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" 21581bb050SLi Zefan #include "inode-map.h" 2262b99540SQu Wenruo #include "qgroup.h" 23cdccee99SLiu Bo #include "print-tree.h" 2486736342SJosef Bacik #include "delalloc-space.h" 25aac0023cSJosef Bacik #include "block-group.h" 2619b546d7SQu Wenruo #include "backref.h" 27e9a28dc5SQu Wenruo #include "misc.h" 285d4f98a2SYan Zheng 295d4f98a2SYan Zheng /* 300c891389SQu Wenruo * Relocation overview 310c891389SQu Wenruo * 320c891389SQu Wenruo * [What does relocation do] 330c891389SQu Wenruo * 340c891389SQu Wenruo * The objective of relocation is to relocate all extents of the target block 350c891389SQu Wenruo * group to other block groups. 360c891389SQu Wenruo * This is utilized by resize (shrink only), profile converting, compacting 370c891389SQu Wenruo * space, or balance routine to spread chunks over devices. 380c891389SQu Wenruo * 390c891389SQu Wenruo * Before | After 400c891389SQu Wenruo * ------------------------------------------------------------------ 410c891389SQu Wenruo * BG A: 10 data extents | BG A: deleted 420c891389SQu Wenruo * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated) 430c891389SQu Wenruo * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated) 440c891389SQu Wenruo * 450c891389SQu Wenruo * [How does relocation work] 460c891389SQu Wenruo * 470c891389SQu Wenruo * 1. Mark the target block group read-only 480c891389SQu Wenruo * New extents won't be allocated from the target block group. 490c891389SQu Wenruo * 500c891389SQu Wenruo * 2.1 Record each extent in the target block group 510c891389SQu Wenruo * To build a proper map of extents to be relocated. 520c891389SQu Wenruo * 530c891389SQu Wenruo * 2.2 Build data reloc tree and reloc trees 540c891389SQu Wenruo * Data reloc tree will contain an inode, recording all newly relocated 550c891389SQu Wenruo * data extents. 560c891389SQu Wenruo * There will be only one data reloc tree for one data block group. 570c891389SQu Wenruo * 580c891389SQu Wenruo * Reloc tree will be a special snapshot of its source tree, containing 590c891389SQu Wenruo * relocated tree blocks. 600c891389SQu Wenruo * Each tree referring to a tree block in target block group will get its 610c891389SQu Wenruo * reloc tree built. 620c891389SQu Wenruo * 630c891389SQu Wenruo * 2.3 Swap source tree with its corresponding reloc tree 640c891389SQu Wenruo * Each involved tree only refers to new extents after swap. 650c891389SQu Wenruo * 660c891389SQu Wenruo * 3. Cleanup reloc trees and data reloc tree. 670c891389SQu Wenruo * As old extents in the target block group are still referenced by reloc 680c891389SQu Wenruo * trees, we need to clean them up before really freeing the target block 690c891389SQu Wenruo * group. 700c891389SQu Wenruo * 710c891389SQu Wenruo * The main complexity is in steps 2.2 and 2.3. 720c891389SQu Wenruo * 730c891389SQu Wenruo * The entry point of relocation is relocate_block_group() function. 740c891389SQu Wenruo */ 750c891389SQu Wenruo 760647bf56SWang Shilong #define RELOCATION_RESERVED_NODES 256 772a979612SQu Wenruo /* 785d4f98a2SYan Zheng * map address of tree root to tree 795d4f98a2SYan Zheng */ 805d4f98a2SYan Zheng struct mapping_node { 81e9a28dc5SQu Wenruo struct { 825d4f98a2SYan Zheng struct rb_node rb_node; 835d4f98a2SYan Zheng u64 bytenr; 84e9a28dc5SQu Wenruo }; /* Use rb_simle_node for search/insert */ 855d4f98a2SYan Zheng void *data; 865d4f98a2SYan Zheng }; 875d4f98a2SYan Zheng 885d4f98a2SYan Zheng struct mapping_tree { 895d4f98a2SYan Zheng struct rb_root rb_root; 905d4f98a2SYan Zheng spinlock_t lock; 915d4f98a2SYan Zheng }; 925d4f98a2SYan Zheng 935d4f98a2SYan Zheng /* 945d4f98a2SYan Zheng * present a tree block to process 955d4f98a2SYan Zheng */ 965d4f98a2SYan Zheng struct tree_block { 97e9a28dc5SQu Wenruo struct { 985d4f98a2SYan Zheng struct rb_node rb_node; 995d4f98a2SYan Zheng u64 bytenr; 100e9a28dc5SQu Wenruo }; /* Use rb_simple_node for search/insert */ 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); 672c974c464SWang Shilong if (!node) 673c974c464SWang Shilong return; 674c974c464SWang Shilong BUG_ON((struct btrfs_root *)node->data != root); 675389305b2SQu Wenruo } 676c974c464SWang Shilong 677f44deb74SJosef Bacik /* 678f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 679f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 680f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 681f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 682f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 683f44deb74SJosef Bacik * messing with the list is in charge of the reference. 684f44deb74SJosef Bacik */ 6850b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 686f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 687f44deb74SJosef Bacik put_ref = true; 688c974c464SWang Shilong list_del_init(&root->root_list); 689f44deb74SJosef Bacik } 6900b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 691f44deb74SJosef Bacik if (put_ref) 692f44deb74SJosef Bacik btrfs_put_root(root); 693c974c464SWang Shilong kfree(node); 694c974c464SWang Shilong } 695c974c464SWang Shilong 696c974c464SWang Shilong /* 697c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 698c974c464SWang Shilong * mapping 699c974c464SWang Shilong */ 700ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 701c974c464SWang Shilong { 7020b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 703c974c464SWang Shilong struct rb_node *rb_node; 704c974c464SWang Shilong struct mapping_node *node = NULL; 7050b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 706c974c464SWang Shilong 707c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 708e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 709ea287ab1SJosef Bacik root->commit_root->start); 7105d4f98a2SYan Zheng if (rb_node) { 7115d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 7125d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 7135d4f98a2SYan Zheng } 7145d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 7155d4f98a2SYan Zheng 7168f71f3e0SLiu Bo if (!node) 7178f71f3e0SLiu Bo return 0; 7185d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 7195d4f98a2SYan Zheng 7205d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 721ea287ab1SJosef Bacik node->bytenr = root->node->start; 722e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 7235d4f98a2SYan Zheng node->bytenr, &node->rb_node); 7245d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 72543c04fb1SJeff Mahoney if (rb_node) 726982c92cbSQu Wenruo btrfs_backref_panic(fs_info, node->bytenr, -EEXIST); 7275d4f98a2SYan Zheng return 0; 7285d4f98a2SYan Zheng } 7295d4f98a2SYan Zheng 7303fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 7313fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 7325d4f98a2SYan Zheng { 7330b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 7345d4f98a2SYan Zheng struct btrfs_root *reloc_root; 7355d4f98a2SYan Zheng struct extent_buffer *eb; 7365d4f98a2SYan Zheng struct btrfs_root_item *root_item; 7375d4f98a2SYan Zheng struct btrfs_key root_key; 7385d4f98a2SYan Zheng int ret; 7395d4f98a2SYan Zheng 7405d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 7415d4f98a2SYan Zheng BUG_ON(!root_item); 7425d4f98a2SYan Zheng 7435d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 7445d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 7453fd0a558SYan, Zheng root_key.offset = objectid; 7465d4f98a2SYan Zheng 7473fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 748054570a1SFilipe Manana u64 commit_root_gen; 749054570a1SFilipe Manana 7503fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 7515d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 7525d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 7535d4f98a2SYan Zheng BUG_ON(ret); 754054570a1SFilipe Manana /* 755054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 756054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 757054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 758054570a1SFilipe Manana * either inside the critical section of a transaction commit 759054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 760054570a1SFilipe Manana * it's created before the transaction commit is started. 761054570a1SFilipe Manana */ 762054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 763054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 7643fd0a558SYan, Zheng } else { 7653fd0a558SYan, Zheng /* 7663fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 7673fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 7683fd0a558SYan, Zheng * modified after it was created have RELOC flag 7693fd0a558SYan, Zheng * set in their headers. so it's OK to not update 7703fd0a558SYan, Zheng * the 'last_snapshot'. 7713fd0a558SYan, Zheng */ 7723fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 7733fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 7743fd0a558SYan, Zheng BUG_ON(ret); 7753fd0a558SYan, Zheng } 7763fd0a558SYan, Zheng 7775d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 7785d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 7795d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 7805d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 7813fd0a558SYan, Zheng 7823fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 7833fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 7843fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 7853fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 7865d4f98a2SYan Zheng root_item->drop_level = 0; 7873fd0a558SYan, Zheng } 7885d4f98a2SYan Zheng 7895d4f98a2SYan Zheng btrfs_tree_unlock(eb); 7905d4f98a2SYan Zheng free_extent_buffer(eb); 7915d4f98a2SYan Zheng 7920b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 7935d4f98a2SYan Zheng &root_key, root_item); 7945d4f98a2SYan Zheng BUG_ON(ret); 7955d4f98a2SYan Zheng kfree(root_item); 7965d4f98a2SYan Zheng 7973dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 7985d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 79992a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 8005d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 8013fd0a558SYan, Zheng return reloc_root; 8023fd0a558SYan, Zheng } 8033fd0a558SYan, Zheng 8043fd0a558SYan, Zheng /* 8053fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 8063fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 807f44deb74SJosef Bacik * 808f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 809f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 8103fd0a558SYan, Zheng */ 8113fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 8123fd0a558SYan, Zheng struct btrfs_root *root) 8133fd0a558SYan, Zheng { 8140b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 8153fd0a558SYan, Zheng struct btrfs_root *reloc_root; 8160b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 81720dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 8183fd0a558SYan, Zheng int clear_rsv = 0; 819ffd7b339SJeff Mahoney int ret; 8203fd0a558SYan, Zheng 821aec7db3bSJosef Bacik if (!rc) 8222abc726aSJosef Bacik return 0; 8232abc726aSJosef Bacik 8241fac4a54SQu Wenruo /* 8251fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 8261fac4a54SQu Wenruo * create/update the dead reloc tree 8271fac4a54SQu Wenruo */ 8286282675eSQu Wenruo if (reloc_root_is_dead(root)) 8291fac4a54SQu Wenruo return 0; 8301fac4a54SQu Wenruo 831aec7db3bSJosef Bacik /* 832aec7db3bSJosef Bacik * This is subtle but important. We do not do 833aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 834aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 835aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 836aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 837aec7db3bSJosef Bacik * in. 838aec7db3bSJosef Bacik */ 8393fd0a558SYan, Zheng if (root->reloc_root) { 8403fd0a558SYan, Zheng reloc_root = root->reloc_root; 8413fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 8423fd0a558SYan, Zheng return 0; 8433fd0a558SYan, Zheng } 8443fd0a558SYan, Zheng 845aec7db3bSJosef Bacik /* 846aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 847aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 848aec7db3bSJosef Bacik */ 849aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 850aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 851aec7db3bSJosef Bacik return 0; 852aec7db3bSJosef Bacik 85320dd2cbfSMiao Xie if (!trans->reloc_reserved) { 85420dd2cbfSMiao Xie rsv = trans->block_rsv; 8553fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 8563fd0a558SYan, Zheng clear_rsv = 1; 8573fd0a558SYan, Zheng } 8583fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 8593fd0a558SYan, Zheng if (clear_rsv) 86020dd2cbfSMiao Xie trans->block_rsv = rsv; 8615d4f98a2SYan Zheng 862ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 863ffd7b339SJeff Mahoney BUG_ON(ret < 0); 864f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 8655d4f98a2SYan Zheng return 0; 8665d4f98a2SYan Zheng } 8675d4f98a2SYan Zheng 8685d4f98a2SYan Zheng /* 8695d4f98a2SYan Zheng * update root item of reloc tree 8705d4f98a2SYan Zheng */ 8715d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 8725d4f98a2SYan Zheng struct btrfs_root *root) 8735d4f98a2SYan Zheng { 8740b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 8755d4f98a2SYan Zheng struct btrfs_root *reloc_root; 8765d4f98a2SYan Zheng struct btrfs_root_item *root_item; 8775d4f98a2SYan Zheng int ret; 8785d4f98a2SYan Zheng 8796282675eSQu Wenruo if (!have_reloc_root(root)) 8807585717fSChris Mason goto out; 8815d4f98a2SYan Zheng 8825d4f98a2SYan Zheng reloc_root = root->reloc_root; 8835d4f98a2SYan Zheng root_item = &reloc_root->root_item; 8845d4f98a2SYan Zheng 885f44deb74SJosef Bacik /* 886f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 887f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 888f44deb74SJosef Bacik * hold it while we update the reloc root. 889f44deb74SJosef Bacik */ 890f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 891f44deb74SJosef Bacik 892d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 8930b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 8943fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 895d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 8966282675eSQu Wenruo /* 8976282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 8986282675eSQu Wenruo * have_reloc_root will not touch it from now on. 8996282675eSQu Wenruo */ 9006282675eSQu Wenruo smp_wmb(); 901c974c464SWang Shilong __del_reloc_root(reloc_root); 9025d4f98a2SYan Zheng } 9035d4f98a2SYan Zheng 9045d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 905ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 9065d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 9075d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 9085d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 9095d4f98a2SYan Zheng } 9105d4f98a2SYan Zheng 9110b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 9125d4f98a2SYan Zheng &reloc_root->root_key, root_item); 9135d4f98a2SYan Zheng BUG_ON(ret); 914f44deb74SJosef Bacik btrfs_put_root(reloc_root); 9157585717fSChris Mason out: 9165d4f98a2SYan Zheng return 0; 9175d4f98a2SYan Zheng } 9185d4f98a2SYan Zheng 9195d4f98a2SYan Zheng /* 9205d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 9215d4f98a2SYan Zheng * in a subvolume 9225d4f98a2SYan Zheng */ 9235d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 9245d4f98a2SYan Zheng { 9255d4f98a2SYan Zheng struct rb_node *node; 9265d4f98a2SYan Zheng struct rb_node *prev; 9275d4f98a2SYan Zheng struct btrfs_inode *entry; 9285d4f98a2SYan Zheng struct inode *inode; 9295d4f98a2SYan Zheng 9305d4f98a2SYan Zheng spin_lock(&root->inode_lock); 9315d4f98a2SYan Zheng again: 9325d4f98a2SYan Zheng node = root->inode_tree.rb_node; 9335d4f98a2SYan Zheng prev = NULL; 9345d4f98a2SYan Zheng while (node) { 9355d4f98a2SYan Zheng prev = node; 9365d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9375d4f98a2SYan Zheng 9384a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 9395d4f98a2SYan Zheng node = node->rb_left; 9404a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 9415d4f98a2SYan Zheng node = node->rb_right; 9425d4f98a2SYan Zheng else 9435d4f98a2SYan Zheng break; 9445d4f98a2SYan Zheng } 9455d4f98a2SYan Zheng if (!node) { 9465d4f98a2SYan Zheng while (prev) { 9475d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 9484a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 9495d4f98a2SYan Zheng node = prev; 9505d4f98a2SYan Zheng break; 9515d4f98a2SYan Zheng } 9525d4f98a2SYan Zheng prev = rb_next(prev); 9535d4f98a2SYan Zheng } 9545d4f98a2SYan Zheng } 9555d4f98a2SYan Zheng while (node) { 9565d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9575d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 9585d4f98a2SYan Zheng if (inode) { 9595d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9605d4f98a2SYan Zheng return inode; 9615d4f98a2SYan Zheng } 9625d4f98a2SYan Zheng 9634a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 9645d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 9655d4f98a2SYan Zheng goto again; 9665d4f98a2SYan Zheng 9675d4f98a2SYan Zheng node = rb_next(node); 9685d4f98a2SYan Zheng } 9695d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9705d4f98a2SYan Zheng return NULL; 9715d4f98a2SYan Zheng } 9725d4f98a2SYan Zheng 9735d4f98a2SYan Zheng /* 9745d4f98a2SYan Zheng * get new location of data 9755d4f98a2SYan Zheng */ 9765d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 9775d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 9785d4f98a2SYan Zheng { 9795d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 9805d4f98a2SYan Zheng struct btrfs_path *path; 9815d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 9825d4f98a2SYan Zheng struct extent_buffer *leaf; 9835d4f98a2SYan Zheng int ret; 9845d4f98a2SYan Zheng 9855d4f98a2SYan Zheng path = btrfs_alloc_path(); 9865d4f98a2SYan Zheng if (!path) 9875d4f98a2SYan Zheng return -ENOMEM; 9885d4f98a2SYan Zheng 9895d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 990f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 991f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 9925d4f98a2SYan Zheng if (ret < 0) 9935d4f98a2SYan Zheng goto out; 9945d4f98a2SYan Zheng if (ret > 0) { 9955d4f98a2SYan Zheng ret = -ENOENT; 9965d4f98a2SYan Zheng goto out; 9975d4f98a2SYan Zheng } 9985d4f98a2SYan Zheng 9995d4f98a2SYan Zheng leaf = path->nodes[0]; 10005d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 10015d4f98a2SYan Zheng struct btrfs_file_extent_item); 10025d4f98a2SYan Zheng 10035d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 10045d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 10055d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 10065d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 10075d4f98a2SYan Zheng 10085d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 100983d4cfd4SJosef Bacik ret = -EINVAL; 10105d4f98a2SYan Zheng goto out; 10115d4f98a2SYan Zheng } 10125d4f98a2SYan Zheng 10135d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10145d4f98a2SYan Zheng ret = 0; 10155d4f98a2SYan Zheng out: 10165d4f98a2SYan Zheng btrfs_free_path(path); 10175d4f98a2SYan Zheng return ret; 10185d4f98a2SYan Zheng } 10195d4f98a2SYan Zheng 10205d4f98a2SYan Zheng /* 10215d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 10225d4f98a2SYan Zheng * the new locations. 10235d4f98a2SYan Zheng */ 10243fd0a558SYan, Zheng static noinline_for_stack 10253fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 10265d4f98a2SYan Zheng struct reloc_control *rc, 10275d4f98a2SYan Zheng struct btrfs_root *root, 10283fd0a558SYan, Zheng struct extent_buffer *leaf) 10295d4f98a2SYan Zheng { 10300b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 10315d4f98a2SYan Zheng struct btrfs_key key; 10325d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10335d4f98a2SYan Zheng struct inode *inode = NULL; 10345d4f98a2SYan Zheng u64 parent; 10355d4f98a2SYan Zheng u64 bytenr; 10363fd0a558SYan, Zheng u64 new_bytenr = 0; 10375d4f98a2SYan Zheng u64 num_bytes; 10385d4f98a2SYan Zheng u64 end; 10395d4f98a2SYan Zheng u32 nritems; 10405d4f98a2SYan Zheng u32 i; 104183d4cfd4SJosef Bacik int ret = 0; 10425d4f98a2SYan Zheng int first = 1; 10435d4f98a2SYan Zheng int dirty = 0; 10445d4f98a2SYan Zheng 10455d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 10465d4f98a2SYan Zheng return 0; 10475d4f98a2SYan Zheng 10485d4f98a2SYan Zheng /* reloc trees always use full backref */ 10495d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 10505d4f98a2SYan Zheng parent = leaf->start; 10515d4f98a2SYan Zheng else 10525d4f98a2SYan Zheng parent = 0; 10535d4f98a2SYan Zheng 10545d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 10555d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 105682fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 105782fa113fSQu Wenruo 10585d4f98a2SYan Zheng cond_resched(); 10595d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 10605d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 10615d4f98a2SYan Zheng continue; 10625d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 10635d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 10645d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 10655d4f98a2SYan Zheng continue; 10665d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10675d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 10685d4f98a2SYan Zheng if (bytenr == 0) 10695d4f98a2SYan Zheng continue; 10709569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 10719569cc20SQu Wenruo rc->block_group->length)) 10725d4f98a2SYan Zheng continue; 10735d4f98a2SYan Zheng 10745d4f98a2SYan Zheng /* 10755d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 10765d4f98a2SYan Zheng * to complete and drop the extent cache 10775d4f98a2SYan Zheng */ 10785d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 10795d4f98a2SYan Zheng if (first) { 10805d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 10815d4f98a2SYan Zheng first = 0; 10824a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 10833fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 10845d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 10855d4f98a2SYan Zheng } 10864a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 10875d4f98a2SYan Zheng end = key.offset + 10885d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 10895d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 10900b246afaSJeff Mahoney fs_info->sectorsize)); 10910b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 10925d4f98a2SYan Zheng end--; 10935d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1094d0082371SJeff Mahoney key.offset, end); 10955d4f98a2SYan Zheng if (!ret) 10965d4f98a2SYan Zheng continue; 10975d4f98a2SYan Zheng 1098dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1099dcdbc059SNikolay Borisov key.offset, end, 1); 11005d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1101d0082371SJeff Mahoney key.offset, end); 11025d4f98a2SYan Zheng } 11035d4f98a2SYan Zheng } 11045d4f98a2SYan Zheng 11055d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 11065d4f98a2SYan Zheng bytenr, num_bytes); 110783d4cfd4SJosef Bacik if (ret) { 110883d4cfd4SJosef Bacik /* 110983d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 111083d4cfd4SJosef Bacik * in the file extent yet. 111183d4cfd4SJosef Bacik */ 111283d4cfd4SJosef Bacik break; 11133fd0a558SYan, Zheng } 11145d4f98a2SYan Zheng 11155d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 11165d4f98a2SYan Zheng dirty = 1; 11175d4f98a2SYan Zheng 11185d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 111982fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 112082fa113fSQu Wenruo num_bytes, parent); 112182fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 112282fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1123b06c4bf5SFilipe Manana key.objectid, key.offset); 112482fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 112583d4cfd4SJosef Bacik if (ret) { 112666642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 112783d4cfd4SJosef Bacik break; 112883d4cfd4SJosef Bacik } 11295d4f98a2SYan Zheng 1130ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1131ffd4bb2aSQu Wenruo num_bytes, parent); 1132ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1133ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1134b06c4bf5SFilipe Manana key.objectid, key.offset); 1135ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 113683d4cfd4SJosef Bacik if (ret) { 113766642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 113883d4cfd4SJosef Bacik break; 113983d4cfd4SJosef Bacik } 11405d4f98a2SYan Zheng } 11415d4f98a2SYan Zheng if (dirty) 11425d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 11433fd0a558SYan, Zheng if (inode) 11443fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 114583d4cfd4SJosef Bacik return ret; 11465d4f98a2SYan Zheng } 11475d4f98a2SYan Zheng 11485d4f98a2SYan Zheng static noinline_for_stack 11495d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 11505d4f98a2SYan Zheng struct btrfs_path *path, int level) 11515d4f98a2SYan Zheng { 11525d4f98a2SYan Zheng struct btrfs_disk_key key1; 11535d4f98a2SYan Zheng struct btrfs_disk_key key2; 11545d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 11555d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 11565d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 11575d4f98a2SYan Zheng } 11585d4f98a2SYan Zheng 11595d4f98a2SYan Zheng /* 11605d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 11615d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 11625d4f98a2SYan Zheng * reloc tree was create can be replaced. 11635d4f98a2SYan Zheng * 11645d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 11655d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 11665d4f98a2SYan Zheng * errors, a negative error number is returned. 11675d4f98a2SYan Zheng */ 11683fd0a558SYan, Zheng static noinline_for_stack 11693d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 11705d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 11715d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 11725d4f98a2SYan Zheng int lowest_level, int max_level) 11735d4f98a2SYan Zheng { 11740b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 11755d4f98a2SYan Zheng struct extent_buffer *eb; 11765d4f98a2SYan Zheng struct extent_buffer *parent; 117782fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 11785d4f98a2SYan Zheng struct btrfs_key key; 11795d4f98a2SYan Zheng u64 old_bytenr; 11805d4f98a2SYan Zheng u64 new_bytenr; 11815d4f98a2SYan Zheng u64 old_ptr_gen; 11825d4f98a2SYan Zheng u64 new_ptr_gen; 11835d4f98a2SYan Zheng u64 last_snapshot; 11845d4f98a2SYan Zheng u32 blocksize; 11853fd0a558SYan, Zheng int cow = 0; 11865d4f98a2SYan Zheng int level; 11875d4f98a2SYan Zheng int ret; 11885d4f98a2SYan Zheng int slot; 11895d4f98a2SYan Zheng 11905d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 11915d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 11925d4f98a2SYan Zheng 11935d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 11943fd0a558SYan, Zheng again: 11955d4f98a2SYan Zheng slot = path->slots[lowest_level]; 11965d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 11975d4f98a2SYan Zheng 11985d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 11998bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 12005d4f98a2SYan Zheng level = btrfs_header_level(eb); 12015d4f98a2SYan Zheng 12025d4f98a2SYan Zheng if (level < lowest_level) { 12035d4f98a2SYan Zheng btrfs_tree_unlock(eb); 12045d4f98a2SYan Zheng free_extent_buffer(eb); 12055d4f98a2SYan Zheng return 0; 12065d4f98a2SYan Zheng } 12075d4f98a2SYan Zheng 12083fd0a558SYan, Zheng if (cow) { 12095d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 12105d4f98a2SYan Zheng BUG_ON(ret); 12113fd0a558SYan, Zheng } 12128bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 12135d4f98a2SYan Zheng 12145d4f98a2SYan Zheng if (next_key) { 12155d4f98a2SYan Zheng next_key->objectid = (u64)-1; 12165d4f98a2SYan Zheng next_key->type = (u8)-1; 12175d4f98a2SYan Zheng next_key->offset = (u64)-1; 12185d4f98a2SYan Zheng } 12195d4f98a2SYan Zheng 12205d4f98a2SYan Zheng parent = eb; 12215d4f98a2SYan Zheng while (1) { 1222581c1760SQu Wenruo struct btrfs_key first_key; 1223581c1760SQu Wenruo 12245d4f98a2SYan Zheng level = btrfs_header_level(parent); 12255d4f98a2SYan Zheng BUG_ON(level < lowest_level); 12265d4f98a2SYan Zheng 1227e3b83361SQu Wenruo ret = btrfs_bin_search(parent, &key, &slot); 1228cbca7d59SFilipe Manana if (ret < 0) 1229cbca7d59SFilipe Manana break; 12305d4f98a2SYan Zheng if (ret && slot > 0) 12315d4f98a2SYan Zheng slot--; 12325d4f98a2SYan Zheng 12335d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 12345d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 12355d4f98a2SYan Zheng 12365d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 12370b246afaSJeff Mahoney blocksize = fs_info->nodesize; 12385d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 123917515f1bSQu Wenruo btrfs_node_key_to_cpu(parent, &first_key, 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 1264581c1760SQu Wenruo eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen, 1265581c1760SQu Wenruo level - 1, &first_key); 126664c043deSLiu Bo if (IS_ERR(eb)) { 126764c043deSLiu Bo ret = PTR_ERR(eb); 1268264813acSLiu Bo break; 126964c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 127064c043deSLiu Bo ret = -EIO; 1271416bc658SJosef Bacik free_extent_buffer(eb); 1272379cde74SStefan Behrens break; 1273416bc658SJosef Bacik } 12745d4f98a2SYan Zheng btrfs_tree_lock(eb); 12753fd0a558SYan, Zheng if (cow) { 12765d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 12775d4f98a2SYan Zheng slot, &eb); 12785d4f98a2SYan Zheng BUG_ON(ret); 12795d4f98a2SYan Zheng } 12808bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 12815d4f98a2SYan Zheng 12825d4f98a2SYan Zheng btrfs_tree_unlock(parent); 12835d4f98a2SYan Zheng free_extent_buffer(parent); 12845d4f98a2SYan Zheng 12855d4f98a2SYan Zheng parent = eb; 12865d4f98a2SYan Zheng continue; 12875d4f98a2SYan Zheng } 12885d4f98a2SYan Zheng 12893fd0a558SYan, Zheng if (!cow) { 12903fd0a558SYan, Zheng btrfs_tree_unlock(parent); 12913fd0a558SYan, Zheng free_extent_buffer(parent); 12923fd0a558SYan, Zheng cow = 1; 12933fd0a558SYan, Zheng goto again; 12943fd0a558SYan, Zheng } 12953fd0a558SYan, Zheng 12965d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 12975d4f98a2SYan Zheng path->slots[level]); 1298b3b4aa74SDavid Sterba btrfs_release_path(path); 12995d4f98a2SYan Zheng 13005d4f98a2SYan Zheng path->lowest_level = level; 13015d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 13025d4f98a2SYan Zheng path->lowest_level = 0; 13035d4f98a2SYan Zheng BUG_ON(ret); 13045d4f98a2SYan Zheng 13055d4f98a2SYan Zheng /* 1306824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 1307824d8dffSQu Wenruo * 1308824d8dffSQu Wenruo * We must trace both trees. 1309824d8dffSQu Wenruo * 1) Tree reloc subtree 1310824d8dffSQu Wenruo * If not traced, we will leak data numbers 1311824d8dffSQu Wenruo * 2) Fs subtree 1312824d8dffSQu Wenruo * If not traced, we will double count old data 1313f616f5cdSQu Wenruo * 1314f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 1315f616f5cdSQu Wenruo * the swapped tree blocks. 1316f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 1317f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 1318824d8dffSQu Wenruo */ 1319370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 1320370a11b8SQu Wenruo rc->block_group, parent, slot, 1321370a11b8SQu Wenruo path->nodes[level], path->slots[level], 1322370a11b8SQu Wenruo last_snapshot); 1323370a11b8SQu Wenruo if (ret < 0) 1324370a11b8SQu Wenruo break; 1325824d8dffSQu Wenruo /* 13265d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 13275d4f98a2SYan Zheng */ 13285d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 13295d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 13305d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 13315d4f98a2SYan Zheng 13325d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 13335d4f98a2SYan Zheng path->slots[level], old_bytenr); 13345d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 13355d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 13365d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 13375d4f98a2SYan Zheng 133882fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 133982fa113fSQu Wenruo blocksize, path->nodes[level]->start); 134082fa113fSQu Wenruo ref.skip_qgroup = true; 134182fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 134282fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13435d4f98a2SYan Zheng BUG_ON(ret); 134482fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 134582fa113fSQu Wenruo blocksize, 0); 134682fa113fSQu Wenruo ref.skip_qgroup = true; 134782fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 134882fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13495d4f98a2SYan Zheng BUG_ON(ret); 13505d4f98a2SYan Zheng 1351ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 1352ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 1353ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 1354ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1355ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13565d4f98a2SYan Zheng BUG_ON(ret); 13575d4f98a2SYan Zheng 1358ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 1359ffd4bb2aSQu Wenruo blocksize, 0); 1360ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 1361ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1362ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13635d4f98a2SYan Zheng BUG_ON(ret); 13645d4f98a2SYan Zheng 13655d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 13665d4f98a2SYan Zheng 13675d4f98a2SYan Zheng ret = level; 13685d4f98a2SYan Zheng break; 13695d4f98a2SYan Zheng } 13705d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13715d4f98a2SYan Zheng free_extent_buffer(parent); 13725d4f98a2SYan Zheng return ret; 13735d4f98a2SYan Zheng } 13745d4f98a2SYan Zheng 13755d4f98a2SYan Zheng /* 13765d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 13775d4f98a2SYan Zheng */ 13785d4f98a2SYan Zheng static noinline_for_stack 13795d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 13805d4f98a2SYan Zheng int *level) 13815d4f98a2SYan Zheng { 13825d4f98a2SYan Zheng struct extent_buffer *eb; 13835d4f98a2SYan Zheng int i; 13845d4f98a2SYan Zheng u64 last_snapshot; 13855d4f98a2SYan Zheng u32 nritems; 13865d4f98a2SYan Zheng 13875d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 13885d4f98a2SYan Zheng 13895d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 13905d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 13915d4f98a2SYan Zheng path->nodes[i] = NULL; 13925d4f98a2SYan Zheng } 13935d4f98a2SYan Zheng 13945d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 13955d4f98a2SYan Zheng eb = path->nodes[i]; 13965d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 13975d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 13985d4f98a2SYan Zheng path->slots[i]++; 13995d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 14005d4f98a2SYan Zheng last_snapshot) 14015d4f98a2SYan Zheng continue; 14025d4f98a2SYan Zheng 14035d4f98a2SYan Zheng *level = i; 14045d4f98a2SYan Zheng return 0; 14055d4f98a2SYan Zheng } 14065d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14075d4f98a2SYan Zheng path->nodes[i] = NULL; 14085d4f98a2SYan Zheng } 14095d4f98a2SYan Zheng return 1; 14105d4f98a2SYan Zheng } 14115d4f98a2SYan Zheng 14125d4f98a2SYan Zheng /* 14135d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 14145d4f98a2SYan Zheng */ 14155d4f98a2SYan Zheng static noinline_for_stack 14165d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14175d4f98a2SYan Zheng int *level) 14185d4f98a2SYan Zheng { 14192ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14205d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 14215d4f98a2SYan Zheng int i; 14225d4f98a2SYan Zheng u64 bytenr; 14235d4f98a2SYan Zheng u64 ptr_gen = 0; 14245d4f98a2SYan Zheng u64 last_snapshot; 14255d4f98a2SYan Zheng u32 nritems; 14265d4f98a2SYan Zheng 14275d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14285d4f98a2SYan Zheng 14295d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 1430581c1760SQu Wenruo struct btrfs_key first_key; 1431581c1760SQu Wenruo 14325d4f98a2SYan Zheng eb = path->nodes[i]; 14335d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14345d4f98a2SYan Zheng while (path->slots[i] < nritems) { 14355d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 14365d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 14375d4f98a2SYan Zheng break; 14385d4f98a2SYan Zheng path->slots[i]++; 14395d4f98a2SYan Zheng } 14405d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 14415d4f98a2SYan Zheng if (i == *level) 14425d4f98a2SYan Zheng break; 14435d4f98a2SYan Zheng *level = i + 1; 14445d4f98a2SYan Zheng return 0; 14455d4f98a2SYan Zheng } 14465d4f98a2SYan Zheng if (i == 1) { 14475d4f98a2SYan Zheng *level = i; 14485d4f98a2SYan Zheng return 0; 14495d4f98a2SYan Zheng } 14505d4f98a2SYan Zheng 14515d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 1452581c1760SQu Wenruo btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]); 1453581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1, 1454581c1760SQu Wenruo &first_key); 145564c043deSLiu Bo if (IS_ERR(eb)) { 145664c043deSLiu Bo return PTR_ERR(eb); 145764c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 1458416bc658SJosef Bacik free_extent_buffer(eb); 1459416bc658SJosef Bacik return -EIO; 1460416bc658SJosef Bacik } 14615d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 14625d4f98a2SYan Zheng path->nodes[i - 1] = eb; 14635d4f98a2SYan Zheng path->slots[i - 1] = 0; 14645d4f98a2SYan Zheng } 14655d4f98a2SYan Zheng return 1; 14665d4f98a2SYan Zheng } 14675d4f98a2SYan Zheng 14685d4f98a2SYan Zheng /* 14695d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 14705d4f98a2SYan Zheng * [min_key, max_key) 14715d4f98a2SYan Zheng */ 14725d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 14735d4f98a2SYan Zheng struct btrfs_key *min_key, 14745d4f98a2SYan Zheng struct btrfs_key *max_key) 14755d4f98a2SYan Zheng { 14760b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14775d4f98a2SYan Zheng struct inode *inode = NULL; 14785d4f98a2SYan Zheng u64 objectid; 14795d4f98a2SYan Zheng u64 start, end; 148033345d01SLi Zefan u64 ino; 14815d4f98a2SYan Zheng 14825d4f98a2SYan Zheng objectid = min_key->objectid; 14835d4f98a2SYan Zheng while (1) { 14845d4f98a2SYan Zheng cond_resched(); 14855d4f98a2SYan Zheng iput(inode); 14865d4f98a2SYan Zheng 14875d4f98a2SYan Zheng if (objectid > max_key->objectid) 14885d4f98a2SYan Zheng break; 14895d4f98a2SYan Zheng 14905d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 14915d4f98a2SYan Zheng if (!inode) 14925d4f98a2SYan Zheng break; 14934a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 14945d4f98a2SYan Zheng 149533345d01SLi Zefan if (ino > max_key->objectid) { 14965d4f98a2SYan Zheng iput(inode); 14975d4f98a2SYan Zheng break; 14985d4f98a2SYan Zheng } 14995d4f98a2SYan Zheng 150033345d01SLi Zefan objectid = ino + 1; 15015d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 15025d4f98a2SYan Zheng continue; 15035d4f98a2SYan Zheng 150433345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 15055d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 15065d4f98a2SYan Zheng continue; 15075d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 15085d4f98a2SYan Zheng start = 0; 15095d4f98a2SYan Zheng else { 15105d4f98a2SYan Zheng start = min_key->offset; 15110b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 15125d4f98a2SYan Zheng } 15135d4f98a2SYan Zheng } else { 15145d4f98a2SYan Zheng start = 0; 15155d4f98a2SYan Zheng } 15165d4f98a2SYan Zheng 151733345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 15185d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 15195d4f98a2SYan Zheng continue; 15205d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 15215d4f98a2SYan Zheng end = (u64)-1; 15225d4f98a2SYan Zheng } else { 15235d4f98a2SYan Zheng if (max_key->offset == 0) 15245d4f98a2SYan Zheng continue; 15255d4f98a2SYan Zheng end = max_key->offset; 15260b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 15275d4f98a2SYan Zheng end--; 15285d4f98a2SYan Zheng } 15295d4f98a2SYan Zheng } else { 15305d4f98a2SYan Zheng end = (u64)-1; 15315d4f98a2SYan Zheng } 15325d4f98a2SYan Zheng 15335d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 1534d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 1535dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 1536d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 15375d4f98a2SYan Zheng } 15385d4f98a2SYan Zheng return 0; 15395d4f98a2SYan Zheng } 15405d4f98a2SYan Zheng 15415d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 15425d4f98a2SYan Zheng struct btrfs_key *key) 15435d4f98a2SYan Zheng 15445d4f98a2SYan Zheng { 15455d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 15465d4f98a2SYan Zheng if (!path->nodes[level]) 15475d4f98a2SYan Zheng break; 15485d4f98a2SYan Zheng if (path->slots[level] + 1 < 15495d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 15505d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 15515d4f98a2SYan Zheng path->slots[level] + 1); 15525d4f98a2SYan Zheng return 0; 15535d4f98a2SYan Zheng } 15545d4f98a2SYan Zheng level++; 15555d4f98a2SYan Zheng } 15565d4f98a2SYan Zheng return 1; 15575d4f98a2SYan Zheng } 15585d4f98a2SYan Zheng 15595d4f98a2SYan Zheng /* 1560d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 1561d2311e69SQu Wenruo */ 1562d2311e69SQu Wenruo static void insert_dirty_subvol(struct btrfs_trans_handle *trans, 1563d2311e69SQu Wenruo struct reloc_control *rc, 1564d2311e69SQu Wenruo struct btrfs_root *root) 1565d2311e69SQu Wenruo { 1566d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1567d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 1568d2311e69SQu Wenruo 1569d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 1570d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 1571d2311e69SQu Wenruo ASSERT(reloc_root); 1572d2311e69SQu Wenruo 1573d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 1574d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 1575d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 1576d2311e69SQu Wenruo reloc_root_item->drop_level = 0; 1577d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 1578d2311e69SQu Wenruo btrfs_update_reloc_root(trans, root); 1579d2311e69SQu Wenruo 1580d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 158100246528SJosef Bacik btrfs_grab_root(root); 1582d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 1583d2311e69SQu Wenruo } 1584d2311e69SQu Wenruo } 1585d2311e69SQu Wenruo 1586d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 1587d2311e69SQu Wenruo { 1588d2311e69SQu Wenruo struct btrfs_root *root; 1589d2311e69SQu Wenruo struct btrfs_root *next; 1590d2311e69SQu Wenruo int ret = 0; 159130d40577SQu Wenruo int ret2; 1592d2311e69SQu Wenruo 1593d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 1594d2311e69SQu Wenruo reloc_dirty_list) { 159530d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 159630d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 1597d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1598d2311e69SQu Wenruo 1599d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 1600d2311e69SQu Wenruo root->reloc_root = NULL; 16016282675eSQu Wenruo /* 16026282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 16036282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 16046282675eSQu Wenruo */ 16056282675eSQu Wenruo smp_wmb(); 16061fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 1607f28de8d8SJosef Bacik if (reloc_root) { 1608f44deb74SJosef Bacik /* 1609f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 1610f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 1611f44deb74SJosef Bacik * drop the ref ourselves. 1612f44deb74SJosef Bacik */ 1613f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 1614f44deb74SJosef Bacik if (ret2 < 0) { 1615f44deb74SJosef Bacik btrfs_put_root(reloc_root); 1616f44deb74SJosef Bacik if (!ret) 1617f28de8d8SJosef Bacik ret = ret2; 1618f28de8d8SJosef Bacik } 1619f44deb74SJosef Bacik } 162000246528SJosef Bacik btrfs_put_root(root); 162130d40577SQu Wenruo } else { 162230d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 16230078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 1624f44deb74SJosef Bacik if (ret2 < 0) { 1625f44deb74SJosef Bacik btrfs_put_root(root); 1626f44deb74SJosef Bacik if (!ret) 162730d40577SQu Wenruo ret = ret2; 162830d40577SQu Wenruo } 1629d2311e69SQu Wenruo } 1630f44deb74SJosef Bacik } 1631d2311e69SQu Wenruo return ret; 1632d2311e69SQu Wenruo } 1633d2311e69SQu Wenruo 1634d2311e69SQu Wenruo /* 16355d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 16365d4f98a2SYan Zheng * fs tree. 16375d4f98a2SYan Zheng */ 16385d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 16395d4f98a2SYan Zheng struct btrfs_root *root) 16405d4f98a2SYan Zheng { 16410b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 16425d4f98a2SYan Zheng struct btrfs_key key; 16435d4f98a2SYan Zheng struct btrfs_key next_key; 16449e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 16455d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16465d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16475d4f98a2SYan Zheng struct btrfs_path *path; 16483fd0a558SYan, Zheng struct extent_buffer *leaf; 16495d4f98a2SYan Zheng int level; 16505d4f98a2SYan Zheng int max_level; 16515d4f98a2SYan Zheng int replaced = 0; 16525d4f98a2SYan Zheng int ret; 16535d4f98a2SYan Zheng int err = 0; 16543fd0a558SYan, Zheng u32 min_reserved; 16555d4f98a2SYan Zheng 16565d4f98a2SYan Zheng path = btrfs_alloc_path(); 16575d4f98a2SYan Zheng if (!path) 16585d4f98a2SYan Zheng return -ENOMEM; 1659e4058b54SDavid Sterba path->reada = READA_FORWARD; 16605d4f98a2SYan Zheng 16615d4f98a2SYan Zheng reloc_root = root->reloc_root; 16625d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16635d4f98a2SYan Zheng 16645d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 16655d4f98a2SYan Zheng level = btrfs_root_level(root_item); 166667439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 16675d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 16685d4f98a2SYan Zheng path->slots[level] = 0; 16695d4f98a2SYan Zheng } else { 16705d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 16715d4f98a2SYan Zheng 16725d4f98a2SYan Zheng level = root_item->drop_level; 16735d4f98a2SYan Zheng BUG_ON(level == 0); 16745d4f98a2SYan Zheng path->lowest_level = level; 16755d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 167633c66f43SYan Zheng path->lowest_level = 0; 16775d4f98a2SYan Zheng if (ret < 0) { 16785d4f98a2SYan Zheng btrfs_free_path(path); 16795d4f98a2SYan Zheng return ret; 16805d4f98a2SYan Zheng } 16815d4f98a2SYan Zheng 16825d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 16835d4f98a2SYan Zheng path->slots[level]); 16845d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 16855d4f98a2SYan Zheng 16865d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 16875d4f98a2SYan Zheng } 16885d4f98a2SYan Zheng 168944d354abSQu Wenruo /* 169044d354abSQu Wenruo * In merge_reloc_root(), we modify the upper level pointer to swap the 169144d354abSQu Wenruo * tree blocks between reloc tree and subvolume tree. Thus for tree 169244d354abSQu Wenruo * block COW, we COW at most from level 1 to root level for each tree. 169344d354abSQu Wenruo * 169444d354abSQu Wenruo * Thus the needed metadata size is at most root_level * nodesize, 169544d354abSQu Wenruo * and * 2 since we have two trees to COW. 169644d354abSQu Wenruo */ 169744d354abSQu Wenruo min_reserved = fs_info->nodesize * btrfs_root_level(root_item) * 2; 16985d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 16995d4f98a2SYan Zheng 17005d4f98a2SYan Zheng while (1) { 170108e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 170244d354abSQu Wenruo BTRFS_RESERVE_FLUSH_LIMIT); 17033fd0a558SYan, Zheng if (ret) { 17049e6a0c52SJosef Bacik err = ret; 17059e6a0c52SJosef Bacik goto out; 17063fd0a558SYan, Zheng } 17079e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 17089e6a0c52SJosef Bacik if (IS_ERR(trans)) { 17099e6a0c52SJosef Bacik err = PTR_ERR(trans); 17109e6a0c52SJosef Bacik trans = NULL; 17119e6a0c52SJosef Bacik goto out; 17129e6a0c52SJosef Bacik } 17132abc726aSJosef Bacik 17142abc726aSJosef Bacik /* 17152abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 17162abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 17172abc726aSJosef Bacik * 17182abc726aSJosef Bacik * But that's ok, we started the trans handle on our 17192abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 17202abc726aSJosef Bacik * dirty list. At commit time we'll still call 17212abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 17222abc726aSJosef Bacik * appropriately. 17232abc726aSJosef Bacik */ 17242abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 17259e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 17263fd0a558SYan, Zheng 17273fd0a558SYan, Zheng replaced = 0; 17285d4f98a2SYan Zheng max_level = level; 17295d4f98a2SYan Zheng 17305d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 17315d4f98a2SYan Zheng if (ret < 0) { 17325d4f98a2SYan Zheng err = ret; 17335d4f98a2SYan Zheng goto out; 17345d4f98a2SYan Zheng } 17355d4f98a2SYan Zheng if (ret > 0) 17365d4f98a2SYan Zheng break; 17375d4f98a2SYan Zheng 17385d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 17395d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 17405d4f98a2SYan Zheng ret = 0; 17415d4f98a2SYan Zheng } else { 17423d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 17433fd0a558SYan, Zheng &next_key, level, max_level); 17445d4f98a2SYan Zheng } 17455d4f98a2SYan Zheng if (ret < 0) { 17465d4f98a2SYan Zheng err = ret; 17475d4f98a2SYan Zheng goto out; 17485d4f98a2SYan Zheng } 17495d4f98a2SYan Zheng 17505d4f98a2SYan Zheng if (ret > 0) { 17515d4f98a2SYan Zheng level = ret; 17525d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17535d4f98a2SYan Zheng path->slots[level]); 17545d4f98a2SYan Zheng replaced = 1; 17555d4f98a2SYan Zheng } 17565d4f98a2SYan Zheng 17575d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 17585d4f98a2SYan Zheng if (ret > 0) 17595d4f98a2SYan Zheng break; 17605d4f98a2SYan Zheng 17615d4f98a2SYan Zheng BUG_ON(level == 0); 17625d4f98a2SYan Zheng /* 17635d4f98a2SYan Zheng * save the merging progress in the drop_progress. 17645d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 17655d4f98a2SYan Zheng */ 17665d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 17675d4f98a2SYan Zheng path->slots[level]); 17685d4f98a2SYan Zheng root_item->drop_level = level; 17695d4f98a2SYan Zheng 17703a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 17719e6a0c52SJosef Bacik trans = NULL; 17725d4f98a2SYan Zheng 17732ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 17745d4f98a2SYan Zheng 17755d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17765d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17775d4f98a2SYan Zheng } 17785d4f98a2SYan Zheng 17795d4f98a2SYan Zheng /* 17805d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 17815d4f98a2SYan Zheng * relocated and the block is tree root. 17825d4f98a2SYan Zheng */ 17835d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 17845d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 17855d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 17865d4f98a2SYan Zheng free_extent_buffer(leaf); 17875d4f98a2SYan Zheng if (ret < 0) 17885d4f98a2SYan Zheng err = ret; 17895d4f98a2SYan Zheng out: 17905d4f98a2SYan Zheng btrfs_free_path(path); 17915d4f98a2SYan Zheng 1792d2311e69SQu Wenruo if (err == 0) 1793d2311e69SQu Wenruo insert_dirty_subvol(trans, rc, root); 17945d4f98a2SYan Zheng 17959e6a0c52SJosef Bacik if (trans) 17963a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 17975d4f98a2SYan Zheng 17982ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 17995d4f98a2SYan Zheng 18005d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 18015d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 18025d4f98a2SYan Zheng 18035d4f98a2SYan Zheng return err; 18045d4f98a2SYan Zheng } 18055d4f98a2SYan Zheng 18063fd0a558SYan, Zheng static noinline_for_stack 18073fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 18085d4f98a2SYan Zheng { 18093fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 18100b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 18113fd0a558SYan, Zheng struct btrfs_root *reloc_root; 18125d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 18133fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 18143fd0a558SYan, Zheng u64 num_bytes = 0; 18153fd0a558SYan, Zheng int ret; 18163fd0a558SYan, Zheng 18170b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 18180b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 18193fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 18200b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 18217585717fSChris Mason 18223fd0a558SYan, Zheng again: 18233fd0a558SYan, Zheng if (!err) { 18243fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 182508e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 182608e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 18273fd0a558SYan, Zheng if (ret) 18283fd0a558SYan, Zheng err = ret; 18293fd0a558SYan, Zheng } 18303fd0a558SYan, Zheng 18317a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 18323612b495STsutomu Itoh if (IS_ERR(trans)) { 18333612b495STsutomu Itoh if (!err) 18342ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 183563f018beSNikolay Borisov num_bytes, NULL); 18363612b495STsutomu Itoh return PTR_ERR(trans); 18373612b495STsutomu Itoh } 18383fd0a558SYan, Zheng 18393fd0a558SYan, Zheng if (!err) { 18403fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 18413a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18422ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 184363f018beSNikolay Borisov num_bytes, NULL); 18443fd0a558SYan, Zheng goto again; 18453fd0a558SYan, Zheng } 18463fd0a558SYan, Zheng } 18473fd0a558SYan, Zheng 18483fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 18493fd0a558SYan, Zheng 18503fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 18513fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 18523fd0a558SYan, Zheng struct btrfs_root, root_list); 18533fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 18543fd0a558SYan, Zheng 1855a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1856a820feb5SDavid Sterba false); 18573fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 18583fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 18593fd0a558SYan, Zheng 18603fd0a558SYan, Zheng /* 18613fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 18623fd0a558SYan, Zheng * knows it should resumes merging 18633fd0a558SYan, Zheng */ 18643fd0a558SYan, Zheng if (!err) 18653fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 18663fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 18673fd0a558SYan, Zheng 18683fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 186900246528SJosef Bacik btrfs_put_root(root); 18703fd0a558SYan, Zheng } 18713fd0a558SYan, Zheng 18723fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 18733fd0a558SYan, Zheng 18743fd0a558SYan, Zheng if (!err) 18753a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 18763fd0a558SYan, Zheng else 18773a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18783fd0a558SYan, Zheng return err; 18793fd0a558SYan, Zheng } 18803fd0a558SYan, Zheng 18813fd0a558SYan, Zheng static noinline_for_stack 1882aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 1883aca1bba6SLiu Bo { 1884a7571232SNikolay Borisov struct btrfs_root *reloc_root, *tmp; 1885aca1bba6SLiu Bo 1886a7571232SNikolay Borisov list_for_each_entry_safe(reloc_root, tmp, list, root_list) 1887bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 1888aca1bba6SLiu Bo } 1889aca1bba6SLiu Bo 1890aca1bba6SLiu Bo static noinline_for_stack 189194404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 18923fd0a558SYan, Zheng { 18930b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 18945d4f98a2SYan Zheng struct btrfs_root *root; 18955d4f98a2SYan Zheng struct btrfs_root *reloc_root; 18963fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 18973fd0a558SYan, Zheng int found = 0; 1898aca1bba6SLiu Bo int ret = 0; 18993fd0a558SYan, Zheng again: 19003fd0a558SYan, Zheng root = rc->extent_root; 19017585717fSChris Mason 19027585717fSChris Mason /* 19037585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 19047585717fSChris Mason * we have to make sure nobody is in the middle of 19057585717fSChris Mason * adding their roots to the list while we are 19067585717fSChris Mason * doing this splice 19077585717fSChris Mason */ 19080b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 19093fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 19100b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 19115d4f98a2SYan Zheng 19123fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 19133fd0a558SYan, Zheng found = 1; 19143fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 19153fd0a558SYan, Zheng struct btrfs_root, root_list); 19165d4f98a2SYan Zheng 1917a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1918a820feb5SDavid Sterba false); 19195d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 19205d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 19215d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 19223fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 192300246528SJosef Bacik btrfs_put_root(root); 1924b37b39cdSJosef Bacik if (ret) { 192525e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 192625e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 192725e293c2SWang Shilong &reloc_roots); 1928aca1bba6SLiu Bo goto out; 1929b37b39cdSJosef Bacik } 19303fd0a558SYan, Zheng } else { 193151415b6cSQu Wenruo if (!IS_ERR(root)) { 193251415b6cSQu Wenruo if (root->reloc_root == reloc_root) { 193351415b6cSQu Wenruo root->reloc_root = NULL; 193451415b6cSQu Wenruo btrfs_put_root(reloc_root); 193551415b6cSQu Wenruo } 19361dae7e0eSQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, 19371dae7e0eSQu Wenruo &root->state); 193851415b6cSQu Wenruo btrfs_put_root(root); 193951415b6cSQu Wenruo } 194051415b6cSQu Wenruo 19413fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 194230d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 194330d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 194430d40577SQu Wenruo &rc->dirty_subvol_roots); 19453fd0a558SYan, Zheng } 19465d4f98a2SYan Zheng } 19475d4f98a2SYan Zheng 19483fd0a558SYan, Zheng if (found) { 19493fd0a558SYan, Zheng found = 0; 19503fd0a558SYan, Zheng goto again; 19515d4f98a2SYan Zheng } 1952aca1bba6SLiu Bo out: 1953aca1bba6SLiu Bo if (ret) { 19540b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 1955aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 1956467bb1d2SWang Shilong 1957467bb1d2SWang Shilong /* new reloc root may be added */ 19580b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 1959467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 19600b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 1961467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 1962aca1bba6SLiu Bo } 1963aca1bba6SLiu Bo 19647b7b7431SJosef Bacik /* 19657b7b7431SJosef Bacik * We used to have 19667b7b7431SJosef Bacik * 19677b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 19687b7b7431SJosef Bacik * 19697b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 19707b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 19717b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 19727b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 19737b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 19747b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 19757b7b7431SJosef Bacik * will be cleaned up on unmount. 19767b7b7431SJosef Bacik * 19777b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 19787b7b7431SJosef Bacik */ 19795d4f98a2SYan Zheng } 19805d4f98a2SYan Zheng 19815d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 19825d4f98a2SYan Zheng { 19835d4f98a2SYan Zheng struct tree_block *block; 19845d4f98a2SYan Zheng struct rb_node *rb_node; 19855d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 19865d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 19875d4f98a2SYan Zheng rb_erase(rb_node, blocks); 19885d4f98a2SYan Zheng kfree(block); 19895d4f98a2SYan Zheng } 19905d4f98a2SYan Zheng } 19915d4f98a2SYan Zheng 19925d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 19935d4f98a2SYan Zheng struct btrfs_root *reloc_root) 19945d4f98a2SYan Zheng { 19950b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 19965d4f98a2SYan Zheng struct btrfs_root *root; 1997442b1ac5SJosef Bacik int ret; 19985d4f98a2SYan Zheng 19995d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 20005d4f98a2SYan Zheng return 0; 20015d4f98a2SYan Zheng 2002a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false); 20035d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 20045d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 2005442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 200600246528SJosef Bacik btrfs_put_root(root); 20075d4f98a2SYan Zheng 2008442b1ac5SJosef Bacik return ret; 20095d4f98a2SYan Zheng } 20105d4f98a2SYan Zheng 20113fd0a558SYan, Zheng static noinline_for_stack 20123fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 20133fd0a558SYan, Zheng struct reloc_control *rc, 2014a26195a5SQu Wenruo struct btrfs_backref_node *node, 2015a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 20165d4f98a2SYan Zheng { 2017a26195a5SQu Wenruo struct btrfs_backref_node *next; 20185d4f98a2SYan Zheng struct btrfs_root *root; 20193fd0a558SYan, Zheng int index = 0; 20203fd0a558SYan, Zheng 20215d4f98a2SYan Zheng next = node; 20225d4f98a2SYan Zheng while (1) { 20235d4f98a2SYan Zheng cond_resched(); 20245d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 20255d4f98a2SYan Zheng root = next->root; 20263fd0a558SYan, Zheng BUG_ON(!root); 202792a7cc42SQu Wenruo BUG_ON(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)); 20285d4f98a2SYan Zheng 20295d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 20305d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 20315d4f98a2SYan Zheng break; 20325d4f98a2SYan Zheng } 20335d4f98a2SYan Zheng 20345d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 20353fd0a558SYan, Zheng root = root->reloc_root; 20363fd0a558SYan, Zheng 20373fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 20383fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 20393fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 20403fd0a558SYan, Zheng next->new_bytenr = root->node->start; 204100246528SJosef Bacik btrfs_put_root(next->root); 204200246528SJosef Bacik next->root = btrfs_grab_root(root); 20430b530bc5SJosef Bacik ASSERT(next->root); 20443fd0a558SYan, Zheng list_add_tail(&next->list, 20453fd0a558SYan, Zheng &rc->backref_cache.changed); 20469569cc20SQu Wenruo mark_block_processed(rc, next); 20475d4f98a2SYan Zheng break; 20485d4f98a2SYan Zheng } 20495d4f98a2SYan Zheng 20503fd0a558SYan, Zheng WARN_ON(1); 20515d4f98a2SYan Zheng root = NULL; 20525d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 20535d4f98a2SYan Zheng if (!next || next->level <= node->level) 20545d4f98a2SYan Zheng break; 20555d4f98a2SYan Zheng } 20563fd0a558SYan, Zheng if (!root) 20573fd0a558SYan, Zheng return NULL; 20585d4f98a2SYan Zheng 20593fd0a558SYan, Zheng next = node; 20603fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 20613fd0a558SYan, Zheng while (1) { 20623fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 20633fd0a558SYan, Zheng if (--index < 0) 20643fd0a558SYan, Zheng break; 20653fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 20663fd0a558SYan, Zheng } 20675d4f98a2SYan Zheng return root; 20685d4f98a2SYan Zheng } 20695d4f98a2SYan Zheng 20703fd0a558SYan, Zheng /* 207192a7cc42SQu Wenruo * Select a tree root for relocation. 207292a7cc42SQu Wenruo * 207392a7cc42SQu Wenruo * Return NULL if the block is not shareable. We should use do_relocation() in 207492a7cc42SQu Wenruo * this case. 207592a7cc42SQu Wenruo * 207692a7cc42SQu Wenruo * Return a tree root pointer if the block is shareable. 207792a7cc42SQu Wenruo * Return -ENOENT if the block is root of reloc tree. 20783fd0a558SYan, Zheng */ 20795d4f98a2SYan Zheng static noinline_for_stack 2080a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 20815d4f98a2SYan Zheng { 2082a26195a5SQu Wenruo struct btrfs_backref_node *next; 20833fd0a558SYan, Zheng struct btrfs_root *root; 20843fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2085a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 20863fd0a558SYan, Zheng int index = 0; 20873fd0a558SYan, Zheng 20883fd0a558SYan, Zheng next = node; 20893fd0a558SYan, Zheng while (1) { 20903fd0a558SYan, Zheng cond_resched(); 20913fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 20923fd0a558SYan, Zheng root = next->root; 20933fd0a558SYan, Zheng BUG_ON(!root); 20943fd0a558SYan, Zheng 209592a7cc42SQu Wenruo /* No other choice for non-shareable tree */ 209692a7cc42SQu Wenruo if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 20973fd0a558SYan, Zheng return root; 20983fd0a558SYan, Zheng 20993fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 21003fd0a558SYan, Zheng fs_root = root; 21013fd0a558SYan, Zheng 21023fd0a558SYan, Zheng if (next != node) 21033fd0a558SYan, Zheng return NULL; 21043fd0a558SYan, Zheng 21053fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 21063fd0a558SYan, Zheng if (!next || next->level <= node->level) 21073fd0a558SYan, Zheng break; 21083fd0a558SYan, Zheng } 21093fd0a558SYan, Zheng 21103fd0a558SYan, Zheng if (!fs_root) 21113fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 21123fd0a558SYan, Zheng return fs_root; 21135d4f98a2SYan Zheng } 21145d4f98a2SYan Zheng 21155d4f98a2SYan Zheng static noinline_for_stack 21163fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2117a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 21185d4f98a2SYan Zheng { 21190b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2120a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2121a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2122a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 21233fd0a558SYan, Zheng u64 num_bytes = 0; 21243fd0a558SYan, Zheng int index = 0; 21255d4f98a2SYan Zheng 21263fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 21273fd0a558SYan, Zheng 21283fd0a558SYan, Zheng while (next) { 21293fd0a558SYan, Zheng cond_resched(); 21305d4f98a2SYan Zheng while (1) { 21313fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 21325d4f98a2SYan Zheng break; 21335d4f98a2SYan Zheng 21340b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 21353fd0a558SYan, Zheng 21363fd0a558SYan, Zheng if (list_empty(&next->upper)) 21373fd0a558SYan, Zheng break; 21383fd0a558SYan, Zheng 21393fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2140a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 21413fd0a558SYan, Zheng edges[index++] = edge; 21423fd0a558SYan, Zheng next = edge->node[UPPER]; 21435d4f98a2SYan Zheng } 21443fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 21453fd0a558SYan, Zheng } 21463fd0a558SYan, Zheng return num_bytes; 21473fd0a558SYan, Zheng } 21483fd0a558SYan, Zheng 21493fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 21503fd0a558SYan, Zheng struct reloc_control *rc, 2151a26195a5SQu Wenruo struct btrfs_backref_node *node) 21523fd0a558SYan, Zheng { 21533fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2154da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 21553fd0a558SYan, Zheng u64 num_bytes; 21563fd0a558SYan, Zheng int ret; 21570647bf56SWang Shilong u64 tmp; 21583fd0a558SYan, Zheng 21593fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 21603fd0a558SYan, Zheng 21613fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 21620647bf56SWang Shilong rc->reserved_bytes += num_bytes; 21638ca17f0fSJosef Bacik 21648ca17f0fSJosef Bacik /* 21658ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 21668ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 21678ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 21688ca17f0fSJosef Bacik */ 21690647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 21708ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 21713fd0a558SYan, Zheng if (ret) { 2172da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 21730647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 21740647bf56SWang Shilong tmp <<= 1; 21750647bf56SWang Shilong /* 21760647bf56SWang Shilong * only one thread can access block_rsv at this point, 21770647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 21780647bf56SWang Shilong * we expand more reservation size here to allow enough 217952042d8eSAndrea Gelmini * space for relocation and we will return earlier in 21800647bf56SWang Shilong * enospc case. 21810647bf56SWang Shilong */ 2182da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 21830647bf56SWang Shilong RELOCATION_RESERVED_NODES; 21848ca17f0fSJosef Bacik return -EAGAIN; 21853fd0a558SYan, Zheng } 21863fd0a558SYan, Zheng 21873fd0a558SYan, Zheng return 0; 21883fd0a558SYan, Zheng } 21893fd0a558SYan, Zheng 21905d4f98a2SYan Zheng /* 21915d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 21925d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 21935d4f98a2SYan Zheng * 21945d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 21955d4f98a2SYan Zheng * in that case this function just updates pointers. 21965d4f98a2SYan Zheng */ 21975d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 21983fd0a558SYan, Zheng struct reloc_control *rc, 2199a26195a5SQu Wenruo struct btrfs_backref_node *node, 22005d4f98a2SYan Zheng struct btrfs_key *key, 22015d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 22025d4f98a2SYan Zheng { 22032ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2204a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2205a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2206a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 22075d4f98a2SYan Zheng struct btrfs_root *root; 22085d4f98a2SYan Zheng struct extent_buffer *eb; 22095d4f98a2SYan Zheng u32 blocksize; 22105d4f98a2SYan Zheng u64 bytenr; 22115d4f98a2SYan Zheng u64 generation; 22125d4f98a2SYan Zheng int slot; 22135d4f98a2SYan Zheng int ret; 22145d4f98a2SYan Zheng int err = 0; 22155d4f98a2SYan Zheng 22165d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 22175d4f98a2SYan Zheng 22185d4f98a2SYan Zheng path->lowest_level = node->level + 1; 22193fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 22205d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 2221581c1760SQu Wenruo struct btrfs_key first_key; 222282fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2223581c1760SQu Wenruo 22245d4f98a2SYan Zheng cond_resched(); 22255d4f98a2SYan Zheng 22265d4f98a2SYan Zheng upper = edge->node[UPPER]; 2227dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 22283fd0a558SYan, Zheng BUG_ON(!root); 22295d4f98a2SYan Zheng 22303fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 22313fd0a558SYan, Zheng if (!lowest) { 2232e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 2233cbca7d59SFilipe Manana if (ret < 0) { 2234cbca7d59SFilipe Manana err = ret; 2235cbca7d59SFilipe Manana goto next; 2236cbca7d59SFilipe Manana } 22373fd0a558SYan, Zheng BUG_ON(ret); 22383fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 22393fd0a558SYan, Zheng if (node->eb->start == bytenr) 22403fd0a558SYan, Zheng goto next; 22413fd0a558SYan, Zheng } 2242b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 22433fd0a558SYan, Zheng } 22445d4f98a2SYan Zheng 22455d4f98a2SYan Zheng if (!upper->eb) { 22465d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 22473561b9dbSLiu Bo if (ret) { 22483561b9dbSLiu Bo if (ret < 0) 22495d4f98a2SYan Zheng err = ret; 22503561b9dbSLiu Bo else 22513561b9dbSLiu Bo err = -ENOENT; 22523561b9dbSLiu Bo 22533561b9dbSLiu Bo btrfs_release_path(path); 22545d4f98a2SYan Zheng break; 22555d4f98a2SYan Zheng } 22565d4f98a2SYan Zheng 22573fd0a558SYan, Zheng if (!upper->eb) { 22583fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 22593fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 22603fd0a558SYan, Zheng } else { 22613fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 22623fd0a558SYan, Zheng } 22633fd0a558SYan, Zheng 22643fd0a558SYan, Zheng upper->locked = 1; 22653fd0a558SYan, Zheng path->locks[upper->level] = 0; 22663fd0a558SYan, Zheng 22675d4f98a2SYan Zheng slot = path->slots[upper->level]; 2268b3b4aa74SDavid Sterba btrfs_release_path(path); 22695d4f98a2SYan Zheng } else { 2270e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 2271cbca7d59SFilipe Manana if (ret < 0) { 2272cbca7d59SFilipe Manana err = ret; 2273cbca7d59SFilipe Manana goto next; 2274cbca7d59SFilipe Manana } 22755d4f98a2SYan Zheng BUG_ON(ret); 22765d4f98a2SYan Zheng } 22775d4f98a2SYan Zheng 22785d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 22793fd0a558SYan, Zheng if (lowest) { 22804547f4d8SLiu Bo if (bytenr != node->bytenr) { 22814547f4d8SLiu Bo btrfs_err(root->fs_info, 22824547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 22834547f4d8SLiu Bo bytenr, node->bytenr, slot, 22844547f4d8SLiu Bo upper->eb->start); 22854547f4d8SLiu Bo err = -EIO; 22864547f4d8SLiu Bo goto next; 22874547f4d8SLiu Bo } 22885d4f98a2SYan Zheng } else { 22893fd0a558SYan, Zheng if (node->eb->start == bytenr) 22903fd0a558SYan, Zheng goto next; 22915d4f98a2SYan Zheng } 22925d4f98a2SYan Zheng 2293da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 22945d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 2295581c1760SQu Wenruo btrfs_node_key_to_cpu(upper->eb, &first_key, slot); 2296581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, generation, 2297581c1760SQu Wenruo upper->level - 1, &first_key); 229864c043deSLiu Bo if (IS_ERR(eb)) { 229964c043deSLiu Bo err = PTR_ERR(eb); 230064c043deSLiu Bo goto next; 230164c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2302416bc658SJosef Bacik free_extent_buffer(eb); 230397d9a8a4STsutomu Itoh err = -EIO; 230497d9a8a4STsutomu Itoh goto next; 230597d9a8a4STsutomu Itoh } 23065d4f98a2SYan Zheng btrfs_tree_lock(eb); 23078bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 23085d4f98a2SYan Zheng 23095d4f98a2SYan Zheng if (!node->eb) { 23105d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 23115d4f98a2SYan Zheng slot, &eb); 23123fd0a558SYan, Zheng btrfs_tree_unlock(eb); 23133fd0a558SYan, Zheng free_extent_buffer(eb); 23145d4f98a2SYan Zheng if (ret < 0) { 23155d4f98a2SYan Zheng err = ret; 23163fd0a558SYan, Zheng goto next; 23175d4f98a2SYan Zheng } 23183fd0a558SYan, Zheng BUG_ON(node->eb != eb); 23195d4f98a2SYan Zheng } else { 23205d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 23215d4f98a2SYan Zheng node->eb->start); 23225d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 23235d4f98a2SYan Zheng trans->transid); 23245d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 23255d4f98a2SYan Zheng 232682fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 23275d4f98a2SYan Zheng node->eb->start, blocksize, 232882fa113fSQu Wenruo upper->eb->start); 232982fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 233082fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 233182fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 233282fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 23335d4f98a2SYan Zheng BUG_ON(ret); 23345d4f98a2SYan Zheng 23355d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 23365d4f98a2SYan Zheng BUG_ON(ret); 23375d4f98a2SYan Zheng } 23383fd0a558SYan, Zheng next: 23393fd0a558SYan, Zheng if (!upper->pending) 2340b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 23413fd0a558SYan, Zheng else 2342b0fe7078SQu Wenruo btrfs_backref_unlock_node_buffer(upper); 23433fd0a558SYan, Zheng if (err) 23443fd0a558SYan, Zheng break; 23455d4f98a2SYan Zheng } 23463fd0a558SYan, Zheng 23473fd0a558SYan, Zheng if (!err && node->pending) { 2348b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 23493fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 23503fd0a558SYan, Zheng node->pending = 0; 23515d4f98a2SYan Zheng } 23523fd0a558SYan, Zheng 23535d4f98a2SYan Zheng path->lowest_level = 0; 23543fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 23555d4f98a2SYan Zheng return err; 23565d4f98a2SYan Zheng } 23575d4f98a2SYan Zheng 23585d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 23593fd0a558SYan, Zheng struct reloc_control *rc, 2360a26195a5SQu Wenruo struct btrfs_backref_node *node, 23615d4f98a2SYan Zheng struct btrfs_path *path) 23625d4f98a2SYan Zheng { 23635d4f98a2SYan Zheng struct btrfs_key key; 23645d4f98a2SYan Zheng 23655d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 23663fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 23675d4f98a2SYan Zheng } 23685d4f98a2SYan Zheng 23695d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 23703fd0a558SYan, Zheng struct reloc_control *rc, 23713fd0a558SYan, Zheng struct btrfs_path *path, int err) 23725d4f98a2SYan Zheng { 23733fd0a558SYan, Zheng LIST_HEAD(list); 2374a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 2375a26195a5SQu Wenruo struct btrfs_backref_node *node; 23765d4f98a2SYan Zheng int level; 23775d4f98a2SYan Zheng int ret; 23785d4f98a2SYan Zheng 23795d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 23805d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 23815d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 2382a26195a5SQu Wenruo struct btrfs_backref_node, list); 23833fd0a558SYan, Zheng list_move_tail(&node->list, &list); 23843fd0a558SYan, Zheng BUG_ON(!node->pending); 23855d4f98a2SYan Zheng 23863fd0a558SYan, Zheng if (!err) { 23873fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 23885d4f98a2SYan Zheng if (ret < 0) 23895d4f98a2SYan Zheng err = ret; 23905d4f98a2SYan Zheng } 23915d4f98a2SYan Zheng } 23923fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 23933fd0a558SYan, Zheng } 23945d4f98a2SYan Zheng return err; 23955d4f98a2SYan Zheng } 23965d4f98a2SYan Zheng 23975d4f98a2SYan Zheng /* 23985d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 23995d4f98a2SYan Zheng * as processed. 24005d4f98a2SYan Zheng */ 24015d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 2402a26195a5SQu Wenruo struct btrfs_backref_node *node) 24035d4f98a2SYan Zheng { 2404a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2405a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2406a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24075d4f98a2SYan Zheng int index = 0; 24085d4f98a2SYan Zheng 24095d4f98a2SYan Zheng while (next) { 24105d4f98a2SYan Zheng cond_resched(); 24115d4f98a2SYan Zheng while (1) { 24125d4f98a2SYan Zheng if (next->processed) 24135d4f98a2SYan Zheng break; 24145d4f98a2SYan Zheng 24159569cc20SQu Wenruo mark_block_processed(rc, next); 24165d4f98a2SYan Zheng 24175d4f98a2SYan Zheng if (list_empty(&next->upper)) 24185d4f98a2SYan Zheng break; 24195d4f98a2SYan Zheng 24205d4f98a2SYan Zheng edge = list_entry(next->upper.next, 2421a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 24225d4f98a2SYan Zheng edges[index++] = edge; 24235d4f98a2SYan Zheng next = edge->node[UPPER]; 24245d4f98a2SYan Zheng } 24255d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 24265d4f98a2SYan Zheng } 24275d4f98a2SYan Zheng } 24285d4f98a2SYan Zheng 24297476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 24305d4f98a2SYan Zheng { 2431da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 24327476dfdaSDavid Sterba 24335d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 24349655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 24355d4f98a2SYan Zheng return 1; 24365d4f98a2SYan Zheng return 0; 24375d4f98a2SYan Zheng } 24385d4f98a2SYan Zheng 24392ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 24405d4f98a2SYan Zheng struct tree_block *block) 24415d4f98a2SYan Zheng { 24425d4f98a2SYan Zheng struct extent_buffer *eb; 24435d4f98a2SYan Zheng 2444581c1760SQu Wenruo eb = read_tree_block(fs_info, block->bytenr, block->key.offset, 2445581c1760SQu Wenruo block->level, NULL); 244664c043deSLiu Bo if (IS_ERR(eb)) { 244764c043deSLiu Bo return PTR_ERR(eb); 244864c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2449416bc658SJosef Bacik free_extent_buffer(eb); 2450416bc658SJosef Bacik return -EIO; 2451416bc658SJosef Bacik } 24525d4f98a2SYan Zheng if (block->level == 0) 24535d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 24545d4f98a2SYan Zheng else 24555d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 24565d4f98a2SYan Zheng free_extent_buffer(eb); 24575d4f98a2SYan Zheng block->key_ready = 1; 24585d4f98a2SYan Zheng return 0; 24595d4f98a2SYan Zheng } 24605d4f98a2SYan Zheng 24615d4f98a2SYan Zheng /* 24625d4f98a2SYan Zheng * helper function to relocate a tree block 24635d4f98a2SYan Zheng */ 24645d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 24655d4f98a2SYan Zheng struct reloc_control *rc, 2466a26195a5SQu Wenruo struct btrfs_backref_node *node, 24675d4f98a2SYan Zheng struct btrfs_key *key, 24685d4f98a2SYan Zheng struct btrfs_path *path) 24695d4f98a2SYan Zheng { 24705d4f98a2SYan Zheng struct btrfs_root *root; 24713fd0a558SYan, Zheng int ret = 0; 24725d4f98a2SYan Zheng 24733fd0a558SYan, Zheng if (!node) 24745d4f98a2SYan Zheng return 0; 24753fd0a558SYan, Zheng 24765f6b2e5cSJosef Bacik /* 24775f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 24785f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 24795f6b2e5cSJosef Bacik */ 24805f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 24815f6b2e5cSJosef Bacik if (ret) 24825f6b2e5cSJosef Bacik goto out; 24835f6b2e5cSJosef Bacik 24843fd0a558SYan, Zheng BUG_ON(node->processed); 2485147d256eSZhaolei root = select_one_root(node); 24863fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 24873fd0a558SYan, Zheng update_processed_blocks(rc, node); 24883fd0a558SYan, Zheng goto out; 24895d4f98a2SYan Zheng } 24905d4f98a2SYan Zheng 24913fd0a558SYan, Zheng if (root) { 249292a7cc42SQu Wenruo if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 24933fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 24943fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 24953fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 24963fd0a558SYan, Zheng root = root->reloc_root; 24973fd0a558SYan, Zheng node->new_bytenr = root->node->start; 249800246528SJosef Bacik btrfs_put_root(node->root); 249900246528SJosef Bacik node->root = btrfs_grab_root(root); 25000b530bc5SJosef Bacik ASSERT(node->root); 25013fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 25023fd0a558SYan, Zheng } else { 25035d4f98a2SYan Zheng path->lowest_level = node->level; 25045d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2505b3b4aa74SDavid Sterba btrfs_release_path(path); 25063fd0a558SYan, Zheng if (ret > 0) 25075d4f98a2SYan Zheng ret = 0; 25083fd0a558SYan, Zheng } 25093fd0a558SYan, Zheng if (!ret) 25103fd0a558SYan, Zheng update_processed_blocks(rc, node); 25113fd0a558SYan, Zheng } else { 25123fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 25133fd0a558SYan, Zheng } 25145d4f98a2SYan Zheng out: 25150647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 2516023acb07SQu Wenruo btrfs_backref_cleanup_node(&rc->backref_cache, node); 25175d4f98a2SYan Zheng return ret; 25185d4f98a2SYan Zheng } 25195d4f98a2SYan Zheng 25205d4f98a2SYan Zheng /* 25215d4f98a2SYan Zheng * relocate a list of blocks 25225d4f98a2SYan Zheng */ 25235d4f98a2SYan Zheng static noinline_for_stack 25245d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 25255d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 25265d4f98a2SYan Zheng { 25272ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2528a26195a5SQu Wenruo struct btrfs_backref_node *node; 25295d4f98a2SYan Zheng struct btrfs_path *path; 25305d4f98a2SYan Zheng struct tree_block *block; 253198ff7b94SQu Wenruo struct tree_block *next; 25325d4f98a2SYan Zheng int ret; 25335d4f98a2SYan Zheng int err = 0; 25345d4f98a2SYan Zheng 25355d4f98a2SYan Zheng path = btrfs_alloc_path(); 2536e1a12670SLiu Bo if (!path) { 2537e1a12670SLiu Bo err = -ENOMEM; 253834c2b290SDavid Sterba goto out_free_blocks; 2539e1a12670SLiu Bo } 25405d4f98a2SYan Zheng 254198ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 254298ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 25435d4f98a2SYan Zheng if (!block->key_ready) 25442ff7e61eSJeff Mahoney readahead_tree_block(fs_info, block->bytenr); 25455d4f98a2SYan Zheng } 25465d4f98a2SYan Zheng 254798ff7b94SQu Wenruo /* Get first keys */ 254898ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 254934c2b290SDavid Sterba if (!block->key_ready) { 25502ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 255134c2b290SDavid Sterba if (err) 255234c2b290SDavid Sterba goto out_free_path; 255334c2b290SDavid Sterba } 25545d4f98a2SYan Zheng } 25555d4f98a2SYan Zheng 255698ff7b94SQu Wenruo /* Do tree relocation */ 255798ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 25583fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 25595d4f98a2SYan Zheng block->level, block->bytenr); 25605d4f98a2SYan Zheng if (IS_ERR(node)) { 25615d4f98a2SYan Zheng err = PTR_ERR(node); 25625d4f98a2SYan Zheng goto out; 25635d4f98a2SYan Zheng } 25645d4f98a2SYan Zheng 25655d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 25665d4f98a2SYan Zheng path); 25675d4f98a2SYan Zheng if (ret < 0) { 25685d4f98a2SYan Zheng err = ret; 256950dbbb71SJosef Bacik break; 25705d4f98a2SYan Zheng } 25715d4f98a2SYan Zheng } 25725d4f98a2SYan Zheng out: 25733fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 25745d4f98a2SYan Zheng 257534c2b290SDavid Sterba out_free_path: 25765d4f98a2SYan Zheng btrfs_free_path(path); 257734c2b290SDavid Sterba out_free_blocks: 2578e1a12670SLiu Bo free_block_list(blocks); 25795d4f98a2SYan Zheng return err; 25805d4f98a2SYan Zheng } 25815d4f98a2SYan Zheng 2582056d9becSNikolay Borisov static noinline_for_stack int prealloc_file_extent_cluster( 2583056d9becSNikolay Borisov struct btrfs_inode *inode, 2584efa56464SYan, Zheng struct file_extent_cluster *cluster) 2585efa56464SYan, Zheng { 2586efa56464SYan, Zheng u64 alloc_hint = 0; 2587efa56464SYan, Zheng u64 start; 2588efa56464SYan, Zheng u64 end; 2589056d9becSNikolay Borisov u64 offset = inode->index_cnt; 2590efa56464SYan, Zheng u64 num_bytes; 25914e9d0d01SNikolay Borisov int nr; 2592efa56464SYan, Zheng int ret = 0; 2593dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 2594dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 2595214e61d0SNikolay Borisov u64 cur_offset = prealloc_start; 2596efa56464SYan, Zheng 2597efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2598056d9becSNikolay Borisov ret = btrfs_alloc_data_chunk_ondemand(inode, 2599dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 2600efa56464SYan, Zheng if (ret) 2601214e61d0SNikolay Borisov return ret; 2602efa56464SYan, Zheng 2603056d9becSNikolay Borisov inode_lock(&inode->vfs_inode); 26044e9d0d01SNikolay Borisov for (nr = 0; nr < cluster->nr; nr++) { 2605efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2606efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2607efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2608efa56464SYan, Zheng else 2609efa56464SYan, Zheng end = cluster->end - offset; 2610efa56464SYan, Zheng 2611056d9becSNikolay Borisov lock_extent(&inode->io_tree, start, end); 2612efa56464SYan, Zheng num_bytes = end + 1 - start; 2613056d9becSNikolay Borisov ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start, 2614efa56464SYan, Zheng num_bytes, num_bytes, 2615efa56464SYan, Zheng end + 1, &alloc_hint); 261618513091SWang Xiaoguang cur_offset = end + 1; 2617056d9becSNikolay Borisov unlock_extent(&inode->io_tree, start, end); 2618efa56464SYan, Zheng if (ret) 2619efa56464SYan, Zheng break; 2620efa56464SYan, Zheng } 2621056d9becSNikolay Borisov inode_unlock(&inode->vfs_inode); 2622214e61d0SNikolay Borisov 262318513091SWang Xiaoguang if (cur_offset < prealloc_end) 2624056d9becSNikolay Borisov btrfs_free_reserved_data_space_noquota(inode->root->fs_info, 2625a89ef455SFilipe Manana prealloc_end + 1 - cur_offset); 2626efa56464SYan, Zheng return ret; 2627efa56464SYan, Zheng } 2628efa56464SYan, Zheng 2629efa56464SYan, Zheng static noinline_for_stack 26300257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 26310257bb82SYan, Zheng u64 block_start) 26325d4f98a2SYan Zheng { 26335d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 26345d4f98a2SYan Zheng struct extent_map *em; 26350257bb82SYan, Zheng int ret = 0; 26365d4f98a2SYan Zheng 2637172ddd60SDavid Sterba em = alloc_extent_map(); 26380257bb82SYan, Zheng if (!em) 26390257bb82SYan, Zheng return -ENOMEM; 26400257bb82SYan, Zheng 26415d4f98a2SYan Zheng em->start = start; 26420257bb82SYan, Zheng em->len = end + 1 - start; 26430257bb82SYan, Zheng em->block_len = em->len; 26440257bb82SYan, Zheng em->block_start = block_start; 26455d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 26465d4f98a2SYan Zheng 2647d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 26485d4f98a2SYan Zheng while (1) { 2649890871beSChris Mason write_lock(&em_tree->lock); 265009a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 2651890871beSChris Mason write_unlock(&em_tree->lock); 26525d4f98a2SYan Zheng if (ret != -EEXIST) { 26535d4f98a2SYan Zheng free_extent_map(em); 26545d4f98a2SYan Zheng break; 26555d4f98a2SYan Zheng } 2656dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 26575d4f98a2SYan Zheng } 2658d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 26590257bb82SYan, Zheng return ret; 26600257bb82SYan, Zheng } 26615d4f98a2SYan Zheng 2662726a3421SQu Wenruo /* 2663726a3421SQu Wenruo * Allow error injection to test balance cancellation 2664726a3421SQu Wenruo */ 2665726a3421SQu Wenruo int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 2666726a3421SQu Wenruo { 26675cb502f4SQu Wenruo return atomic_read(&fs_info->balance_cancel_req) || 26685cb502f4SQu Wenruo fatal_signal_pending(current); 2669726a3421SQu Wenruo } 2670726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 2671726a3421SQu Wenruo 26720257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 26730257bb82SYan, Zheng struct file_extent_cluster *cluster) 26740257bb82SYan, Zheng { 26752ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 26760257bb82SYan, Zheng u64 page_start; 26770257bb82SYan, Zheng u64 page_end; 26780257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 26790257bb82SYan, Zheng unsigned long index; 26800257bb82SYan, Zheng unsigned long last_index; 26810257bb82SYan, Zheng struct page *page; 26820257bb82SYan, Zheng struct file_ra_state *ra; 26833b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 26840257bb82SYan, Zheng int nr = 0; 26850257bb82SYan, Zheng int ret = 0; 26860257bb82SYan, Zheng 26870257bb82SYan, Zheng if (!cluster->nr) 26880257bb82SYan, Zheng return 0; 26890257bb82SYan, Zheng 26900257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 26910257bb82SYan, Zheng if (!ra) 26920257bb82SYan, Zheng return -ENOMEM; 26930257bb82SYan, Zheng 2694056d9becSNikolay Borisov ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster); 26950257bb82SYan, Zheng if (ret) 2696efa56464SYan, Zheng goto out; 26970257bb82SYan, Zheng 26980257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 26990257bb82SYan, Zheng 2700efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2701efa56464SYan, Zheng cluster->end - offset, cluster->start); 2702efa56464SYan, Zheng if (ret) 2703efa56464SYan, Zheng goto out; 2704efa56464SYan, Zheng 270509cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 270609cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 27070257bb82SYan, Zheng while (index <= last_index) { 27089f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 27099f3db423SNikolay Borisov PAGE_SIZE); 2710efa56464SYan, Zheng if (ret) 2711efa56464SYan, Zheng goto out; 2712efa56464SYan, Zheng 27130257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 27140257bb82SYan, Zheng if (!page) { 27150257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 27160257bb82SYan, Zheng ra, NULL, index, 27170257bb82SYan, Zheng last_index + 1 - index); 2718a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 27193b16a4e3SJosef Bacik mask); 27200257bb82SYan, Zheng if (!page) { 2721691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 272243b18595SQu Wenruo PAGE_SIZE, true); 272344db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 27248702ba93SQu Wenruo PAGE_SIZE); 27250257bb82SYan, Zheng ret = -ENOMEM; 2726efa56464SYan, Zheng goto out; 27270257bb82SYan, Zheng } 27280257bb82SYan, Zheng } 27290257bb82SYan, Zheng 27300257bb82SYan, Zheng if (PageReadahead(page)) { 27310257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 27320257bb82SYan, Zheng ra, NULL, page, index, 27330257bb82SYan, Zheng last_index + 1 - index); 27340257bb82SYan, Zheng } 27350257bb82SYan, Zheng 27360257bb82SYan, Zheng if (!PageUptodate(page)) { 27370257bb82SYan, Zheng btrfs_readpage(NULL, page); 27380257bb82SYan, Zheng lock_page(page); 27390257bb82SYan, Zheng if (!PageUptodate(page)) { 27400257bb82SYan, Zheng unlock_page(page); 274109cbfeafSKirill A. Shutemov put_page(page); 2742691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 274343b18595SQu Wenruo PAGE_SIZE, true); 27448b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 27458702ba93SQu Wenruo PAGE_SIZE); 27460257bb82SYan, Zheng ret = -EIO; 2747efa56464SYan, Zheng goto out; 27480257bb82SYan, Zheng } 27490257bb82SYan, Zheng } 27500257bb82SYan, Zheng 27514eee4fa4SMiao Xie page_start = page_offset(page); 275209cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 27530257bb82SYan, Zheng 2754d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 27550257bb82SYan, Zheng 27560257bb82SYan, Zheng set_page_extent_mapped(page); 27570257bb82SYan, Zheng 27580257bb82SYan, Zheng if (nr < cluster->nr && 27590257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 27600257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 27610257bb82SYan, Zheng page_start, page_end, 2762ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 27630257bb82SYan, Zheng nr++; 27640257bb82SYan, Zheng } 27650257bb82SYan, Zheng 2766c2566f22SNikolay Borisov ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, 2767c2566f22SNikolay Borisov page_end, 0, NULL); 2768765f3cebSNikolay Borisov if (ret) { 2769765f3cebSNikolay Borisov unlock_page(page); 2770765f3cebSNikolay Borisov put_page(page); 2771765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 277243b18595SQu Wenruo PAGE_SIZE, true); 2773765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 27748702ba93SQu Wenruo PAGE_SIZE); 2775765f3cebSNikolay Borisov 2776765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 2777765f3cebSNikolay Borisov page_start, page_end, 2778765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 2779765f3cebSNikolay Borisov goto out; 2780765f3cebSNikolay Borisov 2781765f3cebSNikolay Borisov } 27820257bb82SYan, Zheng set_page_dirty(page); 27830257bb82SYan, Zheng 27840257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 2785d0082371SJeff Mahoney page_start, page_end); 27860257bb82SYan, Zheng unlock_page(page); 278709cbfeafSKirill A. Shutemov put_page(page); 27880257bb82SYan, Zheng 27890257bb82SYan, Zheng index++; 27908702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 2791efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 27922ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 27937f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 27947f913c7cSQu Wenruo ret = -ECANCELED; 27957f913c7cSQu Wenruo goto out; 27967f913c7cSQu Wenruo } 27970257bb82SYan, Zheng } 27980257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 2799efa56464SYan, Zheng out: 28000257bb82SYan, Zheng kfree(ra); 28010257bb82SYan, Zheng return ret; 28020257bb82SYan, Zheng } 28030257bb82SYan, Zheng 28040257bb82SYan, Zheng static noinline_for_stack 28050257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 28060257bb82SYan, Zheng struct file_extent_cluster *cluster) 28070257bb82SYan, Zheng { 28080257bb82SYan, Zheng int ret; 28090257bb82SYan, Zheng 28100257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 28110257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 28120257bb82SYan, Zheng if (ret) 28130257bb82SYan, Zheng return ret; 28140257bb82SYan, Zheng cluster->nr = 0; 28150257bb82SYan, Zheng } 28160257bb82SYan, Zheng 28170257bb82SYan, Zheng if (!cluster->nr) 28180257bb82SYan, Zheng cluster->start = extent_key->objectid; 28190257bb82SYan, Zheng else 28200257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 28210257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 28220257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 28230257bb82SYan, Zheng cluster->nr++; 28240257bb82SYan, Zheng 28250257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 28260257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 28270257bb82SYan, Zheng if (ret) 28280257bb82SYan, Zheng return ret; 28290257bb82SYan, Zheng cluster->nr = 0; 28300257bb82SYan, Zheng } 28310257bb82SYan, Zheng return 0; 28325d4f98a2SYan Zheng } 28335d4f98a2SYan Zheng 28345d4f98a2SYan Zheng /* 28355d4f98a2SYan Zheng * helper to add a tree block to the list. 28365d4f98a2SYan Zheng * the major work is getting the generation and level of the block 28375d4f98a2SYan Zheng */ 28385d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 28395d4f98a2SYan Zheng struct btrfs_key *extent_key, 28405d4f98a2SYan Zheng struct btrfs_path *path, 28415d4f98a2SYan Zheng struct rb_root *blocks) 28425d4f98a2SYan Zheng { 28435d4f98a2SYan Zheng struct extent_buffer *eb; 28445d4f98a2SYan Zheng struct btrfs_extent_item *ei; 28455d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 28465d4f98a2SYan Zheng struct tree_block *block; 28475d4f98a2SYan Zheng struct rb_node *rb_node; 28485d4f98a2SYan Zheng u32 item_size; 28495d4f98a2SYan Zheng int level = -1; 28507fdf4b60SWang Shilong u64 generation; 28515d4f98a2SYan Zheng 28525d4f98a2SYan Zheng eb = path->nodes[0]; 28535d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 28545d4f98a2SYan Zheng 28553173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 28563173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 28575d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 28585d4f98a2SYan Zheng struct btrfs_extent_item); 28593173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 28605d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 28615d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 28625d4f98a2SYan Zheng } else { 28633173a18fSJosef Bacik level = (int)extent_key->offset; 28643173a18fSJosef Bacik } 28653173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 28666d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 2867ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 2868ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 2869ba3c2b19SNikolay Borisov return -EINVAL; 28703173a18fSJosef Bacik } else { 28715d4f98a2SYan Zheng BUG(); 28725d4f98a2SYan Zheng } 28735d4f98a2SYan Zheng 2874b3b4aa74SDavid Sterba btrfs_release_path(path); 28755d4f98a2SYan Zheng 28765d4f98a2SYan Zheng BUG_ON(level == -1); 28775d4f98a2SYan Zheng 28785d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 28795d4f98a2SYan Zheng if (!block) 28805d4f98a2SYan Zheng return -ENOMEM; 28815d4f98a2SYan Zheng 28825d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 2883da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 28845d4f98a2SYan Zheng block->key.offset = generation; 28855d4f98a2SYan Zheng block->level = level; 28865d4f98a2SYan Zheng block->key_ready = 0; 28875d4f98a2SYan Zheng 2888e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 288943c04fb1SJeff Mahoney if (rb_node) 2890982c92cbSQu Wenruo btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr, 2891982c92cbSQu Wenruo -EEXIST); 28925d4f98a2SYan Zheng 28935d4f98a2SYan Zheng return 0; 28945d4f98a2SYan Zheng } 28955d4f98a2SYan Zheng 28965d4f98a2SYan Zheng /* 28975d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 28985d4f98a2SYan Zheng */ 28995d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 29005d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 29015d4f98a2SYan Zheng struct rb_root *blocks) 29025d4f98a2SYan Zheng { 29030b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 29045d4f98a2SYan Zheng struct btrfs_path *path; 29055d4f98a2SYan Zheng struct btrfs_key key; 29065d4f98a2SYan Zheng int ret; 29070b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 29085d4f98a2SYan Zheng 29097476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 29105d4f98a2SYan Zheng return 0; 29115d4f98a2SYan Zheng 2912e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 29135d4f98a2SYan Zheng return 0; 29145d4f98a2SYan Zheng 29155d4f98a2SYan Zheng path = btrfs_alloc_path(); 29165d4f98a2SYan Zheng if (!path) 29175d4f98a2SYan Zheng return -ENOMEM; 2918aee68ee5SJosef Bacik again: 29195d4f98a2SYan Zheng key.objectid = bytenr; 2920aee68ee5SJosef Bacik if (skinny) { 2921aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 2922aee68ee5SJosef Bacik key.offset = (u64)-1; 2923aee68ee5SJosef Bacik } else { 29245d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 29255d4f98a2SYan Zheng key.offset = blocksize; 2926aee68ee5SJosef Bacik } 29275d4f98a2SYan Zheng 29285d4f98a2SYan Zheng path->search_commit_root = 1; 29295d4f98a2SYan Zheng path->skip_locking = 1; 29305d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 29315d4f98a2SYan Zheng if (ret < 0) 29325d4f98a2SYan Zheng goto out; 29335d4f98a2SYan Zheng 2934aee68ee5SJosef Bacik if (ret > 0 && skinny) { 2935aee68ee5SJosef Bacik if (path->slots[0]) { 2936aee68ee5SJosef Bacik path->slots[0]--; 2937aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 2938aee68ee5SJosef Bacik path->slots[0]); 29393173a18fSJosef Bacik if (key.objectid == bytenr && 2940aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 2941aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 2942aee68ee5SJosef Bacik key.offset == blocksize))) 29433173a18fSJosef Bacik ret = 0; 29443173a18fSJosef Bacik } 2945aee68ee5SJosef Bacik 2946aee68ee5SJosef Bacik if (ret) { 2947aee68ee5SJosef Bacik skinny = false; 2948aee68ee5SJosef Bacik btrfs_release_path(path); 2949aee68ee5SJosef Bacik goto again; 2950aee68ee5SJosef Bacik } 2951aee68ee5SJosef Bacik } 2952cdccee99SLiu Bo if (ret) { 2953cdccee99SLiu Bo ASSERT(ret == 1); 2954cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 2955cdccee99SLiu Bo btrfs_err(fs_info, 2956cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 2957cdccee99SLiu Bo bytenr); 2958cdccee99SLiu Bo WARN_ON(1); 2959cdccee99SLiu Bo ret = -EINVAL; 2960cdccee99SLiu Bo goto out; 2961cdccee99SLiu Bo } 29623173a18fSJosef Bacik 29635d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 29645d4f98a2SYan Zheng out: 29655d4f98a2SYan Zheng btrfs_free_path(path); 29665d4f98a2SYan Zheng return ret; 29675d4f98a2SYan Zheng } 29685d4f98a2SYan Zheng 29690af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 297032da5386SDavid Sterba struct btrfs_block_group *block_group, 29711bbc621eSChris Mason struct inode *inode, 29721bbc621eSChris Mason u64 ino) 29730af3d00bSJosef Bacik { 29740af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 29750af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 29760af3d00bSJosef Bacik int ret = 0; 29770af3d00bSJosef Bacik 29780af3d00bSJosef Bacik if (inode) 29790af3d00bSJosef Bacik goto truncate; 29800af3d00bSJosef Bacik 29810202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, ino, root); 29822e19f1f9SAl Viro if (IS_ERR(inode)) 29830af3d00bSJosef Bacik return -ENOENT; 29840af3d00bSJosef Bacik 29850af3d00bSJosef Bacik truncate: 29862ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 29877b61cd92SMiao Xie &fs_info->global_block_rsv); 29887b61cd92SMiao Xie if (ret) 29897b61cd92SMiao Xie goto out; 29907b61cd92SMiao Xie 29917a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 29920af3d00bSJosef Bacik if (IS_ERR(trans)) { 29933612b495STsutomu Itoh ret = PTR_ERR(trans); 29940af3d00bSJosef Bacik goto out; 29950af3d00bSJosef Bacik } 29960af3d00bSJosef Bacik 299777ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 29980af3d00bSJosef Bacik 29993a45bb20SJeff Mahoney btrfs_end_transaction(trans); 30002ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 30010af3d00bSJosef Bacik out: 30020af3d00bSJosef Bacik iput(inode); 30030af3d00bSJosef Bacik return ret; 30040af3d00bSJosef Bacik } 30050af3d00bSJosef Bacik 30065d4f98a2SYan Zheng /* 300719b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 300819b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 30095d4f98a2SYan Zheng */ 301019b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 301119b546d7SQu Wenruo struct btrfs_block_group *block_group, 301219b546d7SQu Wenruo u64 data_bytenr) 30135d4f98a2SYan Zheng { 301419b546d7SQu Wenruo u64 space_cache_ino; 301519b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 30165d4f98a2SYan Zheng struct btrfs_key key; 301719b546d7SQu Wenruo bool found = false; 301819b546d7SQu Wenruo int i; 30195d4f98a2SYan Zheng int ret; 30205d4f98a2SYan Zheng 302119b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 302219b546d7SQu Wenruo return 0; 30235d4f98a2SYan Zheng 302419b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 302519b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 302619b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 302719b546d7SQu Wenruo continue; 302819b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 302919b546d7SQu Wenruo if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG && 303019b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 303119b546d7SQu Wenruo found = true; 303219b546d7SQu Wenruo space_cache_ino = key.objectid; 303319b546d7SQu Wenruo break; 303419b546d7SQu Wenruo } 303519b546d7SQu Wenruo } 303619b546d7SQu Wenruo if (!found) 303719b546d7SQu Wenruo return -ENOENT; 303819b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 303919b546d7SQu Wenruo space_cache_ino); 30400af3d00bSJosef Bacik return ret; 30415d4f98a2SYan Zheng } 30425d4f98a2SYan Zheng 30435d4f98a2SYan Zheng /* 30442c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 30455d4f98a2SYan Zheng */ 30465d4f98a2SYan Zheng static noinline_for_stack 30475d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 30485d4f98a2SYan Zheng struct btrfs_key *extent_key, 30495d4f98a2SYan Zheng struct btrfs_path *path, 30505d4f98a2SYan Zheng struct rb_root *blocks) 30515d4f98a2SYan Zheng { 305219b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 305319b546d7SQu Wenruo struct ulist *leaves = NULL; 305419b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 305519b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 305619b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3057647f63bdSFilipe David Borba Manana int ret = 0; 30585d4f98a2SYan Zheng 3059b3b4aa74SDavid Sterba btrfs_release_path(path); 306019b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 306119b546d7SQu Wenruo 0, &leaves, NULL, true); 306219b546d7SQu Wenruo if (ret < 0) 306319b546d7SQu Wenruo return ret; 306419b546d7SQu Wenruo 306519b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 306619b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 306719b546d7SQu Wenruo struct extent_buffer *eb; 306819b546d7SQu Wenruo 306919b546d7SQu Wenruo eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL); 307019b546d7SQu Wenruo if (IS_ERR(eb)) { 307119b546d7SQu Wenruo ret = PTR_ERR(eb); 307219b546d7SQu Wenruo break; 307319b546d7SQu Wenruo } 307419b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 307519b546d7SQu Wenruo extent_key->objectid); 307619b546d7SQu Wenruo free_extent_buffer(eb); 307719b546d7SQu Wenruo if (ret < 0) 307819b546d7SQu Wenruo break; 307919b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 308019b546d7SQu Wenruo if (ret < 0) 308119b546d7SQu Wenruo break; 308219b546d7SQu Wenruo } 308319b546d7SQu Wenruo if (ret < 0) 30845d4f98a2SYan Zheng free_block_list(blocks); 308519b546d7SQu Wenruo ulist_free(leaves); 308619b546d7SQu Wenruo return ret; 30875d4f98a2SYan Zheng } 30885d4f98a2SYan Zheng 30895d4f98a2SYan Zheng /* 30902c016dc2SLiu Bo * helper to find next unprocessed extent 30915d4f98a2SYan Zheng */ 30925d4f98a2SYan Zheng static noinline_for_stack 3093147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 30943fd0a558SYan, Zheng struct btrfs_key *extent_key) 30955d4f98a2SYan Zheng { 30960b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 30975d4f98a2SYan Zheng struct btrfs_key key; 30985d4f98a2SYan Zheng struct extent_buffer *leaf; 30995d4f98a2SYan Zheng u64 start, end, last; 31005d4f98a2SYan Zheng int ret; 31015d4f98a2SYan Zheng 3102b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 31035d4f98a2SYan Zheng while (1) { 31045d4f98a2SYan Zheng cond_resched(); 31055d4f98a2SYan Zheng if (rc->search_start >= last) { 31065d4f98a2SYan Zheng ret = 1; 31075d4f98a2SYan Zheng break; 31085d4f98a2SYan Zheng } 31095d4f98a2SYan Zheng 31105d4f98a2SYan Zheng key.objectid = rc->search_start; 31115d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31125d4f98a2SYan Zheng key.offset = 0; 31135d4f98a2SYan Zheng 31145d4f98a2SYan Zheng path->search_commit_root = 1; 31155d4f98a2SYan Zheng path->skip_locking = 1; 31165d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 31175d4f98a2SYan Zheng 0, 0); 31185d4f98a2SYan Zheng if (ret < 0) 31195d4f98a2SYan Zheng break; 31205d4f98a2SYan Zheng next: 31215d4f98a2SYan Zheng leaf = path->nodes[0]; 31225d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 31235d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 31245d4f98a2SYan Zheng if (ret != 0) 31255d4f98a2SYan Zheng break; 31265d4f98a2SYan Zheng leaf = path->nodes[0]; 31275d4f98a2SYan Zheng } 31285d4f98a2SYan Zheng 31295d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 31305d4f98a2SYan Zheng if (key.objectid >= last) { 31315d4f98a2SYan Zheng ret = 1; 31325d4f98a2SYan Zheng break; 31335d4f98a2SYan Zheng } 31345d4f98a2SYan Zheng 31353173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 31363173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 31373173a18fSJosef Bacik path->slots[0]++; 31383173a18fSJosef Bacik goto next; 31393173a18fSJosef Bacik } 31403173a18fSJosef Bacik 31413173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 31425d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 31435d4f98a2SYan Zheng path->slots[0]++; 31445d4f98a2SYan Zheng goto next; 31455d4f98a2SYan Zheng } 31465d4f98a2SYan Zheng 31473173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 31480b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 31493173a18fSJosef Bacik rc->search_start) { 31503173a18fSJosef Bacik path->slots[0]++; 31513173a18fSJosef Bacik goto next; 31523173a18fSJosef Bacik } 31533173a18fSJosef Bacik 31545d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 31555d4f98a2SYan Zheng key.objectid, &start, &end, 3156e6138876SJosef Bacik EXTENT_DIRTY, NULL); 31575d4f98a2SYan Zheng 31585d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3159b3b4aa74SDavid Sterba btrfs_release_path(path); 31605d4f98a2SYan Zheng rc->search_start = end + 1; 31615d4f98a2SYan Zheng } else { 31623173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 31635d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 31643173a18fSJosef Bacik else 31653173a18fSJosef Bacik rc->search_start = key.objectid + 31660b246afaSJeff Mahoney fs_info->nodesize; 31673fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 31685d4f98a2SYan Zheng return 0; 31695d4f98a2SYan Zheng } 31705d4f98a2SYan Zheng } 3171b3b4aa74SDavid Sterba btrfs_release_path(path); 31725d4f98a2SYan Zheng return ret; 31735d4f98a2SYan Zheng } 31745d4f98a2SYan Zheng 31755d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 31765d4f98a2SYan Zheng { 31775d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31787585717fSChris Mason 31797585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 31805d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 31817585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 31825d4f98a2SYan Zheng } 31835d4f98a2SYan Zheng 31845d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 31855d4f98a2SYan Zheng { 31865d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31877585717fSChris Mason 31887585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 31895d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 31907585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 31915d4f98a2SYan Zheng } 31925d4f98a2SYan Zheng 31935d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 31945d4f98a2SYan Zheng { 31955d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 31965d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 31975d4f98a2SYan Zheng return 1; 31985d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 31995d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 32005d4f98a2SYan Zheng return 1; 32015d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 32025d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 32035d4f98a2SYan Zheng return 1; 32045d4f98a2SYan Zheng return 0; 32055d4f98a2SYan Zheng } 32065d4f98a2SYan Zheng 32073fd0a558SYan, Zheng static noinline_for_stack 32083fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 32093fd0a558SYan, Zheng { 32103fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3211ac2fabacSJosef Bacik int ret; 32123fd0a558SYan, Zheng 32132ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 321466d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 32153fd0a558SYan, Zheng if (!rc->block_rsv) 32163fd0a558SYan, Zheng return -ENOMEM; 32173fd0a558SYan, Zheng 32183fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3219b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 32203fd0a558SYan, Zheng rc->extents_found = 0; 32213fd0a558SYan, Zheng rc->nodes_relocated = 0; 32223fd0a558SYan, Zheng rc->merging_rsv_size = 0; 32230647bf56SWang Shilong rc->reserved_bytes = 0; 3224da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 32250647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3226ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3227ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3228ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3229ac2fabacSJosef Bacik if (ret) 3230ac2fabacSJosef Bacik return ret; 32313fd0a558SYan, Zheng 32323fd0a558SYan, Zheng rc->create_reloc_tree = 1; 32333fd0a558SYan, Zheng set_reloc_control(rc); 32343fd0a558SYan, Zheng 32357a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 323628818947SLiu Bo if (IS_ERR(trans)) { 323728818947SLiu Bo unset_reloc_control(rc); 323828818947SLiu Bo /* 323928818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 324028818947SLiu Bo * cleanup. And callers are responsible to free the above 324128818947SLiu Bo * block rsv. 324228818947SLiu Bo */ 324328818947SLiu Bo return PTR_ERR(trans); 324428818947SLiu Bo } 32453a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 32463fd0a558SYan, Zheng return 0; 32473fd0a558SYan, Zheng } 324876dda93cSYan, Zheng 32495d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 32505d4f98a2SYan Zheng { 32512ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 32525d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 32535d4f98a2SYan Zheng struct btrfs_key key; 32545d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 32555d4f98a2SYan Zheng struct btrfs_path *path; 32565d4f98a2SYan Zheng struct btrfs_extent_item *ei; 32575d4f98a2SYan Zheng u64 flags; 32585d4f98a2SYan Zheng u32 item_size; 32595d4f98a2SYan Zheng int ret; 32605d4f98a2SYan Zheng int err = 0; 3261c87f08caSChris Mason int progress = 0; 32625d4f98a2SYan Zheng 32635d4f98a2SYan Zheng path = btrfs_alloc_path(); 32643fd0a558SYan, Zheng if (!path) 32655d4f98a2SYan Zheng return -ENOMEM; 3266e4058b54SDavid Sterba path->reada = READA_FORWARD; 32673fd0a558SYan, Zheng 32683fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 32693fd0a558SYan, Zheng if (ret) { 32703fd0a558SYan, Zheng err = ret; 32713fd0a558SYan, Zheng goto out_free; 32722423fdfbSJiri Slaby } 32735d4f98a2SYan Zheng 32745d4f98a2SYan Zheng while (1) { 32750647bf56SWang Shilong rc->reserved_bytes = 0; 32760647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 32770647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 32780647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 32790647bf56SWang Shilong if (ret) { 32800647bf56SWang Shilong err = ret; 32810647bf56SWang Shilong break; 32820647bf56SWang Shilong } 3283c87f08caSChris Mason progress++; 3284a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 32850f788c58SLiu Bo if (IS_ERR(trans)) { 32860f788c58SLiu Bo err = PTR_ERR(trans); 32870f788c58SLiu Bo trans = NULL; 32880f788c58SLiu Bo break; 32890f788c58SLiu Bo } 3290c87f08caSChris Mason restart: 32913fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 32923a45bb20SJeff Mahoney btrfs_end_transaction(trans); 329342a657f5SPan Bian trans = NULL; 32943fd0a558SYan, Zheng continue; 32953fd0a558SYan, Zheng } 32963fd0a558SYan, Zheng 3297147d256eSZhaolei ret = find_next_extent(rc, path, &key); 32985d4f98a2SYan Zheng if (ret < 0) 32995d4f98a2SYan Zheng err = ret; 33005d4f98a2SYan Zheng if (ret != 0) 33015d4f98a2SYan Zheng break; 33025d4f98a2SYan Zheng 33035d4f98a2SYan Zheng rc->extents_found++; 33045d4f98a2SYan Zheng 33055d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 33065d4f98a2SYan Zheng struct btrfs_extent_item); 33073fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 33085d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 33095d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 33105d4f98a2SYan Zheng ret = check_extent_flags(flags); 33115d4f98a2SYan Zheng BUG_ON(ret); 33126d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3313ba3c2b19SNikolay Borisov err = -EINVAL; 3314ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 3315ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 3316ba3c2b19SNikolay Borisov break; 33175d4f98a2SYan Zheng } else { 33185d4f98a2SYan Zheng BUG(); 33195d4f98a2SYan Zheng } 33205d4f98a2SYan Zheng 33215d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 33225d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 33235d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 33245d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 33255d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 33265d4f98a2SYan Zheng } else { 3327b3b4aa74SDavid Sterba btrfs_release_path(path); 33285d4f98a2SYan Zheng ret = 0; 33295d4f98a2SYan Zheng } 33305d4f98a2SYan Zheng if (ret < 0) { 33313fd0a558SYan, Zheng err = ret; 33325d4f98a2SYan Zheng break; 33335d4f98a2SYan Zheng } 33345d4f98a2SYan Zheng 33355d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 33365d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 33375d4f98a2SYan Zheng if (ret < 0) { 33383fd0a558SYan, Zheng if (ret != -EAGAIN) { 33395d4f98a2SYan Zheng err = ret; 33405d4f98a2SYan Zheng break; 33415d4f98a2SYan Zheng } 33423fd0a558SYan, Zheng rc->extents_found--; 33433fd0a558SYan, Zheng rc->search_start = key.objectid; 33443fd0a558SYan, Zheng } 33455d4f98a2SYan Zheng } 33465d4f98a2SYan Zheng 33473a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 33482ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 33493fd0a558SYan, Zheng trans = NULL; 33505d4f98a2SYan Zheng 33515d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 33525d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 33535d4f98a2SYan Zheng rc->found_file_extent = 1; 33540257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 33553fd0a558SYan, Zheng &key, &rc->cluster); 33565d4f98a2SYan Zheng if (ret < 0) { 33575d4f98a2SYan Zheng err = ret; 33585d4f98a2SYan Zheng break; 33595d4f98a2SYan Zheng } 33605d4f98a2SYan Zheng } 3361f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 3362f31ea088SQu Wenruo err = -ECANCELED; 3363f31ea088SQu Wenruo break; 3364f31ea088SQu Wenruo } 33655d4f98a2SYan Zheng } 3366c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 336743a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 33689689457bSShilong Wang if (ret == 1) { 3369c87f08caSChris Mason err = 0; 3370c87f08caSChris Mason progress = 0; 3371c87f08caSChris Mason goto restart; 3372c87f08caSChris Mason } 3373c87f08caSChris Mason } 33743fd0a558SYan, Zheng 3375b3b4aa74SDavid Sterba btrfs_release_path(path); 337691166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 33775d4f98a2SYan Zheng 33785d4f98a2SYan Zheng if (trans) { 33793a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 33802ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 33815d4f98a2SYan Zheng } 33825d4f98a2SYan Zheng 33830257bb82SYan, Zheng if (!err) { 33843fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 33853fd0a558SYan, Zheng &rc->cluster); 33860257bb82SYan, Zheng if (ret < 0) 33870257bb82SYan, Zheng err = ret; 33880257bb82SYan, Zheng } 33890257bb82SYan, Zheng 33903fd0a558SYan, Zheng rc->create_reloc_tree = 0; 33913fd0a558SYan, Zheng set_reloc_control(rc); 33920257bb82SYan, Zheng 339313fe1bdbSQu Wenruo btrfs_backref_release_cache(&rc->backref_cache); 339463f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 33955d4f98a2SYan Zheng 33967f913c7cSQu Wenruo /* 33977f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 33987f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 33997f913c7cSQu Wenruo * 34007f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 34017f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 34027f913c7cSQu Wenruo * merge_reloc_roots() 34037f913c7cSQu Wenruo */ 34043fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 34055d4f98a2SYan Zheng 34065d4f98a2SYan Zheng merge_reloc_roots(rc); 34075d4f98a2SYan Zheng 34083fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 34095d4f98a2SYan Zheng unset_reloc_control(rc); 341063f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 34115d4f98a2SYan Zheng 34125d4f98a2SYan Zheng /* get rid of pinned extents */ 34137a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 341462b99540SQu Wenruo if (IS_ERR(trans)) { 34153612b495STsutomu Itoh err = PTR_ERR(trans); 341662b99540SQu Wenruo goto out_free; 341762b99540SQu Wenruo } 34183a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 34196217b0faSJosef Bacik out_free: 3420d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3421d2311e69SQu Wenruo if (ret < 0 && !err) 3422d2311e69SQu Wenruo err = ret; 34232ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 34243fd0a558SYan, Zheng btrfs_free_path(path); 34255d4f98a2SYan Zheng return err; 34265d4f98a2SYan Zheng } 34275d4f98a2SYan Zheng 34285d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 34290257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 34305d4f98a2SYan Zheng { 34315d4f98a2SYan Zheng struct btrfs_path *path; 34325d4f98a2SYan Zheng struct btrfs_inode_item *item; 34335d4f98a2SYan Zheng struct extent_buffer *leaf; 34345d4f98a2SYan Zheng int ret; 34355d4f98a2SYan Zheng 34365d4f98a2SYan Zheng path = btrfs_alloc_path(); 34375d4f98a2SYan Zheng if (!path) 34385d4f98a2SYan Zheng return -ENOMEM; 34395d4f98a2SYan Zheng 34405d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 34415d4f98a2SYan Zheng if (ret) 34425d4f98a2SYan Zheng goto out; 34435d4f98a2SYan Zheng 34445d4f98a2SYan Zheng leaf = path->nodes[0]; 34455d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 3446b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 34475d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 34480257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 34495d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 34503fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 34513fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 34525d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 34535d4f98a2SYan Zheng out: 34545d4f98a2SYan Zheng btrfs_free_path(path); 34555d4f98a2SYan Zheng return ret; 34565d4f98a2SYan Zheng } 34575d4f98a2SYan Zheng 34585d4f98a2SYan Zheng /* 34595d4f98a2SYan Zheng * helper to create inode for data relocation. 34605d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 34615d4f98a2SYan Zheng */ 34623fd0a558SYan, Zheng static noinline_for_stack 34633fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 346432da5386SDavid Sterba struct btrfs_block_group *group) 34655d4f98a2SYan Zheng { 34665d4f98a2SYan Zheng struct inode *inode = NULL; 34675d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 34685d4f98a2SYan Zheng struct btrfs_root *root; 34694624900dSZhaolei u64 objectid; 34705d4f98a2SYan Zheng int err = 0; 34715d4f98a2SYan Zheng 3472aeb935a4SQu Wenruo root = btrfs_grab_root(fs_info->data_reloc_root); 3473a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 347476deacf0SJosef Bacik if (IS_ERR(trans)) { 347500246528SJosef Bacik btrfs_put_root(root); 34763fd0a558SYan, Zheng return ERR_CAST(trans); 347776deacf0SJosef Bacik } 34785d4f98a2SYan Zheng 3479581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 34805d4f98a2SYan Zheng if (err) 34815d4f98a2SYan Zheng goto out; 34825d4f98a2SYan Zheng 34830257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 34845d4f98a2SYan Zheng BUG_ON(err); 34855d4f98a2SYan Zheng 34860202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, objectid, root); 34872e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 3488b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 34895d4f98a2SYan Zheng 349073f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 34915d4f98a2SYan Zheng out: 349200246528SJosef Bacik btrfs_put_root(root); 34933a45bb20SJeff Mahoney btrfs_end_transaction(trans); 34942ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 34955d4f98a2SYan Zheng if (err) { 34965d4f98a2SYan Zheng if (inode) 34975d4f98a2SYan Zheng iput(inode); 34985d4f98a2SYan Zheng inode = ERR_PTR(err); 34995d4f98a2SYan Zheng } 35005d4f98a2SYan Zheng return inode; 35015d4f98a2SYan Zheng } 35025d4f98a2SYan Zheng 3503c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 35043fd0a558SYan, Zheng { 35053fd0a558SYan, Zheng struct reloc_control *rc; 35063fd0a558SYan, Zheng 35073fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 35083fd0a558SYan, Zheng if (!rc) 35093fd0a558SYan, Zheng return NULL; 35103fd0a558SYan, Zheng 35113fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 3512d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 3513584fb121SQu Wenruo btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1); 35143fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 351543eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 351643eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 35173fd0a558SYan, Zheng return rc; 35183fd0a558SYan, Zheng } 35193fd0a558SYan, Zheng 35201a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 35211a0afa0eSJosef Bacik { 35221a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 35231a0afa0eSJosef Bacik 35241a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 35251a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 35261a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 35271a0afa0eSJosef Bacik kfree(node); 35281a0afa0eSJosef Bacik 35291a0afa0eSJosef Bacik kfree(rc); 35301a0afa0eSJosef Bacik } 35311a0afa0eSJosef Bacik 35325d4f98a2SYan Zheng /* 3533ebce0e01SAdam Borowski * Print the block group being relocated 3534ebce0e01SAdam Borowski */ 3535ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 353632da5386SDavid Sterba struct btrfs_block_group *block_group) 3537ebce0e01SAdam Borowski { 3538f89e09cfSAnand Jain char buf[128] = {'\0'}; 3539ebce0e01SAdam Borowski 3540f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 3541ebce0e01SAdam Borowski 3542ebce0e01SAdam Borowski btrfs_info(fs_info, 3543ebce0e01SAdam Borowski "relocating block group %llu flags %s", 3544b3470b5dSDavid Sterba block_group->start, buf); 3545ebce0e01SAdam Borowski } 3546ebce0e01SAdam Borowski 3547430640e3SQu Wenruo static const char *stage_to_string(int stage) 3548430640e3SQu Wenruo { 3549430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 3550430640e3SQu Wenruo return "move data extents"; 3551430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 3552430640e3SQu Wenruo return "update data pointers"; 3553430640e3SQu Wenruo return "unknown"; 3554430640e3SQu Wenruo } 3555430640e3SQu Wenruo 3556ebce0e01SAdam Borowski /* 35575d4f98a2SYan Zheng * function to relocate all extents in a block group. 35585d4f98a2SYan Zheng */ 35596bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 35605d4f98a2SYan Zheng { 356132da5386SDavid Sterba struct btrfs_block_group *bg; 35626bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 35635d4f98a2SYan Zheng struct reloc_control *rc; 35640af3d00bSJosef Bacik struct inode *inode; 35650af3d00bSJosef Bacik struct btrfs_path *path; 35665d4f98a2SYan Zheng int ret; 3567f0486c68SYan, Zheng int rw = 0; 35685d4f98a2SYan Zheng int err = 0; 35695d4f98a2SYan Zheng 3570eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 3571eede2bf3SOmar Sandoval if (!bg) 3572eede2bf3SOmar Sandoval return -ENOENT; 3573eede2bf3SOmar Sandoval 3574eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 3575eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 3576eede2bf3SOmar Sandoval return -ETXTBSY; 3577eede2bf3SOmar Sandoval } 3578eede2bf3SOmar Sandoval 3579c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 3580eede2bf3SOmar Sandoval if (!rc) { 3581eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 35825d4f98a2SYan Zheng return -ENOMEM; 3583eede2bf3SOmar Sandoval } 35845d4f98a2SYan Zheng 3585f0486c68SYan, Zheng rc->extent_root = extent_root; 3586eede2bf3SOmar Sandoval rc->block_group = bg; 35875d4f98a2SYan Zheng 3588b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 3589f0486c68SYan, Zheng if (ret) { 3590f0486c68SYan, Zheng err = ret; 3591f0486c68SYan, Zheng goto out; 3592f0486c68SYan, Zheng } 3593f0486c68SYan, Zheng rw = 1; 3594f0486c68SYan, Zheng 35950af3d00bSJosef Bacik path = btrfs_alloc_path(); 35960af3d00bSJosef Bacik if (!path) { 35970af3d00bSJosef Bacik err = -ENOMEM; 35980af3d00bSJosef Bacik goto out; 35990af3d00bSJosef Bacik } 36000af3d00bSJosef Bacik 36017949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 36020af3d00bSJosef Bacik btrfs_free_path(path); 36030af3d00bSJosef Bacik 36040af3d00bSJosef Bacik if (!IS_ERR(inode)) 36051bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 36060af3d00bSJosef Bacik else 36070af3d00bSJosef Bacik ret = PTR_ERR(inode); 36080af3d00bSJosef Bacik 36090af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 36100af3d00bSJosef Bacik err = ret; 36110af3d00bSJosef Bacik goto out; 36120af3d00bSJosef Bacik } 36130af3d00bSJosef Bacik 36145d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 36155d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 36165d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 36175d4f98a2SYan Zheng rc->data_inode = NULL; 36185d4f98a2SYan Zheng goto out; 36195d4f98a2SYan Zheng } 36205d4f98a2SYan Zheng 36210b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 36225d4f98a2SYan Zheng 36239cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 3624f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 36256374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 3626b3470b5dSDavid Sterba rc->block_group->start, 3627b3470b5dSDavid Sterba rc->block_group->length); 36285d4f98a2SYan Zheng 36295d4f98a2SYan Zheng while (1) { 3630430640e3SQu Wenruo int finishes_stage; 3631430640e3SQu Wenruo 363276dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 36335d4f98a2SYan Zheng ret = relocate_block_group(rc); 363476dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 3635ff612ba7SJosef Bacik if (ret < 0) 36365d4f98a2SYan Zheng err = ret; 3637ff612ba7SJosef Bacik 3638430640e3SQu Wenruo finishes_stage = rc->stage; 3639ff612ba7SJosef Bacik /* 3640ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 3641ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 3642ff612ba7SJosef Bacik * different block group we could end up hitting the 3643ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 3644ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 3645ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 3646ff612ba7SJosef Bacik * out of the loop if we hit an error. 3647ff612ba7SJosef Bacik */ 3648ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 3649ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 3650ff612ba7SJosef Bacik (u64)-1); 3651ff612ba7SJosef Bacik if (ret) 3652ff612ba7SJosef Bacik err = ret; 3653ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 3654ff612ba7SJosef Bacik 0, -1); 3655ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 36565d4f98a2SYan Zheng } 36575d4f98a2SYan Zheng 3658ff612ba7SJosef Bacik if (err < 0) 3659ff612ba7SJosef Bacik goto out; 3660ff612ba7SJosef Bacik 36615d4f98a2SYan Zheng if (rc->extents_found == 0) 36625d4f98a2SYan Zheng break; 36635d4f98a2SYan Zheng 3664430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 3665430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 36665d4f98a2SYan Zheng } 36675d4f98a2SYan Zheng 36685d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 36695d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 3670bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 36715d4f98a2SYan Zheng out: 3672f0486c68SYan, Zheng if (err && rw) 36732ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 36745d4f98a2SYan Zheng iput(rc->data_inode); 36755d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 36761a0afa0eSJosef Bacik free_reloc_control(rc); 36775d4f98a2SYan Zheng return err; 36785d4f98a2SYan Zheng } 36795d4f98a2SYan Zheng 368076dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 368176dda93cSYan, Zheng { 36820b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 368376dda93cSYan, Zheng struct btrfs_trans_handle *trans; 368479787eaaSJeff Mahoney int ret, err; 368576dda93cSYan, Zheng 36860b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 368779787eaaSJeff Mahoney if (IS_ERR(trans)) 368879787eaaSJeff Mahoney return PTR_ERR(trans); 368976dda93cSYan, Zheng 369076dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 369176dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 369276dda93cSYan, Zheng root->root_item.drop_level = 0; 369376dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 36940b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 369576dda93cSYan, Zheng &root->root_key, &root->root_item); 369676dda93cSYan, Zheng 36973a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 369879787eaaSJeff Mahoney if (err) 369979787eaaSJeff Mahoney return err; 370079787eaaSJeff Mahoney return ret; 370176dda93cSYan, Zheng } 370276dda93cSYan, Zheng 37035d4f98a2SYan Zheng /* 37045d4f98a2SYan Zheng * recover relocation interrupted by system crash. 37055d4f98a2SYan Zheng * 37065d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 37075d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 37085d4f98a2SYan Zheng */ 37095d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 37105d4f98a2SYan Zheng { 37110b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 37125d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 37135d4f98a2SYan Zheng struct btrfs_key key; 37145d4f98a2SYan Zheng struct btrfs_root *fs_root; 37155d4f98a2SYan Zheng struct btrfs_root *reloc_root; 37165d4f98a2SYan Zheng struct btrfs_path *path; 37175d4f98a2SYan Zheng struct extent_buffer *leaf; 37185d4f98a2SYan Zheng struct reloc_control *rc = NULL; 37195d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 37205d4f98a2SYan Zheng int ret; 37215d4f98a2SYan Zheng int err = 0; 37225d4f98a2SYan Zheng 37235d4f98a2SYan Zheng path = btrfs_alloc_path(); 37245d4f98a2SYan Zheng if (!path) 37255d4f98a2SYan Zheng return -ENOMEM; 3726e4058b54SDavid Sterba path->reada = READA_BACK; 37275d4f98a2SYan Zheng 37285d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 37295d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 37305d4f98a2SYan Zheng key.offset = (u64)-1; 37315d4f98a2SYan Zheng 37325d4f98a2SYan Zheng while (1) { 37330b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 37345d4f98a2SYan Zheng path, 0, 0); 37355d4f98a2SYan Zheng if (ret < 0) { 37365d4f98a2SYan Zheng err = ret; 37375d4f98a2SYan Zheng goto out; 37385d4f98a2SYan Zheng } 37395d4f98a2SYan Zheng if (ret > 0) { 37405d4f98a2SYan Zheng if (path->slots[0] == 0) 37415d4f98a2SYan Zheng break; 37425d4f98a2SYan Zheng path->slots[0]--; 37435d4f98a2SYan Zheng } 37445d4f98a2SYan Zheng leaf = path->nodes[0]; 37455d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 3746b3b4aa74SDavid Sterba btrfs_release_path(path); 37475d4f98a2SYan Zheng 37485d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 37495d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 37505d4f98a2SYan Zheng break; 37515d4f98a2SYan Zheng 37523dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 37535d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 37545d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 37555d4f98a2SYan Zheng goto out; 37565d4f98a2SYan Zheng } 37575d4f98a2SYan Zheng 375892a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 37595d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 37605d4f98a2SYan Zheng 37615d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 3762a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, 3763a820feb5SDavid Sterba reloc_root->root_key.offset, false); 37645d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 376576dda93cSYan, Zheng ret = PTR_ERR(fs_root); 376676dda93cSYan, Zheng if (ret != -ENOENT) { 376776dda93cSYan, Zheng err = ret; 37685d4f98a2SYan Zheng goto out; 37695d4f98a2SYan Zheng } 377079787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 377179787eaaSJeff Mahoney if (ret < 0) { 377279787eaaSJeff Mahoney err = ret; 377379787eaaSJeff Mahoney goto out; 377479787eaaSJeff Mahoney } 3775932fd26dSJosef Bacik } else { 377600246528SJosef Bacik btrfs_put_root(fs_root); 377776dda93cSYan, Zheng } 37785d4f98a2SYan Zheng } 37795d4f98a2SYan Zheng 37805d4f98a2SYan Zheng if (key.offset == 0) 37815d4f98a2SYan Zheng break; 37825d4f98a2SYan Zheng 37835d4f98a2SYan Zheng key.offset--; 37845d4f98a2SYan Zheng } 3785b3b4aa74SDavid Sterba btrfs_release_path(path); 37865d4f98a2SYan Zheng 37875d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 37885d4f98a2SYan Zheng goto out; 37895d4f98a2SYan Zheng 3790c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 37915d4f98a2SYan Zheng if (!rc) { 37925d4f98a2SYan Zheng err = -ENOMEM; 37935d4f98a2SYan Zheng goto out; 37945d4f98a2SYan Zheng } 37955d4f98a2SYan Zheng 37960b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 37975d4f98a2SYan Zheng 37985d4f98a2SYan Zheng set_reloc_control(rc); 37995d4f98a2SYan Zheng 38007a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 38013612b495STsutomu Itoh if (IS_ERR(trans)) { 38023612b495STsutomu Itoh err = PTR_ERR(trans); 3803fb2d83eeSJosef Bacik goto out_unset; 38043612b495STsutomu Itoh } 38053fd0a558SYan, Zheng 38063fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 38073fd0a558SYan, Zheng 38085d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 38095d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 38105d4f98a2SYan Zheng struct btrfs_root, root_list); 38115d4f98a2SYan Zheng list_del(&reloc_root->root_list); 38125d4f98a2SYan Zheng 38135d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 38145d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 38155d4f98a2SYan Zheng &rc->reloc_roots); 38165d4f98a2SYan Zheng continue; 38175d4f98a2SYan Zheng } 38185d4f98a2SYan Zheng 3819a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 3820a820feb5SDavid Sterba false); 382179787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 382279787eaaSJeff Mahoney err = PTR_ERR(fs_root); 3823ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 38241402d17dSXiyu Yang btrfs_end_transaction(trans); 3825fb2d83eeSJosef Bacik goto out_unset; 382679787eaaSJeff Mahoney } 38275d4f98a2SYan Zheng 3828ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 382979787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 3830f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 383100246528SJosef Bacik btrfs_put_root(fs_root); 38325d4f98a2SYan Zheng } 38335d4f98a2SYan Zheng 38343a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 383579787eaaSJeff Mahoney if (err) 3836fb2d83eeSJosef Bacik goto out_unset; 38375d4f98a2SYan Zheng 38385d4f98a2SYan Zheng merge_reloc_roots(rc); 38395d4f98a2SYan Zheng 38405d4f98a2SYan Zheng unset_reloc_control(rc); 38415d4f98a2SYan Zheng 38427a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 384362b99540SQu Wenruo if (IS_ERR(trans)) { 38443612b495STsutomu Itoh err = PTR_ERR(trans); 38456217b0faSJosef Bacik goto out_clean; 384662b99540SQu Wenruo } 38473a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 38486217b0faSJosef Bacik out_clean: 3849d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3850d2311e69SQu Wenruo if (ret < 0 && !err) 3851d2311e69SQu Wenruo err = ret; 3852fb2d83eeSJosef Bacik out_unset: 3853fb2d83eeSJosef Bacik unset_reloc_control(rc); 38541a0afa0eSJosef Bacik free_reloc_control(rc); 38553612b495STsutomu Itoh out: 3856aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 3857aca1bba6SLiu Bo 38585d4f98a2SYan Zheng btrfs_free_path(path); 38595d4f98a2SYan Zheng 38605d4f98a2SYan Zheng if (err == 0) { 38615d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 3862aeb935a4SQu Wenruo fs_root = btrfs_grab_root(fs_info->data_reloc_root); 3863aeb935a4SQu Wenruo ASSERT(fs_root); 386466b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 386500246528SJosef Bacik btrfs_put_root(fs_root); 3866932fd26dSJosef Bacik } 38675d4f98a2SYan Zheng return err; 38685d4f98a2SYan Zheng } 38695d4f98a2SYan Zheng 38705d4f98a2SYan Zheng /* 38715d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 38725d4f98a2SYan Zheng * 38735d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 38745d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 38755d4f98a2SYan Zheng */ 38767bfa9535SNikolay Borisov int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len) 38775d4f98a2SYan Zheng { 38787bfa9535SNikolay Borisov struct btrfs_fs_info *fs_info = inode->root->fs_info; 38795d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 38805d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 38815d4f98a2SYan Zheng int ret; 38825d4f98a2SYan Zheng u64 disk_bytenr; 38834577b014SJosef Bacik u64 new_bytenr; 38845d4f98a2SYan Zheng LIST_HEAD(list); 38855d4f98a2SYan Zheng 38867bfa9535SNikolay Borisov ordered = btrfs_lookup_ordered_extent(inode, file_pos); 3887bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 38885d4f98a2SYan Zheng 38897bfa9535SNikolay Borisov disk_bytenr = file_pos + inode->index_cnt; 38900b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 3891a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 389279787eaaSJeff Mahoney if (ret) 389379787eaaSJeff Mahoney goto out; 38945d4f98a2SYan Zheng 38955d4f98a2SYan Zheng while (!list_empty(&list)) { 38965d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 38975d4f98a2SYan Zheng list_del_init(&sums->list); 38985d4f98a2SYan Zheng 38994577b014SJosef Bacik /* 39004577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 39014577b014SJosef Bacik * We need to do this because we will read in entire prealloc 39024577b014SJosef Bacik * extents but we may have written to say the middle of the 39034577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 39044577b014SJosef Bacik * the right disk offset. 39054577b014SJosef Bacik * 39064577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 39074577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 39084577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 39094577b014SJosef Bacik * disk length. 39104577b014SJosef Bacik */ 3911bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 39124577b014SJosef Bacik sums->bytenr = new_bytenr; 39135d4f98a2SYan Zheng 3914f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 39155d4f98a2SYan Zheng } 391679787eaaSJeff Mahoney out: 39175d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 3918411fc6bcSAndi Kleen return ret; 39195d4f98a2SYan Zheng } 39203fd0a558SYan, Zheng 392183d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 39223fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 39233fd0a558SYan, Zheng struct extent_buffer *cow) 39243fd0a558SYan, Zheng { 39250b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 39263fd0a558SYan, Zheng struct reloc_control *rc; 3927a26195a5SQu Wenruo struct btrfs_backref_node *node; 39283fd0a558SYan, Zheng int first_cow = 0; 39293fd0a558SYan, Zheng int level; 393083d4cfd4SJosef Bacik int ret = 0; 39313fd0a558SYan, Zheng 39320b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 39333fd0a558SYan, Zheng if (!rc) 393483d4cfd4SJosef Bacik return 0; 39353fd0a558SYan, Zheng 39363fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 39373fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 39383fd0a558SYan, Zheng 39393fd0a558SYan, Zheng level = btrfs_header_level(buf); 39403fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 39413fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 39423fd0a558SYan, Zheng first_cow = 1; 39433fd0a558SYan, Zheng 39443fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 39453fd0a558SYan, Zheng rc->create_reloc_tree) { 39463fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 39473fd0a558SYan, Zheng 39483fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 39493fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 39503fd0a558SYan, Zheng node->new_bytenr != buf->start); 39513fd0a558SYan, Zheng 3952b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 395367439dadSDavid Sterba atomic_inc(&cow->refs); 39543fd0a558SYan, Zheng node->eb = cow; 39553fd0a558SYan, Zheng node->new_bytenr = cow->start; 39563fd0a558SYan, Zheng 39573fd0a558SYan, Zheng if (!node->pending) { 39583fd0a558SYan, Zheng list_move_tail(&node->list, 39593fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 39603fd0a558SYan, Zheng node->pending = 1; 39613fd0a558SYan, Zheng } 39623fd0a558SYan, Zheng 39633fd0a558SYan, Zheng if (first_cow) 39649569cc20SQu Wenruo mark_block_processed(rc, node); 39653fd0a558SYan, Zheng 39663fd0a558SYan, Zheng if (first_cow && level > 0) 39673fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 39683fd0a558SYan, Zheng } 39693fd0a558SYan, Zheng 397083d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 39713fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 397283d4cfd4SJosef Bacik return ret; 39733fd0a558SYan, Zheng } 39743fd0a558SYan, Zheng 39753fd0a558SYan, Zheng /* 39763fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 397701327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 39783fd0a558SYan, Zheng */ 3979147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 39803fd0a558SYan, Zheng u64 *bytes_to_reserve) 39813fd0a558SYan, Zheng { 398210995c04SQu Wenruo struct btrfs_root *root = pending->root; 398310995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 39843fd0a558SYan, Zheng 39856282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 39863fd0a558SYan, Zheng return; 39873fd0a558SYan, Zheng 39883fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 39893fd0a558SYan, Zheng return; 39903fd0a558SYan, Zheng 39913fd0a558SYan, Zheng root = root->reloc_root; 39923fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 39933fd0a558SYan, Zheng /* 39943fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 39953fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 39963fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 39973fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 39983fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 39993fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 40003fd0a558SYan, Zheng * tree may use more space than it frees. so we need 40013fd0a558SYan, Zheng * reserve extra space. 40023fd0a558SYan, Zheng */ 40033fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 40043fd0a558SYan, Zheng } 40053fd0a558SYan, Zheng 40063fd0a558SYan, Zheng /* 40073fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 40083fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4009f44deb74SJosef Bacik * 4010f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4011f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4012f44deb74SJosef Bacik * rc->reloc_roots. 40133fd0a558SYan, Zheng */ 401449b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 40153fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 40163fd0a558SYan, Zheng { 40173fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 40183fd0a558SYan, Zheng struct btrfs_root *reloc_root; 40193fd0a558SYan, Zheng struct btrfs_root *new_root; 402010995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 40213fd0a558SYan, Zheng int ret; 40223fd0a558SYan, Zheng 40236282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 402449b25e05SJeff Mahoney return 0; 40253fd0a558SYan, Zheng 40263fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 40273fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 40283fd0a558SYan, Zheng 40293fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 40303fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 40313fd0a558SYan, Zheng rc->block_rsv, 40323a584174SLu Fengqi rc->nodes_relocated, true); 403349b25e05SJeff Mahoney if (ret) 403449b25e05SJeff Mahoney return ret; 40353fd0a558SYan, Zheng } 40363fd0a558SYan, Zheng 40373fd0a558SYan, Zheng new_root = pending->snap; 40383fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 40393fd0a558SYan, Zheng new_root->root_key.objectid); 404049b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 404149b25e05SJeff Mahoney return PTR_ERR(reloc_root); 40423fd0a558SYan, Zheng 4043ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4044ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4045f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 40463fd0a558SYan, Zheng 404749b25e05SJeff Mahoney if (rc->create_reloc_tree) 40483fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 404949b25e05SJeff Mahoney return ret; 40503fd0a558SYan, Zheng } 4051