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; 7365d4f98a2SYan Zheng int ret; 7375d4f98a2SYan Zheng 7385d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 7395d4f98a2SYan Zheng BUG_ON(!root_item); 7405d4f98a2SYan Zheng 7415d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 7425d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 7433fd0a558SYan, Zheng root_key.offset = objectid; 7445d4f98a2SYan Zheng 7453fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 746054570a1SFilipe Manana u64 commit_root_gen; 747054570a1SFilipe Manana 7483fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 7495d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 7505d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 7515d4f98a2SYan Zheng BUG_ON(ret); 752054570a1SFilipe Manana /* 753054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 754054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 755054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 756054570a1SFilipe Manana * either inside the critical section of a transaction commit 757054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 758054570a1SFilipe Manana * it's created before the transaction commit is started. 759054570a1SFilipe Manana */ 760054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 761054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 7623fd0a558SYan, Zheng } else { 7633fd0a558SYan, Zheng /* 7643fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 7653fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 7663fd0a558SYan, Zheng * modified after it was created have RELOC flag 7673fd0a558SYan, Zheng * set in their headers. so it's OK to not update 7683fd0a558SYan, Zheng * the 'last_snapshot'. 7693fd0a558SYan, Zheng */ 7703fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 7713fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 7723fd0a558SYan, Zheng BUG_ON(ret); 7733fd0a558SYan, Zheng } 7743fd0a558SYan, Zheng 7755d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 7765d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 7775d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 7785d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 7793fd0a558SYan, Zheng 7803fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 7813fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 7823fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 7833fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 784c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, 0); 7853fd0a558SYan, Zheng } 7865d4f98a2SYan Zheng 7875d4f98a2SYan Zheng btrfs_tree_unlock(eb); 7885d4f98a2SYan Zheng free_extent_buffer(eb); 7895d4f98a2SYan Zheng 7900b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 7915d4f98a2SYan Zheng &root_key, root_item); 7925d4f98a2SYan Zheng BUG_ON(ret); 7935d4f98a2SYan Zheng kfree(root_item); 7945d4f98a2SYan Zheng 7953dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 7965d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 79792a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 7985d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 7993fd0a558SYan, Zheng return reloc_root; 8003fd0a558SYan, Zheng } 8013fd0a558SYan, Zheng 8023fd0a558SYan, Zheng /* 8033fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 8043fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 805f44deb74SJosef Bacik * 806f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 807f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 8083fd0a558SYan, Zheng */ 8093fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 8103fd0a558SYan, Zheng struct btrfs_root *root) 8113fd0a558SYan, Zheng { 8120b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 8133fd0a558SYan, Zheng struct btrfs_root *reloc_root; 8140b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 81520dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 8163fd0a558SYan, Zheng int clear_rsv = 0; 817ffd7b339SJeff Mahoney int ret; 8183fd0a558SYan, Zheng 819aec7db3bSJosef Bacik if (!rc) 8202abc726aSJosef Bacik return 0; 8212abc726aSJosef Bacik 8221fac4a54SQu Wenruo /* 8231fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 8241fac4a54SQu Wenruo * create/update the dead reloc tree 8251fac4a54SQu Wenruo */ 8266282675eSQu Wenruo if (reloc_root_is_dead(root)) 8271fac4a54SQu Wenruo return 0; 8281fac4a54SQu Wenruo 829aec7db3bSJosef Bacik /* 830aec7db3bSJosef Bacik * This is subtle but important. We do not do 831aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 832aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 833aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 834aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 835aec7db3bSJosef Bacik * in. 836aec7db3bSJosef Bacik */ 8373fd0a558SYan, Zheng if (root->reloc_root) { 8383fd0a558SYan, Zheng reloc_root = root->reloc_root; 8393fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 8403fd0a558SYan, Zheng return 0; 8413fd0a558SYan, Zheng } 8423fd0a558SYan, Zheng 843aec7db3bSJosef Bacik /* 844aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 845aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 846aec7db3bSJosef Bacik */ 847aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 848aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 849aec7db3bSJosef Bacik return 0; 850aec7db3bSJosef Bacik 85120dd2cbfSMiao Xie if (!trans->reloc_reserved) { 85220dd2cbfSMiao Xie rsv = trans->block_rsv; 8533fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 8543fd0a558SYan, Zheng clear_rsv = 1; 8553fd0a558SYan, Zheng } 8563fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 8573fd0a558SYan, Zheng if (clear_rsv) 85820dd2cbfSMiao Xie trans->block_rsv = rsv; 859*00bb36a0SJosef Bacik if (IS_ERR(reloc_root)) 860*00bb36a0SJosef Bacik return PTR_ERR(reloc_root); 8615d4f98a2SYan Zheng 862ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 863*00bb36a0SJosef Bacik if (ret) { 864*00bb36a0SJosef Bacik /* Pairs with create_reloc_root */ 865*00bb36a0SJosef Bacik btrfs_put_root(reloc_root); 866*00bb36a0SJosef Bacik return ret; 867*00bb36a0SJosef Bacik } 868f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 8695d4f98a2SYan Zheng return 0; 8705d4f98a2SYan Zheng } 8715d4f98a2SYan Zheng 8725d4f98a2SYan Zheng /* 8735d4f98a2SYan Zheng * update root item of reloc tree 8745d4f98a2SYan Zheng */ 8755d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 8765d4f98a2SYan Zheng struct btrfs_root *root) 8775d4f98a2SYan Zheng { 8780b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 8795d4f98a2SYan Zheng struct btrfs_root *reloc_root; 8805d4f98a2SYan Zheng struct btrfs_root_item *root_item; 8815d4f98a2SYan Zheng int ret; 8825d4f98a2SYan Zheng 8836282675eSQu Wenruo if (!have_reloc_root(root)) 8847585717fSChris Mason goto out; 8855d4f98a2SYan Zheng 8865d4f98a2SYan Zheng reloc_root = root->reloc_root; 8875d4f98a2SYan Zheng root_item = &reloc_root->root_item; 8885d4f98a2SYan Zheng 889f44deb74SJosef Bacik /* 890f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 891f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 892f44deb74SJosef Bacik * hold it while we update the reloc root. 893f44deb74SJosef Bacik */ 894f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 895f44deb74SJosef Bacik 896d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 8970b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 8983fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 899d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 9006282675eSQu Wenruo /* 9016282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 9026282675eSQu Wenruo * have_reloc_root will not touch it from now on. 9036282675eSQu Wenruo */ 9046282675eSQu Wenruo smp_wmb(); 905c974c464SWang Shilong __del_reloc_root(reloc_root); 9065d4f98a2SYan Zheng } 9075d4f98a2SYan Zheng 9085d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 909ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 9105d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 9115d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 9125d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 9135d4f98a2SYan Zheng } 9145d4f98a2SYan Zheng 9150b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 9165d4f98a2SYan Zheng &reloc_root->root_key, root_item); 9175d4f98a2SYan Zheng BUG_ON(ret); 918f44deb74SJosef Bacik btrfs_put_root(reloc_root); 9197585717fSChris Mason out: 9205d4f98a2SYan Zheng return 0; 9215d4f98a2SYan Zheng } 9225d4f98a2SYan Zheng 9235d4f98a2SYan Zheng /* 9245d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 9255d4f98a2SYan Zheng * in a subvolume 9265d4f98a2SYan Zheng */ 9275d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 9285d4f98a2SYan Zheng { 9295d4f98a2SYan Zheng struct rb_node *node; 9305d4f98a2SYan Zheng struct rb_node *prev; 9315d4f98a2SYan Zheng struct btrfs_inode *entry; 9325d4f98a2SYan Zheng struct inode *inode; 9335d4f98a2SYan Zheng 9345d4f98a2SYan Zheng spin_lock(&root->inode_lock); 9355d4f98a2SYan Zheng again: 9365d4f98a2SYan Zheng node = root->inode_tree.rb_node; 9375d4f98a2SYan Zheng prev = NULL; 9385d4f98a2SYan Zheng while (node) { 9395d4f98a2SYan Zheng prev = node; 9405d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9415d4f98a2SYan Zheng 9424a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 9435d4f98a2SYan Zheng node = node->rb_left; 9444a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 9455d4f98a2SYan Zheng node = node->rb_right; 9465d4f98a2SYan Zheng else 9475d4f98a2SYan Zheng break; 9485d4f98a2SYan Zheng } 9495d4f98a2SYan Zheng if (!node) { 9505d4f98a2SYan Zheng while (prev) { 9515d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 9524a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 9535d4f98a2SYan Zheng node = prev; 9545d4f98a2SYan Zheng break; 9555d4f98a2SYan Zheng } 9565d4f98a2SYan Zheng prev = rb_next(prev); 9575d4f98a2SYan Zheng } 9585d4f98a2SYan Zheng } 9595d4f98a2SYan Zheng while (node) { 9605d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9615d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 9625d4f98a2SYan Zheng if (inode) { 9635d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9645d4f98a2SYan Zheng return inode; 9655d4f98a2SYan Zheng } 9665d4f98a2SYan Zheng 9674a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 9685d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 9695d4f98a2SYan Zheng goto again; 9705d4f98a2SYan Zheng 9715d4f98a2SYan Zheng node = rb_next(node); 9725d4f98a2SYan Zheng } 9735d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9745d4f98a2SYan Zheng return NULL; 9755d4f98a2SYan Zheng } 9765d4f98a2SYan Zheng 9775d4f98a2SYan Zheng /* 9785d4f98a2SYan Zheng * get new location of data 9795d4f98a2SYan Zheng */ 9805d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 9815d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 9825d4f98a2SYan Zheng { 9835d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 9845d4f98a2SYan Zheng struct btrfs_path *path; 9855d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 9865d4f98a2SYan Zheng struct extent_buffer *leaf; 9875d4f98a2SYan Zheng int ret; 9885d4f98a2SYan Zheng 9895d4f98a2SYan Zheng path = btrfs_alloc_path(); 9905d4f98a2SYan Zheng if (!path) 9915d4f98a2SYan Zheng return -ENOMEM; 9925d4f98a2SYan Zheng 9935d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 994f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 995f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 9965d4f98a2SYan Zheng if (ret < 0) 9975d4f98a2SYan Zheng goto out; 9985d4f98a2SYan Zheng if (ret > 0) { 9995d4f98a2SYan Zheng ret = -ENOENT; 10005d4f98a2SYan Zheng goto out; 10015d4f98a2SYan Zheng } 10025d4f98a2SYan Zheng 10035d4f98a2SYan Zheng leaf = path->nodes[0]; 10045d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 10055d4f98a2SYan Zheng struct btrfs_file_extent_item); 10065d4f98a2SYan Zheng 10075d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 10085d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 10095d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 10105d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 10115d4f98a2SYan Zheng 10125d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 101383d4cfd4SJosef Bacik ret = -EINVAL; 10145d4f98a2SYan Zheng goto out; 10155d4f98a2SYan Zheng } 10165d4f98a2SYan Zheng 10175d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10185d4f98a2SYan Zheng ret = 0; 10195d4f98a2SYan Zheng out: 10205d4f98a2SYan Zheng btrfs_free_path(path); 10215d4f98a2SYan Zheng return ret; 10225d4f98a2SYan Zheng } 10235d4f98a2SYan Zheng 10245d4f98a2SYan Zheng /* 10255d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 10265d4f98a2SYan Zheng * the new locations. 10275d4f98a2SYan Zheng */ 10283fd0a558SYan, Zheng static noinline_for_stack 10293fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 10305d4f98a2SYan Zheng struct reloc_control *rc, 10315d4f98a2SYan Zheng struct btrfs_root *root, 10323fd0a558SYan, Zheng struct extent_buffer *leaf) 10335d4f98a2SYan Zheng { 10340b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 10355d4f98a2SYan Zheng struct btrfs_key key; 10365d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10375d4f98a2SYan Zheng struct inode *inode = NULL; 10385d4f98a2SYan Zheng u64 parent; 10395d4f98a2SYan Zheng u64 bytenr; 10403fd0a558SYan, Zheng u64 new_bytenr = 0; 10415d4f98a2SYan Zheng u64 num_bytes; 10425d4f98a2SYan Zheng u64 end; 10435d4f98a2SYan Zheng u32 nritems; 10445d4f98a2SYan Zheng u32 i; 104583d4cfd4SJosef Bacik int ret = 0; 10465d4f98a2SYan Zheng int first = 1; 10475d4f98a2SYan Zheng int dirty = 0; 10485d4f98a2SYan Zheng 10495d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 10505d4f98a2SYan Zheng return 0; 10515d4f98a2SYan Zheng 10525d4f98a2SYan Zheng /* reloc trees always use full backref */ 10535d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 10545d4f98a2SYan Zheng parent = leaf->start; 10555d4f98a2SYan Zheng else 10565d4f98a2SYan Zheng parent = 0; 10575d4f98a2SYan Zheng 10585d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 10595d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 106082fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 106182fa113fSQu Wenruo 10625d4f98a2SYan Zheng cond_resched(); 10635d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 10645d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 10655d4f98a2SYan Zheng continue; 10665d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 10675d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 10685d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 10695d4f98a2SYan Zheng continue; 10705d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10715d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 10725d4f98a2SYan Zheng if (bytenr == 0) 10735d4f98a2SYan Zheng continue; 10749569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 10759569cc20SQu Wenruo rc->block_group->length)) 10765d4f98a2SYan Zheng continue; 10775d4f98a2SYan Zheng 10785d4f98a2SYan Zheng /* 10795d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 10805d4f98a2SYan Zheng * to complete and drop the extent cache 10815d4f98a2SYan Zheng */ 10825d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 10835d4f98a2SYan Zheng if (first) { 10845d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 10855d4f98a2SYan Zheng first = 0; 10864a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 10873fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 10885d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 10895d4f98a2SYan Zheng } 10904a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 10915d4f98a2SYan Zheng end = key.offset + 10925d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 10935d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 10940b246afaSJeff Mahoney fs_info->sectorsize)); 10950b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 10965d4f98a2SYan Zheng end--; 10975d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1098d0082371SJeff Mahoney key.offset, end); 10995d4f98a2SYan Zheng if (!ret) 11005d4f98a2SYan Zheng continue; 11015d4f98a2SYan Zheng 1102dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1103dcdbc059SNikolay Borisov key.offset, end, 1); 11045d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1105d0082371SJeff Mahoney key.offset, end); 11065d4f98a2SYan Zheng } 11075d4f98a2SYan Zheng } 11085d4f98a2SYan Zheng 11095d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 11105d4f98a2SYan Zheng bytenr, num_bytes); 111183d4cfd4SJosef Bacik if (ret) { 111283d4cfd4SJosef Bacik /* 111383d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 111483d4cfd4SJosef Bacik * in the file extent yet. 111583d4cfd4SJosef Bacik */ 111683d4cfd4SJosef Bacik break; 11173fd0a558SYan, Zheng } 11185d4f98a2SYan Zheng 11195d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 11205d4f98a2SYan Zheng dirty = 1; 11215d4f98a2SYan Zheng 11225d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 112382fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 112482fa113fSQu Wenruo num_bytes, parent); 112582fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 112682fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1127b06c4bf5SFilipe Manana key.objectid, key.offset); 112882fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 112983d4cfd4SJosef Bacik if (ret) { 113066642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 113183d4cfd4SJosef Bacik break; 113283d4cfd4SJosef Bacik } 11335d4f98a2SYan Zheng 1134ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1135ffd4bb2aSQu Wenruo num_bytes, parent); 1136ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1137ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1138b06c4bf5SFilipe Manana key.objectid, key.offset); 1139ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 114083d4cfd4SJosef Bacik if (ret) { 114166642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 114283d4cfd4SJosef Bacik break; 114383d4cfd4SJosef Bacik } 11445d4f98a2SYan Zheng } 11455d4f98a2SYan Zheng if (dirty) 11465d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 11473fd0a558SYan, Zheng if (inode) 11483fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 114983d4cfd4SJosef Bacik return ret; 11505d4f98a2SYan Zheng } 11515d4f98a2SYan Zheng 11525d4f98a2SYan Zheng static noinline_for_stack 11535d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 11545d4f98a2SYan Zheng struct btrfs_path *path, int level) 11555d4f98a2SYan Zheng { 11565d4f98a2SYan Zheng struct btrfs_disk_key key1; 11575d4f98a2SYan Zheng struct btrfs_disk_key key2; 11585d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 11595d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 11605d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 11615d4f98a2SYan Zheng } 11625d4f98a2SYan Zheng 11635d4f98a2SYan Zheng /* 11645d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 11655d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 11665d4f98a2SYan Zheng * reloc tree was create can be replaced. 11675d4f98a2SYan Zheng * 11685d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 11695d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 11705d4f98a2SYan Zheng * errors, a negative error number is returned. 11715d4f98a2SYan Zheng */ 11723fd0a558SYan, Zheng static noinline_for_stack 11733d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 11745d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 11755d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 11765d4f98a2SYan Zheng int lowest_level, int max_level) 11775d4f98a2SYan Zheng { 11780b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 11795d4f98a2SYan Zheng struct extent_buffer *eb; 11805d4f98a2SYan Zheng struct extent_buffer *parent; 118182fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 11825d4f98a2SYan Zheng struct btrfs_key key; 11835d4f98a2SYan Zheng u64 old_bytenr; 11845d4f98a2SYan Zheng u64 new_bytenr; 11855d4f98a2SYan Zheng u64 old_ptr_gen; 11865d4f98a2SYan Zheng u64 new_ptr_gen; 11875d4f98a2SYan Zheng u64 last_snapshot; 11885d4f98a2SYan Zheng u32 blocksize; 11893fd0a558SYan, Zheng int cow = 0; 11905d4f98a2SYan Zheng int level; 11915d4f98a2SYan Zheng int ret; 11925d4f98a2SYan Zheng int slot; 11935d4f98a2SYan Zheng 11945d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 11955d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 11965d4f98a2SYan Zheng 11975d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 11983fd0a558SYan, Zheng again: 11995d4f98a2SYan Zheng slot = path->slots[lowest_level]; 12005d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 12015d4f98a2SYan Zheng 12025d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 12035d4f98a2SYan Zheng level = btrfs_header_level(eb); 12045d4f98a2SYan Zheng 12055d4f98a2SYan Zheng if (level < lowest_level) { 12065d4f98a2SYan Zheng btrfs_tree_unlock(eb); 12075d4f98a2SYan Zheng free_extent_buffer(eb); 12085d4f98a2SYan Zheng return 0; 12095d4f98a2SYan Zheng } 12105d4f98a2SYan Zheng 12113fd0a558SYan, Zheng if (cow) { 12129631e4ccSJosef Bacik ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb, 12139631e4ccSJosef Bacik BTRFS_NESTING_COW); 12145d4f98a2SYan Zheng BUG_ON(ret); 12153fd0a558SYan, Zheng } 12165d4f98a2SYan Zheng 12175d4f98a2SYan Zheng if (next_key) { 12185d4f98a2SYan Zheng next_key->objectid = (u64)-1; 12195d4f98a2SYan Zheng next_key->type = (u8)-1; 12205d4f98a2SYan Zheng next_key->offset = (u64)-1; 12215d4f98a2SYan Zheng } 12225d4f98a2SYan Zheng 12235d4f98a2SYan Zheng parent = eb; 12245d4f98a2SYan Zheng while (1) { 12255d4f98a2SYan Zheng level = btrfs_header_level(parent); 12265d4f98a2SYan Zheng BUG_ON(level < lowest_level); 12275d4f98a2SYan Zheng 1228e3b83361SQu Wenruo ret = btrfs_bin_search(parent, &key, &slot); 1229cbca7d59SFilipe Manana if (ret < 0) 1230cbca7d59SFilipe Manana break; 12315d4f98a2SYan Zheng if (ret && slot > 0) 12325d4f98a2SYan Zheng slot--; 12335d4f98a2SYan Zheng 12345d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 12355d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 12365d4f98a2SYan Zheng 12375d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 12380b246afaSJeff Mahoney blocksize = fs_info->nodesize; 12395d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 12405d4f98a2SYan Zheng 12415d4f98a2SYan Zheng if (level <= max_level) { 12425d4f98a2SYan Zheng eb = path->nodes[level]; 12435d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 12445d4f98a2SYan Zheng path->slots[level]); 12455d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 12465d4f98a2SYan Zheng path->slots[level]); 12475d4f98a2SYan Zheng } else { 12485d4f98a2SYan Zheng new_bytenr = 0; 12495d4f98a2SYan Zheng new_ptr_gen = 0; 12505d4f98a2SYan Zheng } 12515d4f98a2SYan Zheng 1252fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 12535d4f98a2SYan Zheng ret = level; 12545d4f98a2SYan Zheng break; 12555d4f98a2SYan Zheng } 12565d4f98a2SYan Zheng 12575d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 12585d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 12593fd0a558SYan, Zheng if (level <= lowest_level) { 12605d4f98a2SYan Zheng ret = 0; 12615d4f98a2SYan Zheng break; 12625d4f98a2SYan Zheng } 12635d4f98a2SYan Zheng 12646b3426beSJosef Bacik eb = btrfs_read_node_slot(parent, slot); 126564c043deSLiu Bo if (IS_ERR(eb)) { 126664c043deSLiu Bo ret = PTR_ERR(eb); 1267264813acSLiu Bo break; 1268416bc658SJosef Bacik } 12695d4f98a2SYan Zheng btrfs_tree_lock(eb); 12703fd0a558SYan, Zheng if (cow) { 12715d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 12729631e4ccSJosef Bacik slot, &eb, 12739631e4ccSJosef Bacik BTRFS_NESTING_COW); 12745d4f98a2SYan Zheng BUG_ON(ret); 12755d4f98a2SYan Zheng } 12765d4f98a2SYan Zheng 12775d4f98a2SYan Zheng btrfs_tree_unlock(parent); 12785d4f98a2SYan Zheng free_extent_buffer(parent); 12795d4f98a2SYan Zheng 12805d4f98a2SYan Zheng parent = eb; 12815d4f98a2SYan Zheng continue; 12825d4f98a2SYan Zheng } 12835d4f98a2SYan Zheng 12843fd0a558SYan, Zheng if (!cow) { 12853fd0a558SYan, Zheng btrfs_tree_unlock(parent); 12863fd0a558SYan, Zheng free_extent_buffer(parent); 12873fd0a558SYan, Zheng cow = 1; 12883fd0a558SYan, Zheng goto again; 12893fd0a558SYan, Zheng } 12903fd0a558SYan, Zheng 12915d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 12925d4f98a2SYan Zheng path->slots[level]); 1293b3b4aa74SDavid Sterba btrfs_release_path(path); 12945d4f98a2SYan Zheng 12955d4f98a2SYan Zheng path->lowest_level = level; 12965d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 12975d4f98a2SYan Zheng path->lowest_level = 0; 12985d4f98a2SYan Zheng BUG_ON(ret); 12995d4f98a2SYan Zheng 13005d4f98a2SYan Zheng /* 1301824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 1302824d8dffSQu Wenruo * 1303824d8dffSQu Wenruo * We must trace both trees. 1304824d8dffSQu Wenruo * 1) Tree reloc subtree 1305824d8dffSQu Wenruo * If not traced, we will leak data numbers 1306824d8dffSQu Wenruo * 2) Fs subtree 1307824d8dffSQu Wenruo * If not traced, we will double count old data 1308f616f5cdSQu Wenruo * 1309f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 1310f616f5cdSQu Wenruo * the swapped tree blocks. 1311f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 1312f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 1313824d8dffSQu Wenruo */ 1314370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 1315370a11b8SQu Wenruo rc->block_group, parent, slot, 1316370a11b8SQu Wenruo path->nodes[level], path->slots[level], 1317370a11b8SQu Wenruo last_snapshot); 1318370a11b8SQu Wenruo if (ret < 0) 1319370a11b8SQu Wenruo break; 1320824d8dffSQu Wenruo /* 13215d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 13225d4f98a2SYan Zheng */ 13235d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 13245d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 13255d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 13265d4f98a2SYan Zheng 13275d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 13285d4f98a2SYan Zheng path->slots[level], old_bytenr); 13295d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 13305d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 13315d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 13325d4f98a2SYan Zheng 133382fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 133482fa113fSQu Wenruo blocksize, path->nodes[level]->start); 133582fa113fSQu Wenruo ref.skip_qgroup = true; 133682fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 133782fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13385d4f98a2SYan Zheng BUG_ON(ret); 133982fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 134082fa113fSQu Wenruo blocksize, 0); 134182fa113fSQu Wenruo ref.skip_qgroup = true; 134282fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 134382fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13445d4f98a2SYan Zheng BUG_ON(ret); 13455d4f98a2SYan Zheng 1346ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 1347ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 1348ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 1349ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1350ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13515d4f98a2SYan Zheng BUG_ON(ret); 13525d4f98a2SYan Zheng 1353ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 1354ffd4bb2aSQu Wenruo blocksize, 0); 1355ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 1356ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1357ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13585d4f98a2SYan Zheng BUG_ON(ret); 13595d4f98a2SYan Zheng 13605d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 13615d4f98a2SYan Zheng 13625d4f98a2SYan Zheng ret = level; 13635d4f98a2SYan Zheng break; 13645d4f98a2SYan Zheng } 13655d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13665d4f98a2SYan Zheng free_extent_buffer(parent); 13675d4f98a2SYan Zheng return ret; 13685d4f98a2SYan Zheng } 13695d4f98a2SYan Zheng 13705d4f98a2SYan Zheng /* 13715d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 13725d4f98a2SYan Zheng */ 13735d4f98a2SYan Zheng static noinline_for_stack 13745d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 13755d4f98a2SYan Zheng int *level) 13765d4f98a2SYan Zheng { 13775d4f98a2SYan Zheng struct extent_buffer *eb; 13785d4f98a2SYan Zheng int i; 13795d4f98a2SYan Zheng u64 last_snapshot; 13805d4f98a2SYan Zheng u32 nritems; 13815d4f98a2SYan Zheng 13825d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 13835d4f98a2SYan Zheng 13845d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 13855d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 13865d4f98a2SYan Zheng path->nodes[i] = NULL; 13875d4f98a2SYan Zheng } 13885d4f98a2SYan Zheng 13895d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 13905d4f98a2SYan Zheng eb = path->nodes[i]; 13915d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 13925d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 13935d4f98a2SYan Zheng path->slots[i]++; 13945d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 13955d4f98a2SYan Zheng last_snapshot) 13965d4f98a2SYan Zheng continue; 13975d4f98a2SYan Zheng 13985d4f98a2SYan Zheng *level = i; 13995d4f98a2SYan Zheng return 0; 14005d4f98a2SYan Zheng } 14015d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14025d4f98a2SYan Zheng path->nodes[i] = NULL; 14035d4f98a2SYan Zheng } 14045d4f98a2SYan Zheng return 1; 14055d4f98a2SYan Zheng } 14065d4f98a2SYan Zheng 14075d4f98a2SYan Zheng /* 14085d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 14095d4f98a2SYan Zheng */ 14105d4f98a2SYan Zheng static noinline_for_stack 14115d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14125d4f98a2SYan Zheng int *level) 14135d4f98a2SYan Zheng { 14145d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 14155d4f98a2SYan Zheng int i; 14165d4f98a2SYan Zheng u64 ptr_gen = 0; 14175d4f98a2SYan Zheng u64 last_snapshot; 14185d4f98a2SYan Zheng u32 nritems; 14195d4f98a2SYan Zheng 14205d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14215d4f98a2SYan Zheng 14225d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 14235d4f98a2SYan Zheng eb = path->nodes[i]; 14245d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14255d4f98a2SYan Zheng while (path->slots[i] < nritems) { 14265d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 14275d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 14285d4f98a2SYan Zheng break; 14295d4f98a2SYan Zheng path->slots[i]++; 14305d4f98a2SYan Zheng } 14315d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 14325d4f98a2SYan Zheng if (i == *level) 14335d4f98a2SYan Zheng break; 14345d4f98a2SYan Zheng *level = i + 1; 14355d4f98a2SYan Zheng return 0; 14365d4f98a2SYan Zheng } 14375d4f98a2SYan Zheng if (i == 1) { 14385d4f98a2SYan Zheng *level = i; 14395d4f98a2SYan Zheng return 0; 14405d4f98a2SYan Zheng } 14415d4f98a2SYan Zheng 14428ef385bbSJosef Bacik eb = btrfs_read_node_slot(eb, path->slots[i]); 14438ef385bbSJosef Bacik if (IS_ERR(eb)) 144464c043deSLiu Bo return PTR_ERR(eb); 14455d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 14465d4f98a2SYan Zheng path->nodes[i - 1] = eb; 14475d4f98a2SYan Zheng path->slots[i - 1] = 0; 14485d4f98a2SYan Zheng } 14495d4f98a2SYan Zheng return 1; 14505d4f98a2SYan Zheng } 14515d4f98a2SYan Zheng 14525d4f98a2SYan Zheng /* 14535d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 14545d4f98a2SYan Zheng * [min_key, max_key) 14555d4f98a2SYan Zheng */ 14565d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 14575d4f98a2SYan Zheng struct btrfs_key *min_key, 14585d4f98a2SYan Zheng struct btrfs_key *max_key) 14595d4f98a2SYan Zheng { 14600b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14615d4f98a2SYan Zheng struct inode *inode = NULL; 14625d4f98a2SYan Zheng u64 objectid; 14635d4f98a2SYan Zheng u64 start, end; 146433345d01SLi Zefan u64 ino; 14655d4f98a2SYan Zheng 14665d4f98a2SYan Zheng objectid = min_key->objectid; 14675d4f98a2SYan Zheng while (1) { 14685d4f98a2SYan Zheng cond_resched(); 14695d4f98a2SYan Zheng iput(inode); 14705d4f98a2SYan Zheng 14715d4f98a2SYan Zheng if (objectid > max_key->objectid) 14725d4f98a2SYan Zheng break; 14735d4f98a2SYan Zheng 14745d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 14755d4f98a2SYan Zheng if (!inode) 14765d4f98a2SYan Zheng break; 14774a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 14785d4f98a2SYan Zheng 147933345d01SLi Zefan if (ino > max_key->objectid) { 14805d4f98a2SYan Zheng iput(inode); 14815d4f98a2SYan Zheng break; 14825d4f98a2SYan Zheng } 14835d4f98a2SYan Zheng 148433345d01SLi Zefan objectid = ino + 1; 14855d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 14865d4f98a2SYan Zheng continue; 14875d4f98a2SYan Zheng 148833345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 14895d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 14905d4f98a2SYan Zheng continue; 14915d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 14925d4f98a2SYan Zheng start = 0; 14935d4f98a2SYan Zheng else { 14945d4f98a2SYan Zheng start = min_key->offset; 14950b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 14965d4f98a2SYan Zheng } 14975d4f98a2SYan Zheng } else { 14985d4f98a2SYan Zheng start = 0; 14995d4f98a2SYan Zheng } 15005d4f98a2SYan Zheng 150133345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 15025d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 15035d4f98a2SYan Zheng continue; 15045d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 15055d4f98a2SYan Zheng end = (u64)-1; 15065d4f98a2SYan Zheng } else { 15075d4f98a2SYan Zheng if (max_key->offset == 0) 15085d4f98a2SYan Zheng continue; 15095d4f98a2SYan Zheng end = max_key->offset; 15100b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 15115d4f98a2SYan Zheng end--; 15125d4f98a2SYan Zheng } 15135d4f98a2SYan Zheng } else { 15145d4f98a2SYan Zheng end = (u64)-1; 15155d4f98a2SYan Zheng } 15165d4f98a2SYan Zheng 15175d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 1518d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 1519dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 1520d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 15215d4f98a2SYan Zheng } 15225d4f98a2SYan Zheng return 0; 15235d4f98a2SYan Zheng } 15245d4f98a2SYan Zheng 15255d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 15265d4f98a2SYan Zheng struct btrfs_key *key) 15275d4f98a2SYan Zheng 15285d4f98a2SYan Zheng { 15295d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 15305d4f98a2SYan Zheng if (!path->nodes[level]) 15315d4f98a2SYan Zheng break; 15325d4f98a2SYan Zheng if (path->slots[level] + 1 < 15335d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 15345d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 15355d4f98a2SYan Zheng path->slots[level] + 1); 15365d4f98a2SYan Zheng return 0; 15375d4f98a2SYan Zheng } 15385d4f98a2SYan Zheng level++; 15395d4f98a2SYan Zheng } 15405d4f98a2SYan Zheng return 1; 15415d4f98a2SYan Zheng } 15425d4f98a2SYan Zheng 15435d4f98a2SYan Zheng /* 1544d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 1545d2311e69SQu Wenruo */ 1546d2311e69SQu Wenruo static void insert_dirty_subvol(struct btrfs_trans_handle *trans, 1547d2311e69SQu Wenruo struct reloc_control *rc, 1548d2311e69SQu Wenruo struct btrfs_root *root) 1549d2311e69SQu Wenruo { 1550d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1551d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 1552d2311e69SQu Wenruo 1553d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 1554d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 1555d2311e69SQu Wenruo ASSERT(reloc_root); 1556d2311e69SQu Wenruo 1557d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 1558d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 1559d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 1560c8422684SDavid Sterba btrfs_set_root_drop_level(reloc_root_item, 0); 1561d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 1562d2311e69SQu Wenruo btrfs_update_reloc_root(trans, root); 1563d2311e69SQu Wenruo 1564d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 156500246528SJosef Bacik btrfs_grab_root(root); 1566d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 1567d2311e69SQu Wenruo } 1568d2311e69SQu Wenruo } 1569d2311e69SQu Wenruo 1570d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 1571d2311e69SQu Wenruo { 1572d2311e69SQu Wenruo struct btrfs_root *root; 1573d2311e69SQu Wenruo struct btrfs_root *next; 1574d2311e69SQu Wenruo int ret = 0; 157530d40577SQu Wenruo int ret2; 1576d2311e69SQu Wenruo 1577d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 1578d2311e69SQu Wenruo reloc_dirty_list) { 157930d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 158030d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 1581d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1582d2311e69SQu Wenruo 1583d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 1584d2311e69SQu Wenruo root->reloc_root = NULL; 15856282675eSQu Wenruo /* 15866282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 15876282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 15886282675eSQu Wenruo */ 15896282675eSQu Wenruo smp_wmb(); 15901fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 1591f28de8d8SJosef Bacik if (reloc_root) { 1592f44deb74SJosef Bacik /* 1593f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 1594f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 1595f44deb74SJosef Bacik * drop the ref ourselves. 1596f44deb74SJosef Bacik */ 1597f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 1598f44deb74SJosef Bacik if (ret2 < 0) { 1599f44deb74SJosef Bacik btrfs_put_root(reloc_root); 1600f44deb74SJosef Bacik if (!ret) 1601f28de8d8SJosef Bacik ret = ret2; 1602f28de8d8SJosef Bacik } 1603f44deb74SJosef Bacik } 160400246528SJosef Bacik btrfs_put_root(root); 160530d40577SQu Wenruo } else { 160630d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 16070078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 1608f44deb74SJosef Bacik if (ret2 < 0) { 1609f44deb74SJosef Bacik btrfs_put_root(root); 1610f44deb74SJosef Bacik if (!ret) 161130d40577SQu Wenruo ret = ret2; 161230d40577SQu Wenruo } 1613d2311e69SQu Wenruo } 1614f44deb74SJosef Bacik } 1615d2311e69SQu Wenruo return ret; 1616d2311e69SQu Wenruo } 1617d2311e69SQu Wenruo 1618d2311e69SQu Wenruo /* 16195d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 16205d4f98a2SYan Zheng * fs tree. 16215d4f98a2SYan Zheng */ 16225d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 16235d4f98a2SYan Zheng struct btrfs_root *root) 16245d4f98a2SYan Zheng { 16250b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 16265d4f98a2SYan Zheng struct btrfs_key key; 16275d4f98a2SYan Zheng struct btrfs_key next_key; 16289e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 16295d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16305d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16315d4f98a2SYan Zheng struct btrfs_path *path; 16323fd0a558SYan, Zheng struct extent_buffer *leaf; 1633fca3a45dSJosef Bacik int reserve_level; 16345d4f98a2SYan Zheng int level; 16355d4f98a2SYan Zheng int max_level; 16365d4f98a2SYan Zheng int replaced = 0; 1637c6a592f2SNikolay Borisov int ret = 0; 16383fd0a558SYan, Zheng u32 min_reserved; 16395d4f98a2SYan Zheng 16405d4f98a2SYan Zheng path = btrfs_alloc_path(); 16415d4f98a2SYan Zheng if (!path) 16425d4f98a2SYan Zheng return -ENOMEM; 1643e4058b54SDavid Sterba path->reada = READA_FORWARD; 16445d4f98a2SYan Zheng 16455d4f98a2SYan Zheng reloc_root = root->reloc_root; 16465d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16475d4f98a2SYan Zheng 16485d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 16495d4f98a2SYan Zheng level = btrfs_root_level(root_item); 165067439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 16515d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 16525d4f98a2SYan Zheng path->slots[level] = 0; 16535d4f98a2SYan Zheng } else { 16545d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 16555d4f98a2SYan Zheng 1656c8422684SDavid Sterba level = btrfs_root_drop_level(root_item); 16575d4f98a2SYan Zheng BUG_ON(level == 0); 16585d4f98a2SYan Zheng path->lowest_level = level; 16595d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 166033c66f43SYan Zheng path->lowest_level = 0; 16615d4f98a2SYan Zheng if (ret < 0) { 16625d4f98a2SYan Zheng btrfs_free_path(path); 16635d4f98a2SYan Zheng return ret; 16645d4f98a2SYan Zheng } 16655d4f98a2SYan Zheng 16665d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 16675d4f98a2SYan Zheng path->slots[level]); 16685d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 16695d4f98a2SYan Zheng 16705d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 16715d4f98a2SYan Zheng } 16725d4f98a2SYan Zheng 167344d354abSQu Wenruo /* 167444d354abSQu Wenruo * In merge_reloc_root(), we modify the upper level pointer to swap the 167544d354abSQu Wenruo * tree blocks between reloc tree and subvolume tree. Thus for tree 167644d354abSQu Wenruo * block COW, we COW at most from level 1 to root level for each tree. 167744d354abSQu Wenruo * 167844d354abSQu Wenruo * Thus the needed metadata size is at most root_level * nodesize, 167944d354abSQu Wenruo * and * 2 since we have two trees to COW. 168044d354abSQu Wenruo */ 1681fca3a45dSJosef Bacik reserve_level = max_t(int, 1, btrfs_root_level(root_item)); 1682fca3a45dSJosef Bacik min_reserved = fs_info->nodesize * reserve_level * 2; 16835d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 16845d4f98a2SYan Zheng 16855d4f98a2SYan Zheng while (1) { 168608e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 168744d354abSQu Wenruo BTRFS_RESERVE_FLUSH_LIMIT); 1688c6a592f2SNikolay Borisov if (ret) 16899e6a0c52SJosef Bacik goto out; 16909e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 16919e6a0c52SJosef Bacik if (IS_ERR(trans)) { 1692c6a592f2SNikolay Borisov ret = PTR_ERR(trans); 16939e6a0c52SJosef Bacik trans = NULL; 16949e6a0c52SJosef Bacik goto out; 16959e6a0c52SJosef Bacik } 16962abc726aSJosef Bacik 16972abc726aSJosef Bacik /* 16982abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 16992abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 17002abc726aSJosef Bacik * 17012abc726aSJosef Bacik * But that's ok, we started the trans handle on our 17022abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 17032abc726aSJosef Bacik * dirty list. At commit time we'll still call 17042abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 17052abc726aSJosef Bacik * appropriately. 17062abc726aSJosef Bacik */ 17072abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 17089e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 17093fd0a558SYan, Zheng 17103fd0a558SYan, Zheng replaced = 0; 17115d4f98a2SYan Zheng max_level = level; 17125d4f98a2SYan Zheng 17135d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 1714c6a592f2SNikolay Borisov if (ret < 0) 17155d4f98a2SYan Zheng goto out; 17165d4f98a2SYan Zheng if (ret > 0) 17175d4f98a2SYan Zheng break; 17185d4f98a2SYan Zheng 17195d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 17205d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 17215d4f98a2SYan Zheng ret = 0; 17225d4f98a2SYan Zheng } else { 17233d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 17243fd0a558SYan, Zheng &next_key, level, max_level); 17255d4f98a2SYan Zheng } 1726c6a592f2SNikolay Borisov if (ret < 0) 17275d4f98a2SYan Zheng goto out; 17285d4f98a2SYan Zheng if (ret > 0) { 17295d4f98a2SYan Zheng level = ret; 17305d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17315d4f98a2SYan Zheng path->slots[level]); 17325d4f98a2SYan Zheng replaced = 1; 17335d4f98a2SYan Zheng } 17345d4f98a2SYan Zheng 17355d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 17365d4f98a2SYan Zheng if (ret > 0) 17375d4f98a2SYan Zheng break; 17385d4f98a2SYan Zheng 17395d4f98a2SYan Zheng BUG_ON(level == 0); 17405d4f98a2SYan Zheng /* 17415d4f98a2SYan Zheng * save the merging progress in the drop_progress. 17425d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 17435d4f98a2SYan Zheng */ 17445d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 17455d4f98a2SYan Zheng path->slots[level]); 1746c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, level); 17475d4f98a2SYan Zheng 17483a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 17499e6a0c52SJosef Bacik trans = NULL; 17505d4f98a2SYan Zheng 17512ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 17525d4f98a2SYan Zheng 17535d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17545d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17555d4f98a2SYan Zheng } 17565d4f98a2SYan Zheng 17575d4f98a2SYan Zheng /* 17585d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 17595d4f98a2SYan Zheng * relocated and the block is tree root. 17605d4f98a2SYan Zheng */ 17615d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 17629631e4ccSJosef Bacik ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf, 17639631e4ccSJosef Bacik BTRFS_NESTING_COW); 17645d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 17655d4f98a2SYan Zheng free_extent_buffer(leaf); 17665d4f98a2SYan Zheng out: 17675d4f98a2SYan Zheng btrfs_free_path(path); 17685d4f98a2SYan Zheng 1769c6a592f2SNikolay Borisov if (ret == 0) 1770d2311e69SQu Wenruo insert_dirty_subvol(trans, rc, root); 17715d4f98a2SYan Zheng 17729e6a0c52SJosef Bacik if (trans) 17733a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 17745d4f98a2SYan Zheng 17752ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 17765d4f98a2SYan Zheng 17775d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17785d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17795d4f98a2SYan Zheng 1780c6a592f2SNikolay Borisov return ret; 17815d4f98a2SYan Zheng } 17825d4f98a2SYan Zheng 17833fd0a558SYan, Zheng static noinline_for_stack 17843fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 17855d4f98a2SYan Zheng { 17863fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 17870b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 17883fd0a558SYan, Zheng struct btrfs_root *reloc_root; 17895d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 17903fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 17913fd0a558SYan, Zheng u64 num_bytes = 0; 17923fd0a558SYan, Zheng int ret; 17933fd0a558SYan, Zheng 17940b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 17950b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 17963fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 17970b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 17987585717fSChris Mason 17993fd0a558SYan, Zheng again: 18003fd0a558SYan, Zheng if (!err) { 18013fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 180208e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 180308e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 18043fd0a558SYan, Zheng if (ret) 18053fd0a558SYan, Zheng err = ret; 18063fd0a558SYan, Zheng } 18073fd0a558SYan, Zheng 18087a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 18093612b495STsutomu Itoh if (IS_ERR(trans)) { 18103612b495STsutomu Itoh if (!err) 18112ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 181263f018beSNikolay Borisov num_bytes, NULL); 18133612b495STsutomu Itoh return PTR_ERR(trans); 18143612b495STsutomu Itoh } 18153fd0a558SYan, Zheng 18163fd0a558SYan, Zheng if (!err) { 18173fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 18183a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18192ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 182063f018beSNikolay Borisov num_bytes, NULL); 18213fd0a558SYan, Zheng goto again; 18223fd0a558SYan, Zheng } 18233fd0a558SYan, Zheng } 18243fd0a558SYan, Zheng 18253fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 18263fd0a558SYan, Zheng 18273fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 18283fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 18293fd0a558SYan, Zheng struct btrfs_root, root_list); 18303fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 18313fd0a558SYan, Zheng 1832a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1833a820feb5SDavid Sterba false); 18343fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 18353fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 18363fd0a558SYan, Zheng 18373fd0a558SYan, Zheng /* 18383fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 18393fd0a558SYan, Zheng * knows it should resumes merging 18403fd0a558SYan, Zheng */ 18413fd0a558SYan, Zheng if (!err) 18423fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 18433fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 18443fd0a558SYan, Zheng 18453fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 184600246528SJosef Bacik btrfs_put_root(root); 18473fd0a558SYan, Zheng } 18483fd0a558SYan, Zheng 18493fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 18503fd0a558SYan, Zheng 18513fd0a558SYan, Zheng if (!err) 18523a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 18533fd0a558SYan, Zheng else 18543a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18553fd0a558SYan, Zheng return err; 18563fd0a558SYan, Zheng } 18573fd0a558SYan, Zheng 18583fd0a558SYan, Zheng static noinline_for_stack 1859aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 1860aca1bba6SLiu Bo { 1861a7571232SNikolay Borisov struct btrfs_root *reloc_root, *tmp; 1862aca1bba6SLiu Bo 1863a7571232SNikolay Borisov list_for_each_entry_safe(reloc_root, tmp, list, root_list) 1864bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 1865aca1bba6SLiu Bo } 1866aca1bba6SLiu Bo 1867aca1bba6SLiu Bo static noinline_for_stack 186894404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 18693fd0a558SYan, Zheng { 18700b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 18715d4f98a2SYan Zheng struct btrfs_root *root; 18725d4f98a2SYan Zheng struct btrfs_root *reloc_root; 18733fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 18743fd0a558SYan, Zheng int found = 0; 1875aca1bba6SLiu Bo int ret = 0; 18763fd0a558SYan, Zheng again: 18773fd0a558SYan, Zheng root = rc->extent_root; 18787585717fSChris Mason 18797585717fSChris Mason /* 18807585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 18817585717fSChris Mason * we have to make sure nobody is in the middle of 18827585717fSChris Mason * adding their roots to the list while we are 18837585717fSChris Mason * doing this splice 18847585717fSChris Mason */ 18850b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 18863fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 18870b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 18885d4f98a2SYan Zheng 18893fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 18903fd0a558SYan, Zheng found = 1; 18913fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 18923fd0a558SYan, Zheng struct btrfs_root, root_list); 18935d4f98a2SYan Zheng 1894a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1895a820feb5SDavid Sterba false); 18965d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 18975d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 18985d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 18993fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 190000246528SJosef Bacik btrfs_put_root(root); 1901b37b39cdSJosef Bacik if (ret) { 190225e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 190325e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 190425e293c2SWang Shilong &reloc_roots); 1905aca1bba6SLiu Bo goto out; 1906b37b39cdSJosef Bacik } 19073fd0a558SYan, Zheng } else { 190851415b6cSQu Wenruo if (!IS_ERR(root)) { 190951415b6cSQu Wenruo if (root->reloc_root == reloc_root) { 191051415b6cSQu Wenruo root->reloc_root = NULL; 191151415b6cSQu Wenruo btrfs_put_root(reloc_root); 191251415b6cSQu Wenruo } 19131dae7e0eSQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, 19141dae7e0eSQu Wenruo &root->state); 191551415b6cSQu Wenruo btrfs_put_root(root); 191651415b6cSQu Wenruo } 191751415b6cSQu Wenruo 19183fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 191930d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 192030d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 192130d40577SQu Wenruo &rc->dirty_subvol_roots); 19223fd0a558SYan, Zheng } 19235d4f98a2SYan Zheng } 19245d4f98a2SYan Zheng 19253fd0a558SYan, Zheng if (found) { 19263fd0a558SYan, Zheng found = 0; 19273fd0a558SYan, Zheng goto again; 19285d4f98a2SYan Zheng } 1929aca1bba6SLiu Bo out: 1930aca1bba6SLiu Bo if (ret) { 19310b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 1932aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 1933467bb1d2SWang Shilong 1934467bb1d2SWang Shilong /* new reloc root may be added */ 19350b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 1936467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 19370b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 1938467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 1939aca1bba6SLiu Bo } 1940aca1bba6SLiu Bo 19417b7b7431SJosef Bacik /* 19427b7b7431SJosef Bacik * We used to have 19437b7b7431SJosef Bacik * 19447b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 19457b7b7431SJosef Bacik * 19467b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 19477b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 19487b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 19497b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 19507b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 19517b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 19527b7b7431SJosef Bacik * will be cleaned up on unmount. 19537b7b7431SJosef Bacik * 19547b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 19557b7b7431SJosef Bacik */ 19565d4f98a2SYan Zheng } 19575d4f98a2SYan Zheng 19585d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 19595d4f98a2SYan Zheng { 19605d4f98a2SYan Zheng struct tree_block *block; 19615d4f98a2SYan Zheng struct rb_node *rb_node; 19625d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 19635d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 19645d4f98a2SYan Zheng rb_erase(rb_node, blocks); 19655d4f98a2SYan Zheng kfree(block); 19665d4f98a2SYan Zheng } 19675d4f98a2SYan Zheng } 19685d4f98a2SYan Zheng 19695d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 19705d4f98a2SYan Zheng struct btrfs_root *reloc_root) 19715d4f98a2SYan Zheng { 19720b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 19735d4f98a2SYan Zheng struct btrfs_root *root; 1974442b1ac5SJosef Bacik int ret; 19755d4f98a2SYan Zheng 19765d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 19775d4f98a2SYan Zheng return 0; 19785d4f98a2SYan Zheng 1979a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false); 1980404bccbcSJosef Bacik 1981404bccbcSJosef Bacik /* 1982404bccbcSJosef Bacik * This should succeed, since we can't have a reloc root without having 1983404bccbcSJosef Bacik * already looked up the actual root and created the reloc root for this 1984404bccbcSJosef Bacik * root. 1985404bccbcSJosef Bacik * 1986404bccbcSJosef Bacik * However if there's some sort of corruption where we have a ref to a 1987404bccbcSJosef Bacik * reloc root without a corresponding root this could return ENOENT. 1988404bccbcSJosef Bacik */ 1989404bccbcSJosef Bacik if (IS_ERR(root)) { 1990404bccbcSJosef Bacik ASSERT(0); 1991404bccbcSJosef Bacik return PTR_ERR(root); 1992404bccbcSJosef Bacik } 1993404bccbcSJosef Bacik if (root->reloc_root != reloc_root) { 1994404bccbcSJosef Bacik ASSERT(0); 1995404bccbcSJosef Bacik btrfs_err(fs_info, 1996404bccbcSJosef Bacik "root %llu has two reloc roots associated with it", 1997404bccbcSJosef Bacik reloc_root->root_key.offset); 1998404bccbcSJosef Bacik btrfs_put_root(root); 1999404bccbcSJosef Bacik return -EUCLEAN; 2000404bccbcSJosef Bacik } 2001442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 200200246528SJosef Bacik btrfs_put_root(root); 20035d4f98a2SYan Zheng 2004442b1ac5SJosef Bacik return ret; 20055d4f98a2SYan Zheng } 20065d4f98a2SYan Zheng 20073fd0a558SYan, Zheng static noinline_for_stack 20083fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 20093fd0a558SYan, Zheng struct reloc_control *rc, 2010a26195a5SQu Wenruo struct btrfs_backref_node *node, 2011a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 20125d4f98a2SYan Zheng { 2013a26195a5SQu Wenruo struct btrfs_backref_node *next; 20145d4f98a2SYan Zheng struct btrfs_root *root; 20153fd0a558SYan, Zheng int index = 0; 201692de551bSJosef Bacik int ret; 20173fd0a558SYan, Zheng 20185d4f98a2SYan Zheng next = node; 20195d4f98a2SYan Zheng while (1) { 20205d4f98a2SYan Zheng cond_resched(); 20215d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 20225d4f98a2SYan Zheng root = next->root; 20238ee66afeSJosef Bacik 20248ee66afeSJosef Bacik /* 20258ee66afeSJosef Bacik * If there is no root, then our references for this block are 20268ee66afeSJosef Bacik * incomplete, as we should be able to walk all the way up to a 20278ee66afeSJosef Bacik * block that is owned by a root. 20288ee66afeSJosef Bacik * 20298ee66afeSJosef Bacik * This path is only for SHAREABLE roots, so if we come upon a 20308ee66afeSJosef Bacik * non-SHAREABLE root then we have backrefs that resolve 20318ee66afeSJosef Bacik * improperly. 20328ee66afeSJosef Bacik * 20338ee66afeSJosef Bacik * Both of these cases indicate file system corruption, or a bug 20348ee66afeSJosef Bacik * in the backref walking code. 20358ee66afeSJosef Bacik */ 20368ee66afeSJosef Bacik if (!root) { 20378ee66afeSJosef Bacik ASSERT(0); 20388ee66afeSJosef Bacik btrfs_err(trans->fs_info, 20398ee66afeSJosef Bacik "bytenr %llu doesn't have a backref path ending in a root", 20408ee66afeSJosef Bacik node->bytenr); 20418ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 20428ee66afeSJosef Bacik } 20438ee66afeSJosef Bacik if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 20448ee66afeSJosef Bacik ASSERT(0); 20458ee66afeSJosef Bacik btrfs_err(trans->fs_info, 20468ee66afeSJosef Bacik "bytenr %llu has multiple refs with one ending in a non-shareable root", 20478ee66afeSJosef Bacik node->bytenr); 20488ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 20498ee66afeSJosef Bacik } 20505d4f98a2SYan Zheng 20515d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 205292de551bSJosef Bacik ret = record_reloc_root_in_trans(trans, root); 205392de551bSJosef Bacik if (ret) 205492de551bSJosef Bacik return ERR_PTR(ret); 20555d4f98a2SYan Zheng break; 20565d4f98a2SYan Zheng } 20575d4f98a2SYan Zheng 205892de551bSJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 205992de551bSJosef Bacik if (ret) 206092de551bSJosef Bacik return ERR_PTR(ret); 20613fd0a558SYan, Zheng root = root->reloc_root; 20623fd0a558SYan, Zheng 20633fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 20648ee66afeSJosef Bacik /* 20658ee66afeSJosef Bacik * We just created the reloc root, so we shouldn't have 20668ee66afeSJosef Bacik * ->new_bytenr set and this shouldn't be in the changed 20678ee66afeSJosef Bacik * list. If it is then we have multiple roots pointing 20688ee66afeSJosef Bacik * at the same bytenr which indicates corruption, or 20698ee66afeSJosef Bacik * we've made a mistake in the backref walking code. 20708ee66afeSJosef Bacik */ 20718ee66afeSJosef Bacik ASSERT(next->new_bytenr == 0); 20728ee66afeSJosef Bacik ASSERT(list_empty(&next->list)); 20738ee66afeSJosef Bacik if (next->new_bytenr || !list_empty(&next->list)) { 20748ee66afeSJosef Bacik btrfs_err(trans->fs_info, 20758ee66afeSJosef Bacik "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu", 20768ee66afeSJosef Bacik node->bytenr, next->bytenr); 20778ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 20788ee66afeSJosef Bacik } 20798ee66afeSJosef Bacik 20803fd0a558SYan, Zheng next->new_bytenr = root->node->start; 208100246528SJosef Bacik btrfs_put_root(next->root); 208200246528SJosef Bacik next->root = btrfs_grab_root(root); 20830b530bc5SJosef Bacik ASSERT(next->root); 20843fd0a558SYan, Zheng list_add_tail(&next->list, 20853fd0a558SYan, Zheng &rc->backref_cache.changed); 20869569cc20SQu Wenruo mark_block_processed(rc, next); 20875d4f98a2SYan Zheng break; 20885d4f98a2SYan Zheng } 20895d4f98a2SYan Zheng 20903fd0a558SYan, Zheng WARN_ON(1); 20915d4f98a2SYan Zheng root = NULL; 20925d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 20935d4f98a2SYan Zheng if (!next || next->level <= node->level) 20945d4f98a2SYan Zheng break; 20955d4f98a2SYan Zheng } 2096cbdc2ebcSJosef Bacik if (!root) { 2097cbdc2ebcSJosef Bacik /* 2098cbdc2ebcSJosef Bacik * This can happen if there's fs corruption or if there's a bug 2099cbdc2ebcSJosef Bacik * in the backref lookup code. 2100cbdc2ebcSJosef Bacik */ 2101cbdc2ebcSJosef Bacik ASSERT(0); 2102cbdc2ebcSJosef Bacik return ERR_PTR(-ENOENT); 2103cbdc2ebcSJosef Bacik } 21045d4f98a2SYan Zheng 21053fd0a558SYan, Zheng next = node; 21063fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 21073fd0a558SYan, Zheng while (1) { 21083fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 21093fd0a558SYan, Zheng if (--index < 0) 21103fd0a558SYan, Zheng break; 21113fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 21123fd0a558SYan, Zheng } 21135d4f98a2SYan Zheng return root; 21145d4f98a2SYan Zheng } 21155d4f98a2SYan Zheng 21163fd0a558SYan, Zheng /* 211792a7cc42SQu Wenruo * Select a tree root for relocation. 211892a7cc42SQu Wenruo * 211992a7cc42SQu Wenruo * Return NULL if the block is not shareable. We should use do_relocation() in 212092a7cc42SQu Wenruo * this case. 212192a7cc42SQu Wenruo * 212292a7cc42SQu Wenruo * Return a tree root pointer if the block is shareable. 212392a7cc42SQu Wenruo * Return -ENOENT if the block is root of reloc tree. 21243fd0a558SYan, Zheng */ 21255d4f98a2SYan Zheng static noinline_for_stack 2126a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 21275d4f98a2SYan Zheng { 2128a26195a5SQu Wenruo struct btrfs_backref_node *next; 21293fd0a558SYan, Zheng struct btrfs_root *root; 21303fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2131a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 21323fd0a558SYan, Zheng int index = 0; 21333fd0a558SYan, Zheng 21343fd0a558SYan, Zheng next = node; 21353fd0a558SYan, Zheng while (1) { 21363fd0a558SYan, Zheng cond_resched(); 21373fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 21383fd0a558SYan, Zheng root = next->root; 21393fd0a558SYan, Zheng BUG_ON(!root); 21403fd0a558SYan, Zheng 214192a7cc42SQu Wenruo /* No other choice for non-shareable tree */ 214292a7cc42SQu Wenruo if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 21433fd0a558SYan, Zheng return root; 21443fd0a558SYan, Zheng 21453fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 21463fd0a558SYan, Zheng fs_root = root; 21473fd0a558SYan, Zheng 21483fd0a558SYan, Zheng if (next != node) 21493fd0a558SYan, Zheng return NULL; 21503fd0a558SYan, Zheng 21513fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 21523fd0a558SYan, Zheng if (!next || next->level <= node->level) 21533fd0a558SYan, Zheng break; 21543fd0a558SYan, Zheng } 21553fd0a558SYan, Zheng 21563fd0a558SYan, Zheng if (!fs_root) 21573fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 21583fd0a558SYan, Zheng return fs_root; 21595d4f98a2SYan Zheng } 21605d4f98a2SYan Zheng 21615d4f98a2SYan Zheng static noinline_for_stack 21623fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2163a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 21645d4f98a2SYan Zheng { 21650b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2166a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2167a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2168a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 21693fd0a558SYan, Zheng u64 num_bytes = 0; 21703fd0a558SYan, Zheng int index = 0; 21715d4f98a2SYan Zheng 21723fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 21733fd0a558SYan, Zheng 21743fd0a558SYan, Zheng while (next) { 21753fd0a558SYan, Zheng cond_resched(); 21765d4f98a2SYan Zheng while (1) { 21773fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 21785d4f98a2SYan Zheng break; 21795d4f98a2SYan Zheng 21800b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 21813fd0a558SYan, Zheng 21823fd0a558SYan, Zheng if (list_empty(&next->upper)) 21833fd0a558SYan, Zheng break; 21843fd0a558SYan, Zheng 21853fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2186a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 21873fd0a558SYan, Zheng edges[index++] = edge; 21883fd0a558SYan, Zheng next = edge->node[UPPER]; 21895d4f98a2SYan Zheng } 21903fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 21913fd0a558SYan, Zheng } 21923fd0a558SYan, Zheng return num_bytes; 21933fd0a558SYan, Zheng } 21943fd0a558SYan, Zheng 21953fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 21963fd0a558SYan, Zheng struct reloc_control *rc, 2197a26195a5SQu Wenruo struct btrfs_backref_node *node) 21983fd0a558SYan, Zheng { 21993fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2200da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 22013fd0a558SYan, Zheng u64 num_bytes; 22023fd0a558SYan, Zheng int ret; 22030647bf56SWang Shilong u64 tmp; 22043fd0a558SYan, Zheng 22053fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 22063fd0a558SYan, Zheng 22073fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 22080647bf56SWang Shilong rc->reserved_bytes += num_bytes; 22098ca17f0fSJosef Bacik 22108ca17f0fSJosef Bacik /* 22118ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 22128ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 22138ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 22148ca17f0fSJosef Bacik */ 22150647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 22168ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 22173fd0a558SYan, Zheng if (ret) { 2218da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 22190647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 22200647bf56SWang Shilong tmp <<= 1; 22210647bf56SWang Shilong /* 22220647bf56SWang Shilong * only one thread can access block_rsv at this point, 22230647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 22240647bf56SWang Shilong * we expand more reservation size here to allow enough 222552042d8eSAndrea Gelmini * space for relocation and we will return earlier in 22260647bf56SWang Shilong * enospc case. 22270647bf56SWang Shilong */ 2228da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 22290647bf56SWang Shilong RELOCATION_RESERVED_NODES; 22308ca17f0fSJosef Bacik return -EAGAIN; 22313fd0a558SYan, Zheng } 22323fd0a558SYan, Zheng 22333fd0a558SYan, Zheng return 0; 22343fd0a558SYan, Zheng } 22353fd0a558SYan, Zheng 22365d4f98a2SYan Zheng /* 22375d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 22385d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 22395d4f98a2SYan Zheng * 22405d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 22415d4f98a2SYan Zheng * in that case this function just updates pointers. 22425d4f98a2SYan Zheng */ 22435d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 22443fd0a558SYan, Zheng struct reloc_control *rc, 2245a26195a5SQu Wenruo struct btrfs_backref_node *node, 22465d4f98a2SYan Zheng struct btrfs_key *key, 22475d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 22485d4f98a2SYan Zheng { 2249a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2250a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2251a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 22525d4f98a2SYan Zheng struct btrfs_root *root; 22535d4f98a2SYan Zheng struct extent_buffer *eb; 22545d4f98a2SYan Zheng u32 blocksize; 22555d4f98a2SYan Zheng u64 bytenr; 22565d4f98a2SYan Zheng int slot; 22578df01fddSNikolay Borisov int ret = 0; 22585d4f98a2SYan Zheng 2259ffe30dd8SJosef Bacik /* 2260ffe30dd8SJosef Bacik * If we are lowest then this is the first time we're processing this 2261ffe30dd8SJosef Bacik * block, and thus shouldn't have an eb associated with it yet. 2262ffe30dd8SJosef Bacik */ 2263ffe30dd8SJosef Bacik ASSERT(!lowest || !node->eb); 22645d4f98a2SYan Zheng 22655d4f98a2SYan Zheng path->lowest_level = node->level + 1; 22663fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 22675d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 226882fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2269581c1760SQu Wenruo 22705d4f98a2SYan Zheng cond_resched(); 22715d4f98a2SYan Zheng 22725d4f98a2SYan Zheng upper = edge->node[UPPER]; 2273dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 2274cbdc2ebcSJosef Bacik if (IS_ERR(root)) { 2275cbdc2ebcSJosef Bacik ret = PTR_ERR(root); 2276cbdc2ebcSJosef Bacik goto next; 2277cbdc2ebcSJosef Bacik } 22785d4f98a2SYan Zheng 22793fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 22803fd0a558SYan, Zheng if (!lowest) { 2281e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 22828df01fddSNikolay Borisov if (ret < 0) 2283cbca7d59SFilipe Manana goto next; 22843fd0a558SYan, Zheng BUG_ON(ret); 22853fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 22863fd0a558SYan, Zheng if (node->eb->start == bytenr) 22873fd0a558SYan, Zheng goto next; 22883fd0a558SYan, Zheng } 2289b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 22903fd0a558SYan, Zheng } 22915d4f98a2SYan Zheng 22925d4f98a2SYan Zheng if (!upper->eb) { 22935d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 22943561b9dbSLiu Bo if (ret) { 22958df01fddSNikolay Borisov if (ret > 0) 22968df01fddSNikolay Borisov ret = -ENOENT; 22973561b9dbSLiu Bo 22983561b9dbSLiu Bo btrfs_release_path(path); 22995d4f98a2SYan Zheng break; 23005d4f98a2SYan Zheng } 23015d4f98a2SYan Zheng 23023fd0a558SYan, Zheng if (!upper->eb) { 23033fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 23043fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 23053fd0a558SYan, Zheng } else { 23063fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 23073fd0a558SYan, Zheng } 23083fd0a558SYan, Zheng 23093fd0a558SYan, Zheng upper->locked = 1; 23103fd0a558SYan, Zheng path->locks[upper->level] = 0; 23113fd0a558SYan, Zheng 23125d4f98a2SYan Zheng slot = path->slots[upper->level]; 2313b3b4aa74SDavid Sterba btrfs_release_path(path); 23145d4f98a2SYan Zheng } else { 2315e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 23168df01fddSNikolay Borisov if (ret < 0) 2317cbca7d59SFilipe Manana goto next; 23185d4f98a2SYan Zheng BUG_ON(ret); 23195d4f98a2SYan Zheng } 23205d4f98a2SYan Zheng 23215d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 23223fd0a558SYan, Zheng if (lowest) { 23234547f4d8SLiu Bo if (bytenr != node->bytenr) { 23244547f4d8SLiu Bo btrfs_err(root->fs_info, 23254547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 23264547f4d8SLiu Bo bytenr, node->bytenr, slot, 23274547f4d8SLiu Bo upper->eb->start); 23288df01fddSNikolay Borisov ret = -EIO; 23294547f4d8SLiu Bo goto next; 23304547f4d8SLiu Bo } 23315d4f98a2SYan Zheng } else { 23323fd0a558SYan, Zheng if (node->eb->start == bytenr) 23333fd0a558SYan, Zheng goto next; 23345d4f98a2SYan Zheng } 23355d4f98a2SYan Zheng 2336da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 2337c9752536SJosef Bacik eb = btrfs_read_node_slot(upper->eb, slot); 233864c043deSLiu Bo if (IS_ERR(eb)) { 23398df01fddSNikolay Borisov ret = PTR_ERR(eb); 234064c043deSLiu Bo goto next; 234197d9a8a4STsutomu Itoh } 23425d4f98a2SYan Zheng btrfs_tree_lock(eb); 23435d4f98a2SYan Zheng 23445d4f98a2SYan Zheng if (!node->eb) { 23455d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 23469631e4ccSJosef Bacik slot, &eb, BTRFS_NESTING_COW); 23473fd0a558SYan, Zheng btrfs_tree_unlock(eb); 23483fd0a558SYan, Zheng free_extent_buffer(eb); 23498df01fddSNikolay Borisov if (ret < 0) 23503fd0a558SYan, Zheng goto next; 2351ffe30dd8SJosef Bacik /* 2352ffe30dd8SJosef Bacik * We've just COWed this block, it should have updated 2353ffe30dd8SJosef Bacik * the correct backref node entry. 2354ffe30dd8SJosef Bacik */ 2355ffe30dd8SJosef Bacik ASSERT(node->eb == eb); 23565d4f98a2SYan Zheng } else { 23575d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 23585d4f98a2SYan Zheng node->eb->start); 23595d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 23605d4f98a2SYan Zheng trans->transid); 23615d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 23625d4f98a2SYan Zheng 236382fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 23645d4f98a2SYan Zheng node->eb->start, blocksize, 236582fa113fSQu Wenruo upper->eb->start); 236682fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 236782fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 236882fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 236982fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 23705d4f98a2SYan Zheng BUG_ON(ret); 23715d4f98a2SYan Zheng 23725d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 23735d4f98a2SYan Zheng BUG_ON(ret); 23745d4f98a2SYan Zheng } 23753fd0a558SYan, Zheng next: 23763fd0a558SYan, Zheng if (!upper->pending) 2377b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 23783fd0a558SYan, Zheng else 2379b0fe7078SQu Wenruo btrfs_backref_unlock_node_buffer(upper); 23808df01fddSNikolay Borisov if (ret) 23813fd0a558SYan, Zheng break; 23825d4f98a2SYan Zheng } 23833fd0a558SYan, Zheng 23848df01fddSNikolay Borisov if (!ret && node->pending) { 2385b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 23863fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 23873fd0a558SYan, Zheng node->pending = 0; 23885d4f98a2SYan Zheng } 23893fd0a558SYan, Zheng 23905d4f98a2SYan Zheng path->lowest_level = 0; 2391ffe30dd8SJosef Bacik 2392ffe30dd8SJosef Bacik /* 2393ffe30dd8SJosef Bacik * We should have allocated all of our space in the block rsv and thus 2394ffe30dd8SJosef Bacik * shouldn't ENOSPC. 2395ffe30dd8SJosef Bacik */ 2396ffe30dd8SJosef Bacik ASSERT(ret != -ENOSPC); 23978df01fddSNikolay Borisov return ret; 23985d4f98a2SYan Zheng } 23995d4f98a2SYan Zheng 24005d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 24013fd0a558SYan, Zheng struct reloc_control *rc, 2402a26195a5SQu Wenruo struct btrfs_backref_node *node, 24035d4f98a2SYan Zheng struct btrfs_path *path) 24045d4f98a2SYan Zheng { 24055d4f98a2SYan Zheng struct btrfs_key key; 24065d4f98a2SYan Zheng 24075d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 24083fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 24095d4f98a2SYan Zheng } 24105d4f98a2SYan Zheng 24115d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 24123fd0a558SYan, Zheng struct reloc_control *rc, 24133fd0a558SYan, Zheng struct btrfs_path *path, int err) 24145d4f98a2SYan Zheng { 24153fd0a558SYan, Zheng LIST_HEAD(list); 2416a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 2417a26195a5SQu Wenruo struct btrfs_backref_node *node; 24185d4f98a2SYan Zheng int level; 24195d4f98a2SYan Zheng int ret; 24205d4f98a2SYan Zheng 24215d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 24225d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 24235d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 2424a26195a5SQu Wenruo struct btrfs_backref_node, list); 24253fd0a558SYan, Zheng list_move_tail(&node->list, &list); 24263fd0a558SYan, Zheng BUG_ON(!node->pending); 24275d4f98a2SYan Zheng 24283fd0a558SYan, Zheng if (!err) { 24293fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 24305d4f98a2SYan Zheng if (ret < 0) 24315d4f98a2SYan Zheng err = ret; 24325d4f98a2SYan Zheng } 24335d4f98a2SYan Zheng } 24343fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 24353fd0a558SYan, Zheng } 24365d4f98a2SYan Zheng return err; 24375d4f98a2SYan Zheng } 24385d4f98a2SYan Zheng 24395d4f98a2SYan Zheng /* 24405d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 24415d4f98a2SYan Zheng * as processed. 24425d4f98a2SYan Zheng */ 24435d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 2444a26195a5SQu Wenruo struct btrfs_backref_node *node) 24455d4f98a2SYan Zheng { 2446a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2447a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2448a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24495d4f98a2SYan Zheng int index = 0; 24505d4f98a2SYan Zheng 24515d4f98a2SYan Zheng while (next) { 24525d4f98a2SYan Zheng cond_resched(); 24535d4f98a2SYan Zheng while (1) { 24545d4f98a2SYan Zheng if (next->processed) 24555d4f98a2SYan Zheng break; 24565d4f98a2SYan Zheng 24579569cc20SQu Wenruo mark_block_processed(rc, next); 24585d4f98a2SYan Zheng 24595d4f98a2SYan Zheng if (list_empty(&next->upper)) 24605d4f98a2SYan Zheng break; 24615d4f98a2SYan Zheng 24625d4f98a2SYan Zheng edge = list_entry(next->upper.next, 2463a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 24645d4f98a2SYan Zheng edges[index++] = edge; 24655d4f98a2SYan Zheng next = edge->node[UPPER]; 24665d4f98a2SYan Zheng } 24675d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 24685d4f98a2SYan Zheng } 24695d4f98a2SYan Zheng } 24705d4f98a2SYan Zheng 24717476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 24725d4f98a2SYan Zheng { 2473da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 24747476dfdaSDavid Sterba 24755d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 24769655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 24775d4f98a2SYan Zheng return 1; 24785d4f98a2SYan Zheng return 0; 24795d4f98a2SYan Zheng } 24805d4f98a2SYan Zheng 24812ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 24825d4f98a2SYan Zheng struct tree_block *block) 24835d4f98a2SYan Zheng { 24845d4f98a2SYan Zheng struct extent_buffer *eb; 24855d4f98a2SYan Zheng 2486f7ba2d37SJosef Bacik eb = read_tree_block(fs_info, block->bytenr, block->owner, 2487f7ba2d37SJosef Bacik block->key.offset, block->level, NULL); 248864c043deSLiu Bo if (IS_ERR(eb)) { 248964c043deSLiu Bo return PTR_ERR(eb); 249064c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2491416bc658SJosef Bacik free_extent_buffer(eb); 2492416bc658SJosef Bacik return -EIO; 2493416bc658SJosef Bacik } 24945d4f98a2SYan Zheng if (block->level == 0) 24955d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 24965d4f98a2SYan Zheng else 24975d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 24985d4f98a2SYan Zheng free_extent_buffer(eb); 24995d4f98a2SYan Zheng block->key_ready = 1; 25005d4f98a2SYan Zheng return 0; 25015d4f98a2SYan Zheng } 25025d4f98a2SYan Zheng 25035d4f98a2SYan Zheng /* 25045d4f98a2SYan Zheng * helper function to relocate a tree block 25055d4f98a2SYan Zheng */ 25065d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 25075d4f98a2SYan Zheng struct reloc_control *rc, 2508a26195a5SQu Wenruo struct btrfs_backref_node *node, 25095d4f98a2SYan Zheng struct btrfs_key *key, 25105d4f98a2SYan Zheng struct btrfs_path *path) 25115d4f98a2SYan Zheng { 25125d4f98a2SYan Zheng struct btrfs_root *root; 25133fd0a558SYan, Zheng int ret = 0; 25145d4f98a2SYan Zheng 25153fd0a558SYan, Zheng if (!node) 25165d4f98a2SYan Zheng return 0; 25173fd0a558SYan, Zheng 25185f6b2e5cSJosef Bacik /* 25195f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 25205f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 25215f6b2e5cSJosef Bacik */ 25225f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 25235f6b2e5cSJosef Bacik if (ret) 25245f6b2e5cSJosef Bacik goto out; 25255f6b2e5cSJosef Bacik 25263fd0a558SYan, Zheng BUG_ON(node->processed); 2527147d256eSZhaolei root = select_one_root(node); 25283fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 25293fd0a558SYan, Zheng update_processed_blocks(rc, node); 25303fd0a558SYan, Zheng goto out; 25315d4f98a2SYan Zheng } 25325d4f98a2SYan Zheng 25333fd0a558SYan, Zheng if (root) { 253492a7cc42SQu Wenruo if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 25351c7bfa15SJosef Bacik /* 25361c7bfa15SJosef Bacik * This block was the root block of a root, and this is 25371c7bfa15SJosef Bacik * the first time we're processing the block and thus it 25381c7bfa15SJosef Bacik * should not have had the ->new_bytenr modified and 25391c7bfa15SJosef Bacik * should have not been included on the changed list. 25401c7bfa15SJosef Bacik * 25411c7bfa15SJosef Bacik * However in the case of corruption we could have 25421c7bfa15SJosef Bacik * multiple refs pointing to the same block improperly, 25431c7bfa15SJosef Bacik * and thus we would trip over these checks. ASSERT() 25441c7bfa15SJosef Bacik * for the developer case, because it could indicate a 25451c7bfa15SJosef Bacik * bug in the backref code, however error out for a 25461c7bfa15SJosef Bacik * normal user in the case of corruption. 25471c7bfa15SJosef Bacik */ 25481c7bfa15SJosef Bacik ASSERT(node->new_bytenr == 0); 25491c7bfa15SJosef Bacik ASSERT(list_empty(&node->list)); 25501c7bfa15SJosef Bacik if (node->new_bytenr || !list_empty(&node->list)) { 25511c7bfa15SJosef Bacik btrfs_err(root->fs_info, 25521c7bfa15SJosef Bacik "bytenr %llu has improper references to it", 25531c7bfa15SJosef Bacik node->bytenr); 25541c7bfa15SJosef Bacik ret = -EUCLEAN; 25551c7bfa15SJosef Bacik goto out; 25561c7bfa15SJosef Bacik } 2557d18c7bd9SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 2558d18c7bd9SJosef Bacik if (ret) 2559d18c7bd9SJosef Bacik goto out; 25603fd0a558SYan, Zheng root = root->reloc_root; 25613fd0a558SYan, Zheng node->new_bytenr = root->node->start; 256200246528SJosef Bacik btrfs_put_root(node->root); 256300246528SJosef Bacik node->root = btrfs_grab_root(root); 25640b530bc5SJosef Bacik ASSERT(node->root); 25653fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 25663fd0a558SYan, Zheng } else { 25675d4f98a2SYan Zheng path->lowest_level = node->level; 25685d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2569b3b4aa74SDavid Sterba btrfs_release_path(path); 25703fd0a558SYan, Zheng if (ret > 0) 25715d4f98a2SYan Zheng ret = 0; 25723fd0a558SYan, Zheng } 25733fd0a558SYan, Zheng if (!ret) 25743fd0a558SYan, Zheng update_processed_blocks(rc, node); 25753fd0a558SYan, Zheng } else { 25763fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 25773fd0a558SYan, Zheng } 25785d4f98a2SYan Zheng out: 25790647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 2580023acb07SQu Wenruo btrfs_backref_cleanup_node(&rc->backref_cache, node); 25815d4f98a2SYan Zheng return ret; 25825d4f98a2SYan Zheng } 25835d4f98a2SYan Zheng 25845d4f98a2SYan Zheng /* 25855d4f98a2SYan Zheng * relocate a list of blocks 25865d4f98a2SYan Zheng */ 25875d4f98a2SYan Zheng static noinline_for_stack 25885d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 25895d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 25905d4f98a2SYan Zheng { 25912ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2592a26195a5SQu Wenruo struct btrfs_backref_node *node; 25935d4f98a2SYan Zheng struct btrfs_path *path; 25945d4f98a2SYan Zheng struct tree_block *block; 259598ff7b94SQu Wenruo struct tree_block *next; 25965d4f98a2SYan Zheng int ret; 25975d4f98a2SYan Zheng int err = 0; 25985d4f98a2SYan Zheng 25995d4f98a2SYan Zheng path = btrfs_alloc_path(); 2600e1a12670SLiu Bo if (!path) { 2601e1a12670SLiu Bo err = -ENOMEM; 260234c2b290SDavid Sterba goto out_free_blocks; 2603e1a12670SLiu Bo } 26045d4f98a2SYan Zheng 260598ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 260698ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 26075d4f98a2SYan Zheng if (!block->key_ready) 2608f7ba2d37SJosef Bacik btrfs_readahead_tree_block(fs_info, block->bytenr, 2609f7ba2d37SJosef Bacik block->owner, 0, 26103fbaf258SJosef Bacik block->level); 26115d4f98a2SYan Zheng } 26125d4f98a2SYan Zheng 261398ff7b94SQu Wenruo /* Get first keys */ 261498ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 261534c2b290SDavid Sterba if (!block->key_ready) { 26162ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 261734c2b290SDavid Sterba if (err) 261834c2b290SDavid Sterba goto out_free_path; 261934c2b290SDavid Sterba } 26205d4f98a2SYan Zheng } 26215d4f98a2SYan Zheng 262298ff7b94SQu Wenruo /* Do tree relocation */ 262398ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 26243fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 26255d4f98a2SYan Zheng block->level, block->bytenr); 26265d4f98a2SYan Zheng if (IS_ERR(node)) { 26275d4f98a2SYan Zheng err = PTR_ERR(node); 26285d4f98a2SYan Zheng goto out; 26295d4f98a2SYan Zheng } 26305d4f98a2SYan Zheng 26315d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 26325d4f98a2SYan Zheng path); 26335d4f98a2SYan Zheng if (ret < 0) { 26345d4f98a2SYan Zheng err = ret; 263550dbbb71SJosef Bacik break; 26365d4f98a2SYan Zheng } 26375d4f98a2SYan Zheng } 26385d4f98a2SYan Zheng out: 26393fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 26405d4f98a2SYan Zheng 264134c2b290SDavid Sterba out_free_path: 26425d4f98a2SYan Zheng btrfs_free_path(path); 264334c2b290SDavid Sterba out_free_blocks: 2644e1a12670SLiu Bo free_block_list(blocks); 26455d4f98a2SYan Zheng return err; 26465d4f98a2SYan Zheng } 26475d4f98a2SYan Zheng 2648056d9becSNikolay Borisov static noinline_for_stack int prealloc_file_extent_cluster( 2649056d9becSNikolay Borisov struct btrfs_inode *inode, 2650efa56464SYan, Zheng struct file_extent_cluster *cluster) 2651efa56464SYan, Zheng { 2652efa56464SYan, Zheng u64 alloc_hint = 0; 2653efa56464SYan, Zheng u64 start; 2654efa56464SYan, Zheng u64 end; 2655056d9becSNikolay Borisov u64 offset = inode->index_cnt; 2656efa56464SYan, Zheng u64 num_bytes; 26574e9d0d01SNikolay Borisov int nr; 2658efa56464SYan, Zheng int ret = 0; 2659dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 2660dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 2661214e61d0SNikolay Borisov u64 cur_offset = prealloc_start; 2662efa56464SYan, Zheng 2663efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2664056d9becSNikolay Borisov ret = btrfs_alloc_data_chunk_ondemand(inode, 2665dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 2666efa56464SYan, Zheng if (ret) 2667214e61d0SNikolay Borisov return ret; 2668efa56464SYan, Zheng 266932430c61SNaohiro Aota /* 267032430c61SNaohiro Aota * On a zoned filesystem, we cannot preallocate the file region. 267132430c61SNaohiro Aota * Instead, we dirty and fiemap_write the region. 267232430c61SNaohiro Aota */ 267332430c61SNaohiro Aota if (btrfs_is_zoned(inode->root->fs_info)) { 267432430c61SNaohiro Aota struct btrfs_root *root = inode->root; 267532430c61SNaohiro Aota struct btrfs_trans_handle *trans; 267632430c61SNaohiro Aota 267732430c61SNaohiro Aota end = cluster->end - offset + 1; 267832430c61SNaohiro Aota trans = btrfs_start_transaction(root, 1); 267932430c61SNaohiro Aota if (IS_ERR(trans)) 268032430c61SNaohiro Aota return PTR_ERR(trans); 268132430c61SNaohiro Aota 268232430c61SNaohiro Aota inode->vfs_inode.i_ctime = current_time(&inode->vfs_inode); 268332430c61SNaohiro Aota i_size_write(&inode->vfs_inode, end); 268432430c61SNaohiro Aota ret = btrfs_update_inode(trans, root, inode); 268532430c61SNaohiro Aota if (ret) { 268632430c61SNaohiro Aota btrfs_abort_transaction(trans, ret); 268732430c61SNaohiro Aota btrfs_end_transaction(trans); 268832430c61SNaohiro Aota return ret; 268932430c61SNaohiro Aota } 269032430c61SNaohiro Aota 269132430c61SNaohiro Aota return btrfs_end_transaction(trans); 269232430c61SNaohiro Aota } 269332430c61SNaohiro Aota 269464708539SJosef Bacik btrfs_inode_lock(&inode->vfs_inode, 0); 26954e9d0d01SNikolay Borisov for (nr = 0; nr < cluster->nr; nr++) { 2696efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2697efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2698efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2699efa56464SYan, Zheng else 2700efa56464SYan, Zheng end = cluster->end - offset; 2701efa56464SYan, Zheng 2702056d9becSNikolay Borisov lock_extent(&inode->io_tree, start, end); 2703efa56464SYan, Zheng num_bytes = end + 1 - start; 2704056d9becSNikolay Borisov ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start, 2705efa56464SYan, Zheng num_bytes, num_bytes, 2706efa56464SYan, Zheng end + 1, &alloc_hint); 270718513091SWang Xiaoguang cur_offset = end + 1; 2708056d9becSNikolay Borisov unlock_extent(&inode->io_tree, start, end); 2709efa56464SYan, Zheng if (ret) 2710efa56464SYan, Zheng break; 2711efa56464SYan, Zheng } 271264708539SJosef Bacik btrfs_inode_unlock(&inode->vfs_inode, 0); 2713214e61d0SNikolay Borisov 271418513091SWang Xiaoguang if (cur_offset < prealloc_end) 2715056d9becSNikolay Borisov btrfs_free_reserved_data_space_noquota(inode->root->fs_info, 2716a89ef455SFilipe Manana prealloc_end + 1 - cur_offset); 2717efa56464SYan, Zheng return ret; 2718efa56464SYan, Zheng } 2719efa56464SYan, Zheng 2720efa56464SYan, Zheng static noinline_for_stack 27210257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 27220257bb82SYan, Zheng u64 block_start) 27235d4f98a2SYan Zheng { 27245d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 27255d4f98a2SYan Zheng struct extent_map *em; 27260257bb82SYan, Zheng int ret = 0; 27275d4f98a2SYan Zheng 2728172ddd60SDavid Sterba em = alloc_extent_map(); 27290257bb82SYan, Zheng if (!em) 27300257bb82SYan, Zheng return -ENOMEM; 27310257bb82SYan, Zheng 27325d4f98a2SYan Zheng em->start = start; 27330257bb82SYan, Zheng em->len = end + 1 - start; 27340257bb82SYan, Zheng em->block_len = em->len; 27350257bb82SYan, Zheng em->block_start = block_start; 27365d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 27375d4f98a2SYan Zheng 2738d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 27395d4f98a2SYan Zheng while (1) { 2740890871beSChris Mason write_lock(&em_tree->lock); 274109a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 2742890871beSChris Mason write_unlock(&em_tree->lock); 27435d4f98a2SYan Zheng if (ret != -EEXIST) { 27445d4f98a2SYan Zheng free_extent_map(em); 27455d4f98a2SYan Zheng break; 27465d4f98a2SYan Zheng } 2747dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 27485d4f98a2SYan Zheng } 2749d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 27500257bb82SYan, Zheng return ret; 27510257bb82SYan, Zheng } 27525d4f98a2SYan Zheng 2753726a3421SQu Wenruo /* 2754726a3421SQu Wenruo * Allow error injection to test balance cancellation 2755726a3421SQu Wenruo */ 27561fec12a5SJosef Bacik noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 2757726a3421SQu Wenruo { 27585cb502f4SQu Wenruo return atomic_read(&fs_info->balance_cancel_req) || 27595cb502f4SQu Wenruo fatal_signal_pending(current); 2760726a3421SQu Wenruo } 2761726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 2762726a3421SQu Wenruo 27630257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 27640257bb82SYan, Zheng struct file_extent_cluster *cluster) 27650257bb82SYan, Zheng { 27662ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 27670257bb82SYan, Zheng u64 page_start; 27680257bb82SYan, Zheng u64 page_end; 27690257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 27700257bb82SYan, Zheng unsigned long index; 27710257bb82SYan, Zheng unsigned long last_index; 27720257bb82SYan, Zheng struct page *page; 27730257bb82SYan, Zheng struct file_ra_state *ra; 27743b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 27750257bb82SYan, Zheng int nr = 0; 27760257bb82SYan, Zheng int ret = 0; 27770257bb82SYan, Zheng 27780257bb82SYan, Zheng if (!cluster->nr) 27790257bb82SYan, Zheng return 0; 27800257bb82SYan, Zheng 27810257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 27820257bb82SYan, Zheng if (!ra) 27830257bb82SYan, Zheng return -ENOMEM; 27840257bb82SYan, Zheng 2785056d9becSNikolay Borisov ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster); 27860257bb82SYan, Zheng if (ret) 2787efa56464SYan, Zheng goto out; 27880257bb82SYan, Zheng 27890257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 27900257bb82SYan, Zheng 2791efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2792efa56464SYan, Zheng cluster->end - offset, cluster->start); 2793efa56464SYan, Zheng if (ret) 2794efa56464SYan, Zheng goto out; 2795efa56464SYan, Zheng 279609cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 279709cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 27980257bb82SYan, Zheng while (index <= last_index) { 27999f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 28009f3db423SNikolay Borisov PAGE_SIZE); 2801efa56464SYan, Zheng if (ret) 2802efa56464SYan, Zheng goto out; 2803efa56464SYan, Zheng 28040257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 28050257bb82SYan, Zheng if (!page) { 28060257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 28070257bb82SYan, Zheng ra, NULL, index, 28080257bb82SYan, Zheng last_index + 1 - index); 2809a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 28103b16a4e3SJosef Bacik mask); 28110257bb82SYan, Zheng if (!page) { 2812691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 281343b18595SQu Wenruo PAGE_SIZE, true); 281444db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 28158702ba93SQu Wenruo PAGE_SIZE); 28160257bb82SYan, Zheng ret = -ENOMEM; 2817efa56464SYan, Zheng goto out; 28180257bb82SYan, Zheng } 28190257bb82SYan, Zheng } 282032443de3SQu Wenruo ret = set_page_extent_mapped(page); 282132443de3SQu Wenruo if (ret < 0) { 282232443de3SQu Wenruo btrfs_delalloc_release_metadata(BTRFS_I(inode), 282332443de3SQu Wenruo PAGE_SIZE, true); 282432443de3SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 282532443de3SQu Wenruo unlock_page(page); 282632443de3SQu Wenruo put_page(page); 282732443de3SQu Wenruo goto out; 282832443de3SQu Wenruo } 28290257bb82SYan, Zheng 28300257bb82SYan, Zheng if (PageReadahead(page)) { 28310257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 28320257bb82SYan, Zheng ra, NULL, page, index, 28330257bb82SYan, Zheng last_index + 1 - index); 28340257bb82SYan, Zheng } 28350257bb82SYan, Zheng 28360257bb82SYan, Zheng if (!PageUptodate(page)) { 28370257bb82SYan, Zheng btrfs_readpage(NULL, page); 28380257bb82SYan, Zheng lock_page(page); 28390257bb82SYan, Zheng if (!PageUptodate(page)) { 28400257bb82SYan, Zheng unlock_page(page); 284109cbfeafSKirill A. Shutemov put_page(page); 2842691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 284343b18595SQu Wenruo PAGE_SIZE, true); 28448b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 28458702ba93SQu Wenruo PAGE_SIZE); 28460257bb82SYan, Zheng ret = -EIO; 2847efa56464SYan, Zheng goto out; 28480257bb82SYan, Zheng } 28490257bb82SYan, Zheng } 28500257bb82SYan, Zheng 28514eee4fa4SMiao Xie page_start = page_offset(page); 285209cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 28530257bb82SYan, Zheng 2854d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 28550257bb82SYan, Zheng 28560257bb82SYan, Zheng if (nr < cluster->nr && 28570257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 28580257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 28590257bb82SYan, Zheng page_start, page_end, 2860ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 28610257bb82SYan, Zheng nr++; 28620257bb82SYan, Zheng } 28630257bb82SYan, Zheng 2864c2566f22SNikolay Borisov ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, 2865c2566f22SNikolay Borisov page_end, 0, NULL); 2866765f3cebSNikolay Borisov if (ret) { 2867765f3cebSNikolay Borisov unlock_page(page); 2868765f3cebSNikolay Borisov put_page(page); 2869765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 287043b18595SQu Wenruo PAGE_SIZE, true); 2871765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 28728702ba93SQu Wenruo PAGE_SIZE); 2873765f3cebSNikolay Borisov 2874765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 2875765f3cebSNikolay Borisov page_start, page_end, 2876765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 2877765f3cebSNikolay Borisov goto out; 2878765f3cebSNikolay Borisov 2879765f3cebSNikolay Borisov } 28800257bb82SYan, Zheng set_page_dirty(page); 28810257bb82SYan, Zheng 28820257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 2883d0082371SJeff Mahoney page_start, page_end); 28840257bb82SYan, Zheng unlock_page(page); 288509cbfeafSKirill A. Shutemov put_page(page); 28860257bb82SYan, Zheng 28870257bb82SYan, Zheng index++; 28888702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 2889efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 28902ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 28917f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 28927f913c7cSQu Wenruo ret = -ECANCELED; 28937f913c7cSQu Wenruo goto out; 28947f913c7cSQu Wenruo } 28950257bb82SYan, Zheng } 28960257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 289732430c61SNaohiro Aota if (btrfs_is_zoned(fs_info) && !ret) 289832430c61SNaohiro Aota ret = btrfs_wait_ordered_range(inode, 0, (u64)-1); 2899efa56464SYan, Zheng out: 29000257bb82SYan, Zheng kfree(ra); 29010257bb82SYan, Zheng return ret; 29020257bb82SYan, Zheng } 29030257bb82SYan, Zheng 29040257bb82SYan, Zheng static noinline_for_stack 29050257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 29060257bb82SYan, Zheng struct file_extent_cluster *cluster) 29070257bb82SYan, Zheng { 29080257bb82SYan, Zheng int ret; 29090257bb82SYan, Zheng 29100257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 29110257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 29120257bb82SYan, Zheng if (ret) 29130257bb82SYan, Zheng return ret; 29140257bb82SYan, Zheng cluster->nr = 0; 29150257bb82SYan, Zheng } 29160257bb82SYan, Zheng 29170257bb82SYan, Zheng if (!cluster->nr) 29180257bb82SYan, Zheng cluster->start = extent_key->objectid; 29190257bb82SYan, Zheng else 29200257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 29210257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 29220257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 29230257bb82SYan, Zheng cluster->nr++; 29240257bb82SYan, Zheng 29250257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 29260257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 29270257bb82SYan, Zheng if (ret) 29280257bb82SYan, Zheng return ret; 29290257bb82SYan, Zheng cluster->nr = 0; 29300257bb82SYan, Zheng } 29310257bb82SYan, Zheng return 0; 29325d4f98a2SYan Zheng } 29335d4f98a2SYan Zheng 29345d4f98a2SYan Zheng /* 29355d4f98a2SYan Zheng * helper to add a tree block to the list. 29365d4f98a2SYan Zheng * the major work is getting the generation and level of the block 29375d4f98a2SYan Zheng */ 29385d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 29395d4f98a2SYan Zheng struct btrfs_key *extent_key, 29405d4f98a2SYan Zheng struct btrfs_path *path, 29415d4f98a2SYan Zheng struct rb_root *blocks) 29425d4f98a2SYan Zheng { 29435d4f98a2SYan Zheng struct extent_buffer *eb; 29445d4f98a2SYan Zheng struct btrfs_extent_item *ei; 29455d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 29465d4f98a2SYan Zheng struct tree_block *block; 29475d4f98a2SYan Zheng struct rb_node *rb_node; 29485d4f98a2SYan Zheng u32 item_size; 29495d4f98a2SYan Zheng int level = -1; 29507fdf4b60SWang Shilong u64 generation; 2951f7ba2d37SJosef Bacik u64 owner = 0; 29525d4f98a2SYan Zheng 29535d4f98a2SYan Zheng eb = path->nodes[0]; 29545d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 29555d4f98a2SYan Zheng 29563173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 29573173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 2958f7ba2d37SJosef Bacik unsigned long ptr = 0, end; 2959f7ba2d37SJosef Bacik 29605d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 29615d4f98a2SYan Zheng struct btrfs_extent_item); 2962f7ba2d37SJosef Bacik end = (unsigned long)ei + item_size; 29633173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 29645d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 29655d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 2966f7ba2d37SJosef Bacik ptr = (unsigned long)(bi + 1); 29675d4f98a2SYan Zheng } else { 29683173a18fSJosef Bacik level = (int)extent_key->offset; 2969f7ba2d37SJosef Bacik ptr = (unsigned long)(ei + 1); 29703173a18fSJosef Bacik } 29713173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 2972f7ba2d37SJosef Bacik 2973f7ba2d37SJosef Bacik /* 2974f7ba2d37SJosef Bacik * We're reading random blocks without knowing their owner ahead 2975f7ba2d37SJosef Bacik * of time. This is ok most of the time, as all reloc roots and 2976f7ba2d37SJosef Bacik * fs roots have the same lock type. However normal trees do 2977f7ba2d37SJosef Bacik * not, and the only way to know ahead of time is to read the 2978f7ba2d37SJosef Bacik * inline ref offset. We know it's an fs root if 2979f7ba2d37SJosef Bacik * 2980f7ba2d37SJosef Bacik * 1. There's more than one ref. 2981f7ba2d37SJosef Bacik * 2. There's a SHARED_DATA_REF_KEY set. 2982f7ba2d37SJosef Bacik * 3. FULL_BACKREF is set on the flags. 2983f7ba2d37SJosef Bacik * 2984f7ba2d37SJosef Bacik * Otherwise it's safe to assume that the ref offset == the 2985f7ba2d37SJosef Bacik * owner of this block, so we can use that when calling 2986f7ba2d37SJosef Bacik * read_tree_block. 2987f7ba2d37SJosef Bacik */ 2988f7ba2d37SJosef Bacik if (btrfs_extent_refs(eb, ei) == 1 && 2989f7ba2d37SJosef Bacik !(btrfs_extent_flags(eb, ei) & 2990f7ba2d37SJosef Bacik BTRFS_BLOCK_FLAG_FULL_BACKREF) && 2991f7ba2d37SJosef Bacik ptr < end) { 2992f7ba2d37SJosef Bacik struct btrfs_extent_inline_ref *iref; 2993f7ba2d37SJosef Bacik int type; 2994f7ba2d37SJosef Bacik 2995f7ba2d37SJosef Bacik iref = (struct btrfs_extent_inline_ref *)ptr; 2996f7ba2d37SJosef Bacik type = btrfs_get_extent_inline_ref_type(eb, iref, 2997f7ba2d37SJosef Bacik BTRFS_REF_TYPE_BLOCK); 2998f7ba2d37SJosef Bacik if (type == BTRFS_REF_TYPE_INVALID) 2999f7ba2d37SJosef Bacik return -EINVAL; 3000f7ba2d37SJosef Bacik if (type == BTRFS_TREE_BLOCK_REF_KEY) 3001f7ba2d37SJosef Bacik owner = btrfs_extent_inline_ref_offset(eb, iref); 3002f7ba2d37SJosef Bacik } 30036d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3004ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3005ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3006ba3c2b19SNikolay Borisov return -EINVAL; 30073173a18fSJosef Bacik } else { 30085d4f98a2SYan Zheng BUG(); 30095d4f98a2SYan Zheng } 30105d4f98a2SYan Zheng 3011b3b4aa74SDavid Sterba btrfs_release_path(path); 30125d4f98a2SYan Zheng 30135d4f98a2SYan Zheng BUG_ON(level == -1); 30145d4f98a2SYan Zheng 30155d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 30165d4f98a2SYan Zheng if (!block) 30175d4f98a2SYan Zheng return -ENOMEM; 30185d4f98a2SYan Zheng 30195d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3020da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 30215d4f98a2SYan Zheng block->key.offset = generation; 30225d4f98a2SYan Zheng block->level = level; 30235d4f98a2SYan Zheng block->key_ready = 0; 3024f7ba2d37SJosef Bacik block->owner = owner; 30255d4f98a2SYan Zheng 3026e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 302743c04fb1SJeff Mahoney if (rb_node) 3028982c92cbSQu Wenruo btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr, 3029982c92cbSQu Wenruo -EEXIST); 30305d4f98a2SYan Zheng 30315d4f98a2SYan Zheng return 0; 30325d4f98a2SYan Zheng } 30335d4f98a2SYan Zheng 30345d4f98a2SYan Zheng /* 30355d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 30365d4f98a2SYan Zheng */ 30375d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 30385d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 30395d4f98a2SYan Zheng struct rb_root *blocks) 30405d4f98a2SYan Zheng { 30410b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 30425d4f98a2SYan Zheng struct btrfs_path *path; 30435d4f98a2SYan Zheng struct btrfs_key key; 30445d4f98a2SYan Zheng int ret; 30450b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 30465d4f98a2SYan Zheng 30477476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 30485d4f98a2SYan Zheng return 0; 30495d4f98a2SYan Zheng 3050e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 30515d4f98a2SYan Zheng return 0; 30525d4f98a2SYan Zheng 30535d4f98a2SYan Zheng path = btrfs_alloc_path(); 30545d4f98a2SYan Zheng if (!path) 30555d4f98a2SYan Zheng return -ENOMEM; 3056aee68ee5SJosef Bacik again: 30575d4f98a2SYan Zheng key.objectid = bytenr; 3058aee68ee5SJosef Bacik if (skinny) { 3059aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3060aee68ee5SJosef Bacik key.offset = (u64)-1; 3061aee68ee5SJosef Bacik } else { 30625d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 30635d4f98a2SYan Zheng key.offset = blocksize; 3064aee68ee5SJosef Bacik } 30655d4f98a2SYan Zheng 30665d4f98a2SYan Zheng path->search_commit_root = 1; 30675d4f98a2SYan Zheng path->skip_locking = 1; 30685d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 30695d4f98a2SYan Zheng if (ret < 0) 30705d4f98a2SYan Zheng goto out; 30715d4f98a2SYan Zheng 3072aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3073aee68ee5SJosef Bacik if (path->slots[0]) { 3074aee68ee5SJosef Bacik path->slots[0]--; 3075aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3076aee68ee5SJosef Bacik path->slots[0]); 30773173a18fSJosef Bacik if (key.objectid == bytenr && 3078aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3079aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3080aee68ee5SJosef Bacik key.offset == blocksize))) 30813173a18fSJosef Bacik ret = 0; 30823173a18fSJosef Bacik } 3083aee68ee5SJosef Bacik 3084aee68ee5SJosef Bacik if (ret) { 3085aee68ee5SJosef Bacik skinny = false; 3086aee68ee5SJosef Bacik btrfs_release_path(path); 3087aee68ee5SJosef Bacik goto again; 3088aee68ee5SJosef Bacik } 3089aee68ee5SJosef Bacik } 3090cdccee99SLiu Bo if (ret) { 3091cdccee99SLiu Bo ASSERT(ret == 1); 3092cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3093cdccee99SLiu Bo btrfs_err(fs_info, 3094cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3095cdccee99SLiu Bo bytenr); 3096cdccee99SLiu Bo WARN_ON(1); 3097cdccee99SLiu Bo ret = -EINVAL; 3098cdccee99SLiu Bo goto out; 3099cdccee99SLiu Bo } 31003173a18fSJosef Bacik 31015d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 31025d4f98a2SYan Zheng out: 31035d4f98a2SYan Zheng btrfs_free_path(path); 31045d4f98a2SYan Zheng return ret; 31055d4f98a2SYan Zheng } 31065d4f98a2SYan Zheng 31070af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 310832da5386SDavid Sterba struct btrfs_block_group *block_group, 31091bbc621eSChris Mason struct inode *inode, 31101bbc621eSChris Mason u64 ino) 31110af3d00bSJosef Bacik { 31120af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 31130af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 31140af3d00bSJosef Bacik int ret = 0; 31150af3d00bSJosef Bacik 31160af3d00bSJosef Bacik if (inode) 31170af3d00bSJosef Bacik goto truncate; 31180af3d00bSJosef Bacik 31190202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, ino, root); 31202e19f1f9SAl Viro if (IS_ERR(inode)) 31210af3d00bSJosef Bacik return -ENOENT; 31220af3d00bSJosef Bacik 31230af3d00bSJosef Bacik truncate: 31242ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 31257b61cd92SMiao Xie &fs_info->global_block_rsv); 31267b61cd92SMiao Xie if (ret) 31277b61cd92SMiao Xie goto out; 31287b61cd92SMiao Xie 31297a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 31300af3d00bSJosef Bacik if (IS_ERR(trans)) { 31313612b495STsutomu Itoh ret = PTR_ERR(trans); 31320af3d00bSJosef Bacik goto out; 31330af3d00bSJosef Bacik } 31340af3d00bSJosef Bacik 313577ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 31360af3d00bSJosef Bacik 31373a45bb20SJeff Mahoney btrfs_end_transaction(trans); 31382ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 31390af3d00bSJosef Bacik out: 31400af3d00bSJosef Bacik iput(inode); 31410af3d00bSJosef Bacik return ret; 31420af3d00bSJosef Bacik } 31430af3d00bSJosef Bacik 31445d4f98a2SYan Zheng /* 314519b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 314619b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 31475d4f98a2SYan Zheng */ 314819b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 314919b546d7SQu Wenruo struct btrfs_block_group *block_group, 315019b546d7SQu Wenruo u64 data_bytenr) 31515d4f98a2SYan Zheng { 315219b546d7SQu Wenruo u64 space_cache_ino; 315319b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 31545d4f98a2SYan Zheng struct btrfs_key key; 315519b546d7SQu Wenruo bool found = false; 315619b546d7SQu Wenruo int i; 31575d4f98a2SYan Zheng int ret; 31585d4f98a2SYan Zheng 315919b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 316019b546d7SQu Wenruo return 0; 31615d4f98a2SYan Zheng 316219b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 316350e31ef4SQu Wenruo u8 type; 316450e31ef4SQu Wenruo 316519b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 316619b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 316719b546d7SQu Wenruo continue; 316819b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 316950e31ef4SQu Wenruo type = btrfs_file_extent_type(leaf, ei); 317050e31ef4SQu Wenruo 317150e31ef4SQu Wenruo if ((type == BTRFS_FILE_EXTENT_REG || 317250e31ef4SQu Wenruo type == BTRFS_FILE_EXTENT_PREALLOC) && 317319b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 317419b546d7SQu Wenruo found = true; 317519b546d7SQu Wenruo space_cache_ino = key.objectid; 317619b546d7SQu Wenruo break; 317719b546d7SQu Wenruo } 317819b546d7SQu Wenruo } 317919b546d7SQu Wenruo if (!found) 318019b546d7SQu Wenruo return -ENOENT; 318119b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 318219b546d7SQu Wenruo space_cache_ino); 31830af3d00bSJosef Bacik return ret; 31845d4f98a2SYan Zheng } 31855d4f98a2SYan Zheng 31865d4f98a2SYan Zheng /* 31872c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 31885d4f98a2SYan Zheng */ 31895d4f98a2SYan Zheng static noinline_for_stack 31905d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 31915d4f98a2SYan Zheng struct btrfs_key *extent_key, 31925d4f98a2SYan Zheng struct btrfs_path *path, 31935d4f98a2SYan Zheng struct rb_root *blocks) 31945d4f98a2SYan Zheng { 319519b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 319619b546d7SQu Wenruo struct ulist *leaves = NULL; 319719b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 319819b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 319919b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3200647f63bdSFilipe David Borba Manana int ret = 0; 32015d4f98a2SYan Zheng 3202b3b4aa74SDavid Sterba btrfs_release_path(path); 320319b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 320419b546d7SQu Wenruo 0, &leaves, NULL, true); 320519b546d7SQu Wenruo if (ret < 0) 320619b546d7SQu Wenruo return ret; 320719b546d7SQu Wenruo 320819b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 320919b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 321019b546d7SQu Wenruo struct extent_buffer *eb; 321119b546d7SQu Wenruo 32121b7ec85eSJosef Bacik eb = read_tree_block(fs_info, ref_node->val, 0, 0, 0, NULL); 321319b546d7SQu Wenruo if (IS_ERR(eb)) { 321419b546d7SQu Wenruo ret = PTR_ERR(eb); 321519b546d7SQu Wenruo break; 321619b546d7SQu Wenruo } 321719b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 321819b546d7SQu Wenruo extent_key->objectid); 321919b546d7SQu Wenruo free_extent_buffer(eb); 322019b546d7SQu Wenruo if (ret < 0) 322119b546d7SQu Wenruo break; 322219b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 322319b546d7SQu Wenruo if (ret < 0) 322419b546d7SQu Wenruo break; 322519b546d7SQu Wenruo } 322619b546d7SQu Wenruo if (ret < 0) 32275d4f98a2SYan Zheng free_block_list(blocks); 322819b546d7SQu Wenruo ulist_free(leaves); 322919b546d7SQu Wenruo return ret; 32305d4f98a2SYan Zheng } 32315d4f98a2SYan Zheng 32325d4f98a2SYan Zheng /* 32332c016dc2SLiu Bo * helper to find next unprocessed extent 32345d4f98a2SYan Zheng */ 32355d4f98a2SYan Zheng static noinline_for_stack 3236147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 32373fd0a558SYan, Zheng struct btrfs_key *extent_key) 32385d4f98a2SYan Zheng { 32390b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 32405d4f98a2SYan Zheng struct btrfs_key key; 32415d4f98a2SYan Zheng struct extent_buffer *leaf; 32425d4f98a2SYan Zheng u64 start, end, last; 32435d4f98a2SYan Zheng int ret; 32445d4f98a2SYan Zheng 3245b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 32465d4f98a2SYan Zheng while (1) { 32475d4f98a2SYan Zheng cond_resched(); 32485d4f98a2SYan Zheng if (rc->search_start >= last) { 32495d4f98a2SYan Zheng ret = 1; 32505d4f98a2SYan Zheng break; 32515d4f98a2SYan Zheng } 32525d4f98a2SYan Zheng 32535d4f98a2SYan Zheng key.objectid = rc->search_start; 32545d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 32555d4f98a2SYan Zheng key.offset = 0; 32565d4f98a2SYan Zheng 32575d4f98a2SYan Zheng path->search_commit_root = 1; 32585d4f98a2SYan Zheng path->skip_locking = 1; 32595d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 32605d4f98a2SYan Zheng 0, 0); 32615d4f98a2SYan Zheng if (ret < 0) 32625d4f98a2SYan Zheng break; 32635d4f98a2SYan Zheng next: 32645d4f98a2SYan Zheng leaf = path->nodes[0]; 32655d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 32665d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 32675d4f98a2SYan Zheng if (ret != 0) 32685d4f98a2SYan Zheng break; 32695d4f98a2SYan Zheng leaf = path->nodes[0]; 32705d4f98a2SYan Zheng } 32715d4f98a2SYan Zheng 32725d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 32735d4f98a2SYan Zheng if (key.objectid >= last) { 32745d4f98a2SYan Zheng ret = 1; 32755d4f98a2SYan Zheng break; 32765d4f98a2SYan Zheng } 32775d4f98a2SYan Zheng 32783173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 32793173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 32803173a18fSJosef Bacik path->slots[0]++; 32813173a18fSJosef Bacik goto next; 32823173a18fSJosef Bacik } 32833173a18fSJosef Bacik 32843173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 32855d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 32865d4f98a2SYan Zheng path->slots[0]++; 32875d4f98a2SYan Zheng goto next; 32885d4f98a2SYan Zheng } 32895d4f98a2SYan Zheng 32903173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 32910b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 32923173a18fSJosef Bacik rc->search_start) { 32933173a18fSJosef Bacik path->slots[0]++; 32943173a18fSJosef Bacik goto next; 32953173a18fSJosef Bacik } 32963173a18fSJosef Bacik 32975d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 32985d4f98a2SYan Zheng key.objectid, &start, &end, 3299e6138876SJosef Bacik EXTENT_DIRTY, NULL); 33005d4f98a2SYan Zheng 33015d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3302b3b4aa74SDavid Sterba btrfs_release_path(path); 33035d4f98a2SYan Zheng rc->search_start = end + 1; 33045d4f98a2SYan Zheng } else { 33053173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 33065d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 33073173a18fSJosef Bacik else 33083173a18fSJosef Bacik rc->search_start = key.objectid + 33090b246afaSJeff Mahoney fs_info->nodesize; 33103fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 33115d4f98a2SYan Zheng return 0; 33125d4f98a2SYan Zheng } 33135d4f98a2SYan Zheng } 3314b3b4aa74SDavid Sterba btrfs_release_path(path); 33155d4f98a2SYan Zheng return ret; 33165d4f98a2SYan Zheng } 33175d4f98a2SYan Zheng 33185d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 33195d4f98a2SYan Zheng { 33205d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33217585717fSChris Mason 33227585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 33235d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 33247585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 33255d4f98a2SYan Zheng } 33265d4f98a2SYan Zheng 33275d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 33285d4f98a2SYan Zheng { 33295d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33307585717fSChris Mason 33317585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 33325d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 33337585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 33345d4f98a2SYan Zheng } 33355d4f98a2SYan Zheng 33365d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 33375d4f98a2SYan Zheng { 33385d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 33395d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 33405d4f98a2SYan Zheng return 1; 33415d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 33425d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 33435d4f98a2SYan Zheng return 1; 33445d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 33455d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 33465d4f98a2SYan Zheng return 1; 33475d4f98a2SYan Zheng return 0; 33485d4f98a2SYan Zheng } 33495d4f98a2SYan Zheng 33503fd0a558SYan, Zheng static noinline_for_stack 33513fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 33523fd0a558SYan, Zheng { 33533fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3354ac2fabacSJosef Bacik int ret; 33553fd0a558SYan, Zheng 33562ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 335766d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 33583fd0a558SYan, Zheng if (!rc->block_rsv) 33593fd0a558SYan, Zheng return -ENOMEM; 33603fd0a558SYan, Zheng 33613fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3362b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 33633fd0a558SYan, Zheng rc->extents_found = 0; 33643fd0a558SYan, Zheng rc->nodes_relocated = 0; 33653fd0a558SYan, Zheng rc->merging_rsv_size = 0; 33660647bf56SWang Shilong rc->reserved_bytes = 0; 3367da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 33680647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3369ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3370ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3371ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3372ac2fabacSJosef Bacik if (ret) 3373ac2fabacSJosef Bacik return ret; 33743fd0a558SYan, Zheng 33753fd0a558SYan, Zheng rc->create_reloc_tree = 1; 33763fd0a558SYan, Zheng set_reloc_control(rc); 33773fd0a558SYan, Zheng 33787a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 337928818947SLiu Bo if (IS_ERR(trans)) { 338028818947SLiu Bo unset_reloc_control(rc); 338128818947SLiu Bo /* 338228818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 338328818947SLiu Bo * cleanup. And callers are responsible to free the above 338428818947SLiu Bo * block rsv. 338528818947SLiu Bo */ 338628818947SLiu Bo return PTR_ERR(trans); 338728818947SLiu Bo } 33883a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 33893fd0a558SYan, Zheng return 0; 33903fd0a558SYan, Zheng } 339176dda93cSYan, Zheng 33925d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 33935d4f98a2SYan Zheng { 33942ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33955d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 33965d4f98a2SYan Zheng struct btrfs_key key; 33975d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 33985d4f98a2SYan Zheng struct btrfs_path *path; 33995d4f98a2SYan Zheng struct btrfs_extent_item *ei; 34005d4f98a2SYan Zheng u64 flags; 34015d4f98a2SYan Zheng u32 item_size; 34025d4f98a2SYan Zheng int ret; 34035d4f98a2SYan Zheng int err = 0; 3404c87f08caSChris Mason int progress = 0; 34055d4f98a2SYan Zheng 34065d4f98a2SYan Zheng path = btrfs_alloc_path(); 34073fd0a558SYan, Zheng if (!path) 34085d4f98a2SYan Zheng return -ENOMEM; 3409e4058b54SDavid Sterba path->reada = READA_FORWARD; 34103fd0a558SYan, Zheng 34113fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 34123fd0a558SYan, Zheng if (ret) { 34133fd0a558SYan, Zheng err = ret; 34143fd0a558SYan, Zheng goto out_free; 34152423fdfbSJiri Slaby } 34165d4f98a2SYan Zheng 34175d4f98a2SYan Zheng while (1) { 34180647bf56SWang Shilong rc->reserved_bytes = 0; 34190647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 34200647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 34210647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 34220647bf56SWang Shilong if (ret) { 34230647bf56SWang Shilong err = ret; 34240647bf56SWang Shilong break; 34250647bf56SWang Shilong } 3426c87f08caSChris Mason progress++; 3427a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 34280f788c58SLiu Bo if (IS_ERR(trans)) { 34290f788c58SLiu Bo err = PTR_ERR(trans); 34300f788c58SLiu Bo trans = NULL; 34310f788c58SLiu Bo break; 34320f788c58SLiu Bo } 3433c87f08caSChris Mason restart: 34343fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 34353a45bb20SJeff Mahoney btrfs_end_transaction(trans); 343642a657f5SPan Bian trans = NULL; 34373fd0a558SYan, Zheng continue; 34383fd0a558SYan, Zheng } 34393fd0a558SYan, Zheng 3440147d256eSZhaolei ret = find_next_extent(rc, path, &key); 34415d4f98a2SYan Zheng if (ret < 0) 34425d4f98a2SYan Zheng err = ret; 34435d4f98a2SYan Zheng if (ret != 0) 34445d4f98a2SYan Zheng break; 34455d4f98a2SYan Zheng 34465d4f98a2SYan Zheng rc->extents_found++; 34475d4f98a2SYan Zheng 34485d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 34495d4f98a2SYan Zheng struct btrfs_extent_item); 34503fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 34515d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 34525d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 34535d4f98a2SYan Zheng ret = check_extent_flags(flags); 34545d4f98a2SYan Zheng BUG_ON(ret); 34556d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3456ba3c2b19SNikolay Borisov err = -EINVAL; 3457ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 3458ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 3459ba3c2b19SNikolay Borisov break; 34605d4f98a2SYan Zheng } else { 34615d4f98a2SYan Zheng BUG(); 34625d4f98a2SYan Zheng } 34635d4f98a2SYan Zheng 34645d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 34655d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 34665d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 34675d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 34685d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 34695d4f98a2SYan Zheng } else { 3470b3b4aa74SDavid Sterba btrfs_release_path(path); 34715d4f98a2SYan Zheng ret = 0; 34725d4f98a2SYan Zheng } 34735d4f98a2SYan Zheng if (ret < 0) { 34743fd0a558SYan, Zheng err = ret; 34755d4f98a2SYan Zheng break; 34765d4f98a2SYan Zheng } 34775d4f98a2SYan Zheng 34785d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 34795d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 34805d4f98a2SYan Zheng if (ret < 0) { 34813fd0a558SYan, Zheng if (ret != -EAGAIN) { 34825d4f98a2SYan Zheng err = ret; 34835d4f98a2SYan Zheng break; 34845d4f98a2SYan Zheng } 34853fd0a558SYan, Zheng rc->extents_found--; 34863fd0a558SYan, Zheng rc->search_start = key.objectid; 34873fd0a558SYan, Zheng } 34885d4f98a2SYan Zheng } 34895d4f98a2SYan Zheng 34903a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 34912ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 34923fd0a558SYan, Zheng trans = NULL; 34935d4f98a2SYan Zheng 34945d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 34955d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 34965d4f98a2SYan Zheng rc->found_file_extent = 1; 34970257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 34983fd0a558SYan, Zheng &key, &rc->cluster); 34995d4f98a2SYan Zheng if (ret < 0) { 35005d4f98a2SYan Zheng err = ret; 35015d4f98a2SYan Zheng break; 35025d4f98a2SYan Zheng } 35035d4f98a2SYan Zheng } 3504f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 3505f31ea088SQu Wenruo err = -ECANCELED; 3506f31ea088SQu Wenruo break; 3507f31ea088SQu Wenruo } 35085d4f98a2SYan Zheng } 3509c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 351043a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 35119689457bSShilong Wang if (ret == 1) { 3512c87f08caSChris Mason err = 0; 3513c87f08caSChris Mason progress = 0; 3514c87f08caSChris Mason goto restart; 3515c87f08caSChris Mason } 3516c87f08caSChris Mason } 35173fd0a558SYan, Zheng 3518b3b4aa74SDavid Sterba btrfs_release_path(path); 351991166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 35205d4f98a2SYan Zheng 35215d4f98a2SYan Zheng if (trans) { 35223a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 35232ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 35245d4f98a2SYan Zheng } 35255d4f98a2SYan Zheng 35260257bb82SYan, Zheng if (!err) { 35273fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 35283fd0a558SYan, Zheng &rc->cluster); 35290257bb82SYan, Zheng if (ret < 0) 35300257bb82SYan, Zheng err = ret; 35310257bb82SYan, Zheng } 35320257bb82SYan, Zheng 35333fd0a558SYan, Zheng rc->create_reloc_tree = 0; 35343fd0a558SYan, Zheng set_reloc_control(rc); 35350257bb82SYan, Zheng 353613fe1bdbSQu Wenruo btrfs_backref_release_cache(&rc->backref_cache); 353763f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 35385d4f98a2SYan Zheng 35397f913c7cSQu Wenruo /* 35407f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 35417f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 35427f913c7cSQu Wenruo * 35437f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 35447f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 35457f913c7cSQu Wenruo * merge_reloc_roots() 35467f913c7cSQu Wenruo */ 35473fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 35485d4f98a2SYan Zheng 35495d4f98a2SYan Zheng merge_reloc_roots(rc); 35505d4f98a2SYan Zheng 35513fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 35525d4f98a2SYan Zheng unset_reloc_control(rc); 355363f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 35545d4f98a2SYan Zheng 35555d4f98a2SYan Zheng /* get rid of pinned extents */ 35567a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 355762b99540SQu Wenruo if (IS_ERR(trans)) { 35583612b495STsutomu Itoh err = PTR_ERR(trans); 355962b99540SQu Wenruo goto out_free; 356062b99540SQu Wenruo } 35613a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 35626217b0faSJosef Bacik out_free: 3563d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3564d2311e69SQu Wenruo if (ret < 0 && !err) 3565d2311e69SQu Wenruo err = ret; 35662ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 35673fd0a558SYan, Zheng btrfs_free_path(path); 35685d4f98a2SYan Zheng return err; 35695d4f98a2SYan Zheng } 35705d4f98a2SYan Zheng 35715d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 35720257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 35735d4f98a2SYan Zheng { 35745d4f98a2SYan Zheng struct btrfs_path *path; 35755d4f98a2SYan Zheng struct btrfs_inode_item *item; 35765d4f98a2SYan Zheng struct extent_buffer *leaf; 357732430c61SNaohiro Aota u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC; 35785d4f98a2SYan Zheng int ret; 35795d4f98a2SYan Zheng 358032430c61SNaohiro Aota if (btrfs_is_zoned(trans->fs_info)) 358132430c61SNaohiro Aota flags &= ~BTRFS_INODE_PREALLOC; 358232430c61SNaohiro Aota 35835d4f98a2SYan Zheng path = btrfs_alloc_path(); 35845d4f98a2SYan Zheng if (!path) 35855d4f98a2SYan Zheng return -ENOMEM; 35865d4f98a2SYan Zheng 35875d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 35885d4f98a2SYan Zheng if (ret) 35895d4f98a2SYan Zheng goto out; 35905d4f98a2SYan Zheng 35915d4f98a2SYan Zheng leaf = path->nodes[0]; 35925d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 3593b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 35945d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 35950257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 35965d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 359732430c61SNaohiro Aota btrfs_set_inode_flags(leaf, item, flags); 35985d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 35995d4f98a2SYan Zheng out: 36005d4f98a2SYan Zheng btrfs_free_path(path); 36015d4f98a2SYan Zheng return ret; 36025d4f98a2SYan Zheng } 36035d4f98a2SYan Zheng 36045d4f98a2SYan Zheng /* 36055d4f98a2SYan Zheng * helper to create inode for data relocation. 36065d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 36075d4f98a2SYan Zheng */ 36083fd0a558SYan, Zheng static noinline_for_stack 36093fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 361032da5386SDavid Sterba struct btrfs_block_group *group) 36115d4f98a2SYan Zheng { 36125d4f98a2SYan Zheng struct inode *inode = NULL; 36135d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 36145d4f98a2SYan Zheng struct btrfs_root *root; 36154624900dSZhaolei u64 objectid; 36165d4f98a2SYan Zheng int err = 0; 36175d4f98a2SYan Zheng 3618aeb935a4SQu Wenruo root = btrfs_grab_root(fs_info->data_reloc_root); 3619a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 362076deacf0SJosef Bacik if (IS_ERR(trans)) { 362100246528SJosef Bacik btrfs_put_root(root); 36223fd0a558SYan, Zheng return ERR_CAST(trans); 362376deacf0SJosef Bacik } 36245d4f98a2SYan Zheng 3625543068a2SNikolay Borisov err = btrfs_get_free_objectid(root, &objectid); 36265d4f98a2SYan Zheng if (err) 36275d4f98a2SYan Zheng goto out; 36285d4f98a2SYan Zheng 36290257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 36305d4f98a2SYan Zheng BUG_ON(err); 36315d4f98a2SYan Zheng 36320202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, objectid, root); 36332e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 3634b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 36355d4f98a2SYan Zheng 363673f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 36375d4f98a2SYan Zheng out: 363800246528SJosef Bacik btrfs_put_root(root); 36393a45bb20SJeff Mahoney btrfs_end_transaction(trans); 36402ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 36415d4f98a2SYan Zheng if (err) { 36425d4f98a2SYan Zheng if (inode) 36435d4f98a2SYan Zheng iput(inode); 36445d4f98a2SYan Zheng inode = ERR_PTR(err); 36455d4f98a2SYan Zheng } 36465d4f98a2SYan Zheng return inode; 36475d4f98a2SYan Zheng } 36485d4f98a2SYan Zheng 3649c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 36503fd0a558SYan, Zheng { 36513fd0a558SYan, Zheng struct reloc_control *rc; 36523fd0a558SYan, Zheng 36533fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 36543fd0a558SYan, Zheng if (!rc) 36553fd0a558SYan, Zheng return NULL; 36563fd0a558SYan, Zheng 36573fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 3658d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 3659584fb121SQu Wenruo btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1); 36603fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 366143eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 366243eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 36633fd0a558SYan, Zheng return rc; 36643fd0a558SYan, Zheng } 36653fd0a558SYan, Zheng 36661a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 36671a0afa0eSJosef Bacik { 36681a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 36691a0afa0eSJosef Bacik 36701a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 36711a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 36721a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 36731a0afa0eSJosef Bacik kfree(node); 36741a0afa0eSJosef Bacik 36751a0afa0eSJosef Bacik kfree(rc); 36761a0afa0eSJosef Bacik } 36771a0afa0eSJosef Bacik 36785d4f98a2SYan Zheng /* 3679ebce0e01SAdam Borowski * Print the block group being relocated 3680ebce0e01SAdam Borowski */ 3681ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 368232da5386SDavid Sterba struct btrfs_block_group *block_group) 3683ebce0e01SAdam Borowski { 3684f89e09cfSAnand Jain char buf[128] = {'\0'}; 3685ebce0e01SAdam Borowski 3686f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 3687ebce0e01SAdam Borowski 3688ebce0e01SAdam Borowski btrfs_info(fs_info, 3689ebce0e01SAdam Borowski "relocating block group %llu flags %s", 3690b3470b5dSDavid Sterba block_group->start, buf); 3691ebce0e01SAdam Borowski } 3692ebce0e01SAdam Borowski 3693430640e3SQu Wenruo static const char *stage_to_string(int stage) 3694430640e3SQu Wenruo { 3695430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 3696430640e3SQu Wenruo return "move data extents"; 3697430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 3698430640e3SQu Wenruo return "update data pointers"; 3699430640e3SQu Wenruo return "unknown"; 3700430640e3SQu Wenruo } 3701430640e3SQu Wenruo 3702ebce0e01SAdam Borowski /* 37035d4f98a2SYan Zheng * function to relocate all extents in a block group. 37045d4f98a2SYan Zheng */ 37056bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 37065d4f98a2SYan Zheng { 370732da5386SDavid Sterba struct btrfs_block_group *bg; 37086bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 37095d4f98a2SYan Zheng struct reloc_control *rc; 37100af3d00bSJosef Bacik struct inode *inode; 37110af3d00bSJosef Bacik struct btrfs_path *path; 37125d4f98a2SYan Zheng int ret; 3713f0486c68SYan, Zheng int rw = 0; 37145d4f98a2SYan Zheng int err = 0; 37155d4f98a2SYan Zheng 3716eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 3717eede2bf3SOmar Sandoval if (!bg) 3718eede2bf3SOmar Sandoval return -ENOENT; 3719eede2bf3SOmar Sandoval 3720eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 3721eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 3722eede2bf3SOmar Sandoval return -ETXTBSY; 3723eede2bf3SOmar Sandoval } 3724eede2bf3SOmar Sandoval 3725c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 3726eede2bf3SOmar Sandoval if (!rc) { 3727eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 37285d4f98a2SYan Zheng return -ENOMEM; 3729eede2bf3SOmar Sandoval } 37305d4f98a2SYan Zheng 3731f0486c68SYan, Zheng rc->extent_root = extent_root; 3732eede2bf3SOmar Sandoval rc->block_group = bg; 37335d4f98a2SYan Zheng 3734b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 3735f0486c68SYan, Zheng if (ret) { 3736f0486c68SYan, Zheng err = ret; 3737f0486c68SYan, Zheng goto out; 3738f0486c68SYan, Zheng } 3739f0486c68SYan, Zheng rw = 1; 3740f0486c68SYan, Zheng 37410af3d00bSJosef Bacik path = btrfs_alloc_path(); 37420af3d00bSJosef Bacik if (!path) { 37430af3d00bSJosef Bacik err = -ENOMEM; 37440af3d00bSJosef Bacik goto out; 37450af3d00bSJosef Bacik } 37460af3d00bSJosef Bacik 37477949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 37480af3d00bSJosef Bacik btrfs_free_path(path); 37490af3d00bSJosef Bacik 37500af3d00bSJosef Bacik if (!IS_ERR(inode)) 37511bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 37520af3d00bSJosef Bacik else 37530af3d00bSJosef Bacik ret = PTR_ERR(inode); 37540af3d00bSJosef Bacik 37550af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 37560af3d00bSJosef Bacik err = ret; 37570af3d00bSJosef Bacik goto out; 37580af3d00bSJosef Bacik } 37590af3d00bSJosef Bacik 37605d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 37615d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 37625d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 37635d4f98a2SYan Zheng rc->data_inode = NULL; 37645d4f98a2SYan Zheng goto out; 37655d4f98a2SYan Zheng } 37665d4f98a2SYan Zheng 37670b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 37685d4f98a2SYan Zheng 37699cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 3770f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 37716374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 3772b3470b5dSDavid Sterba rc->block_group->start, 3773b3470b5dSDavid Sterba rc->block_group->length); 37745d4f98a2SYan Zheng 37755d4f98a2SYan Zheng while (1) { 3776430640e3SQu Wenruo int finishes_stage; 3777430640e3SQu Wenruo 377876dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 37795d4f98a2SYan Zheng ret = relocate_block_group(rc); 378076dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 3781ff612ba7SJosef Bacik if (ret < 0) 37825d4f98a2SYan Zheng err = ret; 3783ff612ba7SJosef Bacik 3784430640e3SQu Wenruo finishes_stage = rc->stage; 3785ff612ba7SJosef Bacik /* 3786ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 3787ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 3788ff612ba7SJosef Bacik * different block group we could end up hitting the 3789ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 3790ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 3791ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 3792ff612ba7SJosef Bacik * out of the loop if we hit an error. 3793ff612ba7SJosef Bacik */ 3794ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 3795ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 3796ff612ba7SJosef Bacik (u64)-1); 3797ff612ba7SJosef Bacik if (ret) 3798ff612ba7SJosef Bacik err = ret; 3799ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 3800ff612ba7SJosef Bacik 0, -1); 3801ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 38025d4f98a2SYan Zheng } 38035d4f98a2SYan Zheng 3804ff612ba7SJosef Bacik if (err < 0) 3805ff612ba7SJosef Bacik goto out; 3806ff612ba7SJosef Bacik 38075d4f98a2SYan Zheng if (rc->extents_found == 0) 38085d4f98a2SYan Zheng break; 38095d4f98a2SYan Zheng 3810430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 3811430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 38125d4f98a2SYan Zheng } 38135d4f98a2SYan Zheng 38145d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 38155d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 3816bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 38175d4f98a2SYan Zheng out: 3818f0486c68SYan, Zheng if (err && rw) 38192ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 38205d4f98a2SYan Zheng iput(rc->data_inode); 38215d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 38221a0afa0eSJosef Bacik free_reloc_control(rc); 38235d4f98a2SYan Zheng return err; 38245d4f98a2SYan Zheng } 38255d4f98a2SYan Zheng 382676dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 382776dda93cSYan, Zheng { 38280b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 382976dda93cSYan, Zheng struct btrfs_trans_handle *trans; 383079787eaaSJeff Mahoney int ret, err; 383176dda93cSYan, Zheng 38320b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 383379787eaaSJeff Mahoney if (IS_ERR(trans)) 383479787eaaSJeff Mahoney return PTR_ERR(trans); 383576dda93cSYan, Zheng 383676dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 383776dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 3838c8422684SDavid Sterba btrfs_set_root_drop_level(&root->root_item, 0); 383976dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 38400b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 384176dda93cSYan, Zheng &root->root_key, &root->root_item); 384276dda93cSYan, Zheng 38433a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 384479787eaaSJeff Mahoney if (err) 384579787eaaSJeff Mahoney return err; 384679787eaaSJeff Mahoney return ret; 384776dda93cSYan, Zheng } 384876dda93cSYan, Zheng 38495d4f98a2SYan Zheng /* 38505d4f98a2SYan Zheng * recover relocation interrupted by system crash. 38515d4f98a2SYan Zheng * 38525d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 38535d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 38545d4f98a2SYan Zheng */ 38555d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 38565d4f98a2SYan Zheng { 38570b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 38585d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 38595d4f98a2SYan Zheng struct btrfs_key key; 38605d4f98a2SYan Zheng struct btrfs_root *fs_root; 38615d4f98a2SYan Zheng struct btrfs_root *reloc_root; 38625d4f98a2SYan Zheng struct btrfs_path *path; 38635d4f98a2SYan Zheng struct extent_buffer *leaf; 38645d4f98a2SYan Zheng struct reloc_control *rc = NULL; 38655d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 38665d4f98a2SYan Zheng int ret; 38675d4f98a2SYan Zheng int err = 0; 38685d4f98a2SYan Zheng 38695d4f98a2SYan Zheng path = btrfs_alloc_path(); 38705d4f98a2SYan Zheng if (!path) 38715d4f98a2SYan Zheng return -ENOMEM; 3872e4058b54SDavid Sterba path->reada = READA_BACK; 38735d4f98a2SYan Zheng 38745d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 38755d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 38765d4f98a2SYan Zheng key.offset = (u64)-1; 38775d4f98a2SYan Zheng 38785d4f98a2SYan Zheng while (1) { 38790b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 38805d4f98a2SYan Zheng path, 0, 0); 38815d4f98a2SYan Zheng if (ret < 0) { 38825d4f98a2SYan Zheng err = ret; 38835d4f98a2SYan Zheng goto out; 38845d4f98a2SYan Zheng } 38855d4f98a2SYan Zheng if (ret > 0) { 38865d4f98a2SYan Zheng if (path->slots[0] == 0) 38875d4f98a2SYan Zheng break; 38885d4f98a2SYan Zheng path->slots[0]--; 38895d4f98a2SYan Zheng } 38905d4f98a2SYan Zheng leaf = path->nodes[0]; 38915d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 3892b3b4aa74SDavid Sterba btrfs_release_path(path); 38935d4f98a2SYan Zheng 38945d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 38955d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 38965d4f98a2SYan Zheng break; 38975d4f98a2SYan Zheng 38983dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 38995d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 39005d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 39015d4f98a2SYan Zheng goto out; 39025d4f98a2SYan Zheng } 39035d4f98a2SYan Zheng 390492a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 39055d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 39065d4f98a2SYan Zheng 39075d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 3908a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, 3909a820feb5SDavid Sterba reloc_root->root_key.offset, false); 39105d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 391176dda93cSYan, Zheng ret = PTR_ERR(fs_root); 391276dda93cSYan, Zheng if (ret != -ENOENT) { 391376dda93cSYan, Zheng err = ret; 39145d4f98a2SYan Zheng goto out; 39155d4f98a2SYan Zheng } 391679787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 391779787eaaSJeff Mahoney if (ret < 0) { 391879787eaaSJeff Mahoney err = ret; 391979787eaaSJeff Mahoney goto out; 392079787eaaSJeff Mahoney } 3921932fd26dSJosef Bacik } else { 392200246528SJosef Bacik btrfs_put_root(fs_root); 392376dda93cSYan, Zheng } 39245d4f98a2SYan Zheng } 39255d4f98a2SYan Zheng 39265d4f98a2SYan Zheng if (key.offset == 0) 39275d4f98a2SYan Zheng break; 39285d4f98a2SYan Zheng 39295d4f98a2SYan Zheng key.offset--; 39305d4f98a2SYan Zheng } 3931b3b4aa74SDavid Sterba btrfs_release_path(path); 39325d4f98a2SYan Zheng 39335d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 39345d4f98a2SYan Zheng goto out; 39355d4f98a2SYan Zheng 3936c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 39375d4f98a2SYan Zheng if (!rc) { 39385d4f98a2SYan Zheng err = -ENOMEM; 39395d4f98a2SYan Zheng goto out; 39405d4f98a2SYan Zheng } 39415d4f98a2SYan Zheng 39420b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 39435d4f98a2SYan Zheng 39445d4f98a2SYan Zheng set_reloc_control(rc); 39455d4f98a2SYan Zheng 39467a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 39473612b495STsutomu Itoh if (IS_ERR(trans)) { 39483612b495STsutomu Itoh err = PTR_ERR(trans); 3949fb2d83eeSJosef Bacik goto out_unset; 39503612b495STsutomu Itoh } 39513fd0a558SYan, Zheng 39523fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 39533fd0a558SYan, Zheng 39545d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 39555d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 39565d4f98a2SYan Zheng struct btrfs_root, root_list); 39575d4f98a2SYan Zheng list_del(&reloc_root->root_list); 39585d4f98a2SYan Zheng 39595d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 39605d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 39615d4f98a2SYan Zheng &rc->reloc_roots); 39625d4f98a2SYan Zheng continue; 39635d4f98a2SYan Zheng } 39645d4f98a2SYan Zheng 3965a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 3966a820feb5SDavid Sterba false); 396779787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 396879787eaaSJeff Mahoney err = PTR_ERR(fs_root); 3969ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 39701402d17dSXiyu Yang btrfs_end_transaction(trans); 3971fb2d83eeSJosef Bacik goto out_unset; 397279787eaaSJeff Mahoney } 39735d4f98a2SYan Zheng 3974ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 397579787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 3976f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 397700246528SJosef Bacik btrfs_put_root(fs_root); 39785d4f98a2SYan Zheng } 39795d4f98a2SYan Zheng 39803a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 398179787eaaSJeff Mahoney if (err) 3982fb2d83eeSJosef Bacik goto out_unset; 39835d4f98a2SYan Zheng 39845d4f98a2SYan Zheng merge_reloc_roots(rc); 39855d4f98a2SYan Zheng 39865d4f98a2SYan Zheng unset_reloc_control(rc); 39875d4f98a2SYan Zheng 39887a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 398962b99540SQu Wenruo if (IS_ERR(trans)) { 39903612b495STsutomu Itoh err = PTR_ERR(trans); 39916217b0faSJosef Bacik goto out_clean; 399262b99540SQu Wenruo } 39933a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 39946217b0faSJosef Bacik out_clean: 3995d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3996d2311e69SQu Wenruo if (ret < 0 && !err) 3997d2311e69SQu Wenruo err = ret; 3998fb2d83eeSJosef Bacik out_unset: 3999fb2d83eeSJosef Bacik unset_reloc_control(rc); 40001a0afa0eSJosef Bacik free_reloc_control(rc); 40013612b495STsutomu Itoh out: 4002aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4003aca1bba6SLiu Bo 40045d4f98a2SYan Zheng btrfs_free_path(path); 40055d4f98a2SYan Zheng 40065d4f98a2SYan Zheng if (err == 0) { 40075d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 4008aeb935a4SQu Wenruo fs_root = btrfs_grab_root(fs_info->data_reloc_root); 4009aeb935a4SQu Wenruo ASSERT(fs_root); 401066b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 401100246528SJosef Bacik btrfs_put_root(fs_root); 4012932fd26dSJosef Bacik } 40135d4f98a2SYan Zheng return err; 40145d4f98a2SYan Zheng } 40155d4f98a2SYan Zheng 40165d4f98a2SYan Zheng /* 40175d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 40185d4f98a2SYan Zheng * 40195d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 40205d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 40215d4f98a2SYan Zheng */ 40227bfa9535SNikolay Borisov int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len) 40235d4f98a2SYan Zheng { 40247bfa9535SNikolay Borisov struct btrfs_fs_info *fs_info = inode->root->fs_info; 40255d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 40265d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 40275d4f98a2SYan Zheng int ret; 40285d4f98a2SYan Zheng u64 disk_bytenr; 40294577b014SJosef Bacik u64 new_bytenr; 40305d4f98a2SYan Zheng LIST_HEAD(list); 40315d4f98a2SYan Zheng 40327bfa9535SNikolay Borisov ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4033bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 40345d4f98a2SYan Zheng 40357bfa9535SNikolay Borisov disk_bytenr = file_pos + inode->index_cnt; 40360b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4037a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 403879787eaaSJeff Mahoney if (ret) 403979787eaaSJeff Mahoney goto out; 40405d4f98a2SYan Zheng 40415d4f98a2SYan Zheng while (!list_empty(&list)) { 40425d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 40435d4f98a2SYan Zheng list_del_init(&sums->list); 40445d4f98a2SYan Zheng 40454577b014SJosef Bacik /* 40464577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 40474577b014SJosef Bacik * We need to do this because we will read in entire prealloc 40484577b014SJosef Bacik * extents but we may have written to say the middle of the 40494577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 40504577b014SJosef Bacik * the right disk offset. 40514577b014SJosef Bacik * 40524577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 40534577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 40544577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 40554577b014SJosef Bacik * disk length. 40564577b014SJosef Bacik */ 4057bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 40584577b014SJosef Bacik sums->bytenr = new_bytenr; 40595d4f98a2SYan Zheng 4060f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 40615d4f98a2SYan Zheng } 406279787eaaSJeff Mahoney out: 40635d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4064411fc6bcSAndi Kleen return ret; 40655d4f98a2SYan Zheng } 40663fd0a558SYan, Zheng 406783d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 40683fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 40693fd0a558SYan, Zheng struct extent_buffer *cow) 40703fd0a558SYan, Zheng { 40710b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 40723fd0a558SYan, Zheng struct reloc_control *rc; 4073a26195a5SQu Wenruo struct btrfs_backref_node *node; 40743fd0a558SYan, Zheng int first_cow = 0; 40753fd0a558SYan, Zheng int level; 407683d4cfd4SJosef Bacik int ret = 0; 40773fd0a558SYan, Zheng 40780b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 40793fd0a558SYan, Zheng if (!rc) 408083d4cfd4SJosef Bacik return 0; 40813fd0a558SYan, Zheng 40823fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 40833fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 40843fd0a558SYan, Zheng 40853fd0a558SYan, Zheng level = btrfs_header_level(buf); 40863fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 40873fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 40883fd0a558SYan, Zheng first_cow = 1; 40893fd0a558SYan, Zheng 40903fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 40913fd0a558SYan, Zheng rc->create_reloc_tree) { 40923fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 40933fd0a558SYan, Zheng 40943fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 40953fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 40963fd0a558SYan, Zheng node->new_bytenr != buf->start); 40973fd0a558SYan, Zheng 4098b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 409967439dadSDavid Sterba atomic_inc(&cow->refs); 41003fd0a558SYan, Zheng node->eb = cow; 41013fd0a558SYan, Zheng node->new_bytenr = cow->start; 41023fd0a558SYan, Zheng 41033fd0a558SYan, Zheng if (!node->pending) { 41043fd0a558SYan, Zheng list_move_tail(&node->list, 41053fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 41063fd0a558SYan, Zheng node->pending = 1; 41073fd0a558SYan, Zheng } 41083fd0a558SYan, Zheng 41093fd0a558SYan, Zheng if (first_cow) 41109569cc20SQu Wenruo mark_block_processed(rc, node); 41113fd0a558SYan, Zheng 41123fd0a558SYan, Zheng if (first_cow && level > 0) 41133fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 41143fd0a558SYan, Zheng } 41153fd0a558SYan, Zheng 411683d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 41173fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 411883d4cfd4SJosef Bacik return ret; 41193fd0a558SYan, Zheng } 41203fd0a558SYan, Zheng 41213fd0a558SYan, Zheng /* 41223fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 412301327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 41243fd0a558SYan, Zheng */ 4125147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 41263fd0a558SYan, Zheng u64 *bytes_to_reserve) 41273fd0a558SYan, Zheng { 412810995c04SQu Wenruo struct btrfs_root *root = pending->root; 412910995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 41303fd0a558SYan, Zheng 41316282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 41323fd0a558SYan, Zheng return; 41333fd0a558SYan, Zheng 41343fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 41353fd0a558SYan, Zheng return; 41363fd0a558SYan, Zheng 41373fd0a558SYan, Zheng root = root->reloc_root; 41383fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 41393fd0a558SYan, Zheng /* 41403fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 41413fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 41423fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 41433fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 41443fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 41453fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 41463fd0a558SYan, Zheng * tree may use more space than it frees. so we need 41473fd0a558SYan, Zheng * reserve extra space. 41483fd0a558SYan, Zheng */ 41493fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 41503fd0a558SYan, Zheng } 41513fd0a558SYan, Zheng 41523fd0a558SYan, Zheng /* 41533fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 41543fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4155f44deb74SJosef Bacik * 4156f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4157f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4158f44deb74SJosef Bacik * rc->reloc_roots. 41593fd0a558SYan, Zheng */ 416049b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 41613fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 41623fd0a558SYan, Zheng { 41633fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 41643fd0a558SYan, Zheng struct btrfs_root *reloc_root; 41653fd0a558SYan, Zheng struct btrfs_root *new_root; 416610995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 41673fd0a558SYan, Zheng int ret; 41683fd0a558SYan, Zheng 41696282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 417049b25e05SJeff Mahoney return 0; 41713fd0a558SYan, Zheng 41723fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 41733fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 41743fd0a558SYan, Zheng 41753fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 41763fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 41773fd0a558SYan, Zheng rc->block_rsv, 41783a584174SLu Fengqi rc->nodes_relocated, true); 417949b25e05SJeff Mahoney if (ret) 418049b25e05SJeff Mahoney return ret; 41813fd0a558SYan, Zheng } 41823fd0a558SYan, Zheng 41833fd0a558SYan, Zheng new_root = pending->snap; 41843fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 41853fd0a558SYan, Zheng new_root->root_key.objectid); 418649b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 418749b25e05SJeff Mahoney return PTR_ERR(reloc_root); 41883fd0a558SYan, Zheng 4189ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4190ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4191f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 41923fd0a558SYan, Zheng 419349b25e05SJeff Mahoney if (rc->create_reloc_tree) 41943fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 419549b25e05SJeff Mahoney return ret; 41963fd0a558SYan, Zheng } 4197