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)) 906592fbcd5SJosef Bacik return 0; 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); 939f44deb74SJosef Bacik btrfs_put_root(reloc_root); 940592fbcd5SJosef Bacik return ret; 9415d4f98a2SYan Zheng } 9425d4f98a2SYan Zheng 9435d4f98a2SYan Zheng /* 9445d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 9455d4f98a2SYan Zheng * in a subvolume 9465d4f98a2SYan Zheng */ 9475d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 9485d4f98a2SYan Zheng { 9495d4f98a2SYan Zheng struct rb_node *node; 9505d4f98a2SYan Zheng struct rb_node *prev; 9515d4f98a2SYan Zheng struct btrfs_inode *entry; 9525d4f98a2SYan Zheng struct inode *inode; 9535d4f98a2SYan Zheng 9545d4f98a2SYan Zheng spin_lock(&root->inode_lock); 9555d4f98a2SYan Zheng again: 9565d4f98a2SYan Zheng node = root->inode_tree.rb_node; 9575d4f98a2SYan Zheng prev = NULL; 9585d4f98a2SYan Zheng while (node) { 9595d4f98a2SYan Zheng prev = node; 9605d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9615d4f98a2SYan Zheng 9624a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 9635d4f98a2SYan Zheng node = node->rb_left; 9644a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 9655d4f98a2SYan Zheng node = node->rb_right; 9665d4f98a2SYan Zheng else 9675d4f98a2SYan Zheng break; 9685d4f98a2SYan Zheng } 9695d4f98a2SYan Zheng if (!node) { 9705d4f98a2SYan Zheng while (prev) { 9715d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 9724a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 9735d4f98a2SYan Zheng node = prev; 9745d4f98a2SYan Zheng break; 9755d4f98a2SYan Zheng } 9765d4f98a2SYan Zheng prev = rb_next(prev); 9775d4f98a2SYan Zheng } 9785d4f98a2SYan Zheng } 9795d4f98a2SYan Zheng while (node) { 9805d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9815d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 9825d4f98a2SYan Zheng if (inode) { 9835d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9845d4f98a2SYan Zheng return inode; 9855d4f98a2SYan Zheng } 9865d4f98a2SYan Zheng 9874a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 9885d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 9895d4f98a2SYan Zheng goto again; 9905d4f98a2SYan Zheng 9915d4f98a2SYan Zheng node = rb_next(node); 9925d4f98a2SYan Zheng } 9935d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9945d4f98a2SYan Zheng return NULL; 9955d4f98a2SYan Zheng } 9965d4f98a2SYan Zheng 9975d4f98a2SYan Zheng /* 9985d4f98a2SYan Zheng * get new location of data 9995d4f98a2SYan Zheng */ 10005d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 10015d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 10025d4f98a2SYan Zheng { 10035d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 10045d4f98a2SYan Zheng struct btrfs_path *path; 10055d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10065d4f98a2SYan Zheng struct extent_buffer *leaf; 10075d4f98a2SYan Zheng int ret; 10085d4f98a2SYan Zheng 10095d4f98a2SYan Zheng path = btrfs_alloc_path(); 10105d4f98a2SYan Zheng if (!path) 10115d4f98a2SYan Zheng return -ENOMEM; 10125d4f98a2SYan Zheng 10135d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 1014f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 1015f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 10165d4f98a2SYan Zheng if (ret < 0) 10175d4f98a2SYan Zheng goto out; 10185d4f98a2SYan Zheng if (ret > 0) { 10195d4f98a2SYan Zheng ret = -ENOENT; 10205d4f98a2SYan Zheng goto out; 10215d4f98a2SYan Zheng } 10225d4f98a2SYan Zheng 10235d4f98a2SYan Zheng leaf = path->nodes[0]; 10245d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 10255d4f98a2SYan Zheng struct btrfs_file_extent_item); 10265d4f98a2SYan Zheng 10275d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 10285d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 10295d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 10305d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 10315d4f98a2SYan Zheng 10325d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 103383d4cfd4SJosef Bacik ret = -EINVAL; 10345d4f98a2SYan Zheng goto out; 10355d4f98a2SYan Zheng } 10365d4f98a2SYan Zheng 10375d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10385d4f98a2SYan Zheng ret = 0; 10395d4f98a2SYan Zheng out: 10405d4f98a2SYan Zheng btrfs_free_path(path); 10415d4f98a2SYan Zheng return ret; 10425d4f98a2SYan Zheng } 10435d4f98a2SYan Zheng 10445d4f98a2SYan Zheng /* 10455d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 10465d4f98a2SYan Zheng * the new locations. 10475d4f98a2SYan Zheng */ 10483fd0a558SYan, Zheng static noinline_for_stack 10493fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 10505d4f98a2SYan Zheng struct reloc_control *rc, 10515d4f98a2SYan Zheng struct btrfs_root *root, 10523fd0a558SYan, Zheng struct extent_buffer *leaf) 10535d4f98a2SYan Zheng { 10540b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 10555d4f98a2SYan Zheng struct btrfs_key key; 10565d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10575d4f98a2SYan Zheng struct inode *inode = NULL; 10585d4f98a2SYan Zheng u64 parent; 10595d4f98a2SYan Zheng u64 bytenr; 10603fd0a558SYan, Zheng u64 new_bytenr = 0; 10615d4f98a2SYan Zheng u64 num_bytes; 10625d4f98a2SYan Zheng u64 end; 10635d4f98a2SYan Zheng u32 nritems; 10645d4f98a2SYan Zheng u32 i; 106583d4cfd4SJosef Bacik int ret = 0; 10665d4f98a2SYan Zheng int first = 1; 10675d4f98a2SYan Zheng int dirty = 0; 10685d4f98a2SYan Zheng 10695d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 10705d4f98a2SYan Zheng return 0; 10715d4f98a2SYan Zheng 10725d4f98a2SYan Zheng /* reloc trees always use full backref */ 10735d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 10745d4f98a2SYan Zheng parent = leaf->start; 10755d4f98a2SYan Zheng else 10765d4f98a2SYan Zheng parent = 0; 10775d4f98a2SYan Zheng 10785d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 10795d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 108082fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 108182fa113fSQu Wenruo 10825d4f98a2SYan Zheng cond_resched(); 10835d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 10845d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 10855d4f98a2SYan Zheng continue; 10865d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 10875d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 10885d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 10895d4f98a2SYan Zheng continue; 10905d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10915d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 10925d4f98a2SYan Zheng if (bytenr == 0) 10935d4f98a2SYan Zheng continue; 10949569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 10959569cc20SQu Wenruo rc->block_group->length)) 10965d4f98a2SYan Zheng continue; 10975d4f98a2SYan Zheng 10985d4f98a2SYan Zheng /* 10995d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 11005d4f98a2SYan Zheng * to complete and drop the extent cache 11015d4f98a2SYan Zheng */ 11025d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 11035d4f98a2SYan Zheng if (first) { 11045d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11055d4f98a2SYan Zheng first = 0; 11064a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 11073fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 11085d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11095d4f98a2SYan Zheng } 11104a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 11115d4f98a2SYan Zheng end = key.offset + 11125d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 11135d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 11140b246afaSJeff Mahoney fs_info->sectorsize)); 11150b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 11165d4f98a2SYan Zheng end--; 11175d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1118d0082371SJeff Mahoney key.offset, end); 11195d4f98a2SYan Zheng if (!ret) 11205d4f98a2SYan Zheng continue; 11215d4f98a2SYan Zheng 1122dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1123dcdbc059SNikolay Borisov key.offset, end, 1); 11245d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1125d0082371SJeff Mahoney key.offset, end); 11265d4f98a2SYan Zheng } 11275d4f98a2SYan Zheng } 11285d4f98a2SYan Zheng 11295d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 11305d4f98a2SYan Zheng bytenr, num_bytes); 113183d4cfd4SJosef Bacik if (ret) { 113283d4cfd4SJosef Bacik /* 113383d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 113483d4cfd4SJosef Bacik * in the file extent yet. 113583d4cfd4SJosef Bacik */ 113683d4cfd4SJosef Bacik break; 11373fd0a558SYan, Zheng } 11385d4f98a2SYan Zheng 11395d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 11405d4f98a2SYan Zheng dirty = 1; 11415d4f98a2SYan Zheng 11425d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 114382fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 114482fa113fSQu Wenruo num_bytes, parent); 114582fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 114682fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1147b06c4bf5SFilipe Manana key.objectid, key.offset); 114882fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 114983d4cfd4SJosef Bacik if (ret) { 115066642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 115183d4cfd4SJosef Bacik break; 115283d4cfd4SJosef Bacik } 11535d4f98a2SYan Zheng 1154ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1155ffd4bb2aSQu Wenruo num_bytes, parent); 1156ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1157ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1158b06c4bf5SFilipe Manana key.objectid, key.offset); 1159ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 116083d4cfd4SJosef Bacik if (ret) { 116166642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 116283d4cfd4SJosef Bacik break; 116383d4cfd4SJosef Bacik } 11645d4f98a2SYan Zheng } 11655d4f98a2SYan Zheng if (dirty) 11665d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 11673fd0a558SYan, Zheng if (inode) 11683fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 116983d4cfd4SJosef Bacik return ret; 11705d4f98a2SYan Zheng } 11715d4f98a2SYan Zheng 11725d4f98a2SYan Zheng static noinline_for_stack 11735d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 11745d4f98a2SYan Zheng struct btrfs_path *path, int level) 11755d4f98a2SYan Zheng { 11765d4f98a2SYan Zheng struct btrfs_disk_key key1; 11775d4f98a2SYan Zheng struct btrfs_disk_key key2; 11785d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 11795d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 11805d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 11815d4f98a2SYan Zheng } 11825d4f98a2SYan Zheng 11835d4f98a2SYan Zheng /* 11845d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 11855d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 11865d4f98a2SYan Zheng * reloc tree was create can be replaced. 11875d4f98a2SYan Zheng * 11885d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 11895d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 11905d4f98a2SYan Zheng * errors, a negative error number is returned. 11915d4f98a2SYan Zheng */ 11923fd0a558SYan, Zheng static noinline_for_stack 11933d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 11945d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 11955d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 11965d4f98a2SYan Zheng int lowest_level, int max_level) 11975d4f98a2SYan Zheng { 11980b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 11995d4f98a2SYan Zheng struct extent_buffer *eb; 12005d4f98a2SYan Zheng struct extent_buffer *parent; 120182fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 12025d4f98a2SYan Zheng struct btrfs_key key; 12035d4f98a2SYan Zheng u64 old_bytenr; 12045d4f98a2SYan Zheng u64 new_bytenr; 12055d4f98a2SYan Zheng u64 old_ptr_gen; 12065d4f98a2SYan Zheng u64 new_ptr_gen; 12075d4f98a2SYan Zheng u64 last_snapshot; 12085d4f98a2SYan Zheng u32 blocksize; 12093fd0a558SYan, Zheng int cow = 0; 12105d4f98a2SYan Zheng int level; 12115d4f98a2SYan Zheng int ret; 12125d4f98a2SYan Zheng int slot; 12135d4f98a2SYan Zheng 12147a9213a9SJosef Bacik ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 12157a9213a9SJosef Bacik ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 12165d4f98a2SYan Zheng 12175d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 12183fd0a558SYan, Zheng again: 12195d4f98a2SYan Zheng slot = path->slots[lowest_level]; 12205d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 12215d4f98a2SYan Zheng 12225d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 12235d4f98a2SYan Zheng level = btrfs_header_level(eb); 12245d4f98a2SYan Zheng 12255d4f98a2SYan Zheng if (level < lowest_level) { 12265d4f98a2SYan Zheng btrfs_tree_unlock(eb); 12275d4f98a2SYan Zheng free_extent_buffer(eb); 12285d4f98a2SYan Zheng return 0; 12295d4f98a2SYan Zheng } 12305d4f98a2SYan Zheng 12313fd0a558SYan, Zheng if (cow) { 12329631e4ccSJosef Bacik ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb, 12339631e4ccSJosef Bacik BTRFS_NESTING_COW); 123445b87c5dSJosef Bacik if (ret) { 123545b87c5dSJosef Bacik btrfs_tree_unlock(eb); 123645b87c5dSJosef Bacik free_extent_buffer(eb); 123745b87c5dSJosef Bacik return ret; 123845b87c5dSJosef Bacik } 12393fd0a558SYan, Zheng } 12405d4f98a2SYan Zheng 12415d4f98a2SYan Zheng if (next_key) { 12425d4f98a2SYan Zheng next_key->objectid = (u64)-1; 12435d4f98a2SYan Zheng next_key->type = (u8)-1; 12445d4f98a2SYan Zheng next_key->offset = (u64)-1; 12455d4f98a2SYan Zheng } 12465d4f98a2SYan Zheng 12475d4f98a2SYan Zheng parent = eb; 12485d4f98a2SYan Zheng while (1) { 12495d4f98a2SYan Zheng level = btrfs_header_level(parent); 12507a9213a9SJosef Bacik ASSERT(level >= lowest_level); 12515d4f98a2SYan Zheng 1252e3b83361SQu Wenruo ret = btrfs_bin_search(parent, &key, &slot); 1253cbca7d59SFilipe Manana if (ret < 0) 1254cbca7d59SFilipe Manana break; 12555d4f98a2SYan Zheng if (ret && slot > 0) 12565d4f98a2SYan Zheng slot--; 12575d4f98a2SYan Zheng 12585d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 12595d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 12605d4f98a2SYan Zheng 12615d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 12620b246afaSJeff Mahoney blocksize = fs_info->nodesize; 12635d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 12645d4f98a2SYan Zheng 12655d4f98a2SYan Zheng if (level <= max_level) { 12665d4f98a2SYan Zheng eb = path->nodes[level]; 12675d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 12685d4f98a2SYan Zheng path->slots[level]); 12695d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 12705d4f98a2SYan Zheng path->slots[level]); 12715d4f98a2SYan Zheng } else { 12725d4f98a2SYan Zheng new_bytenr = 0; 12735d4f98a2SYan Zheng new_ptr_gen = 0; 12745d4f98a2SYan Zheng } 12755d4f98a2SYan Zheng 1276fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 12775d4f98a2SYan Zheng ret = level; 12785d4f98a2SYan Zheng break; 12795d4f98a2SYan Zheng } 12805d4f98a2SYan Zheng 12815d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 12825d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 12833fd0a558SYan, Zheng if (level <= lowest_level) { 12845d4f98a2SYan Zheng ret = 0; 12855d4f98a2SYan Zheng break; 12865d4f98a2SYan Zheng } 12875d4f98a2SYan Zheng 12886b3426beSJosef Bacik eb = btrfs_read_node_slot(parent, slot); 128964c043deSLiu Bo if (IS_ERR(eb)) { 129064c043deSLiu Bo ret = PTR_ERR(eb); 1291264813acSLiu Bo break; 1292416bc658SJosef Bacik } 12935d4f98a2SYan Zheng btrfs_tree_lock(eb); 12943fd0a558SYan, Zheng if (cow) { 12955d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 12969631e4ccSJosef Bacik slot, &eb, 12979631e4ccSJosef Bacik BTRFS_NESTING_COW); 129845b87c5dSJosef Bacik if (ret) { 129945b87c5dSJosef Bacik btrfs_tree_unlock(eb); 130045b87c5dSJosef Bacik free_extent_buffer(eb); 130145b87c5dSJosef Bacik break; 130245b87c5dSJosef Bacik } 13035d4f98a2SYan Zheng } 13045d4f98a2SYan Zheng 13055d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13065d4f98a2SYan Zheng free_extent_buffer(parent); 13075d4f98a2SYan Zheng 13085d4f98a2SYan Zheng parent = eb; 13095d4f98a2SYan Zheng continue; 13105d4f98a2SYan Zheng } 13115d4f98a2SYan Zheng 13123fd0a558SYan, Zheng if (!cow) { 13133fd0a558SYan, Zheng btrfs_tree_unlock(parent); 13143fd0a558SYan, Zheng free_extent_buffer(parent); 13153fd0a558SYan, Zheng cow = 1; 13163fd0a558SYan, Zheng goto again; 13173fd0a558SYan, Zheng } 13183fd0a558SYan, Zheng 13195d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 13205d4f98a2SYan Zheng path->slots[level]); 1321b3b4aa74SDavid Sterba btrfs_release_path(path); 13225d4f98a2SYan Zheng 13235d4f98a2SYan Zheng path->lowest_level = level; 13245d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 13255d4f98a2SYan Zheng path->lowest_level = 0; 1326*0e9873e2SJosef Bacik if (ret) { 1327*0e9873e2SJosef Bacik if (ret > 0) 1328*0e9873e2SJosef Bacik ret = -ENOENT; 1329*0e9873e2SJosef Bacik break; 1330*0e9873e2SJosef Bacik } 13315d4f98a2SYan Zheng 13325d4f98a2SYan Zheng /* 1333824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 1334824d8dffSQu Wenruo * 1335824d8dffSQu Wenruo * We must trace both trees. 1336824d8dffSQu Wenruo * 1) Tree reloc subtree 1337824d8dffSQu Wenruo * If not traced, we will leak data numbers 1338824d8dffSQu Wenruo * 2) Fs subtree 1339824d8dffSQu Wenruo * If not traced, we will double count old data 1340f616f5cdSQu Wenruo * 1341f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 1342f616f5cdSQu Wenruo * the swapped tree blocks. 1343f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 1344f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 1345824d8dffSQu Wenruo */ 1346370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 1347370a11b8SQu Wenruo rc->block_group, parent, slot, 1348370a11b8SQu Wenruo path->nodes[level], path->slots[level], 1349370a11b8SQu Wenruo last_snapshot); 1350370a11b8SQu Wenruo if (ret < 0) 1351370a11b8SQu Wenruo break; 1352824d8dffSQu Wenruo /* 13535d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 13545d4f98a2SYan Zheng */ 13555d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 13565d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 13575d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 13585d4f98a2SYan Zheng 13595d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 13605d4f98a2SYan Zheng path->slots[level], old_bytenr); 13615d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 13625d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 13635d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 13645d4f98a2SYan Zheng 136582fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 136682fa113fSQu Wenruo blocksize, path->nodes[level]->start); 136782fa113fSQu Wenruo ref.skip_qgroup = true; 136882fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 136982fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13705d4f98a2SYan Zheng BUG_ON(ret); 137182fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 137282fa113fSQu Wenruo blocksize, 0); 137382fa113fSQu Wenruo ref.skip_qgroup = true; 137482fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 137582fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13765d4f98a2SYan Zheng BUG_ON(ret); 13775d4f98a2SYan Zheng 1378ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 1379ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 1380ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 1381ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1382ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13835d4f98a2SYan Zheng BUG_ON(ret); 13845d4f98a2SYan Zheng 1385ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 1386ffd4bb2aSQu Wenruo blocksize, 0); 1387ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 1388ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1389ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13905d4f98a2SYan Zheng BUG_ON(ret); 13915d4f98a2SYan Zheng 13925d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 13935d4f98a2SYan Zheng 13945d4f98a2SYan Zheng ret = level; 13955d4f98a2SYan Zheng break; 13965d4f98a2SYan Zheng } 13975d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13985d4f98a2SYan Zheng free_extent_buffer(parent); 13995d4f98a2SYan Zheng return ret; 14005d4f98a2SYan Zheng } 14015d4f98a2SYan Zheng 14025d4f98a2SYan Zheng /* 14035d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 14045d4f98a2SYan Zheng */ 14055d4f98a2SYan Zheng static noinline_for_stack 14065d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14075d4f98a2SYan Zheng int *level) 14085d4f98a2SYan Zheng { 14095d4f98a2SYan Zheng struct extent_buffer *eb; 14105d4f98a2SYan Zheng int i; 14115d4f98a2SYan Zheng u64 last_snapshot; 14125d4f98a2SYan Zheng u32 nritems; 14135d4f98a2SYan Zheng 14145d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14155d4f98a2SYan Zheng 14165d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 14175d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14185d4f98a2SYan Zheng path->nodes[i] = NULL; 14195d4f98a2SYan Zheng } 14205d4f98a2SYan Zheng 14215d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 14225d4f98a2SYan Zheng eb = path->nodes[i]; 14235d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14245d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 14255d4f98a2SYan Zheng path->slots[i]++; 14265d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 14275d4f98a2SYan Zheng last_snapshot) 14285d4f98a2SYan Zheng continue; 14295d4f98a2SYan Zheng 14305d4f98a2SYan Zheng *level = i; 14315d4f98a2SYan Zheng return 0; 14325d4f98a2SYan Zheng } 14335d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14345d4f98a2SYan Zheng path->nodes[i] = NULL; 14355d4f98a2SYan Zheng } 14365d4f98a2SYan Zheng return 1; 14375d4f98a2SYan Zheng } 14385d4f98a2SYan Zheng 14395d4f98a2SYan Zheng /* 14405d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 14415d4f98a2SYan Zheng */ 14425d4f98a2SYan Zheng static noinline_for_stack 14435d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14445d4f98a2SYan Zheng int *level) 14455d4f98a2SYan Zheng { 14465d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 14475d4f98a2SYan Zheng int i; 14485d4f98a2SYan Zheng u64 ptr_gen = 0; 14495d4f98a2SYan Zheng u64 last_snapshot; 14505d4f98a2SYan Zheng u32 nritems; 14515d4f98a2SYan Zheng 14525d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14535d4f98a2SYan Zheng 14545d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 14555d4f98a2SYan Zheng eb = path->nodes[i]; 14565d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14575d4f98a2SYan Zheng while (path->slots[i] < nritems) { 14585d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 14595d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 14605d4f98a2SYan Zheng break; 14615d4f98a2SYan Zheng path->slots[i]++; 14625d4f98a2SYan Zheng } 14635d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 14645d4f98a2SYan Zheng if (i == *level) 14655d4f98a2SYan Zheng break; 14665d4f98a2SYan Zheng *level = i + 1; 14675d4f98a2SYan Zheng return 0; 14685d4f98a2SYan Zheng } 14695d4f98a2SYan Zheng if (i == 1) { 14705d4f98a2SYan Zheng *level = i; 14715d4f98a2SYan Zheng return 0; 14725d4f98a2SYan Zheng } 14735d4f98a2SYan Zheng 14748ef385bbSJosef Bacik eb = btrfs_read_node_slot(eb, path->slots[i]); 14758ef385bbSJosef Bacik if (IS_ERR(eb)) 147664c043deSLiu Bo return PTR_ERR(eb); 14775d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 14785d4f98a2SYan Zheng path->nodes[i - 1] = eb; 14795d4f98a2SYan Zheng path->slots[i - 1] = 0; 14805d4f98a2SYan Zheng } 14815d4f98a2SYan Zheng return 1; 14825d4f98a2SYan Zheng } 14835d4f98a2SYan Zheng 14845d4f98a2SYan Zheng /* 14855d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 14865d4f98a2SYan Zheng * [min_key, max_key) 14875d4f98a2SYan Zheng */ 14885d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 14895d4f98a2SYan Zheng struct btrfs_key *min_key, 14905d4f98a2SYan Zheng struct btrfs_key *max_key) 14915d4f98a2SYan Zheng { 14920b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14935d4f98a2SYan Zheng struct inode *inode = NULL; 14945d4f98a2SYan Zheng u64 objectid; 14955d4f98a2SYan Zheng u64 start, end; 149633345d01SLi Zefan u64 ino; 14975d4f98a2SYan Zheng 14985d4f98a2SYan Zheng objectid = min_key->objectid; 14995d4f98a2SYan Zheng while (1) { 15005d4f98a2SYan Zheng cond_resched(); 15015d4f98a2SYan Zheng iput(inode); 15025d4f98a2SYan Zheng 15035d4f98a2SYan Zheng if (objectid > max_key->objectid) 15045d4f98a2SYan Zheng break; 15055d4f98a2SYan Zheng 15065d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 15075d4f98a2SYan Zheng if (!inode) 15085d4f98a2SYan Zheng break; 15094a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 15105d4f98a2SYan Zheng 151133345d01SLi Zefan if (ino > max_key->objectid) { 15125d4f98a2SYan Zheng iput(inode); 15135d4f98a2SYan Zheng break; 15145d4f98a2SYan Zheng } 15155d4f98a2SYan Zheng 151633345d01SLi Zefan objectid = ino + 1; 15175d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 15185d4f98a2SYan Zheng continue; 15195d4f98a2SYan Zheng 152033345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 15215d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 15225d4f98a2SYan Zheng continue; 15235d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 15245d4f98a2SYan Zheng start = 0; 15255d4f98a2SYan Zheng else { 15265d4f98a2SYan Zheng start = min_key->offset; 15270b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 15285d4f98a2SYan Zheng } 15295d4f98a2SYan Zheng } else { 15305d4f98a2SYan Zheng start = 0; 15315d4f98a2SYan Zheng } 15325d4f98a2SYan Zheng 153333345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 15345d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 15355d4f98a2SYan Zheng continue; 15365d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 15375d4f98a2SYan Zheng end = (u64)-1; 15385d4f98a2SYan Zheng } else { 15395d4f98a2SYan Zheng if (max_key->offset == 0) 15405d4f98a2SYan Zheng continue; 15415d4f98a2SYan Zheng end = max_key->offset; 15420b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 15435d4f98a2SYan Zheng end--; 15445d4f98a2SYan Zheng } 15455d4f98a2SYan Zheng } else { 15465d4f98a2SYan Zheng end = (u64)-1; 15475d4f98a2SYan Zheng } 15485d4f98a2SYan Zheng 15495d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 1550d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 1551dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 1552d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 15535d4f98a2SYan Zheng } 15545d4f98a2SYan Zheng return 0; 15555d4f98a2SYan Zheng } 15565d4f98a2SYan Zheng 15575d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 15585d4f98a2SYan Zheng struct btrfs_key *key) 15595d4f98a2SYan Zheng 15605d4f98a2SYan Zheng { 15615d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 15625d4f98a2SYan Zheng if (!path->nodes[level]) 15635d4f98a2SYan Zheng break; 15645d4f98a2SYan Zheng if (path->slots[level] + 1 < 15655d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 15665d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 15675d4f98a2SYan Zheng path->slots[level] + 1); 15685d4f98a2SYan Zheng return 0; 15695d4f98a2SYan Zheng } 15705d4f98a2SYan Zheng level++; 15715d4f98a2SYan Zheng } 15725d4f98a2SYan Zheng return 1; 15735d4f98a2SYan Zheng } 15745d4f98a2SYan Zheng 15755d4f98a2SYan Zheng /* 1576d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 1577d2311e69SQu Wenruo */ 1578ac54da6cSJosef Bacik static int insert_dirty_subvol(struct btrfs_trans_handle *trans, 1579d2311e69SQu Wenruo struct reloc_control *rc, 1580d2311e69SQu Wenruo struct btrfs_root *root) 1581d2311e69SQu Wenruo { 1582d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1583d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 15847934133fSJosef Bacik int ret; 1585d2311e69SQu Wenruo 1586d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 1587d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 1588d2311e69SQu Wenruo ASSERT(reloc_root); 1589d2311e69SQu Wenruo 1590d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 1591d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 1592d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 1593c8422684SDavid Sterba btrfs_set_root_drop_level(reloc_root_item, 0); 1594d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 15957934133fSJosef Bacik ret = btrfs_update_reloc_root(trans, root); 15967934133fSJosef Bacik if (ret) 15977934133fSJosef Bacik return ret; 1598d2311e69SQu Wenruo 1599d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 160000246528SJosef Bacik btrfs_grab_root(root); 1601d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 1602d2311e69SQu Wenruo } 1603ac54da6cSJosef Bacik 1604ac54da6cSJosef Bacik return 0; 1605d2311e69SQu Wenruo } 1606d2311e69SQu Wenruo 1607d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 1608d2311e69SQu Wenruo { 1609d2311e69SQu Wenruo struct btrfs_root *root; 1610d2311e69SQu Wenruo struct btrfs_root *next; 1611d2311e69SQu Wenruo int ret = 0; 161230d40577SQu Wenruo int ret2; 1613d2311e69SQu Wenruo 1614d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 1615d2311e69SQu Wenruo reloc_dirty_list) { 161630d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 161730d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 1618d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1619d2311e69SQu Wenruo 1620d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 1621d2311e69SQu Wenruo root->reloc_root = NULL; 16226282675eSQu Wenruo /* 16236282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 16246282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 16256282675eSQu Wenruo */ 16266282675eSQu Wenruo smp_wmb(); 16271fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 1628f28de8d8SJosef Bacik if (reloc_root) { 1629f44deb74SJosef Bacik /* 1630f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 1631f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 1632f44deb74SJosef Bacik * drop the ref ourselves. 1633f44deb74SJosef Bacik */ 1634f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 1635f44deb74SJosef Bacik if (ret2 < 0) { 1636f44deb74SJosef Bacik btrfs_put_root(reloc_root); 1637f44deb74SJosef Bacik if (!ret) 1638f28de8d8SJosef Bacik ret = ret2; 1639f28de8d8SJosef Bacik } 1640f44deb74SJosef Bacik } 164100246528SJosef Bacik btrfs_put_root(root); 164230d40577SQu Wenruo } else { 164330d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 16440078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 1645f44deb74SJosef Bacik if (ret2 < 0) { 1646f44deb74SJosef Bacik btrfs_put_root(root); 1647f44deb74SJosef Bacik if (!ret) 164830d40577SQu Wenruo ret = ret2; 164930d40577SQu Wenruo } 1650d2311e69SQu Wenruo } 1651f44deb74SJosef Bacik } 1652d2311e69SQu Wenruo return ret; 1653d2311e69SQu Wenruo } 1654d2311e69SQu Wenruo 1655d2311e69SQu Wenruo /* 16565d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 16575d4f98a2SYan Zheng * fs tree. 16585d4f98a2SYan Zheng */ 16595d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 16605d4f98a2SYan Zheng struct btrfs_root *root) 16615d4f98a2SYan Zheng { 16620b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 16635d4f98a2SYan Zheng struct btrfs_key key; 16645d4f98a2SYan Zheng struct btrfs_key next_key; 16659e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 16665d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16675d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16685d4f98a2SYan Zheng struct btrfs_path *path; 16693fd0a558SYan, Zheng struct extent_buffer *leaf; 1670fca3a45dSJosef Bacik int reserve_level; 16715d4f98a2SYan Zheng int level; 16725d4f98a2SYan Zheng int max_level; 16735d4f98a2SYan Zheng int replaced = 0; 1674c6a592f2SNikolay Borisov int ret = 0; 16753fd0a558SYan, Zheng u32 min_reserved; 16765d4f98a2SYan Zheng 16775d4f98a2SYan Zheng path = btrfs_alloc_path(); 16785d4f98a2SYan Zheng if (!path) 16795d4f98a2SYan Zheng return -ENOMEM; 1680e4058b54SDavid Sterba path->reada = READA_FORWARD; 16815d4f98a2SYan Zheng 16825d4f98a2SYan Zheng reloc_root = root->reloc_root; 16835d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16845d4f98a2SYan Zheng 16855d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 16865d4f98a2SYan Zheng level = btrfs_root_level(root_item); 168767439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 16885d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 16895d4f98a2SYan Zheng path->slots[level] = 0; 16905d4f98a2SYan Zheng } else { 16915d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 16925d4f98a2SYan Zheng 1693c8422684SDavid Sterba level = btrfs_root_drop_level(root_item); 16945d4f98a2SYan Zheng BUG_ON(level == 0); 16955d4f98a2SYan Zheng path->lowest_level = level; 16965d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 169733c66f43SYan Zheng path->lowest_level = 0; 16985d4f98a2SYan Zheng if (ret < 0) { 16995d4f98a2SYan Zheng btrfs_free_path(path); 17005d4f98a2SYan Zheng return ret; 17015d4f98a2SYan Zheng } 17025d4f98a2SYan Zheng 17035d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 17045d4f98a2SYan Zheng path->slots[level]); 17055d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 17065d4f98a2SYan Zheng 17075d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 17085d4f98a2SYan Zheng } 17095d4f98a2SYan Zheng 171044d354abSQu Wenruo /* 171144d354abSQu Wenruo * In merge_reloc_root(), we modify the upper level pointer to swap the 171244d354abSQu Wenruo * tree blocks between reloc tree and subvolume tree. Thus for tree 171344d354abSQu Wenruo * block COW, we COW at most from level 1 to root level for each tree. 171444d354abSQu Wenruo * 171544d354abSQu Wenruo * Thus the needed metadata size is at most root_level * nodesize, 171644d354abSQu Wenruo * and * 2 since we have two trees to COW. 171744d354abSQu Wenruo */ 1718fca3a45dSJosef Bacik reserve_level = max_t(int, 1, btrfs_root_level(root_item)); 1719fca3a45dSJosef Bacik min_reserved = fs_info->nodesize * reserve_level * 2; 17205d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 17215d4f98a2SYan Zheng 17225d4f98a2SYan Zheng while (1) { 172308e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 172444d354abSQu Wenruo BTRFS_RESERVE_FLUSH_LIMIT); 1725c6a592f2SNikolay Borisov if (ret) 17269e6a0c52SJosef Bacik goto out; 17279e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 17289e6a0c52SJosef Bacik if (IS_ERR(trans)) { 1729c6a592f2SNikolay Borisov ret = PTR_ERR(trans); 17309e6a0c52SJosef Bacik trans = NULL; 17319e6a0c52SJosef Bacik goto out; 17329e6a0c52SJosef Bacik } 17332abc726aSJosef Bacik 17342abc726aSJosef Bacik /* 17352abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 17362abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 17372abc726aSJosef Bacik * 17382abc726aSJosef Bacik * But that's ok, we started the trans handle on our 17392abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 17402abc726aSJosef Bacik * dirty list. At commit time we'll still call 17412abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 17422abc726aSJosef Bacik * appropriately. 17432abc726aSJosef Bacik */ 17442abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 17459e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 17463fd0a558SYan, Zheng 17473fd0a558SYan, Zheng replaced = 0; 17485d4f98a2SYan Zheng max_level = level; 17495d4f98a2SYan Zheng 17505d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 1751c6a592f2SNikolay Borisov if (ret < 0) 17525d4f98a2SYan Zheng goto out; 17535d4f98a2SYan Zheng if (ret > 0) 17545d4f98a2SYan Zheng break; 17555d4f98a2SYan Zheng 17565d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 17575d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 17585d4f98a2SYan Zheng ret = 0; 17595d4f98a2SYan Zheng } else { 17603d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 17613fd0a558SYan, Zheng &next_key, level, max_level); 17625d4f98a2SYan Zheng } 1763c6a592f2SNikolay Borisov if (ret < 0) 17645d4f98a2SYan Zheng goto out; 17655d4f98a2SYan Zheng if (ret > 0) { 17665d4f98a2SYan Zheng level = ret; 17675d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17685d4f98a2SYan Zheng path->slots[level]); 17695d4f98a2SYan Zheng replaced = 1; 17705d4f98a2SYan Zheng } 17715d4f98a2SYan Zheng 17725d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 17735d4f98a2SYan Zheng if (ret > 0) 17745d4f98a2SYan Zheng break; 17755d4f98a2SYan Zheng 17765d4f98a2SYan Zheng BUG_ON(level == 0); 17775d4f98a2SYan Zheng /* 17785d4f98a2SYan Zheng * save the merging progress in the drop_progress. 17795d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 17805d4f98a2SYan Zheng */ 17815d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 17825d4f98a2SYan Zheng path->slots[level]); 1783c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, level); 17845d4f98a2SYan Zheng 17853a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 17869e6a0c52SJosef Bacik trans = NULL; 17875d4f98a2SYan Zheng 17882ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 17895d4f98a2SYan Zheng 17905d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17915d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17925d4f98a2SYan Zheng } 17935d4f98a2SYan Zheng 17945d4f98a2SYan Zheng /* 17955d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 17965d4f98a2SYan Zheng * relocated and the block is tree root. 17975d4f98a2SYan Zheng */ 17985d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 17999631e4ccSJosef Bacik ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf, 18009631e4ccSJosef Bacik BTRFS_NESTING_COW); 18015d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 18025d4f98a2SYan Zheng free_extent_buffer(leaf); 18035d4f98a2SYan Zheng out: 18045d4f98a2SYan Zheng btrfs_free_path(path); 18055d4f98a2SYan Zheng 1806ac54da6cSJosef Bacik if (ret == 0) { 1807ac54da6cSJosef Bacik ret = insert_dirty_subvol(trans, rc, root); 1808ac54da6cSJosef Bacik if (ret) 1809ac54da6cSJosef Bacik btrfs_abort_transaction(trans, ret); 1810ac54da6cSJosef Bacik } 18115d4f98a2SYan Zheng 18129e6a0c52SJosef Bacik if (trans) 18133a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 18145d4f98a2SYan Zheng 18152ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 18165d4f98a2SYan Zheng 18175d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 18185d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 18195d4f98a2SYan Zheng 1820c6a592f2SNikolay Borisov return ret; 18215d4f98a2SYan Zheng } 18225d4f98a2SYan Zheng 18233fd0a558SYan, Zheng static noinline_for_stack 18243fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 18255d4f98a2SYan Zheng { 18263fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 18270b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 18283fd0a558SYan, Zheng struct btrfs_root *reloc_root; 18295d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 18303fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 18313fd0a558SYan, Zheng u64 num_bytes = 0; 18323fd0a558SYan, Zheng int ret; 18333fd0a558SYan, Zheng 18340b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 18350b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 18363fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 18370b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 18387585717fSChris Mason 18393fd0a558SYan, Zheng again: 18403fd0a558SYan, Zheng if (!err) { 18413fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 184208e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 184308e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 18443fd0a558SYan, Zheng if (ret) 18453fd0a558SYan, Zheng err = ret; 18463fd0a558SYan, Zheng } 18473fd0a558SYan, Zheng 18487a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 18493612b495STsutomu Itoh if (IS_ERR(trans)) { 18503612b495STsutomu Itoh if (!err) 18512ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 185263f018beSNikolay Borisov num_bytes, NULL); 18533612b495STsutomu Itoh return PTR_ERR(trans); 18543612b495STsutomu Itoh } 18553fd0a558SYan, Zheng 18563fd0a558SYan, Zheng if (!err) { 18573fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 18583a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18592ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 186063f018beSNikolay Borisov num_bytes, NULL); 18613fd0a558SYan, Zheng goto again; 18623fd0a558SYan, Zheng } 18633fd0a558SYan, Zheng } 18643fd0a558SYan, Zheng 18653fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 18663fd0a558SYan, Zheng 18673fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 18683fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 18693fd0a558SYan, Zheng struct btrfs_root, root_list); 18703fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 18713fd0a558SYan, Zheng 1872a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1873a820feb5SDavid Sterba false); 18743fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 18753fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 18763fd0a558SYan, Zheng 18773fd0a558SYan, Zheng /* 18783fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 18793fd0a558SYan, Zheng * knows it should resumes merging 18803fd0a558SYan, Zheng */ 18813fd0a558SYan, Zheng if (!err) 18823fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 1883bbae13f8SJosef Bacik ret = btrfs_update_reloc_root(trans, root); 18843fd0a558SYan, Zheng 1885bbae13f8SJosef Bacik /* 1886bbae13f8SJosef Bacik * Even if we have an error we need this reloc root back on our 1887bbae13f8SJosef Bacik * list so we can clean up properly. 1888bbae13f8SJosef Bacik */ 18893fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 189000246528SJosef Bacik btrfs_put_root(root); 1891bbae13f8SJosef Bacik 1892bbae13f8SJosef Bacik if (ret) { 1893bbae13f8SJosef Bacik btrfs_abort_transaction(trans, ret); 1894bbae13f8SJosef Bacik if (!err) 1895bbae13f8SJosef Bacik err = ret; 1896bbae13f8SJosef Bacik break; 1897bbae13f8SJosef Bacik } 18983fd0a558SYan, Zheng } 18993fd0a558SYan, Zheng 19003fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 19013fd0a558SYan, Zheng 19023fd0a558SYan, Zheng if (!err) 19033a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 19043fd0a558SYan, Zheng else 19053a45bb20SJeff Mahoney btrfs_end_transaction(trans); 19063fd0a558SYan, Zheng return err; 19073fd0a558SYan, Zheng } 19083fd0a558SYan, Zheng 19093fd0a558SYan, Zheng static noinline_for_stack 1910aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 1911aca1bba6SLiu Bo { 1912a7571232SNikolay Borisov struct btrfs_root *reloc_root, *tmp; 1913aca1bba6SLiu Bo 1914a7571232SNikolay Borisov list_for_each_entry_safe(reloc_root, tmp, list, root_list) 1915bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 1916aca1bba6SLiu Bo } 1917aca1bba6SLiu Bo 1918aca1bba6SLiu Bo static noinline_for_stack 191994404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 19203fd0a558SYan, Zheng { 19210b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 19225d4f98a2SYan Zheng struct btrfs_root *root; 19235d4f98a2SYan Zheng struct btrfs_root *reloc_root; 19243fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 19253fd0a558SYan, Zheng int found = 0; 1926aca1bba6SLiu Bo int ret = 0; 19273fd0a558SYan, Zheng again: 19283fd0a558SYan, Zheng root = rc->extent_root; 19297585717fSChris Mason 19307585717fSChris Mason /* 19317585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 19327585717fSChris Mason * we have to make sure nobody is in the middle of 19337585717fSChris Mason * adding their roots to the list while we are 19347585717fSChris Mason * doing this splice 19357585717fSChris Mason */ 19360b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 19373fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 19380b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 19395d4f98a2SYan Zheng 19403fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 19413fd0a558SYan, Zheng found = 1; 19423fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 19433fd0a558SYan, Zheng struct btrfs_root, root_list); 19445d4f98a2SYan Zheng 1945a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1946a820feb5SDavid Sterba false); 19475d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 19485d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 19495d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 19503fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 195100246528SJosef Bacik btrfs_put_root(root); 1952b37b39cdSJosef Bacik if (ret) { 195325e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 195425e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 195525e293c2SWang Shilong &reloc_roots); 1956aca1bba6SLiu Bo goto out; 1957b37b39cdSJosef Bacik } 19583fd0a558SYan, Zheng } else { 195951415b6cSQu Wenruo if (!IS_ERR(root)) { 196051415b6cSQu Wenruo if (root->reloc_root == reloc_root) { 196151415b6cSQu Wenruo root->reloc_root = NULL; 196251415b6cSQu Wenruo btrfs_put_root(reloc_root); 196351415b6cSQu Wenruo } 19641dae7e0eSQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, 19651dae7e0eSQu Wenruo &root->state); 196651415b6cSQu Wenruo btrfs_put_root(root); 196751415b6cSQu Wenruo } 196851415b6cSQu Wenruo 19693fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 197030d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 197130d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 197230d40577SQu Wenruo &rc->dirty_subvol_roots); 19733fd0a558SYan, Zheng } 19745d4f98a2SYan Zheng } 19755d4f98a2SYan Zheng 19763fd0a558SYan, Zheng if (found) { 19773fd0a558SYan, Zheng found = 0; 19783fd0a558SYan, Zheng goto again; 19795d4f98a2SYan Zheng } 1980aca1bba6SLiu Bo out: 1981aca1bba6SLiu Bo if (ret) { 19820b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 1983aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 1984467bb1d2SWang Shilong 1985467bb1d2SWang Shilong /* new reloc root may be added */ 19860b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 1987467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 19880b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 1989467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 1990aca1bba6SLiu Bo } 1991aca1bba6SLiu Bo 19927b7b7431SJosef Bacik /* 19937b7b7431SJosef Bacik * We used to have 19947b7b7431SJosef Bacik * 19957b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 19967b7b7431SJosef Bacik * 19977b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 19987b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 19997b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 20007b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 20017b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 20027b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 20037b7b7431SJosef Bacik * will be cleaned up on unmount. 20047b7b7431SJosef Bacik * 20057b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 20067b7b7431SJosef Bacik */ 20075d4f98a2SYan Zheng } 20085d4f98a2SYan Zheng 20095d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 20105d4f98a2SYan Zheng { 20115d4f98a2SYan Zheng struct tree_block *block; 20125d4f98a2SYan Zheng struct rb_node *rb_node; 20135d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 20145d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 20155d4f98a2SYan Zheng rb_erase(rb_node, blocks); 20165d4f98a2SYan Zheng kfree(block); 20175d4f98a2SYan Zheng } 20185d4f98a2SYan Zheng } 20195d4f98a2SYan Zheng 20205d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 20215d4f98a2SYan Zheng struct btrfs_root *reloc_root) 20225d4f98a2SYan Zheng { 20230b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 20245d4f98a2SYan Zheng struct btrfs_root *root; 2025442b1ac5SJosef Bacik int ret; 20265d4f98a2SYan Zheng 20275d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 20285d4f98a2SYan Zheng return 0; 20295d4f98a2SYan Zheng 2030a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false); 2031404bccbcSJosef Bacik 2032404bccbcSJosef Bacik /* 2033404bccbcSJosef Bacik * This should succeed, since we can't have a reloc root without having 2034404bccbcSJosef Bacik * already looked up the actual root and created the reloc root for this 2035404bccbcSJosef Bacik * root. 2036404bccbcSJosef Bacik * 2037404bccbcSJosef Bacik * However if there's some sort of corruption where we have a ref to a 2038404bccbcSJosef Bacik * reloc root without a corresponding root this could return ENOENT. 2039404bccbcSJosef Bacik */ 2040404bccbcSJosef Bacik if (IS_ERR(root)) { 2041404bccbcSJosef Bacik ASSERT(0); 2042404bccbcSJosef Bacik return PTR_ERR(root); 2043404bccbcSJosef Bacik } 2044404bccbcSJosef Bacik if (root->reloc_root != reloc_root) { 2045404bccbcSJosef Bacik ASSERT(0); 2046404bccbcSJosef Bacik btrfs_err(fs_info, 2047404bccbcSJosef Bacik "root %llu has two reloc roots associated with it", 2048404bccbcSJosef Bacik reloc_root->root_key.offset); 2049404bccbcSJosef Bacik btrfs_put_root(root); 2050404bccbcSJosef Bacik return -EUCLEAN; 2051404bccbcSJosef Bacik } 2052442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 205300246528SJosef Bacik btrfs_put_root(root); 20545d4f98a2SYan Zheng 2055442b1ac5SJosef Bacik return ret; 20565d4f98a2SYan Zheng } 20575d4f98a2SYan Zheng 20583fd0a558SYan, Zheng static noinline_for_stack 20593fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 20603fd0a558SYan, Zheng struct reloc_control *rc, 2061a26195a5SQu Wenruo struct btrfs_backref_node *node, 2062a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 20635d4f98a2SYan Zheng { 2064a26195a5SQu Wenruo struct btrfs_backref_node *next; 20655d4f98a2SYan Zheng struct btrfs_root *root; 20663fd0a558SYan, Zheng int index = 0; 206792de551bSJosef Bacik int ret; 20683fd0a558SYan, Zheng 20695d4f98a2SYan Zheng next = node; 20705d4f98a2SYan Zheng while (1) { 20715d4f98a2SYan Zheng cond_resched(); 20725d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 20735d4f98a2SYan Zheng root = next->root; 20748ee66afeSJosef Bacik 20758ee66afeSJosef Bacik /* 20768ee66afeSJosef Bacik * If there is no root, then our references for this block are 20778ee66afeSJosef Bacik * incomplete, as we should be able to walk all the way up to a 20788ee66afeSJosef Bacik * block that is owned by a root. 20798ee66afeSJosef Bacik * 20808ee66afeSJosef Bacik * This path is only for SHAREABLE roots, so if we come upon a 20818ee66afeSJosef Bacik * non-SHAREABLE root then we have backrefs that resolve 20828ee66afeSJosef Bacik * improperly. 20838ee66afeSJosef Bacik * 20848ee66afeSJosef Bacik * Both of these cases indicate file system corruption, or a bug 20858ee66afeSJosef Bacik * in the backref walking code. 20868ee66afeSJosef Bacik */ 20878ee66afeSJosef Bacik if (!root) { 20888ee66afeSJosef Bacik ASSERT(0); 20898ee66afeSJosef Bacik btrfs_err(trans->fs_info, 20908ee66afeSJosef Bacik "bytenr %llu doesn't have a backref path ending in a root", 20918ee66afeSJosef Bacik node->bytenr); 20928ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 20938ee66afeSJosef Bacik } 20948ee66afeSJosef Bacik if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 20958ee66afeSJosef Bacik ASSERT(0); 20968ee66afeSJosef Bacik btrfs_err(trans->fs_info, 20978ee66afeSJosef Bacik "bytenr %llu has multiple refs with one ending in a non-shareable root", 20988ee66afeSJosef Bacik node->bytenr); 20998ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 21008ee66afeSJosef Bacik } 21015d4f98a2SYan Zheng 21025d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 210392de551bSJosef Bacik ret = record_reloc_root_in_trans(trans, root); 210492de551bSJosef Bacik if (ret) 210592de551bSJosef Bacik return ERR_PTR(ret); 21065d4f98a2SYan Zheng break; 21075d4f98a2SYan Zheng } 21085d4f98a2SYan Zheng 210992de551bSJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 211092de551bSJosef Bacik if (ret) 211192de551bSJosef Bacik return ERR_PTR(ret); 21123fd0a558SYan, Zheng root = root->reloc_root; 21133fd0a558SYan, Zheng 211439200e59SJosef Bacik /* 211539200e59SJosef Bacik * We could have raced with another thread which failed, so 211639200e59SJosef Bacik * root->reloc_root may not be set, return ENOENT in this case. 211739200e59SJosef Bacik */ 211839200e59SJosef Bacik if (!root) 211939200e59SJosef Bacik return ERR_PTR(-ENOENT); 212039200e59SJosef Bacik 21213fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 21228ee66afeSJosef Bacik /* 21238ee66afeSJosef Bacik * We just created the reloc root, so we shouldn't have 21248ee66afeSJosef Bacik * ->new_bytenr set and this shouldn't be in the changed 21258ee66afeSJosef Bacik * list. If it is then we have multiple roots pointing 21268ee66afeSJosef Bacik * at the same bytenr which indicates corruption, or 21278ee66afeSJosef Bacik * we've made a mistake in the backref walking code. 21288ee66afeSJosef Bacik */ 21298ee66afeSJosef Bacik ASSERT(next->new_bytenr == 0); 21308ee66afeSJosef Bacik ASSERT(list_empty(&next->list)); 21318ee66afeSJosef Bacik if (next->new_bytenr || !list_empty(&next->list)) { 21328ee66afeSJosef Bacik btrfs_err(trans->fs_info, 21338ee66afeSJosef Bacik "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu", 21348ee66afeSJosef Bacik node->bytenr, next->bytenr); 21358ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 21368ee66afeSJosef Bacik } 21378ee66afeSJosef Bacik 21383fd0a558SYan, Zheng next->new_bytenr = root->node->start; 213900246528SJosef Bacik btrfs_put_root(next->root); 214000246528SJosef Bacik next->root = btrfs_grab_root(root); 21410b530bc5SJosef Bacik ASSERT(next->root); 21423fd0a558SYan, Zheng list_add_tail(&next->list, 21433fd0a558SYan, Zheng &rc->backref_cache.changed); 21449569cc20SQu Wenruo mark_block_processed(rc, next); 21455d4f98a2SYan Zheng break; 21465d4f98a2SYan Zheng } 21475d4f98a2SYan Zheng 21483fd0a558SYan, Zheng WARN_ON(1); 21495d4f98a2SYan Zheng root = NULL; 21505d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 21515d4f98a2SYan Zheng if (!next || next->level <= node->level) 21525d4f98a2SYan Zheng break; 21535d4f98a2SYan Zheng } 2154cbdc2ebcSJosef Bacik if (!root) { 2155cbdc2ebcSJosef Bacik /* 2156cbdc2ebcSJosef Bacik * This can happen if there's fs corruption or if there's a bug 2157cbdc2ebcSJosef Bacik * in the backref lookup code. 2158cbdc2ebcSJosef Bacik */ 2159cbdc2ebcSJosef Bacik ASSERT(0); 2160cbdc2ebcSJosef Bacik return ERR_PTR(-ENOENT); 2161cbdc2ebcSJosef Bacik } 21625d4f98a2SYan Zheng 21633fd0a558SYan, Zheng next = node; 21643fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 21653fd0a558SYan, Zheng while (1) { 21663fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 21673fd0a558SYan, Zheng if (--index < 0) 21683fd0a558SYan, Zheng break; 21693fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 21703fd0a558SYan, Zheng } 21715d4f98a2SYan Zheng return root; 21725d4f98a2SYan Zheng } 21735d4f98a2SYan Zheng 21743fd0a558SYan, Zheng /* 217592a7cc42SQu Wenruo * Select a tree root for relocation. 217692a7cc42SQu Wenruo * 217792a7cc42SQu Wenruo * Return NULL if the block is not shareable. We should use do_relocation() in 217892a7cc42SQu Wenruo * this case. 217992a7cc42SQu Wenruo * 218092a7cc42SQu Wenruo * Return a tree root pointer if the block is shareable. 218192a7cc42SQu Wenruo * Return -ENOENT if the block is root of reloc tree. 21823fd0a558SYan, Zheng */ 21835d4f98a2SYan Zheng static noinline_for_stack 2184a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 21855d4f98a2SYan Zheng { 2186a26195a5SQu Wenruo struct btrfs_backref_node *next; 21873fd0a558SYan, Zheng struct btrfs_root *root; 21883fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2189a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 21903fd0a558SYan, Zheng int index = 0; 21913fd0a558SYan, Zheng 21923fd0a558SYan, Zheng next = node; 21933fd0a558SYan, Zheng while (1) { 21943fd0a558SYan, Zheng cond_resched(); 21953fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 21963fd0a558SYan, Zheng root = next->root; 21973fd0a558SYan, Zheng BUG_ON(!root); 21983fd0a558SYan, Zheng 219992a7cc42SQu Wenruo /* No other choice for non-shareable tree */ 220092a7cc42SQu Wenruo if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 22013fd0a558SYan, Zheng return root; 22023fd0a558SYan, Zheng 22033fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 22043fd0a558SYan, Zheng fs_root = root; 22053fd0a558SYan, Zheng 22063fd0a558SYan, Zheng if (next != node) 22073fd0a558SYan, Zheng return NULL; 22083fd0a558SYan, Zheng 22093fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 22103fd0a558SYan, Zheng if (!next || next->level <= node->level) 22113fd0a558SYan, Zheng break; 22123fd0a558SYan, Zheng } 22133fd0a558SYan, Zheng 22143fd0a558SYan, Zheng if (!fs_root) 22153fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 22163fd0a558SYan, Zheng return fs_root; 22175d4f98a2SYan Zheng } 22185d4f98a2SYan Zheng 22195d4f98a2SYan Zheng static noinline_for_stack 22203fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2221a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 22225d4f98a2SYan Zheng { 22230b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2224a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2225a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2226a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 22273fd0a558SYan, Zheng u64 num_bytes = 0; 22283fd0a558SYan, Zheng int index = 0; 22295d4f98a2SYan Zheng 22303fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 22313fd0a558SYan, Zheng 22323fd0a558SYan, Zheng while (next) { 22333fd0a558SYan, Zheng cond_resched(); 22345d4f98a2SYan Zheng while (1) { 22353fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 22365d4f98a2SYan Zheng break; 22375d4f98a2SYan Zheng 22380b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 22393fd0a558SYan, Zheng 22403fd0a558SYan, Zheng if (list_empty(&next->upper)) 22413fd0a558SYan, Zheng break; 22423fd0a558SYan, Zheng 22433fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2244a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 22453fd0a558SYan, Zheng edges[index++] = edge; 22463fd0a558SYan, Zheng next = edge->node[UPPER]; 22475d4f98a2SYan Zheng } 22483fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 22493fd0a558SYan, Zheng } 22503fd0a558SYan, Zheng return num_bytes; 22513fd0a558SYan, Zheng } 22523fd0a558SYan, Zheng 22533fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 22543fd0a558SYan, Zheng struct reloc_control *rc, 2255a26195a5SQu Wenruo struct btrfs_backref_node *node) 22563fd0a558SYan, Zheng { 22573fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2258da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 22593fd0a558SYan, Zheng u64 num_bytes; 22603fd0a558SYan, Zheng int ret; 22610647bf56SWang Shilong u64 tmp; 22623fd0a558SYan, Zheng 22633fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 22643fd0a558SYan, Zheng 22653fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 22660647bf56SWang Shilong rc->reserved_bytes += num_bytes; 22678ca17f0fSJosef Bacik 22688ca17f0fSJosef Bacik /* 22698ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 22708ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 22718ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 22728ca17f0fSJosef Bacik */ 22730647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 22748ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 22753fd0a558SYan, Zheng if (ret) { 2276da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 22770647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 22780647bf56SWang Shilong tmp <<= 1; 22790647bf56SWang Shilong /* 22800647bf56SWang Shilong * only one thread can access block_rsv at this point, 22810647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 22820647bf56SWang Shilong * we expand more reservation size here to allow enough 228352042d8eSAndrea Gelmini * space for relocation and we will return earlier in 22840647bf56SWang Shilong * enospc case. 22850647bf56SWang Shilong */ 2286da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 22870647bf56SWang Shilong RELOCATION_RESERVED_NODES; 22888ca17f0fSJosef Bacik return -EAGAIN; 22893fd0a558SYan, Zheng } 22903fd0a558SYan, Zheng 22913fd0a558SYan, Zheng return 0; 22923fd0a558SYan, Zheng } 22933fd0a558SYan, Zheng 22945d4f98a2SYan Zheng /* 22955d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 22965d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 22975d4f98a2SYan Zheng * 22985d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 22995d4f98a2SYan Zheng * in that case this function just updates pointers. 23005d4f98a2SYan Zheng */ 23015d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 23023fd0a558SYan, Zheng struct reloc_control *rc, 2303a26195a5SQu Wenruo struct btrfs_backref_node *node, 23045d4f98a2SYan Zheng struct btrfs_key *key, 23055d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 23065d4f98a2SYan Zheng { 2307a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2308a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2309a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 23105d4f98a2SYan Zheng struct btrfs_root *root; 23115d4f98a2SYan Zheng struct extent_buffer *eb; 23125d4f98a2SYan Zheng u32 blocksize; 23135d4f98a2SYan Zheng u64 bytenr; 23145d4f98a2SYan Zheng int slot; 23158df01fddSNikolay Borisov int ret = 0; 23165d4f98a2SYan Zheng 2317ffe30dd8SJosef Bacik /* 2318ffe30dd8SJosef Bacik * If we are lowest then this is the first time we're processing this 2319ffe30dd8SJosef Bacik * block, and thus shouldn't have an eb associated with it yet. 2320ffe30dd8SJosef Bacik */ 2321ffe30dd8SJosef Bacik ASSERT(!lowest || !node->eb); 23225d4f98a2SYan Zheng 23235d4f98a2SYan Zheng path->lowest_level = node->level + 1; 23243fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 23255d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 232682fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2327581c1760SQu Wenruo 23285d4f98a2SYan Zheng cond_resched(); 23295d4f98a2SYan Zheng 23305d4f98a2SYan Zheng upper = edge->node[UPPER]; 2331dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 2332cbdc2ebcSJosef Bacik if (IS_ERR(root)) { 2333cbdc2ebcSJosef Bacik ret = PTR_ERR(root); 2334cbdc2ebcSJosef Bacik goto next; 2335cbdc2ebcSJosef Bacik } 23365d4f98a2SYan Zheng 23373fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 23383fd0a558SYan, Zheng if (!lowest) { 2339e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 23408df01fddSNikolay Borisov if (ret < 0) 2341cbca7d59SFilipe Manana goto next; 23423fd0a558SYan, Zheng BUG_ON(ret); 23433fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 23443fd0a558SYan, Zheng if (node->eb->start == bytenr) 23453fd0a558SYan, Zheng goto next; 23463fd0a558SYan, Zheng } 2347b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 23483fd0a558SYan, Zheng } 23495d4f98a2SYan Zheng 23505d4f98a2SYan Zheng if (!upper->eb) { 23515d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 23523561b9dbSLiu Bo if (ret) { 23538df01fddSNikolay Borisov if (ret > 0) 23548df01fddSNikolay Borisov ret = -ENOENT; 23553561b9dbSLiu Bo 23563561b9dbSLiu Bo btrfs_release_path(path); 23575d4f98a2SYan Zheng break; 23585d4f98a2SYan Zheng } 23595d4f98a2SYan Zheng 23603fd0a558SYan, Zheng if (!upper->eb) { 23613fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 23623fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 23633fd0a558SYan, Zheng } else { 23643fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 23653fd0a558SYan, Zheng } 23663fd0a558SYan, Zheng 23673fd0a558SYan, Zheng upper->locked = 1; 23683fd0a558SYan, Zheng path->locks[upper->level] = 0; 23693fd0a558SYan, Zheng 23705d4f98a2SYan Zheng slot = path->slots[upper->level]; 2371b3b4aa74SDavid Sterba btrfs_release_path(path); 23725d4f98a2SYan Zheng } else { 2373e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 23748df01fddSNikolay Borisov if (ret < 0) 2375cbca7d59SFilipe Manana goto next; 23765d4f98a2SYan Zheng BUG_ON(ret); 23775d4f98a2SYan Zheng } 23785d4f98a2SYan Zheng 23795d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 23803fd0a558SYan, Zheng if (lowest) { 23814547f4d8SLiu Bo if (bytenr != node->bytenr) { 23824547f4d8SLiu Bo btrfs_err(root->fs_info, 23834547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 23844547f4d8SLiu Bo bytenr, node->bytenr, slot, 23854547f4d8SLiu Bo upper->eb->start); 23868df01fddSNikolay Borisov ret = -EIO; 23874547f4d8SLiu Bo goto next; 23884547f4d8SLiu Bo } 23895d4f98a2SYan Zheng } else { 23903fd0a558SYan, Zheng if (node->eb->start == bytenr) 23913fd0a558SYan, Zheng goto next; 23925d4f98a2SYan Zheng } 23935d4f98a2SYan Zheng 2394da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 2395c9752536SJosef Bacik eb = btrfs_read_node_slot(upper->eb, slot); 239664c043deSLiu Bo if (IS_ERR(eb)) { 23978df01fddSNikolay Borisov ret = PTR_ERR(eb); 239864c043deSLiu Bo goto next; 239997d9a8a4STsutomu Itoh } 24005d4f98a2SYan Zheng btrfs_tree_lock(eb); 24015d4f98a2SYan Zheng 24025d4f98a2SYan Zheng if (!node->eb) { 24035d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 24049631e4ccSJosef Bacik slot, &eb, BTRFS_NESTING_COW); 24053fd0a558SYan, Zheng btrfs_tree_unlock(eb); 24063fd0a558SYan, Zheng free_extent_buffer(eb); 24078df01fddSNikolay Borisov if (ret < 0) 24083fd0a558SYan, Zheng goto next; 2409ffe30dd8SJosef Bacik /* 2410ffe30dd8SJosef Bacik * We've just COWed this block, it should have updated 2411ffe30dd8SJosef Bacik * the correct backref node entry. 2412ffe30dd8SJosef Bacik */ 2413ffe30dd8SJosef Bacik ASSERT(node->eb == eb); 24145d4f98a2SYan Zheng } else { 24155d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 24165d4f98a2SYan Zheng node->eb->start); 24175d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 24185d4f98a2SYan Zheng trans->transid); 24195d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 24205d4f98a2SYan Zheng 242182fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 24225d4f98a2SYan Zheng node->eb->start, blocksize, 242382fa113fSQu Wenruo upper->eb->start); 242482fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 242582fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 242682fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 242782fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 24285d4f98a2SYan Zheng BUG_ON(ret); 24295d4f98a2SYan Zheng 24305d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 24315d4f98a2SYan Zheng BUG_ON(ret); 24325d4f98a2SYan Zheng } 24333fd0a558SYan, Zheng next: 24343fd0a558SYan, Zheng if (!upper->pending) 2435b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 24363fd0a558SYan, Zheng else 2437b0fe7078SQu Wenruo btrfs_backref_unlock_node_buffer(upper); 24388df01fddSNikolay Borisov if (ret) 24393fd0a558SYan, Zheng break; 24405d4f98a2SYan Zheng } 24413fd0a558SYan, Zheng 24428df01fddSNikolay Borisov if (!ret && node->pending) { 2443b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 24443fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 24453fd0a558SYan, Zheng node->pending = 0; 24465d4f98a2SYan Zheng } 24473fd0a558SYan, Zheng 24485d4f98a2SYan Zheng path->lowest_level = 0; 2449ffe30dd8SJosef Bacik 2450ffe30dd8SJosef Bacik /* 2451ffe30dd8SJosef Bacik * We should have allocated all of our space in the block rsv and thus 2452ffe30dd8SJosef Bacik * shouldn't ENOSPC. 2453ffe30dd8SJosef Bacik */ 2454ffe30dd8SJosef Bacik ASSERT(ret != -ENOSPC); 24558df01fddSNikolay Borisov return ret; 24565d4f98a2SYan Zheng } 24575d4f98a2SYan Zheng 24585d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 24593fd0a558SYan, Zheng struct reloc_control *rc, 2460a26195a5SQu Wenruo struct btrfs_backref_node *node, 24615d4f98a2SYan Zheng struct btrfs_path *path) 24625d4f98a2SYan Zheng { 24635d4f98a2SYan Zheng struct btrfs_key key; 24645d4f98a2SYan Zheng 24655d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 24663fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 24675d4f98a2SYan Zheng } 24685d4f98a2SYan Zheng 24695d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 24703fd0a558SYan, Zheng struct reloc_control *rc, 24713fd0a558SYan, Zheng struct btrfs_path *path, int err) 24725d4f98a2SYan Zheng { 24733fd0a558SYan, Zheng LIST_HEAD(list); 2474a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 2475a26195a5SQu Wenruo struct btrfs_backref_node *node; 24765d4f98a2SYan Zheng int level; 24775d4f98a2SYan Zheng int ret; 24785d4f98a2SYan Zheng 24795d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 24805d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 24815d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 2482a26195a5SQu Wenruo struct btrfs_backref_node, list); 24833fd0a558SYan, Zheng list_move_tail(&node->list, &list); 24843fd0a558SYan, Zheng BUG_ON(!node->pending); 24855d4f98a2SYan Zheng 24863fd0a558SYan, Zheng if (!err) { 24873fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 24885d4f98a2SYan Zheng if (ret < 0) 24895d4f98a2SYan Zheng err = ret; 24905d4f98a2SYan Zheng } 24915d4f98a2SYan Zheng } 24923fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 24933fd0a558SYan, Zheng } 24945d4f98a2SYan Zheng return err; 24955d4f98a2SYan Zheng } 24965d4f98a2SYan Zheng 24975d4f98a2SYan Zheng /* 24985d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 24995d4f98a2SYan Zheng * as processed. 25005d4f98a2SYan Zheng */ 25015d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 2502a26195a5SQu Wenruo struct btrfs_backref_node *node) 25035d4f98a2SYan Zheng { 2504a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2505a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2506a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 25075d4f98a2SYan Zheng int index = 0; 25085d4f98a2SYan Zheng 25095d4f98a2SYan Zheng while (next) { 25105d4f98a2SYan Zheng cond_resched(); 25115d4f98a2SYan Zheng while (1) { 25125d4f98a2SYan Zheng if (next->processed) 25135d4f98a2SYan Zheng break; 25145d4f98a2SYan Zheng 25159569cc20SQu Wenruo mark_block_processed(rc, next); 25165d4f98a2SYan Zheng 25175d4f98a2SYan Zheng if (list_empty(&next->upper)) 25185d4f98a2SYan Zheng break; 25195d4f98a2SYan Zheng 25205d4f98a2SYan Zheng edge = list_entry(next->upper.next, 2521a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 25225d4f98a2SYan Zheng edges[index++] = edge; 25235d4f98a2SYan Zheng next = edge->node[UPPER]; 25245d4f98a2SYan Zheng } 25255d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 25265d4f98a2SYan Zheng } 25275d4f98a2SYan Zheng } 25285d4f98a2SYan Zheng 25297476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 25305d4f98a2SYan Zheng { 2531da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 25327476dfdaSDavid Sterba 25335d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 25349655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 25355d4f98a2SYan Zheng return 1; 25365d4f98a2SYan Zheng return 0; 25375d4f98a2SYan Zheng } 25385d4f98a2SYan Zheng 25392ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 25405d4f98a2SYan Zheng struct tree_block *block) 25415d4f98a2SYan Zheng { 25425d4f98a2SYan Zheng struct extent_buffer *eb; 25435d4f98a2SYan Zheng 2544f7ba2d37SJosef Bacik eb = read_tree_block(fs_info, block->bytenr, block->owner, 2545f7ba2d37SJosef Bacik block->key.offset, block->level, NULL); 254664c043deSLiu Bo if (IS_ERR(eb)) { 254764c043deSLiu Bo return PTR_ERR(eb); 254864c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2549416bc658SJosef Bacik free_extent_buffer(eb); 2550416bc658SJosef Bacik return -EIO; 2551416bc658SJosef Bacik } 25525d4f98a2SYan Zheng if (block->level == 0) 25535d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 25545d4f98a2SYan Zheng else 25555d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 25565d4f98a2SYan Zheng free_extent_buffer(eb); 25575d4f98a2SYan Zheng block->key_ready = 1; 25585d4f98a2SYan Zheng return 0; 25595d4f98a2SYan Zheng } 25605d4f98a2SYan Zheng 25615d4f98a2SYan Zheng /* 25625d4f98a2SYan Zheng * helper function to relocate a tree block 25635d4f98a2SYan Zheng */ 25645d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 25655d4f98a2SYan Zheng struct reloc_control *rc, 2566a26195a5SQu Wenruo struct btrfs_backref_node *node, 25675d4f98a2SYan Zheng struct btrfs_key *key, 25685d4f98a2SYan Zheng struct btrfs_path *path) 25695d4f98a2SYan Zheng { 25705d4f98a2SYan Zheng struct btrfs_root *root; 25713fd0a558SYan, Zheng int ret = 0; 25725d4f98a2SYan Zheng 25733fd0a558SYan, Zheng if (!node) 25745d4f98a2SYan Zheng return 0; 25753fd0a558SYan, Zheng 25765f6b2e5cSJosef Bacik /* 25775f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 25785f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 25795f6b2e5cSJosef Bacik */ 25805f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 25815f6b2e5cSJosef Bacik if (ret) 25825f6b2e5cSJosef Bacik goto out; 25835f6b2e5cSJosef Bacik 25843fd0a558SYan, Zheng BUG_ON(node->processed); 2585147d256eSZhaolei root = select_one_root(node); 25863fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 25873fd0a558SYan, Zheng update_processed_blocks(rc, node); 25883fd0a558SYan, Zheng goto out; 25895d4f98a2SYan Zheng } 25905d4f98a2SYan Zheng 25913fd0a558SYan, Zheng if (root) { 259292a7cc42SQu Wenruo if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 25931c7bfa15SJosef Bacik /* 25941c7bfa15SJosef Bacik * This block was the root block of a root, and this is 25951c7bfa15SJosef Bacik * the first time we're processing the block and thus it 25961c7bfa15SJosef Bacik * should not have had the ->new_bytenr modified and 25971c7bfa15SJosef Bacik * should have not been included on the changed list. 25981c7bfa15SJosef Bacik * 25991c7bfa15SJosef Bacik * However in the case of corruption we could have 26001c7bfa15SJosef Bacik * multiple refs pointing to the same block improperly, 26011c7bfa15SJosef Bacik * and thus we would trip over these checks. ASSERT() 26021c7bfa15SJosef Bacik * for the developer case, because it could indicate a 26031c7bfa15SJosef Bacik * bug in the backref code, however error out for a 26041c7bfa15SJosef Bacik * normal user in the case of corruption. 26051c7bfa15SJosef Bacik */ 26061c7bfa15SJosef Bacik ASSERT(node->new_bytenr == 0); 26071c7bfa15SJosef Bacik ASSERT(list_empty(&node->list)); 26081c7bfa15SJosef Bacik if (node->new_bytenr || !list_empty(&node->list)) { 26091c7bfa15SJosef Bacik btrfs_err(root->fs_info, 26101c7bfa15SJosef Bacik "bytenr %llu has improper references to it", 26111c7bfa15SJosef Bacik node->bytenr); 26121c7bfa15SJosef Bacik ret = -EUCLEAN; 26131c7bfa15SJosef Bacik goto out; 26141c7bfa15SJosef Bacik } 2615d18c7bd9SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 2616d18c7bd9SJosef Bacik if (ret) 2617d18c7bd9SJosef Bacik goto out; 261839200e59SJosef Bacik /* 261939200e59SJosef Bacik * Another thread could have failed, need to check if we 262039200e59SJosef Bacik * have reloc_root actually set. 262139200e59SJosef Bacik */ 262239200e59SJosef Bacik if (!root->reloc_root) { 262339200e59SJosef Bacik ret = -ENOENT; 262439200e59SJosef Bacik goto out; 262539200e59SJosef Bacik } 26263fd0a558SYan, Zheng root = root->reloc_root; 26273fd0a558SYan, Zheng node->new_bytenr = root->node->start; 262800246528SJosef Bacik btrfs_put_root(node->root); 262900246528SJosef Bacik node->root = btrfs_grab_root(root); 26300b530bc5SJosef Bacik ASSERT(node->root); 26313fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 26323fd0a558SYan, Zheng } else { 26335d4f98a2SYan Zheng path->lowest_level = node->level; 26345d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2635b3b4aa74SDavid Sterba btrfs_release_path(path); 26363fd0a558SYan, Zheng if (ret > 0) 26375d4f98a2SYan Zheng ret = 0; 26383fd0a558SYan, Zheng } 26393fd0a558SYan, Zheng if (!ret) 26403fd0a558SYan, Zheng update_processed_blocks(rc, node); 26413fd0a558SYan, Zheng } else { 26423fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 26433fd0a558SYan, Zheng } 26445d4f98a2SYan Zheng out: 26450647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 2646023acb07SQu Wenruo btrfs_backref_cleanup_node(&rc->backref_cache, node); 26475d4f98a2SYan Zheng return ret; 26485d4f98a2SYan Zheng } 26495d4f98a2SYan Zheng 26505d4f98a2SYan Zheng /* 26515d4f98a2SYan Zheng * relocate a list of blocks 26525d4f98a2SYan Zheng */ 26535d4f98a2SYan Zheng static noinline_for_stack 26545d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 26555d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 26565d4f98a2SYan Zheng { 26572ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2658a26195a5SQu Wenruo struct btrfs_backref_node *node; 26595d4f98a2SYan Zheng struct btrfs_path *path; 26605d4f98a2SYan Zheng struct tree_block *block; 266198ff7b94SQu Wenruo struct tree_block *next; 26625d4f98a2SYan Zheng int ret; 26635d4f98a2SYan Zheng int err = 0; 26645d4f98a2SYan Zheng 26655d4f98a2SYan Zheng path = btrfs_alloc_path(); 2666e1a12670SLiu Bo if (!path) { 2667e1a12670SLiu Bo err = -ENOMEM; 266834c2b290SDavid Sterba goto out_free_blocks; 2669e1a12670SLiu Bo } 26705d4f98a2SYan Zheng 267198ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 267298ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 26735d4f98a2SYan Zheng if (!block->key_ready) 2674f7ba2d37SJosef Bacik btrfs_readahead_tree_block(fs_info, block->bytenr, 2675f7ba2d37SJosef Bacik block->owner, 0, 26763fbaf258SJosef Bacik block->level); 26775d4f98a2SYan Zheng } 26785d4f98a2SYan Zheng 267998ff7b94SQu Wenruo /* Get first keys */ 268098ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 268134c2b290SDavid Sterba if (!block->key_ready) { 26822ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 268334c2b290SDavid Sterba if (err) 268434c2b290SDavid Sterba goto out_free_path; 268534c2b290SDavid Sterba } 26865d4f98a2SYan Zheng } 26875d4f98a2SYan Zheng 268898ff7b94SQu Wenruo /* Do tree relocation */ 268998ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 26903fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 26915d4f98a2SYan Zheng block->level, block->bytenr); 26925d4f98a2SYan Zheng if (IS_ERR(node)) { 26935d4f98a2SYan Zheng err = PTR_ERR(node); 26945d4f98a2SYan Zheng goto out; 26955d4f98a2SYan Zheng } 26965d4f98a2SYan Zheng 26975d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 26985d4f98a2SYan Zheng path); 26995d4f98a2SYan Zheng if (ret < 0) { 27005d4f98a2SYan Zheng err = ret; 270150dbbb71SJosef Bacik break; 27025d4f98a2SYan Zheng } 27035d4f98a2SYan Zheng } 27045d4f98a2SYan Zheng out: 27053fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 27065d4f98a2SYan Zheng 270734c2b290SDavid Sterba out_free_path: 27085d4f98a2SYan Zheng btrfs_free_path(path); 270934c2b290SDavid Sterba out_free_blocks: 2710e1a12670SLiu Bo free_block_list(blocks); 27115d4f98a2SYan Zheng return err; 27125d4f98a2SYan Zheng } 27135d4f98a2SYan Zheng 2714056d9becSNikolay Borisov static noinline_for_stack int prealloc_file_extent_cluster( 2715056d9becSNikolay Borisov struct btrfs_inode *inode, 2716efa56464SYan, Zheng struct file_extent_cluster *cluster) 2717efa56464SYan, Zheng { 2718efa56464SYan, Zheng u64 alloc_hint = 0; 2719efa56464SYan, Zheng u64 start; 2720efa56464SYan, Zheng u64 end; 2721056d9becSNikolay Borisov u64 offset = inode->index_cnt; 2722efa56464SYan, Zheng u64 num_bytes; 27234e9d0d01SNikolay Borisov int nr; 2724efa56464SYan, Zheng int ret = 0; 2725dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 2726dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 2727214e61d0SNikolay Borisov u64 cur_offset = prealloc_start; 2728efa56464SYan, Zheng 2729efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2730056d9becSNikolay Borisov ret = btrfs_alloc_data_chunk_ondemand(inode, 2731dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 2732efa56464SYan, Zheng if (ret) 2733214e61d0SNikolay Borisov return ret; 2734efa56464SYan, Zheng 273532430c61SNaohiro Aota /* 273632430c61SNaohiro Aota * On a zoned filesystem, we cannot preallocate the file region. 273732430c61SNaohiro Aota * Instead, we dirty and fiemap_write the region. 273832430c61SNaohiro Aota */ 273932430c61SNaohiro Aota if (btrfs_is_zoned(inode->root->fs_info)) { 274032430c61SNaohiro Aota struct btrfs_root *root = inode->root; 274132430c61SNaohiro Aota struct btrfs_trans_handle *trans; 274232430c61SNaohiro Aota 274332430c61SNaohiro Aota end = cluster->end - offset + 1; 274432430c61SNaohiro Aota trans = btrfs_start_transaction(root, 1); 274532430c61SNaohiro Aota if (IS_ERR(trans)) 274632430c61SNaohiro Aota return PTR_ERR(trans); 274732430c61SNaohiro Aota 274832430c61SNaohiro Aota inode->vfs_inode.i_ctime = current_time(&inode->vfs_inode); 274932430c61SNaohiro Aota i_size_write(&inode->vfs_inode, end); 275032430c61SNaohiro Aota ret = btrfs_update_inode(trans, root, inode); 275132430c61SNaohiro Aota if (ret) { 275232430c61SNaohiro Aota btrfs_abort_transaction(trans, ret); 275332430c61SNaohiro Aota btrfs_end_transaction(trans); 275432430c61SNaohiro Aota return ret; 275532430c61SNaohiro Aota } 275632430c61SNaohiro Aota 275732430c61SNaohiro Aota return btrfs_end_transaction(trans); 275832430c61SNaohiro Aota } 275932430c61SNaohiro Aota 276064708539SJosef Bacik btrfs_inode_lock(&inode->vfs_inode, 0); 27614e9d0d01SNikolay Borisov for (nr = 0; nr < cluster->nr; nr++) { 2762efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2763efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2764efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2765efa56464SYan, Zheng else 2766efa56464SYan, Zheng end = cluster->end - offset; 2767efa56464SYan, Zheng 2768056d9becSNikolay Borisov lock_extent(&inode->io_tree, start, end); 2769efa56464SYan, Zheng num_bytes = end + 1 - start; 2770056d9becSNikolay Borisov ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start, 2771efa56464SYan, Zheng num_bytes, num_bytes, 2772efa56464SYan, Zheng end + 1, &alloc_hint); 277318513091SWang Xiaoguang cur_offset = end + 1; 2774056d9becSNikolay Borisov unlock_extent(&inode->io_tree, start, end); 2775efa56464SYan, Zheng if (ret) 2776efa56464SYan, Zheng break; 2777efa56464SYan, Zheng } 277864708539SJosef Bacik btrfs_inode_unlock(&inode->vfs_inode, 0); 2779214e61d0SNikolay Borisov 278018513091SWang Xiaoguang if (cur_offset < prealloc_end) 2781056d9becSNikolay Borisov btrfs_free_reserved_data_space_noquota(inode->root->fs_info, 2782a89ef455SFilipe Manana prealloc_end + 1 - cur_offset); 2783efa56464SYan, Zheng return ret; 2784efa56464SYan, Zheng } 2785efa56464SYan, Zheng 2786efa56464SYan, Zheng static noinline_for_stack 27870257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 27880257bb82SYan, Zheng u64 block_start) 27895d4f98a2SYan Zheng { 27905d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 27915d4f98a2SYan Zheng struct extent_map *em; 27920257bb82SYan, Zheng int ret = 0; 27935d4f98a2SYan Zheng 2794172ddd60SDavid Sterba em = alloc_extent_map(); 27950257bb82SYan, Zheng if (!em) 27960257bb82SYan, Zheng return -ENOMEM; 27970257bb82SYan, Zheng 27985d4f98a2SYan Zheng em->start = start; 27990257bb82SYan, Zheng em->len = end + 1 - start; 28000257bb82SYan, Zheng em->block_len = em->len; 28010257bb82SYan, Zheng em->block_start = block_start; 28025d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 28035d4f98a2SYan Zheng 2804d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 28055d4f98a2SYan Zheng while (1) { 2806890871beSChris Mason write_lock(&em_tree->lock); 280709a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 2808890871beSChris Mason write_unlock(&em_tree->lock); 28095d4f98a2SYan Zheng if (ret != -EEXIST) { 28105d4f98a2SYan Zheng free_extent_map(em); 28115d4f98a2SYan Zheng break; 28125d4f98a2SYan Zheng } 2813dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 28145d4f98a2SYan Zheng } 2815d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 28160257bb82SYan, Zheng return ret; 28170257bb82SYan, Zheng } 28185d4f98a2SYan Zheng 2819726a3421SQu Wenruo /* 2820726a3421SQu Wenruo * Allow error injection to test balance cancellation 2821726a3421SQu Wenruo */ 28221fec12a5SJosef Bacik noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 2823726a3421SQu Wenruo { 28245cb502f4SQu Wenruo return atomic_read(&fs_info->balance_cancel_req) || 28255cb502f4SQu Wenruo fatal_signal_pending(current); 2826726a3421SQu Wenruo } 2827726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 2828726a3421SQu Wenruo 28290257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 28300257bb82SYan, Zheng struct file_extent_cluster *cluster) 28310257bb82SYan, Zheng { 28322ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 28330257bb82SYan, Zheng u64 page_start; 28340257bb82SYan, Zheng u64 page_end; 28350257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 28360257bb82SYan, Zheng unsigned long index; 28370257bb82SYan, Zheng unsigned long last_index; 28380257bb82SYan, Zheng struct page *page; 28390257bb82SYan, Zheng struct file_ra_state *ra; 28403b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 28410257bb82SYan, Zheng int nr = 0; 28420257bb82SYan, Zheng int ret = 0; 28430257bb82SYan, Zheng 28440257bb82SYan, Zheng if (!cluster->nr) 28450257bb82SYan, Zheng return 0; 28460257bb82SYan, Zheng 28470257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 28480257bb82SYan, Zheng if (!ra) 28490257bb82SYan, Zheng return -ENOMEM; 28500257bb82SYan, Zheng 2851056d9becSNikolay Borisov ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster); 28520257bb82SYan, Zheng if (ret) 2853efa56464SYan, Zheng goto out; 28540257bb82SYan, Zheng 28550257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 28560257bb82SYan, Zheng 2857efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2858efa56464SYan, Zheng cluster->end - offset, cluster->start); 2859efa56464SYan, Zheng if (ret) 2860efa56464SYan, Zheng goto out; 2861efa56464SYan, Zheng 286209cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 286309cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 28640257bb82SYan, Zheng while (index <= last_index) { 28659f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 28669f3db423SNikolay Borisov PAGE_SIZE); 2867efa56464SYan, Zheng if (ret) 2868efa56464SYan, Zheng goto out; 2869efa56464SYan, Zheng 28700257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 28710257bb82SYan, Zheng if (!page) { 28720257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 28730257bb82SYan, Zheng ra, NULL, index, 28740257bb82SYan, Zheng last_index + 1 - index); 2875a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 28763b16a4e3SJosef Bacik mask); 28770257bb82SYan, Zheng if (!page) { 2878691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 287943b18595SQu Wenruo PAGE_SIZE, true); 288044db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 28818702ba93SQu Wenruo PAGE_SIZE); 28820257bb82SYan, Zheng ret = -ENOMEM; 2883efa56464SYan, Zheng goto out; 28840257bb82SYan, Zheng } 28850257bb82SYan, Zheng } 288632443de3SQu Wenruo ret = set_page_extent_mapped(page); 288732443de3SQu Wenruo if (ret < 0) { 288832443de3SQu Wenruo btrfs_delalloc_release_metadata(BTRFS_I(inode), 288932443de3SQu Wenruo PAGE_SIZE, true); 289032443de3SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 289132443de3SQu Wenruo unlock_page(page); 289232443de3SQu Wenruo put_page(page); 289332443de3SQu Wenruo goto out; 289432443de3SQu Wenruo } 28950257bb82SYan, Zheng 28960257bb82SYan, Zheng if (PageReadahead(page)) { 28970257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 28980257bb82SYan, Zheng ra, NULL, page, index, 28990257bb82SYan, Zheng last_index + 1 - index); 29000257bb82SYan, Zheng } 29010257bb82SYan, Zheng 29020257bb82SYan, Zheng if (!PageUptodate(page)) { 29030257bb82SYan, Zheng btrfs_readpage(NULL, page); 29040257bb82SYan, Zheng lock_page(page); 29050257bb82SYan, Zheng if (!PageUptodate(page)) { 29060257bb82SYan, Zheng unlock_page(page); 290709cbfeafSKirill A. Shutemov put_page(page); 2908691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 290943b18595SQu Wenruo PAGE_SIZE, true); 29108b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 29118702ba93SQu Wenruo PAGE_SIZE); 29120257bb82SYan, Zheng ret = -EIO; 2913efa56464SYan, Zheng goto out; 29140257bb82SYan, Zheng } 29150257bb82SYan, Zheng } 29160257bb82SYan, Zheng 29174eee4fa4SMiao Xie page_start = page_offset(page); 291809cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 29190257bb82SYan, Zheng 2920d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 29210257bb82SYan, Zheng 29220257bb82SYan, Zheng if (nr < cluster->nr && 29230257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 29240257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 29250257bb82SYan, Zheng page_start, page_end, 2926ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 29270257bb82SYan, Zheng nr++; 29280257bb82SYan, Zheng } 29290257bb82SYan, Zheng 2930c2566f22SNikolay Borisov ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, 2931c2566f22SNikolay Borisov page_end, 0, NULL); 2932765f3cebSNikolay Borisov if (ret) { 2933765f3cebSNikolay Borisov unlock_page(page); 2934765f3cebSNikolay Borisov put_page(page); 2935765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 293643b18595SQu Wenruo PAGE_SIZE, true); 2937765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 29388702ba93SQu Wenruo PAGE_SIZE); 2939765f3cebSNikolay Borisov 2940765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 2941765f3cebSNikolay Borisov page_start, page_end, 2942765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 2943765f3cebSNikolay Borisov goto out; 2944765f3cebSNikolay Borisov 2945765f3cebSNikolay Borisov } 29460257bb82SYan, Zheng set_page_dirty(page); 29470257bb82SYan, Zheng 29480257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 2949d0082371SJeff Mahoney page_start, page_end); 29500257bb82SYan, Zheng unlock_page(page); 295109cbfeafSKirill A. Shutemov put_page(page); 29520257bb82SYan, Zheng 29530257bb82SYan, Zheng index++; 29548702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 2955efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 29562ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 29577f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 29587f913c7cSQu Wenruo ret = -ECANCELED; 29597f913c7cSQu Wenruo goto out; 29607f913c7cSQu Wenruo } 29610257bb82SYan, Zheng } 29620257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 296332430c61SNaohiro Aota if (btrfs_is_zoned(fs_info) && !ret) 296432430c61SNaohiro Aota ret = btrfs_wait_ordered_range(inode, 0, (u64)-1); 2965efa56464SYan, Zheng out: 29660257bb82SYan, Zheng kfree(ra); 29670257bb82SYan, Zheng return ret; 29680257bb82SYan, Zheng } 29690257bb82SYan, Zheng 29700257bb82SYan, Zheng static noinline_for_stack 29710257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 29720257bb82SYan, Zheng struct file_extent_cluster *cluster) 29730257bb82SYan, Zheng { 29740257bb82SYan, Zheng int ret; 29750257bb82SYan, Zheng 29760257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 29770257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 29780257bb82SYan, Zheng if (ret) 29790257bb82SYan, Zheng return ret; 29800257bb82SYan, Zheng cluster->nr = 0; 29810257bb82SYan, Zheng } 29820257bb82SYan, Zheng 29830257bb82SYan, Zheng if (!cluster->nr) 29840257bb82SYan, Zheng cluster->start = extent_key->objectid; 29850257bb82SYan, Zheng else 29860257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 29870257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 29880257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 29890257bb82SYan, Zheng cluster->nr++; 29900257bb82SYan, Zheng 29910257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 29920257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 29930257bb82SYan, Zheng if (ret) 29940257bb82SYan, Zheng return ret; 29950257bb82SYan, Zheng cluster->nr = 0; 29960257bb82SYan, Zheng } 29970257bb82SYan, Zheng return 0; 29985d4f98a2SYan Zheng } 29995d4f98a2SYan Zheng 30005d4f98a2SYan Zheng /* 30015d4f98a2SYan Zheng * helper to add a tree block to the list. 30025d4f98a2SYan Zheng * the major work is getting the generation and level of the block 30035d4f98a2SYan Zheng */ 30045d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 30055d4f98a2SYan Zheng struct btrfs_key *extent_key, 30065d4f98a2SYan Zheng struct btrfs_path *path, 30075d4f98a2SYan Zheng struct rb_root *blocks) 30085d4f98a2SYan Zheng { 30095d4f98a2SYan Zheng struct extent_buffer *eb; 30105d4f98a2SYan Zheng struct btrfs_extent_item *ei; 30115d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 30125d4f98a2SYan Zheng struct tree_block *block; 30135d4f98a2SYan Zheng struct rb_node *rb_node; 30145d4f98a2SYan Zheng u32 item_size; 30155d4f98a2SYan Zheng int level = -1; 30167fdf4b60SWang Shilong u64 generation; 3017f7ba2d37SJosef Bacik u64 owner = 0; 30185d4f98a2SYan Zheng 30195d4f98a2SYan Zheng eb = path->nodes[0]; 30205d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 30215d4f98a2SYan Zheng 30223173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 30233173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 3024f7ba2d37SJosef Bacik unsigned long ptr = 0, end; 3025f7ba2d37SJosef Bacik 30265d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 30275d4f98a2SYan Zheng struct btrfs_extent_item); 3028f7ba2d37SJosef Bacik end = (unsigned long)ei + item_size; 30293173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 30305d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 30315d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 3032f7ba2d37SJosef Bacik ptr = (unsigned long)(bi + 1); 30335d4f98a2SYan Zheng } else { 30343173a18fSJosef Bacik level = (int)extent_key->offset; 3035f7ba2d37SJosef Bacik ptr = (unsigned long)(ei + 1); 30363173a18fSJosef Bacik } 30373173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 3038f7ba2d37SJosef Bacik 3039f7ba2d37SJosef Bacik /* 3040f7ba2d37SJosef Bacik * We're reading random blocks without knowing their owner ahead 3041f7ba2d37SJosef Bacik * of time. This is ok most of the time, as all reloc roots and 3042f7ba2d37SJosef Bacik * fs roots have the same lock type. However normal trees do 3043f7ba2d37SJosef Bacik * not, and the only way to know ahead of time is to read the 3044f7ba2d37SJosef Bacik * inline ref offset. We know it's an fs root if 3045f7ba2d37SJosef Bacik * 3046f7ba2d37SJosef Bacik * 1. There's more than one ref. 3047f7ba2d37SJosef Bacik * 2. There's a SHARED_DATA_REF_KEY set. 3048f7ba2d37SJosef Bacik * 3. FULL_BACKREF is set on the flags. 3049f7ba2d37SJosef Bacik * 3050f7ba2d37SJosef Bacik * Otherwise it's safe to assume that the ref offset == the 3051f7ba2d37SJosef Bacik * owner of this block, so we can use that when calling 3052f7ba2d37SJosef Bacik * read_tree_block. 3053f7ba2d37SJosef Bacik */ 3054f7ba2d37SJosef Bacik if (btrfs_extent_refs(eb, ei) == 1 && 3055f7ba2d37SJosef Bacik !(btrfs_extent_flags(eb, ei) & 3056f7ba2d37SJosef Bacik BTRFS_BLOCK_FLAG_FULL_BACKREF) && 3057f7ba2d37SJosef Bacik ptr < end) { 3058f7ba2d37SJosef Bacik struct btrfs_extent_inline_ref *iref; 3059f7ba2d37SJosef Bacik int type; 3060f7ba2d37SJosef Bacik 3061f7ba2d37SJosef Bacik iref = (struct btrfs_extent_inline_ref *)ptr; 3062f7ba2d37SJosef Bacik type = btrfs_get_extent_inline_ref_type(eb, iref, 3063f7ba2d37SJosef Bacik BTRFS_REF_TYPE_BLOCK); 3064f7ba2d37SJosef Bacik if (type == BTRFS_REF_TYPE_INVALID) 3065f7ba2d37SJosef Bacik return -EINVAL; 3066f7ba2d37SJosef Bacik if (type == BTRFS_TREE_BLOCK_REF_KEY) 3067f7ba2d37SJosef Bacik owner = btrfs_extent_inline_ref_offset(eb, iref); 3068f7ba2d37SJosef Bacik } 30696d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3070ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3071ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3072ba3c2b19SNikolay Borisov return -EINVAL; 30733173a18fSJosef Bacik } else { 30745d4f98a2SYan Zheng BUG(); 30755d4f98a2SYan Zheng } 30765d4f98a2SYan Zheng 3077b3b4aa74SDavid Sterba btrfs_release_path(path); 30785d4f98a2SYan Zheng 30795d4f98a2SYan Zheng BUG_ON(level == -1); 30805d4f98a2SYan Zheng 30815d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 30825d4f98a2SYan Zheng if (!block) 30835d4f98a2SYan Zheng return -ENOMEM; 30845d4f98a2SYan Zheng 30855d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3086da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 30875d4f98a2SYan Zheng block->key.offset = generation; 30885d4f98a2SYan Zheng block->level = level; 30895d4f98a2SYan Zheng block->key_ready = 0; 3090f7ba2d37SJosef Bacik block->owner = owner; 30915d4f98a2SYan Zheng 3092e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 309343c04fb1SJeff Mahoney if (rb_node) 3094982c92cbSQu Wenruo btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr, 3095982c92cbSQu Wenruo -EEXIST); 30965d4f98a2SYan Zheng 30975d4f98a2SYan Zheng return 0; 30985d4f98a2SYan Zheng } 30995d4f98a2SYan Zheng 31005d4f98a2SYan Zheng /* 31015d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 31025d4f98a2SYan Zheng */ 31035d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 31045d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 31055d4f98a2SYan Zheng struct rb_root *blocks) 31065d4f98a2SYan Zheng { 31070b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31085d4f98a2SYan Zheng struct btrfs_path *path; 31095d4f98a2SYan Zheng struct btrfs_key key; 31105d4f98a2SYan Zheng int ret; 31110b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 31125d4f98a2SYan Zheng 31137476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 31145d4f98a2SYan Zheng return 0; 31155d4f98a2SYan Zheng 3116e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 31175d4f98a2SYan Zheng return 0; 31185d4f98a2SYan Zheng 31195d4f98a2SYan Zheng path = btrfs_alloc_path(); 31205d4f98a2SYan Zheng if (!path) 31215d4f98a2SYan Zheng return -ENOMEM; 3122aee68ee5SJosef Bacik again: 31235d4f98a2SYan Zheng key.objectid = bytenr; 3124aee68ee5SJosef Bacik if (skinny) { 3125aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3126aee68ee5SJosef Bacik key.offset = (u64)-1; 3127aee68ee5SJosef Bacik } else { 31285d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31295d4f98a2SYan Zheng key.offset = blocksize; 3130aee68ee5SJosef Bacik } 31315d4f98a2SYan Zheng 31325d4f98a2SYan Zheng path->search_commit_root = 1; 31335d4f98a2SYan Zheng path->skip_locking = 1; 31345d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 31355d4f98a2SYan Zheng if (ret < 0) 31365d4f98a2SYan Zheng goto out; 31375d4f98a2SYan Zheng 3138aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3139aee68ee5SJosef Bacik if (path->slots[0]) { 3140aee68ee5SJosef Bacik path->slots[0]--; 3141aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3142aee68ee5SJosef Bacik path->slots[0]); 31433173a18fSJosef Bacik if (key.objectid == bytenr && 3144aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3145aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3146aee68ee5SJosef Bacik key.offset == blocksize))) 31473173a18fSJosef Bacik ret = 0; 31483173a18fSJosef Bacik } 3149aee68ee5SJosef Bacik 3150aee68ee5SJosef Bacik if (ret) { 3151aee68ee5SJosef Bacik skinny = false; 3152aee68ee5SJosef Bacik btrfs_release_path(path); 3153aee68ee5SJosef Bacik goto again; 3154aee68ee5SJosef Bacik } 3155aee68ee5SJosef Bacik } 3156cdccee99SLiu Bo if (ret) { 3157cdccee99SLiu Bo ASSERT(ret == 1); 3158cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3159cdccee99SLiu Bo btrfs_err(fs_info, 3160cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3161cdccee99SLiu Bo bytenr); 3162cdccee99SLiu Bo WARN_ON(1); 3163cdccee99SLiu Bo ret = -EINVAL; 3164cdccee99SLiu Bo goto out; 3165cdccee99SLiu Bo } 31663173a18fSJosef Bacik 31675d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 31685d4f98a2SYan Zheng out: 31695d4f98a2SYan Zheng btrfs_free_path(path); 31705d4f98a2SYan Zheng return ret; 31715d4f98a2SYan Zheng } 31725d4f98a2SYan Zheng 31730af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 317432da5386SDavid Sterba struct btrfs_block_group *block_group, 31751bbc621eSChris Mason struct inode *inode, 31761bbc621eSChris Mason u64 ino) 31770af3d00bSJosef Bacik { 31780af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 31790af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 31800af3d00bSJosef Bacik int ret = 0; 31810af3d00bSJosef Bacik 31820af3d00bSJosef Bacik if (inode) 31830af3d00bSJosef Bacik goto truncate; 31840af3d00bSJosef Bacik 31850202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, ino, root); 31862e19f1f9SAl Viro if (IS_ERR(inode)) 31870af3d00bSJosef Bacik return -ENOENT; 31880af3d00bSJosef Bacik 31890af3d00bSJosef Bacik truncate: 31902ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 31917b61cd92SMiao Xie &fs_info->global_block_rsv); 31927b61cd92SMiao Xie if (ret) 31937b61cd92SMiao Xie goto out; 31947b61cd92SMiao Xie 31957a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 31960af3d00bSJosef Bacik if (IS_ERR(trans)) { 31973612b495STsutomu Itoh ret = PTR_ERR(trans); 31980af3d00bSJosef Bacik goto out; 31990af3d00bSJosef Bacik } 32000af3d00bSJosef Bacik 320177ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 32020af3d00bSJosef Bacik 32033a45bb20SJeff Mahoney btrfs_end_transaction(trans); 32042ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 32050af3d00bSJosef Bacik out: 32060af3d00bSJosef Bacik iput(inode); 32070af3d00bSJosef Bacik return ret; 32080af3d00bSJosef Bacik } 32090af3d00bSJosef Bacik 32105d4f98a2SYan Zheng /* 321119b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 321219b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 32135d4f98a2SYan Zheng */ 321419b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 321519b546d7SQu Wenruo struct btrfs_block_group *block_group, 321619b546d7SQu Wenruo u64 data_bytenr) 32175d4f98a2SYan Zheng { 321819b546d7SQu Wenruo u64 space_cache_ino; 321919b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 32205d4f98a2SYan Zheng struct btrfs_key key; 322119b546d7SQu Wenruo bool found = false; 322219b546d7SQu Wenruo int i; 32235d4f98a2SYan Zheng int ret; 32245d4f98a2SYan Zheng 322519b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 322619b546d7SQu Wenruo return 0; 32275d4f98a2SYan Zheng 322819b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 322950e31ef4SQu Wenruo u8 type; 323050e31ef4SQu Wenruo 323119b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 323219b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 323319b546d7SQu Wenruo continue; 323419b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 323550e31ef4SQu Wenruo type = btrfs_file_extent_type(leaf, ei); 323650e31ef4SQu Wenruo 323750e31ef4SQu Wenruo if ((type == BTRFS_FILE_EXTENT_REG || 323850e31ef4SQu Wenruo type == BTRFS_FILE_EXTENT_PREALLOC) && 323919b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 324019b546d7SQu Wenruo found = true; 324119b546d7SQu Wenruo space_cache_ino = key.objectid; 324219b546d7SQu Wenruo break; 324319b546d7SQu Wenruo } 324419b546d7SQu Wenruo } 324519b546d7SQu Wenruo if (!found) 324619b546d7SQu Wenruo return -ENOENT; 324719b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 324819b546d7SQu Wenruo space_cache_ino); 32490af3d00bSJosef Bacik return ret; 32505d4f98a2SYan Zheng } 32515d4f98a2SYan Zheng 32525d4f98a2SYan Zheng /* 32532c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 32545d4f98a2SYan Zheng */ 32555d4f98a2SYan Zheng static noinline_for_stack 32565d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 32575d4f98a2SYan Zheng struct btrfs_key *extent_key, 32585d4f98a2SYan Zheng struct btrfs_path *path, 32595d4f98a2SYan Zheng struct rb_root *blocks) 32605d4f98a2SYan Zheng { 326119b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 326219b546d7SQu Wenruo struct ulist *leaves = NULL; 326319b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 326419b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 326519b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3266647f63bdSFilipe David Borba Manana int ret = 0; 32675d4f98a2SYan Zheng 3268b3b4aa74SDavid Sterba btrfs_release_path(path); 326919b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 327019b546d7SQu Wenruo 0, &leaves, NULL, true); 327119b546d7SQu Wenruo if (ret < 0) 327219b546d7SQu Wenruo return ret; 327319b546d7SQu Wenruo 327419b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 327519b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 327619b546d7SQu Wenruo struct extent_buffer *eb; 327719b546d7SQu Wenruo 32781b7ec85eSJosef Bacik eb = read_tree_block(fs_info, ref_node->val, 0, 0, 0, NULL); 327919b546d7SQu Wenruo if (IS_ERR(eb)) { 328019b546d7SQu Wenruo ret = PTR_ERR(eb); 328119b546d7SQu Wenruo break; 328219b546d7SQu Wenruo } 328319b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 328419b546d7SQu Wenruo extent_key->objectid); 328519b546d7SQu Wenruo free_extent_buffer(eb); 328619b546d7SQu Wenruo if (ret < 0) 328719b546d7SQu Wenruo break; 328819b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 328919b546d7SQu Wenruo if (ret < 0) 329019b546d7SQu Wenruo break; 329119b546d7SQu Wenruo } 329219b546d7SQu Wenruo if (ret < 0) 32935d4f98a2SYan Zheng free_block_list(blocks); 329419b546d7SQu Wenruo ulist_free(leaves); 329519b546d7SQu Wenruo return ret; 32965d4f98a2SYan Zheng } 32975d4f98a2SYan Zheng 32985d4f98a2SYan Zheng /* 32992c016dc2SLiu Bo * helper to find next unprocessed extent 33005d4f98a2SYan Zheng */ 33015d4f98a2SYan Zheng static noinline_for_stack 3302147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 33033fd0a558SYan, Zheng struct btrfs_key *extent_key) 33045d4f98a2SYan Zheng { 33050b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33065d4f98a2SYan Zheng struct btrfs_key key; 33075d4f98a2SYan Zheng struct extent_buffer *leaf; 33085d4f98a2SYan Zheng u64 start, end, last; 33095d4f98a2SYan Zheng int ret; 33105d4f98a2SYan Zheng 3311b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 33125d4f98a2SYan Zheng while (1) { 33135d4f98a2SYan Zheng cond_resched(); 33145d4f98a2SYan Zheng if (rc->search_start >= last) { 33155d4f98a2SYan Zheng ret = 1; 33165d4f98a2SYan Zheng break; 33175d4f98a2SYan Zheng } 33185d4f98a2SYan Zheng 33195d4f98a2SYan Zheng key.objectid = rc->search_start; 33205d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 33215d4f98a2SYan Zheng key.offset = 0; 33225d4f98a2SYan Zheng 33235d4f98a2SYan Zheng path->search_commit_root = 1; 33245d4f98a2SYan Zheng path->skip_locking = 1; 33255d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 33265d4f98a2SYan Zheng 0, 0); 33275d4f98a2SYan Zheng if (ret < 0) 33285d4f98a2SYan Zheng break; 33295d4f98a2SYan Zheng next: 33305d4f98a2SYan Zheng leaf = path->nodes[0]; 33315d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 33325d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 33335d4f98a2SYan Zheng if (ret != 0) 33345d4f98a2SYan Zheng break; 33355d4f98a2SYan Zheng leaf = path->nodes[0]; 33365d4f98a2SYan Zheng } 33375d4f98a2SYan Zheng 33385d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 33395d4f98a2SYan Zheng if (key.objectid >= last) { 33405d4f98a2SYan Zheng ret = 1; 33415d4f98a2SYan Zheng break; 33425d4f98a2SYan Zheng } 33435d4f98a2SYan Zheng 33443173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 33453173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 33463173a18fSJosef Bacik path->slots[0]++; 33473173a18fSJosef Bacik goto next; 33483173a18fSJosef Bacik } 33493173a18fSJosef Bacik 33503173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 33515d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 33525d4f98a2SYan Zheng path->slots[0]++; 33535d4f98a2SYan Zheng goto next; 33545d4f98a2SYan Zheng } 33555d4f98a2SYan Zheng 33563173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 33570b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 33583173a18fSJosef Bacik rc->search_start) { 33593173a18fSJosef Bacik path->slots[0]++; 33603173a18fSJosef Bacik goto next; 33613173a18fSJosef Bacik } 33623173a18fSJosef Bacik 33635d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 33645d4f98a2SYan Zheng key.objectid, &start, &end, 3365e6138876SJosef Bacik EXTENT_DIRTY, NULL); 33665d4f98a2SYan Zheng 33675d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3368b3b4aa74SDavid Sterba btrfs_release_path(path); 33695d4f98a2SYan Zheng rc->search_start = end + 1; 33705d4f98a2SYan Zheng } else { 33713173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 33725d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 33733173a18fSJosef Bacik else 33743173a18fSJosef Bacik rc->search_start = key.objectid + 33750b246afaSJeff Mahoney fs_info->nodesize; 33763fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 33775d4f98a2SYan Zheng return 0; 33785d4f98a2SYan Zheng } 33795d4f98a2SYan Zheng } 3380b3b4aa74SDavid Sterba btrfs_release_path(path); 33815d4f98a2SYan Zheng return ret; 33825d4f98a2SYan Zheng } 33835d4f98a2SYan Zheng 33845d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 33855d4f98a2SYan Zheng { 33865d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33877585717fSChris Mason 33887585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 33895d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 33907585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 33915d4f98a2SYan Zheng } 33925d4f98a2SYan Zheng 33935d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 33945d4f98a2SYan Zheng { 33955d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33967585717fSChris Mason 33977585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 33985d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 33997585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 34005d4f98a2SYan Zheng } 34015d4f98a2SYan Zheng 34025d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 34035d4f98a2SYan Zheng { 34045d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 34055d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 34065d4f98a2SYan Zheng return 1; 34075d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 34085d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 34095d4f98a2SYan Zheng return 1; 34105d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 34115d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 34125d4f98a2SYan Zheng return 1; 34135d4f98a2SYan Zheng return 0; 34145d4f98a2SYan Zheng } 34155d4f98a2SYan Zheng 34163fd0a558SYan, Zheng static noinline_for_stack 34173fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 34183fd0a558SYan, Zheng { 34193fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3420ac2fabacSJosef Bacik int ret; 34213fd0a558SYan, Zheng 34222ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 342366d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 34243fd0a558SYan, Zheng if (!rc->block_rsv) 34253fd0a558SYan, Zheng return -ENOMEM; 34263fd0a558SYan, Zheng 34273fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3428b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 34293fd0a558SYan, Zheng rc->extents_found = 0; 34303fd0a558SYan, Zheng rc->nodes_relocated = 0; 34313fd0a558SYan, Zheng rc->merging_rsv_size = 0; 34320647bf56SWang Shilong rc->reserved_bytes = 0; 3433da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 34340647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3435ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3436ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3437ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3438ac2fabacSJosef Bacik if (ret) 3439ac2fabacSJosef Bacik return ret; 34403fd0a558SYan, Zheng 34413fd0a558SYan, Zheng rc->create_reloc_tree = 1; 34423fd0a558SYan, Zheng set_reloc_control(rc); 34433fd0a558SYan, Zheng 34447a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 344528818947SLiu Bo if (IS_ERR(trans)) { 344628818947SLiu Bo unset_reloc_control(rc); 344728818947SLiu Bo /* 344828818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 344928818947SLiu Bo * cleanup. And callers are responsible to free the above 345028818947SLiu Bo * block rsv. 345128818947SLiu Bo */ 345228818947SLiu Bo return PTR_ERR(trans); 345328818947SLiu Bo } 34543a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 34553fd0a558SYan, Zheng return 0; 34563fd0a558SYan, Zheng } 345776dda93cSYan, Zheng 34585d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 34595d4f98a2SYan Zheng { 34602ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 34615d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 34625d4f98a2SYan Zheng struct btrfs_key key; 34635d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 34645d4f98a2SYan Zheng struct btrfs_path *path; 34655d4f98a2SYan Zheng struct btrfs_extent_item *ei; 34665d4f98a2SYan Zheng u64 flags; 34675d4f98a2SYan Zheng u32 item_size; 34685d4f98a2SYan Zheng int ret; 34695d4f98a2SYan Zheng int err = 0; 3470c87f08caSChris Mason int progress = 0; 34715d4f98a2SYan Zheng 34725d4f98a2SYan Zheng path = btrfs_alloc_path(); 34733fd0a558SYan, Zheng if (!path) 34745d4f98a2SYan Zheng return -ENOMEM; 3475e4058b54SDavid Sterba path->reada = READA_FORWARD; 34763fd0a558SYan, Zheng 34773fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 34783fd0a558SYan, Zheng if (ret) { 34793fd0a558SYan, Zheng err = ret; 34803fd0a558SYan, Zheng goto out_free; 34812423fdfbSJiri Slaby } 34825d4f98a2SYan Zheng 34835d4f98a2SYan Zheng while (1) { 34840647bf56SWang Shilong rc->reserved_bytes = 0; 34850647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 34860647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 34870647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 34880647bf56SWang Shilong if (ret) { 34890647bf56SWang Shilong err = ret; 34900647bf56SWang Shilong break; 34910647bf56SWang Shilong } 3492c87f08caSChris Mason progress++; 3493a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 34940f788c58SLiu Bo if (IS_ERR(trans)) { 34950f788c58SLiu Bo err = PTR_ERR(trans); 34960f788c58SLiu Bo trans = NULL; 34970f788c58SLiu Bo break; 34980f788c58SLiu Bo } 3499c87f08caSChris Mason restart: 35003fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 35013a45bb20SJeff Mahoney btrfs_end_transaction(trans); 350242a657f5SPan Bian trans = NULL; 35033fd0a558SYan, Zheng continue; 35043fd0a558SYan, Zheng } 35053fd0a558SYan, Zheng 3506147d256eSZhaolei ret = find_next_extent(rc, path, &key); 35075d4f98a2SYan Zheng if (ret < 0) 35085d4f98a2SYan Zheng err = ret; 35095d4f98a2SYan Zheng if (ret != 0) 35105d4f98a2SYan Zheng break; 35115d4f98a2SYan Zheng 35125d4f98a2SYan Zheng rc->extents_found++; 35135d4f98a2SYan Zheng 35145d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 35155d4f98a2SYan Zheng struct btrfs_extent_item); 35163fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 35175d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 35185d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 35195d4f98a2SYan Zheng ret = check_extent_flags(flags); 35205d4f98a2SYan Zheng BUG_ON(ret); 35216d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3522ba3c2b19SNikolay Borisov err = -EINVAL; 3523ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 3524ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 3525ba3c2b19SNikolay Borisov break; 35265d4f98a2SYan Zheng } else { 35275d4f98a2SYan Zheng BUG(); 35285d4f98a2SYan Zheng } 35295d4f98a2SYan Zheng 35305d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 35315d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 35325d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 35335d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 35345d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 35355d4f98a2SYan Zheng } else { 3536b3b4aa74SDavid Sterba btrfs_release_path(path); 35375d4f98a2SYan Zheng ret = 0; 35385d4f98a2SYan Zheng } 35395d4f98a2SYan Zheng if (ret < 0) { 35403fd0a558SYan, Zheng err = ret; 35415d4f98a2SYan Zheng break; 35425d4f98a2SYan Zheng } 35435d4f98a2SYan Zheng 35445d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 35455d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 35465d4f98a2SYan Zheng if (ret < 0) { 35473fd0a558SYan, Zheng if (ret != -EAGAIN) { 35485d4f98a2SYan Zheng err = ret; 35495d4f98a2SYan Zheng break; 35505d4f98a2SYan Zheng } 35513fd0a558SYan, Zheng rc->extents_found--; 35523fd0a558SYan, Zheng rc->search_start = key.objectid; 35533fd0a558SYan, Zheng } 35545d4f98a2SYan Zheng } 35555d4f98a2SYan Zheng 35563a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 35572ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 35583fd0a558SYan, Zheng trans = NULL; 35595d4f98a2SYan Zheng 35605d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 35615d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 35625d4f98a2SYan Zheng rc->found_file_extent = 1; 35630257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 35643fd0a558SYan, Zheng &key, &rc->cluster); 35655d4f98a2SYan Zheng if (ret < 0) { 35665d4f98a2SYan Zheng err = ret; 35675d4f98a2SYan Zheng break; 35685d4f98a2SYan Zheng } 35695d4f98a2SYan Zheng } 3570f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 3571f31ea088SQu Wenruo err = -ECANCELED; 3572f31ea088SQu Wenruo break; 3573f31ea088SQu Wenruo } 35745d4f98a2SYan Zheng } 3575c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 357643a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 35779689457bSShilong Wang if (ret == 1) { 3578c87f08caSChris Mason err = 0; 3579c87f08caSChris Mason progress = 0; 3580c87f08caSChris Mason goto restart; 3581c87f08caSChris Mason } 3582c87f08caSChris Mason } 35833fd0a558SYan, Zheng 3584b3b4aa74SDavid Sterba btrfs_release_path(path); 358591166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 35865d4f98a2SYan Zheng 35875d4f98a2SYan Zheng if (trans) { 35883a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 35892ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 35905d4f98a2SYan Zheng } 35915d4f98a2SYan Zheng 35920257bb82SYan, Zheng if (!err) { 35933fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 35943fd0a558SYan, Zheng &rc->cluster); 35950257bb82SYan, Zheng if (ret < 0) 35960257bb82SYan, Zheng err = ret; 35970257bb82SYan, Zheng } 35980257bb82SYan, Zheng 35993fd0a558SYan, Zheng rc->create_reloc_tree = 0; 36003fd0a558SYan, Zheng set_reloc_control(rc); 36010257bb82SYan, Zheng 360213fe1bdbSQu Wenruo btrfs_backref_release_cache(&rc->backref_cache); 360363f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 36045d4f98a2SYan Zheng 36057f913c7cSQu Wenruo /* 36067f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 36077f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 36087f913c7cSQu Wenruo * 36097f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 36107f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 36117f913c7cSQu Wenruo * merge_reloc_roots() 36127f913c7cSQu Wenruo */ 36133fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 36145d4f98a2SYan Zheng 36155d4f98a2SYan Zheng merge_reloc_roots(rc); 36165d4f98a2SYan Zheng 36173fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 36185d4f98a2SYan Zheng unset_reloc_control(rc); 361963f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 36205d4f98a2SYan Zheng 36215d4f98a2SYan Zheng /* get rid of pinned extents */ 36227a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 362362b99540SQu Wenruo if (IS_ERR(trans)) { 36243612b495STsutomu Itoh err = PTR_ERR(trans); 362562b99540SQu Wenruo goto out_free; 362662b99540SQu Wenruo } 36273a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 36286217b0faSJosef Bacik out_free: 3629d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3630d2311e69SQu Wenruo if (ret < 0 && !err) 3631d2311e69SQu Wenruo err = ret; 36322ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 36333fd0a558SYan, Zheng btrfs_free_path(path); 36345d4f98a2SYan Zheng return err; 36355d4f98a2SYan Zheng } 36365d4f98a2SYan Zheng 36375d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 36380257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 36395d4f98a2SYan Zheng { 36405d4f98a2SYan Zheng struct btrfs_path *path; 36415d4f98a2SYan Zheng struct btrfs_inode_item *item; 36425d4f98a2SYan Zheng struct extent_buffer *leaf; 364332430c61SNaohiro Aota u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC; 36445d4f98a2SYan Zheng int ret; 36455d4f98a2SYan Zheng 364632430c61SNaohiro Aota if (btrfs_is_zoned(trans->fs_info)) 364732430c61SNaohiro Aota flags &= ~BTRFS_INODE_PREALLOC; 364832430c61SNaohiro Aota 36495d4f98a2SYan Zheng path = btrfs_alloc_path(); 36505d4f98a2SYan Zheng if (!path) 36515d4f98a2SYan Zheng return -ENOMEM; 36525d4f98a2SYan Zheng 36535d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 36545d4f98a2SYan Zheng if (ret) 36555d4f98a2SYan Zheng goto out; 36565d4f98a2SYan Zheng 36575d4f98a2SYan Zheng leaf = path->nodes[0]; 36585d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 3659b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 36605d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 36610257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 36625d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 366332430c61SNaohiro Aota btrfs_set_inode_flags(leaf, item, flags); 36645d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 36655d4f98a2SYan Zheng out: 36665d4f98a2SYan Zheng btrfs_free_path(path); 36675d4f98a2SYan Zheng return ret; 36685d4f98a2SYan Zheng } 36695d4f98a2SYan Zheng 36705d4f98a2SYan Zheng /* 36715d4f98a2SYan Zheng * helper to create inode for data relocation. 36725d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 36735d4f98a2SYan Zheng */ 36743fd0a558SYan, Zheng static noinline_for_stack 36753fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 367632da5386SDavid Sterba struct btrfs_block_group *group) 36775d4f98a2SYan Zheng { 36785d4f98a2SYan Zheng struct inode *inode = NULL; 36795d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 36805d4f98a2SYan Zheng struct btrfs_root *root; 36814624900dSZhaolei u64 objectid; 36825d4f98a2SYan Zheng int err = 0; 36835d4f98a2SYan Zheng 3684aeb935a4SQu Wenruo root = btrfs_grab_root(fs_info->data_reloc_root); 3685a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 368676deacf0SJosef Bacik if (IS_ERR(trans)) { 368700246528SJosef Bacik btrfs_put_root(root); 36883fd0a558SYan, Zheng return ERR_CAST(trans); 368976deacf0SJosef Bacik } 36905d4f98a2SYan Zheng 3691543068a2SNikolay Borisov err = btrfs_get_free_objectid(root, &objectid); 36925d4f98a2SYan Zheng if (err) 36935d4f98a2SYan Zheng goto out; 36945d4f98a2SYan Zheng 36950257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 36965d4f98a2SYan Zheng BUG_ON(err); 36975d4f98a2SYan Zheng 36980202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, objectid, root); 36992e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 3700b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 37015d4f98a2SYan Zheng 370273f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 37035d4f98a2SYan Zheng out: 370400246528SJosef Bacik btrfs_put_root(root); 37053a45bb20SJeff Mahoney btrfs_end_transaction(trans); 37062ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 37075d4f98a2SYan Zheng if (err) { 37085d4f98a2SYan Zheng if (inode) 37095d4f98a2SYan Zheng iput(inode); 37105d4f98a2SYan Zheng inode = ERR_PTR(err); 37115d4f98a2SYan Zheng } 37125d4f98a2SYan Zheng return inode; 37135d4f98a2SYan Zheng } 37145d4f98a2SYan Zheng 3715c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 37163fd0a558SYan, Zheng { 37173fd0a558SYan, Zheng struct reloc_control *rc; 37183fd0a558SYan, Zheng 37193fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 37203fd0a558SYan, Zheng if (!rc) 37213fd0a558SYan, Zheng return NULL; 37223fd0a558SYan, Zheng 37233fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 3724d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 3725584fb121SQu Wenruo btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1); 37263fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 372743eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 372843eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 37293fd0a558SYan, Zheng return rc; 37303fd0a558SYan, Zheng } 37313fd0a558SYan, Zheng 37321a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 37331a0afa0eSJosef Bacik { 37341a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 37351a0afa0eSJosef Bacik 37361a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 37371a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 37381a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 37391a0afa0eSJosef Bacik kfree(node); 37401a0afa0eSJosef Bacik 37411a0afa0eSJosef Bacik kfree(rc); 37421a0afa0eSJosef Bacik } 37431a0afa0eSJosef Bacik 37445d4f98a2SYan Zheng /* 3745ebce0e01SAdam Borowski * Print the block group being relocated 3746ebce0e01SAdam Borowski */ 3747ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 374832da5386SDavid Sterba struct btrfs_block_group *block_group) 3749ebce0e01SAdam Borowski { 3750f89e09cfSAnand Jain char buf[128] = {'\0'}; 3751ebce0e01SAdam Borowski 3752f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 3753ebce0e01SAdam Borowski 3754ebce0e01SAdam Borowski btrfs_info(fs_info, 3755ebce0e01SAdam Borowski "relocating block group %llu flags %s", 3756b3470b5dSDavid Sterba block_group->start, buf); 3757ebce0e01SAdam Borowski } 3758ebce0e01SAdam Borowski 3759430640e3SQu Wenruo static const char *stage_to_string(int stage) 3760430640e3SQu Wenruo { 3761430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 3762430640e3SQu Wenruo return "move data extents"; 3763430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 3764430640e3SQu Wenruo return "update data pointers"; 3765430640e3SQu Wenruo return "unknown"; 3766430640e3SQu Wenruo } 3767430640e3SQu Wenruo 3768ebce0e01SAdam Borowski /* 37695d4f98a2SYan Zheng * function to relocate all extents in a block group. 37705d4f98a2SYan Zheng */ 37716bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 37725d4f98a2SYan Zheng { 377332da5386SDavid Sterba struct btrfs_block_group *bg; 37746bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 37755d4f98a2SYan Zheng struct reloc_control *rc; 37760af3d00bSJosef Bacik struct inode *inode; 37770af3d00bSJosef Bacik struct btrfs_path *path; 37785d4f98a2SYan Zheng int ret; 3779f0486c68SYan, Zheng int rw = 0; 37805d4f98a2SYan Zheng int err = 0; 37815d4f98a2SYan Zheng 3782eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 3783eede2bf3SOmar Sandoval if (!bg) 3784eede2bf3SOmar Sandoval return -ENOENT; 3785eede2bf3SOmar Sandoval 3786eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 3787eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 3788eede2bf3SOmar Sandoval return -ETXTBSY; 3789eede2bf3SOmar Sandoval } 3790eede2bf3SOmar Sandoval 3791c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 3792eede2bf3SOmar Sandoval if (!rc) { 3793eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 37945d4f98a2SYan Zheng return -ENOMEM; 3795eede2bf3SOmar Sandoval } 37965d4f98a2SYan Zheng 3797f0486c68SYan, Zheng rc->extent_root = extent_root; 3798eede2bf3SOmar Sandoval rc->block_group = bg; 37995d4f98a2SYan Zheng 3800b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 3801f0486c68SYan, Zheng if (ret) { 3802f0486c68SYan, Zheng err = ret; 3803f0486c68SYan, Zheng goto out; 3804f0486c68SYan, Zheng } 3805f0486c68SYan, Zheng rw = 1; 3806f0486c68SYan, Zheng 38070af3d00bSJosef Bacik path = btrfs_alloc_path(); 38080af3d00bSJosef Bacik if (!path) { 38090af3d00bSJosef Bacik err = -ENOMEM; 38100af3d00bSJosef Bacik goto out; 38110af3d00bSJosef Bacik } 38120af3d00bSJosef Bacik 38137949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 38140af3d00bSJosef Bacik btrfs_free_path(path); 38150af3d00bSJosef Bacik 38160af3d00bSJosef Bacik if (!IS_ERR(inode)) 38171bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 38180af3d00bSJosef Bacik else 38190af3d00bSJosef Bacik ret = PTR_ERR(inode); 38200af3d00bSJosef Bacik 38210af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 38220af3d00bSJosef Bacik err = ret; 38230af3d00bSJosef Bacik goto out; 38240af3d00bSJosef Bacik } 38250af3d00bSJosef Bacik 38265d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 38275d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 38285d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 38295d4f98a2SYan Zheng rc->data_inode = NULL; 38305d4f98a2SYan Zheng goto out; 38315d4f98a2SYan Zheng } 38325d4f98a2SYan Zheng 38330b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 38345d4f98a2SYan Zheng 38359cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 3836f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 38376374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 3838b3470b5dSDavid Sterba rc->block_group->start, 3839b3470b5dSDavid Sterba rc->block_group->length); 38405d4f98a2SYan Zheng 38415d4f98a2SYan Zheng while (1) { 3842430640e3SQu Wenruo int finishes_stage; 3843430640e3SQu Wenruo 384476dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 38455d4f98a2SYan Zheng ret = relocate_block_group(rc); 384676dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 3847ff612ba7SJosef Bacik if (ret < 0) 38485d4f98a2SYan Zheng err = ret; 3849ff612ba7SJosef Bacik 3850430640e3SQu Wenruo finishes_stage = rc->stage; 3851ff612ba7SJosef Bacik /* 3852ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 3853ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 3854ff612ba7SJosef Bacik * different block group we could end up hitting the 3855ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 3856ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 3857ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 3858ff612ba7SJosef Bacik * out of the loop if we hit an error. 3859ff612ba7SJosef Bacik */ 3860ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 3861ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 3862ff612ba7SJosef Bacik (u64)-1); 3863ff612ba7SJosef Bacik if (ret) 3864ff612ba7SJosef Bacik err = ret; 3865ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 3866ff612ba7SJosef Bacik 0, -1); 3867ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 38685d4f98a2SYan Zheng } 38695d4f98a2SYan Zheng 3870ff612ba7SJosef Bacik if (err < 0) 3871ff612ba7SJosef Bacik goto out; 3872ff612ba7SJosef Bacik 38735d4f98a2SYan Zheng if (rc->extents_found == 0) 38745d4f98a2SYan Zheng break; 38755d4f98a2SYan Zheng 3876430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 3877430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 38785d4f98a2SYan Zheng } 38795d4f98a2SYan Zheng 38805d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 38815d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 3882bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 38835d4f98a2SYan Zheng out: 3884f0486c68SYan, Zheng if (err && rw) 38852ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 38865d4f98a2SYan Zheng iput(rc->data_inode); 38875d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 38881a0afa0eSJosef Bacik free_reloc_control(rc); 38895d4f98a2SYan Zheng return err; 38905d4f98a2SYan Zheng } 38915d4f98a2SYan Zheng 389276dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 389376dda93cSYan, Zheng { 38940b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 389576dda93cSYan, Zheng struct btrfs_trans_handle *trans; 389679787eaaSJeff Mahoney int ret, err; 389776dda93cSYan, Zheng 38980b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 389979787eaaSJeff Mahoney if (IS_ERR(trans)) 390079787eaaSJeff Mahoney return PTR_ERR(trans); 390176dda93cSYan, Zheng 390276dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 390376dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 3904c8422684SDavid Sterba btrfs_set_root_drop_level(&root->root_item, 0); 390576dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 39060b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 390776dda93cSYan, Zheng &root->root_key, &root->root_item); 390876dda93cSYan, Zheng 39093a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 391079787eaaSJeff Mahoney if (err) 391179787eaaSJeff Mahoney return err; 391279787eaaSJeff Mahoney return ret; 391376dda93cSYan, Zheng } 391476dda93cSYan, Zheng 39155d4f98a2SYan Zheng /* 39165d4f98a2SYan Zheng * recover relocation interrupted by system crash. 39175d4f98a2SYan Zheng * 39185d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 39195d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 39205d4f98a2SYan Zheng */ 39215d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 39225d4f98a2SYan Zheng { 39230b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 39245d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 39255d4f98a2SYan Zheng struct btrfs_key key; 39265d4f98a2SYan Zheng struct btrfs_root *fs_root; 39275d4f98a2SYan Zheng struct btrfs_root *reloc_root; 39285d4f98a2SYan Zheng struct btrfs_path *path; 39295d4f98a2SYan Zheng struct extent_buffer *leaf; 39305d4f98a2SYan Zheng struct reloc_control *rc = NULL; 39315d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 39325d4f98a2SYan Zheng int ret; 39335d4f98a2SYan Zheng int err = 0; 39345d4f98a2SYan Zheng 39355d4f98a2SYan Zheng path = btrfs_alloc_path(); 39365d4f98a2SYan Zheng if (!path) 39375d4f98a2SYan Zheng return -ENOMEM; 3938e4058b54SDavid Sterba path->reada = READA_BACK; 39395d4f98a2SYan Zheng 39405d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 39415d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 39425d4f98a2SYan Zheng key.offset = (u64)-1; 39435d4f98a2SYan Zheng 39445d4f98a2SYan Zheng while (1) { 39450b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 39465d4f98a2SYan Zheng path, 0, 0); 39475d4f98a2SYan Zheng if (ret < 0) { 39485d4f98a2SYan Zheng err = ret; 39495d4f98a2SYan Zheng goto out; 39505d4f98a2SYan Zheng } 39515d4f98a2SYan Zheng if (ret > 0) { 39525d4f98a2SYan Zheng if (path->slots[0] == 0) 39535d4f98a2SYan Zheng break; 39545d4f98a2SYan Zheng path->slots[0]--; 39555d4f98a2SYan Zheng } 39565d4f98a2SYan Zheng leaf = path->nodes[0]; 39575d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 3958b3b4aa74SDavid Sterba btrfs_release_path(path); 39595d4f98a2SYan Zheng 39605d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 39615d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 39625d4f98a2SYan Zheng break; 39635d4f98a2SYan Zheng 39643dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 39655d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 39665d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 39675d4f98a2SYan Zheng goto out; 39685d4f98a2SYan Zheng } 39695d4f98a2SYan Zheng 397092a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 39715d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 39725d4f98a2SYan Zheng 39735d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 3974a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, 3975a820feb5SDavid Sterba reloc_root->root_key.offset, false); 39765d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 397776dda93cSYan, Zheng ret = PTR_ERR(fs_root); 397876dda93cSYan, Zheng if (ret != -ENOENT) { 397976dda93cSYan, Zheng err = ret; 39805d4f98a2SYan Zheng goto out; 39815d4f98a2SYan Zheng } 398279787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 398379787eaaSJeff Mahoney if (ret < 0) { 398479787eaaSJeff Mahoney err = ret; 398579787eaaSJeff Mahoney goto out; 398679787eaaSJeff Mahoney } 3987932fd26dSJosef Bacik } else { 398800246528SJosef Bacik btrfs_put_root(fs_root); 398976dda93cSYan, Zheng } 39905d4f98a2SYan Zheng } 39915d4f98a2SYan Zheng 39925d4f98a2SYan Zheng if (key.offset == 0) 39935d4f98a2SYan Zheng break; 39945d4f98a2SYan Zheng 39955d4f98a2SYan Zheng key.offset--; 39965d4f98a2SYan Zheng } 3997b3b4aa74SDavid Sterba btrfs_release_path(path); 39985d4f98a2SYan Zheng 39995d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 40005d4f98a2SYan Zheng goto out; 40015d4f98a2SYan Zheng 4002c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 40035d4f98a2SYan Zheng if (!rc) { 40045d4f98a2SYan Zheng err = -ENOMEM; 40055d4f98a2SYan Zheng goto out; 40065d4f98a2SYan Zheng } 40075d4f98a2SYan Zheng 40080b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 40095d4f98a2SYan Zheng 40105d4f98a2SYan Zheng set_reloc_control(rc); 40115d4f98a2SYan Zheng 40127a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 40133612b495STsutomu Itoh if (IS_ERR(trans)) { 40143612b495STsutomu Itoh err = PTR_ERR(trans); 4015fb2d83eeSJosef Bacik goto out_unset; 40163612b495STsutomu Itoh } 40173fd0a558SYan, Zheng 40183fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 40193fd0a558SYan, Zheng 40205d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 40215d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 40225d4f98a2SYan Zheng struct btrfs_root, root_list); 40235d4f98a2SYan Zheng list_del(&reloc_root->root_list); 40245d4f98a2SYan Zheng 40255d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 40265d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 40275d4f98a2SYan Zheng &rc->reloc_roots); 40285d4f98a2SYan Zheng continue; 40295d4f98a2SYan Zheng } 40305d4f98a2SYan Zheng 4031a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 4032a820feb5SDavid Sterba false); 403379787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 403479787eaaSJeff Mahoney err = PTR_ERR(fs_root); 4035ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 40361402d17dSXiyu Yang btrfs_end_transaction(trans); 4037fb2d83eeSJosef Bacik goto out_unset; 403879787eaaSJeff Mahoney } 40395d4f98a2SYan Zheng 4040ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 404179787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 4042f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 404300246528SJosef Bacik btrfs_put_root(fs_root); 40445d4f98a2SYan Zheng } 40455d4f98a2SYan Zheng 40463a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 404779787eaaSJeff Mahoney if (err) 4048fb2d83eeSJosef Bacik goto out_unset; 40495d4f98a2SYan Zheng 40505d4f98a2SYan Zheng merge_reloc_roots(rc); 40515d4f98a2SYan Zheng 40525d4f98a2SYan Zheng unset_reloc_control(rc); 40535d4f98a2SYan Zheng 40547a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 405562b99540SQu Wenruo if (IS_ERR(trans)) { 40563612b495STsutomu Itoh err = PTR_ERR(trans); 40576217b0faSJosef Bacik goto out_clean; 405862b99540SQu Wenruo } 40593a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 40606217b0faSJosef Bacik out_clean: 4061d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4062d2311e69SQu Wenruo if (ret < 0 && !err) 4063d2311e69SQu Wenruo err = ret; 4064fb2d83eeSJosef Bacik out_unset: 4065fb2d83eeSJosef Bacik unset_reloc_control(rc); 40661a0afa0eSJosef Bacik free_reloc_control(rc); 40673612b495STsutomu Itoh out: 4068aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4069aca1bba6SLiu Bo 40705d4f98a2SYan Zheng btrfs_free_path(path); 40715d4f98a2SYan Zheng 40725d4f98a2SYan Zheng if (err == 0) { 40735d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 4074aeb935a4SQu Wenruo fs_root = btrfs_grab_root(fs_info->data_reloc_root); 4075aeb935a4SQu Wenruo ASSERT(fs_root); 407666b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 407700246528SJosef Bacik btrfs_put_root(fs_root); 4078932fd26dSJosef Bacik } 40795d4f98a2SYan Zheng return err; 40805d4f98a2SYan Zheng } 40815d4f98a2SYan Zheng 40825d4f98a2SYan Zheng /* 40835d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 40845d4f98a2SYan Zheng * 40855d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 40865d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 40875d4f98a2SYan Zheng */ 40887bfa9535SNikolay Borisov int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len) 40895d4f98a2SYan Zheng { 40907bfa9535SNikolay Borisov struct btrfs_fs_info *fs_info = inode->root->fs_info; 40915d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 40925d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 40935d4f98a2SYan Zheng int ret; 40945d4f98a2SYan Zheng u64 disk_bytenr; 40954577b014SJosef Bacik u64 new_bytenr; 40965d4f98a2SYan Zheng LIST_HEAD(list); 40975d4f98a2SYan Zheng 40987bfa9535SNikolay Borisov ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4099bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 41005d4f98a2SYan Zheng 41017bfa9535SNikolay Borisov disk_bytenr = file_pos + inode->index_cnt; 41020b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4103a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 410479787eaaSJeff Mahoney if (ret) 410579787eaaSJeff Mahoney goto out; 41065d4f98a2SYan Zheng 41075d4f98a2SYan Zheng while (!list_empty(&list)) { 41085d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 41095d4f98a2SYan Zheng list_del_init(&sums->list); 41105d4f98a2SYan Zheng 41114577b014SJosef Bacik /* 41124577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 41134577b014SJosef Bacik * We need to do this because we will read in entire prealloc 41144577b014SJosef Bacik * extents but we may have written to say the middle of the 41154577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 41164577b014SJosef Bacik * the right disk offset. 41174577b014SJosef Bacik * 41184577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 41194577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 41204577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 41214577b014SJosef Bacik * disk length. 41224577b014SJosef Bacik */ 4123bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 41244577b014SJosef Bacik sums->bytenr = new_bytenr; 41255d4f98a2SYan Zheng 4126f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 41275d4f98a2SYan Zheng } 412879787eaaSJeff Mahoney out: 41295d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4130411fc6bcSAndi Kleen return ret; 41315d4f98a2SYan Zheng } 41323fd0a558SYan, Zheng 413383d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 41343fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 41353fd0a558SYan, Zheng struct extent_buffer *cow) 41363fd0a558SYan, Zheng { 41370b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 41383fd0a558SYan, Zheng struct reloc_control *rc; 4139a26195a5SQu Wenruo struct btrfs_backref_node *node; 41403fd0a558SYan, Zheng int first_cow = 0; 41413fd0a558SYan, Zheng int level; 414283d4cfd4SJosef Bacik int ret = 0; 41433fd0a558SYan, Zheng 41440b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 41453fd0a558SYan, Zheng if (!rc) 414683d4cfd4SJosef Bacik return 0; 41473fd0a558SYan, Zheng 41483fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 41493fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 41503fd0a558SYan, Zheng 41513fd0a558SYan, Zheng level = btrfs_header_level(buf); 41523fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 41533fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 41543fd0a558SYan, Zheng first_cow = 1; 41553fd0a558SYan, Zheng 41563fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 41573fd0a558SYan, Zheng rc->create_reloc_tree) { 41583fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 41593fd0a558SYan, Zheng 41603fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 41613fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 41623fd0a558SYan, Zheng node->new_bytenr != buf->start); 41633fd0a558SYan, Zheng 4164b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 416567439dadSDavid Sterba atomic_inc(&cow->refs); 41663fd0a558SYan, Zheng node->eb = cow; 41673fd0a558SYan, Zheng node->new_bytenr = cow->start; 41683fd0a558SYan, Zheng 41693fd0a558SYan, Zheng if (!node->pending) { 41703fd0a558SYan, Zheng list_move_tail(&node->list, 41713fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 41723fd0a558SYan, Zheng node->pending = 1; 41733fd0a558SYan, Zheng } 41743fd0a558SYan, Zheng 41753fd0a558SYan, Zheng if (first_cow) 41769569cc20SQu Wenruo mark_block_processed(rc, node); 41773fd0a558SYan, Zheng 41783fd0a558SYan, Zheng if (first_cow && level > 0) 41793fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 41803fd0a558SYan, Zheng } 41813fd0a558SYan, Zheng 418283d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 41833fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 418483d4cfd4SJosef Bacik return ret; 41853fd0a558SYan, Zheng } 41863fd0a558SYan, Zheng 41873fd0a558SYan, Zheng /* 41883fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 418901327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 41903fd0a558SYan, Zheng */ 4191147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 41923fd0a558SYan, Zheng u64 *bytes_to_reserve) 41933fd0a558SYan, Zheng { 419410995c04SQu Wenruo struct btrfs_root *root = pending->root; 419510995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 41963fd0a558SYan, Zheng 41976282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 41983fd0a558SYan, Zheng return; 41993fd0a558SYan, Zheng 42003fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 42013fd0a558SYan, Zheng return; 42023fd0a558SYan, Zheng 42033fd0a558SYan, Zheng root = root->reloc_root; 42043fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 42053fd0a558SYan, Zheng /* 42063fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 42073fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 42083fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 42093fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 42103fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 42113fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 42123fd0a558SYan, Zheng * tree may use more space than it frees. so we need 42133fd0a558SYan, Zheng * reserve extra space. 42143fd0a558SYan, Zheng */ 42153fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 42163fd0a558SYan, Zheng } 42173fd0a558SYan, Zheng 42183fd0a558SYan, Zheng /* 42193fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 42203fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4221f44deb74SJosef Bacik * 4222f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4223f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4224f44deb74SJosef Bacik * rc->reloc_roots. 42253fd0a558SYan, Zheng */ 422649b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 42273fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 42283fd0a558SYan, Zheng { 42293fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 42303fd0a558SYan, Zheng struct btrfs_root *reloc_root; 42313fd0a558SYan, Zheng struct btrfs_root *new_root; 423210995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 42333fd0a558SYan, Zheng int ret; 42343fd0a558SYan, Zheng 42356282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 423649b25e05SJeff Mahoney return 0; 42373fd0a558SYan, Zheng 42383fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 42393fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 42403fd0a558SYan, Zheng 42413fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 42423fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 42433fd0a558SYan, Zheng rc->block_rsv, 42443a584174SLu Fengqi rc->nodes_relocated, true); 424549b25e05SJeff Mahoney if (ret) 424649b25e05SJeff Mahoney return ret; 42473fd0a558SYan, Zheng } 42483fd0a558SYan, Zheng 42493fd0a558SYan, Zheng new_root = pending->snap; 42503fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 42513fd0a558SYan, Zheng new_root->root_key.objectid); 425249b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 425349b25e05SJeff Mahoney return PTR_ERR(reloc_root); 42543fd0a558SYan, Zheng 4255ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4256ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4257f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 42583fd0a558SYan, Zheng 425949b25e05SJeff Mahoney if (rc->create_reloc_tree) 42603fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 426149b25e05SJeff Mahoney return ret; 42623fd0a558SYan, Zheng } 4263