJosef Bacik * fs roots have the same lock type. However normal trees do 3022f7ba2d37SJosef Bacik * not, and the only way to know ahead of time is to read the 3023f7ba2d37SJosef Bacik * inline ref offset. We know it's an fs root if 3024f7ba2d37SJosef Bacik * 3025f7ba2d37SJosef Bacik * 1. There's more than one ref. 3026f7ba2d37SJosef Bacik * 2. There's a SHARED_DATA_REF_KEY set. 3027f7ba2d37SJosef Bacik * 3. FULL_BACKREF is set on the flags. 3028f7ba2d37SJosef Bacik * 3029f7ba2d37SJosef Bacik * Otherwise it's safe to assume that the ref offset == the 3030f7ba2d37SJosef Bacik * owner of this block, so we can use that when calling 3031f7ba2d37SJosef Bacik * read_tree_block. 3032f7ba2d37SJosef Bacik */ 3033f7ba2d37SJosef Bacik if (btrfs_extent_refs(eb, ei) == 1 && 3034f7ba2d37SJosef Bacik !(btrfs_extent_flags(eb, ei) & 3035f7ba2d37SJosef Bacik BTRFS_BLOCK_FLAG_FULL_BACKREF) && 3036f7ba2d37SJosef Bacik ptr < end) { 3037f7ba2d37SJosef Bacik struct btrfs_extent_inline_ref *iref; 3038f7ba2d37SJosef Bacik int type; 3039f7ba2d37SJosef Bacik 3040f7ba2d37SJosef Bacik iref = (struct btrfs_extent_inline_ref *)ptr; 3041f7ba2d37SJosef Bacik type = btrfs_get_extent_inline_ref_type(eb, iref, 3042f7ba2d37SJosef Bacik BTRFS_REF_TYPE_BLOCK); 3043f7ba2d37SJosef Bacik if (type == BTRFS_REF_TYPE_INVALID) 3044f7ba2d37SJosef Bacik return -EINVAL; 3045f7ba2d37SJosef Bacik if (type == BTRFS_TREE_BLOCK_REF_KEY) 3046f7ba2d37SJosef Bacik owner = btrfs_extent_inline_ref_offset(eb, iref); 3047f7ba2d37SJosef Bacik } 30486d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3049ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3050ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3051ba3c2b19SNikolay Borisov return -EINVAL; 30523173a18fSJosef Bacik } else { 30535d4f98a2SYan Zheng BUG(); 30545d4f98a2SYan Zheng } 30555d4f98a2SYan Zheng 3056b3b4aa74SDavid Sterba btrfs_release_path(path); 30575d4f98a2SYan Zheng 30585d4f98a2SYan Zheng BUG_ON(level == -1); 30595d4f98a2SYan Zheng 30605d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 30615d4f98a2SYan Zheng if (!block) 30625d4f98a2SYan Zheng return -ENOMEM; 30635d4f98a2SYan Zheng 30645d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3065da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 30665d4f98a2SYan Zheng block->key.offset = generation; 30675d4f98a2SYan Zheng block->level = level; 30685d4f98a2SYan Zheng block->key_ready = 0; 3069f7ba2d37SJosef Bacik block->owner = owner; 30705d4f98a2SYan Zheng 3071e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 307243c04fb1SJeff Mahoney if (rb_node) 3073982c92cbSQu Wenruo btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr, 3074982c92cbSQu Wenruo -EEXIST); 30755d4f98a2SYan Zheng 30765d4f98a2SYan Zheng return 0; 30775d4f98a2SYan Zheng } 30785d4f98a2SYan Zheng 30795d4f98a2SYan Zheng /* 30805d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 30815d4f98a2SYan Zheng */ 30825d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 30835d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 30845d4f98a2SYan Zheng struct rb_root *blocks) 30855d4f98a2SYan Zheng { 30860b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 30875d4f98a2SYan Zheng struct btrfs_path *path; 30885d4f98a2SYan Zheng struct btrfs_key key; 30895d4f98a2SYan Zheng int ret; 30900b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 30915d4f98a2SYan Zheng 30927476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 30935d4f98a2SYan Zheng return 0; 30945d4f98a2SYan Zheng 3095e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 30965d4f98a2SYan Zheng return 0; 30975d4f98a2SYan Zheng 30985d4f98a2SYan Zheng path = btrfs_alloc_path(); 30995d4f98a2SYan Zheng if (!path) 31005d4f98a2SYan Zheng return -ENOMEM; 3101aee68ee5SJosef Bacik again: 31025d4f98a2SYan Zheng key.objectid = bytenr; 3103aee68ee5SJosef Bacik if (skinny) { 3104aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3105aee68ee5SJosef Bacik key.offset = (u64)-1; 3106aee68ee5SJosef Bacik } else { 31075d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31085d4f98a2SYan Zheng key.offset = blocksize; 3109aee68ee5SJosef Bacik } 31105d4f98a2SYan Zheng 31115d4f98a2SYan Zheng path->search_commit_root = 1; 31125d4f98a2SYan Zheng path->skip_locking = 1; 31135d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 31145d4f98a2SYan Zheng if (ret < 0) 31155d4f98a2SYan Zheng goto out; 31165d4f98a2SYan Zheng 3117aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3118aee68ee5SJosef Bacik if (path->slots[0]) { 3119aee68ee5SJosef Bacik path->slots[0]--; 3120aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3121aee68ee5SJosef Bacik path->slots[0]); 31223173a18fSJosef Bacik if (key.objectid == bytenr && 3123aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3124aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3125aee68ee5SJosef Bacik key.offset == blocksize))) 31263173a18fSJosef Bacik ret = 0; 31273173a18fSJosef Bacik } 3128aee68ee5SJosef Bacik 3129aee68ee5SJosef Bacik if (ret) { 3130aee68ee5SJosef Bacik skinny = false; 3131aee68ee5SJosef Bacik btrfs_release_path(path); 3132aee68ee5SJosef Bacik goto again; 3133aee68ee5SJosef Bacik } 3134aee68ee5SJosef Bacik } 3135cdccee99SLiu Bo if (ret) { 3136cdccee99SLiu Bo ASSERT(ret == 1); 3137cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3138cdccee99SLiu Bo btrfs_err(fs_info, 3139cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3140cdccee99SLiu Bo bytenr); 3141cdccee99SLiu Bo WARN_ON(1); 3142cdccee99SLiu Bo ret = -EINVAL; 3143cdccee99SLiu Bo goto out; 3144cdccee99SLiu Bo } 31453173a18fSJosef Bacik 31465d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 31475d4f98a2SYan Zheng out: 31485d4f98a2SYan Zheng btrfs_free_path(path); 31495d4f98a2SYan Zheng return ret; 31505d4f98a2SYan Zheng } 31515d4f98a2SYan Zheng 31520af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 315332da5386SDavid Sterba struct btrfs_block_group *block_group, 31541bbc621eSChris Mason struct inode *inode, 31551bbc621eSChris Mason u64 ino) 31560af3d00bSJosef Bacik { 31570af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 31580af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 31590af3d00bSJosef Bacik int ret = 0; 31600af3d00bSJosef Bacik 31610af3d00bSJosef Bacik if (inode) 31620af3d00bSJosef Bacik goto truncate; 31630af3d00bSJosef Bacik 31640202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, ino, root); 31652e19f1f9SAl Viro if (IS_ERR(inode)) 31660af3d00bSJosef Bacik return -ENOENT; 31670af3d00bSJosef Bacik 31680af3d00bSJosef Bacik truncate: 31692ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 31707b61cd92SMiao Xie &fs_info->global_block_rsv); 31717b61cd92SMiao Xie if (ret) 31727b61cd92SMiao Xie goto out; 31737b61cd92SMiao Xie 31747a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 31750af3d00bSJosef Bacik if (IS_ERR(trans)) { 31763612b495STsutomu Itoh ret = PTR_ERR(trans); 31770af3d00bSJosef Bacik goto out; 31780af3d00bSJosef Bacik } 31790af3d00bSJosef Bacik 318077ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 31810af3d00bSJosef Bacik 31823a45bb20SJeff Mahoney btrfs_end_transaction(trans); 31832ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 31840af3d00bSJosef Bacik out: 31850af3d00bSJosef Bacik iput(inode); 31860af3d00bSJosef Bacik return ret; 31870af3d00bSJosef Bacik } 31880af3d00bSJosef Bacik 31895d4f98a2SYan Zheng /* 319019b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 319119b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 31925d4f98a2SYan Zheng */ 319319b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 319419b546d7SQu Wenruo struct btrfs_block_group *block_group, 319519b546d7SQu Wenruo u64 data_bytenr) 31965d4f98a2SYan Zheng { 319719b546d7SQu Wenruo u64 space_cache_ino; 319819b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 31995d4f98a2SYan Zheng struct btrfs_key key; 320019b546d7SQu Wenruo bool found = false; 320119b546d7SQu Wenruo int i; 32025d4f98a2SYan Zheng int ret; 32035d4f98a2SYan Zheng 320419b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 320519b546d7SQu Wenruo return 0; 32065d4f98a2SYan Zheng 320719b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 320850e31ef4SQu Wenruo u8 type; 320950e31ef4SQu Wenruo 321019b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 321119b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 321219b546d7SQu Wenruo continue; 321319b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 321450e31ef4SQu Wenruo type = btrfs_file_extent_type(leaf, ei); 321550e31ef4SQu Wenruo 321650e31ef4SQu Wenruo if ((type == BTRFS_FILE_EXTENT_REG || 321750e31ef4SQu Wenruo type == BTRFS_FILE_EXTENT_PREALLOC) && 321819b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 321919b546d7SQu Wenruo found = true; 322019b546d7SQu Wenruo space_cache_ino = key.objectid; 322119b546d7SQu Wenruo break; 322219b546d7SQu Wenruo } 322319b546d7SQu Wenruo } 322419b546d7SQu Wenruo if (!found) 322519b546d7SQu Wenruo return -ENOENT; 322619b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 322719b546d7SQu Wenruo space_cache_ino); 32280af3d00bSJosef Bacik return ret; 32295d4f98a2SYan Zheng } 32305d4f98a2SYan Zheng 32315d4f98a2SYan Zheng /* 32322c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 32335d4f98a2SYan Zheng */ 32345d4f98a2SYan Zheng static noinline_for_stack 32355d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 32365d4f98a2SYan Zheng struct btrfs_key *extent_key, 32375d4f98a2SYan Zheng struct btrfs_path *path, 32385d4f98a2SYan Zheng struct rb_root *blocks) 32395d4f98a2SYan Zheng { 324019b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 324119b546d7SQu Wenruo struct ulist *leaves = NULL; 324219b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 324319b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 324419b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3245647f63bdSFilipe David Borba Manana int ret = 0; 32465d4f98a2SYan Zheng 3247b3b4aa74SDavid Sterba btrfs_release_path(path); 324819b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 324919b546d7SQu Wenruo 0, &leaves, NULL, true); 325019b546d7SQu Wenruo if (ret < 0) 325119b546d7SQu Wenruo return ret; 325219b546d7SQu Wenruo 325319b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 325419b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 325519b546d7SQu Wenruo struct extent_buffer *eb; 325619b546d7SQu Wenruo 32571b7ec85eSJosef Bacik eb = read_tree_block(fs_info, ref_node->val, 0, 0, 0, NULL); 325819b546d7SQu Wenruo if (IS_ERR(eb)) { 325919b546d7SQu Wenruo ret = PTR_ERR(eb); 326019b546d7SQu Wenruo break; 326119b546d7SQu Wenruo } 326219b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 326319b546d7SQu Wenruo extent_key->objectid); 326419b546d7SQu Wenruo free_extent_buffer(eb); 326519b546d7SQu Wenruo if (ret < 0) 326619b546d7SQu Wenruo break; 326719b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 326819b546d7SQu Wenruo if (ret < 0) 326919b546d7SQu Wenruo break; 327019b546d7SQu Wenruo } 327119b546d7SQu Wenruo if (ret < 0) 32725d4f98a2SYan Zheng free_block_list(blocks); 327319b546d7SQu Wenruo ulist_free(leaves); 327419b546d7SQu Wenruo return ret; 32755d4f98a2SYan Zheng } 32765d4f98a2SYan Zheng 32775d4f98a2SYan Zheng /* 32782c016dc2SLiu Bo * helper to find next unprocessed extent 32795d4f98a2SYan Zheng */ 32805d4f98a2SYan Zheng static noinline_for_stack 3281147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 32823fd0a558SYan, Zheng struct btrfs_key *extent_key) 32835d4f98a2SYan Zheng { 32840b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 32855d4f98a2SYan Zheng struct btrfs_key key; 32865d4f98a2SYan Zheng struct extent_buffer *leaf; 32875d4f98a2SYan Zheng u64 start, end, last; 32885d4f98a2SYan Zheng int ret; 32895d4f98a2SYan Zheng 3290b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 32915d4f98a2SYan Zheng while (1) { 32925d4f98a2SYan Zheng cond_resched(); 32935d4f98a2SYan Zheng if (rc->search_start >= last) { 32945d4f98a2SYan Zheng ret = 1; 32955d4f98a2SYan Zheng break; 32965d4f98a2SYan Zheng } 32975d4f98a2SYan Zheng 32985d4f98a2SYan Zheng key.objectid = rc->search_start; 32995d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 33005d4f98a2SYan Zheng key.offset = 0; 33015d4f98a2SYan Zheng 33025d4f98a2SYan Zheng path->search_commit_root = 1; 33035d4f98a2SYan Zheng path->skip_locking = 1; 33045d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 33055d4f98a2SYan Zheng 0, 0); 33065d4f98a2SYan Zheng if (ret < 0) 33075d4f98a2SYan Zheng break; 33085d4f98a2SYan Zheng next: 33095d4f98a2SYan Zheng leaf = path->nodes[0]; 33105d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 33115d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 33125d4f98a2SYan Zheng if (ret != 0) 33135d4f98a2SYan Zheng break; 33145d4f98a2SYan Zheng leaf = path->nodes[0]; 33155d4f98a2SYan Zheng } 33165d4f98a2SYan Zheng 33175d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 33185d4f98a2SYan Zheng if (key.objectid >= last) { 33195d4f98a2SYan Zheng ret = 1; 33205d4f98a2SYan Zheng break; 33215d4f98a2SYan Zheng } 33225d4f98a2SYan Zheng 33233173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 33243173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 33253173a18fSJosef Bacik path->slots[0]++; 33263173a18fSJosef Bacik goto next; 33273173a18fSJosef Bacik } 33283173a18fSJosef Bacik 33293173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 33305d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 33315d4f98a2SYan Zheng path->slots[0]++; 33325d4f98a2SYan Zheng goto next; 33335d4f98a2SYan Zheng } 33345d4f98a2SYan Zheng 33353173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 33360b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 33373173a18fSJosef Bacik rc->search_start) { 33383173a18fSJosef Bacik path->slots[0]++; 33393173a18fSJosef Bacik goto next; 33403173a18fSJosef Bacik } 33413173a18fSJosef Bacik 33425d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 33435d4f98a2SYan Zheng key.objectid, &start, &end, 3344e6138876SJosef Bacik EXTENT_DIRTY, NULL); 33455d4f98a2SYan Zheng 33465d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3347b3b4aa74SDavid Sterba btrfs_release_path(path); 33485d4f98a2SYan Zheng rc->search_start = end + 1; 33495d4f98a2SYan Zheng } else { 33503173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 33515d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 33523173a18fSJosef Bacik else 33533173a18fSJosef Bacik rc->search_start = key.objectid + 33540b246afaSJeff Mahoney fs_info->nodesize; 33553fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 33565d4f98a2SYan Zheng return 0; 33575d4f98a2SYan Zheng } 33585d4f98a2SYan Zheng } 3359b3b4aa74SDavid Sterba btrfs_release_path(path); 33605d4f98a2SYan Zheng return ret; 33615d4f98a2SYan Zheng } 33625d4f98a2SYan Zheng 33635d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 33645d4f98a2SYan Zheng { 33655d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33667585717fSChris Mason 33677585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 33685d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 33697585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 33705d4f98a2SYan Zheng } 33715d4f98a2SYan Zheng 33725d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 33735d4f98a2SYan Zheng { 33745d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33757585717fSChris Mason 33767585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 33775d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 33787585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 33795d4f98a2SYan Zheng } 33805d4f98a2SYan Zheng 33815d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 33825d4f98a2SYan Zheng { 33835d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 33845d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 33855d4f98a2SYan Zheng return 1; 33865d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 33875d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 33885d4f98a2SYan Zheng return 1; 33895d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 33905d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 33915d4f98a2SYan Zheng return 1; 33925d4f98a2SYan Zheng return 0; 33935d4f98a2SYan Zheng } 33945d4f98a2SYan Zheng 33953fd0a558SYan, Zheng static noinline_for_stack 33963fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 33973fd0a558SYan, Zheng { 33983fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3399ac2fabacSJosef Bacik int ret; 34003fd0a558SYan, Zheng 34012ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 340266d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 34033fd0a558SYan, Zheng if (!rc->block_rsv) 34043fd0a558SYan, Zheng return -ENOMEM; 34053fd0a558SYan, Zheng 34063fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3407b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 34083fd0a558SYan, Zheng rc->extents_found = 0; 34093fd0a558SYan, Zheng rc->nodes_relocated = 0; 34103fd0a558SYan, Zheng rc->merging_rsv_size = 0; 34110647bf56SWang Shilong rc->reserved_bytes = 0; 3412da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 34130647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3414ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3415ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3416ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3417ac2fabacSJosef Bacik if (ret) 3418ac2fabacSJosef Bacik return ret; 34193fd0a558SYan, Zheng 34203fd0a558SYan, Zheng rc->create_reloc_tree = 1; 34213fd0a558SYan, Zheng set_reloc_control(rc); 34223fd0a558SYan, Zheng 34237a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 342428818947SLiu Bo if (IS_ERR(trans)) { 342528818947SLiu Bo unset_reloc_control(rc); 342628818947SLiu Bo /* 342728818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 342828818947SLiu Bo * cleanup. And callers are responsible to free the above 342928818947SLiu Bo * block rsv. 343028818947SLiu Bo */ 343128818947SLiu Bo return PTR_ERR(trans); 343228818947SLiu Bo } 34333a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 34343fd0a558SYan, Zheng return 0; 34353fd0a558SYan, Zheng } 343676dda93cSYan, Zheng 34375d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 34385d4f98a2SYan Zheng { 34392ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 34405d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 34415d4f98a2SYan Zheng struct btrfs_key key; 34425d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 34435d4f98a2SYan Zheng struct btrfs_path *path; 34445d4f98a2SYan Zheng struct btrfs_extent_item *ei; 34455d4f98a2SYan Zheng u64 flags; 34465d4f98a2SYan Zheng u32 item_size; 34475d4f98a2SYan Zheng int ret; 34485d4f98a2SYan Zheng int err = 0; 3449c87f08caSChris Mason int progress = 0; 34505d4f98a2SYan Zheng 34515d4f98a2SYan Zheng path = btrfs_alloc_path(); 34523fd0a558SYan, Zheng if (!path) 34535d4f98a2SYan Zheng return -ENOMEM; 3454e4058b54SDavid Sterba path->reada = READA_FORWARD; 34553fd0a558SYan, Zheng 34563fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 34573fd0a558SYan, Zheng if (ret) { 34583fd0a558SYan, Zheng err = ret; 34593fd0a558SYan, Zheng goto out_free; 34602423fdfbSJiri Slaby } 34615d4f98a2SYan Zheng 34625d4f98a2SYan Zheng while (1) { 34630647bf56SWang Shilong rc->reserved_bytes = 0; 34640647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 34650647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 34660647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 34670647bf56SWang Shilong if (ret) { 34680647bf56SWang Shilong err = ret; 34690647bf56SWang Shilong break; 34700647bf56SWang Shilong } 3471c87f08caSChris Mason progress++; 3472a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 34730f788c58SLiu Bo if (IS_ERR(trans)) { 34740f788c58SLiu Bo err = PTR_ERR(trans); 34750f788c58SLiu Bo trans = NULL; 34760f788c58SLiu Bo break; 34770f788c58SLiu Bo } 3478c87f08caSChris Mason restart: 34793fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 34803a45bb20SJeff Mahoney btrfs_end_transaction(trans); 348142a657f5SPan Bian trans = NULL; 34823fd0a558SYan, Zheng continue; 34833fd0a558SYan, Zheng } 34843fd0a558SYan, Zheng 3485147d256eSZhaolei ret = find_next_extent(rc, path, &key); 34865d4f98a2SYan Zheng if (ret < 0) 34875d4f98a2SYan Zheng err = ret; 34885d4f98a2SYan Zheng if (ret != 0) 34895d4f98a2SYan Zheng break; 34905d4f98a2SYan Zheng 34915d4f98a2SYan Zheng rc->extents_found++; 34925d4f98a2SYan Zheng 34935d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 34945d4f98a2SYan Zheng struct btrfs_extent_item); 34953fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 34965d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 34975d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 34985d4f98a2SYan Zheng ret = check_extent_flags(flags); 34995d4f98a2SYan Zheng BUG_ON(ret); 35006d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3501ba3c2b19SNikolay Borisov err = -EINVAL; 3502ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 3503ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 3504ba3c2b19SNikolay Borisov break; 35055d4f98a2SYan Zheng } else { 35065d4f98a2SYan Zheng BUG(); 35075d4f98a2SYan Zheng } 35085d4f98a2SYan Zheng 35095d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 35105d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 35115d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 35125d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 35135d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 35145d4f98a2SYan Zheng } else { 3515b3b4aa74SDavid Sterba btrfs_release_path(path); 35165d4f98a2SYan Zheng ret = 0; 35175d4f98a2SYan Zheng } 35185d4f98a2SYan Zheng if (ret < 0) { 35193fd0a558SYan, Zheng err = ret; 35205d4f98a2SYan Zheng break; 35215d4f98a2SYan Zheng } 35225d4f98a2SYan Zheng 35235d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 35245d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 35255d4f98a2SYan Zheng if (ret < 0) { 35263fd0a558SYan, Zheng if (ret != -EAGAIN) { 35275d4f98a2SYan Zheng err = ret; 35285d4f98a2SYan Zheng break; 35295d4f98a2SYan Zheng } 35303fd0a558SYan, Zheng rc->extents_found--; 35313fd0a558SYan, Zheng rc->search_start = key.objectid; 35323fd0a558SYan, Zheng } 35335d4f98a2SYan Zheng } 35345d4f98a2SYan Zheng 35353a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 35362ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 35373fd0a558SYan, Zheng trans = NULL; 35385d4f98a2SYan Zheng 35395d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 35405d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 35415d4f98a2SYan Zheng rc->found_file_extent = 1; 35420257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 35433fd0a558SYan, Zheng &key, &rc->cluster); 35445d4f98a2SYan Zheng if (ret < 0) { 35455d4f98a2SYan Zheng err = ret; 35465d4f98a2SYan Zheng break; 35475d4f98a2SYan Zheng } 35485d4f98a2SYan Zheng } 3549f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 3550f31ea088SQu Wenruo err = -ECANCELED; 3551f31ea088SQu Wenruo break; 3552f31ea088SQu Wenruo } 35535d4f98a2SYan Zheng } 3554c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 355543a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 35569689457bSShilong Wang if (ret == 1) { 3557c87f08caSChris Mason err = 0; 3558c87f08caSChris Mason progress = 0; 3559c87f08caSChris Mason goto restart; 3560c87f08caSChris Mason } 3561c87f08caSChris Mason } 35623fd0a558SYan, Zheng 3563b3b4aa74SDavid Sterba btrfs_release_path(path); 356491166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 35655d4f98a2SYan Zheng 35665d4f98a2SYan Zheng if (trans) { 35673a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 35682ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 35695d4f98a2SYan Zheng } 35705d4f98a2SYan Zheng 35710257bb82SYan, Zheng if (!err) { 35723fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 35733fd0a558SYan, Zheng &rc->cluster); 35740257bb82SYan, Zheng if (ret < 0) 35750257bb82SYan, Zheng err = ret; 35760257bb82SYan, Zheng } 35770257bb82SYan, Zheng 35783fd0a558SYan, Zheng rc->create_reloc_tree = 0; 35793fd0a558SYan, Zheng set_reloc_control(rc); 35800257bb82SYan, Zheng 358113fe1bdbSQu Wenruo btrfs_backref_release_cache(&rc->backref_cache); 358263f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 35835d4f98a2SYan Zheng 35847f913c7cSQu Wenruo /* 35857f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 35867f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 35877f913c7cSQu Wenruo * 35887f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 35897f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 35907f913c7cSQu Wenruo * merge_reloc_roots() 35917f913c7cSQu Wenruo */ 35923fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 35935d4f98a2SYan Zheng 35945d4f98a2SYan Zheng merge_reloc_roots(rc); 35955d4f98a2SYan Zheng 35963fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 35975d4f98a2SYan Zheng unset_reloc_control(rc); 359863f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 35995d4f98a2SYan Zheng 36005d4f98a2SYan Zheng /* get rid of pinned extents */ 36017a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 360262b99540SQu Wenruo if (IS_ERR(trans)) { 36033612b495STsutomu Itoh err = PTR_ERR(trans); 360462b99540SQu Wenruo goto out_free; 360562b99540SQu Wenruo } 36063a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 36076217b0faSJosef Bacik out_free: 3608d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3609d2311e69SQu Wenruo if (ret < 0 && !err) 3610d2311e69SQu Wenruo err = ret; 36112ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 36123fd0a558SYan, Zheng btrfs_free_path(path); 36135d4f98a2SYan Zheng return err; 36145d4f98a2SYan Zheng } 36155d4f98a2SYan Zheng 36165d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 36170257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 36185d4f98a2SYan Zheng { 36195d4f98a2SYan Zheng struct btrfs_path *path; 36205d4f98a2SYan Zheng struct btrfs_inode_item *item; 36215d4f98a2SYan Zheng struct extent_buffer *leaf; 362232430c61SNaohiro Aota u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC; 36235d4f98a2SYan Zheng int ret; 36245d4f98a2SYan Zheng 362532430c61SNaohiro Aota if (btrfs_is_zoned(trans->fs_info)) 362632430c61SNaohiro Aota flags &= ~BTRFS_INODE_PREALLOC; 362732430c61SNaohiro Aota 36285d4f98a2SYan Zheng path = btrfs_alloc_path(); 36295d4f98a2SYan Zheng if (!path) 36305d4f98a2SYan Zheng return -ENOMEM; 36315d4f98a2SYan Zheng 36325d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 36335d4f98a2SYan Zheng if (ret) 36345d4f98a2SYan Zheng goto out; 36355d4f98a2SYan Zheng 36365d4f98a2SYan Zheng leaf = path->nodes[0]; 36375d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 3638b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 36395d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 36400257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 36415d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 364232430c61SNaohiro Aota btrfs_set_inode_flags(leaf, item, flags); 36435d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 36445d4f98a2SYan Zheng out: 36455d4f98a2SYan Zheng btrfs_free_path(path); 36465d4f98a2SYan Zheng return ret; 36475d4f98a2SYan Zheng } 36485d4f98a2SYan Zheng 36495d4f98a2SYan Zheng /* 36505d4f98a2SYan Zheng * helper to create inode for data relocation. 36515d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 36525d4f98a2SYan Zheng */ 36533fd0a558SYan, Zheng static noinline_for_stack 36543fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 365532da5386SDavid Sterba struct btrfs_block_group *group) 36565d4f98a2SYan Zheng { 36575d4f98a2SYan Zheng struct inode *inode = NULL; 36585d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 36595d4f98a2SYan Zheng struct btrfs_root *root; 36604624900dSZhaolei u64 objectid; 36615d4f98a2SYan Zheng int err = 0; 36625d4f98a2SYan Zheng 3663aeb935a4SQu Wenruo root = btrfs_grab_root(fs_info->data_reloc_root); 3664a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 366576deacf0SJosef Bacik if (IS_ERR(trans)) { 366600246528SJosef Bacik btrfs_put_root(root); 36673fd0a558SYan, Zheng return ERR_CAST(trans); 366876deacf0SJosef Bacik } 36695d4f98a2SYan Zheng 3670543068a2SNikolay Borisov err = btrfs_get_free_objectid(root, &objectid); 36715d4f98a2SYan Zheng if (err) 36725d4f98a2SYan Zheng goto out; 36735d4f98a2SYan Zheng 36740257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 36755d4f98a2SYan Zheng BUG_ON(err); 36765d4f98a2SYan Zheng 36770202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, objectid, root); 36782e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 3679b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 36805d4f98a2SYan Zheng 368173f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 36825d4f98a2SYan Zheng out: 368300246528SJosef Bacik btrfs_put_root(root); 36843a45bb20SJeff Mahoney btrfs_end_transaction(trans); 36852ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 36865d4f98a2SYan Zheng if (err) { 36875d4f98a2SYan Zheng if (inode) 36885d4f98a2SYan Zheng iput(inode); 36895d4f98a2SYan Zheng inode = ERR_PTR(err); 36905d4f98a2SYan Zheng } 36915d4f98a2SYan Zheng return inode; 36925d4f98a2SYan Zheng } 36935d4f98a2SYan Zheng 3694c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 36953fd0a558SYan, Zheng { 36963fd0a558SYan, Zheng struct reloc_control *rc; 36973fd0a558SYan, Zheng 36983fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 36993fd0a558SYan, Zheng if (!rc) 37003fd0a558SYan, Zheng return NULL; 37013fd0a558SYan, Zheng 37023fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 3703d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 3704584fb121SQu Wenruo btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1); 37053fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 370643eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 370743eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 37083fd0a558SYan, Zheng return rc; 37093fd0a558SYan, Zheng } 37103fd0a558SYan, Zheng 37111a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 37121a0afa0eSJosef Bacik { 37131a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 37141a0afa0eSJosef Bacik 37151a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 37161a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 37171a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 37181a0afa0eSJosef Bacik kfree(node); 37191a0afa0eSJosef Bacik 37201a0afa0eSJosef Bacik kfree(rc); 37211a0afa0eSJosef Bacik } 37221a0afa0eSJosef Bacik 37235d4f98a2SYan Zheng /* 3724ebce0e01SAdam Borowski * Print the block group being relocated 3725ebce0e01SAdam Borowski */ 3726ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 372732da5386SDavid Sterba struct btrfs_block_group *block_group) 3728ebce0e01SAdam Borowski { 3729f89e09cfSAnand Jain char buf[128] = {'\0'}; 3730ebce0e01SAdam Borowski 3731f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 3732ebce0e01SAdam Borowski 3733ebce0e01SAdam Borowski btrfs_info(fs_info, 3734ebce0e01SAdam Borowski "relocating block group %llu flags %s", 3735b3470b5dSDavid Sterba block_group->start, buf); 3736ebce0e01SAdam Borowski } 3737ebce0e01SAdam Borowski 3738430640e3SQu Wenruo static const char *stage_to_string(int stage) 3739430640e3SQu Wenruo { 3740430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 3741430640e3SQu Wenruo return "move data extents"; 3742430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 3743430640e3SQu Wenruo return "update data pointers"; 3744430640e3SQu Wenruo return "unknown"; 3745430640e3SQu Wenruo } 3746430640e3SQu Wenruo 3747ebce0e01SAdam Borowski /* 37485d4f98a2SYan Zheng * function to relocate all extents in a block group. 37495d4f98a2SYan Zheng */ 37506bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 37515d4f98a2SYan Zheng { 375232da5386SDavid Sterba struct btrfs_block_group *bg; 37536bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 37545d4f98a2SYan Zheng struct reloc_control *rc; 37550af3d00bSJosef Bacik struct inode *inode; 37560af3d00bSJosef Bacik struct btrfs_path *path; 37575d4f98a2SYan Zheng int ret; 3758f0486c68SYan, Zheng int rw = 0; 37595d4f98a2SYan Zheng int err = 0; 37605d4f98a2SYan Zheng 3761eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 3762eede2bf3SOmar Sandoval if (!bg) 3763eede2bf3SOmar Sandoval return -ENOENT; 3764eede2bf3SOmar Sandoval 3765eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 3766eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 3767eede2bf3SOmar Sandoval return -ETXTBSY; 3768eede2bf3SOmar Sandoval } 3769eede2bf3SOmar Sandoval 3770c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 3771eede2bf3SOmar Sandoval if (!rc) { 3772eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 37735d4f98a2SYan Zheng return -ENOMEM; 3774eede2bf3SOmar Sandoval } 37755d4f98a2SYan Zheng 3776f0486c68SYan, Zheng rc->extent_root = extent_root; 3777eede2bf3SOmar Sandoval rc->block_group = bg; 37785d4f98a2SYan Zheng 3779b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 3780f0486c68SYan, Zheng if (ret) { 3781f0486c68SYan, Zheng err = ret; 3782f0486c68SYan, Zheng goto out; 3783f0486c68SYan, Zheng } 3784f0486c68SYan, Zheng rw = 1; 3785f0486c68SYan, Zheng 37860af3d00bSJosef Bacik path = btrfs_alloc_path(); 37870af3d00bSJosef Bacik if (!path) { 37880af3d00bSJosef Bacik err = -ENOMEM; 37890af3d00bSJosef Bacik goto out; 37900af3d00bSJosef Bacik } 37910af3d00bSJosef Bacik 37927949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 37930af3d00bSJosef Bacik btrfs_free_path(path); 37940af3d00bSJosef Bacik 37950af3d00bSJosef Bacik if (!IS_ERR(inode)) 37961bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 37970af3d00bSJosef Bacik else 37980af3d00bSJosef Bacik ret = PTR_ERR(inode); 37990af3d00bSJosef Bacik 38000af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 38010af3d00bSJosef Bacik err = ret; 38020af3d00bSJosef Bacik goto out; 38030af3d00bSJosef Bacik } 38040af3d00bSJosef Bacik 38055d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 38065d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 38075d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 38085d4f98a2SYan Zheng rc->data_inode = NULL; 38095d4f98a2SYan Zheng goto out; 38105d4f98a2SYan Zheng } 38115d4f98a2SYan Zheng 38120b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 38135d4f98a2SYan Zheng 38149cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 3815f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 38166374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 3817b3470b5dSDavid Sterba rc->block_group->start, 3818b3470b5dSDavid Sterba rc->block_group->length); 38195d4f98a2SYan Zheng 38205d4f98a2SYan Zheng while (1) { 3821430640e3SQu Wenruo int finishes_stage; 3822430640e3SQu Wenruo 382376dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 38245d4f98a2SYan Zheng ret = relocate_block_group(rc); 382576dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 3826ff612ba7SJosef Bacik if (ret < 0) 38275d4f98a2SYan Zheng err = ret; 3828ff612ba7SJosef Bacik 3829430640e3SQu Wenruo finishes_stage = rc->stage; 3830ff612ba7SJosef Bacik /* 3831ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 3832ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 3833ff612ba7SJosef Bacik * different block group we could end up hitting the 3834ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 3835ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 3836ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 3837ff612ba7SJosef Bacik * out of the loop if we hit an error. 3838ff612ba7SJosef Bacik */ 3839ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 3840ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 3841ff612ba7SJosef Bacik (u64)-1); 3842ff612ba7SJosef Bacik if (ret) 3843ff612ba7SJosef Bacik err = ret; 3844ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 3845ff612ba7SJosef Bacik 0, -1); 3846ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 38475d4f98a2SYan Zheng } 38485d4f98a2SYan Zheng 3849ff612ba7SJosef Bacik if (err < 0) 3850ff612ba7SJosef Bacik goto out; 3851ff612ba7SJosef Bacik 38525d4f98a2SYan Zheng if (rc->extents_found == 0) 38535d4f98a2SYan Zheng break; 38545d4f98a2SYan Zheng 3855430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 3856430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 38575d4f98a2SYan Zheng } 38585d4f98a2SYan Zheng 38595d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 38605d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 3861bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 38625d4f98a2SYan Zheng out: 3863f0486c68SYan, Zheng if (err && rw) 38642ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 38655d4f98a2SYan Zheng iput(rc->data_inode); 38665d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 38671a0afa0eSJosef Bacik free_reloc_control(rc); 38685d4f98a2SYan Zheng return err; 38695d4f98a2SYan Zheng } 38705d4f98a2SYan Zheng 387176dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 387276dda93cSYan, Zheng { 38730b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 387476dda93cSYan, Zheng struct btrfs_trans_handle *trans; 387579787eaaSJeff Mahoney int ret, err; 387676dda93cSYan, Zheng 38770b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 387879787eaaSJeff Mahoney if (IS_ERR(trans)) 387979787eaaSJeff Mahoney return PTR_ERR(trans); 388076dda93cSYan, Zheng 388176dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 388276dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 3883c8422684SDavid Sterba btrfs_set_root_drop_level(&root->root_item, 0); 388476dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 38850b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 388676dda93cSYan, Zheng &root->root_key, &root->root_item); 388776dda93cSYan, Zheng 38883a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 388979787eaaSJeff Mahoney if (err) 389079787eaaSJeff Mahoney return err; 389179787eaaSJeff Mahoney return ret; 389276dda93cSYan, Zheng } 389376dda93cSYan, Zheng 38945d4f98a2SYan Zheng /* 38955d4f98a2SYan Zheng * recover relocation interrupted by system crash. 38965d4f98a2SYan Zheng * 38975d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 38985d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 38995d4f98a2SYan Zheng */ 39005d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 39015d4f98a2SYan Zheng { 39020b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 39035d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 39045d4f98a2SYan Zheng struct btrfs_key key; 39055d4f98a2SYan Zheng struct btrfs_root *fs_root; 39065d4f98a2SYan Zheng struct btrfs_root *reloc_root; 39075d4f98a2SYan Zheng struct btrfs_path *path; 39085d4f98a2SYan Zheng struct extent_buffer *leaf; 39095d4f98a2SYan Zheng struct reloc_control *rc = NULL; 39105d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 39115d4f98a2SYan Zheng int ret; 39125d4f98a2SYan Zheng int err = 0; 39135d4f98a2SYan Zheng 39145d4f98a2SYan Zheng path = btrfs_alloc_path(); 39155d4f98a2SYan Zheng if (!path) 39165d4f98a2SYan Zheng return -ENOMEM; 3917e4058b54SDavid Sterba path->reada = READA_BACK; 39185d4f98a2SYan Zheng 39195d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 39205d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 39215d4f98a2SYan Zheng key.offset = (u64)-1; 39225d4f98a2SYan Zheng 39235d4f98a2SYan Zheng while (1) { 39240b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 39255d4f98a2SYan Zheng path, 0, 0); 39265d4f98a2SYan Zheng if (ret < 0) { 39275d4f98a2SYan Zheng err = ret; 39285d4f98a2SYan Zheng goto out; 39295d4f98a2SYan Zheng } 39305d4f98a2SYan Zheng if (ret > 0) { 39315d4f98a2SYan Zheng if (path->slots[0] == 0) 39325d4f98a2SYan Zheng break; 39335d4f98a2SYan Zheng path->slots[0]--; 39345d4f98a2SYan Zheng } 39355d4f98a2SYan Zheng leaf = path->nodes[0]; 39365d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 3937b3b4aa74SDavid Sterba btrfs_release_path(path); 39385d4f98a2SYan Zheng 39395d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 39405d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 39415d4f98a2SYan Zheng break; 39425d4f98a2SYan Zheng 39433dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 39445d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 39455d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 39465d4f98a2SYan Zheng goto out; 39475d4f98a2SYan Zheng } 39485d4f98a2SYan Zheng 394992a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 39505d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 39515d4f98a2SYan Zheng 39525d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 3953a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, 3954a820feb5SDavid Sterba reloc_root->root_key.offset, false); 39555d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 395676dda93cSYan, Zheng ret = PTR_ERR(fs_root); 395776dda93cSYan, Zheng if (ret != -ENOENT) { 395876dda93cSYan, Zheng err = ret; 39595d4f98a2SYan Zheng goto out; 39605d4f98a2SYan Zheng } 396179787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 396279787eaaSJeff Mahoney if (ret < 0) { 396379787eaaSJeff Mahoney err = ret; 396479787eaaSJeff Mahoney goto out; 396579787eaaSJeff Mahoney } 3966932fd26dSJosef Bacik } else { 396700246528SJosef Bacik btrfs_put_root(fs_root); 396876dda93cSYan, Zheng } 39695d4f98a2SYan Zheng } 39705d4f98a2SYan Zheng 39715d4f98a2SYan Zheng if (key.offset == 0) 39725d4f98a2SYan Zheng break; 39735d4f98a2SYan Zheng 39745d4f98a2SYan Zheng key.offset--; 39755d4f98a2SYan Zheng } 3976b3b4aa74SDavid Sterba btrfs_release_path(path); 39775d4f98a2SYan Zheng 39785d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 39795d4f98a2SYan Zheng goto out; 39805d4f98a2SYan Zheng 3981c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 39825d4f98a2SYan Zheng if (!rc) { 39835d4f98a2SYan Zheng err = -ENOMEM; 39845d4f98a2SYan Zheng goto out; 39855d4f98a2SYan Zheng } 39865d4f98a2SYan Zheng 39870b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 39885d4f98a2SYan Zheng 39895d4f98a2SYan Zheng set_reloc_control(rc); 39905d4f98a2SYan Zheng 39917a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 39923612b495STsutomu Itoh if (IS_ERR(trans)) { 39933612b495STsutomu Itoh err = PTR_ERR(trans); 3994fb2d83eeSJosef Bacik goto out_unset; 39953612b495STsutomu Itoh } 39963fd0a558SYan, Zheng 39973fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 39983fd0a558SYan, Zheng 39995d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 40005d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 40015d4f98a2SYan Zheng struct btrfs_root, root_list); 40025d4f98a2SYan Zheng list_del(&reloc_root->root_list); 40035d4f98a2SYan Zheng 40045d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 40055d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 40065d4f98a2SYan Zheng &rc->reloc_roots); 40075d4f98a2SYan Zheng continue; 40085d4f98a2SYan Zheng } 40095d4f98a2SYan Zheng 4010a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 4011a820feb5SDavid Sterba false); 401279787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 401379787eaaSJeff Mahoney err = PTR_ERR(fs_root); 4014ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 40151402d17dSXiyu Yang btrfs_end_transaction(trans); 4016fb2d83eeSJosef Bacik goto out_unset; 401779787eaaSJeff Mahoney } 40185d4f98a2SYan Zheng 4019ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 402079787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 4021f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 402200246528SJosef Bacik btrfs_put_root(fs_root); 40235d4f98a2SYan Zheng } 40245d4f98a2SYan Zheng 40253a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 402679787eaaSJeff Mahoney if (err) 4027fb2d83eeSJosef Bacik goto out_unset; 40285d4f98a2SYan Zheng 40295d4f98a2SYan Zheng merge_reloc_roots(rc); 40305d4f98a2SYan Zheng 40315d4f98a2SYan Zheng unset_reloc_control(rc); 40325d4f98a2SYan Zheng 40337a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 403462b99540SQu Wenruo if (IS_ERR(trans)) { 40353612b495STsutomu Itoh err = PTR_ERR(trans); 40366217b0faSJosef Bacik goto out_clean; 403762b99540SQu Wenruo } 40383a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 40396217b0faSJosef Bacik out_clean: 4040d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4041d2311e69SQu Wenruo if (ret < 0 && !err) 4042d2311e69SQu Wenruo err = ret; 4043fb2d83eeSJosef Bacik out_unset: 4044fb2d83eeSJosef Bacik unset_reloc_control(rc); 40451a0afa0eSJosef Bacik free_reloc_control(rc); 40463612b495STsutomu Itoh out: 4047aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4048aca1bba6SLiu Bo 40495d4f98a2SYan Zheng btrfs_free_path(path); 40505d4f98a2SYan Zheng 40515d4f98a2SYan Zheng if (err == 0) { 40525d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 4053aeb935a4SQu Wenruo fs_root = btrfs_grab_root(fs_info->data_reloc_root); 4054aeb935a4SQu Wenruo ASSERT(fs_root); 405566b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 405600246528SJosef Bacik btrfs_put_root(fs_root); 4057932fd26dSJosef Bacik } 40585d4f98a2SYan Zheng return err; 40595d4f98a2SYan Zheng } 40605d4f98a2SYan Zheng 40615d4f98a2SYan Zheng /* 40625d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 40635d4f98a2SYan Zheng * 40645d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 40655d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 40665d4f98a2SYan Zheng */ 40677bfa9535SNikolay Borisov int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len) 40685d4f98a2SYan Zheng { 40697bfa9535SNikolay Borisov struct btrfs_fs_info *fs_info = inode->root->fs_info; 40705d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 40715d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 40725d4f98a2SYan Zheng int ret; 40735d4f98a2SYan Zheng u64 disk_bytenr; 40744577b014SJosef Bacik u64 new_bytenr; 40755d4f98a2SYan Zheng LIST_HEAD(list); 40765d4f98a2SYan Zheng 40777bfa9535SNikolay Borisov ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4078bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 40795d4f98a2SYan Zheng 40807bfa9535SNikolay Borisov disk_bytenr = file_pos + inode->index_cnt; 40810b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4082a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 408379787eaaSJeff Mahoney if (ret) 408479787eaaSJeff Mahoney goto out; 40855d4f98a2SYan Zheng 40865d4f98a2SYan Zheng while (!list_empty(&list)) { 40875d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 40885d4f98a2SYan Zheng list_del_init(&sums->list); 40895d4f98a2SYan Zheng 40904577b014SJosef Bacik /* 40914577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 40924577b014SJosef Bacik * We need to do this because we will read in entire prealloc 40934577b014SJosef Bacik * extents but we may have written to say the middle of the 40944577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 40954577b014SJosef Bacik * the right disk offset. 40964577b014SJosef Bacik * 40974577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 40984577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 40994577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 41004577b014SJosef Bacik * disk length. 41014577b014SJosef Bacik */ 4102bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 41034577b014SJosef Bacik sums->bytenr = new_bytenr; 41045d4f98a2SYan Zheng 4105f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 41065d4f98a2SYan Zheng } 410779787eaaSJeff Mahoney out: 41085d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4109411fc6bcSAndi Kleen return ret; 41105d4f98a2SYan Zheng } 41113fd0a558SYan, Zheng 411283d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 41133fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 41143fd0a558SYan, Zheng struct extent_buffer *cow) 41153fd0a558SYan, Zheng { 41160b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 41173fd0a558SYan, Zheng struct reloc_control *rc; 4118a26195a5SQu Wenruo struct btrfs_backref_node *node; 41193fd0a558SYan, Zheng int first_cow = 0; 41203fd0a558SYan, Zheng int level; 412183d4cfd4SJosef Bacik int ret = 0; 41223fd0a558SYan, Zheng 41230b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 41243fd0a558SYan, Zheng if (!rc) 412583d4cfd4SJosef Bacik return 0; 41263fd0a558SYan, Zheng 41273fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 41283fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 41293fd0a558SYan, Zheng 41303fd0a558SYan, Zheng level = btrfs_header_level(buf); 41313fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 41323fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 41333fd0a558SYan, Zheng first_cow = 1; 41343fd0a558SYan, Zheng 41353fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 41363fd0a558SYan, Zheng rc->create_reloc_tree) { 41373fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 41383fd0a558SYan, Zheng 41393fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 41403fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 41413fd0a558SYan, Zheng node->new_bytenr != buf->start); 41423fd0a558SYan, Zheng 4143b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 414467439dadSDavid Sterba atomic_inc(&cow->refs); 41453fd0a558SYan, Zheng node->eb = cow; 41463fd0a558SYan, Zheng node->new_bytenr = cow->start; 41473fd0a558SYan, Zheng 41483fd0a558SYan, Zheng if (!node->pending) { 41493fd0a558SYan, Zheng list_move_tail(&node->list, 41503fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 41513fd0a558SYan, Zheng node->pending = 1; 41523fd0a558SYan, Zheng } 41533fd0a558SYan, Zheng 41543fd0a558SYan, Zheng if (first_cow) 41559569cc20SQu Wenruo mark_block_processed(rc, node); 41563fd0a558SYan, Zheng 41573fd0a558SYan, Zheng if (first_cow && level > 0) 41583fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 41593fd0a558SYan, Zheng } 41603fd0a558SYan, Zheng 416183d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 41623fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 416383d4cfd4SJosef Bacik return ret; 41643fd0a558SYan, Zheng } 41653fd0a558SYan, Zheng 41663fd0a558SYan, Zheng /* 41673fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 416801327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 41693fd0a558SYan, Zheng */ 4170147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 41713fd0a558SYan, Zheng u64 *bytes_to_reserve) 41723fd0a558SYan, Zheng { 417310995c04SQu Wenruo struct btrfs_root *root = pending->root; 417410995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 41753fd0a558SYan, Zheng 41766282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 41773fd0a558SYan, Zheng return; 41783fd0a558SYan, Zheng 41793fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 41803fd0a558SYan, Zheng return; 41813fd0a558SYan, Zheng 41823fd0a558SYan, Zheng root = root->reloc_root; 41833fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 41843fd0a558SYan, Zheng /* 41853fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 41863fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 41873fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 41883fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 41893fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 41903fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 41913fd0a558SYan, Zheng * tree may use more space than it frees. so we need 41923fd0a558SYan, Zheng * reserve extra space. 41933fd0a558SYan, Zheng */ 41943fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 41953fd0a558SYan, Zheng } 41963fd0a558SYan, Zheng 41973fd0a558SYan, Zheng /* 41983fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 41993fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4200f44deb74SJosef Bacik * 4201f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4202f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4203f44deb74SJosef Bacik * rc->reloc_roots. 42043fd0a558SYan, Zheng */ 420549b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 42063fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 42073fd0a558SYan, Zheng { 42083fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 42093fd0a558SYan, Zheng struct btrfs_root *reloc_root; 42103fd0a558SYan, Zheng struct btrfs_root *new_root; 421110995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 42123fd0a558SYan, Zheng int ret; 42133fd0a558SYan, Zheng 42146282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 421549b25e05SJeff Mahoney return 0; 42163fd0a558SYan, Zheng 42173fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 42183fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 42193fd0a558SYan, Zheng 42203fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 42213fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 42223fd0a558SYan, Zheng rc->block_rsv, 42233a584174SLu Fengqi rc->nodes_relocated, true); 422449b25e05SJeff Mahoney if (ret) 422549b25e05SJeff Mahoney return ret; 42263fd0a558SYan, Zheng } 42273fd0a558SYan, Zheng 42283fd0a558SYan, Zheng new_root = pending->snap; 42293fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 42303fd0a558SYan, Zheng new_root->root_key.objectid); 423149b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 423249b25e05SJeff Mahoney return PTR_ERR(reloc_root); 42333fd0a558SYan, Zheng 4234ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4235ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4236f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 42373fd0a558SYan, Zheng 423849b25e05SJeff Mahoney if (rc->create_reloc_tree) 42393fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 424049b25e05SJeff Mahoney return ret; 42413fd0a558SYan, Zheng } 4242