1c1d7c514SDavid Sterba // SPDX-License-Identifier: GPL-2.0 25d4f98a2SYan Zheng /* 35d4f98a2SYan Zheng * Copyright (C) 2009 Oracle. All rights reserved. 45d4f98a2SYan Zheng */ 55d4f98a2SYan Zheng 65d4f98a2SYan Zheng #include <linux/sched.h> 75d4f98a2SYan Zheng #include <linux/pagemap.h> 85d4f98a2SYan Zheng #include <linux/writeback.h> 95d4f98a2SYan Zheng #include <linux/blkdev.h> 105d4f98a2SYan Zheng #include <linux/rbtree.h> 115a0e3ad6STejun Heo #include <linux/slab.h> 12726a3421SQu Wenruo #include <linux/error-injection.h> 135d4f98a2SYan Zheng #include "ctree.h" 145d4f98a2SYan Zheng #include "disk-io.h" 155d4f98a2SYan Zheng #include "transaction.h" 165d4f98a2SYan Zheng #include "volumes.h" 175d4f98a2SYan Zheng #include "locking.h" 185d4f98a2SYan Zheng #include "btrfs_inode.h" 195d4f98a2SYan Zheng #include "async-thread.h" 200af3d00bSJosef Bacik #include "free-space-cache.h" 2162b99540SQu Wenruo #include "qgroup.h" 22cdccee99SLiu Bo #include "print-tree.h" 2386736342SJosef Bacik #include "delalloc-space.h" 24aac0023cSJosef Bacik #include "block-group.h" 2519b546d7SQu Wenruo #include "backref.h" 26e9a28dc5SQu Wenruo #include "misc.h" 275d4f98a2SYan Zheng 285d4f98a2SYan Zheng /* 290c891389SQu Wenruo * Relocation overview 300c891389SQu Wenruo * 310c891389SQu Wenruo * [What does relocation do] 320c891389SQu Wenruo * 330c891389SQu Wenruo * The objective of relocation is to relocate all extents of the target block 340c891389SQu Wenruo * group to other block groups. 350c891389SQu Wenruo * This is utilized by resize (shrink only), profile converting, compacting 360c891389SQu Wenruo * space, or balance routine to spread chunks over devices. 370c891389SQu Wenruo * 380c891389SQu Wenruo * Before | After 390c891389SQu Wenruo * ------------------------------------------------------------------ 400c891389SQu Wenruo * BG A: 10 data extents | BG A: deleted 410c891389SQu Wenruo * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated) 420c891389SQu Wenruo * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated) 430c891389SQu Wenruo * 440c891389SQu Wenruo * [How does relocation work] 450c891389SQu Wenruo * 460c891389SQu Wenruo * 1. Mark the target block group read-only 470c891389SQu Wenruo * New extents won't be allocated from the target block group. 480c891389SQu Wenruo * 490c891389SQu Wenruo * 2.1 Record each extent in the target block group 500c891389SQu Wenruo * To build a proper map of extents to be relocated. 510c891389SQu Wenruo * 520c891389SQu Wenruo * 2.2 Build data reloc tree and reloc trees 530c891389SQu Wenruo * Data reloc tree will contain an inode, recording all newly relocated 540c891389SQu Wenruo * data extents. 550c891389SQu Wenruo * There will be only one data reloc tree for one data block group. 560c891389SQu Wenruo * 570c891389SQu Wenruo * Reloc tree will be a special snapshot of its source tree, containing 580c891389SQu Wenruo * relocated tree blocks. 590c891389SQu Wenruo * Each tree referring to a tree block in target block group will get its 600c891389SQu Wenruo * reloc tree built. 610c891389SQu Wenruo * 620c891389SQu Wenruo * 2.3 Swap source tree with its corresponding reloc tree 630c891389SQu Wenruo * Each involved tree only refers to new extents after swap. 640c891389SQu Wenruo * 650c891389SQu Wenruo * 3. Cleanup reloc trees and data reloc tree. 660c891389SQu Wenruo * As old extents in the target block group are still referenced by reloc 670c891389SQu Wenruo * trees, we need to clean them up before really freeing the target block 680c891389SQu Wenruo * group. 690c891389SQu Wenruo * 700c891389SQu Wenruo * The main complexity is in steps 2.2 and 2.3. 710c891389SQu Wenruo * 720c891389SQu Wenruo * The entry point of relocation is relocate_block_group() function. 730c891389SQu Wenruo */ 740c891389SQu Wenruo 750647bf56SWang Shilong #define RELOCATION_RESERVED_NODES 256 762a979612SQu Wenruo /* 775d4f98a2SYan Zheng * map address of tree root to tree 785d4f98a2SYan Zheng */ 795d4f98a2SYan Zheng struct mapping_node { 80e9a28dc5SQu Wenruo struct { 815d4f98a2SYan Zheng struct rb_node rb_node; 825d4f98a2SYan Zheng u64 bytenr; 83e9a28dc5SQu Wenruo }; /* Use rb_simle_node for search/insert */ 845d4f98a2SYan Zheng void *data; 855d4f98a2SYan Zheng }; 865d4f98a2SYan Zheng 875d4f98a2SYan Zheng struct mapping_tree { 885d4f98a2SYan Zheng struct rb_root rb_root; 895d4f98a2SYan Zheng spinlock_t lock; 905d4f98a2SYan Zheng }; 915d4f98a2SYan Zheng 925d4f98a2SYan Zheng /* 935d4f98a2SYan Zheng * present a tree block to process 945d4f98a2SYan Zheng */ 955d4f98a2SYan Zheng struct tree_block { 96e9a28dc5SQu Wenruo struct { 975d4f98a2SYan Zheng struct rb_node rb_node; 985d4f98a2SYan Zheng u64 bytenr; 99e9a28dc5SQu Wenruo }; /* Use rb_simple_node for search/insert */ 100f7ba2d37SJosef Bacik u64 owner; 1015d4f98a2SYan Zheng struct btrfs_key key; 1025d4f98a2SYan Zheng unsigned int level:8; 1035d4f98a2SYan Zheng unsigned int key_ready:1; 1045d4f98a2SYan Zheng }; 1055d4f98a2SYan Zheng 1060257bb82SYan, Zheng #define MAX_EXTENTS 128 1070257bb82SYan, Zheng 1080257bb82SYan, Zheng struct file_extent_cluster { 1090257bb82SYan, Zheng u64 start; 1100257bb82SYan, Zheng u64 end; 1110257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 1120257bb82SYan, Zheng unsigned int nr; 1130257bb82SYan, Zheng }; 1140257bb82SYan, Zheng 1155d4f98a2SYan Zheng struct reloc_control { 1165d4f98a2SYan Zheng /* block group to relocate */ 11732da5386SDavid Sterba struct btrfs_block_group *block_group; 1185d4f98a2SYan Zheng /* extent tree */ 1195d4f98a2SYan Zheng struct btrfs_root *extent_root; 1205d4f98a2SYan Zheng /* inode for moving data */ 1215d4f98a2SYan Zheng struct inode *data_inode; 1223fd0a558SYan, Zheng 1233fd0a558SYan, Zheng struct btrfs_block_rsv *block_rsv; 1243fd0a558SYan, Zheng 125a26195a5SQu Wenruo struct btrfs_backref_cache backref_cache; 1263fd0a558SYan, Zheng 1273fd0a558SYan, Zheng struct file_extent_cluster cluster; 1285d4f98a2SYan Zheng /* tree blocks have been processed */ 1295d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 1305d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 1315d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 1325d4f98a2SYan Zheng /* list of reloc trees */ 1335d4f98a2SYan Zheng struct list_head reloc_roots; 134d2311e69SQu Wenruo /* list of subvolume trees that get relocated */ 135d2311e69SQu Wenruo struct list_head dirty_subvol_roots; 1363fd0a558SYan, Zheng /* size of metadata reservation for merging reloc trees */ 1373fd0a558SYan, Zheng u64 merging_rsv_size; 1383fd0a558SYan, Zheng /* size of relocated tree nodes */ 1393fd0a558SYan, Zheng u64 nodes_relocated; 1400647bf56SWang Shilong /* reserved size for block group relocation*/ 1410647bf56SWang Shilong u64 reserved_bytes; 1423fd0a558SYan, Zheng 1435d4f98a2SYan Zheng u64 search_start; 1445d4f98a2SYan Zheng u64 extents_found; 1453fd0a558SYan, Zheng 1463fd0a558SYan, Zheng unsigned int stage:8; 1473fd0a558SYan, Zheng unsigned int create_reloc_tree:1; 1483fd0a558SYan, Zheng unsigned int merge_reloc_tree:1; 1495d4f98a2SYan Zheng unsigned int found_file_extent:1; 1505d4f98a2SYan Zheng }; 1515d4f98a2SYan Zheng 1525d4f98a2SYan Zheng /* stages of data relocation */ 1535d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 1545d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 1555d4f98a2SYan Zheng 1569569cc20SQu Wenruo static void mark_block_processed(struct reloc_control *rc, 157a26195a5SQu Wenruo struct btrfs_backref_node *node) 1589569cc20SQu Wenruo { 1599569cc20SQu Wenruo u32 blocksize; 1609569cc20SQu Wenruo 1619569cc20SQu Wenruo if (node->level == 0 || 1629569cc20SQu Wenruo in_range(node->bytenr, rc->block_group->start, 1639569cc20SQu Wenruo rc->block_group->length)) { 1649569cc20SQu Wenruo blocksize = rc->extent_root->fs_info->nodesize; 1659569cc20SQu Wenruo set_extent_bits(&rc->processed_blocks, node->bytenr, 1669569cc20SQu Wenruo node->bytenr + blocksize - 1, EXTENT_DIRTY); 1679569cc20SQu Wenruo } 1689569cc20SQu Wenruo node->processed = 1; 1699569cc20SQu Wenruo } 1709569cc20SQu Wenruo 1715d4f98a2SYan Zheng 1725d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 1735d4f98a2SYan Zheng { 1746bef4d31SEric Paris tree->rb_root = RB_ROOT; 1755d4f98a2SYan Zheng spin_lock_init(&tree->lock); 1765d4f98a2SYan Zheng } 1775d4f98a2SYan Zheng 1785d4f98a2SYan Zheng /* 1795d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 1805d4f98a2SYan Zheng */ 181a26195a5SQu Wenruo static struct btrfs_backref_node *walk_up_backref( 182a26195a5SQu Wenruo struct btrfs_backref_node *node, 183a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 1845d4f98a2SYan Zheng { 185a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1865d4f98a2SYan Zheng int idx = *index; 1875d4f98a2SYan Zheng 1885d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 1895d4f98a2SYan Zheng edge = list_entry(node->upper.next, 190a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 1915d4f98a2SYan Zheng edges[idx++] = edge; 1925d4f98a2SYan Zheng node = edge->node[UPPER]; 1935d4f98a2SYan Zheng } 1943fd0a558SYan, Zheng BUG_ON(node->detached); 1955d4f98a2SYan Zheng *index = idx; 1965d4f98a2SYan Zheng return node; 1975d4f98a2SYan Zheng } 1985d4f98a2SYan Zheng 1995d4f98a2SYan Zheng /* 2005d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 2015d4f98a2SYan Zheng */ 202a26195a5SQu Wenruo static struct btrfs_backref_node *walk_down_backref( 203a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 2045d4f98a2SYan Zheng { 205a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 206a26195a5SQu Wenruo struct btrfs_backref_node *lower; 2075d4f98a2SYan Zheng int idx = *index; 2085d4f98a2SYan Zheng 2095d4f98a2SYan Zheng while (idx > 0) { 2105d4f98a2SYan Zheng edge = edges[idx - 1]; 2115d4f98a2SYan Zheng lower = edge->node[LOWER]; 2125d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 2135d4f98a2SYan Zheng idx--; 2145d4f98a2SYan Zheng continue; 2155d4f98a2SYan Zheng } 2165d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 217a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 2185d4f98a2SYan Zheng edges[idx - 1] = edge; 2195d4f98a2SYan Zheng *index = idx; 2205d4f98a2SYan Zheng return edge->node[UPPER]; 2215d4f98a2SYan Zheng } 2225d4f98a2SYan Zheng *index = 0; 2235d4f98a2SYan Zheng return NULL; 2245d4f98a2SYan Zheng } 2255d4f98a2SYan Zheng 226a26195a5SQu Wenruo static void update_backref_node(struct btrfs_backref_cache *cache, 227a26195a5SQu Wenruo struct btrfs_backref_node *node, u64 bytenr) 2283fd0a558SYan, Zheng { 2293fd0a558SYan, Zheng struct rb_node *rb_node; 2303fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 2313fd0a558SYan, Zheng node->bytenr = bytenr; 232e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node); 23343c04fb1SJeff Mahoney if (rb_node) 234982c92cbSQu Wenruo btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST); 2353fd0a558SYan, Zheng } 2363fd0a558SYan, Zheng 2373fd0a558SYan, Zheng /* 2383fd0a558SYan, Zheng * update backref cache after a transaction commit 2393fd0a558SYan, Zheng */ 2403fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 241a26195a5SQu Wenruo struct btrfs_backref_cache *cache) 2423fd0a558SYan, Zheng { 243a26195a5SQu Wenruo struct btrfs_backref_node *node; 2443fd0a558SYan, Zheng int level = 0; 2453fd0a558SYan, Zheng 2463fd0a558SYan, Zheng if (cache->last_trans == 0) { 2473fd0a558SYan, Zheng cache->last_trans = trans->transid; 2483fd0a558SYan, Zheng return 0; 2493fd0a558SYan, Zheng } 2503fd0a558SYan, Zheng 2513fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 2523fd0a558SYan, Zheng return 0; 2533fd0a558SYan, Zheng 2543fd0a558SYan, Zheng /* 2553fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 2563fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 2573fd0a558SYan, Zheng * so the detached nodes are no longer useful. 2583fd0a558SYan, Zheng */ 2593fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 2603fd0a558SYan, Zheng node = list_entry(cache->detached.next, 261a26195a5SQu Wenruo struct btrfs_backref_node, list); 262023acb07SQu Wenruo btrfs_backref_cleanup_node(cache, node); 2633fd0a558SYan, Zheng } 2643fd0a558SYan, Zheng 2653fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 2663fd0a558SYan, Zheng node = list_entry(cache->changed.next, 267a26195a5SQu Wenruo struct btrfs_backref_node, list); 2683fd0a558SYan, Zheng list_del_init(&node->list); 2693fd0a558SYan, Zheng BUG_ON(node->pending); 2703fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 2713fd0a558SYan, Zheng } 2723fd0a558SYan, Zheng 2733fd0a558SYan, Zheng /* 2743fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 2753fd0a558SYan, Zheng * errors during processing the pending nodes. 2763fd0a558SYan, Zheng */ 2773fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 2783fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 2793fd0a558SYan, Zheng BUG_ON(!node->pending); 2803fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 2813fd0a558SYan, Zheng continue; 2823fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 2833fd0a558SYan, Zheng } 2843fd0a558SYan, Zheng } 2853fd0a558SYan, Zheng 2863fd0a558SYan, Zheng cache->last_trans = 0; 2873fd0a558SYan, Zheng return 1; 2883fd0a558SYan, Zheng } 2893fd0a558SYan, Zheng 2906282675eSQu Wenruo static bool reloc_root_is_dead(struct btrfs_root *root) 2916282675eSQu Wenruo { 2926282675eSQu Wenruo /* 2936282675eSQu Wenruo * Pair with set_bit/clear_bit in clean_dirty_subvols and 2946282675eSQu Wenruo * btrfs_update_reloc_root. We need to see the updated bit before 2956282675eSQu Wenruo * trying to access reloc_root 2966282675eSQu Wenruo */ 2976282675eSQu Wenruo smp_rmb(); 2986282675eSQu Wenruo if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state)) 2996282675eSQu Wenruo return true; 3006282675eSQu Wenruo return false; 3016282675eSQu Wenruo } 3026282675eSQu Wenruo 3036282675eSQu Wenruo /* 3046282675eSQu Wenruo * Check if this subvolume tree has valid reloc tree. 3056282675eSQu Wenruo * 3066282675eSQu Wenruo * Reloc tree after swap is considered dead, thus not considered as valid. 3076282675eSQu Wenruo * This is enough for most callers, as they don't distinguish dead reloc root 30855465730SQu Wenruo * from no reloc root. But btrfs_should_ignore_reloc_root() below is a 30955465730SQu Wenruo * special case. 3106282675eSQu Wenruo */ 3116282675eSQu Wenruo static bool have_reloc_root(struct btrfs_root *root) 3126282675eSQu Wenruo { 3136282675eSQu Wenruo if (reloc_root_is_dead(root)) 3146282675eSQu Wenruo return false; 3156282675eSQu Wenruo if (!root->reloc_root) 3166282675eSQu Wenruo return false; 3176282675eSQu Wenruo return true; 3186282675eSQu Wenruo } 319f2a97a9dSDavid Sterba 32055465730SQu Wenruo int btrfs_should_ignore_reloc_root(struct btrfs_root *root) 3213fd0a558SYan, Zheng { 3223fd0a558SYan, Zheng struct btrfs_root *reloc_root; 3233fd0a558SYan, Zheng 32492a7cc42SQu Wenruo if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 3253fd0a558SYan, Zheng return 0; 3263fd0a558SYan, Zheng 3276282675eSQu Wenruo /* This root has been merged with its reloc tree, we can ignore it */ 3286282675eSQu Wenruo if (reloc_root_is_dead(root)) 3296282675eSQu Wenruo return 1; 3306282675eSQu Wenruo 3313fd0a558SYan, Zheng reloc_root = root->reloc_root; 3323fd0a558SYan, Zheng if (!reloc_root) 3333fd0a558SYan, Zheng return 0; 3343fd0a558SYan, Zheng 3354d4225fcSJosef Bacik if (btrfs_header_generation(reloc_root->commit_root) == 3364d4225fcSJosef Bacik root->fs_info->running_transaction->transid) 3373fd0a558SYan, Zheng return 0; 3383fd0a558SYan, Zheng /* 3393fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 3403fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 3413fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 3423fd0a558SYan, Zheng * relocation. 3433fd0a558SYan, Zheng */ 3443fd0a558SYan, Zheng return 1; 3453fd0a558SYan, Zheng } 34655465730SQu Wenruo 3475d4f98a2SYan Zheng /* 3485d4f98a2SYan Zheng * find reloc tree by address of tree root 3495d4f98a2SYan Zheng */ 3502433bea5SQu Wenruo struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr) 3515d4f98a2SYan Zheng { 3522433bea5SQu Wenruo struct reloc_control *rc = fs_info->reloc_ctl; 3535d4f98a2SYan Zheng struct rb_node *rb_node; 3545d4f98a2SYan Zheng struct mapping_node *node; 3555d4f98a2SYan Zheng struct btrfs_root *root = NULL; 3565d4f98a2SYan Zheng 3572433bea5SQu Wenruo ASSERT(rc); 3585d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 359e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr); 3605d4f98a2SYan Zheng if (rb_node) { 3615d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 3625d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 3635d4f98a2SYan Zheng } 3645d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 36500246528SJosef Bacik return btrfs_grab_root(root); 3665d4f98a2SYan Zheng } 3675d4f98a2SYan Zheng 3685d4f98a2SYan Zheng /* 36929db137bSQu Wenruo * For useless nodes, do two major clean ups: 37029db137bSQu Wenruo * 37129db137bSQu Wenruo * - Cleanup the children edges and nodes 37229db137bSQu Wenruo * If child node is also orphan (no parent) during cleanup, then the child 37329db137bSQu Wenruo * node will also be cleaned up. 37429db137bSQu Wenruo * 37529db137bSQu Wenruo * - Freeing up leaves (level 0), keeps nodes detached 37629db137bSQu Wenruo * For nodes, the node is still cached as "detached" 37729db137bSQu Wenruo * 37829db137bSQu Wenruo * Return false if @node is not in the @useless_nodes list. 37929db137bSQu Wenruo * Return true if @node is in the @useless_nodes list. 38029db137bSQu Wenruo */ 38129db137bSQu Wenruo static bool handle_useless_nodes(struct reloc_control *rc, 382a26195a5SQu Wenruo struct btrfs_backref_node *node) 38329db137bSQu Wenruo { 384a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 38529db137bSQu Wenruo struct list_head *useless_node = &cache->useless_node; 38629db137bSQu Wenruo bool ret = false; 38729db137bSQu Wenruo 38829db137bSQu Wenruo while (!list_empty(useless_node)) { 389a26195a5SQu Wenruo struct btrfs_backref_node *cur; 39029db137bSQu Wenruo 391a26195a5SQu Wenruo cur = list_first_entry(useless_node, struct btrfs_backref_node, 39229db137bSQu Wenruo list); 39329db137bSQu Wenruo list_del_init(&cur->list); 39429db137bSQu Wenruo 39529db137bSQu Wenruo /* Only tree root nodes can be added to @useless_nodes */ 39629db137bSQu Wenruo ASSERT(list_empty(&cur->upper)); 39729db137bSQu Wenruo 39829db137bSQu Wenruo if (cur == node) 39929db137bSQu Wenruo ret = true; 40029db137bSQu Wenruo 40129db137bSQu Wenruo /* The node is the lowest node */ 40229db137bSQu Wenruo if (cur->lowest) { 40329db137bSQu Wenruo list_del_init(&cur->lower); 40429db137bSQu Wenruo cur->lowest = 0; 40529db137bSQu Wenruo } 40629db137bSQu Wenruo 40729db137bSQu Wenruo /* Cleanup the lower edges */ 40829db137bSQu Wenruo while (!list_empty(&cur->lower)) { 409a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 410a26195a5SQu Wenruo struct btrfs_backref_node *lower; 41129db137bSQu Wenruo 41229db137bSQu Wenruo edge = list_entry(cur->lower.next, 413a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 41429db137bSQu Wenruo list_del(&edge->list[UPPER]); 41529db137bSQu Wenruo list_del(&edge->list[LOWER]); 41629db137bSQu Wenruo lower = edge->node[LOWER]; 417741188d3SQu Wenruo btrfs_backref_free_edge(cache, edge); 41829db137bSQu Wenruo 41929db137bSQu Wenruo /* Child node is also orphan, queue for cleanup */ 42029db137bSQu Wenruo if (list_empty(&lower->upper)) 42129db137bSQu Wenruo list_add(&lower->list, useless_node); 42229db137bSQu Wenruo } 42329db137bSQu Wenruo /* Mark this block processed for relocation */ 42429db137bSQu Wenruo mark_block_processed(rc, cur); 42529db137bSQu Wenruo 42629db137bSQu Wenruo /* 42729db137bSQu Wenruo * Backref nodes for tree leaves are deleted from the cache. 42829db137bSQu Wenruo * Backref nodes for upper level tree blocks are left in the 42929db137bSQu Wenruo * cache to avoid unnecessary backref lookup. 43029db137bSQu Wenruo */ 43129db137bSQu Wenruo if (cur->level > 0) { 43229db137bSQu Wenruo list_add(&cur->list, &cache->detached); 43329db137bSQu Wenruo cur->detached = 1; 43429db137bSQu Wenruo } else { 43529db137bSQu Wenruo rb_erase(&cur->rb_node, &cache->rb_root); 436741188d3SQu Wenruo btrfs_backref_free_node(cache, cur); 43729db137bSQu Wenruo } 43829db137bSQu Wenruo } 43929db137bSQu Wenruo return ret; 44029db137bSQu Wenruo } 44129db137bSQu Wenruo 44229db137bSQu Wenruo /* 443e7d571c7SQu Wenruo * Build backref tree for a given tree block. Root of the backref tree 444e7d571c7SQu Wenruo * corresponds the tree block, leaves of the backref tree correspond roots of 445e7d571c7SQu Wenruo * b-trees that reference the tree block. 446e7d571c7SQu Wenruo * 447e7d571c7SQu Wenruo * The basic idea of this function is check backrefs of a given block to find 448e7d571c7SQu Wenruo * upper level blocks that reference the block, and then check backrefs of 449e7d571c7SQu Wenruo * these upper level blocks recursively. The recursion stops when tree root is 450e7d571c7SQu Wenruo * reached or backrefs for the block is cached. 451e7d571c7SQu Wenruo * 452e7d571c7SQu Wenruo * NOTE: if we find that backrefs for a block are cached, we know backrefs for 453e7d571c7SQu Wenruo * all upper level blocks that directly/indirectly reference the block are also 454e7d571c7SQu Wenruo * cached. 455e7d571c7SQu Wenruo */ 456a26195a5SQu Wenruo static noinline_for_stack struct btrfs_backref_node *build_backref_tree( 457e7d571c7SQu Wenruo struct reloc_control *rc, struct btrfs_key *node_key, 458e7d571c7SQu Wenruo int level, u64 bytenr) 459e7d571c7SQu Wenruo { 460e7d571c7SQu Wenruo struct btrfs_backref_iter *iter; 461a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 462e7d571c7SQu Wenruo /* For searching parent of TREE_BLOCK_REF */ 463e7d571c7SQu Wenruo struct btrfs_path *path; 464a26195a5SQu Wenruo struct btrfs_backref_node *cur; 465a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 466a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 467e7d571c7SQu Wenruo int ret; 468e7d571c7SQu Wenruo int err = 0; 469e7d571c7SQu Wenruo 470e7d571c7SQu Wenruo iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS); 471e7d571c7SQu Wenruo if (!iter) 472e7d571c7SQu Wenruo return ERR_PTR(-ENOMEM); 473e7d571c7SQu Wenruo path = btrfs_alloc_path(); 474e7d571c7SQu Wenruo if (!path) { 475e7d571c7SQu Wenruo err = -ENOMEM; 476e7d571c7SQu Wenruo goto out; 477e7d571c7SQu Wenruo } 478e7d571c7SQu Wenruo 479b1818dabSQu Wenruo node = btrfs_backref_alloc_node(cache, bytenr, level); 480e7d571c7SQu Wenruo if (!node) { 481e7d571c7SQu Wenruo err = -ENOMEM; 482e7d571c7SQu Wenruo goto out; 483e7d571c7SQu Wenruo } 484e7d571c7SQu Wenruo 485e7d571c7SQu Wenruo node->lowest = 1; 486e7d571c7SQu Wenruo cur = node; 487e7d571c7SQu Wenruo 488e7d571c7SQu Wenruo /* Breadth-first search to build backref cache */ 489e7d571c7SQu Wenruo do { 4901b60d2ecSQu Wenruo ret = btrfs_backref_add_tree_node(cache, path, iter, node_key, 4911b60d2ecSQu Wenruo cur); 492e7d571c7SQu Wenruo if (ret < 0) { 493e7d571c7SQu Wenruo err = ret; 494e7d571c7SQu Wenruo goto out; 495e7d571c7SQu Wenruo } 496e7d571c7SQu Wenruo edge = list_first_entry_or_null(&cache->pending_edge, 497a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 498e7d571c7SQu Wenruo /* 499e7d571c7SQu Wenruo * The pending list isn't empty, take the first block to 500e7d571c7SQu Wenruo * process 501e7d571c7SQu Wenruo */ 502e7d571c7SQu Wenruo if (edge) { 5035d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 5045d4f98a2SYan Zheng cur = edge->node[UPPER]; 5055d4f98a2SYan Zheng } 506e7d571c7SQu Wenruo } while (edge); 5075d4f98a2SYan Zheng 5081f872924SQu Wenruo /* Finish the upper linkage of newly added edges/nodes */ 509fc997ed0SQu Wenruo ret = btrfs_backref_finish_upper_links(cache, node); 5101f872924SQu Wenruo if (ret < 0) { 5111f872924SQu Wenruo err = ret; 51275bfb9afSJosef Bacik goto out; 51375bfb9afSJosef Bacik } 51475bfb9afSJosef Bacik 51529db137bSQu Wenruo if (handle_useless_nodes(rc, node)) 5163fd0a558SYan, Zheng node = NULL; 5175d4f98a2SYan Zheng out: 51871f572a9SQu Wenruo btrfs_backref_iter_free(iter); 51971f572a9SQu Wenruo btrfs_free_path(path); 5205d4f98a2SYan Zheng if (err) { 5211b23ea18SQu Wenruo btrfs_backref_error_cleanup(cache, node); 5225d4f98a2SYan Zheng return ERR_PTR(err); 5235d4f98a2SYan Zheng } 52475bfb9afSJosef Bacik ASSERT(!node || !node->detached); 52584780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 52684780289SQu Wenruo list_empty(&cache->pending_edge)); 5275d4f98a2SYan Zheng return node; 5285d4f98a2SYan Zheng } 5295d4f98a2SYan Zheng 5305d4f98a2SYan Zheng /* 5313fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 5323fd0a558SYan, Zheng * the backref node is created by cloning backref node that 5333fd0a558SYan, Zheng * corresponds to root of source tree 5343fd0a558SYan, Zheng */ 5353fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 5363fd0a558SYan, Zheng struct reloc_control *rc, 5373fd0a558SYan, Zheng struct btrfs_root *src, 5383fd0a558SYan, Zheng struct btrfs_root *dest) 5393fd0a558SYan, Zheng { 5403fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 541a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 542a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 543a26195a5SQu Wenruo struct btrfs_backref_node *new_node; 544a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 545a26195a5SQu Wenruo struct btrfs_backref_edge *new_edge; 5463fd0a558SYan, Zheng struct rb_node *rb_node; 5473fd0a558SYan, Zheng 5483fd0a558SYan, Zheng if (cache->last_trans > 0) 5493fd0a558SYan, Zheng update_backref_cache(trans, cache); 5503fd0a558SYan, Zheng 551e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start); 5523fd0a558SYan, Zheng if (rb_node) { 553a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, rb_node); 5543fd0a558SYan, Zheng if (node->detached) 5553fd0a558SYan, Zheng node = NULL; 5563fd0a558SYan, Zheng else 5573fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 5583fd0a558SYan, Zheng } 5593fd0a558SYan, Zheng 5603fd0a558SYan, Zheng if (!node) { 561e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, 5623fd0a558SYan, Zheng reloc_root->commit_root->start); 5633fd0a558SYan, Zheng if (rb_node) { 564a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, 5653fd0a558SYan, Zheng rb_node); 5663fd0a558SYan, Zheng BUG_ON(node->detached); 5673fd0a558SYan, Zheng } 5683fd0a558SYan, Zheng } 5693fd0a558SYan, Zheng 5703fd0a558SYan, Zheng if (!node) 5713fd0a558SYan, Zheng return 0; 5723fd0a558SYan, Zheng 573b1818dabSQu Wenruo new_node = btrfs_backref_alloc_node(cache, dest->node->start, 574b1818dabSQu Wenruo node->level); 5753fd0a558SYan, Zheng if (!new_node) 5763fd0a558SYan, Zheng return -ENOMEM; 5773fd0a558SYan, Zheng 5783fd0a558SYan, Zheng new_node->lowest = node->lowest; 5796848ad64SYan, Zheng new_node->checked = 1; 58000246528SJosef Bacik new_node->root = btrfs_grab_root(dest); 5810b530bc5SJosef Bacik ASSERT(new_node->root); 5823fd0a558SYan, Zheng 5833fd0a558SYan, Zheng if (!node->lowest) { 5843fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 58547254d07SQu Wenruo new_edge = btrfs_backref_alloc_edge(cache); 5863fd0a558SYan, Zheng if (!new_edge) 5873fd0a558SYan, Zheng goto fail; 5883fd0a558SYan, Zheng 589f39911e5SQu Wenruo btrfs_backref_link_edge(new_edge, edge->node[LOWER], 590f39911e5SQu Wenruo new_node, LINK_UPPER); 5913fd0a558SYan, Zheng } 59276b9e23dSMiao Xie } else { 59376b9e23dSMiao Xie list_add_tail(&new_node->lower, &cache->leaves); 5943fd0a558SYan, Zheng } 5953fd0a558SYan, Zheng 596e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr, 5973fd0a558SYan, Zheng &new_node->rb_node); 59843c04fb1SJeff Mahoney if (rb_node) 599982c92cbSQu Wenruo btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST); 6003fd0a558SYan, Zheng 6013fd0a558SYan, Zheng if (!new_node->lowest) { 6023fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 6033fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 6043fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 6053fd0a558SYan, Zheng } 6063fd0a558SYan, Zheng } 6073fd0a558SYan, Zheng return 0; 6083fd0a558SYan, Zheng fail: 6093fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 6103fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 611a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 6123fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 613741188d3SQu Wenruo btrfs_backref_free_edge(cache, new_edge); 6143fd0a558SYan, Zheng } 615741188d3SQu Wenruo btrfs_backref_free_node(cache, new_node); 6163fd0a558SYan, Zheng return -ENOMEM; 6173fd0a558SYan, Zheng } 6183fd0a558SYan, Zheng 6193fd0a558SYan, Zheng /* 6205d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 6215d4f98a2SYan Zheng */ 622ffd7b339SJeff Mahoney static int __must_check __add_reloc_root(struct btrfs_root *root) 6235d4f98a2SYan Zheng { 6240b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 6255d4f98a2SYan Zheng struct rb_node *rb_node; 6265d4f98a2SYan Zheng struct mapping_node *node; 6270b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 6285d4f98a2SYan Zheng 6295d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 630ffd7b339SJeff Mahoney if (!node) 631ffd7b339SJeff Mahoney return -ENOMEM; 6325d4f98a2SYan Zheng 633ea287ab1SJosef Bacik node->bytenr = root->commit_root->start; 6345d4f98a2SYan Zheng node->data = root; 6355d4f98a2SYan Zheng 6365d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 637e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 6385d4f98a2SYan Zheng node->bytenr, &node->rb_node); 6395d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 640ffd7b339SJeff Mahoney if (rb_node) { 6410b246afaSJeff Mahoney btrfs_panic(fs_info, -EEXIST, 6425d163e0eSJeff Mahoney "Duplicate root found for start=%llu while inserting into relocation tree", 6435d163e0eSJeff Mahoney node->bytenr); 644ffd7b339SJeff Mahoney } 6455d4f98a2SYan Zheng 6465d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 6475d4f98a2SYan Zheng return 0; 6485d4f98a2SYan Zheng } 6495d4f98a2SYan Zheng 6505d4f98a2SYan Zheng /* 651c974c464SWang Shilong * helper to delete the 'address of tree root -> reloc tree' 6525d4f98a2SYan Zheng * mapping 6535d4f98a2SYan Zheng */ 654c974c464SWang Shilong static void __del_reloc_root(struct btrfs_root *root) 6555d4f98a2SYan Zheng { 6560b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 6575d4f98a2SYan Zheng struct rb_node *rb_node; 6585d4f98a2SYan Zheng struct mapping_node *node = NULL; 6590b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 660f44deb74SJosef Bacik bool put_ref = false; 6615d4f98a2SYan Zheng 66265c6e82bSQu Wenruo if (rc && root->node) { 6635d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 664e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 665ea287ab1SJosef Bacik root->commit_root->start); 666c974c464SWang Shilong if (rb_node) { 667c974c464SWang Shilong node = rb_entry(rb_node, struct mapping_node, rb_node); 668c974c464SWang Shilong rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 669ea287ab1SJosef Bacik RB_CLEAR_NODE(&node->rb_node); 670c974c464SWang Shilong } 671c974c464SWang Shilong spin_unlock(&rc->reloc_root_tree.lock); 672c78a10aeSJosef Bacik ASSERT(!node || (struct btrfs_root *)node->data == root); 673389305b2SQu Wenruo } 674c974c464SWang Shilong 675f44deb74SJosef Bacik /* 676f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 677f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 678f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 679f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 680f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 681f44deb74SJosef Bacik * messing with the list is in charge of the reference. 682f44deb74SJosef Bacik */ 6830b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 684f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 685f44deb74SJosef Bacik put_ref = true; 686c974c464SWang Shilong list_del_init(&root->root_list); 687f44deb74SJosef Bacik } 6880b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 689f44deb74SJosef Bacik if (put_ref) 690f44deb74SJosef Bacik btrfs_put_root(root); 691c974c464SWang Shilong kfree(node); 692c974c464SWang Shilong } 693c974c464SWang Shilong 694c974c464SWang Shilong /* 695c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 696c974c464SWang Shilong * mapping 697c974c464SWang Shilong */ 698ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 699c974c464SWang Shilong { 7000b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 701c974c464SWang Shilong struct rb_node *rb_node; 702c974c464SWang Shilong struct mapping_node *node = NULL; 7030b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 704c974c464SWang Shilong 705c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 706e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 707ea287ab1SJosef Bacik root->commit_root->start); 7085d4f98a2SYan Zheng if (rb_node) { 7095d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 7105d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 7115d4f98a2SYan Zheng } 7125d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 7135d4f98a2SYan Zheng 7148f71f3e0SLiu Bo if (!node) 7158f71f3e0SLiu Bo return 0; 7165d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 7175d4f98a2SYan Zheng 7185d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 719ea287ab1SJosef Bacik node->bytenr = root->node->start; 720e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 7215d4f98a2SYan Zheng node->bytenr, &node->rb_node); 7225d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 72343c04fb1SJeff Mahoney if (rb_node) 724982c92cbSQu Wenruo btrfs_backref_panic(fs_info, node->bytenr, -EEXIST); 7255d4f98a2SYan Zheng return 0; 7265d4f98a2SYan Zheng } 7275d4f98a2SYan Zheng 7283fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 7293fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 7305d4f98a2SYan Zheng { 7310b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 7325d4f98a2SYan Zheng struct btrfs_root *reloc_root; 7335d4f98a2SYan Zheng struct extent_buffer *eb; 7345d4f98a2SYan Zheng struct btrfs_root_item *root_item; 7355d4f98a2SYan Zheng struct btrfs_key root_key; 7365d4f98a2SYan Zheng int ret; 7375d4f98a2SYan Zheng 7385d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 7395d4f98a2SYan Zheng BUG_ON(!root_item); 7405d4f98a2SYan Zheng 7415d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 7425d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 7433fd0a558SYan, Zheng root_key.offset = objectid; 7445d4f98a2SYan Zheng 7453fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 746054570a1SFilipe Manana u64 commit_root_gen; 747054570a1SFilipe Manana 7483fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 7495d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 7505d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 7515d4f98a2SYan Zheng BUG_ON(ret); 752054570a1SFilipe Manana /* 753054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 754054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 755054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 756054570a1SFilipe Manana * either inside the critical section of a transaction commit 757054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 758054570a1SFilipe Manana * it's created before the transaction commit is started. 759054570a1SFilipe Manana */ 760054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 761054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 7623fd0a558SYan, Zheng } else { 7633fd0a558SYan, Zheng /* 7643fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 7653fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 7663fd0a558SYan, Zheng * modified after it was created have RELOC flag 7673fd0a558SYan, Zheng * set in their headers. so it's OK to not update 7683fd0a558SYan, Zheng * the 'last_snapshot'. 7693fd0a558SYan, Zheng */ 7703fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 7713fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 7723fd0a558SYan, Zheng BUG_ON(ret); 7733fd0a558SYan, Zheng } 7743fd0a558SYan, Zheng 7755d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 7765d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 7775d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 7785d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 7793fd0a558SYan, Zheng 7803fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 7813fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 7823fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 7833fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 784c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, 0); 7853fd0a558SYan, Zheng } 7865d4f98a2SYan Zheng 7875d4f98a2SYan Zheng btrfs_tree_unlock(eb); 7885d4f98a2SYan Zheng free_extent_buffer(eb); 7895d4f98a2SYan Zheng 7900b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 7915d4f98a2SYan Zheng &root_key, root_item); 7925d4f98a2SYan Zheng BUG_ON(ret); 7935d4f98a2SYan Zheng kfree(root_item); 7945d4f98a2SYan Zheng 7953dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 7965d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 79792a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 7985d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 7993fd0a558SYan, Zheng return reloc_root; 8003fd0a558SYan, Zheng } 8013fd0a558SYan, Zheng 8023fd0a558SYan, Zheng /* 8033fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 8043fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 805f44deb74SJosef Bacik * 806f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 807f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 8083fd0a558SYan, Zheng */ 8093fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 8103fd0a558SYan, Zheng struct btrfs_root *root) 8113fd0a558SYan, Zheng { 8120b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 8133fd0a558SYan, Zheng struct btrfs_root *reloc_root; 8140b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 81520dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 8163fd0a558SYan, Zheng int clear_rsv = 0; 817ffd7b339SJeff Mahoney int ret; 8183fd0a558SYan, Zheng 819aec7db3bSJosef Bacik if (!rc) 8202abc726aSJosef Bacik return 0; 8212abc726aSJosef Bacik 8221fac4a54SQu Wenruo /* 8231fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 8241fac4a54SQu Wenruo * create/update the dead reloc tree 8251fac4a54SQu Wenruo */ 8266282675eSQu Wenruo if (reloc_root_is_dead(root)) 8271fac4a54SQu Wenruo return 0; 8281fac4a54SQu Wenruo 829aec7db3bSJosef Bacik /* 830aec7db3bSJosef Bacik * This is subtle but important. We do not do 831aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 832aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 833aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 834aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 835aec7db3bSJosef Bacik * in. 836aec7db3bSJosef Bacik */ 8373fd0a558SYan, Zheng if (root->reloc_root) { 8383fd0a558SYan, Zheng reloc_root = root->reloc_root; 8393fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 8403fd0a558SYan, Zheng return 0; 8413fd0a558SYan, Zheng } 8423fd0a558SYan, Zheng 843aec7db3bSJosef Bacik /* 844aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 845aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 846aec7db3bSJosef Bacik */ 847aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 848aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 849aec7db3bSJosef Bacik return 0; 850aec7db3bSJosef Bacik 85120dd2cbfSMiao Xie if (!trans->reloc_reserved) { 85220dd2cbfSMiao Xie rsv = trans->block_rsv; 8533fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 8543fd0a558SYan, Zheng clear_rsv = 1; 8553fd0a558SYan, Zheng } 8563fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 8573fd0a558SYan, Zheng if (clear_rsv) 85820dd2cbfSMiao Xie trans->block_rsv = rsv; 8595d4f98a2SYan Zheng 860ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 861ffd7b339SJeff Mahoney BUG_ON(ret < 0); 862f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 8635d4f98a2SYan Zheng return 0; 8645d4f98a2SYan Zheng } 8655d4f98a2SYan Zheng 8665d4f98a2SYan Zheng /* 8675d4f98a2SYan Zheng * update root item of reloc tree 8685d4f98a2SYan Zheng */ 8695d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 8705d4f98a2SYan Zheng struct btrfs_root *root) 8715d4f98a2SYan Zheng { 8720b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 8735d4f98a2SYan Zheng struct btrfs_root *reloc_root; 8745d4f98a2SYan Zheng struct btrfs_root_item *root_item; 8755d4f98a2SYan Zheng int ret; 8765d4f98a2SYan Zheng 8776282675eSQu Wenruo if (!have_reloc_root(root)) 8787585717fSChris Mason goto out; 8795d4f98a2SYan Zheng 8805d4f98a2SYan Zheng reloc_root = root->reloc_root; 8815d4f98a2SYan Zheng root_item = &reloc_root->root_item; 8825d4f98a2SYan Zheng 883f44deb74SJosef Bacik /* 884f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 885f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 886f44deb74SJosef Bacik * hold it while we update the reloc root. 887f44deb74SJosef Bacik */ 888f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 889f44deb74SJosef Bacik 890d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 8910b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 8923fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 893d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 8946282675eSQu Wenruo /* 8956282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 8966282675eSQu Wenruo * have_reloc_root will not touch it from now on. 8976282675eSQu Wenruo */ 8986282675eSQu Wenruo smp_wmb(); 899c974c464SWang Shilong __del_reloc_root(reloc_root); 9005d4f98a2SYan Zheng } 9015d4f98a2SYan Zheng 9025d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 903ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 9045d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 9055d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 9065d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 9075d4f98a2SYan Zheng } 9085d4f98a2SYan Zheng 9090b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 9105d4f98a2SYan Zheng &reloc_root->root_key, root_item); 9115d4f98a2SYan Zheng BUG_ON(ret); 912f44deb74SJosef Bacik btrfs_put_root(reloc_root); 9137585717fSChris Mason out: 9145d4f98a2SYan Zheng return 0; 9155d4f98a2SYan Zheng } 9165d4f98a2SYan Zheng 9175d4f98a2SYan Zheng /* 9185d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 9195d4f98a2SYan Zheng * in a subvolume 9205d4f98a2SYan Zheng */ 9215d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 9225d4f98a2SYan Zheng { 9235d4f98a2SYan Zheng struct rb_node *node; 9245d4f98a2SYan Zheng struct rb_node *prev; 9255d4f98a2SYan Zheng struct btrfs_inode *entry; 9265d4f98a2SYan Zheng struct inode *inode; 9275d4f98a2SYan Zheng 9285d4f98a2SYan Zheng spin_lock(&root->inode_lock); 9295d4f98a2SYan Zheng again: 9305d4f98a2SYan Zheng node = root->inode_tree.rb_node; 9315d4f98a2SYan Zheng prev = NULL; 9325d4f98a2SYan Zheng while (node) { 9335d4f98a2SYan Zheng prev = node; 9345d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9355d4f98a2SYan Zheng 9364a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 9375d4f98a2SYan Zheng node = node->rb_left; 9384a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 9395d4f98a2SYan Zheng node = node->rb_right; 9405d4f98a2SYan Zheng else 9415d4f98a2SYan Zheng break; 9425d4f98a2SYan Zheng } 9435d4f98a2SYan Zheng if (!node) { 9445d4f98a2SYan Zheng while (prev) { 9455d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 9464a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 9475d4f98a2SYan Zheng node = prev; 9485d4f98a2SYan Zheng break; 9495d4f98a2SYan Zheng } 9505d4f98a2SYan Zheng prev = rb_next(prev); 9515d4f98a2SYan Zheng } 9525d4f98a2SYan Zheng } 9535d4f98a2SYan Zheng while (node) { 9545d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9555d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 9565d4f98a2SYan Zheng if (inode) { 9575d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9585d4f98a2SYan Zheng return inode; 9595d4f98a2SYan Zheng } 9605d4f98a2SYan Zheng 9614a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 9625d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 9635d4f98a2SYan Zheng goto again; 9645d4f98a2SYan Zheng 9655d4f98a2SYan Zheng node = rb_next(node); 9665d4f98a2SYan Zheng } 9675d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9685d4f98a2SYan Zheng return NULL; 9695d4f98a2SYan Zheng } 9705d4f98a2SYan Zheng 9715d4f98a2SYan Zheng /* 9725d4f98a2SYan Zheng * get new location of data 9735d4f98a2SYan Zheng */ 9745d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 9755d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 9765d4f98a2SYan Zheng { 9775d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 9785d4f98a2SYan Zheng struct btrfs_path *path; 9795d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 9805d4f98a2SYan Zheng struct extent_buffer *leaf; 9815d4f98a2SYan Zheng int ret; 9825d4f98a2SYan Zheng 9835d4f98a2SYan Zheng path = btrfs_alloc_path(); 9845d4f98a2SYan Zheng if (!path) 9855d4f98a2SYan Zheng return -ENOMEM; 9865d4f98a2SYan Zheng 9875d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 988f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 989f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 9905d4f98a2SYan Zheng if (ret < 0) 9915d4f98a2SYan Zheng goto out; 9925d4f98a2SYan Zheng if (ret > 0) { 9935d4f98a2SYan Zheng ret = -ENOENT; 9945d4f98a2SYan Zheng goto out; 9955d4f98a2SYan Zheng } 9965d4f98a2SYan Zheng 9975d4f98a2SYan Zheng leaf = path->nodes[0]; 9985d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 9995d4f98a2SYan Zheng struct btrfs_file_extent_item); 10005d4f98a2SYan Zheng 10015d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 10025d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 10035d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 10045d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 10055d4f98a2SYan Zheng 10065d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 100783d4cfd4SJosef Bacik ret = -EINVAL; 10085d4f98a2SYan Zheng goto out; 10095d4f98a2SYan Zheng } 10105d4f98a2SYan Zheng 10115d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10125d4f98a2SYan Zheng ret = 0; 10135d4f98a2SYan Zheng out: 10145d4f98a2SYan Zheng btrfs_free_path(path); 10155d4f98a2SYan Zheng return ret; 10165d4f98a2SYan Zheng } 10175d4f98a2SYan Zheng 10185d4f98a2SYan Zheng /* 10195d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 10205d4f98a2SYan Zheng * the new locations. 10215d4f98a2SYan Zheng */ 10223fd0a558SYan, Zheng static noinline_for_stack 10233fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 10245d4f98a2SYan Zheng struct reloc_control *rc, 10255d4f98a2SYan Zheng struct btrfs_root *root, 10263fd0a558SYan, Zheng struct extent_buffer *leaf) 10275d4f98a2SYan Zheng { 10280b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 10295d4f98a2SYan Zheng struct btrfs_key key; 10305d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10315d4f98a2SYan Zheng struct inode *inode = NULL; 10325d4f98a2SYan Zheng u64 parent; 10335d4f98a2SYan Zheng u64 bytenr; 10343fd0a558SYan, Zheng u64 new_bytenr = 0; 10355d4f98a2SYan Zheng u64 num_bytes; 10365d4f98a2SYan Zheng u64 end; 10375d4f98a2SYan Zheng u32 nritems; 10385d4f98a2SYan Zheng u32 i; 103983d4cfd4SJosef Bacik int ret = 0; 10405d4f98a2SYan Zheng int first = 1; 10415d4f98a2SYan Zheng int dirty = 0; 10425d4f98a2SYan Zheng 10435d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 10445d4f98a2SYan Zheng return 0; 10455d4f98a2SYan Zheng 10465d4f98a2SYan Zheng /* reloc trees always use full backref */ 10475d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 10485d4f98a2SYan Zheng parent = leaf->start; 10495d4f98a2SYan Zheng else 10505d4f98a2SYan Zheng parent = 0; 10515d4f98a2SYan Zheng 10525d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 10535d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 105482fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 105582fa113fSQu Wenruo 10565d4f98a2SYan Zheng cond_resched(); 10575d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 10585d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 10595d4f98a2SYan Zheng continue; 10605d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 10615d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 10625d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 10635d4f98a2SYan Zheng continue; 10645d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10655d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 10665d4f98a2SYan Zheng if (bytenr == 0) 10675d4f98a2SYan Zheng continue; 10689569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 10699569cc20SQu Wenruo rc->block_group->length)) 10705d4f98a2SYan Zheng continue; 10715d4f98a2SYan Zheng 10725d4f98a2SYan Zheng /* 10735d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 10745d4f98a2SYan Zheng * to complete and drop the extent cache 10755d4f98a2SYan Zheng */ 10765d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 10775d4f98a2SYan Zheng if (first) { 10785d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 10795d4f98a2SYan Zheng first = 0; 10804a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 10813fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 10825d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 10835d4f98a2SYan Zheng } 10844a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 10855d4f98a2SYan Zheng end = key.offset + 10865d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 10875d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 10880b246afaSJeff Mahoney fs_info->sectorsize)); 10890b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 10905d4f98a2SYan Zheng end--; 10915d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1092d0082371SJeff Mahoney key.offset, end); 10935d4f98a2SYan Zheng if (!ret) 10945d4f98a2SYan Zheng continue; 10955d4f98a2SYan Zheng 1096dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1097dcdbc059SNikolay Borisov key.offset, end, 1); 10985d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1099d0082371SJeff Mahoney key.offset, end); 11005d4f98a2SYan Zheng } 11015d4f98a2SYan Zheng } 11025d4f98a2SYan Zheng 11035d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 11045d4f98a2SYan Zheng bytenr, num_bytes); 110583d4cfd4SJosef Bacik if (ret) { 110683d4cfd4SJosef Bacik /* 110783d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 110883d4cfd4SJosef Bacik * in the file extent yet. 110983d4cfd4SJosef Bacik */ 111083d4cfd4SJosef Bacik break; 11113fd0a558SYan, Zheng } 11125d4f98a2SYan Zheng 11135d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 11145d4f98a2SYan Zheng dirty = 1; 11155d4f98a2SYan Zheng 11165d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 111782fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 111882fa113fSQu Wenruo num_bytes, parent); 111982fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 112082fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1121b06c4bf5SFilipe Manana key.objectid, key.offset); 112282fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 112383d4cfd4SJosef Bacik if (ret) { 112466642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 112583d4cfd4SJosef Bacik break; 112683d4cfd4SJosef Bacik } 11275d4f98a2SYan Zheng 1128ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1129ffd4bb2aSQu Wenruo num_bytes, parent); 1130ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1131ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1132b06c4bf5SFilipe Manana key.objectid, key.offset); 1133ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 113483d4cfd4SJosef Bacik if (ret) { 113566642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 113683d4cfd4SJosef Bacik break; 113783d4cfd4SJosef Bacik } 11385d4f98a2SYan Zheng } 11395d4f98a2SYan Zheng if (dirty) 11405d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 11413fd0a558SYan, Zheng if (inode) 11423fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 114383d4cfd4SJosef Bacik return ret; 11445d4f98a2SYan Zheng } 11455d4f98a2SYan Zheng 11465d4f98a2SYan Zheng static noinline_for_stack 11475d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 11485d4f98a2SYan Zheng struct btrfs_path *path, int level) 11495d4f98a2SYan Zheng { 11505d4f98a2SYan Zheng struct btrfs_disk_key key1; 11515d4f98a2SYan Zheng struct btrfs_disk_key key2; 11525d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 11535d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 11545d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 11555d4f98a2SYan Zheng } 11565d4f98a2SYan Zheng 11575d4f98a2SYan Zheng /* 11585d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 11595d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 11605d4f98a2SYan Zheng * reloc tree was create can be replaced. 11615d4f98a2SYan Zheng * 11625d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 11635d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 11645d4f98a2SYan Zheng * errors, a negative error number is returned. 11655d4f98a2SYan Zheng */ 11663fd0a558SYan, Zheng static noinline_for_stack 11673d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 11685d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 11695d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 11705d4f98a2SYan Zheng int lowest_level, int max_level) 11715d4f98a2SYan Zheng { 11720b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 11735d4f98a2SYan Zheng struct extent_buffer *eb; 11745d4f98a2SYan Zheng struct extent_buffer *parent; 117582fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 11765d4f98a2SYan Zheng struct btrfs_key key; 11775d4f98a2SYan Zheng u64 old_bytenr; 11785d4f98a2SYan Zheng u64 new_bytenr; 11795d4f98a2SYan Zheng u64 old_ptr_gen; 11805d4f98a2SYan Zheng u64 new_ptr_gen; 11815d4f98a2SYan Zheng u64 last_snapshot; 11825d4f98a2SYan Zheng u32 blocksize; 11833fd0a558SYan, Zheng int cow = 0; 11845d4f98a2SYan Zheng int level; 11855d4f98a2SYan Zheng int ret; 11865d4f98a2SYan Zheng int slot; 11875d4f98a2SYan Zheng 11885d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 11895d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 11905d4f98a2SYan Zheng 11915d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 11923fd0a558SYan, Zheng again: 11935d4f98a2SYan Zheng slot = path->slots[lowest_level]; 11945d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 11955d4f98a2SYan Zheng 11965d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 11975d4f98a2SYan Zheng level = btrfs_header_level(eb); 11985d4f98a2SYan Zheng 11995d4f98a2SYan Zheng if (level < lowest_level) { 12005d4f98a2SYan Zheng btrfs_tree_unlock(eb); 12015d4f98a2SYan Zheng free_extent_buffer(eb); 12025d4f98a2SYan Zheng return 0; 12035d4f98a2SYan Zheng } 12045d4f98a2SYan Zheng 12053fd0a558SYan, Zheng if (cow) { 12069631e4ccSJosef Bacik ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb, 12079631e4ccSJosef Bacik BTRFS_NESTING_COW); 12085d4f98a2SYan Zheng BUG_ON(ret); 12093fd0a558SYan, Zheng } 12105d4f98a2SYan Zheng 12115d4f98a2SYan Zheng if (next_key) { 12125d4f98a2SYan Zheng next_key->objectid = (u64)-1; 12135d4f98a2SYan Zheng next_key->type = (u8)-1; 12145d4f98a2SYan Zheng next_key->offset = (u64)-1; 12155d4f98a2SYan Zheng } 12165d4f98a2SYan Zheng 12175d4f98a2SYan Zheng parent = eb; 12185d4f98a2SYan Zheng while (1) { 12195d4f98a2SYan Zheng level = btrfs_header_level(parent); 12205d4f98a2SYan Zheng BUG_ON(level < lowest_level); 12215d4f98a2SYan Zheng 1222e3b83361SQu Wenruo ret = btrfs_bin_search(parent, &key, &slot); 1223cbca7d59SFilipe Manana if (ret < 0) 1224cbca7d59SFilipe Manana break; 12255d4f98a2SYan Zheng if (ret && slot > 0) 12265d4f98a2SYan Zheng slot--; 12275d4f98a2SYan Zheng 12285d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 12295d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 12305d4f98a2SYan Zheng 12315d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 12320b246afaSJeff Mahoney blocksize = fs_info->nodesize; 12335d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 12345d4f98a2SYan Zheng 12355d4f98a2SYan Zheng if (level <= max_level) { 12365d4f98a2SYan Zheng eb = path->nodes[level]; 12375d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 12385d4f98a2SYan Zheng path->slots[level]); 12395d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 12405d4f98a2SYan Zheng path->slots[level]); 12415d4f98a2SYan Zheng } else { 12425d4f98a2SYan Zheng new_bytenr = 0; 12435d4f98a2SYan Zheng new_ptr_gen = 0; 12445d4f98a2SYan Zheng } 12455d4f98a2SYan Zheng 1246fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 12475d4f98a2SYan Zheng ret = level; 12485d4f98a2SYan Zheng break; 12495d4f98a2SYan Zheng } 12505d4f98a2SYan Zheng 12515d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 12525d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 12533fd0a558SYan, Zheng if (level <= lowest_level) { 12545d4f98a2SYan Zheng ret = 0; 12555d4f98a2SYan Zheng break; 12565d4f98a2SYan Zheng } 12575d4f98a2SYan Zheng 12586b3426beSJosef Bacik eb = btrfs_read_node_slot(parent, slot); 125964c043deSLiu Bo if (IS_ERR(eb)) { 126064c043deSLiu Bo ret = PTR_ERR(eb); 1261264813acSLiu Bo break; 1262416bc658SJosef Bacik } 12635d4f98a2SYan Zheng btrfs_tree_lock(eb); 12643fd0a558SYan, Zheng if (cow) { 12655d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 12669631e4ccSJosef Bacik slot, &eb, 12679631e4ccSJosef Bacik BTRFS_NESTING_COW); 12685d4f98a2SYan Zheng BUG_ON(ret); 12695d4f98a2SYan Zheng } 12705d4f98a2SYan Zheng 12715d4f98a2SYan Zheng btrfs_tree_unlock(parent); 12725d4f98a2SYan Zheng free_extent_buffer(parent); 12735d4f98a2SYan Zheng 12745d4f98a2SYan Zheng parent = eb; 12755d4f98a2SYan Zheng continue; 12765d4f98a2SYan Zheng } 12775d4f98a2SYan Zheng 12783fd0a558SYan, Zheng if (!cow) { 12793fd0a558SYan, Zheng btrfs_tree_unlock(parent); 12803fd0a558SYan, Zheng free_extent_buffer(parent); 12813fd0a558SYan, Zheng cow = 1; 12823fd0a558SYan, Zheng goto again; 12833fd0a558SYan, Zheng } 12843fd0a558SYan, Zheng 12855d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 12865d4f98a2SYan Zheng path->slots[level]); 1287b3b4aa74SDavid Sterba btrfs_release_path(path); 12885d4f98a2SYan Zheng 12895d4f98a2SYan Zheng path->lowest_level = level; 12905d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 12915d4f98a2SYan Zheng path->lowest_level = 0; 12925d4f98a2SYan Zheng BUG_ON(ret); 12935d4f98a2SYan Zheng 12945d4f98a2SYan Zheng /* 1295824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 1296824d8dffSQu Wenruo * 1297824d8dffSQu Wenruo * We must trace both trees. 1298824d8dffSQu Wenruo * 1) Tree reloc subtree 1299824d8dffSQu Wenruo * If not traced, we will leak data numbers 1300824d8dffSQu Wenruo * 2) Fs subtree 1301824d8dffSQu Wenruo * If not traced, we will double count old data 1302f616f5cdSQu Wenruo * 1303f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 1304f616f5cdSQu Wenruo * the swapped tree blocks. 1305f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 1306f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 1307824d8dffSQu Wenruo */ 1308370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 1309370a11b8SQu Wenruo rc->block_group, parent, slot, 1310370a11b8SQu Wenruo path->nodes[level], path->slots[level], 1311370a11b8SQu Wenruo last_snapshot); 1312370a11b8SQu Wenruo if (ret < 0) 1313370a11b8SQu Wenruo break; 1314824d8dffSQu Wenruo /* 13155d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 13165d4f98a2SYan Zheng */ 13175d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 13185d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 13195d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 13205d4f98a2SYan Zheng 13215d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 13225d4f98a2SYan Zheng path->slots[level], old_bytenr); 13235d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 13245d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 13255d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 13265d4f98a2SYan Zheng 132782fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 132882fa113fSQu Wenruo blocksize, path->nodes[level]->start); 132982fa113fSQu Wenruo ref.skip_qgroup = true; 133082fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 133182fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13325d4f98a2SYan Zheng BUG_ON(ret); 133382fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 133482fa113fSQu Wenruo blocksize, 0); 133582fa113fSQu Wenruo ref.skip_qgroup = true; 133682fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 133782fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 13385d4f98a2SYan Zheng BUG_ON(ret); 13395d4f98a2SYan Zheng 1340ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 1341ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 1342ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 1343ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1344ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13455d4f98a2SYan Zheng BUG_ON(ret); 13465d4f98a2SYan Zheng 1347ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 1348ffd4bb2aSQu Wenruo blocksize, 0); 1349ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 1350ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1351ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 13525d4f98a2SYan Zheng BUG_ON(ret); 13535d4f98a2SYan Zheng 13545d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 13555d4f98a2SYan Zheng 13565d4f98a2SYan Zheng ret = level; 13575d4f98a2SYan Zheng break; 13585d4f98a2SYan Zheng } 13595d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13605d4f98a2SYan Zheng free_extent_buffer(parent); 13615d4f98a2SYan Zheng return ret; 13625d4f98a2SYan Zheng } 13635d4f98a2SYan Zheng 13645d4f98a2SYan Zheng /* 13655d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 13665d4f98a2SYan Zheng */ 13675d4f98a2SYan Zheng static noinline_for_stack 13685d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 13695d4f98a2SYan Zheng int *level) 13705d4f98a2SYan Zheng { 13715d4f98a2SYan Zheng struct extent_buffer *eb; 13725d4f98a2SYan Zheng int i; 13735d4f98a2SYan Zheng u64 last_snapshot; 13745d4f98a2SYan Zheng u32 nritems; 13755d4f98a2SYan Zheng 13765d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 13775d4f98a2SYan Zheng 13785d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 13795d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 13805d4f98a2SYan Zheng path->nodes[i] = NULL; 13815d4f98a2SYan Zheng } 13825d4f98a2SYan Zheng 13835d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 13845d4f98a2SYan Zheng eb = path->nodes[i]; 13855d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 13865d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 13875d4f98a2SYan Zheng path->slots[i]++; 13885d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 13895d4f98a2SYan Zheng last_snapshot) 13905d4f98a2SYan Zheng continue; 13915d4f98a2SYan Zheng 13925d4f98a2SYan Zheng *level = i; 13935d4f98a2SYan Zheng return 0; 13945d4f98a2SYan Zheng } 13955d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 13965d4f98a2SYan Zheng path->nodes[i] = NULL; 13975d4f98a2SYan Zheng } 13985d4f98a2SYan Zheng return 1; 13995d4f98a2SYan Zheng } 14005d4f98a2SYan Zheng 14015d4f98a2SYan Zheng /* 14025d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 14035d4f98a2SYan Zheng */ 14045d4f98a2SYan Zheng static noinline_for_stack 14055d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14065d4f98a2SYan Zheng int *level) 14075d4f98a2SYan Zheng { 14085d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 14095d4f98a2SYan Zheng int i; 14105d4f98a2SYan Zheng u64 ptr_gen = 0; 14115d4f98a2SYan Zheng u64 last_snapshot; 14125d4f98a2SYan Zheng u32 nritems; 14135d4f98a2SYan Zheng 14145d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14155d4f98a2SYan Zheng 14165d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 14175d4f98a2SYan Zheng eb = path->nodes[i]; 14185d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14195d4f98a2SYan Zheng while (path->slots[i] < nritems) { 14205d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 14215d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 14225d4f98a2SYan Zheng break; 14235d4f98a2SYan Zheng path->slots[i]++; 14245d4f98a2SYan Zheng } 14255d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 14265d4f98a2SYan Zheng if (i == *level) 14275d4f98a2SYan Zheng break; 14285d4f98a2SYan Zheng *level = i + 1; 14295d4f98a2SYan Zheng return 0; 14305d4f98a2SYan Zheng } 14315d4f98a2SYan Zheng if (i == 1) { 14325d4f98a2SYan Zheng *level = i; 14335d4f98a2SYan Zheng return 0; 14345d4f98a2SYan Zheng } 14355d4f98a2SYan Zheng 14368ef385bbSJosef Bacik eb = btrfs_read_node_slot(eb, path->slots[i]); 14378ef385bbSJosef Bacik if (IS_ERR(eb)) 143864c043deSLiu Bo return PTR_ERR(eb); 14395d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 14405d4f98a2SYan Zheng path->nodes[i - 1] = eb; 14415d4f98a2SYan Zheng path->slots[i - 1] = 0; 14425d4f98a2SYan Zheng } 14435d4f98a2SYan Zheng return 1; 14445d4f98a2SYan Zheng } 14455d4f98a2SYan Zheng 14465d4f98a2SYan Zheng /* 14475d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 14485d4f98a2SYan Zheng * [min_key, max_key) 14495d4f98a2SYan Zheng */ 14505d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 14515d4f98a2SYan Zheng struct btrfs_key *min_key, 14525d4f98a2SYan Zheng struct btrfs_key *max_key) 14535d4f98a2SYan Zheng { 14540b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14555d4f98a2SYan Zheng struct inode *inode = NULL; 14565d4f98a2SYan Zheng u64 objectid; 14575d4f98a2SYan Zheng u64 start, end; 145833345d01SLi Zefan u64 ino; 14595d4f98a2SYan Zheng 14605d4f98a2SYan Zheng objectid = min_key->objectid; 14615d4f98a2SYan Zheng while (1) { 14625d4f98a2SYan Zheng cond_resched(); 14635d4f98a2SYan Zheng iput(inode); 14645d4f98a2SYan Zheng 14655d4f98a2SYan Zheng if (objectid > max_key->objectid) 14665d4f98a2SYan Zheng break; 14675d4f98a2SYan Zheng 14685d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 14695d4f98a2SYan Zheng if (!inode) 14705d4f98a2SYan Zheng break; 14714a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 14725d4f98a2SYan Zheng 147333345d01SLi Zefan if (ino > max_key->objectid) { 14745d4f98a2SYan Zheng iput(inode); 14755d4f98a2SYan Zheng break; 14765d4f98a2SYan Zheng } 14775d4f98a2SYan Zheng 147833345d01SLi Zefan objectid = ino + 1; 14795d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 14805d4f98a2SYan Zheng continue; 14815d4f98a2SYan Zheng 148233345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 14835d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 14845d4f98a2SYan Zheng continue; 14855d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 14865d4f98a2SYan Zheng start = 0; 14875d4f98a2SYan Zheng else { 14885d4f98a2SYan Zheng start = min_key->offset; 14890b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 14905d4f98a2SYan Zheng } 14915d4f98a2SYan Zheng } else { 14925d4f98a2SYan Zheng start = 0; 14935d4f98a2SYan Zheng } 14945d4f98a2SYan Zheng 149533345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 14965d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 14975d4f98a2SYan Zheng continue; 14985d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 14995d4f98a2SYan Zheng end = (u64)-1; 15005d4f98a2SYan Zheng } else { 15015d4f98a2SYan Zheng if (max_key->offset == 0) 15025d4f98a2SYan Zheng continue; 15035d4f98a2SYan Zheng end = max_key->offset; 15040b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 15055d4f98a2SYan Zheng end--; 15065d4f98a2SYan Zheng } 15075d4f98a2SYan Zheng } else { 15085d4f98a2SYan Zheng end = (u64)-1; 15095d4f98a2SYan Zheng } 15105d4f98a2SYan Zheng 15115d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 1512d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 1513dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 1514d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 15155d4f98a2SYan Zheng } 15165d4f98a2SYan Zheng return 0; 15175d4f98a2SYan Zheng } 15185d4f98a2SYan Zheng 15195d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 15205d4f98a2SYan Zheng struct btrfs_key *key) 15215d4f98a2SYan Zheng 15225d4f98a2SYan Zheng { 15235d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 15245d4f98a2SYan Zheng if (!path->nodes[level]) 15255d4f98a2SYan Zheng break; 15265d4f98a2SYan Zheng if (path->slots[level] + 1 < 15275d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 15285d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 15295d4f98a2SYan Zheng path->slots[level] + 1); 15305d4f98a2SYan Zheng return 0; 15315d4f98a2SYan Zheng } 15325d4f98a2SYan Zheng level++; 15335d4f98a2SYan Zheng } 15345d4f98a2SYan Zheng return 1; 15355d4f98a2SYan Zheng } 15365d4f98a2SYan Zheng 15375d4f98a2SYan Zheng /* 1538d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 1539d2311e69SQu Wenruo */ 1540d2311e69SQu Wenruo static void insert_dirty_subvol(struct btrfs_trans_handle *trans, 1541d2311e69SQu Wenruo struct reloc_control *rc, 1542d2311e69SQu Wenruo struct btrfs_root *root) 1543d2311e69SQu Wenruo { 1544d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1545d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 1546d2311e69SQu Wenruo 1547d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 1548d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 1549d2311e69SQu Wenruo ASSERT(reloc_root); 1550d2311e69SQu Wenruo 1551d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 1552d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 1553d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 1554c8422684SDavid Sterba btrfs_set_root_drop_level(reloc_root_item, 0); 1555d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 1556d2311e69SQu Wenruo btrfs_update_reloc_root(trans, root); 1557d2311e69SQu Wenruo 1558d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 155900246528SJosef Bacik btrfs_grab_root(root); 1560d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 1561d2311e69SQu Wenruo } 1562d2311e69SQu Wenruo } 1563d2311e69SQu Wenruo 1564d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 1565d2311e69SQu Wenruo { 1566d2311e69SQu Wenruo struct btrfs_root *root; 1567d2311e69SQu Wenruo struct btrfs_root *next; 1568d2311e69SQu Wenruo int ret = 0; 156930d40577SQu Wenruo int ret2; 1570d2311e69SQu Wenruo 1571d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 1572d2311e69SQu Wenruo reloc_dirty_list) { 157330d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 157430d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 1575d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1576d2311e69SQu Wenruo 1577d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 1578d2311e69SQu Wenruo root->reloc_root = NULL; 15796282675eSQu Wenruo /* 15806282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 15816282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 15826282675eSQu Wenruo */ 15836282675eSQu Wenruo smp_wmb(); 15841fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 1585f28de8d8SJosef Bacik if (reloc_root) { 1586f44deb74SJosef Bacik /* 1587f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 1588f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 1589f44deb74SJosef Bacik * drop the ref ourselves. 1590f44deb74SJosef Bacik */ 1591f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 1592f44deb74SJosef Bacik if (ret2 < 0) { 1593f44deb74SJosef Bacik btrfs_put_root(reloc_root); 1594f44deb74SJosef Bacik if (!ret) 1595f28de8d8SJosef Bacik ret = ret2; 1596f28de8d8SJosef Bacik } 1597f44deb74SJosef Bacik } 159800246528SJosef Bacik btrfs_put_root(root); 159930d40577SQu Wenruo } else { 160030d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 16010078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 1602f44deb74SJosef Bacik if (ret2 < 0) { 1603f44deb74SJosef Bacik btrfs_put_root(root); 1604f44deb74SJosef Bacik if (!ret) 160530d40577SQu Wenruo ret = ret2; 160630d40577SQu Wenruo } 1607d2311e69SQu Wenruo } 1608f44deb74SJosef Bacik } 1609d2311e69SQu Wenruo return ret; 1610d2311e69SQu Wenruo } 1611d2311e69SQu Wenruo 1612d2311e69SQu Wenruo /* 16135d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 16145d4f98a2SYan Zheng * fs tree. 16155d4f98a2SYan Zheng */ 16165d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 16175d4f98a2SYan Zheng struct btrfs_root *root) 16185d4f98a2SYan Zheng { 16190b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 16205d4f98a2SYan Zheng struct btrfs_key key; 16215d4f98a2SYan Zheng struct btrfs_key next_key; 16229e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 16235d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16245d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16255d4f98a2SYan Zheng struct btrfs_path *path; 16263fd0a558SYan, Zheng struct extent_buffer *leaf; 1627fca3a45dSJosef Bacik int reserve_level; 16285d4f98a2SYan Zheng int level; 16295d4f98a2SYan Zheng int max_level; 16305d4f98a2SYan Zheng int replaced = 0; 1631c6a592f2SNikolay Borisov int ret = 0; 16323fd0a558SYan, Zheng u32 min_reserved; 16335d4f98a2SYan Zheng 16345d4f98a2SYan Zheng path = btrfs_alloc_path(); 16355d4f98a2SYan Zheng if (!path) 16365d4f98a2SYan Zheng return -ENOMEM; 1637e4058b54SDavid Sterba path->reada = READA_FORWARD; 16385d4f98a2SYan Zheng 16395d4f98a2SYan Zheng reloc_root = root->reloc_root; 16405d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16415d4f98a2SYan Zheng 16425d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 16435d4f98a2SYan Zheng level = btrfs_root_level(root_item); 164467439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 16455d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 16465d4f98a2SYan Zheng path->slots[level] = 0; 16475d4f98a2SYan Zheng } else { 16485d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 16495d4f98a2SYan Zheng 1650c8422684SDavid Sterba level = btrfs_root_drop_level(root_item); 16515d4f98a2SYan Zheng BUG_ON(level == 0); 16525d4f98a2SYan Zheng path->lowest_level = level; 16535d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 165433c66f43SYan Zheng path->lowest_level = 0; 16555d4f98a2SYan Zheng if (ret < 0) { 16565d4f98a2SYan Zheng btrfs_free_path(path); 16575d4f98a2SYan Zheng return ret; 16585d4f98a2SYan Zheng } 16595d4f98a2SYan Zheng 16605d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 16615d4f98a2SYan Zheng path->slots[level]); 16625d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 16635d4f98a2SYan Zheng 16645d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 16655d4f98a2SYan Zheng } 16665d4f98a2SYan Zheng 166744d354abSQu Wenruo /* 166844d354abSQu Wenruo * In merge_reloc_root(), we modify the upper level pointer to swap the 166944d354abSQu Wenruo * tree blocks between reloc tree and subvolume tree. Thus for tree 167044d354abSQu Wenruo * block COW, we COW at most from level 1 to root level for each tree. 167144d354abSQu Wenruo * 167244d354abSQu Wenruo * Thus the needed metadata size is at most root_level * nodesize, 167344d354abSQu Wenruo * and * 2 since we have two trees to COW. 167444d354abSQu Wenruo */ 1675fca3a45dSJosef Bacik reserve_level = max_t(int, 1, btrfs_root_level(root_item)); 1676fca3a45dSJosef Bacik min_reserved = fs_info->nodesize * reserve_level * 2; 16775d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 16785d4f98a2SYan Zheng 16795d4f98a2SYan Zheng while (1) { 168008e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 168144d354abSQu Wenruo BTRFS_RESERVE_FLUSH_LIMIT); 1682c6a592f2SNikolay Borisov if (ret) 16839e6a0c52SJosef Bacik goto out; 16849e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 16859e6a0c52SJosef Bacik if (IS_ERR(trans)) { 1686c6a592f2SNikolay Borisov ret = PTR_ERR(trans); 16879e6a0c52SJosef Bacik trans = NULL; 16889e6a0c52SJosef Bacik goto out; 16899e6a0c52SJosef Bacik } 16902abc726aSJosef Bacik 16912abc726aSJosef Bacik /* 16922abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 16932abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 16942abc726aSJosef Bacik * 16952abc726aSJosef Bacik * But that's ok, we started the trans handle on our 16962abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 16972abc726aSJosef Bacik * dirty list. At commit time we'll still call 16982abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 16992abc726aSJosef Bacik * appropriately. 17002abc726aSJosef Bacik */ 17012abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 17029e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 17033fd0a558SYan, Zheng 17043fd0a558SYan, Zheng replaced = 0; 17055d4f98a2SYan Zheng max_level = level; 17065d4f98a2SYan Zheng 17075d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 1708c6a592f2SNikolay Borisov if (ret < 0) 17095d4f98a2SYan Zheng goto out; 17105d4f98a2SYan Zheng if (ret > 0) 17115d4f98a2SYan Zheng break; 17125d4f98a2SYan Zheng 17135d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 17145d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 17155d4f98a2SYan Zheng ret = 0; 17165d4f98a2SYan Zheng } else { 17173d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 17183fd0a558SYan, Zheng &next_key, level, max_level); 17195d4f98a2SYan Zheng } 1720c6a592f2SNikolay Borisov if (ret < 0) 17215d4f98a2SYan Zheng goto out; 17225d4f98a2SYan Zheng if (ret > 0) { 17235d4f98a2SYan Zheng level = ret; 17245d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17255d4f98a2SYan Zheng path->slots[level]); 17265d4f98a2SYan Zheng replaced = 1; 17275d4f98a2SYan Zheng } 17285d4f98a2SYan Zheng 17295d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 17305d4f98a2SYan Zheng if (ret > 0) 17315d4f98a2SYan Zheng break; 17325d4f98a2SYan Zheng 17335d4f98a2SYan Zheng BUG_ON(level == 0); 17345d4f98a2SYan Zheng /* 17355d4f98a2SYan Zheng * save the merging progress in the drop_progress. 17365d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 17375d4f98a2SYan Zheng */ 17385d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 17395d4f98a2SYan Zheng path->slots[level]); 1740c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, level); 17415d4f98a2SYan Zheng 17423a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 17439e6a0c52SJosef Bacik trans = NULL; 17445d4f98a2SYan Zheng 17452ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 17465d4f98a2SYan Zheng 17475d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17485d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17495d4f98a2SYan Zheng } 17505d4f98a2SYan Zheng 17515d4f98a2SYan Zheng /* 17525d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 17535d4f98a2SYan Zheng * relocated and the block is tree root. 17545d4f98a2SYan Zheng */ 17555d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 17569631e4ccSJosef Bacik ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf, 17579631e4ccSJosef Bacik BTRFS_NESTING_COW); 17585d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 17595d4f98a2SYan Zheng free_extent_buffer(leaf); 17605d4f98a2SYan Zheng out: 17615d4f98a2SYan Zheng btrfs_free_path(path); 17625d4f98a2SYan Zheng 1763c6a592f2SNikolay Borisov if (ret == 0) 1764d2311e69SQu Wenruo insert_dirty_subvol(trans, rc, root); 17655d4f98a2SYan Zheng 17669e6a0c52SJosef Bacik if (trans) 17673a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 17685d4f98a2SYan Zheng 17692ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 17705d4f98a2SYan Zheng 17715d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17725d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17735d4f98a2SYan Zheng 1774c6a592f2SNikolay Borisov return ret; 17755d4f98a2SYan Zheng } 17765d4f98a2SYan Zheng 17773fd0a558SYan, Zheng static noinline_for_stack 17783fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 17795d4f98a2SYan Zheng { 17803fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 17810b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 17823fd0a558SYan, Zheng struct btrfs_root *reloc_root; 17835d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 17843fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 17853fd0a558SYan, Zheng u64 num_bytes = 0; 17863fd0a558SYan, Zheng int ret; 17873fd0a558SYan, Zheng 17880b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 17890b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 17903fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 17910b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 17927585717fSChris Mason 17933fd0a558SYan, Zheng again: 17943fd0a558SYan, Zheng if (!err) { 17953fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 179608e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 179708e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 17983fd0a558SYan, Zheng if (ret) 17993fd0a558SYan, Zheng err = ret; 18003fd0a558SYan, Zheng } 18013fd0a558SYan, Zheng 18027a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 18033612b495STsutomu Itoh if (IS_ERR(trans)) { 18043612b495STsutomu Itoh if (!err) 18052ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 180663f018beSNikolay Borisov num_bytes, NULL); 18073612b495STsutomu Itoh return PTR_ERR(trans); 18083612b495STsutomu Itoh } 18093fd0a558SYan, Zheng 18103fd0a558SYan, Zheng if (!err) { 18113fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 18123a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18132ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 181463f018beSNikolay Borisov num_bytes, NULL); 18153fd0a558SYan, Zheng goto again; 18163fd0a558SYan, Zheng } 18173fd0a558SYan, Zheng } 18183fd0a558SYan, Zheng 18193fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 18203fd0a558SYan, Zheng 18213fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 18223fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 18233fd0a558SYan, Zheng struct btrfs_root, root_list); 18243fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 18253fd0a558SYan, Zheng 1826a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1827a820feb5SDavid Sterba false); 18283fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 18293fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 18303fd0a558SYan, Zheng 18313fd0a558SYan, Zheng /* 18323fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 18333fd0a558SYan, Zheng * knows it should resumes merging 18343fd0a558SYan, Zheng */ 18353fd0a558SYan, Zheng if (!err) 18363fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 18373fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 18383fd0a558SYan, Zheng 18393fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 184000246528SJosef Bacik btrfs_put_root(root); 18413fd0a558SYan, Zheng } 18423fd0a558SYan, Zheng 18433fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 18443fd0a558SYan, Zheng 18453fd0a558SYan, Zheng if (!err) 18463a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 18473fd0a558SYan, Zheng else 18483a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18493fd0a558SYan, Zheng return err; 18503fd0a558SYan, Zheng } 18513fd0a558SYan, Zheng 18523fd0a558SYan, Zheng static noinline_for_stack 1853aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 1854aca1bba6SLiu Bo { 1855a7571232SNikolay Borisov struct btrfs_root *reloc_root, *tmp; 1856aca1bba6SLiu Bo 1857a7571232SNikolay Borisov list_for_each_entry_safe(reloc_root, tmp, list, root_list) 1858bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 1859aca1bba6SLiu Bo } 1860aca1bba6SLiu Bo 1861aca1bba6SLiu Bo static noinline_for_stack 186294404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 18633fd0a558SYan, Zheng { 18640b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 18655d4f98a2SYan Zheng struct btrfs_root *root; 18665d4f98a2SYan Zheng struct btrfs_root *reloc_root; 18673fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 18683fd0a558SYan, Zheng int found = 0; 1869aca1bba6SLiu Bo int ret = 0; 18703fd0a558SYan, Zheng again: 18713fd0a558SYan, Zheng root = rc->extent_root; 18727585717fSChris Mason 18737585717fSChris Mason /* 18747585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 18757585717fSChris Mason * we have to make sure nobody is in the middle of 18767585717fSChris Mason * adding their roots to the list while we are 18777585717fSChris Mason * doing this splice 18787585717fSChris Mason */ 18790b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 18803fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 18810b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 18825d4f98a2SYan Zheng 18833fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 18843fd0a558SYan, Zheng found = 1; 18853fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 18863fd0a558SYan, Zheng struct btrfs_root, root_list); 18875d4f98a2SYan Zheng 1888a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1889a820feb5SDavid Sterba false); 18905d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 18915d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 18925d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 18933fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 189400246528SJosef Bacik btrfs_put_root(root); 1895b37b39cdSJosef Bacik if (ret) { 189625e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 189725e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 189825e293c2SWang Shilong &reloc_roots); 1899aca1bba6SLiu Bo goto out; 1900b37b39cdSJosef Bacik } 19013fd0a558SYan, Zheng } else { 190251415b6cSQu Wenruo if (!IS_ERR(root)) { 190351415b6cSQu Wenruo if (root->reloc_root == reloc_root) { 190451415b6cSQu Wenruo root->reloc_root = NULL; 190551415b6cSQu Wenruo btrfs_put_root(reloc_root); 190651415b6cSQu Wenruo } 19071dae7e0eSQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, 19081dae7e0eSQu Wenruo &root->state); 190951415b6cSQu Wenruo btrfs_put_root(root); 191051415b6cSQu Wenruo } 191151415b6cSQu Wenruo 19123fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 191330d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 191430d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 191530d40577SQu Wenruo &rc->dirty_subvol_roots); 19163fd0a558SYan, Zheng } 19175d4f98a2SYan Zheng } 19185d4f98a2SYan Zheng 19193fd0a558SYan, Zheng if (found) { 19203fd0a558SYan, Zheng found = 0; 19213fd0a558SYan, Zheng goto again; 19225d4f98a2SYan Zheng } 1923aca1bba6SLiu Bo out: 1924aca1bba6SLiu Bo if (ret) { 19250b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 1926aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 1927467bb1d2SWang Shilong 1928467bb1d2SWang Shilong /* new reloc root may be added */ 19290b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 1930467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 19310b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 1932467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 1933aca1bba6SLiu Bo } 1934aca1bba6SLiu Bo 19357b7b7431SJosef Bacik /* 19367b7b7431SJosef Bacik * We used to have 19377b7b7431SJosef Bacik * 19387b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 19397b7b7431SJosef Bacik * 19407b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 19417b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 19427b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 19437b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 19447b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 19457b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 19467b7b7431SJosef Bacik * will be cleaned up on unmount. 19477b7b7431SJosef Bacik * 19487b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 19497b7b7431SJosef Bacik */ 19505d4f98a2SYan Zheng } 19515d4f98a2SYan Zheng 19525d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 19535d4f98a2SYan Zheng { 19545d4f98a2SYan Zheng struct tree_block *block; 19555d4f98a2SYan Zheng struct rb_node *rb_node; 19565d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 19575d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 19585d4f98a2SYan Zheng rb_erase(rb_node, blocks); 19595d4f98a2SYan Zheng kfree(block); 19605d4f98a2SYan Zheng } 19615d4f98a2SYan Zheng } 19625d4f98a2SYan Zheng 19635d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 19645d4f98a2SYan Zheng struct btrfs_root *reloc_root) 19655d4f98a2SYan Zheng { 19660b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 19675d4f98a2SYan Zheng struct btrfs_root *root; 1968442b1ac5SJosef Bacik int ret; 19695d4f98a2SYan Zheng 19705d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 19715d4f98a2SYan Zheng return 0; 19725d4f98a2SYan Zheng 1973a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false); 19745d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 19755d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 1976442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 197700246528SJosef Bacik btrfs_put_root(root); 19785d4f98a2SYan Zheng 1979442b1ac5SJosef Bacik return ret; 19805d4f98a2SYan Zheng } 19815d4f98a2SYan Zheng 19823fd0a558SYan, Zheng static noinline_for_stack 19833fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 19843fd0a558SYan, Zheng struct reloc_control *rc, 1985a26195a5SQu Wenruo struct btrfs_backref_node *node, 1986a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 19875d4f98a2SYan Zheng { 1988a26195a5SQu Wenruo struct btrfs_backref_node *next; 19895d4f98a2SYan Zheng struct btrfs_root *root; 19903fd0a558SYan, Zheng int index = 0; 19913fd0a558SYan, Zheng 19925d4f98a2SYan Zheng next = node; 19935d4f98a2SYan Zheng while (1) { 19945d4f98a2SYan Zheng cond_resched(); 19955d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 19965d4f98a2SYan Zheng root = next->root; 19973fd0a558SYan, Zheng BUG_ON(!root); 199892a7cc42SQu Wenruo BUG_ON(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)); 19995d4f98a2SYan Zheng 20005d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 20015d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 20025d4f98a2SYan Zheng break; 20035d4f98a2SYan Zheng } 20045d4f98a2SYan Zheng 20055d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 20063fd0a558SYan, Zheng root = root->reloc_root; 20073fd0a558SYan, Zheng 20083fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 20093fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 20103fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 20113fd0a558SYan, Zheng next->new_bytenr = root->node->start; 201200246528SJosef Bacik btrfs_put_root(next->root); 201300246528SJosef Bacik next->root = btrfs_grab_root(root); 20140b530bc5SJosef Bacik ASSERT(next->root); 20153fd0a558SYan, Zheng list_add_tail(&next->list, 20163fd0a558SYan, Zheng &rc->backref_cache.changed); 20179569cc20SQu Wenruo mark_block_processed(rc, next); 20185d4f98a2SYan Zheng break; 20195d4f98a2SYan Zheng } 20205d4f98a2SYan Zheng 20213fd0a558SYan, Zheng WARN_ON(1); 20225d4f98a2SYan Zheng root = NULL; 20235d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 20245d4f98a2SYan Zheng if (!next || next->level <= node->level) 20255d4f98a2SYan Zheng break; 20265d4f98a2SYan Zheng } 20273fd0a558SYan, Zheng if (!root) 20283fd0a558SYan, Zheng return NULL; 20295d4f98a2SYan Zheng 20303fd0a558SYan, Zheng next = node; 20313fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 20323fd0a558SYan, Zheng while (1) { 20333fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 20343fd0a558SYan, Zheng if (--index < 0) 20353fd0a558SYan, Zheng break; 20363fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 20373fd0a558SYan, Zheng } 20385d4f98a2SYan Zheng return root; 20395d4f98a2SYan Zheng } 20405d4f98a2SYan Zheng 20413fd0a558SYan, Zheng /* 204292a7cc42SQu Wenruo * Select a tree root for relocation. 204392a7cc42SQu Wenruo * 204492a7cc42SQu Wenruo * Return NULL if the block is not shareable. We should use do_relocation() in 204592a7cc42SQu Wenruo * this case. 204692a7cc42SQu Wenruo * 204792a7cc42SQu Wenruo * Return a tree root pointer if the block is shareable. 204892a7cc42SQu Wenruo * Return -ENOENT if the block is root of reloc tree. 20493fd0a558SYan, Zheng */ 20505d4f98a2SYan Zheng static noinline_for_stack 2051a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 20525d4f98a2SYan Zheng { 2053a26195a5SQu Wenruo struct btrfs_backref_node *next; 20543fd0a558SYan, Zheng struct btrfs_root *root; 20553fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2056a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 20573fd0a558SYan, Zheng int index = 0; 20583fd0a558SYan, Zheng 20593fd0a558SYan, Zheng next = node; 20603fd0a558SYan, Zheng while (1) { 20613fd0a558SYan, Zheng cond_resched(); 20623fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 20633fd0a558SYan, Zheng root = next->root; 20643fd0a558SYan, Zheng BUG_ON(!root); 20653fd0a558SYan, Zheng 206692a7cc42SQu Wenruo /* No other choice for non-shareable tree */ 206792a7cc42SQu Wenruo if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 20683fd0a558SYan, Zheng return root; 20693fd0a558SYan, Zheng 20703fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 20713fd0a558SYan, Zheng fs_root = root; 20723fd0a558SYan, Zheng 20733fd0a558SYan, Zheng if (next != node) 20743fd0a558SYan, Zheng return NULL; 20753fd0a558SYan, Zheng 20763fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 20773fd0a558SYan, Zheng if (!next || next->level <= node->level) 20783fd0a558SYan, Zheng break; 20793fd0a558SYan, Zheng } 20803fd0a558SYan, Zheng 20813fd0a558SYan, Zheng if (!fs_root) 20823fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 20833fd0a558SYan, Zheng return fs_root; 20845d4f98a2SYan Zheng } 20855d4f98a2SYan Zheng 20865d4f98a2SYan Zheng static noinline_for_stack 20873fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2088a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 20895d4f98a2SYan Zheng { 20900b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2091a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2092a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2093a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 20943fd0a558SYan, Zheng u64 num_bytes = 0; 20953fd0a558SYan, Zheng int index = 0; 20965d4f98a2SYan Zheng 20973fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 20983fd0a558SYan, Zheng 20993fd0a558SYan, Zheng while (next) { 21003fd0a558SYan, Zheng cond_resched(); 21015d4f98a2SYan Zheng while (1) { 21023fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 21035d4f98a2SYan Zheng break; 21045d4f98a2SYan Zheng 21050b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 21063fd0a558SYan, Zheng 21073fd0a558SYan, Zheng if (list_empty(&next->upper)) 21083fd0a558SYan, Zheng break; 21093fd0a558SYan, Zheng 21103fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2111a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 21123fd0a558SYan, Zheng edges[index++] = edge; 21133fd0a558SYan, Zheng next = edge->node[UPPER]; 21145d4f98a2SYan Zheng } 21153fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 21163fd0a558SYan, Zheng } 21173fd0a558SYan, Zheng return num_bytes; 21183fd0a558SYan, Zheng } 21193fd0a558SYan, Zheng 21203fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 21213fd0a558SYan, Zheng struct reloc_control *rc, 2122a26195a5SQu Wenruo struct btrfs_backref_node *node) 21233fd0a558SYan, Zheng { 21243fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2125da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 21263fd0a558SYan, Zheng u64 num_bytes; 21273fd0a558SYan, Zheng int ret; 21280647bf56SWang Shilong u64 tmp; 21293fd0a558SYan, Zheng 21303fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 21313fd0a558SYan, Zheng 21323fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 21330647bf56SWang Shilong rc->reserved_bytes += num_bytes; 21348ca17f0fSJosef Bacik 21358ca17f0fSJosef Bacik /* 21368ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 21378ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 21388ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 21398ca17f0fSJosef Bacik */ 21400647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 21418ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 21423fd0a558SYan, Zheng if (ret) { 2143da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 21440647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 21450647bf56SWang Shilong tmp <<= 1; 21460647bf56SWang Shilong /* 21470647bf56SWang Shilong * only one thread can access block_rsv at this point, 21480647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 21490647bf56SWang Shilong * we expand more reservation size here to allow enough 215052042d8eSAndrea Gelmini * space for relocation and we will return earlier in 21510647bf56SWang Shilong * enospc case. 21520647bf56SWang Shilong */ 2153da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 21540647bf56SWang Shilong RELOCATION_RESERVED_NODES; 21558ca17f0fSJosef Bacik return -EAGAIN; 21563fd0a558SYan, Zheng } 21573fd0a558SYan, Zheng 21583fd0a558SYan, Zheng return 0; 21593fd0a558SYan, Zheng } 21603fd0a558SYan, Zheng 21615d4f98a2SYan Zheng /* 21625d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 21635d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 21645d4f98a2SYan Zheng * 21655d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 21665d4f98a2SYan Zheng * in that case this function just updates pointers. 21675d4f98a2SYan Zheng */ 21685d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 21693fd0a558SYan, Zheng struct reloc_control *rc, 2170a26195a5SQu Wenruo struct btrfs_backref_node *node, 21715d4f98a2SYan Zheng struct btrfs_key *key, 21725d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 21735d4f98a2SYan Zheng { 2174a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2175a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2176a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 21775d4f98a2SYan Zheng struct btrfs_root *root; 21785d4f98a2SYan Zheng struct extent_buffer *eb; 21795d4f98a2SYan Zheng u32 blocksize; 21805d4f98a2SYan Zheng u64 bytenr; 21815d4f98a2SYan Zheng int slot; 21828df01fddSNikolay Borisov int ret = 0; 21835d4f98a2SYan Zheng 21845d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 21855d4f98a2SYan Zheng 21865d4f98a2SYan Zheng path->lowest_level = node->level + 1; 21873fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 21885d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 218982fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2190581c1760SQu Wenruo 21915d4f98a2SYan Zheng cond_resched(); 21925d4f98a2SYan Zheng 21935d4f98a2SYan Zheng upper = edge->node[UPPER]; 2194dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 21953fd0a558SYan, Zheng BUG_ON(!root); 21965d4f98a2SYan Zheng 21973fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 21983fd0a558SYan, Zheng if (!lowest) { 2199e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 22008df01fddSNikolay Borisov if (ret < 0) 2201cbca7d59SFilipe Manana goto next; 22023fd0a558SYan, Zheng BUG_ON(ret); 22033fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 22043fd0a558SYan, Zheng if (node->eb->start == bytenr) 22053fd0a558SYan, Zheng goto next; 22063fd0a558SYan, Zheng } 2207b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 22083fd0a558SYan, Zheng } 22095d4f98a2SYan Zheng 22105d4f98a2SYan Zheng if (!upper->eb) { 22115d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 22123561b9dbSLiu Bo if (ret) { 22138df01fddSNikolay Borisov if (ret > 0) 22148df01fddSNikolay Borisov ret = -ENOENT; 22153561b9dbSLiu Bo 22163561b9dbSLiu Bo btrfs_release_path(path); 22175d4f98a2SYan Zheng break; 22185d4f98a2SYan Zheng } 22195d4f98a2SYan Zheng 22203fd0a558SYan, Zheng if (!upper->eb) { 22213fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 22223fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 22233fd0a558SYan, Zheng } else { 22243fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 22253fd0a558SYan, Zheng } 22263fd0a558SYan, Zheng 22273fd0a558SYan, Zheng upper->locked = 1; 22283fd0a558SYan, Zheng path->locks[upper->level] = 0; 22293fd0a558SYan, Zheng 22305d4f98a2SYan Zheng slot = path->slots[upper->level]; 2231b3b4aa74SDavid Sterba btrfs_release_path(path); 22325d4f98a2SYan Zheng } else { 2233e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 22348df01fddSNikolay Borisov if (ret < 0) 2235cbca7d59SFilipe Manana goto next; 22365d4f98a2SYan Zheng BUG_ON(ret); 22375d4f98a2SYan Zheng } 22385d4f98a2SYan Zheng 22395d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 22403fd0a558SYan, Zheng if (lowest) { 22414547f4d8SLiu Bo if (bytenr != node->bytenr) { 22424547f4d8SLiu Bo btrfs_err(root->fs_info, 22434547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 22444547f4d8SLiu Bo bytenr, node->bytenr, slot, 22454547f4d8SLiu Bo upper->eb->start); 22468df01fddSNikolay Borisov ret = -EIO; 22474547f4d8SLiu Bo goto next; 22484547f4d8SLiu Bo } 22495d4f98a2SYan Zheng } else { 22503fd0a558SYan, Zheng if (node->eb->start == bytenr) 22513fd0a558SYan, Zheng goto next; 22525d4f98a2SYan Zheng } 22535d4f98a2SYan Zheng 2254da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 2255c9752536SJosef Bacik eb = btrfs_read_node_slot(upper->eb, slot); 225664c043deSLiu Bo if (IS_ERR(eb)) { 22578df01fddSNikolay Borisov ret = PTR_ERR(eb); 225864c043deSLiu Bo goto next; 225997d9a8a4STsutomu Itoh } 22605d4f98a2SYan Zheng btrfs_tree_lock(eb); 22615d4f98a2SYan Zheng 22625d4f98a2SYan Zheng if (!node->eb) { 22635d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 22649631e4ccSJosef Bacik slot, &eb, BTRFS_NESTING_COW); 22653fd0a558SYan, Zheng btrfs_tree_unlock(eb); 22663fd0a558SYan, Zheng free_extent_buffer(eb); 22678df01fddSNikolay Borisov if (ret < 0) 22683fd0a558SYan, Zheng goto next; 22693fd0a558SYan, Zheng BUG_ON(node->eb != eb); 22705d4f98a2SYan Zheng } else { 22715d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 22725d4f98a2SYan Zheng node->eb->start); 22735d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 22745d4f98a2SYan Zheng trans->transid); 22755d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 22765d4f98a2SYan Zheng 227782fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 22785d4f98a2SYan Zheng node->eb->start, blocksize, 227982fa113fSQu Wenruo upper->eb->start); 228082fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 228182fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 228282fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 228382fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 22845d4f98a2SYan Zheng BUG_ON(ret); 22855d4f98a2SYan Zheng 22865d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 22875d4f98a2SYan Zheng BUG_ON(ret); 22885d4f98a2SYan Zheng } 22893fd0a558SYan, Zheng next: 22903fd0a558SYan, Zheng if (!upper->pending) 2291b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 22923fd0a558SYan, Zheng else 2293b0fe7078SQu Wenruo btrfs_backref_unlock_node_buffer(upper); 22948df01fddSNikolay Borisov if (ret) 22953fd0a558SYan, Zheng break; 22965d4f98a2SYan Zheng } 22973fd0a558SYan, Zheng 22988df01fddSNikolay Borisov if (!ret && node->pending) { 2299b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 23003fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 23013fd0a558SYan, Zheng node->pending = 0; 23025d4f98a2SYan Zheng } 23033fd0a558SYan, Zheng 23045d4f98a2SYan Zheng path->lowest_level = 0; 23058df01fddSNikolay Borisov BUG_ON(ret == -ENOSPC); 23068df01fddSNikolay Borisov return ret; 23075d4f98a2SYan Zheng } 23085d4f98a2SYan Zheng 23095d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 23103fd0a558SYan, Zheng struct reloc_control *rc, 2311a26195a5SQu Wenruo struct btrfs_backref_node *node, 23125d4f98a2SYan Zheng struct btrfs_path *path) 23135d4f98a2SYan Zheng { 23145d4f98a2SYan Zheng struct btrfs_key key; 23155d4f98a2SYan Zheng 23165d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 23173fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 23185d4f98a2SYan Zheng } 23195d4f98a2SYan Zheng 23205d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 23213fd0a558SYan, Zheng struct reloc_control *rc, 23223fd0a558SYan, Zheng struct btrfs_path *path, int err) 23235d4f98a2SYan Zheng { 23243fd0a558SYan, Zheng LIST_HEAD(list); 2325a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 2326a26195a5SQu Wenruo struct btrfs_backref_node *node; 23275d4f98a2SYan Zheng int level; 23285d4f98a2SYan Zheng int ret; 23295d4f98a2SYan Zheng 23305d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 23315d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 23325d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 2333a26195a5SQu Wenruo struct btrfs_backref_node, list); 23343fd0a558SYan, Zheng list_move_tail(&node->list, &list); 23353fd0a558SYan, Zheng BUG_ON(!node->pending); 23365d4f98a2SYan Zheng 23373fd0a558SYan, Zheng if (!err) { 23383fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 23395d4f98a2SYan Zheng if (ret < 0) 23405d4f98a2SYan Zheng err = ret; 23415d4f98a2SYan Zheng } 23425d4f98a2SYan Zheng } 23433fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 23443fd0a558SYan, Zheng } 23455d4f98a2SYan Zheng return err; 23465d4f98a2SYan Zheng } 23475d4f98a2SYan Zheng 23485d4f98a2SYan Zheng /* 23495d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 23505d4f98a2SYan Zheng * as processed. 23515d4f98a2SYan Zheng */ 23525d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 2353a26195a5SQu Wenruo struct btrfs_backref_node *node) 23545d4f98a2SYan Zheng { 2355a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2356a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2357a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 23585d4f98a2SYan Zheng int index = 0; 23595d4f98a2SYan Zheng 23605d4f98a2SYan Zheng while (next) { 23615d4f98a2SYan Zheng cond_resched(); 23625d4f98a2SYan Zheng while (1) { 23635d4f98a2SYan Zheng if (next->processed) 23645d4f98a2SYan Zheng break; 23655d4f98a2SYan Zheng 23669569cc20SQu Wenruo mark_block_processed(rc, next); 23675d4f98a2SYan Zheng 23685d4f98a2SYan Zheng if (list_empty(&next->upper)) 23695d4f98a2SYan Zheng break; 23705d4f98a2SYan Zheng 23715d4f98a2SYan Zheng edge = list_entry(next->upper.next, 2372a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 23735d4f98a2SYan Zheng edges[index++] = edge; 23745d4f98a2SYan Zheng next = edge->node[UPPER]; 23755d4f98a2SYan Zheng } 23765d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 23775d4f98a2SYan Zheng } 23785d4f98a2SYan Zheng } 23795d4f98a2SYan Zheng 23807476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 23815d4f98a2SYan Zheng { 2382da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 23837476dfdaSDavid Sterba 23845d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 23859655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 23865d4f98a2SYan Zheng return 1; 23875d4f98a2SYan Zheng return 0; 23885d4f98a2SYan Zheng } 23895d4f98a2SYan Zheng 23902ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 23915d4f98a2SYan Zheng struct tree_block *block) 23925d4f98a2SYan Zheng { 23935d4f98a2SYan Zheng struct extent_buffer *eb; 23945d4f98a2SYan Zheng 2395f7ba2d37SJosef Bacik eb = read_tree_block(fs_info, block->bytenr, block->owner, 2396f7ba2d37SJosef Bacik block->key.offset, block->level, NULL); 239764c043deSLiu Bo if (IS_ERR(eb)) { 239864c043deSLiu Bo return PTR_ERR(eb); 239964c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2400416bc658SJosef Bacik free_extent_buffer(eb); 2401416bc658SJosef Bacik return -EIO; 2402416bc658SJosef Bacik } 24035d4f98a2SYan Zheng if (block->level == 0) 24045d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 24055d4f98a2SYan Zheng else 24065d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 24075d4f98a2SYan Zheng free_extent_buffer(eb); 24085d4f98a2SYan Zheng block->key_ready = 1; 24095d4f98a2SYan Zheng return 0; 24105d4f98a2SYan Zheng } 24115d4f98a2SYan Zheng 24125d4f98a2SYan Zheng /* 24135d4f98a2SYan Zheng * helper function to relocate a tree block 24145d4f98a2SYan Zheng */ 24155d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 24165d4f98a2SYan Zheng struct reloc_control *rc, 2417a26195a5SQu Wenruo struct btrfs_backref_node *node, 24185d4f98a2SYan Zheng struct btrfs_key *key, 24195d4f98a2SYan Zheng struct btrfs_path *path) 24205d4f98a2SYan Zheng { 24215d4f98a2SYan Zheng struct btrfs_root *root; 24223fd0a558SYan, Zheng int ret = 0; 24235d4f98a2SYan Zheng 24243fd0a558SYan, Zheng if (!node) 24255d4f98a2SYan Zheng return 0; 24263fd0a558SYan, Zheng 24275f6b2e5cSJosef Bacik /* 24285f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 24295f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 24305f6b2e5cSJosef Bacik */ 24315f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 24325f6b2e5cSJosef Bacik if (ret) 24335f6b2e5cSJosef Bacik goto out; 24345f6b2e5cSJosef Bacik 24353fd0a558SYan, Zheng BUG_ON(node->processed); 2436147d256eSZhaolei root = select_one_root(node); 24373fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 24383fd0a558SYan, Zheng update_processed_blocks(rc, node); 24393fd0a558SYan, Zheng goto out; 24405d4f98a2SYan Zheng } 24415d4f98a2SYan Zheng 24423fd0a558SYan, Zheng if (root) { 244392a7cc42SQu Wenruo if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 24443fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 24453fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 24463fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 24473fd0a558SYan, Zheng root = root->reloc_root; 24483fd0a558SYan, Zheng node->new_bytenr = root->node->start; 244900246528SJosef Bacik btrfs_put_root(node->root); 245000246528SJosef Bacik node->root = btrfs_grab_root(root); 24510b530bc5SJosef Bacik ASSERT(node->root); 24523fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 24533fd0a558SYan, Zheng } else { 24545d4f98a2SYan Zheng path->lowest_level = node->level; 24555d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2456b3b4aa74SDavid Sterba btrfs_release_path(path); 24573fd0a558SYan, Zheng if (ret > 0) 24585d4f98a2SYan Zheng ret = 0; 24593fd0a558SYan, Zheng } 24603fd0a558SYan, Zheng if (!ret) 24613fd0a558SYan, Zheng update_processed_blocks(rc, node); 24623fd0a558SYan, Zheng } else { 24633fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 24643fd0a558SYan, Zheng } 24655d4f98a2SYan Zheng out: 24660647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 2467023acb07SQu Wenruo btrfs_backref_cleanup_node(&rc->backref_cache, node); 24685d4f98a2SYan Zheng return ret; 24695d4f98a2SYan Zheng } 24705d4f98a2SYan Zheng 24715d4f98a2SYan Zheng /* 24725d4f98a2SYan Zheng * relocate a list of blocks 24735d4f98a2SYan Zheng */ 24745d4f98a2SYan Zheng static noinline_for_stack 24755d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 24765d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 24775d4f98a2SYan Zheng { 24782ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2479a26195a5SQu Wenruo struct btrfs_backref_node *node; 24805d4f98a2SYan Zheng struct btrfs_path *path; 24815d4f98a2SYan Zheng struct tree_block *block; 248298ff7b94SQu Wenruo struct tree_block *next; 24835d4f98a2SYan Zheng int ret; 24845d4f98a2SYan Zheng int err = 0; 24855d4f98a2SYan Zheng 24865d4f98a2SYan Zheng path = btrfs_alloc_path(); 2487e1a12670SLiu Bo if (!path) { 2488e1a12670SLiu Bo err = -ENOMEM; 248934c2b290SDavid Sterba goto out_free_blocks; 2490e1a12670SLiu Bo } 24915d4f98a2SYan Zheng 249298ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 249398ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 24945d4f98a2SYan Zheng if (!block->key_ready) 2495f7ba2d37SJosef Bacik btrfs_readahead_tree_block(fs_info, block->bytenr, 2496f7ba2d37SJosef Bacik block->owner, 0, 24973fbaf258SJosef Bacik block->level); 24985d4f98a2SYan Zheng } 24995d4f98a2SYan Zheng 250098ff7b94SQu Wenruo /* Get first keys */ 250198ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 250234c2b290SDavid Sterba if (!block->key_ready) { 25032ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 250434c2b290SDavid Sterba if (err) 250534c2b290SDavid Sterba goto out_free_path; 250634c2b290SDavid Sterba } 25075d4f98a2SYan Zheng } 25085d4f98a2SYan Zheng 250998ff7b94SQu Wenruo /* Do tree relocation */ 251098ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 25113fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 25125d4f98a2SYan Zheng block->level, block->bytenr); 25135d4f98a2SYan Zheng if (IS_ERR(node)) { 25145d4f98a2SYan Zheng err = PTR_ERR(node); 25155d4f98a2SYan Zheng goto out; 25165d4f98a2SYan Zheng } 25175d4f98a2SYan Zheng 25185d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 25195d4f98a2SYan Zheng path); 25205d4f98a2SYan Zheng if (ret < 0) { 25215d4f98a2SYan Zheng err = ret; 252250dbbb71SJosef Bacik break; 25235d4f98a2SYan Zheng } 25245d4f98a2SYan Zheng } 25255d4f98a2SYan Zheng out: 25263fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 25275d4f98a2SYan Zheng 252834c2b290SDavid Sterba out_free_path: 25295d4f98a2SYan Zheng btrfs_free_path(path); 253034c2b290SDavid Sterba out_free_blocks: 2531e1a12670SLiu Bo free_block_list(blocks); 25325d4f98a2SYan Zheng return err; 25335d4f98a2SYan Zheng } 25345d4f98a2SYan Zheng 2535056d9becSNikolay Borisov static noinline_for_stack int prealloc_file_extent_cluster( 2536056d9becSNikolay Borisov struct btrfs_inode *inode, 2537efa56464SYan, Zheng struct file_extent_cluster *cluster) 2538efa56464SYan, Zheng { 2539efa56464SYan, Zheng u64 alloc_hint = 0; 2540efa56464SYan, Zheng u64 start; 2541efa56464SYan, Zheng u64 end; 2542056d9becSNikolay Borisov u64 offset = inode->index_cnt; 2543efa56464SYan, Zheng u64 num_bytes; 25444e9d0d01SNikolay Borisov int nr; 2545efa56464SYan, Zheng int ret = 0; 2546dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 2547dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 2548214e61d0SNikolay Borisov u64 cur_offset = prealloc_start; 2549efa56464SYan, Zheng 2550efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2551056d9becSNikolay Borisov ret = btrfs_alloc_data_chunk_ondemand(inode, 2552dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 2553efa56464SYan, Zheng if (ret) 2554214e61d0SNikolay Borisov return ret; 2555efa56464SYan, Zheng 2556056d9becSNikolay Borisov inode_lock(&inode->vfs_inode); 25574e9d0d01SNikolay Borisov for (nr = 0; nr < cluster->nr; nr++) { 2558efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2559efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2560efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2561efa56464SYan, Zheng else 2562efa56464SYan, Zheng end = cluster->end - offset; 2563efa56464SYan, Zheng 2564056d9becSNikolay Borisov lock_extent(&inode->io_tree, start, end); 2565efa56464SYan, Zheng num_bytes = end + 1 - start; 2566056d9becSNikolay Borisov ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start, 2567efa56464SYan, Zheng num_bytes, num_bytes, 2568efa56464SYan, Zheng end + 1, &alloc_hint); 256918513091SWang Xiaoguang cur_offset = end + 1; 2570056d9becSNikolay Borisov unlock_extent(&inode->io_tree, start, end); 2571efa56464SYan, Zheng if (ret) 2572efa56464SYan, Zheng break; 2573efa56464SYan, Zheng } 2574056d9becSNikolay Borisov inode_unlock(&inode->vfs_inode); 2575214e61d0SNikolay Borisov 257618513091SWang Xiaoguang if (cur_offset < prealloc_end) 2577056d9becSNikolay Borisov btrfs_free_reserved_data_space_noquota(inode->root->fs_info, 2578a89ef455SFilipe Manana prealloc_end + 1 - cur_offset); 2579efa56464SYan, Zheng return ret; 2580efa56464SYan, Zheng } 2581efa56464SYan, Zheng 2582efa56464SYan, Zheng static noinline_for_stack 25830257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 25840257bb82SYan, Zheng u64 block_start) 25855d4f98a2SYan Zheng { 25865d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 25875d4f98a2SYan Zheng struct extent_map *em; 25880257bb82SYan, Zheng int ret = 0; 25895d4f98a2SYan Zheng 2590172ddd60SDavid Sterba em = alloc_extent_map(); 25910257bb82SYan, Zheng if (!em) 25920257bb82SYan, Zheng return -ENOMEM; 25930257bb82SYan, Zheng 25945d4f98a2SYan Zheng em->start = start; 25950257bb82SYan, Zheng em->len = end + 1 - start; 25960257bb82SYan, Zheng em->block_len = em->len; 25970257bb82SYan, Zheng em->block_start = block_start; 25985d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 25995d4f98a2SYan Zheng 2600d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 26015d4f98a2SYan Zheng while (1) { 2602890871beSChris Mason write_lock(&em_tree->lock); 260309a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 2604890871beSChris Mason write_unlock(&em_tree->lock); 26055d4f98a2SYan Zheng if (ret != -EEXIST) { 26065d4f98a2SYan Zheng free_extent_map(em); 26075d4f98a2SYan Zheng break; 26085d4f98a2SYan Zheng } 2609dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 26105d4f98a2SYan Zheng } 2611d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 26120257bb82SYan, Zheng return ret; 26130257bb82SYan, Zheng } 26145d4f98a2SYan Zheng 2615726a3421SQu Wenruo /* 2616726a3421SQu Wenruo * Allow error injection to test balance cancellation 2617726a3421SQu Wenruo */ 26181fec12a5SJosef Bacik noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 2619726a3421SQu Wenruo { 26205cb502f4SQu Wenruo return atomic_read(&fs_info->balance_cancel_req) || 26215cb502f4SQu Wenruo fatal_signal_pending(current); 2622726a3421SQu Wenruo } 2623726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 2624726a3421SQu Wenruo 26250257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 26260257bb82SYan, Zheng struct file_extent_cluster *cluster) 26270257bb82SYan, Zheng { 26282ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 26290257bb82SYan, Zheng u64 page_start; 26300257bb82SYan, Zheng u64 page_end; 26310257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 26320257bb82SYan, Zheng unsigned long index; 26330257bb82SYan, Zheng unsigned long last_index; 26340257bb82SYan, Zheng struct page *page; 26350257bb82SYan, Zheng struct file_ra_state *ra; 26363b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 26370257bb82SYan, Zheng int nr = 0; 26380257bb82SYan, Zheng int ret = 0; 26390257bb82SYan, Zheng 26400257bb82SYan, Zheng if (!cluster->nr) 26410257bb82SYan, Zheng return 0; 26420257bb82SYan, Zheng 26430257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 26440257bb82SYan, Zheng if (!ra) 26450257bb82SYan, Zheng return -ENOMEM; 26460257bb82SYan, Zheng 2647056d9becSNikolay Borisov ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster); 26480257bb82SYan, Zheng if (ret) 2649efa56464SYan, Zheng goto out; 26500257bb82SYan, Zheng 26510257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 26520257bb82SYan, Zheng 2653efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2654efa56464SYan, Zheng cluster->end - offset, cluster->start); 2655efa56464SYan, Zheng if (ret) 2656efa56464SYan, Zheng goto out; 2657efa56464SYan, Zheng 265809cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 265909cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 26600257bb82SYan, Zheng while (index <= last_index) { 26619f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 26629f3db423SNikolay Borisov PAGE_SIZE); 2663efa56464SYan, Zheng if (ret) 2664efa56464SYan, Zheng goto out; 2665efa56464SYan, Zheng 26660257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 26670257bb82SYan, Zheng if (!page) { 26680257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 26690257bb82SYan, Zheng ra, NULL, index, 26700257bb82SYan, Zheng last_index + 1 - index); 2671a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 26723b16a4e3SJosef Bacik mask); 26730257bb82SYan, Zheng if (!page) { 2674691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 267543b18595SQu Wenruo PAGE_SIZE, true); 267644db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 26778702ba93SQu Wenruo PAGE_SIZE); 26780257bb82SYan, Zheng ret = -ENOMEM; 2679efa56464SYan, Zheng goto out; 26800257bb82SYan, Zheng } 26810257bb82SYan, Zheng } 2682*32443de3SQu Wenruo ret = set_page_extent_mapped(page); 2683*32443de3SQu Wenruo if (ret < 0) { 2684*32443de3SQu Wenruo btrfs_delalloc_release_metadata(BTRFS_I(inode), 2685*32443de3SQu Wenruo PAGE_SIZE, true); 2686*32443de3SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 2687*32443de3SQu Wenruo unlock_page(page); 2688*32443de3SQu Wenruo put_page(page); 2689*32443de3SQu Wenruo goto out; 2690*32443de3SQu Wenruo } 26910257bb82SYan, Zheng 26920257bb82SYan, Zheng if (PageReadahead(page)) { 26930257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 26940257bb82SYan, Zheng ra, NULL, page, index, 26950257bb82SYan, Zheng last_index + 1 - index); 26960257bb82SYan, Zheng } 26970257bb82SYan, Zheng 26980257bb82SYan, Zheng if (!PageUptodate(page)) { 26990257bb82SYan, Zheng btrfs_readpage(NULL, page); 27000257bb82SYan, Zheng lock_page(page); 27010257bb82SYan, Zheng if (!PageUptodate(page)) { 27020257bb82SYan, Zheng unlock_page(page); 270309cbfeafSKirill A. Shutemov put_page(page); 2704691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 270543b18595SQu Wenruo PAGE_SIZE, true); 27068b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 27078702ba93SQu Wenruo PAGE_SIZE); 27080257bb82SYan, Zheng ret = -EIO; 2709efa56464SYan, Zheng goto out; 27100257bb82SYan, Zheng } 27110257bb82SYan, Zheng } 27120257bb82SYan, Zheng 27134eee4fa4SMiao Xie page_start = page_offset(page); 271409cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 27150257bb82SYan, Zheng 2716d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 27170257bb82SYan, Zheng 27180257bb82SYan, Zheng if (nr < cluster->nr && 27190257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 27200257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 27210257bb82SYan, Zheng page_start, page_end, 2722ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 27230257bb82SYan, Zheng nr++; 27240257bb82SYan, Zheng } 27250257bb82SYan, Zheng 2726c2566f22SNikolay Borisov ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, 2727c2566f22SNikolay Borisov page_end, 0, NULL); 2728765f3cebSNikolay Borisov if (ret) { 2729765f3cebSNikolay Borisov unlock_page(page); 2730765f3cebSNikolay Borisov put_page(page); 2731765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 273243b18595SQu Wenruo PAGE_SIZE, true); 2733765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 27348702ba93SQu Wenruo PAGE_SIZE); 2735765f3cebSNikolay Borisov 2736765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 2737765f3cebSNikolay Borisov page_start, page_end, 2738765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 2739765f3cebSNikolay Borisov goto out; 2740765f3cebSNikolay Borisov 2741765f3cebSNikolay Borisov } 27420257bb82SYan, Zheng set_page_dirty(page); 27430257bb82SYan, Zheng 27440257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 2745d0082371SJeff Mahoney page_start, page_end); 27460257bb82SYan, Zheng unlock_page(page); 274709cbfeafSKirill A. Shutemov put_page(page); 27480257bb82SYan, Zheng 27490257bb82SYan, Zheng index++; 27508702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 2751efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 27522ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 27537f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 27547f913c7cSQu Wenruo ret = -ECANCELED; 27557f913c7cSQu Wenruo goto out; 27567f913c7cSQu Wenruo } 27570257bb82SYan, Zheng } 27580257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 2759efa56464SYan, Zheng out: 27600257bb82SYan, Zheng kfree(ra); 27610257bb82SYan, Zheng return ret; 27620257bb82SYan, Zheng } 27630257bb82SYan, Zheng 27640257bb82SYan, Zheng static noinline_for_stack 27650257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 27660257bb82SYan, Zheng struct file_extent_cluster *cluster) 27670257bb82SYan, Zheng { 27680257bb82SYan, Zheng int ret; 27690257bb82SYan, Zheng 27700257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 27710257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 27720257bb82SYan, Zheng if (ret) 27730257bb82SYan, Zheng return ret; 27740257bb82SYan, Zheng cluster->nr = 0; 27750257bb82SYan, Zheng } 27760257bb82SYan, Zheng 27770257bb82SYan, Zheng if (!cluster->nr) 27780257bb82SYan, Zheng cluster->start = extent_key->objectid; 27790257bb82SYan, Zheng else 27800257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 27810257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 27820257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 27830257bb82SYan, Zheng cluster->nr++; 27840257bb82SYan, Zheng 27850257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 27860257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 27870257bb82SYan, Zheng if (ret) 27880257bb82SYan, Zheng return ret; 27890257bb82SYan, Zheng cluster->nr = 0; 27900257bb82SYan, Zheng } 27910257bb82SYan, Zheng return 0; 27925d4f98a2SYan Zheng } 27935d4f98a2SYan Zheng 27945d4f98a2SYan Zheng /* 27955d4f98a2SYan Zheng * helper to add a tree block to the list. 27965d4f98a2SYan Zheng * the major work is getting the generation and level of the block 27975d4f98a2SYan Zheng */ 27985d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 27995d4f98a2SYan Zheng struct btrfs_key *extent_key, 28005d4f98a2SYan Zheng struct btrfs_path *path, 28015d4f98a2SYan Zheng struct rb_root *blocks) 28025d4f98a2SYan Zheng { 28035d4f98a2SYan Zheng struct extent_buffer *eb; 28045d4f98a2SYan Zheng struct btrfs_extent_item *ei; 28055d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 28065d4f98a2SYan Zheng struct tree_block *block; 28075d4f98a2SYan Zheng struct rb_node *rb_node; 28085d4f98a2SYan Zheng u32 item_size; 28095d4f98a2SYan Zheng int level = -1; 28107fdf4b60SWang Shilong u64 generation; 2811f7ba2d37SJosef Bacik u64 owner = 0; 28125d4f98a2SYan Zheng 28135d4f98a2SYan Zheng eb = path->nodes[0]; 28145d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 28155d4f98a2SYan Zheng 28163173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 28173173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 2818f7ba2d37SJosef Bacik unsigned long ptr = 0, end; 2819f7ba2d37SJosef Bacik 28205d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 28215d4f98a2SYan Zheng struct btrfs_extent_item); 2822f7ba2d37SJosef Bacik end = (unsigned long)ei + item_size; 28233173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 28245d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 28255d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 2826f7ba2d37SJosef Bacik ptr = (unsigned long)(bi + 1); 28275d4f98a2SYan Zheng } else { 28283173a18fSJosef Bacik level = (int)extent_key->offset; 2829f7ba2d37SJosef Bacik ptr = (unsigned long)(ei + 1); 28303173a18fSJosef Bacik } 28313173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 2832f7ba2d37SJosef Bacik 2833f7ba2d37SJosef Bacik /* 2834f7ba2d37SJosef Bacik * We're reading random blocks without knowing their owner ahead 2835f7ba2d37SJosef Bacik * of time. This is ok most of the time, as all reloc roots and 2836f7ba2d37SJosef Bacik * fs roots have the same lock type. However normal trees do 2837f7ba2d37SJosef Bacik * not, and the only way to know ahead of time is to read the 2838f7ba2d37SJosef Bacik * inline ref offset. We know it's an fs root if 2839f7ba2d37SJosef Bacik * 2840f7ba2d37SJosef Bacik * 1. There's more than one ref. 2841f7ba2d37SJosef Bacik * 2. There's a SHARED_DATA_REF_KEY set. 2842f7ba2d37SJosef Bacik * 3. FULL_BACKREF is set on the flags. 2843f7ba2d37SJosef Bacik * 2844f7ba2d37SJosef Bacik * Otherwise it's safe to assume that the ref offset == the 2845f7ba2d37SJosef Bacik * owner of this block, so we can use that when calling 2846f7ba2d37SJosef Bacik * read_tree_block. 2847f7ba2d37SJosef Bacik */ 2848f7ba2d37SJosef Bacik if (btrfs_extent_refs(eb, ei) == 1 && 2849f7ba2d37SJosef Bacik !(btrfs_extent_flags(eb, ei) & 2850f7ba2d37SJosef Bacik BTRFS_BLOCK_FLAG_FULL_BACKREF) && 2851f7ba2d37SJosef Bacik ptr < end) { 2852f7ba2d37SJosef Bacik struct btrfs_extent_inline_ref *iref; 2853f7ba2d37SJosef Bacik int type; 2854f7ba2d37SJosef Bacik 2855f7ba2d37SJosef Bacik iref = (struct btrfs_extent_inline_ref *)ptr; 2856f7ba2d37SJosef Bacik type = btrfs_get_extent_inline_ref_type(eb, iref, 2857f7ba2d37SJosef Bacik BTRFS_REF_TYPE_BLOCK); 2858f7ba2d37SJosef Bacik if (type == BTRFS_REF_TYPE_INVALID) 2859f7ba2d37SJosef Bacik return -EINVAL; 2860f7ba2d37SJosef Bacik if (type == BTRFS_TREE_BLOCK_REF_KEY) 2861f7ba2d37SJosef Bacik owner = btrfs_extent_inline_ref_offset(eb, iref); 2862f7ba2d37SJosef Bacik } 28636d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 2864ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 2865ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 2866ba3c2b19SNikolay Borisov return -EINVAL; 28673173a18fSJosef Bacik } else { 28685d4f98a2SYan Zheng BUG(); 28695d4f98a2SYan Zheng } 28705d4f98a2SYan Zheng 2871b3b4aa74SDavid Sterba btrfs_release_path(path); 28725d4f98a2SYan Zheng 28735d4f98a2SYan Zheng BUG_ON(level == -1); 28745d4f98a2SYan Zheng 28755d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 28765d4f98a2SYan Zheng if (!block) 28775d4f98a2SYan Zheng return -ENOMEM; 28785d4f98a2SYan Zheng 28795d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 2880da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 28815d4f98a2SYan Zheng block->key.offset = generation; 28825d4f98a2SYan Zheng block->level = level; 28835d4f98a2SYan Zheng block->key_ready = 0; 2884f7ba2d37SJosef Bacik block->owner = owner; 28855d4f98a2SYan Zheng 2886e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 288743c04fb1SJeff Mahoney if (rb_node) 2888982c92cbSQu Wenruo btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr, 2889982c92cbSQu Wenruo -EEXIST); 28905d4f98a2SYan Zheng 28915d4f98a2SYan Zheng return 0; 28925d4f98a2SYan Zheng } 28935d4f98a2SYan Zheng 28945d4f98a2SYan Zheng /* 28955d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 28965d4f98a2SYan Zheng */ 28975d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 28985d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 28995d4f98a2SYan Zheng struct rb_root *blocks) 29005d4f98a2SYan Zheng { 29010b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 29025d4f98a2SYan Zheng struct btrfs_path *path; 29035d4f98a2SYan Zheng struct btrfs_key key; 29045d4f98a2SYan Zheng int ret; 29050b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 29065d4f98a2SYan Zheng 29077476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 29085d4f98a2SYan Zheng return 0; 29095d4f98a2SYan Zheng 2910e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 29115d4f98a2SYan Zheng return 0; 29125d4f98a2SYan Zheng 29135d4f98a2SYan Zheng path = btrfs_alloc_path(); 29145d4f98a2SYan Zheng if (!path) 29155d4f98a2SYan Zheng return -ENOMEM; 2916aee68ee5SJosef Bacik again: 29175d4f98a2SYan Zheng key.objectid = bytenr; 2918aee68ee5SJosef Bacik if (skinny) { 2919aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 2920aee68ee5SJosef Bacik key.offset = (u64)-1; 2921aee68ee5SJosef Bacik } else { 29225d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 29235d4f98a2SYan Zheng key.offset = blocksize; 2924aee68ee5SJosef Bacik } 29255d4f98a2SYan Zheng 29265d4f98a2SYan Zheng path->search_commit_root = 1; 29275d4f98a2SYan Zheng path->skip_locking = 1; 29285d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 29295d4f98a2SYan Zheng if (ret < 0) 29305d4f98a2SYan Zheng goto out; 29315d4f98a2SYan Zheng 2932aee68ee5SJosef Bacik if (ret > 0 && skinny) { 2933aee68ee5SJosef Bacik if (path->slots[0]) { 2934aee68ee5SJosef Bacik path->slots[0]--; 2935aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 2936aee68ee5SJosef Bacik path->slots[0]); 29373173a18fSJosef Bacik if (key.objectid == bytenr && 2938aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 2939aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 2940aee68ee5SJosef Bacik key.offset == blocksize))) 29413173a18fSJosef Bacik ret = 0; 29423173a18fSJosef Bacik } 2943aee68ee5SJosef Bacik 2944aee68ee5SJosef Bacik if (ret) { 2945aee68ee5SJosef Bacik skinny = false; 2946aee68ee5SJosef Bacik btrfs_release_path(path); 2947aee68ee5SJosef Bacik goto again; 2948aee68ee5SJosef Bacik } 2949aee68ee5SJosef Bacik } 2950cdccee99SLiu Bo if (ret) { 2951cdccee99SLiu Bo ASSERT(ret == 1); 2952cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 2953cdccee99SLiu Bo btrfs_err(fs_info, 2954cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 2955cdccee99SLiu Bo bytenr); 2956cdccee99SLiu Bo WARN_ON(1); 2957cdccee99SLiu Bo ret = -EINVAL; 2958cdccee99SLiu Bo goto out; 2959cdccee99SLiu Bo } 29603173a18fSJosef Bacik 29615d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 29625d4f98a2SYan Zheng out: 29635d4f98a2SYan Zheng btrfs_free_path(path); 29645d4f98a2SYan Zheng return ret; 29655d4f98a2SYan Zheng } 29665d4f98a2SYan Zheng 29670af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 296832da5386SDavid Sterba struct btrfs_block_group *block_group, 29691bbc621eSChris Mason struct inode *inode, 29701bbc621eSChris Mason u64 ino) 29710af3d00bSJosef Bacik { 29720af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 29730af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 29740af3d00bSJosef Bacik int ret = 0; 29750af3d00bSJosef Bacik 29760af3d00bSJosef Bacik if (inode) 29770af3d00bSJosef Bacik goto truncate; 29780af3d00bSJosef Bacik 29790202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, ino, root); 29802e19f1f9SAl Viro if (IS_ERR(inode)) 29810af3d00bSJosef Bacik return -ENOENT; 29820af3d00bSJosef Bacik 29830af3d00bSJosef Bacik truncate: 29842ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 29857b61cd92SMiao Xie &fs_info->global_block_rsv); 29867b61cd92SMiao Xie if (ret) 29877b61cd92SMiao Xie goto out; 29887b61cd92SMiao Xie 29897a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 29900af3d00bSJosef Bacik if (IS_ERR(trans)) { 29913612b495STsutomu Itoh ret = PTR_ERR(trans); 29920af3d00bSJosef Bacik goto out; 29930af3d00bSJosef Bacik } 29940af3d00bSJosef Bacik 299577ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 29960af3d00bSJosef Bacik 29973a45bb20SJeff Mahoney btrfs_end_transaction(trans); 29982ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 29990af3d00bSJosef Bacik out: 30000af3d00bSJosef Bacik iput(inode); 30010af3d00bSJosef Bacik return ret; 30020af3d00bSJosef Bacik } 30030af3d00bSJosef Bacik 30045d4f98a2SYan Zheng /* 300519b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 300619b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 30075d4f98a2SYan Zheng */ 300819b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 300919b546d7SQu Wenruo struct btrfs_block_group *block_group, 301019b546d7SQu Wenruo u64 data_bytenr) 30115d4f98a2SYan Zheng { 301219b546d7SQu Wenruo u64 space_cache_ino; 301319b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 30145d4f98a2SYan Zheng struct btrfs_key key; 301519b546d7SQu Wenruo bool found = false; 301619b546d7SQu Wenruo int i; 30175d4f98a2SYan Zheng int ret; 30185d4f98a2SYan Zheng 301919b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 302019b546d7SQu Wenruo return 0; 30215d4f98a2SYan Zheng 302219b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 302350e31ef4SQu Wenruo u8 type; 302450e31ef4SQu Wenruo 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); 302950e31ef4SQu Wenruo type = btrfs_file_extent_type(leaf, ei); 303050e31ef4SQu Wenruo 303150e31ef4SQu Wenruo if ((type == BTRFS_FILE_EXTENT_REG || 303250e31ef4SQu Wenruo type == BTRFS_FILE_EXTENT_PREALLOC) && 303319b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 303419b546d7SQu Wenruo found = true; 303519b546d7SQu Wenruo space_cache_ino = key.objectid; 303619b546d7SQu Wenruo break; 303719b546d7SQu Wenruo } 303819b546d7SQu Wenruo } 303919b546d7SQu Wenruo if (!found) 304019b546d7SQu Wenruo return -ENOENT; 304119b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 304219b546d7SQu Wenruo space_cache_ino); 30430af3d00bSJosef Bacik return ret; 30445d4f98a2SYan Zheng } 30455d4f98a2SYan Zheng 30465d4f98a2SYan Zheng /* 30472c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 30485d4f98a2SYan Zheng */ 30495d4f98a2SYan Zheng static noinline_for_stack 30505d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 30515d4f98a2SYan Zheng struct btrfs_key *extent_key, 30525d4f98a2SYan Zheng struct btrfs_path *path, 30535d4f98a2SYan Zheng struct rb_root *blocks) 30545d4f98a2SYan Zheng { 305519b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 305619b546d7SQu Wenruo struct ulist *leaves = NULL; 305719b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 305819b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 305919b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3060647f63bdSFilipe David Borba Manana int ret = 0; 30615d4f98a2SYan Zheng 3062b3b4aa74SDavid Sterba btrfs_release_path(path); 306319b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 306419b546d7SQu Wenruo 0, &leaves, NULL, true); 306519b546d7SQu Wenruo if (ret < 0) 306619b546d7SQu Wenruo return ret; 306719b546d7SQu Wenruo 306819b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 306919b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 307019b546d7SQu Wenruo struct extent_buffer *eb; 307119b546d7SQu Wenruo 30721b7ec85eSJosef Bacik eb = read_tree_block(fs_info, ref_node->val, 0, 0, 0, NULL); 307319b546d7SQu Wenruo if (IS_ERR(eb)) { 307419b546d7SQu Wenruo ret = PTR_ERR(eb); 307519b546d7SQu Wenruo break; 307619b546d7SQu Wenruo } 307719b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 307819b546d7SQu Wenruo extent_key->objectid); 307919b546d7SQu Wenruo free_extent_buffer(eb); 308019b546d7SQu Wenruo if (ret < 0) 308119b546d7SQu Wenruo break; 308219b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 308319b546d7SQu Wenruo if (ret < 0) 308419b546d7SQu Wenruo break; 308519b546d7SQu Wenruo } 308619b546d7SQu Wenruo if (ret < 0) 30875d4f98a2SYan Zheng free_block_list(blocks); 308819b546d7SQu Wenruo ulist_free(leaves); 308919b546d7SQu Wenruo return ret; 30905d4f98a2SYan Zheng } 30915d4f98a2SYan Zheng 30925d4f98a2SYan Zheng /* 30932c016dc2SLiu Bo * helper to find next unprocessed extent 30945d4f98a2SYan Zheng */ 30955d4f98a2SYan Zheng static noinline_for_stack 3096147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 30973fd0a558SYan, Zheng struct btrfs_key *extent_key) 30985d4f98a2SYan Zheng { 30990b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31005d4f98a2SYan Zheng struct btrfs_key key; 31015d4f98a2SYan Zheng struct extent_buffer *leaf; 31025d4f98a2SYan Zheng u64 start, end, last; 31035d4f98a2SYan Zheng int ret; 31045d4f98a2SYan Zheng 3105b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 31065d4f98a2SYan Zheng while (1) { 31075d4f98a2SYan Zheng cond_resched(); 31085d4f98a2SYan Zheng if (rc->search_start >= last) { 31095d4f98a2SYan Zheng ret = 1; 31105d4f98a2SYan Zheng break; 31115d4f98a2SYan Zheng } 31125d4f98a2SYan Zheng 31135d4f98a2SYan Zheng key.objectid = rc->search_start; 31145d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31155d4f98a2SYan Zheng key.offset = 0; 31165d4f98a2SYan Zheng 31175d4f98a2SYan Zheng path->search_commit_root = 1; 31185d4f98a2SYan Zheng path->skip_locking = 1; 31195d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 31205d4f98a2SYan Zheng 0, 0); 31215d4f98a2SYan Zheng if (ret < 0) 31225d4f98a2SYan Zheng break; 31235d4f98a2SYan Zheng next: 31245d4f98a2SYan Zheng leaf = path->nodes[0]; 31255d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 31265d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 31275d4f98a2SYan Zheng if (ret != 0) 31285d4f98a2SYan Zheng break; 31295d4f98a2SYan Zheng leaf = path->nodes[0]; 31305d4f98a2SYan Zheng } 31315d4f98a2SYan Zheng 31325d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 31335d4f98a2SYan Zheng if (key.objectid >= last) { 31345d4f98a2SYan Zheng ret = 1; 31355d4f98a2SYan Zheng break; 31365d4f98a2SYan Zheng } 31375d4f98a2SYan Zheng 31383173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 31393173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 31403173a18fSJosef Bacik path->slots[0]++; 31413173a18fSJosef Bacik goto next; 31423173a18fSJosef Bacik } 31433173a18fSJosef Bacik 31443173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 31455d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 31465d4f98a2SYan Zheng path->slots[0]++; 31475d4f98a2SYan Zheng goto next; 31485d4f98a2SYan Zheng } 31495d4f98a2SYan Zheng 31503173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 31510b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 31523173a18fSJosef Bacik rc->search_start) { 31533173a18fSJosef Bacik path->slots[0]++; 31543173a18fSJosef Bacik goto next; 31553173a18fSJosef Bacik } 31563173a18fSJosef Bacik 31575d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 31585d4f98a2SYan Zheng key.objectid, &start, &end, 3159e6138876SJosef Bacik EXTENT_DIRTY, NULL); 31605d4f98a2SYan Zheng 31615d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3162b3b4aa74SDavid Sterba btrfs_release_path(path); 31635d4f98a2SYan Zheng rc->search_start = end + 1; 31645d4f98a2SYan Zheng } else { 31653173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 31665d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 31673173a18fSJosef Bacik else 31683173a18fSJosef Bacik rc->search_start = key.objectid + 31690b246afaSJeff Mahoney fs_info->nodesize; 31703fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 31715d4f98a2SYan Zheng return 0; 31725d4f98a2SYan Zheng } 31735d4f98a2SYan Zheng } 3174b3b4aa74SDavid Sterba btrfs_release_path(path); 31755d4f98a2SYan Zheng return ret; 31765d4f98a2SYan Zheng } 31775d4f98a2SYan Zheng 31785d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 31795d4f98a2SYan Zheng { 31805d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31817585717fSChris Mason 31827585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 31835d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 31847585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 31855d4f98a2SYan Zheng } 31865d4f98a2SYan Zheng 31875d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 31885d4f98a2SYan Zheng { 31895d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31907585717fSChris Mason 31917585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 31925d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 31937585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 31945d4f98a2SYan Zheng } 31955d4f98a2SYan Zheng 31965d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 31975d4f98a2SYan Zheng { 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_EXTENT_FLAG_TREE_BLOCK)) 32035d4f98a2SYan Zheng return 1; 32045d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 32055d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 32065d4f98a2SYan Zheng return 1; 32075d4f98a2SYan Zheng return 0; 32085d4f98a2SYan Zheng } 32095d4f98a2SYan Zheng 32103fd0a558SYan, Zheng static noinline_for_stack 32113fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 32123fd0a558SYan, Zheng { 32133fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3214ac2fabacSJosef Bacik int ret; 32153fd0a558SYan, Zheng 32162ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 321766d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 32183fd0a558SYan, Zheng if (!rc->block_rsv) 32193fd0a558SYan, Zheng return -ENOMEM; 32203fd0a558SYan, Zheng 32213fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3222b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 32233fd0a558SYan, Zheng rc->extents_found = 0; 32243fd0a558SYan, Zheng rc->nodes_relocated = 0; 32253fd0a558SYan, Zheng rc->merging_rsv_size = 0; 32260647bf56SWang Shilong rc->reserved_bytes = 0; 3227da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 32280647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3229ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3230ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3231ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3232ac2fabacSJosef Bacik if (ret) 3233ac2fabacSJosef Bacik return ret; 32343fd0a558SYan, Zheng 32353fd0a558SYan, Zheng rc->create_reloc_tree = 1; 32363fd0a558SYan, Zheng set_reloc_control(rc); 32373fd0a558SYan, Zheng 32387a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 323928818947SLiu Bo if (IS_ERR(trans)) { 324028818947SLiu Bo unset_reloc_control(rc); 324128818947SLiu Bo /* 324228818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 324328818947SLiu Bo * cleanup. And callers are responsible to free the above 324428818947SLiu Bo * block rsv. 324528818947SLiu Bo */ 324628818947SLiu Bo return PTR_ERR(trans); 324728818947SLiu Bo } 32483a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 32493fd0a558SYan, Zheng return 0; 32503fd0a558SYan, Zheng } 325176dda93cSYan, Zheng 32525d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 32535d4f98a2SYan Zheng { 32542ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 32555d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 32565d4f98a2SYan Zheng struct btrfs_key key; 32575d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 32585d4f98a2SYan Zheng struct btrfs_path *path; 32595d4f98a2SYan Zheng struct btrfs_extent_item *ei; 32605d4f98a2SYan Zheng u64 flags; 32615d4f98a2SYan Zheng u32 item_size; 32625d4f98a2SYan Zheng int ret; 32635d4f98a2SYan Zheng int err = 0; 3264c87f08caSChris Mason int progress = 0; 32655d4f98a2SYan Zheng 32665d4f98a2SYan Zheng path = btrfs_alloc_path(); 32673fd0a558SYan, Zheng if (!path) 32685d4f98a2SYan Zheng return -ENOMEM; 3269e4058b54SDavid Sterba path->reada = READA_FORWARD; 32703fd0a558SYan, Zheng 32713fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 32723fd0a558SYan, Zheng if (ret) { 32733fd0a558SYan, Zheng err = ret; 32743fd0a558SYan, Zheng goto out_free; 32752423fdfbSJiri Slaby } 32765d4f98a2SYan Zheng 32775d4f98a2SYan Zheng while (1) { 32780647bf56SWang Shilong rc->reserved_bytes = 0; 32790647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 32800647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 32810647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 32820647bf56SWang Shilong if (ret) { 32830647bf56SWang Shilong err = ret; 32840647bf56SWang Shilong break; 32850647bf56SWang Shilong } 3286c87f08caSChris Mason progress++; 3287a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 32880f788c58SLiu Bo if (IS_ERR(trans)) { 32890f788c58SLiu Bo err = PTR_ERR(trans); 32900f788c58SLiu Bo trans = NULL; 32910f788c58SLiu Bo break; 32920f788c58SLiu Bo } 3293c87f08caSChris Mason restart: 32943fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 32953a45bb20SJeff Mahoney btrfs_end_transaction(trans); 329642a657f5SPan Bian trans = NULL; 32973fd0a558SYan, Zheng continue; 32983fd0a558SYan, Zheng } 32993fd0a558SYan, Zheng 3300147d256eSZhaolei ret = find_next_extent(rc, path, &key); 33015d4f98a2SYan Zheng if (ret < 0) 33025d4f98a2SYan Zheng err = ret; 33035d4f98a2SYan Zheng if (ret != 0) 33045d4f98a2SYan Zheng break; 33055d4f98a2SYan Zheng 33065d4f98a2SYan Zheng rc->extents_found++; 33075d4f98a2SYan Zheng 33085d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 33095d4f98a2SYan Zheng struct btrfs_extent_item); 33103fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 33115d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 33125d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 33135d4f98a2SYan Zheng ret = check_extent_flags(flags); 33145d4f98a2SYan Zheng BUG_ON(ret); 33156d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3316ba3c2b19SNikolay Borisov err = -EINVAL; 3317ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 3318ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 3319ba3c2b19SNikolay Borisov break; 33205d4f98a2SYan Zheng } else { 33215d4f98a2SYan Zheng BUG(); 33225d4f98a2SYan Zheng } 33235d4f98a2SYan Zheng 33245d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 33255d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 33265d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 33275d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 33285d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 33295d4f98a2SYan Zheng } else { 3330b3b4aa74SDavid Sterba btrfs_release_path(path); 33315d4f98a2SYan Zheng ret = 0; 33325d4f98a2SYan Zheng } 33335d4f98a2SYan Zheng if (ret < 0) { 33343fd0a558SYan, Zheng err = ret; 33355d4f98a2SYan Zheng break; 33365d4f98a2SYan Zheng } 33375d4f98a2SYan Zheng 33385d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 33395d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 33405d4f98a2SYan Zheng if (ret < 0) { 33413fd0a558SYan, Zheng if (ret != -EAGAIN) { 33425d4f98a2SYan Zheng err = ret; 33435d4f98a2SYan Zheng break; 33445d4f98a2SYan Zheng } 33453fd0a558SYan, Zheng rc->extents_found--; 33463fd0a558SYan, Zheng rc->search_start = key.objectid; 33473fd0a558SYan, Zheng } 33485d4f98a2SYan Zheng } 33495d4f98a2SYan Zheng 33503a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 33512ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 33523fd0a558SYan, Zheng trans = NULL; 33535d4f98a2SYan Zheng 33545d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 33555d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 33565d4f98a2SYan Zheng rc->found_file_extent = 1; 33570257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 33583fd0a558SYan, Zheng &key, &rc->cluster); 33595d4f98a2SYan Zheng if (ret < 0) { 33605d4f98a2SYan Zheng err = ret; 33615d4f98a2SYan Zheng break; 33625d4f98a2SYan Zheng } 33635d4f98a2SYan Zheng } 3364f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 3365f31ea088SQu Wenruo err = -ECANCELED; 3366f31ea088SQu Wenruo break; 3367f31ea088SQu Wenruo } 33685d4f98a2SYan Zheng } 3369c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 337043a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 33719689457bSShilong Wang if (ret == 1) { 3372c87f08caSChris Mason err = 0; 3373c87f08caSChris Mason progress = 0; 3374c87f08caSChris Mason goto restart; 3375c87f08caSChris Mason } 3376c87f08caSChris Mason } 33773fd0a558SYan, Zheng 3378b3b4aa74SDavid Sterba btrfs_release_path(path); 337991166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 33805d4f98a2SYan Zheng 33815d4f98a2SYan Zheng if (trans) { 33823a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 33832ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 33845d4f98a2SYan Zheng } 33855d4f98a2SYan Zheng 33860257bb82SYan, Zheng if (!err) { 33873fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 33883fd0a558SYan, Zheng &rc->cluster); 33890257bb82SYan, Zheng if (ret < 0) 33900257bb82SYan, Zheng err = ret; 33910257bb82SYan, Zheng } 33920257bb82SYan, Zheng 33933fd0a558SYan, Zheng rc->create_reloc_tree = 0; 33943fd0a558SYan, Zheng set_reloc_control(rc); 33950257bb82SYan, Zheng 339613fe1bdbSQu Wenruo btrfs_backref_release_cache(&rc->backref_cache); 339763f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 33985d4f98a2SYan Zheng 33997f913c7cSQu Wenruo /* 34007f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 34017f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 34027f913c7cSQu Wenruo * 34037f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 34047f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 34057f913c7cSQu Wenruo * merge_reloc_roots() 34067f913c7cSQu Wenruo */ 34073fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 34085d4f98a2SYan Zheng 34095d4f98a2SYan Zheng merge_reloc_roots(rc); 34105d4f98a2SYan Zheng 34113fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 34125d4f98a2SYan Zheng unset_reloc_control(rc); 341363f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 34145d4f98a2SYan Zheng 34155d4f98a2SYan Zheng /* get rid of pinned extents */ 34167a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 341762b99540SQu Wenruo if (IS_ERR(trans)) { 34183612b495STsutomu Itoh err = PTR_ERR(trans); 341962b99540SQu Wenruo goto out_free; 342062b99540SQu Wenruo } 34213a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 34226217b0faSJosef Bacik out_free: 3423d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3424d2311e69SQu Wenruo if (ret < 0 && !err) 3425d2311e69SQu Wenruo err = ret; 34262ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 34273fd0a558SYan, Zheng btrfs_free_path(path); 34285d4f98a2SYan Zheng return err; 34295d4f98a2SYan Zheng } 34305d4f98a2SYan Zheng 34315d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 34320257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 34335d4f98a2SYan Zheng { 34345d4f98a2SYan Zheng struct btrfs_path *path; 34355d4f98a2SYan Zheng struct btrfs_inode_item *item; 34365d4f98a2SYan Zheng struct extent_buffer *leaf; 34375d4f98a2SYan Zheng int ret; 34385d4f98a2SYan Zheng 34395d4f98a2SYan Zheng path = btrfs_alloc_path(); 34405d4f98a2SYan Zheng if (!path) 34415d4f98a2SYan Zheng return -ENOMEM; 34425d4f98a2SYan Zheng 34435d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 34445d4f98a2SYan Zheng if (ret) 34455d4f98a2SYan Zheng goto out; 34465d4f98a2SYan Zheng 34475d4f98a2SYan Zheng leaf = path->nodes[0]; 34485d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 3449b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 34505d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 34510257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 34525d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 34533fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 34543fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 34555d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 34565d4f98a2SYan Zheng out: 34575d4f98a2SYan Zheng btrfs_free_path(path); 34585d4f98a2SYan Zheng return ret; 34595d4f98a2SYan Zheng } 34605d4f98a2SYan Zheng 34615d4f98a2SYan Zheng /* 34625d4f98a2SYan Zheng * helper to create inode for data relocation. 34635d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 34645d4f98a2SYan Zheng */ 34653fd0a558SYan, Zheng static noinline_for_stack 34663fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 346732da5386SDavid Sterba struct btrfs_block_group *group) 34685d4f98a2SYan Zheng { 34695d4f98a2SYan Zheng struct inode *inode = NULL; 34705d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 34715d4f98a2SYan Zheng struct btrfs_root *root; 34724624900dSZhaolei u64 objectid; 34735d4f98a2SYan Zheng int err = 0; 34745d4f98a2SYan Zheng 3475aeb935a4SQu Wenruo root = btrfs_grab_root(fs_info->data_reloc_root); 3476a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 347776deacf0SJosef Bacik if (IS_ERR(trans)) { 347800246528SJosef Bacik btrfs_put_root(root); 34793fd0a558SYan, Zheng return ERR_CAST(trans); 348076deacf0SJosef Bacik } 34815d4f98a2SYan Zheng 3482543068a2SNikolay Borisov err = btrfs_get_free_objectid(root, &objectid); 34835d4f98a2SYan Zheng if (err) 34845d4f98a2SYan Zheng goto out; 34855d4f98a2SYan Zheng 34860257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 34875d4f98a2SYan Zheng BUG_ON(err); 34885d4f98a2SYan Zheng 34890202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, objectid, root); 34902e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 3491b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 34925d4f98a2SYan Zheng 349373f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 34945d4f98a2SYan Zheng out: 349500246528SJosef Bacik btrfs_put_root(root); 34963a45bb20SJeff Mahoney btrfs_end_transaction(trans); 34972ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 34985d4f98a2SYan Zheng if (err) { 34995d4f98a2SYan Zheng if (inode) 35005d4f98a2SYan Zheng iput(inode); 35015d4f98a2SYan Zheng inode = ERR_PTR(err); 35025d4f98a2SYan Zheng } 35035d4f98a2SYan Zheng return inode; 35045d4f98a2SYan Zheng } 35055d4f98a2SYan Zheng 3506c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 35073fd0a558SYan, Zheng { 35083fd0a558SYan, Zheng struct reloc_control *rc; 35093fd0a558SYan, Zheng 35103fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 35113fd0a558SYan, Zheng if (!rc) 35123fd0a558SYan, Zheng return NULL; 35133fd0a558SYan, Zheng 35143fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 3515d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 3516584fb121SQu Wenruo btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1); 35173fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 351843eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 351943eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 35203fd0a558SYan, Zheng return rc; 35213fd0a558SYan, Zheng } 35223fd0a558SYan, Zheng 35231a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 35241a0afa0eSJosef Bacik { 35251a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 35261a0afa0eSJosef Bacik 35271a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 35281a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 35291a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 35301a0afa0eSJosef Bacik kfree(node); 35311a0afa0eSJosef Bacik 35321a0afa0eSJosef Bacik kfree(rc); 35331a0afa0eSJosef Bacik } 35341a0afa0eSJosef Bacik 35355d4f98a2SYan Zheng /* 3536ebce0e01SAdam Borowski * Print the block group being relocated 3537ebce0e01SAdam Borowski */ 3538ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 353932da5386SDavid Sterba struct btrfs_block_group *block_group) 3540ebce0e01SAdam Borowski { 3541f89e09cfSAnand Jain char buf[128] = {'\0'}; 3542ebce0e01SAdam Borowski 3543f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 3544ebce0e01SAdam Borowski 3545ebce0e01SAdam Borowski btrfs_info(fs_info, 3546ebce0e01SAdam Borowski "relocating block group %llu flags %s", 3547b3470b5dSDavid Sterba block_group->start, buf); 3548ebce0e01SAdam Borowski } 3549ebce0e01SAdam Borowski 3550430640e3SQu Wenruo static const char *stage_to_string(int stage) 3551430640e3SQu Wenruo { 3552430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 3553430640e3SQu Wenruo return "move data extents"; 3554430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 3555430640e3SQu Wenruo return "update data pointers"; 3556430640e3SQu Wenruo return "unknown"; 3557430640e3SQu Wenruo } 3558430640e3SQu Wenruo 3559ebce0e01SAdam Borowski /* 35605d4f98a2SYan Zheng * function to relocate all extents in a block group. 35615d4f98a2SYan Zheng */ 35626bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 35635d4f98a2SYan Zheng { 356432da5386SDavid Sterba struct btrfs_block_group *bg; 35656bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 35665d4f98a2SYan Zheng struct reloc_control *rc; 35670af3d00bSJosef Bacik struct inode *inode; 35680af3d00bSJosef Bacik struct btrfs_path *path; 35695d4f98a2SYan Zheng int ret; 3570f0486c68SYan, Zheng int rw = 0; 35715d4f98a2SYan Zheng int err = 0; 35725d4f98a2SYan Zheng 3573eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 3574eede2bf3SOmar Sandoval if (!bg) 3575eede2bf3SOmar Sandoval return -ENOENT; 3576eede2bf3SOmar Sandoval 3577eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 3578eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 3579eede2bf3SOmar Sandoval return -ETXTBSY; 3580eede2bf3SOmar Sandoval } 3581eede2bf3SOmar Sandoval 3582c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 3583eede2bf3SOmar Sandoval if (!rc) { 3584eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 35855d4f98a2SYan Zheng return -ENOMEM; 3586eede2bf3SOmar Sandoval } 35875d4f98a2SYan Zheng 3588f0486c68SYan, Zheng rc->extent_root = extent_root; 3589eede2bf3SOmar Sandoval rc->block_group = bg; 35905d4f98a2SYan Zheng 3591b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 3592f0486c68SYan, Zheng if (ret) { 3593f0486c68SYan, Zheng err = ret; 3594f0486c68SYan, Zheng goto out; 3595f0486c68SYan, Zheng } 3596f0486c68SYan, Zheng rw = 1; 3597f0486c68SYan, Zheng 35980af3d00bSJosef Bacik path = btrfs_alloc_path(); 35990af3d00bSJosef Bacik if (!path) { 36000af3d00bSJosef Bacik err = -ENOMEM; 36010af3d00bSJosef Bacik goto out; 36020af3d00bSJosef Bacik } 36030af3d00bSJosef Bacik 36047949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 36050af3d00bSJosef Bacik btrfs_free_path(path); 36060af3d00bSJosef Bacik 36070af3d00bSJosef Bacik if (!IS_ERR(inode)) 36081bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 36090af3d00bSJosef Bacik else 36100af3d00bSJosef Bacik ret = PTR_ERR(inode); 36110af3d00bSJosef Bacik 36120af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 36130af3d00bSJosef Bacik err = ret; 36140af3d00bSJosef Bacik goto out; 36150af3d00bSJosef Bacik } 36160af3d00bSJosef Bacik 36175d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 36185d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 36195d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 36205d4f98a2SYan Zheng rc->data_inode = NULL; 36215d4f98a2SYan Zheng goto out; 36225d4f98a2SYan Zheng } 36235d4f98a2SYan Zheng 36240b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 36255d4f98a2SYan Zheng 36269cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 3627f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 36286374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 3629b3470b5dSDavid Sterba rc->block_group->start, 3630b3470b5dSDavid Sterba rc->block_group->length); 36315d4f98a2SYan Zheng 36325d4f98a2SYan Zheng while (1) { 3633430640e3SQu Wenruo int finishes_stage; 3634430640e3SQu Wenruo 363576dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 36365d4f98a2SYan Zheng ret = relocate_block_group(rc); 363776dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 3638ff612ba7SJosef Bacik if (ret < 0) 36395d4f98a2SYan Zheng err = ret; 3640ff612ba7SJosef Bacik 3641430640e3SQu Wenruo finishes_stage = rc->stage; 3642ff612ba7SJosef Bacik /* 3643ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 3644ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 3645ff612ba7SJosef Bacik * different block group we could end up hitting the 3646ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 3647ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 3648ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 3649ff612ba7SJosef Bacik * out of the loop if we hit an error. 3650ff612ba7SJosef Bacik */ 3651ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 3652ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 3653ff612ba7SJosef Bacik (u64)-1); 3654ff612ba7SJosef Bacik if (ret) 3655ff612ba7SJosef Bacik err = ret; 3656ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 3657ff612ba7SJosef Bacik 0, -1); 3658ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 36595d4f98a2SYan Zheng } 36605d4f98a2SYan Zheng 3661ff612ba7SJosef Bacik if (err < 0) 3662ff612ba7SJosef Bacik goto out; 3663ff612ba7SJosef Bacik 36645d4f98a2SYan Zheng if (rc->extents_found == 0) 36655d4f98a2SYan Zheng break; 36665d4f98a2SYan Zheng 3667430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 3668430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 36695d4f98a2SYan Zheng } 36705d4f98a2SYan Zheng 36715d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 36725d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 3673bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 36745d4f98a2SYan Zheng out: 3675f0486c68SYan, Zheng if (err && rw) 36762ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 36775d4f98a2SYan Zheng iput(rc->data_inode); 36785d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 36791a0afa0eSJosef Bacik free_reloc_control(rc); 36805d4f98a2SYan Zheng return err; 36815d4f98a2SYan Zheng } 36825d4f98a2SYan Zheng 368376dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 368476dda93cSYan, Zheng { 36850b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 368676dda93cSYan, Zheng struct btrfs_trans_handle *trans; 368779787eaaSJeff Mahoney int ret, err; 368876dda93cSYan, Zheng 36890b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 369079787eaaSJeff Mahoney if (IS_ERR(trans)) 369179787eaaSJeff Mahoney return PTR_ERR(trans); 369276dda93cSYan, Zheng 369376dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 369476dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 3695c8422684SDavid Sterba btrfs_set_root_drop_level(&root->root_item, 0); 369676dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 36970b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 369876dda93cSYan, Zheng &root->root_key, &root->root_item); 369976dda93cSYan, Zheng 37003a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 370179787eaaSJeff Mahoney if (err) 370279787eaaSJeff Mahoney return err; 370379787eaaSJeff Mahoney return ret; 370476dda93cSYan, Zheng } 370576dda93cSYan, Zheng 37065d4f98a2SYan Zheng /* 37075d4f98a2SYan Zheng * recover relocation interrupted by system crash. 37085d4f98a2SYan Zheng * 37095d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 37105d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 37115d4f98a2SYan Zheng */ 37125d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 37135d4f98a2SYan Zheng { 37140b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 37155d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 37165d4f98a2SYan Zheng struct btrfs_key key; 37175d4f98a2SYan Zheng struct btrfs_root *fs_root; 37185d4f98a2SYan Zheng struct btrfs_root *reloc_root; 37195d4f98a2SYan Zheng struct btrfs_path *path; 37205d4f98a2SYan Zheng struct extent_buffer *leaf; 37215d4f98a2SYan Zheng struct reloc_control *rc = NULL; 37225d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 37235d4f98a2SYan Zheng int ret; 37245d4f98a2SYan Zheng int err = 0; 37255d4f98a2SYan Zheng 37265d4f98a2SYan Zheng path = btrfs_alloc_path(); 37275d4f98a2SYan Zheng if (!path) 37285d4f98a2SYan Zheng return -ENOMEM; 3729e4058b54SDavid Sterba path->reada = READA_BACK; 37305d4f98a2SYan Zheng 37315d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 37325d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 37335d4f98a2SYan Zheng key.offset = (u64)-1; 37345d4f98a2SYan Zheng 37355d4f98a2SYan Zheng while (1) { 37360b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 37375d4f98a2SYan Zheng path, 0, 0); 37385d4f98a2SYan Zheng if (ret < 0) { 37395d4f98a2SYan Zheng err = ret; 37405d4f98a2SYan Zheng goto out; 37415d4f98a2SYan Zheng } 37425d4f98a2SYan Zheng if (ret > 0) { 37435d4f98a2SYan Zheng if (path->slots[0] == 0) 37445d4f98a2SYan Zheng break; 37455d4f98a2SYan Zheng path->slots[0]--; 37465d4f98a2SYan Zheng } 37475d4f98a2SYan Zheng leaf = path->nodes[0]; 37485d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 3749b3b4aa74SDavid Sterba btrfs_release_path(path); 37505d4f98a2SYan Zheng 37515d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 37525d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 37535d4f98a2SYan Zheng break; 37545d4f98a2SYan Zheng 37553dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 37565d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 37575d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 37585d4f98a2SYan Zheng goto out; 37595d4f98a2SYan Zheng } 37605d4f98a2SYan Zheng 376192a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 37625d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 37635d4f98a2SYan Zheng 37645d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 3765a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, 3766a820feb5SDavid Sterba reloc_root->root_key.offset, false); 37675d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 376876dda93cSYan, Zheng ret = PTR_ERR(fs_root); 376976dda93cSYan, Zheng if (ret != -ENOENT) { 377076dda93cSYan, Zheng err = ret; 37715d4f98a2SYan Zheng goto out; 37725d4f98a2SYan Zheng } 377379787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 377479787eaaSJeff Mahoney if (ret < 0) { 377579787eaaSJeff Mahoney err = ret; 377679787eaaSJeff Mahoney goto out; 377779787eaaSJeff Mahoney } 3778932fd26dSJosef Bacik } else { 377900246528SJosef Bacik btrfs_put_root(fs_root); 378076dda93cSYan, Zheng } 37815d4f98a2SYan Zheng } 37825d4f98a2SYan Zheng 37835d4f98a2SYan Zheng if (key.offset == 0) 37845d4f98a2SYan Zheng break; 37855d4f98a2SYan Zheng 37865d4f98a2SYan Zheng key.offset--; 37875d4f98a2SYan Zheng } 3788b3b4aa74SDavid Sterba btrfs_release_path(path); 37895d4f98a2SYan Zheng 37905d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 37915d4f98a2SYan Zheng goto out; 37925d4f98a2SYan Zheng 3793c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 37945d4f98a2SYan Zheng if (!rc) { 37955d4f98a2SYan Zheng err = -ENOMEM; 37965d4f98a2SYan Zheng goto out; 37975d4f98a2SYan Zheng } 37985d4f98a2SYan Zheng 37990b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 38005d4f98a2SYan Zheng 38015d4f98a2SYan Zheng set_reloc_control(rc); 38025d4f98a2SYan Zheng 38037a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 38043612b495STsutomu Itoh if (IS_ERR(trans)) { 38053612b495STsutomu Itoh err = PTR_ERR(trans); 3806fb2d83eeSJosef Bacik goto out_unset; 38073612b495STsutomu Itoh } 38083fd0a558SYan, Zheng 38093fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 38103fd0a558SYan, Zheng 38115d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 38125d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 38135d4f98a2SYan Zheng struct btrfs_root, root_list); 38145d4f98a2SYan Zheng list_del(&reloc_root->root_list); 38155d4f98a2SYan Zheng 38165d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 38175d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 38185d4f98a2SYan Zheng &rc->reloc_roots); 38195d4f98a2SYan Zheng continue; 38205d4f98a2SYan Zheng } 38215d4f98a2SYan Zheng 3822a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 3823a820feb5SDavid Sterba false); 382479787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 382579787eaaSJeff Mahoney err = PTR_ERR(fs_root); 3826ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 38271402d17dSXiyu Yang btrfs_end_transaction(trans); 3828fb2d83eeSJosef Bacik goto out_unset; 382979787eaaSJeff Mahoney } 38305d4f98a2SYan Zheng 3831ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 383279787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 3833f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 383400246528SJosef Bacik btrfs_put_root(fs_root); 38355d4f98a2SYan Zheng } 38365d4f98a2SYan Zheng 38373a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 383879787eaaSJeff Mahoney if (err) 3839fb2d83eeSJosef Bacik goto out_unset; 38405d4f98a2SYan Zheng 38415d4f98a2SYan Zheng merge_reloc_roots(rc); 38425d4f98a2SYan Zheng 38435d4f98a2SYan Zheng unset_reloc_control(rc); 38445d4f98a2SYan Zheng 38457a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 384662b99540SQu Wenruo if (IS_ERR(trans)) { 38473612b495STsutomu Itoh err = PTR_ERR(trans); 38486217b0faSJosef Bacik goto out_clean; 384962b99540SQu Wenruo } 38503a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 38516217b0faSJosef Bacik out_clean: 3852d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3853d2311e69SQu Wenruo if (ret < 0 && !err) 3854d2311e69SQu Wenruo err = ret; 3855fb2d83eeSJosef Bacik out_unset: 3856fb2d83eeSJosef Bacik unset_reloc_control(rc); 38571a0afa0eSJosef Bacik free_reloc_control(rc); 38583612b495STsutomu Itoh out: 3859aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 3860aca1bba6SLiu Bo 38615d4f98a2SYan Zheng btrfs_free_path(path); 38625d4f98a2SYan Zheng 38635d4f98a2SYan Zheng if (err == 0) { 38645d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 3865aeb935a4SQu Wenruo fs_root = btrfs_grab_root(fs_info->data_reloc_root); 3866aeb935a4SQu Wenruo ASSERT(fs_root); 386766b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 386800246528SJosef Bacik btrfs_put_root(fs_root); 3869932fd26dSJosef Bacik } 38705d4f98a2SYan Zheng return err; 38715d4f98a2SYan Zheng } 38725d4f98a2SYan Zheng 38735d4f98a2SYan Zheng /* 38745d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 38755d4f98a2SYan Zheng * 38765d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 38775d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 38785d4f98a2SYan Zheng */ 38797bfa9535SNikolay Borisov int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len) 38805d4f98a2SYan Zheng { 38817bfa9535SNikolay Borisov struct btrfs_fs_info *fs_info = inode->root->fs_info; 38825d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 38835d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 38845d4f98a2SYan Zheng int ret; 38855d4f98a2SYan Zheng u64 disk_bytenr; 38864577b014SJosef Bacik u64 new_bytenr; 38875d4f98a2SYan Zheng LIST_HEAD(list); 38885d4f98a2SYan Zheng 38897bfa9535SNikolay Borisov ordered = btrfs_lookup_ordered_extent(inode, file_pos); 3890bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 38915d4f98a2SYan Zheng 38927bfa9535SNikolay Borisov disk_bytenr = file_pos + inode->index_cnt; 38930b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 3894a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 389579787eaaSJeff Mahoney if (ret) 389679787eaaSJeff Mahoney goto out; 38975d4f98a2SYan Zheng 38985d4f98a2SYan Zheng while (!list_empty(&list)) { 38995d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 39005d4f98a2SYan Zheng list_del_init(&sums->list); 39015d4f98a2SYan Zheng 39024577b014SJosef Bacik /* 39034577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 39044577b014SJosef Bacik * We need to do this because we will read in entire prealloc 39054577b014SJosef Bacik * extents but we may have written to say the middle of the 39064577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 39074577b014SJosef Bacik * the right disk offset. 39084577b014SJosef Bacik * 39094577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 39104577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 39114577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 39124577b014SJosef Bacik * disk length. 39134577b014SJosef Bacik */ 3914bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 39154577b014SJosef Bacik sums->bytenr = new_bytenr; 39165d4f98a2SYan Zheng 3917f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 39185d4f98a2SYan Zheng } 391979787eaaSJeff Mahoney out: 39205d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 3921411fc6bcSAndi Kleen return ret; 39225d4f98a2SYan Zheng } 39233fd0a558SYan, Zheng 392483d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 39253fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 39263fd0a558SYan, Zheng struct extent_buffer *cow) 39273fd0a558SYan, Zheng { 39280b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 39293fd0a558SYan, Zheng struct reloc_control *rc; 3930a26195a5SQu Wenruo struct btrfs_backref_node *node; 39313fd0a558SYan, Zheng int first_cow = 0; 39323fd0a558SYan, Zheng int level; 393383d4cfd4SJosef Bacik int ret = 0; 39343fd0a558SYan, Zheng 39350b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 39363fd0a558SYan, Zheng if (!rc) 393783d4cfd4SJosef Bacik return 0; 39383fd0a558SYan, Zheng 39393fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 39403fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 39413fd0a558SYan, Zheng 39423fd0a558SYan, Zheng level = btrfs_header_level(buf); 39433fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 39443fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 39453fd0a558SYan, Zheng first_cow = 1; 39463fd0a558SYan, Zheng 39473fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 39483fd0a558SYan, Zheng rc->create_reloc_tree) { 39493fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 39503fd0a558SYan, Zheng 39513fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 39523fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 39533fd0a558SYan, Zheng node->new_bytenr != buf->start); 39543fd0a558SYan, Zheng 3955b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 395667439dadSDavid Sterba atomic_inc(&cow->refs); 39573fd0a558SYan, Zheng node->eb = cow; 39583fd0a558SYan, Zheng node->new_bytenr = cow->start; 39593fd0a558SYan, Zheng 39603fd0a558SYan, Zheng if (!node->pending) { 39613fd0a558SYan, Zheng list_move_tail(&node->list, 39623fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 39633fd0a558SYan, Zheng node->pending = 1; 39643fd0a558SYan, Zheng } 39653fd0a558SYan, Zheng 39663fd0a558SYan, Zheng if (first_cow) 39679569cc20SQu Wenruo mark_block_processed(rc, node); 39683fd0a558SYan, Zheng 39693fd0a558SYan, Zheng if (first_cow && level > 0) 39703fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 39713fd0a558SYan, Zheng } 39723fd0a558SYan, Zheng 397383d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 39743fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 397583d4cfd4SJosef Bacik return ret; 39763fd0a558SYan, Zheng } 39773fd0a558SYan, Zheng 39783fd0a558SYan, Zheng /* 39793fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 398001327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 39813fd0a558SYan, Zheng */ 3982147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 39833fd0a558SYan, Zheng u64 *bytes_to_reserve) 39843fd0a558SYan, Zheng { 398510995c04SQu Wenruo struct btrfs_root *root = pending->root; 398610995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 39873fd0a558SYan, Zheng 39886282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 39893fd0a558SYan, Zheng return; 39903fd0a558SYan, Zheng 39913fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 39923fd0a558SYan, Zheng return; 39933fd0a558SYan, Zheng 39943fd0a558SYan, Zheng root = root->reloc_root; 39953fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 39963fd0a558SYan, Zheng /* 39973fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 39983fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 39993fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 40003fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 40013fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 40023fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 40033fd0a558SYan, Zheng * tree may use more space than it frees. so we need 40043fd0a558SYan, Zheng * reserve extra space. 40053fd0a558SYan, Zheng */ 40063fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 40073fd0a558SYan, Zheng } 40083fd0a558SYan, Zheng 40093fd0a558SYan, Zheng /* 40103fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 40113fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4012f44deb74SJosef Bacik * 4013f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4014f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4015f44deb74SJosef Bacik * rc->reloc_roots. 40163fd0a558SYan, Zheng */ 401749b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 40183fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 40193fd0a558SYan, Zheng { 40203fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 40213fd0a558SYan, Zheng struct btrfs_root *reloc_root; 40223fd0a558SYan, Zheng struct btrfs_root *new_root; 402310995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 40243fd0a558SYan, Zheng int ret; 40253fd0a558SYan, Zheng 40266282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 402749b25e05SJeff Mahoney return 0; 40283fd0a558SYan, Zheng 40293fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 40303fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 40313fd0a558SYan, Zheng 40323fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 40333fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 40343fd0a558SYan, Zheng rc->block_rsv, 40353a584174SLu Fengqi rc->nodes_relocated, true); 403649b25e05SJeff Mahoney if (ret) 403749b25e05SJeff Mahoney return ret; 40383fd0a558SYan, Zheng } 40393fd0a558SYan, Zheng 40403fd0a558SYan, Zheng new_root = pending->snap; 40413fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 40423fd0a558SYan, Zheng new_root->root_key.objectid); 404349b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 404449b25e05SJeff Mahoney return PTR_ERR(reloc_root); 40453fd0a558SYan, Zheng 4046ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4047ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4048f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 40493fd0a558SYan, Zheng 405049b25e05SJeff Mahoney if (rc->create_reloc_tree) 40513fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 405249b25e05SJeff Mahoney return ret; 40533fd0a558SYan, Zheng } 4054