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) { 64157a304cfSJosef Bacik btrfs_err(fs_info, 6425d163e0eSJeff Mahoney "Duplicate root found for start=%llu while inserting into relocation tree", 6435d163e0eSJeff Mahoney node->bytenr); 64457a304cfSJosef Bacik return -EEXIST; 645ffd7b339SJeff Mahoney } 6465d4f98a2SYan Zheng 6475d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 6485d4f98a2SYan Zheng return 0; 6495d4f98a2SYan Zheng } 6505d4f98a2SYan Zheng 6515d4f98a2SYan Zheng /* 652c974c464SWang Shilong * helper to delete the 'address of tree root -> reloc tree' 6535d4f98a2SYan Zheng * mapping 6545d4f98a2SYan Zheng */ 655c974c464SWang Shilong static void __del_reloc_root(struct btrfs_root *root) 6565d4f98a2SYan Zheng { 6570b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 6585d4f98a2SYan Zheng struct rb_node *rb_node; 6595d4f98a2SYan Zheng struct mapping_node *node = NULL; 6600b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 661f44deb74SJosef Bacik bool put_ref = false; 6625d4f98a2SYan Zheng 66365c6e82bSQu Wenruo if (rc && root->node) { 6645d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 665e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 666ea287ab1SJosef Bacik root->commit_root->start); 667c974c464SWang Shilong if (rb_node) { 668c974c464SWang Shilong node = rb_entry(rb_node, struct mapping_node, rb_node); 669c974c464SWang Shilong rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 670ea287ab1SJosef Bacik RB_CLEAR_NODE(&node->rb_node); 671c974c464SWang Shilong } 672c974c464SWang Shilong spin_unlock(&rc->reloc_root_tree.lock); 673c78a10aeSJosef Bacik ASSERT(!node || (struct btrfs_root *)node->data == root); 674389305b2SQu Wenruo } 675c974c464SWang Shilong 676f44deb74SJosef Bacik /* 677f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 678f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 679f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 680f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 681f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 682f44deb74SJosef Bacik * messing with the list is in charge of the reference. 683f44deb74SJosef Bacik */ 6840b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 685f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 686f44deb74SJosef Bacik put_ref = true; 687c974c464SWang Shilong list_del_init(&root->root_list); 688f44deb74SJosef Bacik } 6890b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 690f44deb74SJosef Bacik if (put_ref) 691f44deb74SJosef Bacik btrfs_put_root(root); 692c974c464SWang Shilong kfree(node); 693c974c464SWang Shilong } 694c974c464SWang Shilong 695c974c464SWang Shilong /* 696c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 697c974c464SWang Shilong * mapping 698c974c464SWang Shilong */ 699ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 700c974c464SWang Shilong { 7010b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 702c974c464SWang Shilong struct rb_node *rb_node; 703c974c464SWang Shilong struct mapping_node *node = NULL; 7040b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 705c974c464SWang Shilong 706c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 707e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 708ea287ab1SJosef Bacik root->commit_root->start); 7095d4f98a2SYan Zheng if (rb_node) { 7105d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 7115d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 7125d4f98a2SYan Zheng } 7135d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 7145d4f98a2SYan Zheng 7158f71f3e0SLiu Bo if (!node) 7168f71f3e0SLiu Bo return 0; 7175d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 7185d4f98a2SYan Zheng 7195d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 720ea287ab1SJosef Bacik node->bytenr = root->node->start; 721e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 7225d4f98a2SYan Zheng node->bytenr, &node->rb_node); 7235d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 72443c04fb1SJeff Mahoney if (rb_node) 725982c92cbSQu Wenruo btrfs_backref_panic(fs_info, node->bytenr, -EEXIST); 7265d4f98a2SYan Zheng return 0; 7275d4f98a2SYan Zheng } 7285d4f98a2SYan Zheng 7293fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 7303fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 7315d4f98a2SYan Zheng { 7320b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 7335d4f98a2SYan Zheng struct btrfs_root *reloc_root; 7345d4f98a2SYan Zheng struct extent_buffer *eb; 7355d4f98a2SYan Zheng struct btrfs_root_item *root_item; 7365d4f98a2SYan Zheng struct btrfs_key root_key; 73784c50ba5SJosef Bacik int ret = 0; 73884c50ba5SJosef Bacik bool must_abort = false; 7395d4f98a2SYan Zheng 7405d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 74184c50ba5SJosef Bacik if (!root_item) 74284c50ba5SJosef Bacik return ERR_PTR(-ENOMEM); 7435d4f98a2SYan Zheng 7445d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 7455d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 7463fd0a558SYan, Zheng root_key.offset = objectid; 7475d4f98a2SYan Zheng 7483fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 749054570a1SFilipe Manana u64 commit_root_gen; 750054570a1SFilipe Manana 7513fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 7525d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 7535d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 75484c50ba5SJosef Bacik if (ret) 75584c50ba5SJosef Bacik goto fail; 75684c50ba5SJosef Bacik 757054570a1SFilipe Manana /* 758054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 759054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 760054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 761054570a1SFilipe Manana * either inside the critical section of a transaction commit 762054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 763054570a1SFilipe Manana * it's created before the transaction commit is started. 764054570a1SFilipe Manana */ 765054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 766054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 7673fd0a558SYan, Zheng } else { 7683fd0a558SYan, Zheng /* 7693fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 7703fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 7713fd0a558SYan, Zheng * modified after it was created have RELOC flag 7723fd0a558SYan, Zheng * set in their headers. so it's OK to not update 7733fd0a558SYan, Zheng * the 'last_snapshot'. 7743fd0a558SYan, Zheng */ 7753fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 7763fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 77784c50ba5SJosef Bacik if (ret) 77884c50ba5SJosef Bacik goto fail; 7793fd0a558SYan, Zheng } 7803fd0a558SYan, Zheng 78184c50ba5SJosef Bacik /* 78284c50ba5SJosef Bacik * We have changed references at this point, we must abort the 78384c50ba5SJosef Bacik * transaction if anything fails. 78484c50ba5SJosef Bacik */ 78584c50ba5SJosef Bacik must_abort = true; 78684c50ba5SJosef Bacik 7875d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 7885d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 7895d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 7905d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 7913fd0a558SYan, Zheng 7923fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 7933fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 7943fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 7953fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 796c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, 0); 7973fd0a558SYan, Zheng } 7985d4f98a2SYan Zheng 7995d4f98a2SYan Zheng btrfs_tree_unlock(eb); 8005d4f98a2SYan Zheng free_extent_buffer(eb); 8015d4f98a2SYan Zheng 8020b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 8035d4f98a2SYan Zheng &root_key, root_item); 80484c50ba5SJosef Bacik if (ret) 80584c50ba5SJosef Bacik goto fail; 80684c50ba5SJosef Bacik 8075d4f98a2SYan Zheng kfree(root_item); 8085d4f98a2SYan Zheng 8093dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 81084c50ba5SJosef Bacik if (IS_ERR(reloc_root)) { 81184c50ba5SJosef Bacik ret = PTR_ERR(reloc_root); 81284c50ba5SJosef Bacik goto abort; 81384c50ba5SJosef Bacik } 81492a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 8155d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 8163fd0a558SYan, Zheng return reloc_root; 81784c50ba5SJosef Bacik fail: 81884c50ba5SJosef Bacik kfree(root_item); 81984c50ba5SJosef Bacik abort: 82084c50ba5SJosef Bacik if (must_abort) 82184c50ba5SJosef Bacik btrfs_abort_transaction(trans, ret); 82284c50ba5SJosef Bacik return ERR_PTR(ret); 8233fd0a558SYan, Zheng } 8243fd0a558SYan, Zheng 8253fd0a558SYan, Zheng /* 8263fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 8273fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 828f44deb74SJosef Bacik * 829f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 830f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 8313fd0a558SYan, Zheng */ 8323fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 8333fd0a558SYan, Zheng struct btrfs_root *root) 8343fd0a558SYan, Zheng { 8350b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 8363fd0a558SYan, Zheng struct btrfs_root *reloc_root; 8370b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 83820dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 8393fd0a558SYan, Zheng int clear_rsv = 0; 840ffd7b339SJeff Mahoney int ret; 8413fd0a558SYan, Zheng 842aec7db3bSJosef Bacik if (!rc) 8432abc726aSJosef Bacik return 0; 8442abc726aSJosef Bacik 8451fac4a54SQu Wenruo /* 8461fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 8471fac4a54SQu Wenruo * create/update the dead reloc tree 8481fac4a54SQu Wenruo */ 8496282675eSQu Wenruo if (reloc_root_is_dead(root)) 8501fac4a54SQu Wenruo return 0; 8511fac4a54SQu Wenruo 852aec7db3bSJosef Bacik /* 853aec7db3bSJosef Bacik * This is subtle but important. We do not do 854aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 855aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 856aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 857aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 858aec7db3bSJosef Bacik * in. 859aec7db3bSJosef Bacik */ 8603fd0a558SYan, Zheng if (root->reloc_root) { 8613fd0a558SYan, Zheng reloc_root = root->reloc_root; 8623fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 8633fd0a558SYan, Zheng return 0; 8643fd0a558SYan, Zheng } 8653fd0a558SYan, Zheng 866aec7db3bSJosef Bacik /* 867aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 868aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 869aec7db3bSJosef Bacik */ 870aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 871aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 872aec7db3bSJosef Bacik return 0; 873aec7db3bSJosef Bacik 87420dd2cbfSMiao Xie if (!trans->reloc_reserved) { 87520dd2cbfSMiao Xie rsv = trans->block_rsv; 8763fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 8773fd0a558SYan, Zheng clear_rsv = 1; 8783fd0a558SYan, Zheng } 8793fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 8803fd0a558SYan, Zheng if (clear_rsv) 88120dd2cbfSMiao Xie trans->block_rsv = rsv; 88200bb36a0SJosef Bacik if (IS_ERR(reloc_root)) 88300bb36a0SJosef Bacik return PTR_ERR(reloc_root); 8845d4f98a2SYan Zheng 885ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 88657a304cfSJosef Bacik ASSERT(ret != -EEXIST); 88700bb36a0SJosef Bacik if (ret) { 88800bb36a0SJosef Bacik /* Pairs with create_reloc_root */ 88900bb36a0SJosef Bacik btrfs_put_root(reloc_root); 89000bb36a0SJosef Bacik return ret; 89100bb36a0SJosef Bacik } 892f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 8935d4f98a2SYan Zheng return 0; 8945d4f98a2SYan Zheng } 8955d4f98a2SYan Zheng 8965d4f98a2SYan Zheng /* 8975d4f98a2SYan Zheng * update root item of reloc tree 8985d4f98a2SYan Zheng */ 8995d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 9005d4f98a2SYan Zheng struct btrfs_root *root) 9015d4f98a2SYan Zheng { 9020b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 9035d4f98a2SYan Zheng struct btrfs_root *reloc_root; 9045d4f98a2SYan Zheng struct btrfs_root_item *root_item; 9055d4f98a2SYan Zheng int ret; 9065d4f98a2SYan Zheng 9076282675eSQu Wenruo if (!have_reloc_root(root)) 908592fbcd5SJosef Bacik return 0; 9095d4f98a2SYan Zheng 9105d4f98a2SYan Zheng reloc_root = root->reloc_root; 9115d4f98a2SYan Zheng root_item = &reloc_root->root_item; 9125d4f98a2SYan Zheng 913f44deb74SJosef Bacik /* 914f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 915f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 916f44deb74SJosef Bacik * hold it while we update the reloc root. 917f44deb74SJosef Bacik */ 918f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 919f44deb74SJosef Bacik 920d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 9210b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 9223fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 923d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 9246282675eSQu Wenruo /* 9256282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 9266282675eSQu Wenruo * have_reloc_root will not touch it from now on. 9276282675eSQu Wenruo */ 9286282675eSQu Wenruo smp_wmb(); 929c974c464SWang Shilong __del_reloc_root(reloc_root); 9305d4f98a2SYan Zheng } 9315d4f98a2SYan Zheng 9325d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 933ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 9345d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 9355d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 9365d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 9375d4f98a2SYan Zheng } 9385d4f98a2SYan Zheng 9390b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 9405d4f98a2SYan Zheng &reloc_root->root_key, root_item); 941f44deb74SJosef Bacik btrfs_put_root(reloc_root); 942592fbcd5SJosef Bacik return ret; 9435d4f98a2SYan Zheng } 9445d4f98a2SYan Zheng 9455d4f98a2SYan Zheng /* 9465d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 9475d4f98a2SYan Zheng * in a subvolume 9485d4f98a2SYan Zheng */ 9495d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 9505d4f98a2SYan Zheng { 9515d4f98a2SYan Zheng struct rb_node *node; 9525d4f98a2SYan Zheng struct rb_node *prev; 9535d4f98a2SYan Zheng struct btrfs_inode *entry; 9545d4f98a2SYan Zheng struct inode *inode; 9555d4f98a2SYan Zheng 9565d4f98a2SYan Zheng spin_lock(&root->inode_lock); 9575d4f98a2SYan Zheng again: 9585d4f98a2SYan Zheng node = root->inode_tree.rb_node; 9595d4f98a2SYan Zheng prev = NULL; 9605d4f98a2SYan Zheng while (node) { 9615d4f98a2SYan Zheng prev = node; 9625d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9635d4f98a2SYan Zheng 9644a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 9655d4f98a2SYan Zheng node = node->rb_left; 9664a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 9675d4f98a2SYan Zheng node = node->rb_right; 9685d4f98a2SYan Zheng else 9695d4f98a2SYan Zheng break; 9705d4f98a2SYan Zheng } 9715d4f98a2SYan Zheng if (!node) { 9725d4f98a2SYan Zheng while (prev) { 9735d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 9744a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 9755d4f98a2SYan Zheng node = prev; 9765d4f98a2SYan Zheng break; 9775d4f98a2SYan Zheng } 9785d4f98a2SYan Zheng prev = rb_next(prev); 9795d4f98a2SYan Zheng } 9805d4f98a2SYan Zheng } 9815d4f98a2SYan Zheng while (node) { 9825d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 9835d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 9845d4f98a2SYan Zheng if (inode) { 9855d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9865d4f98a2SYan Zheng return inode; 9875d4f98a2SYan Zheng } 9885d4f98a2SYan Zheng 9894a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 9905d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 9915d4f98a2SYan Zheng goto again; 9925d4f98a2SYan Zheng 9935d4f98a2SYan Zheng node = rb_next(node); 9945d4f98a2SYan Zheng } 9955d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 9965d4f98a2SYan Zheng return NULL; 9975d4f98a2SYan Zheng } 9985d4f98a2SYan Zheng 9995d4f98a2SYan Zheng /* 10005d4f98a2SYan Zheng * get new location of data 10015d4f98a2SYan Zheng */ 10025d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 10035d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 10045d4f98a2SYan Zheng { 10055d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 10065d4f98a2SYan Zheng struct btrfs_path *path; 10075d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10085d4f98a2SYan Zheng struct extent_buffer *leaf; 10095d4f98a2SYan Zheng int ret; 10105d4f98a2SYan Zheng 10115d4f98a2SYan Zheng path = btrfs_alloc_path(); 10125d4f98a2SYan Zheng if (!path) 10135d4f98a2SYan Zheng return -ENOMEM; 10145d4f98a2SYan Zheng 10155d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 1016f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 1017f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 10185d4f98a2SYan Zheng if (ret < 0) 10195d4f98a2SYan Zheng goto out; 10205d4f98a2SYan Zheng if (ret > 0) { 10215d4f98a2SYan Zheng ret = -ENOENT; 10225d4f98a2SYan Zheng goto out; 10235d4f98a2SYan Zheng } 10245d4f98a2SYan Zheng 10255d4f98a2SYan Zheng leaf = path->nodes[0]; 10265d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 10275d4f98a2SYan Zheng struct btrfs_file_extent_item); 10285d4f98a2SYan Zheng 10295d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 10305d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 10315d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 10325d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 10335d4f98a2SYan Zheng 10345d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 103583d4cfd4SJosef Bacik ret = -EINVAL; 10365d4f98a2SYan Zheng goto out; 10375d4f98a2SYan Zheng } 10385d4f98a2SYan Zheng 10395d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10405d4f98a2SYan Zheng ret = 0; 10415d4f98a2SYan Zheng out: 10425d4f98a2SYan Zheng btrfs_free_path(path); 10435d4f98a2SYan Zheng return ret; 10445d4f98a2SYan Zheng } 10455d4f98a2SYan Zheng 10465d4f98a2SYan Zheng /* 10475d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 10485d4f98a2SYan Zheng * the new locations. 10495d4f98a2SYan Zheng */ 10503fd0a558SYan, Zheng static noinline_for_stack 10513fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 10525d4f98a2SYan Zheng struct reloc_control *rc, 10535d4f98a2SYan Zheng struct btrfs_root *root, 10543fd0a558SYan, Zheng struct extent_buffer *leaf) 10555d4f98a2SYan Zheng { 10560b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 10575d4f98a2SYan Zheng struct btrfs_key key; 10585d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10595d4f98a2SYan Zheng struct inode *inode = NULL; 10605d4f98a2SYan Zheng u64 parent; 10615d4f98a2SYan Zheng u64 bytenr; 10623fd0a558SYan, Zheng u64 new_bytenr = 0; 10635d4f98a2SYan Zheng u64 num_bytes; 10645d4f98a2SYan Zheng u64 end; 10655d4f98a2SYan Zheng u32 nritems; 10665d4f98a2SYan Zheng u32 i; 106783d4cfd4SJosef Bacik int ret = 0; 10685d4f98a2SYan Zheng int first = 1; 10695d4f98a2SYan Zheng int dirty = 0; 10705d4f98a2SYan Zheng 10715d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 10725d4f98a2SYan Zheng return 0; 10735d4f98a2SYan Zheng 10745d4f98a2SYan Zheng /* reloc trees always use full backref */ 10755d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 10765d4f98a2SYan Zheng parent = leaf->start; 10775d4f98a2SYan Zheng else 10785d4f98a2SYan Zheng parent = 0; 10795d4f98a2SYan Zheng 10805d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 10815d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 108282fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 108382fa113fSQu Wenruo 10845d4f98a2SYan Zheng cond_resched(); 10855d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 10865d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 10875d4f98a2SYan Zheng continue; 10885d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 10895d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 10905d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 10915d4f98a2SYan Zheng continue; 10925d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10935d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 10945d4f98a2SYan Zheng if (bytenr == 0) 10955d4f98a2SYan Zheng continue; 10969569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 10979569cc20SQu Wenruo rc->block_group->length)) 10985d4f98a2SYan Zheng continue; 10995d4f98a2SYan Zheng 11005d4f98a2SYan Zheng /* 11015d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 11025d4f98a2SYan Zheng * to complete and drop the extent cache 11035d4f98a2SYan Zheng */ 11045d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 11055d4f98a2SYan Zheng if (first) { 11065d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11075d4f98a2SYan Zheng first = 0; 11084a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 11093fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 11105d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11115d4f98a2SYan Zheng } 11124a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 11135d4f98a2SYan Zheng end = key.offset + 11145d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 11155d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 11160b246afaSJeff Mahoney fs_info->sectorsize)); 11170b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 11185d4f98a2SYan Zheng end--; 11195d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1120d0082371SJeff Mahoney key.offset, end); 11215d4f98a2SYan Zheng if (!ret) 11225d4f98a2SYan Zheng continue; 11235d4f98a2SYan Zheng 1124dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1125dcdbc059SNikolay Borisov key.offset, end, 1); 11265d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1127d0082371SJeff Mahoney key.offset, end); 11285d4f98a2SYan Zheng } 11295d4f98a2SYan Zheng } 11305d4f98a2SYan Zheng 11315d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 11325d4f98a2SYan Zheng bytenr, num_bytes); 113383d4cfd4SJosef Bacik if (ret) { 113483d4cfd4SJosef Bacik /* 113583d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 113683d4cfd4SJosef Bacik * in the file extent yet. 113783d4cfd4SJosef Bacik */ 113883d4cfd4SJosef Bacik break; 11393fd0a558SYan, Zheng } 11405d4f98a2SYan Zheng 11415d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 11425d4f98a2SYan Zheng dirty = 1; 11435d4f98a2SYan Zheng 11445d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 114582fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 114682fa113fSQu Wenruo num_bytes, parent); 114782fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 114882fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1149b06c4bf5SFilipe Manana key.objectid, key.offset); 115082fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 115183d4cfd4SJosef Bacik if (ret) { 115266642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 115383d4cfd4SJosef Bacik break; 115483d4cfd4SJosef Bacik } 11555d4f98a2SYan Zheng 1156ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1157ffd4bb2aSQu Wenruo num_bytes, parent); 1158ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1159ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1160b06c4bf5SFilipe Manana key.objectid, key.offset); 1161ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 116283d4cfd4SJosef Bacik if (ret) { 116366642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 116483d4cfd4SJosef Bacik break; 116583d4cfd4SJosef Bacik } 11665d4f98a2SYan Zheng } 11675d4f98a2SYan Zheng if (dirty) 11685d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 11693fd0a558SYan, Zheng if (inode) 11703fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 117183d4cfd4SJosef Bacik return ret; 11725d4f98a2SYan Zheng } 11735d4f98a2SYan Zheng 11745d4f98a2SYan Zheng static noinline_for_stack 11755d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 11765d4f98a2SYan Zheng struct btrfs_path *path, int level) 11775d4f98a2SYan Zheng { 11785d4f98a2SYan Zheng struct btrfs_disk_key key1; 11795d4f98a2SYan Zheng struct btrfs_disk_key key2; 11805d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 11815d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 11825d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 11835d4f98a2SYan Zheng } 11845d4f98a2SYan Zheng 11855d4f98a2SYan Zheng /* 11865d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 11875d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 11885d4f98a2SYan Zheng * reloc tree was create can be replaced. 11895d4f98a2SYan Zheng * 11905d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 11915d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 11925d4f98a2SYan Zheng * errors, a negative error number is returned. 11935d4f98a2SYan Zheng */ 11943fd0a558SYan, Zheng static noinline_for_stack 11953d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 11965d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 11975d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 11985d4f98a2SYan Zheng int lowest_level, int max_level) 11995d4f98a2SYan Zheng { 12000b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 12015d4f98a2SYan Zheng struct extent_buffer *eb; 12025d4f98a2SYan Zheng struct extent_buffer *parent; 120382fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 12045d4f98a2SYan Zheng struct btrfs_key key; 12055d4f98a2SYan Zheng u64 old_bytenr; 12065d4f98a2SYan Zheng u64 new_bytenr; 12075d4f98a2SYan Zheng u64 old_ptr_gen; 12085d4f98a2SYan Zheng u64 new_ptr_gen; 12095d4f98a2SYan Zheng u64 last_snapshot; 12105d4f98a2SYan Zheng u32 blocksize; 12113fd0a558SYan, Zheng int cow = 0; 12125d4f98a2SYan Zheng int level; 12135d4f98a2SYan Zheng int ret; 12145d4f98a2SYan Zheng int slot; 12155d4f98a2SYan Zheng 12167a9213a9SJosef Bacik ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 12177a9213a9SJosef Bacik ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 12185d4f98a2SYan Zheng 12195d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 12203fd0a558SYan, Zheng again: 12215d4f98a2SYan Zheng slot = path->slots[lowest_level]; 12225d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 12235d4f98a2SYan Zheng 12245d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 12255d4f98a2SYan Zheng level = btrfs_header_level(eb); 12265d4f98a2SYan Zheng 12275d4f98a2SYan Zheng if (level < lowest_level) { 12285d4f98a2SYan Zheng btrfs_tree_unlock(eb); 12295d4f98a2SYan Zheng free_extent_buffer(eb); 12305d4f98a2SYan Zheng return 0; 12315d4f98a2SYan Zheng } 12325d4f98a2SYan Zheng 12333fd0a558SYan, Zheng if (cow) { 12349631e4ccSJosef Bacik ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb, 12359631e4ccSJosef Bacik BTRFS_NESTING_COW); 123645b87c5dSJosef Bacik if (ret) { 123745b87c5dSJosef Bacik btrfs_tree_unlock(eb); 123845b87c5dSJosef Bacik free_extent_buffer(eb); 123945b87c5dSJosef Bacik return ret; 124045b87c5dSJosef Bacik } 12413fd0a558SYan, Zheng } 12425d4f98a2SYan Zheng 12435d4f98a2SYan Zheng if (next_key) { 12445d4f98a2SYan Zheng next_key->objectid = (u64)-1; 12455d4f98a2SYan Zheng next_key->type = (u8)-1; 12465d4f98a2SYan Zheng next_key->offset = (u64)-1; 12475d4f98a2SYan Zheng } 12485d4f98a2SYan Zheng 12495d4f98a2SYan Zheng parent = eb; 12505d4f98a2SYan Zheng while (1) { 12515d4f98a2SYan Zheng level = btrfs_header_level(parent); 12527a9213a9SJosef Bacik ASSERT(level >= lowest_level); 12535d4f98a2SYan Zheng 1254e3b83361SQu Wenruo ret = btrfs_bin_search(parent, &key, &slot); 1255cbca7d59SFilipe Manana if (ret < 0) 1256cbca7d59SFilipe Manana break; 12575d4f98a2SYan Zheng if (ret && slot > 0) 12585d4f98a2SYan Zheng slot--; 12595d4f98a2SYan Zheng 12605d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 12615d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 12625d4f98a2SYan Zheng 12635d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 12640b246afaSJeff Mahoney blocksize = fs_info->nodesize; 12655d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 12665d4f98a2SYan Zheng 12675d4f98a2SYan Zheng if (level <= max_level) { 12685d4f98a2SYan Zheng eb = path->nodes[level]; 12695d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 12705d4f98a2SYan Zheng path->slots[level]); 12715d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 12725d4f98a2SYan Zheng path->slots[level]); 12735d4f98a2SYan Zheng } else { 12745d4f98a2SYan Zheng new_bytenr = 0; 12755d4f98a2SYan Zheng new_ptr_gen = 0; 12765d4f98a2SYan Zheng } 12775d4f98a2SYan Zheng 1278fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 12795d4f98a2SYan Zheng ret = level; 12805d4f98a2SYan Zheng break; 12815d4f98a2SYan Zheng } 12825d4f98a2SYan Zheng 12835d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 12845d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 12853fd0a558SYan, Zheng if (level <= lowest_level) { 12865d4f98a2SYan Zheng ret = 0; 12875d4f98a2SYan Zheng break; 12885d4f98a2SYan Zheng } 12895d4f98a2SYan Zheng 12906b3426beSJosef Bacik eb = btrfs_read_node_slot(parent, slot); 129164c043deSLiu Bo if (IS_ERR(eb)) { 129264c043deSLiu Bo ret = PTR_ERR(eb); 1293264813acSLiu Bo break; 1294416bc658SJosef Bacik } 12955d4f98a2SYan Zheng btrfs_tree_lock(eb); 12963fd0a558SYan, Zheng if (cow) { 12975d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 12989631e4ccSJosef Bacik slot, &eb, 12999631e4ccSJosef Bacik BTRFS_NESTING_COW); 130045b87c5dSJosef Bacik if (ret) { 130145b87c5dSJosef Bacik btrfs_tree_unlock(eb); 130245b87c5dSJosef Bacik free_extent_buffer(eb); 130345b87c5dSJosef Bacik break; 130445b87c5dSJosef Bacik } 13055d4f98a2SYan Zheng } 13065d4f98a2SYan Zheng 13075d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13085d4f98a2SYan Zheng free_extent_buffer(parent); 13095d4f98a2SYan Zheng 13105d4f98a2SYan Zheng parent = eb; 13115d4f98a2SYan Zheng continue; 13125d4f98a2SYan Zheng } 13135d4f98a2SYan Zheng 13143fd0a558SYan, Zheng if (!cow) { 13153fd0a558SYan, Zheng btrfs_tree_unlock(parent); 13163fd0a558SYan, Zheng free_extent_buffer(parent); 13173fd0a558SYan, Zheng cow = 1; 13183fd0a558SYan, Zheng goto again; 13193fd0a558SYan, Zheng } 13203fd0a558SYan, Zheng 13215d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 13225d4f98a2SYan Zheng path->slots[level]); 1323b3b4aa74SDavid Sterba btrfs_release_path(path); 13245d4f98a2SYan Zheng 13255d4f98a2SYan Zheng path->lowest_level = level; 13265d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 13275d4f98a2SYan Zheng path->lowest_level = 0; 13280e9873e2SJosef Bacik if (ret) { 13290e9873e2SJosef Bacik if (ret > 0) 13300e9873e2SJosef Bacik ret = -ENOENT; 13310e9873e2SJosef Bacik break; 13320e9873e2SJosef Bacik } 13335d4f98a2SYan Zheng 13345d4f98a2SYan Zheng /* 1335824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 1336824d8dffSQu Wenruo * 1337824d8dffSQu Wenruo * We must trace both trees. 1338824d8dffSQu Wenruo * 1) Tree reloc subtree 1339824d8dffSQu Wenruo * If not traced, we will leak data numbers 1340824d8dffSQu Wenruo * 2) Fs subtree 1341824d8dffSQu Wenruo * If not traced, we will double count old data 1342f616f5cdSQu Wenruo * 1343f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 1344f616f5cdSQu Wenruo * the swapped tree blocks. 1345f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 1346f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 1347824d8dffSQu Wenruo */ 1348370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 1349370a11b8SQu Wenruo rc->block_group, parent, slot, 1350370a11b8SQu Wenruo path->nodes[level], path->slots[level], 1351370a11b8SQu Wenruo last_snapshot); 1352370a11b8SQu Wenruo if (ret < 0) 1353370a11b8SQu Wenruo break; 1354824d8dffSQu Wenruo /* 13555d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 13565d4f98a2SYan Zheng */ 13575d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 13585d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 13595d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 13605d4f98a2SYan Zheng 13615d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 13625d4f98a2SYan Zheng path->slots[level], old_bytenr); 13635d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 13645d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 13655d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 13665d4f98a2SYan Zheng 136782fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 136882fa113fSQu Wenruo blocksize, path->nodes[level]->start); 136982fa113fSQu Wenruo ref.skip_qgroup = true; 137082fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 137182fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 1372253e258cSJosef Bacik if (ret) { 1373253e258cSJosef Bacik btrfs_abort_transaction(trans, ret); 1374253e258cSJosef Bacik break; 1375253e258cSJosef Bacik } 137682fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 137782fa113fSQu Wenruo blocksize, 0); 137882fa113fSQu Wenruo ref.skip_qgroup = true; 137982fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 138082fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 1381253e258cSJosef Bacik if (ret) { 1382253e258cSJosef Bacik btrfs_abort_transaction(trans, ret); 1383253e258cSJosef Bacik break; 1384253e258cSJosef Bacik } 13855d4f98a2SYan Zheng 1386ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 1387ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 1388ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 1389ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1390ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 1391253e258cSJosef Bacik if (ret) { 1392253e258cSJosef Bacik btrfs_abort_transaction(trans, ret); 1393253e258cSJosef Bacik break; 1394253e258cSJosef Bacik } 13955d4f98a2SYan Zheng 1396ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 1397ffd4bb2aSQu Wenruo blocksize, 0); 1398ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 1399ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 1400ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 1401253e258cSJosef Bacik if (ret) { 1402253e258cSJosef Bacik btrfs_abort_transaction(trans, ret); 1403253e258cSJosef Bacik break; 1404253e258cSJosef Bacik } 14055d4f98a2SYan Zheng 14065d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 14075d4f98a2SYan Zheng 14085d4f98a2SYan Zheng ret = level; 14095d4f98a2SYan Zheng break; 14105d4f98a2SYan Zheng } 14115d4f98a2SYan Zheng btrfs_tree_unlock(parent); 14125d4f98a2SYan Zheng free_extent_buffer(parent); 14135d4f98a2SYan Zheng return ret; 14145d4f98a2SYan Zheng } 14155d4f98a2SYan Zheng 14165d4f98a2SYan Zheng /* 14175d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 14185d4f98a2SYan Zheng */ 14195d4f98a2SYan Zheng static noinline_for_stack 14205d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14215d4f98a2SYan Zheng int *level) 14225d4f98a2SYan Zheng { 14235d4f98a2SYan Zheng struct extent_buffer *eb; 14245d4f98a2SYan Zheng int i; 14255d4f98a2SYan Zheng u64 last_snapshot; 14265d4f98a2SYan Zheng u32 nritems; 14275d4f98a2SYan Zheng 14285d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14295d4f98a2SYan Zheng 14305d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 14315d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14325d4f98a2SYan Zheng path->nodes[i] = NULL; 14335d4f98a2SYan Zheng } 14345d4f98a2SYan Zheng 14355d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 14365d4f98a2SYan Zheng eb = path->nodes[i]; 14375d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14385d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 14395d4f98a2SYan Zheng path->slots[i]++; 14405d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 14415d4f98a2SYan Zheng last_snapshot) 14425d4f98a2SYan Zheng continue; 14435d4f98a2SYan Zheng 14445d4f98a2SYan Zheng *level = i; 14455d4f98a2SYan Zheng return 0; 14465d4f98a2SYan Zheng } 14475d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14485d4f98a2SYan Zheng path->nodes[i] = NULL; 14495d4f98a2SYan Zheng } 14505d4f98a2SYan Zheng return 1; 14515d4f98a2SYan Zheng } 14525d4f98a2SYan Zheng 14535d4f98a2SYan Zheng /* 14545d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 14555d4f98a2SYan Zheng */ 14565d4f98a2SYan Zheng static noinline_for_stack 14575d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14585d4f98a2SYan Zheng int *level) 14595d4f98a2SYan Zheng { 14605d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 14615d4f98a2SYan Zheng int i; 14625d4f98a2SYan Zheng u64 ptr_gen = 0; 14635d4f98a2SYan Zheng u64 last_snapshot; 14645d4f98a2SYan Zheng u32 nritems; 14655d4f98a2SYan Zheng 14665d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14675d4f98a2SYan Zheng 14685d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 14695d4f98a2SYan Zheng eb = path->nodes[i]; 14705d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14715d4f98a2SYan Zheng while (path->slots[i] < nritems) { 14725d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 14735d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 14745d4f98a2SYan Zheng break; 14755d4f98a2SYan Zheng path->slots[i]++; 14765d4f98a2SYan Zheng } 14775d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 14785d4f98a2SYan Zheng if (i == *level) 14795d4f98a2SYan Zheng break; 14805d4f98a2SYan Zheng *level = i + 1; 14815d4f98a2SYan Zheng return 0; 14825d4f98a2SYan Zheng } 14835d4f98a2SYan Zheng if (i == 1) { 14845d4f98a2SYan Zheng *level = i; 14855d4f98a2SYan Zheng return 0; 14865d4f98a2SYan Zheng } 14875d4f98a2SYan Zheng 14888ef385bbSJosef Bacik eb = btrfs_read_node_slot(eb, path->slots[i]); 14898ef385bbSJosef Bacik if (IS_ERR(eb)) 149064c043deSLiu Bo return PTR_ERR(eb); 14915d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 14925d4f98a2SYan Zheng path->nodes[i - 1] = eb; 14935d4f98a2SYan Zheng path->slots[i - 1] = 0; 14945d4f98a2SYan Zheng } 14955d4f98a2SYan Zheng return 1; 14965d4f98a2SYan Zheng } 14975d4f98a2SYan Zheng 14985d4f98a2SYan Zheng /* 14995d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 15005d4f98a2SYan Zheng * [min_key, max_key) 15015d4f98a2SYan Zheng */ 15025d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 15035d4f98a2SYan Zheng struct btrfs_key *min_key, 15045d4f98a2SYan Zheng struct btrfs_key *max_key) 15055d4f98a2SYan Zheng { 15060b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15075d4f98a2SYan Zheng struct inode *inode = NULL; 15085d4f98a2SYan Zheng u64 objectid; 15095d4f98a2SYan Zheng u64 start, end; 151033345d01SLi Zefan u64 ino; 15115d4f98a2SYan Zheng 15125d4f98a2SYan Zheng objectid = min_key->objectid; 15135d4f98a2SYan Zheng while (1) { 15145d4f98a2SYan Zheng cond_resched(); 15155d4f98a2SYan Zheng iput(inode); 15165d4f98a2SYan Zheng 15175d4f98a2SYan Zheng if (objectid > max_key->objectid) 15185d4f98a2SYan Zheng break; 15195d4f98a2SYan Zheng 15205d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 15215d4f98a2SYan Zheng if (!inode) 15225d4f98a2SYan Zheng break; 15234a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 15245d4f98a2SYan Zheng 152533345d01SLi Zefan if (ino > max_key->objectid) { 15265d4f98a2SYan Zheng iput(inode); 15275d4f98a2SYan Zheng break; 15285d4f98a2SYan Zheng } 15295d4f98a2SYan Zheng 153033345d01SLi Zefan objectid = ino + 1; 15315d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 15325d4f98a2SYan Zheng continue; 15335d4f98a2SYan Zheng 153433345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 15355d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 15365d4f98a2SYan Zheng continue; 15375d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 15385d4f98a2SYan Zheng start = 0; 15395d4f98a2SYan Zheng else { 15405d4f98a2SYan Zheng start = min_key->offset; 15410b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 15425d4f98a2SYan Zheng } 15435d4f98a2SYan Zheng } else { 15445d4f98a2SYan Zheng start = 0; 15455d4f98a2SYan Zheng } 15465d4f98a2SYan Zheng 154733345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 15485d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 15495d4f98a2SYan Zheng continue; 15505d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 15515d4f98a2SYan Zheng end = (u64)-1; 15525d4f98a2SYan Zheng } else { 15535d4f98a2SYan Zheng if (max_key->offset == 0) 15545d4f98a2SYan Zheng continue; 15555d4f98a2SYan Zheng end = max_key->offset; 15560b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 15575d4f98a2SYan Zheng end--; 15585d4f98a2SYan Zheng } 15595d4f98a2SYan Zheng } else { 15605d4f98a2SYan Zheng end = (u64)-1; 15615d4f98a2SYan Zheng } 15625d4f98a2SYan Zheng 15635d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 1564d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 1565dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 1566d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 15675d4f98a2SYan Zheng } 15685d4f98a2SYan Zheng return 0; 15695d4f98a2SYan Zheng } 15705d4f98a2SYan Zheng 15715d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 15725d4f98a2SYan Zheng struct btrfs_key *key) 15735d4f98a2SYan Zheng 15745d4f98a2SYan Zheng { 15755d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 15765d4f98a2SYan Zheng if (!path->nodes[level]) 15775d4f98a2SYan Zheng break; 15785d4f98a2SYan Zheng if (path->slots[level] + 1 < 15795d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 15805d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 15815d4f98a2SYan Zheng path->slots[level] + 1); 15825d4f98a2SYan Zheng return 0; 15835d4f98a2SYan Zheng } 15845d4f98a2SYan Zheng level++; 15855d4f98a2SYan Zheng } 15865d4f98a2SYan Zheng return 1; 15875d4f98a2SYan Zheng } 15885d4f98a2SYan Zheng 15895d4f98a2SYan Zheng /* 1590d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 1591d2311e69SQu Wenruo */ 1592ac54da6cSJosef Bacik static int insert_dirty_subvol(struct btrfs_trans_handle *trans, 1593d2311e69SQu Wenruo struct reloc_control *rc, 1594d2311e69SQu Wenruo struct btrfs_root *root) 1595d2311e69SQu Wenruo { 1596d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1597d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 15987934133fSJosef Bacik int ret; 1599d2311e69SQu Wenruo 1600d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 1601d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 1602d2311e69SQu Wenruo ASSERT(reloc_root); 1603d2311e69SQu Wenruo 1604d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 1605d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 1606d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 1607c8422684SDavid Sterba btrfs_set_root_drop_level(reloc_root_item, 0); 1608d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 16097934133fSJosef Bacik ret = btrfs_update_reloc_root(trans, root); 16107934133fSJosef Bacik if (ret) 16117934133fSJosef Bacik return ret; 1612d2311e69SQu Wenruo 1613d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 161400246528SJosef Bacik btrfs_grab_root(root); 1615d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 1616d2311e69SQu Wenruo } 1617ac54da6cSJosef Bacik 1618ac54da6cSJosef Bacik return 0; 1619d2311e69SQu Wenruo } 1620d2311e69SQu Wenruo 1621d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 1622d2311e69SQu Wenruo { 1623d2311e69SQu Wenruo struct btrfs_root *root; 1624d2311e69SQu Wenruo struct btrfs_root *next; 1625d2311e69SQu Wenruo int ret = 0; 162630d40577SQu Wenruo int ret2; 1627d2311e69SQu Wenruo 1628d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 1629d2311e69SQu Wenruo reloc_dirty_list) { 163030d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 163130d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 1632d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 1633d2311e69SQu Wenruo 1634d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 1635d2311e69SQu Wenruo root->reloc_root = NULL; 16366282675eSQu Wenruo /* 16376282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 16386282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 16396282675eSQu Wenruo */ 16406282675eSQu Wenruo smp_wmb(); 16411fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 1642f28de8d8SJosef Bacik if (reloc_root) { 1643f44deb74SJosef Bacik /* 1644f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 1645f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 1646f44deb74SJosef Bacik * drop the ref ourselves. 1647f44deb74SJosef Bacik */ 1648f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 1649f44deb74SJosef Bacik if (ret2 < 0) { 1650f44deb74SJosef Bacik btrfs_put_root(reloc_root); 1651f44deb74SJosef Bacik if (!ret) 1652f28de8d8SJosef Bacik ret = ret2; 1653f28de8d8SJosef Bacik } 1654f44deb74SJosef Bacik } 165500246528SJosef Bacik btrfs_put_root(root); 165630d40577SQu Wenruo } else { 165730d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 16580078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 1659f44deb74SJosef Bacik if (ret2 < 0) { 1660f44deb74SJosef Bacik btrfs_put_root(root); 1661f44deb74SJosef Bacik if (!ret) 166230d40577SQu Wenruo ret = ret2; 166330d40577SQu Wenruo } 1664d2311e69SQu Wenruo } 1665f44deb74SJosef Bacik } 1666d2311e69SQu Wenruo return ret; 1667d2311e69SQu Wenruo } 1668d2311e69SQu Wenruo 1669d2311e69SQu Wenruo /* 16705d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 16715d4f98a2SYan Zheng * fs tree. 16725d4f98a2SYan Zheng */ 16735d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 16745d4f98a2SYan Zheng struct btrfs_root *root) 16755d4f98a2SYan Zheng { 16760b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 16775d4f98a2SYan Zheng struct btrfs_key key; 16785d4f98a2SYan Zheng struct btrfs_key next_key; 16799e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 16805d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16815d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16825d4f98a2SYan Zheng struct btrfs_path *path; 16833fd0a558SYan, Zheng struct extent_buffer *leaf; 1684fca3a45dSJosef Bacik int reserve_level; 16855d4f98a2SYan Zheng int level; 16865d4f98a2SYan Zheng int max_level; 16875d4f98a2SYan Zheng int replaced = 0; 1688c6a592f2SNikolay Borisov int ret = 0; 16893fd0a558SYan, Zheng u32 min_reserved; 16905d4f98a2SYan Zheng 16915d4f98a2SYan Zheng path = btrfs_alloc_path(); 16925d4f98a2SYan Zheng if (!path) 16935d4f98a2SYan Zheng return -ENOMEM; 1694e4058b54SDavid Sterba path->reada = READA_FORWARD; 16955d4f98a2SYan Zheng 16965d4f98a2SYan Zheng reloc_root = root->reloc_root; 16975d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16985d4f98a2SYan Zheng 16995d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 17005d4f98a2SYan Zheng level = btrfs_root_level(root_item); 170167439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 17025d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 17035d4f98a2SYan Zheng path->slots[level] = 0; 17045d4f98a2SYan Zheng } else { 17055d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 17065d4f98a2SYan Zheng 1707c8422684SDavid Sterba level = btrfs_root_drop_level(root_item); 17085d4f98a2SYan Zheng BUG_ON(level == 0); 17095d4f98a2SYan Zheng path->lowest_level = level; 17105d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 171133c66f43SYan Zheng path->lowest_level = 0; 17125d4f98a2SYan Zheng if (ret < 0) { 17135d4f98a2SYan Zheng btrfs_free_path(path); 17145d4f98a2SYan Zheng return ret; 17155d4f98a2SYan Zheng } 17165d4f98a2SYan Zheng 17175d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 17185d4f98a2SYan Zheng path->slots[level]); 17195d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 17205d4f98a2SYan Zheng 17215d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 17225d4f98a2SYan Zheng } 17235d4f98a2SYan Zheng 172444d354abSQu Wenruo /* 172544d354abSQu Wenruo * In merge_reloc_root(), we modify the upper level pointer to swap the 172644d354abSQu Wenruo * tree blocks between reloc tree and subvolume tree. Thus for tree 172744d354abSQu Wenruo * block COW, we COW at most from level 1 to root level for each tree. 172844d354abSQu Wenruo * 172944d354abSQu Wenruo * Thus the needed metadata size is at most root_level * nodesize, 173044d354abSQu Wenruo * and * 2 since we have two trees to COW. 173144d354abSQu Wenruo */ 1732fca3a45dSJosef Bacik reserve_level = max_t(int, 1, btrfs_root_level(root_item)); 1733fca3a45dSJosef Bacik min_reserved = fs_info->nodesize * reserve_level * 2; 17345d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 17355d4f98a2SYan Zheng 17365d4f98a2SYan Zheng while (1) { 173708e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 173844d354abSQu Wenruo BTRFS_RESERVE_FLUSH_LIMIT); 1739c6a592f2SNikolay Borisov if (ret) 17409e6a0c52SJosef Bacik goto out; 17419e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 17429e6a0c52SJosef Bacik if (IS_ERR(trans)) { 1743c6a592f2SNikolay Borisov ret = PTR_ERR(trans); 17449e6a0c52SJosef Bacik trans = NULL; 17459e6a0c52SJosef Bacik goto out; 17469e6a0c52SJosef Bacik } 17472abc726aSJosef Bacik 17482abc726aSJosef Bacik /* 17492abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 17502abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 17512abc726aSJosef Bacik * 17522abc726aSJosef Bacik * But that's ok, we started the trans handle on our 17532abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 17542abc726aSJosef Bacik * dirty list. At commit time we'll still call 17552abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 17562abc726aSJosef Bacik * appropriately. 17572abc726aSJosef Bacik */ 17582abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 17599e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 17603fd0a558SYan, Zheng 17613fd0a558SYan, Zheng replaced = 0; 17625d4f98a2SYan Zheng max_level = level; 17635d4f98a2SYan Zheng 17645d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 1765c6a592f2SNikolay Borisov if (ret < 0) 17665d4f98a2SYan Zheng goto out; 17675d4f98a2SYan Zheng if (ret > 0) 17685d4f98a2SYan Zheng break; 17695d4f98a2SYan Zheng 17705d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 17715d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 17725d4f98a2SYan Zheng ret = 0; 17735d4f98a2SYan Zheng } else { 17743d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 17753fd0a558SYan, Zheng &next_key, level, max_level); 17765d4f98a2SYan Zheng } 1777c6a592f2SNikolay Borisov if (ret < 0) 17785d4f98a2SYan Zheng goto out; 17795d4f98a2SYan Zheng if (ret > 0) { 17805d4f98a2SYan Zheng level = ret; 17815d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17825d4f98a2SYan Zheng path->slots[level]); 17835d4f98a2SYan Zheng replaced = 1; 17845d4f98a2SYan Zheng } 17855d4f98a2SYan Zheng 17865d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 17875d4f98a2SYan Zheng if (ret > 0) 17885d4f98a2SYan Zheng break; 17895d4f98a2SYan Zheng 17905d4f98a2SYan Zheng BUG_ON(level == 0); 17915d4f98a2SYan Zheng /* 17925d4f98a2SYan Zheng * save the merging progress in the drop_progress. 17935d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 17945d4f98a2SYan Zheng */ 17955d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 17965d4f98a2SYan Zheng path->slots[level]); 1797c8422684SDavid Sterba btrfs_set_root_drop_level(root_item, level); 17985d4f98a2SYan Zheng 17993a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 18009e6a0c52SJosef Bacik trans = NULL; 18015d4f98a2SYan Zheng 18022ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 18035d4f98a2SYan Zheng 18045d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 18055d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 18065d4f98a2SYan Zheng } 18075d4f98a2SYan Zheng 18085d4f98a2SYan Zheng /* 18095d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 18105d4f98a2SYan Zheng * relocated and the block is tree root. 18115d4f98a2SYan Zheng */ 18125d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 18139631e4ccSJosef Bacik ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf, 18149631e4ccSJosef Bacik BTRFS_NESTING_COW); 18155d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 18165d4f98a2SYan Zheng free_extent_buffer(leaf); 18175d4f98a2SYan Zheng out: 18185d4f98a2SYan Zheng btrfs_free_path(path); 18195d4f98a2SYan Zheng 1820ac54da6cSJosef Bacik if (ret == 0) { 1821ac54da6cSJosef Bacik ret = insert_dirty_subvol(trans, rc, root); 1822ac54da6cSJosef Bacik if (ret) 1823ac54da6cSJosef Bacik btrfs_abort_transaction(trans, ret); 1824ac54da6cSJosef Bacik } 18255d4f98a2SYan Zheng 18269e6a0c52SJosef Bacik if (trans) 18273a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 18285d4f98a2SYan Zheng 18292ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 18305d4f98a2SYan Zheng 18315d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 18325d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 18335d4f98a2SYan Zheng 1834c6a592f2SNikolay Borisov return ret; 18355d4f98a2SYan Zheng } 18365d4f98a2SYan Zheng 18373fd0a558SYan, Zheng static noinline_for_stack 18383fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 18395d4f98a2SYan Zheng { 18403fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 18410b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 18423fd0a558SYan, Zheng struct btrfs_root *reloc_root; 18435d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 18443fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 18453fd0a558SYan, Zheng u64 num_bytes = 0; 18463fd0a558SYan, Zheng int ret; 18473fd0a558SYan, Zheng 18480b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 18490b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 18503fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 18510b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 18527585717fSChris Mason 18533fd0a558SYan, Zheng again: 18543fd0a558SYan, Zheng if (!err) { 18553fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 185608e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 185708e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 18583fd0a558SYan, Zheng if (ret) 18593fd0a558SYan, Zheng err = ret; 18603fd0a558SYan, Zheng } 18613fd0a558SYan, Zheng 18627a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 18633612b495STsutomu Itoh if (IS_ERR(trans)) { 18643612b495STsutomu Itoh if (!err) 18652ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 186663f018beSNikolay Borisov num_bytes, NULL); 18673612b495STsutomu Itoh return PTR_ERR(trans); 18683612b495STsutomu Itoh } 18693fd0a558SYan, Zheng 18703fd0a558SYan, Zheng if (!err) { 18713fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 18723a45bb20SJeff Mahoney btrfs_end_transaction(trans); 18732ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 187463f018beSNikolay Borisov num_bytes, NULL); 18753fd0a558SYan, Zheng goto again; 18763fd0a558SYan, Zheng } 18773fd0a558SYan, Zheng } 18783fd0a558SYan, Zheng 18793fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 18803fd0a558SYan, Zheng 18813fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 18823fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 18833fd0a558SYan, Zheng struct btrfs_root, root_list); 18843fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 18853fd0a558SYan, Zheng 1886a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1887a820feb5SDavid Sterba false); 1888e0b085b0SJosef Bacik if (IS_ERR(root)) { 1889e0b085b0SJosef Bacik /* 1890e0b085b0SJosef Bacik * Even if we have an error we need this reloc root 1891e0b085b0SJosef Bacik * back on our list so we can clean up properly. 1892e0b085b0SJosef Bacik */ 1893e0b085b0SJosef Bacik list_add(&reloc_root->root_list, &reloc_roots); 1894e0b085b0SJosef Bacik btrfs_abort_transaction(trans, (int)PTR_ERR(root)); 1895e0b085b0SJosef Bacik if (!err) 1896e0b085b0SJosef Bacik err = PTR_ERR(root); 1897e0b085b0SJosef Bacik break; 1898e0b085b0SJosef Bacik } 1899e0b085b0SJosef Bacik ASSERT(root->reloc_root == reloc_root); 19003fd0a558SYan, Zheng 19013fd0a558SYan, Zheng /* 19023fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 19033fd0a558SYan, Zheng * knows it should resumes merging 19043fd0a558SYan, Zheng */ 19053fd0a558SYan, Zheng if (!err) 19063fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 1907bbae13f8SJosef Bacik ret = btrfs_update_reloc_root(trans, root); 19083fd0a558SYan, Zheng 1909bbae13f8SJosef Bacik /* 1910bbae13f8SJosef Bacik * Even if we have an error we need this reloc root back on our 1911bbae13f8SJosef Bacik * list so we can clean up properly. 1912bbae13f8SJosef Bacik */ 19133fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 191400246528SJosef Bacik btrfs_put_root(root); 1915bbae13f8SJosef Bacik 1916bbae13f8SJosef Bacik if (ret) { 1917bbae13f8SJosef Bacik btrfs_abort_transaction(trans, ret); 1918bbae13f8SJosef Bacik if (!err) 1919bbae13f8SJosef Bacik err = ret; 1920bbae13f8SJosef Bacik break; 1921bbae13f8SJosef Bacik } 19223fd0a558SYan, Zheng } 19233fd0a558SYan, Zheng 19243fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 19253fd0a558SYan, Zheng 19263fd0a558SYan, Zheng if (!err) 19273a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 19283fd0a558SYan, Zheng else 19293a45bb20SJeff Mahoney btrfs_end_transaction(trans); 19303fd0a558SYan, Zheng return err; 19313fd0a558SYan, Zheng } 19323fd0a558SYan, Zheng 19333fd0a558SYan, Zheng static noinline_for_stack 1934aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 1935aca1bba6SLiu Bo { 1936a7571232SNikolay Borisov struct btrfs_root *reloc_root, *tmp; 1937aca1bba6SLiu Bo 1938a7571232SNikolay Borisov list_for_each_entry_safe(reloc_root, tmp, list, root_list) 1939bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 1940aca1bba6SLiu Bo } 1941aca1bba6SLiu Bo 1942aca1bba6SLiu Bo static noinline_for_stack 194394404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 19443fd0a558SYan, Zheng { 19450b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 19465d4f98a2SYan Zheng struct btrfs_root *root; 19475d4f98a2SYan Zheng struct btrfs_root *reloc_root; 19483fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 19493fd0a558SYan, Zheng int found = 0; 1950aca1bba6SLiu Bo int ret = 0; 19513fd0a558SYan, Zheng again: 19523fd0a558SYan, Zheng root = rc->extent_root; 19537585717fSChris Mason 19547585717fSChris Mason /* 19557585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 19567585717fSChris Mason * we have to make sure nobody is in the middle of 19577585717fSChris Mason * adding their roots to the list while we are 19587585717fSChris Mason * doing this splice 19597585717fSChris Mason */ 19600b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 19613fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 19620b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 19635d4f98a2SYan Zheng 19643fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 19653fd0a558SYan, Zheng found = 1; 19663fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 19673fd0a558SYan, Zheng struct btrfs_root, root_list); 19685d4f98a2SYan Zheng 1969a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 1970a820feb5SDavid Sterba false); 19715d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 19725d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 19735d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 19743fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 197500246528SJosef Bacik btrfs_put_root(root); 1976b37b39cdSJosef Bacik if (ret) { 197725e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 197825e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 197925e293c2SWang Shilong &reloc_roots); 1980aca1bba6SLiu Bo goto out; 1981b37b39cdSJosef Bacik } 19823fd0a558SYan, Zheng } else { 198351415b6cSQu Wenruo if (!IS_ERR(root)) { 198451415b6cSQu Wenruo if (root->reloc_root == reloc_root) { 198551415b6cSQu Wenruo root->reloc_root = NULL; 198651415b6cSQu Wenruo btrfs_put_root(reloc_root); 198751415b6cSQu Wenruo } 19881dae7e0eSQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, 19891dae7e0eSQu Wenruo &root->state); 199051415b6cSQu Wenruo btrfs_put_root(root); 199151415b6cSQu Wenruo } 199251415b6cSQu Wenruo 19933fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 199430d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 199530d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 199630d40577SQu Wenruo &rc->dirty_subvol_roots); 19973fd0a558SYan, Zheng } 19985d4f98a2SYan Zheng } 19995d4f98a2SYan Zheng 20003fd0a558SYan, Zheng if (found) { 20013fd0a558SYan, Zheng found = 0; 20023fd0a558SYan, Zheng goto again; 20035d4f98a2SYan Zheng } 2004aca1bba6SLiu Bo out: 2005aca1bba6SLiu Bo if (ret) { 20060b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 2007aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 2008467bb1d2SWang Shilong 2009467bb1d2SWang Shilong /* new reloc root may be added */ 20100b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 2011467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 20120b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 2013467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 2014aca1bba6SLiu Bo } 2015aca1bba6SLiu Bo 20167b7b7431SJosef Bacik /* 20177b7b7431SJosef Bacik * We used to have 20187b7b7431SJosef Bacik * 20197b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 20207b7b7431SJosef Bacik * 20217b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 20227b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 20237b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 20247b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 20257b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 20267b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 20277b7b7431SJosef Bacik * will be cleaned up on unmount. 20287b7b7431SJosef Bacik * 20297b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 20307b7b7431SJosef Bacik */ 20315d4f98a2SYan Zheng } 20325d4f98a2SYan Zheng 20335d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 20345d4f98a2SYan Zheng { 20355d4f98a2SYan Zheng struct tree_block *block; 20365d4f98a2SYan Zheng struct rb_node *rb_node; 20375d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 20385d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 20395d4f98a2SYan Zheng rb_erase(rb_node, blocks); 20405d4f98a2SYan Zheng kfree(block); 20415d4f98a2SYan Zheng } 20425d4f98a2SYan Zheng } 20435d4f98a2SYan Zheng 20445d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 20455d4f98a2SYan Zheng struct btrfs_root *reloc_root) 20465d4f98a2SYan Zheng { 20470b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 20485d4f98a2SYan Zheng struct btrfs_root *root; 2049442b1ac5SJosef Bacik int ret; 20505d4f98a2SYan Zheng 20515d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 20525d4f98a2SYan Zheng return 0; 20535d4f98a2SYan Zheng 2054a820feb5SDavid Sterba root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false); 2055404bccbcSJosef Bacik 2056404bccbcSJosef Bacik /* 2057404bccbcSJosef Bacik * This should succeed, since we can't have a reloc root without having 2058404bccbcSJosef Bacik * already looked up the actual root and created the reloc root for this 2059404bccbcSJosef Bacik * root. 2060404bccbcSJosef Bacik * 2061404bccbcSJosef Bacik * However if there's some sort of corruption where we have a ref to a 2062404bccbcSJosef Bacik * reloc root without a corresponding root this could return ENOENT. 2063404bccbcSJosef Bacik */ 2064404bccbcSJosef Bacik if (IS_ERR(root)) { 2065404bccbcSJosef Bacik ASSERT(0); 2066404bccbcSJosef Bacik return PTR_ERR(root); 2067404bccbcSJosef Bacik } 2068404bccbcSJosef Bacik if (root->reloc_root != reloc_root) { 2069404bccbcSJosef Bacik ASSERT(0); 2070404bccbcSJosef Bacik btrfs_err(fs_info, 2071404bccbcSJosef Bacik "root %llu has two reloc roots associated with it", 2072404bccbcSJosef Bacik reloc_root->root_key.offset); 2073404bccbcSJosef Bacik btrfs_put_root(root); 2074404bccbcSJosef Bacik return -EUCLEAN; 2075404bccbcSJosef Bacik } 2076442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 207700246528SJosef Bacik btrfs_put_root(root); 20785d4f98a2SYan Zheng 2079442b1ac5SJosef Bacik return ret; 20805d4f98a2SYan Zheng } 20815d4f98a2SYan Zheng 20823fd0a558SYan, Zheng static noinline_for_stack 20833fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 20843fd0a558SYan, Zheng struct reloc_control *rc, 2085a26195a5SQu Wenruo struct btrfs_backref_node *node, 2086a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 20875d4f98a2SYan Zheng { 2088a26195a5SQu Wenruo struct btrfs_backref_node *next; 20895d4f98a2SYan Zheng struct btrfs_root *root; 20903fd0a558SYan, Zheng int index = 0; 209192de551bSJosef Bacik int ret; 20923fd0a558SYan, Zheng 20935d4f98a2SYan Zheng next = node; 20945d4f98a2SYan Zheng while (1) { 20955d4f98a2SYan Zheng cond_resched(); 20965d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 20975d4f98a2SYan Zheng root = next->root; 20988ee66afeSJosef Bacik 20998ee66afeSJosef Bacik /* 21008ee66afeSJosef Bacik * If there is no root, then our references for this block are 21018ee66afeSJosef Bacik * incomplete, as we should be able to walk all the way up to a 21028ee66afeSJosef Bacik * block that is owned by a root. 21038ee66afeSJosef Bacik * 21048ee66afeSJosef Bacik * This path is only for SHAREABLE roots, so if we come upon a 21058ee66afeSJosef Bacik * non-SHAREABLE root then we have backrefs that resolve 21068ee66afeSJosef Bacik * improperly. 21078ee66afeSJosef Bacik * 21088ee66afeSJosef Bacik * Both of these cases indicate file system corruption, or a bug 21098ee66afeSJosef Bacik * in the backref walking code. 21108ee66afeSJosef Bacik */ 21118ee66afeSJosef Bacik if (!root) { 21128ee66afeSJosef Bacik ASSERT(0); 21138ee66afeSJosef Bacik btrfs_err(trans->fs_info, 21148ee66afeSJosef Bacik "bytenr %llu doesn't have a backref path ending in a root", 21158ee66afeSJosef Bacik node->bytenr); 21168ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 21178ee66afeSJosef Bacik } 21188ee66afeSJosef Bacik if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 21198ee66afeSJosef Bacik ASSERT(0); 21208ee66afeSJosef Bacik btrfs_err(trans->fs_info, 21218ee66afeSJosef Bacik "bytenr %llu has multiple refs with one ending in a non-shareable root", 21228ee66afeSJosef Bacik node->bytenr); 21238ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 21248ee66afeSJosef Bacik } 21255d4f98a2SYan Zheng 21265d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 212792de551bSJosef Bacik ret = record_reloc_root_in_trans(trans, root); 212892de551bSJosef Bacik if (ret) 212992de551bSJosef Bacik return ERR_PTR(ret); 21305d4f98a2SYan Zheng break; 21315d4f98a2SYan Zheng } 21325d4f98a2SYan Zheng 213392de551bSJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 213492de551bSJosef Bacik if (ret) 213592de551bSJosef Bacik return ERR_PTR(ret); 21363fd0a558SYan, Zheng root = root->reloc_root; 21373fd0a558SYan, Zheng 213839200e59SJosef Bacik /* 213939200e59SJosef Bacik * We could have raced with another thread which failed, so 214039200e59SJosef Bacik * root->reloc_root may not be set, return ENOENT in this case. 214139200e59SJosef Bacik */ 214239200e59SJosef Bacik if (!root) 214339200e59SJosef Bacik return ERR_PTR(-ENOENT); 214439200e59SJosef Bacik 21453fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 21468ee66afeSJosef Bacik /* 21478ee66afeSJosef Bacik * We just created the reloc root, so we shouldn't have 21488ee66afeSJosef Bacik * ->new_bytenr set and this shouldn't be in the changed 21498ee66afeSJosef Bacik * list. If it is then we have multiple roots pointing 21508ee66afeSJosef Bacik * at the same bytenr which indicates corruption, or 21518ee66afeSJosef Bacik * we've made a mistake in the backref walking code. 21528ee66afeSJosef Bacik */ 21538ee66afeSJosef Bacik ASSERT(next->new_bytenr == 0); 21548ee66afeSJosef Bacik ASSERT(list_empty(&next->list)); 21558ee66afeSJosef Bacik if (next->new_bytenr || !list_empty(&next->list)) { 21568ee66afeSJosef Bacik btrfs_err(trans->fs_info, 21578ee66afeSJosef Bacik "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu", 21588ee66afeSJosef Bacik node->bytenr, next->bytenr); 21598ee66afeSJosef Bacik return ERR_PTR(-EUCLEAN); 21608ee66afeSJosef Bacik } 21618ee66afeSJosef Bacik 21623fd0a558SYan, Zheng next->new_bytenr = root->node->start; 216300246528SJosef Bacik btrfs_put_root(next->root); 216400246528SJosef Bacik next->root = btrfs_grab_root(root); 21650b530bc5SJosef Bacik ASSERT(next->root); 21663fd0a558SYan, Zheng list_add_tail(&next->list, 21673fd0a558SYan, Zheng &rc->backref_cache.changed); 21689569cc20SQu Wenruo mark_block_processed(rc, next); 21695d4f98a2SYan Zheng break; 21705d4f98a2SYan Zheng } 21715d4f98a2SYan Zheng 21723fd0a558SYan, Zheng WARN_ON(1); 21735d4f98a2SYan Zheng root = NULL; 21745d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 21755d4f98a2SYan Zheng if (!next || next->level <= node->level) 21765d4f98a2SYan Zheng break; 21775d4f98a2SYan Zheng } 2178cbdc2ebcSJosef Bacik if (!root) { 2179cbdc2ebcSJosef Bacik /* 2180cbdc2ebcSJosef Bacik * This can happen if there's fs corruption or if there's a bug 2181cbdc2ebcSJosef Bacik * in the backref lookup code. 2182cbdc2ebcSJosef Bacik */ 2183cbdc2ebcSJosef Bacik ASSERT(0); 2184cbdc2ebcSJosef Bacik return ERR_PTR(-ENOENT); 2185cbdc2ebcSJosef Bacik } 21865d4f98a2SYan Zheng 21873fd0a558SYan, Zheng next = node; 21883fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 21893fd0a558SYan, Zheng while (1) { 21903fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 21913fd0a558SYan, Zheng if (--index < 0) 21923fd0a558SYan, Zheng break; 21933fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 21943fd0a558SYan, Zheng } 21955d4f98a2SYan Zheng return root; 21965d4f98a2SYan Zheng } 21975d4f98a2SYan Zheng 21983fd0a558SYan, Zheng /* 219992a7cc42SQu Wenruo * Select a tree root for relocation. 220092a7cc42SQu Wenruo * 220192a7cc42SQu Wenruo * Return NULL if the block is not shareable. We should use do_relocation() in 220292a7cc42SQu Wenruo * this case. 220392a7cc42SQu Wenruo * 220492a7cc42SQu Wenruo * Return a tree root pointer if the block is shareable. 220592a7cc42SQu Wenruo * Return -ENOENT if the block is root of reloc tree. 22063fd0a558SYan, Zheng */ 22075d4f98a2SYan Zheng static noinline_for_stack 2208a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 22095d4f98a2SYan Zheng { 2210a26195a5SQu Wenruo struct btrfs_backref_node *next; 22113fd0a558SYan, Zheng struct btrfs_root *root; 22123fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2213a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 22143fd0a558SYan, Zheng int index = 0; 22153fd0a558SYan, Zheng 22163fd0a558SYan, Zheng next = node; 22173fd0a558SYan, Zheng while (1) { 22183fd0a558SYan, Zheng cond_resched(); 22193fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 22203fd0a558SYan, Zheng root = next->root; 2221*8717cf44SJosef Bacik 2222*8717cf44SJosef Bacik /* 2223*8717cf44SJosef Bacik * This can occur if we have incomplete extent refs leading all 2224*8717cf44SJosef Bacik * the way up a particular path, in this case return -EUCLEAN. 2225*8717cf44SJosef Bacik */ 2226*8717cf44SJosef Bacik if (!root) 2227*8717cf44SJosef Bacik return ERR_PTR(-EUCLEAN); 22283fd0a558SYan, Zheng 222992a7cc42SQu Wenruo /* No other choice for non-shareable tree */ 223092a7cc42SQu Wenruo if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 22313fd0a558SYan, Zheng return root; 22323fd0a558SYan, Zheng 22333fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 22343fd0a558SYan, Zheng fs_root = root; 22353fd0a558SYan, Zheng 22363fd0a558SYan, Zheng if (next != node) 22373fd0a558SYan, Zheng return NULL; 22383fd0a558SYan, Zheng 22393fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 22403fd0a558SYan, Zheng if (!next || next->level <= node->level) 22413fd0a558SYan, Zheng break; 22423fd0a558SYan, Zheng } 22433fd0a558SYan, Zheng 22443fd0a558SYan, Zheng if (!fs_root) 22453fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 22463fd0a558SYan, Zheng return fs_root; 22475d4f98a2SYan Zheng } 22485d4f98a2SYan Zheng 22495d4f98a2SYan Zheng static noinline_for_stack 22503fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2251a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 22525d4f98a2SYan Zheng { 22530b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2254a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2255a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2256a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 22573fd0a558SYan, Zheng u64 num_bytes = 0; 22583fd0a558SYan, Zheng int index = 0; 22595d4f98a2SYan Zheng 22603fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 22613fd0a558SYan, Zheng 22623fd0a558SYan, Zheng while (next) { 22633fd0a558SYan, Zheng cond_resched(); 22645d4f98a2SYan Zheng while (1) { 22653fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 22665d4f98a2SYan Zheng break; 22675d4f98a2SYan Zheng 22680b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 22693fd0a558SYan, Zheng 22703fd0a558SYan, Zheng if (list_empty(&next->upper)) 22713fd0a558SYan, Zheng break; 22723fd0a558SYan, Zheng 22733fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2274a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 22753fd0a558SYan, Zheng edges[index++] = edge; 22763fd0a558SYan, Zheng next = edge->node[UPPER]; 22775d4f98a2SYan Zheng } 22783fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 22793fd0a558SYan, Zheng } 22803fd0a558SYan, Zheng return num_bytes; 22813fd0a558SYan, Zheng } 22823fd0a558SYan, Zheng 22833fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 22843fd0a558SYan, Zheng struct reloc_control *rc, 2285a26195a5SQu Wenruo struct btrfs_backref_node *node) 22863fd0a558SYan, Zheng { 22873fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2288da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 22893fd0a558SYan, Zheng u64 num_bytes; 22903fd0a558SYan, Zheng int ret; 22910647bf56SWang Shilong u64 tmp; 22923fd0a558SYan, Zheng 22933fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 22943fd0a558SYan, Zheng 22953fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 22960647bf56SWang Shilong rc->reserved_bytes += num_bytes; 22978ca17f0fSJosef Bacik 22988ca17f0fSJosef Bacik /* 22998ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 23008ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 23018ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 23028ca17f0fSJosef Bacik */ 23030647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 23048ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 23053fd0a558SYan, Zheng if (ret) { 2306da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 23070647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 23080647bf56SWang Shilong tmp <<= 1; 23090647bf56SWang Shilong /* 23100647bf56SWang Shilong * only one thread can access block_rsv at this point, 23110647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 23120647bf56SWang Shilong * we expand more reservation size here to allow enough 231352042d8eSAndrea Gelmini * space for relocation and we will return earlier in 23140647bf56SWang Shilong * enospc case. 23150647bf56SWang Shilong */ 2316da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 23170647bf56SWang Shilong RELOCATION_RESERVED_NODES; 23188ca17f0fSJosef Bacik return -EAGAIN; 23193fd0a558SYan, Zheng } 23203fd0a558SYan, Zheng 23213fd0a558SYan, Zheng return 0; 23223fd0a558SYan, Zheng } 23233fd0a558SYan, Zheng 23245d4f98a2SYan Zheng /* 23255d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 23265d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 23275d4f98a2SYan Zheng * 23285d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 23295d4f98a2SYan Zheng * in that case this function just updates pointers. 23305d4f98a2SYan Zheng */ 23315d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 23323fd0a558SYan, Zheng struct reloc_control *rc, 2333a26195a5SQu Wenruo struct btrfs_backref_node *node, 23345d4f98a2SYan Zheng struct btrfs_key *key, 23355d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 23365d4f98a2SYan Zheng { 2337a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2338a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2339a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 23405d4f98a2SYan Zheng struct btrfs_root *root; 23415d4f98a2SYan Zheng struct extent_buffer *eb; 23425d4f98a2SYan Zheng u32 blocksize; 23435d4f98a2SYan Zheng u64 bytenr; 23445d4f98a2SYan Zheng int slot; 23458df01fddSNikolay Borisov int ret = 0; 23465d4f98a2SYan Zheng 2347ffe30dd8SJosef Bacik /* 2348ffe30dd8SJosef Bacik * If we are lowest then this is the first time we're processing this 2349ffe30dd8SJosef Bacik * block, and thus shouldn't have an eb associated with it yet. 2350ffe30dd8SJosef Bacik */ 2351ffe30dd8SJosef Bacik ASSERT(!lowest || !node->eb); 23525d4f98a2SYan Zheng 23535d4f98a2SYan Zheng path->lowest_level = node->level + 1; 23543fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 23555d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 235682fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2357581c1760SQu Wenruo 23585d4f98a2SYan Zheng cond_resched(); 23595d4f98a2SYan Zheng 23605d4f98a2SYan Zheng upper = edge->node[UPPER]; 2361dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 2362cbdc2ebcSJosef Bacik if (IS_ERR(root)) { 2363cbdc2ebcSJosef Bacik ret = PTR_ERR(root); 2364cbdc2ebcSJosef Bacik goto next; 2365cbdc2ebcSJosef Bacik } 23665d4f98a2SYan Zheng 23673fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 23683fd0a558SYan, Zheng if (!lowest) { 2369e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 23708df01fddSNikolay Borisov if (ret < 0) 2371cbca7d59SFilipe Manana goto next; 23723fd0a558SYan, Zheng BUG_ON(ret); 23733fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 23743fd0a558SYan, Zheng if (node->eb->start == bytenr) 23753fd0a558SYan, Zheng goto next; 23763fd0a558SYan, Zheng } 2377b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 23783fd0a558SYan, Zheng } 23795d4f98a2SYan Zheng 23805d4f98a2SYan Zheng if (!upper->eb) { 23815d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 23823561b9dbSLiu Bo if (ret) { 23838df01fddSNikolay Borisov if (ret > 0) 23848df01fddSNikolay Borisov ret = -ENOENT; 23853561b9dbSLiu Bo 23863561b9dbSLiu Bo btrfs_release_path(path); 23875d4f98a2SYan Zheng break; 23885d4f98a2SYan Zheng } 23895d4f98a2SYan Zheng 23903fd0a558SYan, Zheng if (!upper->eb) { 23913fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 23923fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 23933fd0a558SYan, Zheng } else { 23943fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 23953fd0a558SYan, Zheng } 23963fd0a558SYan, Zheng 23973fd0a558SYan, Zheng upper->locked = 1; 23983fd0a558SYan, Zheng path->locks[upper->level] = 0; 23993fd0a558SYan, Zheng 24005d4f98a2SYan Zheng slot = path->slots[upper->level]; 2401b3b4aa74SDavid Sterba btrfs_release_path(path); 24025d4f98a2SYan Zheng } else { 2403e3b83361SQu Wenruo ret = btrfs_bin_search(upper->eb, key, &slot); 24048df01fddSNikolay Borisov if (ret < 0) 2405cbca7d59SFilipe Manana goto next; 24065d4f98a2SYan Zheng BUG_ON(ret); 24075d4f98a2SYan Zheng } 24085d4f98a2SYan Zheng 24095d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 24103fd0a558SYan, Zheng if (lowest) { 24114547f4d8SLiu Bo if (bytenr != node->bytenr) { 24124547f4d8SLiu Bo btrfs_err(root->fs_info, 24134547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 24144547f4d8SLiu Bo bytenr, node->bytenr, slot, 24154547f4d8SLiu Bo upper->eb->start); 24168df01fddSNikolay Borisov ret = -EIO; 24174547f4d8SLiu Bo goto next; 24184547f4d8SLiu Bo } 24195d4f98a2SYan Zheng } else { 24203fd0a558SYan, Zheng if (node->eb->start == bytenr) 24213fd0a558SYan, Zheng goto next; 24225d4f98a2SYan Zheng } 24235d4f98a2SYan Zheng 2424da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 2425c9752536SJosef Bacik eb = btrfs_read_node_slot(upper->eb, slot); 242664c043deSLiu Bo if (IS_ERR(eb)) { 24278df01fddSNikolay Borisov ret = PTR_ERR(eb); 242864c043deSLiu Bo goto next; 242997d9a8a4STsutomu Itoh } 24305d4f98a2SYan Zheng btrfs_tree_lock(eb); 24315d4f98a2SYan Zheng 24325d4f98a2SYan Zheng if (!node->eb) { 24335d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 24349631e4ccSJosef Bacik slot, &eb, BTRFS_NESTING_COW); 24353fd0a558SYan, Zheng btrfs_tree_unlock(eb); 24363fd0a558SYan, Zheng free_extent_buffer(eb); 24378df01fddSNikolay Borisov if (ret < 0) 24383fd0a558SYan, Zheng goto next; 2439ffe30dd8SJosef Bacik /* 2440ffe30dd8SJosef Bacik * We've just COWed this block, it should have updated 2441ffe30dd8SJosef Bacik * the correct backref node entry. 2442ffe30dd8SJosef Bacik */ 2443ffe30dd8SJosef Bacik ASSERT(node->eb == eb); 24445d4f98a2SYan Zheng } else { 24455d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 24465d4f98a2SYan Zheng node->eb->start); 24475d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 24485d4f98a2SYan Zheng trans->transid); 24495d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 24505d4f98a2SYan Zheng 245182fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 24525d4f98a2SYan Zheng node->eb->start, blocksize, 245382fa113fSQu Wenruo upper->eb->start); 245482fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 245582fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 245682fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 245782fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 2458eb6b7fb4SJosef Bacik if (!ret) 2459eb6b7fb4SJosef Bacik ret = btrfs_drop_subtree(trans, root, eb, 2460eb6b7fb4SJosef Bacik upper->eb); 2461eb6b7fb4SJosef Bacik if (ret) 2462eb6b7fb4SJosef Bacik btrfs_abort_transaction(trans, ret); 24635d4f98a2SYan Zheng } 24643fd0a558SYan, Zheng next: 24653fd0a558SYan, Zheng if (!upper->pending) 2466b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(upper); 24673fd0a558SYan, Zheng else 2468b0fe7078SQu Wenruo btrfs_backref_unlock_node_buffer(upper); 24698df01fddSNikolay Borisov if (ret) 24703fd0a558SYan, Zheng break; 24715d4f98a2SYan Zheng } 24723fd0a558SYan, Zheng 24738df01fddSNikolay Borisov if (!ret && node->pending) { 2474b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 24753fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 24763fd0a558SYan, Zheng node->pending = 0; 24775d4f98a2SYan Zheng } 24783fd0a558SYan, Zheng 24795d4f98a2SYan Zheng path->lowest_level = 0; 2480ffe30dd8SJosef Bacik 2481ffe30dd8SJosef Bacik /* 2482ffe30dd8SJosef Bacik * We should have allocated all of our space in the block rsv and thus 2483ffe30dd8SJosef Bacik * shouldn't ENOSPC. 2484ffe30dd8SJosef Bacik */ 2485ffe30dd8SJosef Bacik ASSERT(ret != -ENOSPC); 24868df01fddSNikolay Borisov return ret; 24875d4f98a2SYan Zheng } 24885d4f98a2SYan Zheng 24895d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 24903fd0a558SYan, Zheng struct reloc_control *rc, 2491a26195a5SQu Wenruo struct btrfs_backref_node *node, 24925d4f98a2SYan Zheng struct btrfs_path *path) 24935d4f98a2SYan Zheng { 24945d4f98a2SYan Zheng struct btrfs_key key; 24955d4f98a2SYan Zheng 24965d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 24973fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 24985d4f98a2SYan Zheng } 24995d4f98a2SYan Zheng 25005d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 25013fd0a558SYan, Zheng struct reloc_control *rc, 25023fd0a558SYan, Zheng struct btrfs_path *path, int err) 25035d4f98a2SYan Zheng { 25043fd0a558SYan, Zheng LIST_HEAD(list); 2505a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 2506a26195a5SQu Wenruo struct btrfs_backref_node *node; 25075d4f98a2SYan Zheng int level; 25085d4f98a2SYan Zheng int ret; 25095d4f98a2SYan Zheng 25105d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 25115d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 25125d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 2513a26195a5SQu Wenruo struct btrfs_backref_node, list); 25143fd0a558SYan, Zheng list_move_tail(&node->list, &list); 25153fd0a558SYan, Zheng BUG_ON(!node->pending); 25165d4f98a2SYan Zheng 25173fd0a558SYan, Zheng if (!err) { 25183fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 25195d4f98a2SYan Zheng if (ret < 0) 25205d4f98a2SYan Zheng err = ret; 25215d4f98a2SYan Zheng } 25225d4f98a2SYan Zheng } 25233fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 25243fd0a558SYan, Zheng } 25255d4f98a2SYan Zheng return err; 25265d4f98a2SYan Zheng } 25275d4f98a2SYan Zheng 25285d4f98a2SYan Zheng /* 25295d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 25305d4f98a2SYan Zheng * as processed. 25315d4f98a2SYan Zheng */ 25325d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 2533a26195a5SQu Wenruo struct btrfs_backref_node *node) 25345d4f98a2SYan Zheng { 2535a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2536a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2537a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 25385d4f98a2SYan Zheng int index = 0; 25395d4f98a2SYan Zheng 25405d4f98a2SYan Zheng while (next) { 25415d4f98a2SYan Zheng cond_resched(); 25425d4f98a2SYan Zheng while (1) { 25435d4f98a2SYan Zheng if (next->processed) 25445d4f98a2SYan Zheng break; 25455d4f98a2SYan Zheng 25469569cc20SQu Wenruo mark_block_processed(rc, next); 25475d4f98a2SYan Zheng 25485d4f98a2SYan Zheng if (list_empty(&next->upper)) 25495d4f98a2SYan Zheng break; 25505d4f98a2SYan Zheng 25515d4f98a2SYan Zheng edge = list_entry(next->upper.next, 2552a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 25535d4f98a2SYan Zheng edges[index++] = edge; 25545d4f98a2SYan Zheng next = edge->node[UPPER]; 25555d4f98a2SYan Zheng } 25565d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 25575d4f98a2SYan Zheng } 25585d4f98a2SYan Zheng } 25595d4f98a2SYan Zheng 25607476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 25615d4f98a2SYan Zheng { 2562da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 25637476dfdaSDavid Sterba 25645d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 25659655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 25665d4f98a2SYan Zheng return 1; 25675d4f98a2SYan Zheng return 0; 25685d4f98a2SYan Zheng } 25695d4f98a2SYan Zheng 25702ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 25715d4f98a2SYan Zheng struct tree_block *block) 25725d4f98a2SYan Zheng { 25735d4f98a2SYan Zheng struct extent_buffer *eb; 25745d4f98a2SYan Zheng 2575f7ba2d37SJosef Bacik eb = read_tree_block(fs_info, block->bytenr, block->owner, 2576f7ba2d37SJosef Bacik block->key.offset, block->level, NULL); 257764c043deSLiu Bo if (IS_ERR(eb)) { 257864c043deSLiu Bo return PTR_ERR(eb); 257964c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2580416bc658SJosef Bacik free_extent_buffer(eb); 2581416bc658SJosef Bacik return -EIO; 2582416bc658SJosef Bacik } 25835d4f98a2SYan Zheng if (block->level == 0) 25845d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 25855d4f98a2SYan Zheng else 25865d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 25875d4f98a2SYan Zheng free_extent_buffer(eb); 25885d4f98a2SYan Zheng block->key_ready = 1; 25895d4f98a2SYan Zheng return 0; 25905d4f98a2SYan Zheng } 25915d4f98a2SYan Zheng 25925d4f98a2SYan Zheng /* 25935d4f98a2SYan Zheng * helper function to relocate a tree block 25945d4f98a2SYan Zheng */ 25955d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 25965d4f98a2SYan Zheng struct reloc_control *rc, 2597a26195a5SQu Wenruo struct btrfs_backref_node *node, 25985d4f98a2SYan Zheng struct btrfs_key *key, 25995d4f98a2SYan Zheng struct btrfs_path *path) 26005d4f98a2SYan Zheng { 26015d4f98a2SYan Zheng struct btrfs_root *root; 26023fd0a558SYan, Zheng int ret = 0; 26035d4f98a2SYan Zheng 26043fd0a558SYan, Zheng if (!node) 26055d4f98a2SYan Zheng return 0; 26063fd0a558SYan, Zheng 26075f6b2e5cSJosef Bacik /* 26085f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 26095f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 26105f6b2e5cSJosef Bacik */ 26115f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 26125f6b2e5cSJosef Bacik if (ret) 26135f6b2e5cSJosef Bacik goto out; 26145f6b2e5cSJosef Bacik 26153fd0a558SYan, Zheng BUG_ON(node->processed); 2616147d256eSZhaolei root = select_one_root(node); 2617*8717cf44SJosef Bacik if (IS_ERR(root)) { 2618*8717cf44SJosef Bacik ret = PTR_ERR(root); 2619*8717cf44SJosef Bacik 2620*8717cf44SJosef Bacik /* See explanation in select_one_root for the -EUCLEAN case. */ 2621*8717cf44SJosef Bacik ASSERT(ret == -ENOENT); 2622*8717cf44SJosef Bacik if (ret == -ENOENT) { 2623*8717cf44SJosef Bacik ret = 0; 26243fd0a558SYan, Zheng update_processed_blocks(rc, node); 2625*8717cf44SJosef Bacik } 26263fd0a558SYan, Zheng goto out; 26275d4f98a2SYan Zheng } 26285d4f98a2SYan Zheng 26293fd0a558SYan, Zheng if (root) { 263092a7cc42SQu Wenruo if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) { 26311c7bfa15SJosef Bacik /* 26321c7bfa15SJosef Bacik * This block was the root block of a root, and this is 26331c7bfa15SJosef Bacik * the first time we're processing the block and thus it 26341c7bfa15SJosef Bacik * should not have had the ->new_bytenr modified and 26351c7bfa15SJosef Bacik * should have not been included on the changed list. 26361c7bfa15SJosef Bacik * 26371c7bfa15SJosef Bacik * However in the case of corruption we could have 26381c7bfa15SJosef Bacik * multiple refs pointing to the same block improperly, 26391c7bfa15SJosef Bacik * and thus we would trip over these checks. ASSERT() 26401c7bfa15SJosef Bacik * for the developer case, because it could indicate a 26411c7bfa15SJosef Bacik * bug in the backref code, however error out for a 26421c7bfa15SJosef Bacik * normal user in the case of corruption. 26431c7bfa15SJosef Bacik */ 26441c7bfa15SJosef Bacik ASSERT(node->new_bytenr == 0); 26451c7bfa15SJosef Bacik ASSERT(list_empty(&node->list)); 26461c7bfa15SJosef Bacik if (node->new_bytenr || !list_empty(&node->list)) { 26471c7bfa15SJosef Bacik btrfs_err(root->fs_info, 26481c7bfa15SJosef Bacik "bytenr %llu has improper references to it", 26491c7bfa15SJosef Bacik node->bytenr); 26501c7bfa15SJosef Bacik ret = -EUCLEAN; 26511c7bfa15SJosef Bacik goto out; 26521c7bfa15SJosef Bacik } 2653d18c7bd9SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 2654d18c7bd9SJosef Bacik if (ret) 2655d18c7bd9SJosef Bacik goto out; 265639200e59SJosef Bacik /* 265739200e59SJosef Bacik * Another thread could have failed, need to check if we 265839200e59SJosef Bacik * have reloc_root actually set. 265939200e59SJosef Bacik */ 266039200e59SJosef Bacik if (!root->reloc_root) { 266139200e59SJosef Bacik ret = -ENOENT; 266239200e59SJosef Bacik goto out; 266339200e59SJosef Bacik } 26643fd0a558SYan, Zheng root = root->reloc_root; 26653fd0a558SYan, Zheng node->new_bytenr = root->node->start; 266600246528SJosef Bacik btrfs_put_root(node->root); 266700246528SJosef Bacik node->root = btrfs_grab_root(root); 26680b530bc5SJosef Bacik ASSERT(node->root); 26693fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 26703fd0a558SYan, Zheng } else { 26715d4f98a2SYan Zheng path->lowest_level = node->level; 26725d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2673b3b4aa74SDavid Sterba btrfs_release_path(path); 26743fd0a558SYan, Zheng if (ret > 0) 26755d4f98a2SYan Zheng ret = 0; 26763fd0a558SYan, Zheng } 26773fd0a558SYan, Zheng if (!ret) 26783fd0a558SYan, Zheng update_processed_blocks(rc, node); 26793fd0a558SYan, Zheng } else { 26803fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 26813fd0a558SYan, Zheng } 26825d4f98a2SYan Zheng out: 26830647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 2684023acb07SQu Wenruo btrfs_backref_cleanup_node(&rc->backref_cache, node); 26855d4f98a2SYan Zheng return ret; 26865d4f98a2SYan Zheng } 26875d4f98a2SYan Zheng 26885d4f98a2SYan Zheng /* 26895d4f98a2SYan Zheng * relocate a list of blocks 26905d4f98a2SYan Zheng */ 26915d4f98a2SYan Zheng static noinline_for_stack 26925d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 26935d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 26945d4f98a2SYan Zheng { 26952ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2696a26195a5SQu Wenruo struct btrfs_backref_node *node; 26975d4f98a2SYan Zheng struct btrfs_path *path; 26985d4f98a2SYan Zheng struct tree_block *block; 269998ff7b94SQu Wenruo struct tree_block *next; 27005d4f98a2SYan Zheng int ret; 27015d4f98a2SYan Zheng int err = 0; 27025d4f98a2SYan Zheng 27035d4f98a2SYan Zheng path = btrfs_alloc_path(); 2704e1a12670SLiu Bo if (!path) { 2705e1a12670SLiu Bo err = -ENOMEM; 270634c2b290SDavid Sterba goto out_free_blocks; 2707e1a12670SLiu Bo } 27085d4f98a2SYan Zheng 270998ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 271098ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 27115d4f98a2SYan Zheng if (!block->key_ready) 2712f7ba2d37SJosef Bacik btrfs_readahead_tree_block(fs_info, block->bytenr, 2713f7ba2d37SJosef Bacik block->owner, 0, 27143fbaf258SJosef Bacik block->level); 27155d4f98a2SYan Zheng } 27165d4f98a2SYan Zheng 271798ff7b94SQu Wenruo /* Get first keys */ 271898ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 271934c2b290SDavid Sterba if (!block->key_ready) { 27202ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 272134c2b290SDavid Sterba if (err) 272234c2b290SDavid Sterba goto out_free_path; 272334c2b290SDavid Sterba } 27245d4f98a2SYan Zheng } 27255d4f98a2SYan Zheng 272698ff7b94SQu Wenruo /* Do tree relocation */ 272798ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 27283fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 27295d4f98a2SYan Zheng block->level, block->bytenr); 27305d4f98a2SYan Zheng if (IS_ERR(node)) { 27315d4f98a2SYan Zheng err = PTR_ERR(node); 27325d4f98a2SYan Zheng goto out; 27335d4f98a2SYan Zheng } 27345d4f98a2SYan Zheng 27355d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 27365d4f98a2SYan Zheng path); 27375d4f98a2SYan Zheng if (ret < 0) { 27385d4f98a2SYan Zheng err = ret; 273950dbbb71SJosef Bacik break; 27405d4f98a2SYan Zheng } 27415d4f98a2SYan Zheng } 27425d4f98a2SYan Zheng out: 27433fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 27445d4f98a2SYan Zheng 274534c2b290SDavid Sterba out_free_path: 27465d4f98a2SYan Zheng btrfs_free_path(path); 274734c2b290SDavid Sterba out_free_blocks: 2748e1a12670SLiu Bo free_block_list(blocks); 27495d4f98a2SYan Zheng return err; 27505d4f98a2SYan Zheng } 27515d4f98a2SYan Zheng 2752056d9becSNikolay Borisov static noinline_for_stack int prealloc_file_extent_cluster( 2753056d9becSNikolay Borisov struct btrfs_inode *inode, 2754efa56464SYan, Zheng struct file_extent_cluster *cluster) 2755efa56464SYan, Zheng { 2756efa56464SYan, Zheng u64 alloc_hint = 0; 2757efa56464SYan, Zheng u64 start; 2758efa56464SYan, Zheng u64 end; 2759056d9becSNikolay Borisov u64 offset = inode->index_cnt; 2760efa56464SYan, Zheng u64 num_bytes; 27614e9d0d01SNikolay Borisov int nr; 2762efa56464SYan, Zheng int ret = 0; 2763dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 2764dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 2765214e61d0SNikolay Borisov u64 cur_offset = prealloc_start; 2766efa56464SYan, Zheng 2767efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2768056d9becSNikolay Borisov ret = btrfs_alloc_data_chunk_ondemand(inode, 2769dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 2770efa56464SYan, Zheng if (ret) 2771214e61d0SNikolay Borisov return ret; 2772efa56464SYan, Zheng 277332430c61SNaohiro Aota /* 277432430c61SNaohiro Aota * On a zoned filesystem, we cannot preallocate the file region. 277532430c61SNaohiro Aota * Instead, we dirty and fiemap_write the region. 277632430c61SNaohiro Aota */ 277732430c61SNaohiro Aota if (btrfs_is_zoned(inode->root->fs_info)) { 277832430c61SNaohiro Aota struct btrfs_root *root = inode->root; 277932430c61SNaohiro Aota struct btrfs_trans_handle *trans; 278032430c61SNaohiro Aota 278132430c61SNaohiro Aota end = cluster->end - offset + 1; 278232430c61SNaohiro Aota trans = btrfs_start_transaction(root, 1); 278332430c61SNaohiro Aota if (IS_ERR(trans)) 278432430c61SNaohiro Aota return PTR_ERR(trans); 278532430c61SNaohiro Aota 278632430c61SNaohiro Aota inode->vfs_inode.i_ctime = current_time(&inode->vfs_inode); 278732430c61SNaohiro Aota i_size_write(&inode->vfs_inode, end); 278832430c61SNaohiro Aota ret = btrfs_update_inode(trans, root, inode); 278932430c61SNaohiro Aota if (ret) { 279032430c61SNaohiro Aota btrfs_abort_transaction(trans, ret); 279132430c61SNaohiro Aota btrfs_end_transaction(trans); 279232430c61SNaohiro Aota return ret; 279332430c61SNaohiro Aota } 279432430c61SNaohiro Aota 279532430c61SNaohiro Aota return btrfs_end_transaction(trans); 279632430c61SNaohiro Aota } 279732430c61SNaohiro Aota 279864708539SJosef Bacik btrfs_inode_lock(&inode->vfs_inode, 0); 27994e9d0d01SNikolay Borisov for (nr = 0; nr < cluster->nr; nr++) { 2800efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2801efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2802efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2803efa56464SYan, Zheng else 2804efa56464SYan, Zheng end = cluster->end - offset; 2805efa56464SYan, Zheng 2806056d9becSNikolay Borisov lock_extent(&inode->io_tree, start, end); 2807efa56464SYan, Zheng num_bytes = end + 1 - start; 2808056d9becSNikolay Borisov ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start, 2809efa56464SYan, Zheng num_bytes, num_bytes, 2810efa56464SYan, Zheng end + 1, &alloc_hint); 281118513091SWang Xiaoguang cur_offset = end + 1; 2812056d9becSNikolay Borisov unlock_extent(&inode->io_tree, start, end); 2813efa56464SYan, Zheng if (ret) 2814efa56464SYan, Zheng break; 2815efa56464SYan, Zheng } 281664708539SJosef Bacik btrfs_inode_unlock(&inode->vfs_inode, 0); 2817214e61d0SNikolay Borisov 281818513091SWang Xiaoguang if (cur_offset < prealloc_end) 2819056d9becSNikolay Borisov btrfs_free_reserved_data_space_noquota(inode->root->fs_info, 2820a89ef455SFilipe Manana prealloc_end + 1 - cur_offset); 2821efa56464SYan, Zheng return ret; 2822efa56464SYan, Zheng } 2823efa56464SYan, Zheng 2824efa56464SYan, Zheng static noinline_for_stack 28250257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 28260257bb82SYan, Zheng u64 block_start) 28275d4f98a2SYan Zheng { 28285d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 28295d4f98a2SYan Zheng struct extent_map *em; 28300257bb82SYan, Zheng int ret = 0; 28315d4f98a2SYan Zheng 2832172ddd60SDavid Sterba em = alloc_extent_map(); 28330257bb82SYan, Zheng if (!em) 28340257bb82SYan, Zheng return -ENOMEM; 28350257bb82SYan, Zheng 28365d4f98a2SYan Zheng em->start = start; 28370257bb82SYan, Zheng em->len = end + 1 - start; 28380257bb82SYan, Zheng em->block_len = em->len; 28390257bb82SYan, Zheng em->block_start = block_start; 28405d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 28415d4f98a2SYan Zheng 2842d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 28435d4f98a2SYan Zheng while (1) { 2844890871beSChris Mason write_lock(&em_tree->lock); 284509a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 2846890871beSChris Mason write_unlock(&em_tree->lock); 28475d4f98a2SYan Zheng if (ret != -EEXIST) { 28485d4f98a2SYan Zheng free_extent_map(em); 28495d4f98a2SYan Zheng break; 28505d4f98a2SYan Zheng } 2851dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 28525d4f98a2SYan Zheng } 2853d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 28540257bb82SYan, Zheng return ret; 28550257bb82SYan, Zheng } 28565d4f98a2SYan Zheng 2857726a3421SQu Wenruo /* 2858726a3421SQu Wenruo * Allow error injection to test balance cancellation 2859726a3421SQu Wenruo */ 28601fec12a5SJosef Bacik noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 2861726a3421SQu Wenruo { 28625cb502f4SQu Wenruo return atomic_read(&fs_info->balance_cancel_req) || 28635cb502f4SQu Wenruo fatal_signal_pending(current); 2864726a3421SQu Wenruo } 2865726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 2866726a3421SQu Wenruo 28670257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 28680257bb82SYan, Zheng struct file_extent_cluster *cluster) 28690257bb82SYan, Zheng { 28702ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 28710257bb82SYan, Zheng u64 page_start; 28720257bb82SYan, Zheng u64 page_end; 28730257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 28740257bb82SYan, Zheng unsigned long index; 28750257bb82SYan, Zheng unsigned long last_index; 28760257bb82SYan, Zheng struct page *page; 28770257bb82SYan, Zheng struct file_ra_state *ra; 28783b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 28790257bb82SYan, Zheng int nr = 0; 28800257bb82SYan, Zheng int ret = 0; 28810257bb82SYan, Zheng 28820257bb82SYan, Zheng if (!cluster->nr) 28830257bb82SYan, Zheng return 0; 28840257bb82SYan, Zheng 28850257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 28860257bb82SYan, Zheng if (!ra) 28870257bb82SYan, Zheng return -ENOMEM; 28880257bb82SYan, Zheng 2889056d9becSNikolay Borisov ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster); 28900257bb82SYan, Zheng if (ret) 2891efa56464SYan, Zheng goto out; 28920257bb82SYan, Zheng 28930257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 28940257bb82SYan, Zheng 2895efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2896efa56464SYan, Zheng cluster->end - offset, cluster->start); 2897efa56464SYan, Zheng if (ret) 2898efa56464SYan, Zheng goto out; 2899efa56464SYan, Zheng 290009cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 290109cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 29020257bb82SYan, Zheng while (index <= last_index) { 29039f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 29049f3db423SNikolay Borisov PAGE_SIZE); 2905efa56464SYan, Zheng if (ret) 2906efa56464SYan, Zheng goto out; 2907efa56464SYan, Zheng 29080257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 29090257bb82SYan, Zheng if (!page) { 29100257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 29110257bb82SYan, Zheng ra, NULL, index, 29120257bb82SYan, Zheng last_index + 1 - index); 2913a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 29143b16a4e3SJosef Bacik mask); 29150257bb82SYan, Zheng if (!page) { 2916691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 291743b18595SQu Wenruo PAGE_SIZE, true); 291844db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 29198702ba93SQu Wenruo PAGE_SIZE); 29200257bb82SYan, Zheng ret = -ENOMEM; 2921efa56464SYan, Zheng goto out; 29220257bb82SYan, Zheng } 29230257bb82SYan, Zheng } 292432443de3SQu Wenruo ret = set_page_extent_mapped(page); 292532443de3SQu Wenruo if (ret < 0) { 292632443de3SQu Wenruo btrfs_delalloc_release_metadata(BTRFS_I(inode), 292732443de3SQu Wenruo PAGE_SIZE, true); 292832443de3SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 292932443de3SQu Wenruo unlock_page(page); 293032443de3SQu Wenruo put_page(page); 293132443de3SQu Wenruo goto out; 293232443de3SQu Wenruo } 29330257bb82SYan, Zheng 29340257bb82SYan, Zheng if (PageReadahead(page)) { 29350257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 29360257bb82SYan, Zheng ra, NULL, page, index, 29370257bb82SYan, Zheng last_index + 1 - index); 29380257bb82SYan, Zheng } 29390257bb82SYan, Zheng 29400257bb82SYan, Zheng if (!PageUptodate(page)) { 29410257bb82SYan, Zheng btrfs_readpage(NULL, page); 29420257bb82SYan, Zheng lock_page(page); 29430257bb82SYan, Zheng if (!PageUptodate(page)) { 29440257bb82SYan, Zheng unlock_page(page); 294509cbfeafSKirill A. Shutemov put_page(page); 2946691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 294743b18595SQu Wenruo PAGE_SIZE, true); 29488b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 29498702ba93SQu Wenruo PAGE_SIZE); 29500257bb82SYan, Zheng ret = -EIO; 2951efa56464SYan, Zheng goto out; 29520257bb82SYan, Zheng } 29530257bb82SYan, Zheng } 29540257bb82SYan, Zheng 29554eee4fa4SMiao Xie page_start = page_offset(page); 295609cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 29570257bb82SYan, Zheng 2958d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 29590257bb82SYan, Zheng 29600257bb82SYan, Zheng if (nr < cluster->nr && 29610257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 29620257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 29630257bb82SYan, Zheng page_start, page_end, 2964ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 29650257bb82SYan, Zheng nr++; 29660257bb82SYan, Zheng } 29670257bb82SYan, Zheng 2968c2566f22SNikolay Borisov ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, 2969c2566f22SNikolay Borisov page_end, 0, NULL); 2970765f3cebSNikolay Borisov if (ret) { 2971765f3cebSNikolay Borisov unlock_page(page); 2972765f3cebSNikolay Borisov put_page(page); 2973765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 297443b18595SQu Wenruo PAGE_SIZE, true); 2975765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 29768702ba93SQu Wenruo PAGE_SIZE); 2977765f3cebSNikolay Borisov 2978765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 2979765f3cebSNikolay Borisov page_start, page_end, 2980765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 2981765f3cebSNikolay Borisov goto out; 2982765f3cebSNikolay Borisov 2983765f3cebSNikolay Borisov } 29840257bb82SYan, Zheng set_page_dirty(page); 29850257bb82SYan, Zheng 29860257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 2987d0082371SJeff Mahoney page_start, page_end); 29880257bb82SYan, Zheng unlock_page(page); 298909cbfeafSKirill A. Shutemov put_page(page); 29900257bb82SYan, Zheng 29910257bb82SYan, Zheng index++; 29928702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 2993efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 29942ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 29957f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 29967f913c7cSQu Wenruo ret = -ECANCELED; 29977f913c7cSQu Wenruo goto out; 29987f913c7cSQu Wenruo } 29990257bb82SYan, Zheng } 30000257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 300132430c61SNaohiro Aota if (btrfs_is_zoned(fs_info) && !ret) 300232430c61SNaohiro Aota ret = btrfs_wait_ordered_range(inode, 0, (u64)-1); 3003efa56464SYan, Zheng out: 30040257bb82SYan, Zheng kfree(ra); 30050257bb82SYan, Zheng return ret; 30060257bb82SYan, Zheng } 30070257bb82SYan, Zheng 30080257bb82SYan, Zheng static noinline_for_stack 30090257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 30100257bb82SYan, Zheng struct file_extent_cluster *cluster) 30110257bb82SYan, Zheng { 30120257bb82SYan, Zheng int ret; 30130257bb82SYan, Zheng 30140257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 30150257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30160257bb82SYan, Zheng if (ret) 30170257bb82SYan, Zheng return ret; 30180257bb82SYan, Zheng cluster->nr = 0; 30190257bb82SYan, Zheng } 30200257bb82SYan, Zheng 30210257bb82SYan, Zheng if (!cluster->nr) 30220257bb82SYan, Zheng cluster->start = extent_key->objectid; 30230257bb82SYan, Zheng else 30240257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 30250257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 30260257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 30270257bb82SYan, Zheng cluster->nr++; 30280257bb82SYan, Zheng 30290257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 30300257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30310257bb82SYan, Zheng if (ret) 30320257bb82SYan, Zheng return ret; 30330257bb82SYan, Zheng cluster->nr = 0; 30340257bb82SYan, Zheng } 30350257bb82SYan, Zheng return 0; 30365d4f98a2SYan Zheng } 30375d4f98a2SYan Zheng 30385d4f98a2SYan Zheng /* 30395d4f98a2SYan Zheng * helper to add a tree block to the list. 30405d4f98a2SYan Zheng * the major work is getting the generation and level of the block 30415d4f98a2SYan Zheng */ 30425d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 30435d4f98a2SYan Zheng struct btrfs_key *extent_key, 30445d4f98a2SYan Zheng struct btrfs_path *path, 30455d4f98a2SYan Zheng struct rb_root *blocks) 30465d4f98a2SYan Zheng { 30475d4f98a2SYan Zheng struct extent_buffer *eb; 30485d4f98a2SYan Zheng struct btrfs_extent_item *ei; 30495d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 30505d4f98a2SYan Zheng struct tree_block *block; 30515d4f98a2SYan Zheng struct rb_node *rb_node; 30525d4f98a2SYan Zheng u32 item_size; 30535d4f98a2SYan Zheng int level = -1; 30547fdf4b60SWang Shilong u64 generation; 3055f7ba2d37SJosef Bacik u64 owner = 0; 30565d4f98a2SYan Zheng 30575d4f98a2SYan Zheng eb = path->nodes[0]; 30585d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 30595d4f98a2SYan Zheng 30603173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 30613173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 3062f7ba2d37SJosef Bacik unsigned long ptr = 0, end; 3063f7ba2d37SJosef Bacik 30645d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 30655d4f98a2SYan Zheng struct btrfs_extent_item); 3066f7ba2d37SJosef Bacik end = (unsigned long)ei + item_size; 30673173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 30685d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 30695d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 3070f7ba2d37SJosef Bacik ptr = (unsigned long)(bi + 1); 30715d4f98a2SYan Zheng } else { 30723173a18fSJosef Bacik level = (int)extent_key->offset; 3073f7ba2d37SJosef Bacik ptr = (unsigned long)(ei + 1); 30743173a18fSJosef Bacik } 30753173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 3076f7ba2d37SJosef Bacik 3077f7ba2d37SJosef Bacik /* 3078f7ba2d37SJosef Bacik * We're reading random blocks without knowing their owner ahead 3079f7ba2d37SJosef Bacik * of time. This is ok most of the time, as all reloc roots and 3080f7ba2d37SJosef Bacik * fs roots have the same lock type. However normal trees do 3081f7ba2d37SJosef Bacik * not, and the only way to know ahead of time is to read the 3082f7ba2d37SJosef Bacik * inline ref offset. We know it's an fs root if 3083f7ba2d37SJosef Bacik * 3084f7ba2d37SJosef Bacik * 1. There's more than one ref. 3085f7ba2d37SJosef Bacik * 2. There's a SHARED_DATA_REF_KEY set. 3086f7ba2d37SJosef Bacik * 3. FULL_BACKREF is set on the flags. 3087f7ba2d37SJosef Bacik * 3088f7ba2d37SJosef Bacik * Otherwise it's safe to assume that the ref offset == the 3089f7ba2d37SJosef Bacik * owner of this block, so we can use that when calling 3090f7ba2d37SJosef Bacik * read_tree_block. 3091f7ba2d37SJosef Bacik */ 3092f7ba2d37SJosef Bacik if (btrfs_extent_refs(eb, ei) == 1 && 3093f7ba2d37SJosef Bacik !(btrfs_extent_flags(eb, ei) & 3094f7ba2d37SJosef Bacik BTRFS_BLOCK_FLAG_FULL_BACKREF) && 3095f7ba2d37SJosef Bacik ptr < end) { 3096f7ba2d37SJosef Bacik struct btrfs_extent_inline_ref *iref; 3097f7ba2d37SJosef Bacik int type; 3098f7ba2d37SJosef Bacik 3099f7ba2d37SJosef Bacik iref = (struct btrfs_extent_inline_ref *)ptr; 3100f7ba2d37SJosef Bacik type = btrfs_get_extent_inline_ref_type(eb, iref, 3101f7ba2d37SJosef Bacik BTRFS_REF_TYPE_BLOCK); 3102f7ba2d37SJosef Bacik if (type == BTRFS_REF_TYPE_INVALID) 3103f7ba2d37SJosef Bacik return -EINVAL; 3104f7ba2d37SJosef Bacik if (type == BTRFS_TREE_BLOCK_REF_KEY) 3105f7ba2d37SJosef Bacik owner = btrfs_extent_inline_ref_offset(eb, iref); 3106f7ba2d37SJosef Bacik } 31076d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3108ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3109ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3110ba3c2b19SNikolay Borisov return -EINVAL; 31113173a18fSJosef Bacik } else { 31125d4f98a2SYan Zheng BUG(); 31135d4f98a2SYan Zheng } 31145d4f98a2SYan Zheng 3115b3b4aa74SDavid Sterba btrfs_release_path(path); 31165d4f98a2SYan Zheng 31175d4f98a2SYan Zheng BUG_ON(level == -1); 31185d4f98a2SYan Zheng 31195d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 31205d4f98a2SYan Zheng if (!block) 31215d4f98a2SYan Zheng return -ENOMEM; 31225d4f98a2SYan Zheng 31235d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3124da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 31255d4f98a2SYan Zheng block->key.offset = generation; 31265d4f98a2SYan Zheng block->level = level; 31275d4f98a2SYan Zheng block->key_ready = 0; 3128f7ba2d37SJosef Bacik block->owner = owner; 31295d4f98a2SYan Zheng 3130e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 313143c04fb1SJeff Mahoney if (rb_node) 3132982c92cbSQu Wenruo btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr, 3133982c92cbSQu Wenruo -EEXIST); 31345d4f98a2SYan Zheng 31355d4f98a2SYan Zheng return 0; 31365d4f98a2SYan Zheng } 31375d4f98a2SYan Zheng 31385d4f98a2SYan Zheng /* 31395d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 31405d4f98a2SYan Zheng */ 31415d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 31425d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 31435d4f98a2SYan Zheng struct rb_root *blocks) 31445d4f98a2SYan Zheng { 31450b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31465d4f98a2SYan Zheng struct btrfs_path *path; 31475d4f98a2SYan Zheng struct btrfs_key key; 31485d4f98a2SYan Zheng int ret; 31490b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 31505d4f98a2SYan Zheng 31517476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 31525d4f98a2SYan Zheng return 0; 31535d4f98a2SYan Zheng 3154e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 31555d4f98a2SYan Zheng return 0; 31565d4f98a2SYan Zheng 31575d4f98a2SYan Zheng path = btrfs_alloc_path(); 31585d4f98a2SYan Zheng if (!path) 31595d4f98a2SYan Zheng return -ENOMEM; 3160aee68ee5SJosef Bacik again: 31615d4f98a2SYan Zheng key.objectid = bytenr; 3162aee68ee5SJosef Bacik if (skinny) { 3163aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3164aee68ee5SJosef Bacik key.offset = (u64)-1; 3165aee68ee5SJosef Bacik } else { 31665d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31675d4f98a2SYan Zheng key.offset = blocksize; 3168aee68ee5SJosef Bacik } 31695d4f98a2SYan Zheng 31705d4f98a2SYan Zheng path->search_commit_root = 1; 31715d4f98a2SYan Zheng path->skip_locking = 1; 31725d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 31735d4f98a2SYan Zheng if (ret < 0) 31745d4f98a2SYan Zheng goto out; 31755d4f98a2SYan Zheng 3176aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3177aee68ee5SJosef Bacik if (path->slots[0]) { 3178aee68ee5SJosef Bacik path->slots[0]--; 3179aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3180aee68ee5SJosef Bacik path->slots[0]); 31813173a18fSJosef Bacik if (key.objectid == bytenr && 3182aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3183aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3184aee68ee5SJosef Bacik key.offset == blocksize))) 31853173a18fSJosef Bacik ret = 0; 31863173a18fSJosef Bacik } 3187aee68ee5SJosef Bacik 3188aee68ee5SJosef Bacik if (ret) { 3189aee68ee5SJosef Bacik skinny = false; 3190aee68ee5SJosef Bacik btrfs_release_path(path); 3191aee68ee5SJosef Bacik goto again; 3192aee68ee5SJosef Bacik } 3193aee68ee5SJosef Bacik } 3194cdccee99SLiu Bo if (ret) { 3195cdccee99SLiu Bo ASSERT(ret == 1); 3196cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3197cdccee99SLiu Bo btrfs_err(fs_info, 3198cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3199cdccee99SLiu Bo bytenr); 3200cdccee99SLiu Bo WARN_ON(1); 3201cdccee99SLiu Bo ret = -EINVAL; 3202cdccee99SLiu Bo goto out; 3203cdccee99SLiu Bo } 32043173a18fSJosef Bacik 32055d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 32065d4f98a2SYan Zheng out: 32075d4f98a2SYan Zheng btrfs_free_path(path); 32085d4f98a2SYan Zheng return ret; 32095d4f98a2SYan Zheng } 32105d4f98a2SYan Zheng 32110af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 321232da5386SDavid Sterba struct btrfs_block_group *block_group, 32131bbc621eSChris Mason struct inode *inode, 32141bbc621eSChris Mason u64 ino) 32150af3d00bSJosef Bacik { 32160af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 32170af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 32180af3d00bSJosef Bacik int ret = 0; 32190af3d00bSJosef Bacik 32200af3d00bSJosef Bacik if (inode) 32210af3d00bSJosef Bacik goto truncate; 32220af3d00bSJosef Bacik 32230202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, ino, root); 32242e19f1f9SAl Viro if (IS_ERR(inode)) 32250af3d00bSJosef Bacik return -ENOENT; 32260af3d00bSJosef Bacik 32270af3d00bSJosef Bacik truncate: 32282ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 32297b61cd92SMiao Xie &fs_info->global_block_rsv); 32307b61cd92SMiao Xie if (ret) 32317b61cd92SMiao Xie goto out; 32327b61cd92SMiao Xie 32337a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 32340af3d00bSJosef Bacik if (IS_ERR(trans)) { 32353612b495STsutomu Itoh ret = PTR_ERR(trans); 32360af3d00bSJosef Bacik goto out; 32370af3d00bSJosef Bacik } 32380af3d00bSJosef Bacik 323977ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 32400af3d00bSJosef Bacik 32413a45bb20SJeff Mahoney btrfs_end_transaction(trans); 32422ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 32430af3d00bSJosef Bacik out: 32440af3d00bSJosef Bacik iput(inode); 32450af3d00bSJosef Bacik return ret; 32460af3d00bSJosef Bacik } 32470af3d00bSJosef Bacik 32485d4f98a2SYan Zheng /* 324919b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 325019b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 32515d4f98a2SYan Zheng */ 325219b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 325319b546d7SQu Wenruo struct btrfs_block_group *block_group, 325419b546d7SQu Wenruo u64 data_bytenr) 32555d4f98a2SYan Zheng { 325619b546d7SQu Wenruo u64 space_cache_ino; 325719b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 32585d4f98a2SYan Zheng struct btrfs_key key; 325919b546d7SQu Wenruo bool found = false; 326019b546d7SQu Wenruo int i; 32615d4f98a2SYan Zheng int ret; 32625d4f98a2SYan Zheng 326319b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 326419b546d7SQu Wenruo return 0; 32655d4f98a2SYan Zheng 326619b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 326750e31ef4SQu Wenruo u8 type; 326850e31ef4SQu Wenruo 326919b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 327019b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 327119b546d7SQu Wenruo continue; 327219b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 327350e31ef4SQu Wenruo type = btrfs_file_extent_type(leaf, ei); 327450e31ef4SQu Wenruo 327550e31ef4SQu Wenruo if ((type == BTRFS_FILE_EXTENT_REG || 327650e31ef4SQu Wenruo type == BTRFS_FILE_EXTENT_PREALLOC) && 327719b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 327819b546d7SQu Wenruo found = true; 327919b546d7SQu Wenruo space_cache_ino = key.objectid; 328019b546d7SQu Wenruo break; 328119b546d7SQu Wenruo } 328219b546d7SQu Wenruo } 328319b546d7SQu Wenruo if (!found) 328419b546d7SQu Wenruo return -ENOENT; 328519b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 328619b546d7SQu Wenruo space_cache_ino); 32870af3d00bSJosef Bacik return ret; 32885d4f98a2SYan Zheng } 32895d4f98a2SYan Zheng 32905d4f98a2SYan Zheng /* 32912c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 32925d4f98a2SYan Zheng */ 32935d4f98a2SYan Zheng static noinline_for_stack 32945d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 32955d4f98a2SYan Zheng struct btrfs_key *extent_key, 32965d4f98a2SYan Zheng struct btrfs_path *path, 32975d4f98a2SYan Zheng struct rb_root *blocks) 32985d4f98a2SYan Zheng { 329919b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 330019b546d7SQu Wenruo struct ulist *leaves = NULL; 330119b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 330219b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 330319b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3304647f63bdSFilipe David Borba Manana int ret = 0; 33055d4f98a2SYan Zheng 3306b3b4aa74SDavid Sterba btrfs_release_path(path); 330719b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 330819b546d7SQu Wenruo 0, &leaves, NULL, true); 330919b546d7SQu Wenruo if (ret < 0) 331019b546d7SQu Wenruo return ret; 331119b546d7SQu Wenruo 331219b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 331319b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 331419b546d7SQu Wenruo struct extent_buffer *eb; 331519b546d7SQu Wenruo 33161b7ec85eSJosef Bacik eb = read_tree_block(fs_info, ref_node->val, 0, 0, 0, NULL); 331719b546d7SQu Wenruo if (IS_ERR(eb)) { 331819b546d7SQu Wenruo ret = PTR_ERR(eb); 331919b546d7SQu Wenruo break; 332019b546d7SQu Wenruo } 332119b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 332219b546d7SQu Wenruo extent_key->objectid); 332319b546d7SQu Wenruo free_extent_buffer(eb); 332419b546d7SQu Wenruo if (ret < 0) 332519b546d7SQu Wenruo break; 332619b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 332719b546d7SQu Wenruo if (ret < 0) 332819b546d7SQu Wenruo break; 332919b546d7SQu Wenruo } 333019b546d7SQu Wenruo if (ret < 0) 33315d4f98a2SYan Zheng free_block_list(blocks); 333219b546d7SQu Wenruo ulist_free(leaves); 333319b546d7SQu Wenruo return ret; 33345d4f98a2SYan Zheng } 33355d4f98a2SYan Zheng 33365d4f98a2SYan Zheng /* 33372c016dc2SLiu Bo * helper to find next unprocessed extent 33385d4f98a2SYan Zheng */ 33395d4f98a2SYan Zheng static noinline_for_stack 3340147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 33413fd0a558SYan, Zheng struct btrfs_key *extent_key) 33425d4f98a2SYan Zheng { 33430b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 33445d4f98a2SYan Zheng struct btrfs_key key; 33455d4f98a2SYan Zheng struct extent_buffer *leaf; 33465d4f98a2SYan Zheng u64 start, end, last; 33475d4f98a2SYan Zheng int ret; 33485d4f98a2SYan Zheng 3349b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 33505d4f98a2SYan Zheng while (1) { 33515d4f98a2SYan Zheng cond_resched(); 33525d4f98a2SYan Zheng if (rc->search_start >= last) { 33535d4f98a2SYan Zheng ret = 1; 33545d4f98a2SYan Zheng break; 33555d4f98a2SYan Zheng } 33565d4f98a2SYan Zheng 33575d4f98a2SYan Zheng key.objectid = rc->search_start; 33585d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 33595d4f98a2SYan Zheng key.offset = 0; 33605d4f98a2SYan Zheng 33615d4f98a2SYan Zheng path->search_commit_root = 1; 33625d4f98a2SYan Zheng path->skip_locking = 1; 33635d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 33645d4f98a2SYan Zheng 0, 0); 33655d4f98a2SYan Zheng if (ret < 0) 33665d4f98a2SYan Zheng break; 33675d4f98a2SYan Zheng next: 33685d4f98a2SYan Zheng leaf = path->nodes[0]; 33695d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 33705d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 33715d4f98a2SYan Zheng if (ret != 0) 33725d4f98a2SYan Zheng break; 33735d4f98a2SYan Zheng leaf = path->nodes[0]; 33745d4f98a2SYan Zheng } 33755d4f98a2SYan Zheng 33765d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 33775d4f98a2SYan Zheng if (key.objectid >= last) { 33785d4f98a2SYan Zheng ret = 1; 33795d4f98a2SYan Zheng break; 33805d4f98a2SYan Zheng } 33815d4f98a2SYan Zheng 33823173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 33833173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 33843173a18fSJosef Bacik path->slots[0]++; 33853173a18fSJosef Bacik goto next; 33863173a18fSJosef Bacik } 33873173a18fSJosef Bacik 33883173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 33895d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 33905d4f98a2SYan Zheng path->slots[0]++; 33915d4f98a2SYan Zheng goto next; 33925d4f98a2SYan Zheng } 33935d4f98a2SYan Zheng 33943173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 33950b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 33963173a18fSJosef Bacik rc->search_start) { 33973173a18fSJosef Bacik path->slots[0]++; 33983173a18fSJosef Bacik goto next; 33993173a18fSJosef Bacik } 34003173a18fSJosef Bacik 34015d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 34025d4f98a2SYan Zheng key.objectid, &start, &end, 3403e6138876SJosef Bacik EXTENT_DIRTY, NULL); 34045d4f98a2SYan Zheng 34055d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3406b3b4aa74SDavid Sterba btrfs_release_path(path); 34075d4f98a2SYan Zheng rc->search_start = end + 1; 34085d4f98a2SYan Zheng } else { 34093173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 34105d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 34113173a18fSJosef Bacik else 34123173a18fSJosef Bacik rc->search_start = key.objectid + 34130b246afaSJeff Mahoney fs_info->nodesize; 34143fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 34155d4f98a2SYan Zheng return 0; 34165d4f98a2SYan Zheng } 34175d4f98a2SYan Zheng } 3418b3b4aa74SDavid Sterba btrfs_release_path(path); 34195d4f98a2SYan Zheng return ret; 34205d4f98a2SYan Zheng } 34215d4f98a2SYan Zheng 34225d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 34235d4f98a2SYan Zheng { 34245d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 34257585717fSChris Mason 34267585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 34275d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 34287585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 34295d4f98a2SYan Zheng } 34305d4f98a2SYan Zheng 34315d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 34325d4f98a2SYan Zheng { 34335d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 34347585717fSChris Mason 34357585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 34365d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 34377585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 34385d4f98a2SYan Zheng } 34395d4f98a2SYan Zheng 34403fd0a558SYan, Zheng static noinline_for_stack 34413fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 34423fd0a558SYan, Zheng { 34433fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3444ac2fabacSJosef Bacik int ret; 34453fd0a558SYan, Zheng 34462ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 344766d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 34483fd0a558SYan, Zheng if (!rc->block_rsv) 34493fd0a558SYan, Zheng return -ENOMEM; 34503fd0a558SYan, Zheng 34513fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3452b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 34533fd0a558SYan, Zheng rc->extents_found = 0; 34543fd0a558SYan, Zheng rc->nodes_relocated = 0; 34553fd0a558SYan, Zheng rc->merging_rsv_size = 0; 34560647bf56SWang Shilong rc->reserved_bytes = 0; 3457da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 34580647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3459ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3460ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3461ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3462ac2fabacSJosef Bacik if (ret) 3463ac2fabacSJosef Bacik return ret; 34643fd0a558SYan, Zheng 34653fd0a558SYan, Zheng rc->create_reloc_tree = 1; 34663fd0a558SYan, Zheng set_reloc_control(rc); 34673fd0a558SYan, Zheng 34687a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 346928818947SLiu Bo if (IS_ERR(trans)) { 347028818947SLiu Bo unset_reloc_control(rc); 347128818947SLiu Bo /* 347228818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 347328818947SLiu Bo * cleanup. And callers are responsible to free the above 347428818947SLiu Bo * block rsv. 347528818947SLiu Bo */ 347628818947SLiu Bo return PTR_ERR(trans); 347728818947SLiu Bo } 34783a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 34793fd0a558SYan, Zheng return 0; 34803fd0a558SYan, Zheng } 348176dda93cSYan, Zheng 34825d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 34835d4f98a2SYan Zheng { 34842ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 34855d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 34865d4f98a2SYan Zheng struct btrfs_key key; 34875d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 34885d4f98a2SYan Zheng struct btrfs_path *path; 34895d4f98a2SYan Zheng struct btrfs_extent_item *ei; 34905d4f98a2SYan Zheng u64 flags; 34915d4f98a2SYan Zheng int ret; 34925d4f98a2SYan Zheng int err = 0; 3493c87f08caSChris Mason int progress = 0; 34945d4f98a2SYan Zheng 34955d4f98a2SYan Zheng path = btrfs_alloc_path(); 34963fd0a558SYan, Zheng if (!path) 34975d4f98a2SYan Zheng return -ENOMEM; 3498e4058b54SDavid Sterba path->reada = READA_FORWARD; 34993fd0a558SYan, Zheng 35003fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 35013fd0a558SYan, Zheng if (ret) { 35023fd0a558SYan, Zheng err = ret; 35033fd0a558SYan, Zheng goto out_free; 35042423fdfbSJiri Slaby } 35055d4f98a2SYan Zheng 35065d4f98a2SYan Zheng while (1) { 35070647bf56SWang Shilong rc->reserved_bytes = 0; 35080647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 35090647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 35100647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 35110647bf56SWang Shilong if (ret) { 35120647bf56SWang Shilong err = ret; 35130647bf56SWang Shilong break; 35140647bf56SWang Shilong } 3515c87f08caSChris Mason progress++; 3516a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 35170f788c58SLiu Bo if (IS_ERR(trans)) { 35180f788c58SLiu Bo err = PTR_ERR(trans); 35190f788c58SLiu Bo trans = NULL; 35200f788c58SLiu Bo break; 35210f788c58SLiu Bo } 3522c87f08caSChris Mason restart: 35233fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 35243a45bb20SJeff Mahoney btrfs_end_transaction(trans); 352542a657f5SPan Bian trans = NULL; 35263fd0a558SYan, Zheng continue; 35273fd0a558SYan, Zheng } 35283fd0a558SYan, Zheng 3529147d256eSZhaolei ret = find_next_extent(rc, path, &key); 35305d4f98a2SYan Zheng if (ret < 0) 35315d4f98a2SYan Zheng err = ret; 35325d4f98a2SYan Zheng if (ret != 0) 35335d4f98a2SYan Zheng break; 35345d4f98a2SYan Zheng 35355d4f98a2SYan Zheng rc->extents_found++; 35365d4f98a2SYan Zheng 35375d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 35385d4f98a2SYan Zheng struct btrfs_extent_item); 35395d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 35405d4f98a2SYan Zheng 35415d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 35425d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 35435d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 35445d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 35455d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 35465d4f98a2SYan Zheng } else { 3547b3b4aa74SDavid Sterba btrfs_release_path(path); 35485d4f98a2SYan Zheng ret = 0; 35495d4f98a2SYan Zheng } 35505d4f98a2SYan Zheng if (ret < 0) { 35513fd0a558SYan, Zheng err = ret; 35525d4f98a2SYan Zheng break; 35535d4f98a2SYan Zheng } 35545d4f98a2SYan Zheng 35555d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 35565d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 35575d4f98a2SYan Zheng if (ret < 0) { 35583fd0a558SYan, Zheng if (ret != -EAGAIN) { 35595d4f98a2SYan Zheng err = ret; 35605d4f98a2SYan Zheng break; 35615d4f98a2SYan Zheng } 35623fd0a558SYan, Zheng rc->extents_found--; 35633fd0a558SYan, Zheng rc->search_start = key.objectid; 35643fd0a558SYan, Zheng } 35655d4f98a2SYan Zheng } 35665d4f98a2SYan Zheng 35673a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 35682ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 35693fd0a558SYan, Zheng trans = NULL; 35705d4f98a2SYan Zheng 35715d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 35725d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 35735d4f98a2SYan Zheng rc->found_file_extent = 1; 35740257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 35753fd0a558SYan, Zheng &key, &rc->cluster); 35765d4f98a2SYan Zheng if (ret < 0) { 35775d4f98a2SYan Zheng err = ret; 35785d4f98a2SYan Zheng break; 35795d4f98a2SYan Zheng } 35805d4f98a2SYan Zheng } 3581f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 3582f31ea088SQu Wenruo err = -ECANCELED; 3583f31ea088SQu Wenruo break; 3584f31ea088SQu Wenruo } 35855d4f98a2SYan Zheng } 3586c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 358743a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 35889689457bSShilong Wang if (ret == 1) { 3589c87f08caSChris Mason err = 0; 3590c87f08caSChris Mason progress = 0; 3591c87f08caSChris Mason goto restart; 3592c87f08caSChris Mason } 3593c87f08caSChris Mason } 35943fd0a558SYan, Zheng 3595b3b4aa74SDavid Sterba btrfs_release_path(path); 359691166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 35975d4f98a2SYan Zheng 35985d4f98a2SYan Zheng if (trans) { 35993a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 36002ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 36015d4f98a2SYan Zheng } 36025d4f98a2SYan Zheng 36030257bb82SYan, Zheng if (!err) { 36043fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 36053fd0a558SYan, Zheng &rc->cluster); 36060257bb82SYan, Zheng if (ret < 0) 36070257bb82SYan, Zheng err = ret; 36080257bb82SYan, Zheng } 36090257bb82SYan, Zheng 36103fd0a558SYan, Zheng rc->create_reloc_tree = 0; 36113fd0a558SYan, Zheng set_reloc_control(rc); 36120257bb82SYan, Zheng 361313fe1bdbSQu Wenruo btrfs_backref_release_cache(&rc->backref_cache); 361463f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 36155d4f98a2SYan Zheng 36167f913c7cSQu Wenruo /* 36177f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 36187f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 36197f913c7cSQu Wenruo * 36207f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 36217f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 36227f913c7cSQu Wenruo * merge_reloc_roots() 36237f913c7cSQu Wenruo */ 36243fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 36255d4f98a2SYan Zheng 36265d4f98a2SYan Zheng merge_reloc_roots(rc); 36275d4f98a2SYan Zheng 36283fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 36295d4f98a2SYan Zheng unset_reloc_control(rc); 363063f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 36315d4f98a2SYan Zheng 36325d4f98a2SYan Zheng /* get rid of pinned extents */ 36337a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 363462b99540SQu Wenruo if (IS_ERR(trans)) { 36353612b495STsutomu Itoh err = PTR_ERR(trans); 363662b99540SQu Wenruo goto out_free; 363762b99540SQu Wenruo } 36383a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 36396217b0faSJosef Bacik out_free: 3640d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 3641d2311e69SQu Wenruo if (ret < 0 && !err) 3642d2311e69SQu Wenruo err = ret; 36432ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 36443fd0a558SYan, Zheng btrfs_free_path(path); 36455d4f98a2SYan Zheng return err; 36465d4f98a2SYan Zheng } 36475d4f98a2SYan Zheng 36485d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 36490257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 36505d4f98a2SYan Zheng { 36515d4f98a2SYan Zheng struct btrfs_path *path; 36525d4f98a2SYan Zheng struct btrfs_inode_item *item; 36535d4f98a2SYan Zheng struct extent_buffer *leaf; 365432430c61SNaohiro Aota u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC; 36555d4f98a2SYan Zheng int ret; 36565d4f98a2SYan Zheng 365732430c61SNaohiro Aota if (btrfs_is_zoned(trans->fs_info)) 365832430c61SNaohiro Aota flags &= ~BTRFS_INODE_PREALLOC; 365932430c61SNaohiro Aota 36605d4f98a2SYan Zheng path = btrfs_alloc_path(); 36615d4f98a2SYan Zheng if (!path) 36625d4f98a2SYan Zheng return -ENOMEM; 36635d4f98a2SYan Zheng 36645d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 36655d4f98a2SYan Zheng if (ret) 36665d4f98a2SYan Zheng goto out; 36675d4f98a2SYan Zheng 36685d4f98a2SYan Zheng leaf = path->nodes[0]; 36695d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 3670b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 36715d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 36720257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 36735d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 367432430c61SNaohiro Aota btrfs_set_inode_flags(leaf, item, flags); 36755d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 36765d4f98a2SYan Zheng out: 36775d4f98a2SYan Zheng btrfs_free_path(path); 36785d4f98a2SYan Zheng return ret; 36795d4f98a2SYan Zheng } 36805d4f98a2SYan Zheng 3681790c1b8cSJosef Bacik static void delete_orphan_inode(struct btrfs_trans_handle *trans, 3682790c1b8cSJosef Bacik struct btrfs_root *root, u64 objectid) 3683790c1b8cSJosef Bacik { 3684790c1b8cSJosef Bacik struct btrfs_path *path; 3685790c1b8cSJosef Bacik struct btrfs_key key; 3686790c1b8cSJosef Bacik int ret = 0; 3687790c1b8cSJosef Bacik 3688790c1b8cSJosef Bacik path = btrfs_alloc_path(); 3689790c1b8cSJosef Bacik if (!path) { 3690790c1b8cSJosef Bacik ret = -ENOMEM; 3691790c1b8cSJosef Bacik goto out; 3692790c1b8cSJosef Bacik } 3693790c1b8cSJosef Bacik 3694790c1b8cSJosef Bacik key.objectid = objectid; 3695790c1b8cSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 3696790c1b8cSJosef Bacik key.offset = 0; 3697790c1b8cSJosef Bacik ret = btrfs_search_slot(trans, root, &key, path, -1, 1); 3698790c1b8cSJosef Bacik if (ret) { 3699790c1b8cSJosef Bacik if (ret > 0) 3700790c1b8cSJosef Bacik ret = -ENOENT; 3701790c1b8cSJosef Bacik goto out; 3702790c1b8cSJosef Bacik } 3703790c1b8cSJosef Bacik ret = btrfs_del_item(trans, root, path); 3704790c1b8cSJosef Bacik out: 3705790c1b8cSJosef Bacik if (ret) 3706790c1b8cSJosef Bacik btrfs_abort_transaction(trans, ret); 3707790c1b8cSJosef Bacik btrfs_free_path(path); 3708790c1b8cSJosef Bacik } 3709790c1b8cSJosef Bacik 37105d4f98a2SYan Zheng /* 37115d4f98a2SYan Zheng * helper to create inode for data relocation. 37125d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 37135d4f98a2SYan Zheng */ 37143fd0a558SYan, Zheng static noinline_for_stack 37153fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 371632da5386SDavid Sterba struct btrfs_block_group *group) 37175d4f98a2SYan Zheng { 37185d4f98a2SYan Zheng struct inode *inode = NULL; 37195d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 37205d4f98a2SYan Zheng struct btrfs_root *root; 37214624900dSZhaolei u64 objectid; 37225d4f98a2SYan Zheng int err = 0; 37235d4f98a2SYan Zheng 3724aeb935a4SQu Wenruo root = btrfs_grab_root(fs_info->data_reloc_root); 3725a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 372676deacf0SJosef Bacik if (IS_ERR(trans)) { 372700246528SJosef Bacik btrfs_put_root(root); 37283fd0a558SYan, Zheng return ERR_CAST(trans); 372976deacf0SJosef Bacik } 37305d4f98a2SYan Zheng 3731543068a2SNikolay Borisov err = btrfs_get_free_objectid(root, &objectid); 37325d4f98a2SYan Zheng if (err) 37335d4f98a2SYan Zheng goto out; 37345d4f98a2SYan Zheng 37350257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 3736790c1b8cSJosef Bacik if (err) 3737790c1b8cSJosef Bacik goto out; 37385d4f98a2SYan Zheng 37390202e83fSDavid Sterba inode = btrfs_iget(fs_info->sb, objectid, root); 3740790c1b8cSJosef Bacik if (IS_ERR(inode)) { 3741790c1b8cSJosef Bacik delete_orphan_inode(trans, root, objectid); 3742790c1b8cSJosef Bacik err = PTR_ERR(inode); 3743790c1b8cSJosef Bacik inode = NULL; 3744790c1b8cSJosef Bacik goto out; 3745790c1b8cSJosef Bacik } 3746b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 37475d4f98a2SYan Zheng 374873f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 37495d4f98a2SYan Zheng out: 375000246528SJosef Bacik btrfs_put_root(root); 37513a45bb20SJeff Mahoney btrfs_end_transaction(trans); 37522ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 37535d4f98a2SYan Zheng if (err) { 37545d4f98a2SYan Zheng if (inode) 37555d4f98a2SYan Zheng iput(inode); 37565d4f98a2SYan Zheng inode = ERR_PTR(err); 37575d4f98a2SYan Zheng } 37585d4f98a2SYan Zheng return inode; 37595d4f98a2SYan Zheng } 37605d4f98a2SYan Zheng 3761c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 37623fd0a558SYan, Zheng { 37633fd0a558SYan, Zheng struct reloc_control *rc; 37643fd0a558SYan, Zheng 37653fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 37663fd0a558SYan, Zheng if (!rc) 37673fd0a558SYan, Zheng return NULL; 37683fd0a558SYan, Zheng 37693fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 3770d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 3771584fb121SQu Wenruo btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1); 37723fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 377343eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 377443eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 37753fd0a558SYan, Zheng return rc; 37763fd0a558SYan, Zheng } 37773fd0a558SYan, Zheng 37781a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 37791a0afa0eSJosef Bacik { 37801a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 37811a0afa0eSJosef Bacik 37821a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 37831a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 37841a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 37851a0afa0eSJosef Bacik kfree(node); 37861a0afa0eSJosef Bacik 37871a0afa0eSJosef Bacik kfree(rc); 37881a0afa0eSJosef Bacik } 37891a0afa0eSJosef Bacik 37905d4f98a2SYan Zheng /* 3791ebce0e01SAdam Borowski * Print the block group being relocated 3792ebce0e01SAdam Borowski */ 3793ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 379432da5386SDavid Sterba struct btrfs_block_group *block_group) 3795ebce0e01SAdam Borowski { 3796f89e09cfSAnand Jain char buf[128] = {'\0'}; 3797ebce0e01SAdam Borowski 3798f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 3799ebce0e01SAdam Borowski 3800ebce0e01SAdam Borowski btrfs_info(fs_info, 3801ebce0e01SAdam Borowski "relocating block group %llu flags %s", 3802b3470b5dSDavid Sterba block_group->start, buf); 3803ebce0e01SAdam Borowski } 3804ebce0e01SAdam Borowski 3805430640e3SQu Wenruo static const char *stage_to_string(int stage) 3806430640e3SQu Wenruo { 3807430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 3808430640e3SQu Wenruo return "move data extents"; 3809430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 3810430640e3SQu Wenruo return "update data pointers"; 3811430640e3SQu Wenruo return "unknown"; 3812430640e3SQu Wenruo } 3813430640e3SQu Wenruo 3814ebce0e01SAdam Borowski /* 38155d4f98a2SYan Zheng * function to relocate all extents in a block group. 38165d4f98a2SYan Zheng */ 38176bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 38185d4f98a2SYan Zheng { 381932da5386SDavid Sterba struct btrfs_block_group *bg; 38206bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 38215d4f98a2SYan Zheng struct reloc_control *rc; 38220af3d00bSJosef Bacik struct inode *inode; 38230af3d00bSJosef Bacik struct btrfs_path *path; 38245d4f98a2SYan Zheng int ret; 3825f0486c68SYan, Zheng int rw = 0; 38265d4f98a2SYan Zheng int err = 0; 38275d4f98a2SYan Zheng 3828eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 3829eede2bf3SOmar Sandoval if (!bg) 3830eede2bf3SOmar Sandoval return -ENOENT; 3831eede2bf3SOmar Sandoval 3832eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 3833eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 3834eede2bf3SOmar Sandoval return -ETXTBSY; 3835eede2bf3SOmar Sandoval } 3836eede2bf3SOmar Sandoval 3837c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 3838eede2bf3SOmar Sandoval if (!rc) { 3839eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 38405d4f98a2SYan Zheng return -ENOMEM; 3841eede2bf3SOmar Sandoval } 38425d4f98a2SYan Zheng 3843f0486c68SYan, Zheng rc->extent_root = extent_root; 3844eede2bf3SOmar Sandoval rc->block_group = bg; 38455d4f98a2SYan Zheng 3846b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 3847f0486c68SYan, Zheng if (ret) { 3848f0486c68SYan, Zheng err = ret; 3849f0486c68SYan, Zheng goto out; 3850f0486c68SYan, Zheng } 3851f0486c68SYan, Zheng rw = 1; 3852f0486c68SYan, Zheng 38530af3d00bSJosef Bacik path = btrfs_alloc_path(); 38540af3d00bSJosef Bacik if (!path) { 38550af3d00bSJosef Bacik err = -ENOMEM; 38560af3d00bSJosef Bacik goto out; 38570af3d00bSJosef Bacik } 38580af3d00bSJosef Bacik 38597949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 38600af3d00bSJosef Bacik btrfs_free_path(path); 38610af3d00bSJosef Bacik 38620af3d00bSJosef Bacik if (!IS_ERR(inode)) 38631bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 38640af3d00bSJosef Bacik else 38650af3d00bSJosef Bacik ret = PTR_ERR(inode); 38660af3d00bSJosef Bacik 38670af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 38680af3d00bSJosef Bacik err = ret; 38690af3d00bSJosef Bacik goto out; 38700af3d00bSJosef Bacik } 38710af3d00bSJosef Bacik 38725d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 38735d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 38745d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 38755d4f98a2SYan Zheng rc->data_inode = NULL; 38765d4f98a2SYan Zheng goto out; 38775d4f98a2SYan Zheng } 38785d4f98a2SYan Zheng 38790b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 38805d4f98a2SYan Zheng 38819cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 3882f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 38836374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 3884b3470b5dSDavid Sterba rc->block_group->start, 3885b3470b5dSDavid Sterba rc->block_group->length); 38865d4f98a2SYan Zheng 38875d4f98a2SYan Zheng while (1) { 3888430640e3SQu Wenruo int finishes_stage; 3889430640e3SQu Wenruo 389076dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 38915d4f98a2SYan Zheng ret = relocate_block_group(rc); 389276dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 3893ff612ba7SJosef Bacik if (ret < 0) 38945d4f98a2SYan Zheng err = ret; 3895ff612ba7SJosef Bacik 3896430640e3SQu Wenruo finishes_stage = rc->stage; 3897ff612ba7SJosef Bacik /* 3898ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 3899ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 3900ff612ba7SJosef Bacik * different block group we could end up hitting the 3901ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 3902ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 3903ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 3904ff612ba7SJosef Bacik * out of the loop if we hit an error. 3905ff612ba7SJosef Bacik */ 3906ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 3907ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 3908ff612ba7SJosef Bacik (u64)-1); 3909ff612ba7SJosef Bacik if (ret) 3910ff612ba7SJosef Bacik err = ret; 3911ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 3912ff612ba7SJosef Bacik 0, -1); 3913ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 39145d4f98a2SYan Zheng } 39155d4f98a2SYan Zheng 3916ff612ba7SJosef Bacik if (err < 0) 3917ff612ba7SJosef Bacik goto out; 3918ff612ba7SJosef Bacik 39195d4f98a2SYan Zheng if (rc->extents_found == 0) 39205d4f98a2SYan Zheng break; 39215d4f98a2SYan Zheng 3922430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 3923430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 39245d4f98a2SYan Zheng } 39255d4f98a2SYan Zheng 39265d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 39275d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 3928bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 39295d4f98a2SYan Zheng out: 3930f0486c68SYan, Zheng if (err && rw) 39312ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 39325d4f98a2SYan Zheng iput(rc->data_inode); 39335d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 39341a0afa0eSJosef Bacik free_reloc_control(rc); 39355d4f98a2SYan Zheng return err; 39365d4f98a2SYan Zheng } 39375d4f98a2SYan Zheng 393876dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 393976dda93cSYan, Zheng { 39400b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 394176dda93cSYan, Zheng struct btrfs_trans_handle *trans; 394279787eaaSJeff Mahoney int ret, err; 394376dda93cSYan, Zheng 39440b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 394579787eaaSJeff Mahoney if (IS_ERR(trans)) 394679787eaaSJeff Mahoney return PTR_ERR(trans); 394776dda93cSYan, Zheng 394876dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 394976dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 3950c8422684SDavid Sterba btrfs_set_root_drop_level(&root->root_item, 0); 395176dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 39520b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 395376dda93cSYan, Zheng &root->root_key, &root->root_item); 395476dda93cSYan, Zheng 39553a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 395679787eaaSJeff Mahoney if (err) 395779787eaaSJeff Mahoney return err; 395879787eaaSJeff Mahoney return ret; 395976dda93cSYan, Zheng } 396076dda93cSYan, Zheng 39615d4f98a2SYan Zheng /* 39625d4f98a2SYan Zheng * recover relocation interrupted by system crash. 39635d4f98a2SYan Zheng * 39645d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 39655d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 39665d4f98a2SYan Zheng */ 39675d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 39685d4f98a2SYan Zheng { 39690b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 39705d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 39715d4f98a2SYan Zheng struct btrfs_key key; 39725d4f98a2SYan Zheng struct btrfs_root *fs_root; 39735d4f98a2SYan Zheng struct btrfs_root *reloc_root; 39745d4f98a2SYan Zheng struct btrfs_path *path; 39755d4f98a2SYan Zheng struct extent_buffer *leaf; 39765d4f98a2SYan Zheng struct reloc_control *rc = NULL; 39775d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 39785d4f98a2SYan Zheng int ret; 39795d4f98a2SYan Zheng int err = 0; 39805d4f98a2SYan Zheng 39815d4f98a2SYan Zheng path = btrfs_alloc_path(); 39825d4f98a2SYan Zheng if (!path) 39835d4f98a2SYan Zheng return -ENOMEM; 3984e4058b54SDavid Sterba path->reada = READA_BACK; 39855d4f98a2SYan Zheng 39865d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 39875d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 39885d4f98a2SYan Zheng key.offset = (u64)-1; 39895d4f98a2SYan Zheng 39905d4f98a2SYan Zheng while (1) { 39910b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 39925d4f98a2SYan Zheng path, 0, 0); 39935d4f98a2SYan Zheng if (ret < 0) { 39945d4f98a2SYan Zheng err = ret; 39955d4f98a2SYan Zheng goto out; 39965d4f98a2SYan Zheng } 39975d4f98a2SYan Zheng if (ret > 0) { 39985d4f98a2SYan Zheng if (path->slots[0] == 0) 39995d4f98a2SYan Zheng break; 40005d4f98a2SYan Zheng path->slots[0]--; 40015d4f98a2SYan Zheng } 40025d4f98a2SYan Zheng leaf = path->nodes[0]; 40035d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4004b3b4aa74SDavid Sterba btrfs_release_path(path); 40055d4f98a2SYan Zheng 40065d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 40075d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 40085d4f98a2SYan Zheng break; 40095d4f98a2SYan Zheng 40103dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 40115d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 40125d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 40135d4f98a2SYan Zheng goto out; 40145d4f98a2SYan Zheng } 40155d4f98a2SYan Zheng 401692a7cc42SQu Wenruo set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state); 40175d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 40185d4f98a2SYan Zheng 40195d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 4020a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, 4021a820feb5SDavid Sterba reloc_root->root_key.offset, false); 40225d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 402376dda93cSYan, Zheng ret = PTR_ERR(fs_root); 402476dda93cSYan, Zheng if (ret != -ENOENT) { 402576dda93cSYan, Zheng err = ret; 40265d4f98a2SYan Zheng goto out; 40275d4f98a2SYan Zheng } 402879787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 402979787eaaSJeff Mahoney if (ret < 0) { 403079787eaaSJeff Mahoney err = ret; 403179787eaaSJeff Mahoney goto out; 403279787eaaSJeff Mahoney } 4033932fd26dSJosef Bacik } else { 403400246528SJosef Bacik btrfs_put_root(fs_root); 403576dda93cSYan, Zheng } 40365d4f98a2SYan Zheng } 40375d4f98a2SYan Zheng 40385d4f98a2SYan Zheng if (key.offset == 0) 40395d4f98a2SYan Zheng break; 40405d4f98a2SYan Zheng 40415d4f98a2SYan Zheng key.offset--; 40425d4f98a2SYan Zheng } 4043b3b4aa74SDavid Sterba btrfs_release_path(path); 40445d4f98a2SYan Zheng 40455d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 40465d4f98a2SYan Zheng goto out; 40475d4f98a2SYan Zheng 4048c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 40495d4f98a2SYan Zheng if (!rc) { 40505d4f98a2SYan Zheng err = -ENOMEM; 40515d4f98a2SYan Zheng goto out; 40525d4f98a2SYan Zheng } 40535d4f98a2SYan Zheng 40540b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 40555d4f98a2SYan Zheng 40565d4f98a2SYan Zheng set_reloc_control(rc); 40575d4f98a2SYan Zheng 40587a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 40593612b495STsutomu Itoh if (IS_ERR(trans)) { 40603612b495STsutomu Itoh err = PTR_ERR(trans); 4061fb2d83eeSJosef Bacik goto out_unset; 40623612b495STsutomu Itoh } 40633fd0a558SYan, Zheng 40643fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 40653fd0a558SYan, Zheng 40665d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 40675d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 40685d4f98a2SYan Zheng struct btrfs_root, root_list); 40695d4f98a2SYan Zheng list_del(&reloc_root->root_list); 40705d4f98a2SYan Zheng 40715d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 40725d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 40735d4f98a2SYan Zheng &rc->reloc_roots); 40745d4f98a2SYan Zheng continue; 40755d4f98a2SYan Zheng } 40765d4f98a2SYan Zheng 4077a820feb5SDavid Sterba fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, 4078a820feb5SDavid Sterba false); 407979787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 408079787eaaSJeff Mahoney err = PTR_ERR(fs_root); 4081ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 40821402d17dSXiyu Yang btrfs_end_transaction(trans); 4083fb2d83eeSJosef Bacik goto out_unset; 408479787eaaSJeff Mahoney } 40855d4f98a2SYan Zheng 4086ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 408757a304cfSJosef Bacik ASSERT(err != -EEXIST); 40883c925863SJosef Bacik if (err) { 40893c925863SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 40903c925863SJosef Bacik btrfs_put_root(fs_root); 40913c925863SJosef Bacik btrfs_end_transaction(trans); 40923c925863SJosef Bacik goto out_unset; 40933c925863SJosef Bacik } 4094f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 409500246528SJosef Bacik btrfs_put_root(fs_root); 40965d4f98a2SYan Zheng } 40975d4f98a2SYan Zheng 40983a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 409979787eaaSJeff Mahoney if (err) 4100fb2d83eeSJosef Bacik goto out_unset; 41015d4f98a2SYan Zheng 41025d4f98a2SYan Zheng merge_reloc_roots(rc); 41035d4f98a2SYan Zheng 41045d4f98a2SYan Zheng unset_reloc_control(rc); 41055d4f98a2SYan Zheng 41067a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 410762b99540SQu Wenruo if (IS_ERR(trans)) { 41083612b495STsutomu Itoh err = PTR_ERR(trans); 41096217b0faSJosef Bacik goto out_clean; 411062b99540SQu Wenruo } 41113a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 41126217b0faSJosef Bacik out_clean: 4113d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4114d2311e69SQu Wenruo if (ret < 0 && !err) 4115d2311e69SQu Wenruo err = ret; 4116fb2d83eeSJosef Bacik out_unset: 4117fb2d83eeSJosef Bacik unset_reloc_control(rc); 41181a0afa0eSJosef Bacik free_reloc_control(rc); 41193612b495STsutomu Itoh out: 4120aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4121aca1bba6SLiu Bo 41225d4f98a2SYan Zheng btrfs_free_path(path); 41235d4f98a2SYan Zheng 41245d4f98a2SYan Zheng if (err == 0) { 41255d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 4126aeb935a4SQu Wenruo fs_root = btrfs_grab_root(fs_info->data_reloc_root); 4127aeb935a4SQu Wenruo ASSERT(fs_root); 412866b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 412900246528SJosef Bacik btrfs_put_root(fs_root); 4130932fd26dSJosef Bacik } 41315d4f98a2SYan Zheng return err; 41325d4f98a2SYan Zheng } 41335d4f98a2SYan Zheng 41345d4f98a2SYan Zheng /* 41355d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 41365d4f98a2SYan Zheng * 41375d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 41385d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 41395d4f98a2SYan Zheng */ 41407bfa9535SNikolay Borisov int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len) 41415d4f98a2SYan Zheng { 41427bfa9535SNikolay Borisov struct btrfs_fs_info *fs_info = inode->root->fs_info; 41435d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 41445d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 41455d4f98a2SYan Zheng int ret; 41465d4f98a2SYan Zheng u64 disk_bytenr; 41474577b014SJosef Bacik u64 new_bytenr; 41485d4f98a2SYan Zheng LIST_HEAD(list); 41495d4f98a2SYan Zheng 41507bfa9535SNikolay Borisov ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4151bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 41525d4f98a2SYan Zheng 41537bfa9535SNikolay Borisov disk_bytenr = file_pos + inode->index_cnt; 41540b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4155a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 415679787eaaSJeff Mahoney if (ret) 415779787eaaSJeff Mahoney goto out; 41585d4f98a2SYan Zheng 41595d4f98a2SYan Zheng while (!list_empty(&list)) { 41605d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 41615d4f98a2SYan Zheng list_del_init(&sums->list); 41625d4f98a2SYan Zheng 41634577b014SJosef Bacik /* 41644577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 41654577b014SJosef Bacik * We need to do this because we will read in entire prealloc 41664577b014SJosef Bacik * extents but we may have written to say the middle of the 41674577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 41684577b014SJosef Bacik * the right disk offset. 41694577b014SJosef Bacik * 41704577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 41714577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 41724577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 41734577b014SJosef Bacik * disk length. 41744577b014SJosef Bacik */ 4175bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 41764577b014SJosef Bacik sums->bytenr = new_bytenr; 41775d4f98a2SYan Zheng 4178f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 41795d4f98a2SYan Zheng } 418079787eaaSJeff Mahoney out: 41815d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4182411fc6bcSAndi Kleen return ret; 41835d4f98a2SYan Zheng } 41843fd0a558SYan, Zheng 418583d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 41863fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 41873fd0a558SYan, Zheng struct extent_buffer *cow) 41883fd0a558SYan, Zheng { 41890b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 41903fd0a558SYan, Zheng struct reloc_control *rc; 4191a26195a5SQu Wenruo struct btrfs_backref_node *node; 41923fd0a558SYan, Zheng int first_cow = 0; 41933fd0a558SYan, Zheng int level; 419483d4cfd4SJosef Bacik int ret = 0; 41953fd0a558SYan, Zheng 41960b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 41973fd0a558SYan, Zheng if (!rc) 419883d4cfd4SJosef Bacik return 0; 41993fd0a558SYan, Zheng 42003fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 42013fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 42023fd0a558SYan, Zheng 42033fd0a558SYan, Zheng level = btrfs_header_level(buf); 42043fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 42053fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 42063fd0a558SYan, Zheng first_cow = 1; 42073fd0a558SYan, Zheng 42083fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 42093fd0a558SYan, Zheng rc->create_reloc_tree) { 42103fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 42113fd0a558SYan, Zheng 42123fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 42133fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 42143fd0a558SYan, Zheng node->new_bytenr != buf->start); 42153fd0a558SYan, Zheng 4216b0fe7078SQu Wenruo btrfs_backref_drop_node_buffer(node); 421767439dadSDavid Sterba atomic_inc(&cow->refs); 42183fd0a558SYan, Zheng node->eb = cow; 42193fd0a558SYan, Zheng node->new_bytenr = cow->start; 42203fd0a558SYan, Zheng 42213fd0a558SYan, Zheng if (!node->pending) { 42223fd0a558SYan, Zheng list_move_tail(&node->list, 42233fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 42243fd0a558SYan, Zheng node->pending = 1; 42253fd0a558SYan, Zheng } 42263fd0a558SYan, Zheng 42273fd0a558SYan, Zheng if (first_cow) 42289569cc20SQu Wenruo mark_block_processed(rc, node); 42293fd0a558SYan, Zheng 42303fd0a558SYan, Zheng if (first_cow && level > 0) 42313fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 42323fd0a558SYan, Zheng } 42333fd0a558SYan, Zheng 423483d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 42353fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 423683d4cfd4SJosef Bacik return ret; 42373fd0a558SYan, Zheng } 42383fd0a558SYan, Zheng 42393fd0a558SYan, Zheng /* 42403fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 424101327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 42423fd0a558SYan, Zheng */ 4243147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 42443fd0a558SYan, Zheng u64 *bytes_to_reserve) 42453fd0a558SYan, Zheng { 424610995c04SQu Wenruo struct btrfs_root *root = pending->root; 424710995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 42483fd0a558SYan, Zheng 42496282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 42503fd0a558SYan, Zheng return; 42513fd0a558SYan, Zheng 42523fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 42533fd0a558SYan, Zheng return; 42543fd0a558SYan, Zheng 42553fd0a558SYan, Zheng root = root->reloc_root; 42563fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 42573fd0a558SYan, Zheng /* 42583fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 42593fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 42603fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 42613fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 42623fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 42633fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 42643fd0a558SYan, Zheng * tree may use more space than it frees. so we need 42653fd0a558SYan, Zheng * reserve extra space. 42663fd0a558SYan, Zheng */ 42673fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 42683fd0a558SYan, Zheng } 42693fd0a558SYan, Zheng 42703fd0a558SYan, Zheng /* 42713fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 42723fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4273f44deb74SJosef Bacik * 4274f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4275f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4276f44deb74SJosef Bacik * rc->reloc_roots. 42773fd0a558SYan, Zheng */ 427849b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 42793fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 42803fd0a558SYan, Zheng { 42813fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 42823fd0a558SYan, Zheng struct btrfs_root *reloc_root; 42833fd0a558SYan, Zheng struct btrfs_root *new_root; 428410995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 42853fd0a558SYan, Zheng int ret; 42863fd0a558SYan, Zheng 42876282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 428849b25e05SJeff Mahoney return 0; 42893fd0a558SYan, Zheng 42903fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 42913fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 42923fd0a558SYan, Zheng 42933fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 42943fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 42953fd0a558SYan, Zheng rc->block_rsv, 42963a584174SLu Fengqi rc->nodes_relocated, true); 429749b25e05SJeff Mahoney if (ret) 429849b25e05SJeff Mahoney return ret; 42993fd0a558SYan, Zheng } 43003fd0a558SYan, Zheng 43013fd0a558SYan, Zheng new_root = pending->snap; 43023fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 43033fd0a558SYan, Zheng new_root->root_key.objectid); 430449b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 430549b25e05SJeff Mahoney return PTR_ERR(reloc_root); 43063fd0a558SYan, Zheng 4307ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 430857a304cfSJosef Bacik ASSERT(ret != -EEXIST); 43093c925863SJosef Bacik if (ret) { 43103c925863SJosef Bacik /* Pairs with create_reloc_root */ 43113c925863SJosef Bacik btrfs_put_root(reloc_root); 43123c925863SJosef Bacik return ret; 43133c925863SJosef Bacik } 4314f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 43153fd0a558SYan, Zheng 431649b25e05SJeff Mahoney if (rc->create_reloc_tree) 43173fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 431849b25e05SJeff Mahoney return ret; 43193fd0a558SYan, Zheng } 4320