1c1d7c514SDavid Sterba // SPDX-License-Identifier: GPL-2.0 25d4f98a2SYan Zheng /* 35d4f98a2SYan Zheng * Copyright (C) 2009 Oracle. All rights reserved. 45d4f98a2SYan Zheng */ 55d4f98a2SYan Zheng 65d4f98a2SYan Zheng #include <linux/sched.h> 75d4f98a2SYan Zheng #include <linux/pagemap.h> 85d4f98a2SYan Zheng #include <linux/writeback.h> 95d4f98a2SYan Zheng #include <linux/blkdev.h> 105d4f98a2SYan Zheng #include <linux/rbtree.h> 115a0e3ad6STejun Heo #include <linux/slab.h> 12726a3421SQu Wenruo #include <linux/error-injection.h> 135d4f98a2SYan Zheng #include "ctree.h" 145d4f98a2SYan Zheng #include "disk-io.h" 155d4f98a2SYan Zheng #include "transaction.h" 165d4f98a2SYan Zheng #include "volumes.h" 175d4f98a2SYan Zheng #include "locking.h" 185d4f98a2SYan Zheng #include "btrfs_inode.h" 195d4f98a2SYan Zheng #include "async-thread.h" 200af3d00bSJosef Bacik #include "free-space-cache.h" 21581bb050SLi Zefan #include "inode-map.h" 2262b99540SQu Wenruo #include "qgroup.h" 23cdccee99SLiu Bo #include "print-tree.h" 2486736342SJosef Bacik #include "delalloc-space.h" 25aac0023cSJosef Bacik #include "block-group.h" 2619b546d7SQu Wenruo #include "backref.h" 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 750c891389SQu Wenruo /* 765d4f98a2SYan Zheng * backref_node, mapping_node and tree_block start with this 775d4f98a2SYan Zheng */ 785d4f98a2SYan Zheng struct tree_entry { 795d4f98a2SYan Zheng struct rb_node rb_node; 805d4f98a2SYan Zheng u64 bytenr; 815d4f98a2SYan Zheng }; 825d4f98a2SYan Zheng 835d4f98a2SYan Zheng /* 845d4f98a2SYan Zheng * present a tree block in the backref cache 855d4f98a2SYan Zheng */ 865d4f98a2SYan Zheng struct backref_node { 875d4f98a2SYan Zheng struct rb_node rb_node; 885d4f98a2SYan Zheng u64 bytenr; 893fd0a558SYan, Zheng 903fd0a558SYan, Zheng u64 new_bytenr; 913fd0a558SYan, Zheng /* objectid of tree block owner, can be not uptodate */ 925d4f98a2SYan Zheng u64 owner; 933fd0a558SYan, Zheng /* link to pending, changed or detached list */ 943fd0a558SYan, Zheng struct list_head list; 955d4f98a2SYan Zheng /* list of upper level blocks reference this block */ 965d4f98a2SYan Zheng struct list_head upper; 975d4f98a2SYan Zheng /* list of child blocks in the cache */ 985d4f98a2SYan Zheng struct list_head lower; 995d4f98a2SYan Zheng /* NULL if this node is not tree root */ 1005d4f98a2SYan Zheng struct btrfs_root *root; 1015d4f98a2SYan Zheng /* extent buffer got by COW the block */ 1025d4f98a2SYan Zheng struct extent_buffer *eb; 1035d4f98a2SYan Zheng /* level of tree block */ 1045d4f98a2SYan Zheng unsigned int level:8; 1053fd0a558SYan, Zheng /* is the block in non-reference counted tree */ 1063fd0a558SYan, Zheng unsigned int cowonly:1; 1073fd0a558SYan, Zheng /* 1 if no child node in the cache */ 1085d4f98a2SYan Zheng unsigned int lowest:1; 1095d4f98a2SYan Zheng /* is the extent buffer locked */ 1105d4f98a2SYan Zheng unsigned int locked:1; 1115d4f98a2SYan Zheng /* has the block been processed */ 1125d4f98a2SYan Zheng unsigned int processed:1; 1135d4f98a2SYan Zheng /* have backrefs of this block been checked */ 1145d4f98a2SYan Zheng unsigned int checked:1; 1153fd0a558SYan, Zheng /* 1163fd0a558SYan, Zheng * 1 if corresponding block has been cowed but some upper 1173fd0a558SYan, Zheng * level block pointers may not point to the new location 1183fd0a558SYan, Zheng */ 1193fd0a558SYan, Zheng unsigned int pending:1; 1203fd0a558SYan, Zheng /* 1213fd0a558SYan, Zheng * 1 if the backref node isn't connected to any other 1223fd0a558SYan, Zheng * backref node. 1233fd0a558SYan, Zheng */ 1243fd0a558SYan, Zheng unsigned int detached:1; 1252433bea5SQu Wenruo 1262433bea5SQu Wenruo /* 1272433bea5SQu Wenruo * For generic purpose backref cache, where we only care if it's a reloc 1282433bea5SQu Wenruo * root, doesn't care the source subvolid. 1292433bea5SQu Wenruo */ 1302433bea5SQu Wenruo unsigned int is_reloc_root:1; 1315d4f98a2SYan Zheng }; 1325d4f98a2SYan Zheng 1335d4f98a2SYan Zheng /* 1345d4f98a2SYan Zheng * present a block pointer in the backref cache 1355d4f98a2SYan Zheng */ 1365d4f98a2SYan Zheng struct backref_edge { 1375d4f98a2SYan Zheng struct list_head list[2]; 1385d4f98a2SYan Zheng struct backref_node *node[2]; 1395d4f98a2SYan Zheng }; 1405d4f98a2SYan Zheng 1415d4f98a2SYan Zheng #define LOWER 0 1425d4f98a2SYan Zheng #define UPPER 1 1430647bf56SWang Shilong #define RELOCATION_RESERVED_NODES 256 1445d4f98a2SYan Zheng 1455d4f98a2SYan Zheng struct backref_cache { 1465d4f98a2SYan Zheng /* red black tree of all backref nodes in the cache */ 1475d4f98a2SYan Zheng struct rb_root rb_root; 1483fd0a558SYan, Zheng /* for passing backref nodes to btrfs_reloc_cow_block */ 1493fd0a558SYan, Zheng struct backref_node *path[BTRFS_MAX_LEVEL]; 1503fd0a558SYan, Zheng /* 1513fd0a558SYan, Zheng * list of blocks that have been cowed but some block 1523fd0a558SYan, Zheng * pointers in upper level blocks may not reflect the 1533fd0a558SYan, Zheng * new location 1543fd0a558SYan, Zheng */ 1555d4f98a2SYan Zheng struct list_head pending[BTRFS_MAX_LEVEL]; 1563fd0a558SYan, Zheng /* list of backref nodes with no child node */ 1573fd0a558SYan, Zheng struct list_head leaves; 1583fd0a558SYan, Zheng /* list of blocks that have been cowed in current transaction */ 1593fd0a558SYan, Zheng struct list_head changed; 1603fd0a558SYan, Zheng /* list of detached backref node. */ 1613fd0a558SYan, Zheng struct list_head detached; 1623fd0a558SYan, Zheng 1633fd0a558SYan, Zheng u64 last_trans; 1643fd0a558SYan, Zheng 1653fd0a558SYan, Zheng int nr_nodes; 1663fd0a558SYan, Zheng int nr_edges; 16784780289SQu Wenruo 16884780289SQu Wenruo /* The list of unchecked backref edges during backref cache build */ 16984780289SQu Wenruo struct list_head pending_edge; 17084780289SQu Wenruo 17184780289SQu Wenruo /* The list of useless backref nodes during backref cache build */ 17284780289SQu Wenruo struct list_head useless_node; 17333a0f1f7SQu Wenruo 17433a0f1f7SQu Wenruo struct btrfs_fs_info *fs_info; 1752433bea5SQu Wenruo 1762433bea5SQu Wenruo /* 1772433bea5SQu Wenruo * Whether this cache is for relocation 1782433bea5SQu Wenruo * 1792433bea5SQu Wenruo * Reloction backref cache require more info for reloc root compared 1802433bea5SQu Wenruo * to generic backref cache. 1812433bea5SQu Wenruo */ 1822433bea5SQu Wenruo unsigned int is_reloc; 1835d4f98a2SYan Zheng }; 1845d4f98a2SYan Zheng 1855d4f98a2SYan Zheng /* 1865d4f98a2SYan Zheng * map address of tree root to tree 1875d4f98a2SYan Zheng */ 1885d4f98a2SYan Zheng struct mapping_node { 1895d4f98a2SYan Zheng struct rb_node rb_node; 1905d4f98a2SYan Zheng u64 bytenr; 1915d4f98a2SYan Zheng void *data; 1925d4f98a2SYan Zheng }; 1935d4f98a2SYan Zheng 1945d4f98a2SYan Zheng struct mapping_tree { 1955d4f98a2SYan Zheng struct rb_root rb_root; 1965d4f98a2SYan Zheng spinlock_t lock; 1975d4f98a2SYan Zheng }; 1985d4f98a2SYan Zheng 1995d4f98a2SYan Zheng /* 2005d4f98a2SYan Zheng * present a tree block to process 2015d4f98a2SYan Zheng */ 2025d4f98a2SYan Zheng struct tree_block { 2035d4f98a2SYan Zheng struct rb_node rb_node; 2045d4f98a2SYan Zheng u64 bytenr; 2055d4f98a2SYan Zheng struct btrfs_key key; 2065d4f98a2SYan Zheng unsigned int level:8; 2075d4f98a2SYan Zheng unsigned int key_ready:1; 2085d4f98a2SYan Zheng }; 2095d4f98a2SYan Zheng 2100257bb82SYan, Zheng #define MAX_EXTENTS 128 2110257bb82SYan, Zheng 2120257bb82SYan, Zheng struct file_extent_cluster { 2130257bb82SYan, Zheng u64 start; 2140257bb82SYan, Zheng u64 end; 2150257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 2160257bb82SYan, Zheng unsigned int nr; 2170257bb82SYan, Zheng }; 2180257bb82SYan, Zheng 2195d4f98a2SYan Zheng struct reloc_control { 2205d4f98a2SYan Zheng /* block group to relocate */ 22132da5386SDavid Sterba struct btrfs_block_group *block_group; 2225d4f98a2SYan Zheng /* extent tree */ 2235d4f98a2SYan Zheng struct btrfs_root *extent_root; 2245d4f98a2SYan Zheng /* inode for moving data */ 2255d4f98a2SYan Zheng struct inode *data_inode; 2263fd0a558SYan, Zheng 2273fd0a558SYan, Zheng struct btrfs_block_rsv *block_rsv; 2283fd0a558SYan, Zheng 2293fd0a558SYan, Zheng struct backref_cache backref_cache; 2303fd0a558SYan, Zheng 2313fd0a558SYan, Zheng struct file_extent_cluster cluster; 2325d4f98a2SYan Zheng /* tree blocks have been processed */ 2335d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 2345d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 2355d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 2365d4f98a2SYan Zheng /* list of reloc trees */ 2375d4f98a2SYan Zheng struct list_head reloc_roots; 238d2311e69SQu Wenruo /* list of subvolume trees that get relocated */ 239d2311e69SQu Wenruo struct list_head dirty_subvol_roots; 2403fd0a558SYan, Zheng /* size of metadata reservation for merging reloc trees */ 2413fd0a558SYan, Zheng u64 merging_rsv_size; 2423fd0a558SYan, Zheng /* size of relocated tree nodes */ 2433fd0a558SYan, Zheng u64 nodes_relocated; 2440647bf56SWang Shilong /* reserved size for block group relocation*/ 2450647bf56SWang Shilong u64 reserved_bytes; 2463fd0a558SYan, Zheng 2475d4f98a2SYan Zheng u64 search_start; 2485d4f98a2SYan Zheng u64 extents_found; 2493fd0a558SYan, Zheng 2503fd0a558SYan, Zheng unsigned int stage:8; 2513fd0a558SYan, Zheng unsigned int create_reloc_tree:1; 2523fd0a558SYan, Zheng unsigned int merge_reloc_tree:1; 2535d4f98a2SYan Zheng unsigned int found_file_extent:1; 2545d4f98a2SYan Zheng }; 2555d4f98a2SYan Zheng 2565d4f98a2SYan Zheng /* stages of data relocation */ 2575d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 2585d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 2595d4f98a2SYan Zheng 2603fd0a558SYan, Zheng static void remove_backref_node(struct backref_cache *cache, 2613fd0a558SYan, Zheng struct backref_node *node); 2629569cc20SQu Wenruo 2639569cc20SQu Wenruo static void mark_block_processed(struct reloc_control *rc, 2649569cc20SQu Wenruo struct backref_node *node) 2659569cc20SQu Wenruo { 2669569cc20SQu Wenruo u32 blocksize; 2679569cc20SQu Wenruo 2689569cc20SQu Wenruo if (node->level == 0 || 2699569cc20SQu Wenruo in_range(node->bytenr, rc->block_group->start, 2709569cc20SQu Wenruo rc->block_group->length)) { 2719569cc20SQu Wenruo blocksize = rc->extent_root->fs_info->nodesize; 2729569cc20SQu Wenruo set_extent_bits(&rc->processed_blocks, node->bytenr, 2739569cc20SQu Wenruo node->bytenr + blocksize - 1, EXTENT_DIRTY); 2749569cc20SQu Wenruo } 2759569cc20SQu Wenruo node->processed = 1; 2769569cc20SQu Wenruo } 2779569cc20SQu Wenruo 2785d4f98a2SYan Zheng 2795d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 2805d4f98a2SYan Zheng { 2816bef4d31SEric Paris tree->rb_root = RB_ROOT; 2825d4f98a2SYan Zheng spin_lock_init(&tree->lock); 2835d4f98a2SYan Zheng } 2845d4f98a2SYan Zheng 28533a0f1f7SQu Wenruo static void backref_cache_init(struct btrfs_fs_info *fs_info, 2862433bea5SQu Wenruo struct backref_cache *cache, int is_reloc) 2875d4f98a2SYan Zheng { 2885d4f98a2SYan Zheng int i; 2896bef4d31SEric Paris cache->rb_root = RB_ROOT; 2905d4f98a2SYan Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 2915d4f98a2SYan Zheng INIT_LIST_HEAD(&cache->pending[i]); 2923fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->changed); 2933fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->detached); 2943fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->leaves); 29584780289SQu Wenruo INIT_LIST_HEAD(&cache->pending_edge); 29684780289SQu Wenruo INIT_LIST_HEAD(&cache->useless_node); 29733a0f1f7SQu Wenruo cache->fs_info = fs_info; 2982433bea5SQu Wenruo cache->is_reloc = is_reloc; 2995d4f98a2SYan Zheng } 3005d4f98a2SYan Zheng 3013fd0a558SYan, Zheng static void backref_cache_cleanup(struct backref_cache *cache) 3025d4f98a2SYan Zheng { 3033fd0a558SYan, Zheng struct backref_node *node; 3043fd0a558SYan, Zheng int i; 3053fd0a558SYan, Zheng 3063fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 3073fd0a558SYan, Zheng node = list_entry(cache->detached.next, 3083fd0a558SYan, Zheng struct backref_node, list); 3093fd0a558SYan, Zheng remove_backref_node(cache, node); 3103fd0a558SYan, Zheng } 3113fd0a558SYan, Zheng 3123fd0a558SYan, Zheng while (!list_empty(&cache->leaves)) { 3133fd0a558SYan, Zheng node = list_entry(cache->leaves.next, 3143fd0a558SYan, Zheng struct backref_node, lower); 3153fd0a558SYan, Zheng remove_backref_node(cache, node); 3163fd0a558SYan, Zheng } 3173fd0a558SYan, Zheng 3183fd0a558SYan, Zheng cache->last_trans = 0; 3193fd0a558SYan, Zheng 3203fd0a558SYan, Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 321f4907095SLiu Bo ASSERT(list_empty(&cache->pending[i])); 32284780289SQu Wenruo ASSERT(list_empty(&cache->pending_edge)); 32384780289SQu Wenruo ASSERT(list_empty(&cache->useless_node)); 324f4907095SLiu Bo ASSERT(list_empty(&cache->changed)); 325f4907095SLiu Bo ASSERT(list_empty(&cache->detached)); 326f4907095SLiu Bo ASSERT(RB_EMPTY_ROOT(&cache->rb_root)); 327f4907095SLiu Bo ASSERT(!cache->nr_nodes); 328f4907095SLiu Bo ASSERT(!cache->nr_edges); 3293fd0a558SYan, Zheng } 3303fd0a558SYan, Zheng 3313fd0a558SYan, Zheng static struct backref_node *alloc_backref_node(struct backref_cache *cache) 3323fd0a558SYan, Zheng { 3333fd0a558SYan, Zheng struct backref_node *node; 3343fd0a558SYan, Zheng 3353fd0a558SYan, Zheng node = kzalloc(sizeof(*node), GFP_NOFS); 3363fd0a558SYan, Zheng if (node) { 3373fd0a558SYan, Zheng INIT_LIST_HEAD(&node->list); 3385d4f98a2SYan Zheng INIT_LIST_HEAD(&node->upper); 3395d4f98a2SYan Zheng INIT_LIST_HEAD(&node->lower); 3405d4f98a2SYan Zheng RB_CLEAR_NODE(&node->rb_node); 3413fd0a558SYan, Zheng cache->nr_nodes++; 3423fd0a558SYan, Zheng } 3433fd0a558SYan, Zheng return node; 3443fd0a558SYan, Zheng } 3453fd0a558SYan, Zheng 3463fd0a558SYan, Zheng static void free_backref_node(struct backref_cache *cache, 3473fd0a558SYan, Zheng struct backref_node *node) 3483fd0a558SYan, Zheng { 3493fd0a558SYan, Zheng if (node) { 3503fd0a558SYan, Zheng cache->nr_nodes--; 35100246528SJosef Bacik btrfs_put_root(node->root); 3523fd0a558SYan, Zheng kfree(node); 3533fd0a558SYan, Zheng } 3543fd0a558SYan, Zheng } 3553fd0a558SYan, Zheng 3563fd0a558SYan, Zheng static struct backref_edge *alloc_backref_edge(struct backref_cache *cache) 3573fd0a558SYan, Zheng { 3583fd0a558SYan, Zheng struct backref_edge *edge; 3593fd0a558SYan, Zheng 3603fd0a558SYan, Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 3613fd0a558SYan, Zheng if (edge) 3623fd0a558SYan, Zheng cache->nr_edges++; 3633fd0a558SYan, Zheng return edge; 3643fd0a558SYan, Zheng } 3653fd0a558SYan, Zheng 3663fd0a558SYan, Zheng static void free_backref_edge(struct backref_cache *cache, 3673fd0a558SYan, Zheng struct backref_edge *edge) 3683fd0a558SYan, Zheng { 3693fd0a558SYan, Zheng if (edge) { 3703fd0a558SYan, Zheng cache->nr_edges--; 3713fd0a558SYan, Zheng kfree(edge); 3723fd0a558SYan, Zheng } 3735d4f98a2SYan Zheng } 3745d4f98a2SYan Zheng 3755d4f98a2SYan Zheng static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr, 3765d4f98a2SYan Zheng struct rb_node *node) 3775d4f98a2SYan Zheng { 3785d4f98a2SYan Zheng struct rb_node **p = &root->rb_node; 3795d4f98a2SYan Zheng struct rb_node *parent = NULL; 3805d4f98a2SYan Zheng struct tree_entry *entry; 3815d4f98a2SYan Zheng 3825d4f98a2SYan Zheng while (*p) { 3835d4f98a2SYan Zheng parent = *p; 3845d4f98a2SYan Zheng entry = rb_entry(parent, struct tree_entry, rb_node); 3855d4f98a2SYan Zheng 3865d4f98a2SYan Zheng if (bytenr < entry->bytenr) 3875d4f98a2SYan Zheng p = &(*p)->rb_left; 3885d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 3895d4f98a2SYan Zheng p = &(*p)->rb_right; 3905d4f98a2SYan Zheng else 3915d4f98a2SYan Zheng return parent; 3925d4f98a2SYan Zheng } 3935d4f98a2SYan Zheng 3945d4f98a2SYan Zheng rb_link_node(node, parent, p); 3955d4f98a2SYan Zheng rb_insert_color(node, root); 3965d4f98a2SYan Zheng return NULL; 3975d4f98a2SYan Zheng } 3985d4f98a2SYan Zheng 3995d4f98a2SYan Zheng static struct rb_node *tree_search(struct rb_root *root, u64 bytenr) 4005d4f98a2SYan Zheng { 4015d4f98a2SYan Zheng struct rb_node *n = root->rb_node; 4025d4f98a2SYan Zheng struct tree_entry *entry; 4035d4f98a2SYan Zheng 4045d4f98a2SYan Zheng while (n) { 4055d4f98a2SYan Zheng entry = rb_entry(n, struct tree_entry, rb_node); 4065d4f98a2SYan Zheng 4075d4f98a2SYan Zheng if (bytenr < entry->bytenr) 4085d4f98a2SYan Zheng n = n->rb_left; 4095d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 4105d4f98a2SYan Zheng n = n->rb_right; 4115d4f98a2SYan Zheng else 4125d4f98a2SYan Zheng return n; 4135d4f98a2SYan Zheng } 4145d4f98a2SYan Zheng return NULL; 4155d4f98a2SYan Zheng } 4165d4f98a2SYan Zheng 41748a3b636SEric Sandeen static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr) 41843c04fb1SJeff Mahoney { 41943c04fb1SJeff Mahoney 42043c04fb1SJeff Mahoney struct btrfs_fs_info *fs_info = NULL; 42143c04fb1SJeff Mahoney struct backref_node *bnode = rb_entry(rb_node, struct backref_node, 42243c04fb1SJeff Mahoney rb_node); 42343c04fb1SJeff Mahoney if (bnode->root) 42443c04fb1SJeff Mahoney fs_info = bnode->root->fs_info; 4255d163e0eSJeff Mahoney btrfs_panic(fs_info, errno, 4265d163e0eSJeff Mahoney "Inconsistency in backref cache found at offset %llu", 4275d163e0eSJeff Mahoney bytenr); 42843c04fb1SJeff Mahoney } 42943c04fb1SJeff Mahoney 4305d4f98a2SYan Zheng /* 4315d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 4325d4f98a2SYan Zheng */ 4335d4f98a2SYan Zheng static struct backref_node *walk_up_backref(struct backref_node *node, 4345d4f98a2SYan Zheng struct backref_edge *edges[], 4355d4f98a2SYan Zheng int *index) 4365d4f98a2SYan Zheng { 4375d4f98a2SYan Zheng struct backref_edge *edge; 4385d4f98a2SYan Zheng int idx = *index; 4395d4f98a2SYan Zheng 4405d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 4415d4f98a2SYan Zheng edge = list_entry(node->upper.next, 4425d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 4435d4f98a2SYan Zheng edges[idx++] = edge; 4445d4f98a2SYan Zheng node = edge->node[UPPER]; 4455d4f98a2SYan Zheng } 4463fd0a558SYan, Zheng BUG_ON(node->detached); 4475d4f98a2SYan Zheng *index = idx; 4485d4f98a2SYan Zheng return node; 4495d4f98a2SYan Zheng } 4505d4f98a2SYan Zheng 4515d4f98a2SYan Zheng /* 4525d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 4535d4f98a2SYan Zheng */ 4545d4f98a2SYan Zheng static struct backref_node *walk_down_backref(struct backref_edge *edges[], 4555d4f98a2SYan Zheng int *index) 4565d4f98a2SYan Zheng { 4575d4f98a2SYan Zheng struct backref_edge *edge; 4585d4f98a2SYan Zheng struct backref_node *lower; 4595d4f98a2SYan Zheng int idx = *index; 4605d4f98a2SYan Zheng 4615d4f98a2SYan Zheng while (idx > 0) { 4625d4f98a2SYan Zheng edge = edges[idx - 1]; 4635d4f98a2SYan Zheng lower = edge->node[LOWER]; 4645d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 4655d4f98a2SYan Zheng idx--; 4665d4f98a2SYan Zheng continue; 4675d4f98a2SYan Zheng } 4685d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 4695d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 4705d4f98a2SYan Zheng edges[idx - 1] = edge; 4715d4f98a2SYan Zheng *index = idx; 4725d4f98a2SYan Zheng return edge->node[UPPER]; 4735d4f98a2SYan Zheng } 4745d4f98a2SYan Zheng *index = 0; 4755d4f98a2SYan Zheng return NULL; 4765d4f98a2SYan Zheng } 4775d4f98a2SYan Zheng 4783fd0a558SYan, Zheng static void unlock_node_buffer(struct backref_node *node) 4795d4f98a2SYan Zheng { 4805d4f98a2SYan Zheng if (node->locked) { 4815d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 4825d4f98a2SYan Zheng node->locked = 0; 4835d4f98a2SYan Zheng } 4843fd0a558SYan, Zheng } 4853fd0a558SYan, Zheng 4863fd0a558SYan, Zheng static void drop_node_buffer(struct backref_node *node) 4873fd0a558SYan, Zheng { 4883fd0a558SYan, Zheng if (node->eb) { 4893fd0a558SYan, Zheng unlock_node_buffer(node); 4905d4f98a2SYan Zheng free_extent_buffer(node->eb); 4915d4f98a2SYan Zheng node->eb = NULL; 4925d4f98a2SYan Zheng } 4935d4f98a2SYan Zheng } 4945d4f98a2SYan Zheng 4955d4f98a2SYan Zheng static void drop_backref_node(struct backref_cache *tree, 4965d4f98a2SYan Zheng struct backref_node *node) 4975d4f98a2SYan Zheng { 4985d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 4995d4f98a2SYan Zheng 5005d4f98a2SYan Zheng drop_node_buffer(node); 5013fd0a558SYan, Zheng list_del(&node->list); 5025d4f98a2SYan Zheng list_del(&node->lower); 5033fd0a558SYan, Zheng if (!RB_EMPTY_NODE(&node->rb_node)) 5045d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 5053fd0a558SYan, Zheng free_backref_node(tree, node); 5065d4f98a2SYan Zheng } 5075d4f98a2SYan Zheng 5085d4f98a2SYan Zheng /* 5095d4f98a2SYan Zheng * remove a backref node from the backref cache 5105d4f98a2SYan Zheng */ 5115d4f98a2SYan Zheng static void remove_backref_node(struct backref_cache *cache, 5125d4f98a2SYan Zheng struct backref_node *node) 5135d4f98a2SYan Zheng { 5145d4f98a2SYan Zheng struct backref_node *upper; 5155d4f98a2SYan Zheng struct backref_edge *edge; 5165d4f98a2SYan Zheng 5175d4f98a2SYan Zheng if (!node) 5185d4f98a2SYan Zheng return; 5195d4f98a2SYan Zheng 5203fd0a558SYan, Zheng BUG_ON(!node->lowest && !node->detached); 5215d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 5225d4f98a2SYan Zheng edge = list_entry(node->upper.next, struct backref_edge, 5235d4f98a2SYan Zheng list[LOWER]); 5245d4f98a2SYan Zheng upper = edge->node[UPPER]; 5255d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 5265d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 5273fd0a558SYan, Zheng free_backref_edge(cache, edge); 5283fd0a558SYan, Zheng 5293fd0a558SYan, Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 5303fd0a558SYan, Zheng BUG_ON(!list_empty(&node->upper)); 5313fd0a558SYan, Zheng drop_backref_node(cache, node); 5323fd0a558SYan, Zheng node = upper; 5333fd0a558SYan, Zheng node->lowest = 1; 5343fd0a558SYan, Zheng continue; 5353fd0a558SYan, Zheng } 5365d4f98a2SYan Zheng /* 5373fd0a558SYan, Zheng * add the node to leaf node list if no other 5385d4f98a2SYan Zheng * child block cached. 5395d4f98a2SYan Zheng */ 5405d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 5413fd0a558SYan, Zheng list_add_tail(&upper->lower, &cache->leaves); 5425d4f98a2SYan Zheng upper->lowest = 1; 5435d4f98a2SYan Zheng } 5445d4f98a2SYan Zheng } 5453fd0a558SYan, Zheng 5465d4f98a2SYan Zheng drop_backref_node(cache, node); 5475d4f98a2SYan Zheng } 5485d4f98a2SYan Zheng 5493fd0a558SYan, Zheng static void update_backref_node(struct backref_cache *cache, 5503fd0a558SYan, Zheng struct backref_node *node, u64 bytenr) 5513fd0a558SYan, Zheng { 5523fd0a558SYan, Zheng struct rb_node *rb_node; 5533fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 5543fd0a558SYan, Zheng node->bytenr = bytenr; 5553fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 55643c04fb1SJeff Mahoney if (rb_node) 55743c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, bytenr); 5583fd0a558SYan, Zheng } 5593fd0a558SYan, Zheng 5603fd0a558SYan, Zheng /* 5613fd0a558SYan, Zheng * update backref cache after a transaction commit 5623fd0a558SYan, Zheng */ 5633fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 5643fd0a558SYan, Zheng struct backref_cache *cache) 5653fd0a558SYan, Zheng { 5663fd0a558SYan, Zheng struct backref_node *node; 5673fd0a558SYan, Zheng int level = 0; 5683fd0a558SYan, Zheng 5693fd0a558SYan, Zheng if (cache->last_trans == 0) { 5703fd0a558SYan, Zheng cache->last_trans = trans->transid; 5713fd0a558SYan, Zheng return 0; 5723fd0a558SYan, Zheng } 5733fd0a558SYan, Zheng 5743fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 5753fd0a558SYan, Zheng return 0; 5763fd0a558SYan, Zheng 5773fd0a558SYan, Zheng /* 5783fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 5793fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 5803fd0a558SYan, Zheng * so the detached nodes are no longer useful. 5813fd0a558SYan, Zheng */ 5823fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 5833fd0a558SYan, Zheng node = list_entry(cache->detached.next, 5843fd0a558SYan, Zheng struct backref_node, list); 5853fd0a558SYan, Zheng remove_backref_node(cache, node); 5863fd0a558SYan, Zheng } 5873fd0a558SYan, Zheng 5883fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 5893fd0a558SYan, Zheng node = list_entry(cache->changed.next, 5903fd0a558SYan, Zheng struct backref_node, list); 5913fd0a558SYan, Zheng list_del_init(&node->list); 5923fd0a558SYan, Zheng BUG_ON(node->pending); 5933fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 5943fd0a558SYan, Zheng } 5953fd0a558SYan, Zheng 5963fd0a558SYan, Zheng /* 5973fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 5983fd0a558SYan, Zheng * errors during processing the pending nodes. 5993fd0a558SYan, Zheng */ 6003fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 6013fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 6023fd0a558SYan, Zheng BUG_ON(!node->pending); 6033fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 6043fd0a558SYan, Zheng continue; 6053fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 6063fd0a558SYan, Zheng } 6073fd0a558SYan, Zheng } 6083fd0a558SYan, Zheng 6093fd0a558SYan, Zheng cache->last_trans = 0; 6103fd0a558SYan, Zheng return 1; 6113fd0a558SYan, Zheng } 6123fd0a558SYan, Zheng 6136282675eSQu Wenruo static bool reloc_root_is_dead(struct btrfs_root *root) 6146282675eSQu Wenruo { 6156282675eSQu Wenruo /* 6166282675eSQu Wenruo * Pair with set_bit/clear_bit in clean_dirty_subvols and 6176282675eSQu Wenruo * btrfs_update_reloc_root. We need to see the updated bit before 6186282675eSQu Wenruo * trying to access reloc_root 6196282675eSQu Wenruo */ 6206282675eSQu Wenruo smp_rmb(); 6216282675eSQu Wenruo if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state)) 6226282675eSQu Wenruo return true; 6236282675eSQu Wenruo return false; 6246282675eSQu Wenruo } 6256282675eSQu Wenruo 6266282675eSQu Wenruo /* 6276282675eSQu Wenruo * Check if this subvolume tree has valid reloc tree. 6286282675eSQu Wenruo * 6296282675eSQu Wenruo * Reloc tree after swap is considered dead, thus not considered as valid. 6306282675eSQu Wenruo * This is enough for most callers, as they don't distinguish dead reloc root 6316282675eSQu Wenruo * from no reloc root. But should_ignore_root() below is a special case. 6326282675eSQu Wenruo */ 6336282675eSQu Wenruo static bool have_reloc_root(struct btrfs_root *root) 6346282675eSQu Wenruo { 6356282675eSQu Wenruo if (reloc_root_is_dead(root)) 6366282675eSQu Wenruo return false; 6376282675eSQu Wenruo if (!root->reloc_root) 6386282675eSQu Wenruo return false; 6396282675eSQu Wenruo return true; 6406282675eSQu Wenruo } 641f2a97a9dSDavid Sterba 6423fd0a558SYan, Zheng static int should_ignore_root(struct btrfs_root *root) 6433fd0a558SYan, Zheng { 6443fd0a558SYan, Zheng struct btrfs_root *reloc_root; 6453fd0a558SYan, Zheng 64627cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 6473fd0a558SYan, Zheng return 0; 6483fd0a558SYan, Zheng 6496282675eSQu Wenruo /* This root has been merged with its reloc tree, we can ignore it */ 6506282675eSQu Wenruo if (reloc_root_is_dead(root)) 6516282675eSQu Wenruo return 1; 6526282675eSQu Wenruo 6533fd0a558SYan, Zheng reloc_root = root->reloc_root; 6543fd0a558SYan, Zheng if (!reloc_root) 6553fd0a558SYan, Zheng return 0; 6563fd0a558SYan, Zheng 6574d4225fcSJosef Bacik if (btrfs_header_generation(reloc_root->commit_root) == 6584d4225fcSJosef Bacik root->fs_info->running_transaction->transid) 6593fd0a558SYan, Zheng return 0; 6603fd0a558SYan, Zheng /* 6613fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 6623fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 6633fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 6643fd0a558SYan, Zheng * relocation. 6653fd0a558SYan, Zheng */ 6663fd0a558SYan, Zheng return 1; 6673fd0a558SYan, Zheng } 6685d4f98a2SYan Zheng /* 6695d4f98a2SYan Zheng * find reloc tree by address of tree root 6705d4f98a2SYan Zheng */ 6712433bea5SQu Wenruo struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr) 6725d4f98a2SYan Zheng { 6732433bea5SQu Wenruo struct reloc_control *rc = fs_info->reloc_ctl; 6745d4f98a2SYan Zheng struct rb_node *rb_node; 6755d4f98a2SYan Zheng struct mapping_node *node; 6765d4f98a2SYan Zheng struct btrfs_root *root = NULL; 6775d4f98a2SYan Zheng 6782433bea5SQu Wenruo ASSERT(rc); 6795d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 6805d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr); 6815d4f98a2SYan Zheng if (rb_node) { 6825d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 6835d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 6845d4f98a2SYan Zheng } 6855d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 68600246528SJosef Bacik return btrfs_grab_root(root); 6875d4f98a2SYan Zheng } 6885d4f98a2SYan Zheng 6895d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 6905d4f98a2SYan Zheng u64 root_objectid) 6915d4f98a2SYan Zheng { 6925d4f98a2SYan Zheng struct btrfs_key key; 6935d4f98a2SYan Zheng 6945d4f98a2SYan Zheng key.objectid = root_objectid; 6955d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 6965d4f98a2SYan Zheng key.offset = (u64)-1; 6975d4f98a2SYan Zheng 698bc44d7c4SJosef Bacik return btrfs_get_fs_root(fs_info, &key, false); 6995d4f98a2SYan Zheng } 7005d4f98a2SYan Zheng 7015d4f98a2SYan Zheng /* 7024007ea87SQu Wenruo * Handle direct tree backref 7034007ea87SQu Wenruo * 7044007ea87SQu Wenruo * Direct tree backref means, the backref item shows its parent bytenr 7054007ea87SQu Wenruo * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined). 7064007ea87SQu Wenruo * 7074007ea87SQu Wenruo * @ref_key: The converted backref key. 7084007ea87SQu Wenruo * For keyed backref, it's the item key. 7094007ea87SQu Wenruo * For inlined backref, objectid is the bytenr, 7104007ea87SQu Wenruo * type is btrfs_inline_ref_type, offset is 7114007ea87SQu Wenruo * btrfs_inline_ref_offset. 7124007ea87SQu Wenruo */ 7134007ea87SQu Wenruo static int handle_direct_tree_backref(struct backref_cache *cache, 7144007ea87SQu Wenruo struct btrfs_key *ref_key, 7154007ea87SQu Wenruo struct backref_node *cur) 7164007ea87SQu Wenruo { 7174007ea87SQu Wenruo struct backref_edge *edge; 7184007ea87SQu Wenruo struct backref_node *upper; 7194007ea87SQu Wenruo struct rb_node *rb_node; 7204007ea87SQu Wenruo 7214007ea87SQu Wenruo ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY); 7224007ea87SQu Wenruo 7234007ea87SQu Wenruo /* Only reloc root uses backref pointing to itself */ 7244007ea87SQu Wenruo if (ref_key->objectid == ref_key->offset) { 7254007ea87SQu Wenruo struct btrfs_root *root; 7264007ea87SQu Wenruo 7274007ea87SQu Wenruo cur->is_reloc_root = 1; 7284007ea87SQu Wenruo /* Only reloc backref cache cares about a specific root */ 7294007ea87SQu Wenruo if (cache->is_reloc) { 7304007ea87SQu Wenruo root = find_reloc_root(cache->fs_info, cur->bytenr); 7314007ea87SQu Wenruo if (WARN_ON(!root)) 7324007ea87SQu Wenruo return -ENOENT; 7334007ea87SQu Wenruo cur->root = root; 7344007ea87SQu Wenruo } else { 7354007ea87SQu Wenruo /* 7364007ea87SQu Wenruo * For generic purpose backref cache, reloc root node 7374007ea87SQu Wenruo * is useless. 7384007ea87SQu Wenruo */ 7394007ea87SQu Wenruo list_add(&cur->list, &cache->useless_node); 7404007ea87SQu Wenruo } 7414007ea87SQu Wenruo return 0; 7424007ea87SQu Wenruo } 7434007ea87SQu Wenruo 7444007ea87SQu Wenruo edge = alloc_backref_edge(cache); 7454007ea87SQu Wenruo if (!edge) 7464007ea87SQu Wenruo return -ENOMEM; 7474007ea87SQu Wenruo 7484007ea87SQu Wenruo rb_node = tree_search(&cache->rb_root, ref_key->offset); 7494007ea87SQu Wenruo if (!rb_node) { 7504007ea87SQu Wenruo /* Parent node not yet cached */ 7514007ea87SQu Wenruo upper = alloc_backref_node(cache); 7524007ea87SQu Wenruo if (!upper) { 7534007ea87SQu Wenruo free_backref_edge(cache, edge); 7544007ea87SQu Wenruo return -ENOMEM; 7554007ea87SQu Wenruo } 7564007ea87SQu Wenruo upper->bytenr = ref_key->offset; 7574007ea87SQu Wenruo upper->level = cur->level + 1; 7584007ea87SQu Wenruo 7594007ea87SQu Wenruo /* 7604007ea87SQu Wenruo * Backrefs for the upper level block isn't cached, add the 7614007ea87SQu Wenruo * block to pending list 7624007ea87SQu Wenruo */ 7634007ea87SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 7644007ea87SQu Wenruo } else { 7654007ea87SQu Wenruo /* Parent node already cached */ 7664007ea87SQu Wenruo upper = rb_entry(rb_node, struct backref_node, rb_node); 7674007ea87SQu Wenruo ASSERT(upper->checked); 7684007ea87SQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 7694007ea87SQu Wenruo } 7704007ea87SQu Wenruo list_add_tail(&edge->list[LOWER], &cur->upper); 7714007ea87SQu Wenruo edge->node[LOWER] = cur; 7724007ea87SQu Wenruo edge->node[UPPER] = upper; 7734007ea87SQu Wenruo return 0; 7744007ea87SQu Wenruo } 7754007ea87SQu Wenruo 7764007ea87SQu Wenruo /* 7775d4f98a2SYan Zheng * build backref tree for a given tree block. root of the backref tree 7785d4f98a2SYan Zheng * corresponds the tree block, leaves of the backref tree correspond 7795d4f98a2SYan Zheng * roots of b-trees that reference the tree block. 7805d4f98a2SYan Zheng * 7815d4f98a2SYan Zheng * the basic idea of this function is check backrefs of a given block 78201327610SNicholas D Steeves * to find upper level blocks that reference the block, and then check 78301327610SNicholas D Steeves * backrefs of these upper level blocks recursively. the recursion stop 7845d4f98a2SYan Zheng * when tree root is reached or backrefs for the block is cached. 7855d4f98a2SYan Zheng * 7865d4f98a2SYan Zheng * NOTE: if we find backrefs for a block are cached, we know backrefs 7875d4f98a2SYan Zheng * for all upper level blocks that directly/indirectly reference the 7885d4f98a2SYan Zheng * block are also cached. 7895d4f98a2SYan Zheng */ 7903fd0a558SYan, Zheng static noinline_for_stack 7913fd0a558SYan, Zheng struct backref_node *build_backref_tree(struct reloc_control *rc, 7925d4f98a2SYan Zheng struct btrfs_key *node_key, 7935d4f98a2SYan Zheng int level, u64 bytenr) 7945d4f98a2SYan Zheng { 79571f572a9SQu Wenruo struct btrfs_backref_iter *iter; 7963fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 79771f572a9SQu Wenruo /* For searching parent of TREE_BLOCK_REF */ 79871f572a9SQu Wenruo struct btrfs_path *path; 7995d4f98a2SYan Zheng struct btrfs_root *root; 8005d4f98a2SYan Zheng struct backref_node *cur; 8015d4f98a2SYan Zheng struct backref_node *upper; 8025d4f98a2SYan Zheng struct backref_node *lower; 8035d4f98a2SYan Zheng struct backref_node *node = NULL; 8045d4f98a2SYan Zheng struct backref_node *exist = NULL; 8055d4f98a2SYan Zheng struct backref_edge *edge; 8065d4f98a2SYan Zheng struct rb_node *rb_node; 8073fd0a558SYan, Zheng int cowonly; 8085d4f98a2SYan Zheng int ret; 8095d4f98a2SYan Zheng int err = 0; 810b6c60c80SJosef Bacik bool need_check = true; 8115d4f98a2SYan Zheng 81271f572a9SQu Wenruo iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS); 81371f572a9SQu Wenruo if (!iter) 81471f572a9SQu Wenruo return ERR_PTR(-ENOMEM); 81571f572a9SQu Wenruo path = btrfs_alloc_path(); 81671f572a9SQu Wenruo if (!path) { 8175d4f98a2SYan Zheng err = -ENOMEM; 8185d4f98a2SYan Zheng goto out; 8195d4f98a2SYan Zheng } 8205d4f98a2SYan Zheng 8213fd0a558SYan, Zheng node = alloc_backref_node(cache); 8225d4f98a2SYan Zheng if (!node) { 8235d4f98a2SYan Zheng err = -ENOMEM; 8245d4f98a2SYan Zheng goto out; 8255d4f98a2SYan Zheng } 8265d4f98a2SYan Zheng 8275d4f98a2SYan Zheng node->bytenr = bytenr; 8285d4f98a2SYan Zheng node->level = level; 8295d4f98a2SYan Zheng node->lowest = 1; 8305d4f98a2SYan Zheng cur = node; 8315d4f98a2SYan Zheng again: 83271f572a9SQu Wenruo ret = btrfs_backref_iter_start(iter, cur->bytenr); 8335d4f98a2SYan Zheng if (ret < 0) { 8345d4f98a2SYan Zheng err = ret; 8355d4f98a2SYan Zheng goto out; 8365d4f98a2SYan Zheng } 8375d4f98a2SYan Zheng 83871f572a9SQu Wenruo /* 83971f572a9SQu Wenruo * We skip the first btrfs_tree_block_info, as we don't use the key 84071f572a9SQu Wenruo * stored in it, but fetch it from the tree block 84171f572a9SQu Wenruo */ 84271f572a9SQu Wenruo if (btrfs_backref_has_tree_block_info(iter)) { 84371f572a9SQu Wenruo ret = btrfs_backref_iter_next(iter); 84471f572a9SQu Wenruo if (ret < 0) { 84571f572a9SQu Wenruo err = ret; 84671f572a9SQu Wenruo goto out; 84771f572a9SQu Wenruo } 84871f572a9SQu Wenruo /* No extra backref? This means the tree block is corrupted */ 84971f572a9SQu Wenruo if (ret > 0) { 85071f572a9SQu Wenruo err = -EUCLEAN; 85171f572a9SQu Wenruo goto out; 85271f572a9SQu Wenruo } 85371f572a9SQu Wenruo } 8545d4f98a2SYan Zheng WARN_ON(cur->checked); 8555d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 8565d4f98a2SYan Zheng /* 85770f23fd6SJustin P. Mattock * the backref was added previously when processing 8585d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 8595d4f98a2SYan Zheng */ 86075bfb9afSJosef Bacik ASSERT(list_is_singular(&cur->upper)); 8615d4f98a2SYan Zheng edge = list_entry(cur->upper.next, struct backref_edge, 8625d4f98a2SYan Zheng list[LOWER]); 86375bfb9afSJosef Bacik ASSERT(list_empty(&edge->list[UPPER])); 8645d4f98a2SYan Zheng exist = edge->node[UPPER]; 8655d4f98a2SYan Zheng /* 8665d4f98a2SYan Zheng * add the upper level block to pending list if we need 8675d4f98a2SYan Zheng * check its backrefs 8685d4f98a2SYan Zheng */ 8695d4f98a2SYan Zheng if (!exist->checked) 87084780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 8715d4f98a2SYan Zheng } else { 8725d4f98a2SYan Zheng exist = NULL; 8735d4f98a2SYan Zheng } 8745d4f98a2SYan Zheng 87571f572a9SQu Wenruo for (; ret == 0; ret = btrfs_backref_iter_next(iter)) { 87671f572a9SQu Wenruo struct extent_buffer *eb; 87771f572a9SQu Wenruo struct btrfs_key key; 8783de28d57SLiu Bo int type; 87971f572a9SQu Wenruo 88071f572a9SQu Wenruo cond_resched(); 88171f572a9SQu Wenruo eb = btrfs_backref_get_eb(iter); 88271f572a9SQu Wenruo 88371f572a9SQu Wenruo key.objectid = iter->bytenr; 88471f572a9SQu Wenruo if (btrfs_backref_iter_is_inline_ref(iter)) { 88571f572a9SQu Wenruo struct btrfs_extent_inline_ref *iref; 88671f572a9SQu Wenruo 88771f572a9SQu Wenruo /* update key for inline back ref */ 88871f572a9SQu Wenruo iref = (struct btrfs_extent_inline_ref *) 88971f572a9SQu Wenruo ((unsigned long)iter->cur_ptr); 8903de28d57SLiu Bo type = btrfs_get_extent_inline_ref_type(eb, iref, 8913de28d57SLiu Bo BTRFS_REF_TYPE_BLOCK); 8923de28d57SLiu Bo if (type == BTRFS_REF_TYPE_INVALID) { 893af431dcbSSu Yue err = -EUCLEAN; 8943de28d57SLiu Bo goto out; 8953de28d57SLiu Bo } 8963de28d57SLiu Bo key.type = type; 8975d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 89871f572a9SQu Wenruo } else { 89971f572a9SQu Wenruo key.type = iter->cur_key.type; 90071f572a9SQu Wenruo key.offset = iter->cur_key.offset; 9015d4f98a2SYan Zheng } 9025d4f98a2SYan Zheng 903fa6ac715SQu Wenruo /* 904fa6ac715SQu Wenruo * Parent node found and matches current inline ref, no need to 905fa6ac715SQu Wenruo * rebuild this node for this inline ref. 906fa6ac715SQu Wenruo */ 9075d4f98a2SYan Zheng if (exist && 9085d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 9095d4f98a2SYan Zheng exist->owner == key.offset) || 9105d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 9115d4f98a2SYan Zheng exist->bytenr == key.offset))) { 9125d4f98a2SYan Zheng exist = NULL; 91371f572a9SQu Wenruo continue; 9145d4f98a2SYan Zheng } 9155d4f98a2SYan Zheng 916fa6ac715SQu Wenruo /* SHARED_BLOCK_REF means key.offset is the parent bytenr */ 9175d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 9184007ea87SQu Wenruo ret = handle_direct_tree_backref(cache, &key, cur); 9194007ea87SQu Wenruo if (ret < 0) { 9204007ea87SQu Wenruo err = ret; 9212433bea5SQu Wenruo goto out; 9222433bea5SQu Wenruo } 92371f572a9SQu Wenruo continue; 9246d8ff4e4SDavid Sterba } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) { 925ba3c2b19SNikolay Borisov err = -EINVAL; 926ba3c2b19SNikolay Borisov btrfs_print_v0_err(rc->extent_root->fs_info); 927ba3c2b19SNikolay Borisov btrfs_handle_fs_error(rc->extent_root->fs_info, err, 928ba3c2b19SNikolay Borisov NULL); 929ba3c2b19SNikolay Borisov goto out; 9305d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 93171f572a9SQu Wenruo continue; 9325d4f98a2SYan Zheng } 9335d4f98a2SYan Zheng 934fa6ac715SQu Wenruo /* 935fa6ac715SQu Wenruo * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset 936fa6ac715SQu Wenruo * means the root objectid. We need to search the tree to get 937fa6ac715SQu Wenruo * its parent bytenr. 938fa6ac715SQu Wenruo */ 9395d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, key.offset); 9405d4f98a2SYan Zheng if (IS_ERR(root)) { 9415d4f98a2SYan Zheng err = PTR_ERR(root); 9425d4f98a2SYan Zheng goto out; 9435d4f98a2SYan Zheng } 9445d4f98a2SYan Zheng 94527cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 9463fd0a558SYan, Zheng cur->cowonly = 1; 9473fd0a558SYan, Zheng 9485d4f98a2SYan Zheng if (btrfs_root_level(&root->root_item) == cur->level) { 9495d4f98a2SYan Zheng /* tree root */ 95075bfb9afSJosef Bacik ASSERT(btrfs_root_bytenr(&root->root_item) == 9515d4f98a2SYan Zheng cur->bytenr); 9520b530bc5SJosef Bacik if (should_ignore_root(root)) { 95300246528SJosef Bacik btrfs_put_root(root); 95484780289SQu Wenruo list_add(&cur->list, &cache->useless_node); 9550b530bc5SJosef Bacik } else { 9565d4f98a2SYan Zheng cur->root = root; 9570b530bc5SJosef Bacik } 9585d4f98a2SYan Zheng break; 9595d4f98a2SYan Zheng } 9605d4f98a2SYan Zheng 9615d4f98a2SYan Zheng level = cur->level + 1; 9625d4f98a2SYan Zheng 963fa6ac715SQu Wenruo /* Search the tree to find parent blocks referring the block. */ 96471f572a9SQu Wenruo path->search_commit_root = 1; 96571f572a9SQu Wenruo path->skip_locking = 1; 96671f572a9SQu Wenruo path->lowest_level = level; 96771f572a9SQu Wenruo ret = btrfs_search_slot(NULL, root, node_key, path, 0, 0); 96871f572a9SQu Wenruo path->lowest_level = 0; 9695d4f98a2SYan Zheng if (ret < 0) { 97000246528SJosef Bacik btrfs_put_root(root); 9715d4f98a2SYan Zheng err = ret; 9725d4f98a2SYan Zheng goto out; 9735d4f98a2SYan Zheng } 97471f572a9SQu Wenruo if (ret > 0 && path->slots[level] > 0) 97571f572a9SQu Wenruo path->slots[level]--; 9765d4f98a2SYan Zheng 97771f572a9SQu Wenruo eb = path->nodes[level]; 97871f572a9SQu Wenruo if (btrfs_node_blockptr(eb, path->slots[level]) != 9793561b9dbSLiu Bo cur->bytenr) { 9803561b9dbSLiu Bo btrfs_err(root->fs_info, 9813561b9dbSLiu Bo "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)", 9824fd786e6SMisono Tomohiro cur->bytenr, level - 1, 9834fd786e6SMisono Tomohiro root->root_key.objectid, 9843561b9dbSLiu Bo node_key->objectid, node_key->type, 9853561b9dbSLiu Bo node_key->offset); 98600246528SJosef Bacik btrfs_put_root(root); 9873561b9dbSLiu Bo err = -ENOENT; 9883561b9dbSLiu Bo goto out; 9893561b9dbSLiu Bo } 9905d4f98a2SYan Zheng lower = cur; 991b6c60c80SJosef Bacik need_check = true; 992fa6ac715SQu Wenruo 993fa6ac715SQu Wenruo /* Add all nodes and edges in the path */ 9945d4f98a2SYan Zheng for (; level < BTRFS_MAX_LEVEL; level++) { 99571f572a9SQu Wenruo if (!path->nodes[level]) { 99675bfb9afSJosef Bacik ASSERT(btrfs_root_bytenr(&root->root_item) == 9975d4f98a2SYan Zheng lower->bytenr); 9980b530bc5SJosef Bacik if (should_ignore_root(root)) { 99900246528SJosef Bacik btrfs_put_root(root); 100084780289SQu Wenruo list_add(&lower->list, 100184780289SQu Wenruo &cache->useless_node); 10020b530bc5SJosef Bacik } else { 10035d4f98a2SYan Zheng lower->root = root; 10040b530bc5SJosef Bacik } 10055d4f98a2SYan Zheng break; 10065d4f98a2SYan Zheng } 10075d4f98a2SYan Zheng 10083fd0a558SYan, Zheng edge = alloc_backref_edge(cache); 10095d4f98a2SYan Zheng if (!edge) { 101000246528SJosef Bacik btrfs_put_root(root); 10115d4f98a2SYan Zheng err = -ENOMEM; 10125d4f98a2SYan Zheng goto out; 10135d4f98a2SYan Zheng } 10145d4f98a2SYan Zheng 101571f572a9SQu Wenruo eb = path->nodes[level]; 10165d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, eb->start); 10175d4f98a2SYan Zheng if (!rb_node) { 10183fd0a558SYan, Zheng upper = alloc_backref_node(cache); 10195d4f98a2SYan Zheng if (!upper) { 102000246528SJosef Bacik btrfs_put_root(root); 10213fd0a558SYan, Zheng free_backref_edge(cache, edge); 10225d4f98a2SYan Zheng err = -ENOMEM; 10235d4f98a2SYan Zheng goto out; 10245d4f98a2SYan Zheng } 10255d4f98a2SYan Zheng upper->bytenr = eb->start; 10265d4f98a2SYan Zheng upper->owner = btrfs_header_owner(eb); 10275d4f98a2SYan Zheng upper->level = lower->level + 1; 102827cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, 102927cdeb70SMiao Xie &root->state)) 10303fd0a558SYan, Zheng upper->cowonly = 1; 10315d4f98a2SYan Zheng 10325d4f98a2SYan Zheng /* 10335d4f98a2SYan Zheng * if we know the block isn't shared 10345d4f98a2SYan Zheng * we can void checking its backrefs. 10355d4f98a2SYan Zheng */ 10365d4f98a2SYan Zheng if (btrfs_block_can_be_shared(root, eb)) 10375d4f98a2SYan Zheng upper->checked = 0; 10385d4f98a2SYan Zheng else 10395d4f98a2SYan Zheng upper->checked = 1; 10405d4f98a2SYan Zheng 10415d4f98a2SYan Zheng /* 10425d4f98a2SYan Zheng * add the block to pending list if we 1043b6c60c80SJosef Bacik * need check its backrefs, we only do this once 1044b6c60c80SJosef Bacik * while walking up a tree as we will catch 1045b6c60c80SJosef Bacik * anything else later on. 10465d4f98a2SYan Zheng */ 1047b6c60c80SJosef Bacik if (!upper->checked && need_check) { 1048b6c60c80SJosef Bacik need_check = false; 10495d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], 105084780289SQu Wenruo &cache->pending_edge); 1051bbe90514SJosef Bacik } else { 1052bbe90514SJosef Bacik if (upper->checked) 1053bbe90514SJosef Bacik need_check = true; 10545d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 1055bbe90514SJosef Bacik } 10565d4f98a2SYan Zheng } else { 10575d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 10585d4f98a2SYan Zheng rb_node); 105975bfb9afSJosef Bacik ASSERT(upper->checked); 10605d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 10613fd0a558SYan, Zheng if (!upper->owner) 10623fd0a558SYan, Zheng upper->owner = btrfs_header_owner(eb); 10635d4f98a2SYan Zheng } 10645d4f98a2SYan Zheng list_add_tail(&edge->list[LOWER], &lower->upper); 10655d4f98a2SYan Zheng edge->node[LOWER] = lower; 10663fd0a558SYan, Zheng edge->node[UPPER] = upper; 10675d4f98a2SYan Zheng 10680b530bc5SJosef Bacik if (rb_node) { 106900246528SJosef Bacik btrfs_put_root(root); 10705d4f98a2SYan Zheng break; 10710b530bc5SJosef Bacik } 10725d4f98a2SYan Zheng lower = upper; 10735d4f98a2SYan Zheng upper = NULL; 10745d4f98a2SYan Zheng } 107571f572a9SQu Wenruo btrfs_release_path(path); 10765d4f98a2SYan Zheng } 107771f572a9SQu Wenruo if (ret < 0) { 107871f572a9SQu Wenruo err = ret; 107971f572a9SQu Wenruo goto out; 10805d4f98a2SYan Zheng } 108171f572a9SQu Wenruo ret = 0; 108271f572a9SQu Wenruo btrfs_backref_iter_release(iter); 10835d4f98a2SYan Zheng 10845d4f98a2SYan Zheng cur->checked = 1; 10855d4f98a2SYan Zheng WARN_ON(exist); 10865d4f98a2SYan Zheng 10875d4f98a2SYan Zheng /* the pending list isn't empty, take the first block to process */ 108884780289SQu Wenruo if (!list_empty(&cache->pending_edge)) { 108984780289SQu Wenruo edge = list_first_entry(&cache->pending_edge, 109084780289SQu Wenruo struct backref_edge, list[UPPER]); 10915d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 10925d4f98a2SYan Zheng cur = edge->node[UPPER]; 10935d4f98a2SYan Zheng goto again; 10945d4f98a2SYan Zheng } 10955d4f98a2SYan Zheng 10965d4f98a2SYan Zheng /* 10975d4f98a2SYan Zheng * everything goes well, connect backref nodes and insert backref nodes 10985d4f98a2SYan Zheng * into the cache. 10995d4f98a2SYan Zheng */ 110075bfb9afSJosef Bacik ASSERT(node->checked); 11013fd0a558SYan, Zheng cowonly = node->cowonly; 11023fd0a558SYan, Zheng if (!cowonly) { 11033fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, 11043fd0a558SYan, Zheng &node->rb_node); 110543c04fb1SJeff Mahoney if (rb_node) 110643c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 11073fd0a558SYan, Zheng list_add_tail(&node->lower, &cache->leaves); 11083fd0a558SYan, Zheng } 11095d4f98a2SYan Zheng 11105d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) 111184780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 11125d4f98a2SYan Zheng 111384780289SQu Wenruo while (!list_empty(&cache->pending_edge)) { 111484780289SQu Wenruo edge = list_first_entry(&cache->pending_edge, 111584780289SQu Wenruo struct backref_edge, list[UPPER]); 11165d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 11175d4f98a2SYan Zheng upper = edge->node[UPPER]; 11183fd0a558SYan, Zheng if (upper->detached) { 11193fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 11203fd0a558SYan, Zheng lower = edge->node[LOWER]; 11213fd0a558SYan, Zheng free_backref_edge(cache, edge); 11223fd0a558SYan, Zheng if (list_empty(&lower->upper)) 112384780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 11243fd0a558SYan, Zheng continue; 11253fd0a558SYan, Zheng } 11265d4f98a2SYan Zheng 11275d4f98a2SYan Zheng if (!RB_EMPTY_NODE(&upper->rb_node)) { 11285d4f98a2SYan Zheng if (upper->lowest) { 11295d4f98a2SYan Zheng list_del_init(&upper->lower); 11305d4f98a2SYan Zheng upper->lowest = 0; 11315d4f98a2SYan Zheng } 11325d4f98a2SYan Zheng 11335d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 11345d4f98a2SYan Zheng continue; 11355d4f98a2SYan Zheng } 11365d4f98a2SYan Zheng 113775bfb9afSJosef Bacik if (!upper->checked) { 113875bfb9afSJosef Bacik /* 113975bfb9afSJosef Bacik * Still want to blow up for developers since this is a 114075bfb9afSJosef Bacik * logic bug. 114175bfb9afSJosef Bacik */ 114275bfb9afSJosef Bacik ASSERT(0); 114375bfb9afSJosef Bacik err = -EINVAL; 114475bfb9afSJosef Bacik goto out; 114575bfb9afSJosef Bacik } 114675bfb9afSJosef Bacik if (cowonly != upper->cowonly) { 114775bfb9afSJosef Bacik ASSERT(0); 114875bfb9afSJosef Bacik err = -EINVAL; 114975bfb9afSJosef Bacik goto out; 115075bfb9afSJosef Bacik } 115175bfb9afSJosef Bacik 11523fd0a558SYan, Zheng if (!cowonly) { 11535d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, upper->bytenr, 11545d4f98a2SYan Zheng &upper->rb_node); 115543c04fb1SJeff Mahoney if (rb_node) 115643c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, 115743c04fb1SJeff Mahoney upper->bytenr); 11583fd0a558SYan, Zheng } 11595d4f98a2SYan Zheng 11605d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 11615d4f98a2SYan Zheng 11625d4f98a2SYan Zheng list_for_each_entry(edge, &upper->upper, list[LOWER]) 116384780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 11645d4f98a2SYan Zheng } 11653fd0a558SYan, Zheng /* 11663fd0a558SYan, Zheng * process useless backref nodes. backref nodes for tree leaves 11673fd0a558SYan, Zheng * are deleted from the cache. backref nodes for upper level 11683fd0a558SYan, Zheng * tree blocks are left in the cache to avoid unnecessary backref 11693fd0a558SYan, Zheng * lookup. 11703fd0a558SYan, Zheng */ 117184780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 117284780289SQu Wenruo upper = list_first_entry(&cache->useless_node, 117384780289SQu Wenruo struct backref_node, list); 11743fd0a558SYan, Zheng list_del_init(&upper->list); 117575bfb9afSJosef Bacik ASSERT(list_empty(&upper->upper)); 11763fd0a558SYan, Zheng if (upper == node) 11773fd0a558SYan, Zheng node = NULL; 11783fd0a558SYan, Zheng if (upper->lowest) { 11793fd0a558SYan, Zheng list_del_init(&upper->lower); 11803fd0a558SYan, Zheng upper->lowest = 0; 11813fd0a558SYan, Zheng } 11823fd0a558SYan, Zheng while (!list_empty(&upper->lower)) { 118384780289SQu Wenruo edge = list_first_entry(&upper->lower, 11843fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 11853fd0a558SYan, Zheng list_del(&edge->list[UPPER]); 11863fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 11873fd0a558SYan, Zheng lower = edge->node[LOWER]; 11883fd0a558SYan, Zheng free_backref_edge(cache, edge); 11893fd0a558SYan, Zheng 11903fd0a558SYan, Zheng if (list_empty(&lower->upper)) 119184780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 11923fd0a558SYan, Zheng } 11939569cc20SQu Wenruo mark_block_processed(rc, upper); 11943fd0a558SYan, Zheng if (upper->level > 0) { 11953fd0a558SYan, Zheng list_add(&upper->list, &cache->detached); 11963fd0a558SYan, Zheng upper->detached = 1; 11973fd0a558SYan, Zheng } else { 11983fd0a558SYan, Zheng rb_erase(&upper->rb_node, &cache->rb_root); 11993fd0a558SYan, Zheng free_backref_node(cache, upper); 12003fd0a558SYan, Zheng } 12013fd0a558SYan, Zheng } 12025d4f98a2SYan Zheng out: 120371f572a9SQu Wenruo btrfs_backref_iter_free(iter); 120471f572a9SQu Wenruo btrfs_free_path(path); 12055d4f98a2SYan Zheng if (err) { 120684780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 120784780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 120875bfb9afSJosef Bacik struct backref_node, list); 120975bfb9afSJosef Bacik list_del_init(&lower->list); 12103fd0a558SYan, Zheng } 121184780289SQu Wenruo while (!list_empty(&cache->pending_edge)) { 121284780289SQu Wenruo edge = list_first_entry(&cache->pending_edge, 121384780289SQu Wenruo struct backref_edge, list[UPPER]); 121475bfb9afSJosef Bacik list_del(&edge->list[UPPER]); 12153fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 121675bfb9afSJosef Bacik lower = edge->node[LOWER]; 12175d4f98a2SYan Zheng upper = edge->node[UPPER]; 12183fd0a558SYan, Zheng free_backref_edge(cache, edge); 121975bfb9afSJosef Bacik 122075bfb9afSJosef Bacik /* 122175bfb9afSJosef Bacik * Lower is no longer linked to any upper backref nodes 122275bfb9afSJosef Bacik * and isn't in the cache, we can free it ourselves. 122375bfb9afSJosef Bacik */ 122475bfb9afSJosef Bacik if (list_empty(&lower->upper) && 122575bfb9afSJosef Bacik RB_EMPTY_NODE(&lower->rb_node)) 122684780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 122775bfb9afSJosef Bacik 122875bfb9afSJosef Bacik if (!RB_EMPTY_NODE(&upper->rb_node)) 122975bfb9afSJosef Bacik continue; 123075bfb9afSJosef Bacik 123101327610SNicholas D Steeves /* Add this guy's upper edges to the list to process */ 123275bfb9afSJosef Bacik list_for_each_entry(edge, &upper->upper, list[LOWER]) 123384780289SQu Wenruo list_add_tail(&edge->list[UPPER], 123484780289SQu Wenruo &cache->pending_edge); 123575bfb9afSJosef Bacik if (list_empty(&upper->upper)) 123684780289SQu Wenruo list_add(&upper->list, &cache->useless_node); 123775bfb9afSJosef Bacik } 123875bfb9afSJosef Bacik 123984780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 124084780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 124175bfb9afSJosef Bacik struct backref_node, list); 124275bfb9afSJosef Bacik list_del_init(&lower->list); 12430fd8c3daSLiu Bo if (lower == node) 12440fd8c3daSLiu Bo node = NULL; 124575bfb9afSJosef Bacik free_backref_node(cache, lower); 12465d4f98a2SYan Zheng } 12470fd8c3daSLiu Bo 12488e19c973SJosef Bacik remove_backref_node(cache, node); 124984780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 125084780289SQu Wenruo list_empty(&cache->pending_edge)); 12515d4f98a2SYan Zheng return ERR_PTR(err); 12525d4f98a2SYan Zheng } 125375bfb9afSJosef Bacik ASSERT(!node || !node->detached); 125484780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 125584780289SQu Wenruo list_empty(&cache->pending_edge)); 12565d4f98a2SYan Zheng return node; 12575d4f98a2SYan Zheng } 12585d4f98a2SYan Zheng 12595d4f98a2SYan Zheng /* 12603fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 12613fd0a558SYan, Zheng * the backref node is created by cloning backref node that 12623fd0a558SYan, Zheng * corresponds to root of source tree 12633fd0a558SYan, Zheng */ 12643fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 12653fd0a558SYan, Zheng struct reloc_control *rc, 12663fd0a558SYan, Zheng struct btrfs_root *src, 12673fd0a558SYan, Zheng struct btrfs_root *dest) 12683fd0a558SYan, Zheng { 12693fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 12703fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 12713fd0a558SYan, Zheng struct backref_node *node = NULL; 12723fd0a558SYan, Zheng struct backref_node *new_node; 12733fd0a558SYan, Zheng struct backref_edge *edge; 12743fd0a558SYan, Zheng struct backref_edge *new_edge; 12753fd0a558SYan, Zheng struct rb_node *rb_node; 12763fd0a558SYan, Zheng 12773fd0a558SYan, Zheng if (cache->last_trans > 0) 12783fd0a558SYan, Zheng update_backref_cache(trans, cache); 12793fd0a558SYan, Zheng 12803fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, src->commit_root->start); 12813fd0a558SYan, Zheng if (rb_node) { 12823fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, rb_node); 12833fd0a558SYan, Zheng if (node->detached) 12843fd0a558SYan, Zheng node = NULL; 12853fd0a558SYan, Zheng else 12863fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 12873fd0a558SYan, Zheng } 12883fd0a558SYan, Zheng 12893fd0a558SYan, Zheng if (!node) { 12903fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, 12913fd0a558SYan, Zheng reloc_root->commit_root->start); 12923fd0a558SYan, Zheng if (rb_node) { 12933fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, 12943fd0a558SYan, Zheng rb_node); 12953fd0a558SYan, Zheng BUG_ON(node->detached); 12963fd0a558SYan, Zheng } 12973fd0a558SYan, Zheng } 12983fd0a558SYan, Zheng 12993fd0a558SYan, Zheng if (!node) 13003fd0a558SYan, Zheng return 0; 13013fd0a558SYan, Zheng 13023fd0a558SYan, Zheng new_node = alloc_backref_node(cache); 13033fd0a558SYan, Zheng if (!new_node) 13043fd0a558SYan, Zheng return -ENOMEM; 13053fd0a558SYan, Zheng 13063fd0a558SYan, Zheng new_node->bytenr = dest->node->start; 13073fd0a558SYan, Zheng new_node->level = node->level; 13083fd0a558SYan, Zheng new_node->lowest = node->lowest; 13096848ad64SYan, Zheng new_node->checked = 1; 131000246528SJosef Bacik new_node->root = btrfs_grab_root(dest); 13110b530bc5SJosef Bacik ASSERT(new_node->root); 13123fd0a558SYan, Zheng 13133fd0a558SYan, Zheng if (!node->lowest) { 13143fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 13153fd0a558SYan, Zheng new_edge = alloc_backref_edge(cache); 13163fd0a558SYan, Zheng if (!new_edge) 13173fd0a558SYan, Zheng goto fail; 13183fd0a558SYan, Zheng 13193fd0a558SYan, Zheng new_edge->node[UPPER] = new_node; 13203fd0a558SYan, Zheng new_edge->node[LOWER] = edge->node[LOWER]; 13213fd0a558SYan, Zheng list_add_tail(&new_edge->list[UPPER], 13223fd0a558SYan, Zheng &new_node->lower); 13233fd0a558SYan, Zheng } 132476b9e23dSMiao Xie } else { 132576b9e23dSMiao Xie list_add_tail(&new_node->lower, &cache->leaves); 13263fd0a558SYan, Zheng } 13273fd0a558SYan, Zheng 13283fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, new_node->bytenr, 13293fd0a558SYan, Zheng &new_node->rb_node); 133043c04fb1SJeff Mahoney if (rb_node) 133143c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, new_node->bytenr); 13323fd0a558SYan, Zheng 13333fd0a558SYan, Zheng if (!new_node->lowest) { 13343fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 13353fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 13363fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 13373fd0a558SYan, Zheng } 13383fd0a558SYan, Zheng } 13393fd0a558SYan, Zheng return 0; 13403fd0a558SYan, Zheng fail: 13413fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 13423fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 13433fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 13443fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 13453fd0a558SYan, Zheng free_backref_edge(cache, new_edge); 13463fd0a558SYan, Zheng } 13473fd0a558SYan, Zheng free_backref_node(cache, new_node); 13483fd0a558SYan, Zheng return -ENOMEM; 13493fd0a558SYan, Zheng } 13503fd0a558SYan, Zheng 13513fd0a558SYan, Zheng /* 13525d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 13535d4f98a2SYan Zheng */ 1354ffd7b339SJeff Mahoney static int __must_check __add_reloc_root(struct btrfs_root *root) 13555d4f98a2SYan Zheng { 13560b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 13575d4f98a2SYan Zheng struct rb_node *rb_node; 13585d4f98a2SYan Zheng struct mapping_node *node; 13590b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 13605d4f98a2SYan Zheng 13615d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 1362ffd7b339SJeff Mahoney if (!node) 1363ffd7b339SJeff Mahoney return -ENOMEM; 13645d4f98a2SYan Zheng 1365ea287ab1SJosef Bacik node->bytenr = root->commit_root->start; 13665d4f98a2SYan Zheng node->data = root; 13675d4f98a2SYan Zheng 13685d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 13695d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 13705d4f98a2SYan Zheng node->bytenr, &node->rb_node); 13715d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 1372ffd7b339SJeff Mahoney if (rb_node) { 13730b246afaSJeff Mahoney btrfs_panic(fs_info, -EEXIST, 13745d163e0eSJeff Mahoney "Duplicate root found for start=%llu while inserting into relocation tree", 13755d163e0eSJeff Mahoney node->bytenr); 1376ffd7b339SJeff Mahoney } 13775d4f98a2SYan Zheng 13785d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 13795d4f98a2SYan Zheng return 0; 13805d4f98a2SYan Zheng } 13815d4f98a2SYan Zheng 13825d4f98a2SYan Zheng /* 1383c974c464SWang Shilong * helper to delete the 'address of tree root -> reloc tree' 13845d4f98a2SYan Zheng * mapping 13855d4f98a2SYan Zheng */ 1386c974c464SWang Shilong static void __del_reloc_root(struct btrfs_root *root) 13875d4f98a2SYan Zheng { 13880b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 13895d4f98a2SYan Zheng struct rb_node *rb_node; 13905d4f98a2SYan Zheng struct mapping_node *node = NULL; 13910b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1392f44deb74SJosef Bacik bool put_ref = false; 13935d4f98a2SYan Zheng 139465c6e82bSQu Wenruo if (rc && root->node) { 13955d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 13965d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, 1397ea287ab1SJosef Bacik root->commit_root->start); 1398c974c464SWang Shilong if (rb_node) { 1399c974c464SWang Shilong node = rb_entry(rb_node, struct mapping_node, rb_node); 1400c974c464SWang Shilong rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 1401ea287ab1SJosef Bacik RB_CLEAR_NODE(&node->rb_node); 1402c974c464SWang Shilong } 1403c974c464SWang Shilong spin_unlock(&rc->reloc_root_tree.lock); 1404c974c464SWang Shilong if (!node) 1405c974c464SWang Shilong return; 1406c974c464SWang Shilong BUG_ON((struct btrfs_root *)node->data != root); 1407389305b2SQu Wenruo } 1408c974c464SWang Shilong 1409f44deb74SJosef Bacik /* 1410f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 1411f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 1412f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 1413f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 1414f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 1415f44deb74SJosef Bacik * messing with the list is in charge of the reference. 1416f44deb74SJosef Bacik */ 14170b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 1418f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 1419f44deb74SJosef Bacik put_ref = true; 1420c974c464SWang Shilong list_del_init(&root->root_list); 1421f44deb74SJosef Bacik } 14220b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 1423f44deb74SJosef Bacik if (put_ref) 1424f44deb74SJosef Bacik btrfs_put_root(root); 1425c974c464SWang Shilong kfree(node); 1426c974c464SWang Shilong } 1427c974c464SWang Shilong 1428c974c464SWang Shilong /* 1429c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 1430c974c464SWang Shilong * mapping 1431c974c464SWang Shilong */ 1432ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 1433c974c464SWang Shilong { 14340b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 1435c974c464SWang Shilong struct rb_node *rb_node; 1436c974c464SWang Shilong struct mapping_node *node = NULL; 14370b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1438c974c464SWang Shilong 1439c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 1440c974c464SWang Shilong rb_node = tree_search(&rc->reloc_root_tree.rb_root, 1441ea287ab1SJosef Bacik root->commit_root->start); 14425d4f98a2SYan Zheng if (rb_node) { 14435d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 14445d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 14455d4f98a2SYan Zheng } 14465d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 14475d4f98a2SYan Zheng 14488f71f3e0SLiu Bo if (!node) 14498f71f3e0SLiu Bo return 0; 14505d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 14515d4f98a2SYan Zheng 14525d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1453ea287ab1SJosef Bacik node->bytenr = root->node->start; 14545d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 14555d4f98a2SYan Zheng node->bytenr, &node->rb_node); 14565d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 145743c04fb1SJeff Mahoney if (rb_node) 145843c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 14595d4f98a2SYan Zheng return 0; 14605d4f98a2SYan Zheng } 14615d4f98a2SYan Zheng 14623fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 14633fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 14645d4f98a2SYan Zheng { 14650b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14665d4f98a2SYan Zheng struct btrfs_root *reloc_root; 14675d4f98a2SYan Zheng struct extent_buffer *eb; 14685d4f98a2SYan Zheng struct btrfs_root_item *root_item; 14695d4f98a2SYan Zheng struct btrfs_key root_key; 14705d4f98a2SYan Zheng int ret; 14715d4f98a2SYan Zheng 14725d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 14735d4f98a2SYan Zheng BUG_ON(!root_item); 14745d4f98a2SYan Zheng 14755d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 14765d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 14773fd0a558SYan, Zheng root_key.offset = objectid; 14785d4f98a2SYan Zheng 14793fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 1480054570a1SFilipe Manana u64 commit_root_gen; 1481054570a1SFilipe Manana 14823fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 14835d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 14845d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 14855d4f98a2SYan Zheng BUG_ON(ret); 1486054570a1SFilipe Manana /* 1487054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 1488054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 1489054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 1490054570a1SFilipe Manana * either inside the critical section of a transaction commit 1491054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 1492054570a1SFilipe Manana * it's created before the transaction commit is started. 1493054570a1SFilipe Manana */ 1494054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 1495054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 14963fd0a558SYan, Zheng } else { 14973fd0a558SYan, Zheng /* 14983fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 14993fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 15003fd0a558SYan, Zheng * modified after it was created have RELOC flag 15013fd0a558SYan, Zheng * set in their headers. so it's OK to not update 15023fd0a558SYan, Zheng * the 'last_snapshot'. 15033fd0a558SYan, Zheng */ 15043fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 15053fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 15063fd0a558SYan, Zheng BUG_ON(ret); 15073fd0a558SYan, Zheng } 15083fd0a558SYan, Zheng 15095d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 15105d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 15115d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 15125d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 15133fd0a558SYan, Zheng 15143fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 15153fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 15163fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 15173fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 15185d4f98a2SYan Zheng root_item->drop_level = 0; 15193fd0a558SYan, Zheng } 15205d4f98a2SYan Zheng 15215d4f98a2SYan Zheng btrfs_tree_unlock(eb); 15225d4f98a2SYan Zheng free_extent_buffer(eb); 15235d4f98a2SYan Zheng 15240b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 15255d4f98a2SYan Zheng &root_key, root_item); 15265d4f98a2SYan Zheng BUG_ON(ret); 15275d4f98a2SYan Zheng kfree(root_item); 15285d4f98a2SYan Zheng 15293dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 15305d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 15313dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 15325d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 15333fd0a558SYan, Zheng return reloc_root; 15343fd0a558SYan, Zheng } 15353fd0a558SYan, Zheng 15363fd0a558SYan, Zheng /* 15373fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 15383fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 1539f44deb74SJosef Bacik * 1540f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 1541f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 15423fd0a558SYan, Zheng */ 15433fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 15443fd0a558SYan, Zheng struct btrfs_root *root) 15453fd0a558SYan, Zheng { 15460b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15473fd0a558SYan, Zheng struct btrfs_root *reloc_root; 15480b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 154920dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 15503fd0a558SYan, Zheng int clear_rsv = 0; 1551ffd7b339SJeff Mahoney int ret; 15523fd0a558SYan, Zheng 1553aec7db3bSJosef Bacik if (!rc) 15542abc726aSJosef Bacik return 0; 15552abc726aSJosef Bacik 15561fac4a54SQu Wenruo /* 15571fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 15581fac4a54SQu Wenruo * create/update the dead reloc tree 15591fac4a54SQu Wenruo */ 15606282675eSQu Wenruo if (reloc_root_is_dead(root)) 15611fac4a54SQu Wenruo return 0; 15621fac4a54SQu Wenruo 1563aec7db3bSJosef Bacik /* 1564aec7db3bSJosef Bacik * This is subtle but important. We do not do 1565aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 1566aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 1567aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 1568aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 1569aec7db3bSJosef Bacik * in. 1570aec7db3bSJosef Bacik */ 15713fd0a558SYan, Zheng if (root->reloc_root) { 15723fd0a558SYan, Zheng reloc_root = root->reloc_root; 15733fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 15743fd0a558SYan, Zheng return 0; 15753fd0a558SYan, Zheng } 15763fd0a558SYan, Zheng 1577aec7db3bSJosef Bacik /* 1578aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 1579aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 1580aec7db3bSJosef Bacik */ 1581aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 1582aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 1583aec7db3bSJosef Bacik return 0; 1584aec7db3bSJosef Bacik 158520dd2cbfSMiao Xie if (!trans->reloc_reserved) { 158620dd2cbfSMiao Xie rsv = trans->block_rsv; 15873fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 15883fd0a558SYan, Zheng clear_rsv = 1; 15893fd0a558SYan, Zheng } 15903fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 15913fd0a558SYan, Zheng if (clear_rsv) 159220dd2cbfSMiao Xie trans->block_rsv = rsv; 15935d4f98a2SYan Zheng 1594ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 1595ffd7b339SJeff Mahoney BUG_ON(ret < 0); 1596f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 15975d4f98a2SYan Zheng return 0; 15985d4f98a2SYan Zheng } 15995d4f98a2SYan Zheng 16005d4f98a2SYan Zheng /* 16015d4f98a2SYan Zheng * update root item of reloc tree 16025d4f98a2SYan Zheng */ 16035d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 16045d4f98a2SYan Zheng struct btrfs_root *root) 16055d4f98a2SYan Zheng { 16060b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 16075d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16085d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16095d4f98a2SYan Zheng int ret; 16105d4f98a2SYan Zheng 16116282675eSQu Wenruo if (!have_reloc_root(root)) 16127585717fSChris Mason goto out; 16135d4f98a2SYan Zheng 16145d4f98a2SYan Zheng reloc_root = root->reloc_root; 16155d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16165d4f98a2SYan Zheng 1617f44deb74SJosef Bacik /* 1618f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 1619f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 1620f44deb74SJosef Bacik * hold it while we update the reloc root. 1621f44deb74SJosef Bacik */ 1622f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 1623f44deb74SJosef Bacik 1624d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 16250b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 16263fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 1627d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 16286282675eSQu Wenruo /* 16296282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 16306282675eSQu Wenruo * have_reloc_root will not touch it from now on. 16316282675eSQu Wenruo */ 16326282675eSQu Wenruo smp_wmb(); 1633c974c464SWang Shilong __del_reloc_root(reloc_root); 16345d4f98a2SYan Zheng } 16355d4f98a2SYan Zheng 16365d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 1637ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 16385d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 16395d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 16405d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 16415d4f98a2SYan Zheng } 16425d4f98a2SYan Zheng 16430b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 16445d4f98a2SYan Zheng &reloc_root->root_key, root_item); 16455d4f98a2SYan Zheng BUG_ON(ret); 1646f44deb74SJosef Bacik btrfs_put_root(reloc_root); 16477585717fSChris Mason out: 16485d4f98a2SYan Zheng return 0; 16495d4f98a2SYan Zheng } 16505d4f98a2SYan Zheng 16515d4f98a2SYan Zheng /* 16525d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 16535d4f98a2SYan Zheng * in a subvolume 16545d4f98a2SYan Zheng */ 16555d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 16565d4f98a2SYan Zheng { 16575d4f98a2SYan Zheng struct rb_node *node; 16585d4f98a2SYan Zheng struct rb_node *prev; 16595d4f98a2SYan Zheng struct btrfs_inode *entry; 16605d4f98a2SYan Zheng struct inode *inode; 16615d4f98a2SYan Zheng 16625d4f98a2SYan Zheng spin_lock(&root->inode_lock); 16635d4f98a2SYan Zheng again: 16645d4f98a2SYan Zheng node = root->inode_tree.rb_node; 16655d4f98a2SYan Zheng prev = NULL; 16665d4f98a2SYan Zheng while (node) { 16675d4f98a2SYan Zheng prev = node; 16685d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 16695d4f98a2SYan Zheng 16704a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 16715d4f98a2SYan Zheng node = node->rb_left; 16724a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 16735d4f98a2SYan Zheng node = node->rb_right; 16745d4f98a2SYan Zheng else 16755d4f98a2SYan Zheng break; 16765d4f98a2SYan Zheng } 16775d4f98a2SYan Zheng if (!node) { 16785d4f98a2SYan Zheng while (prev) { 16795d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 16804a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 16815d4f98a2SYan Zheng node = prev; 16825d4f98a2SYan Zheng break; 16835d4f98a2SYan Zheng } 16845d4f98a2SYan Zheng prev = rb_next(prev); 16855d4f98a2SYan Zheng } 16865d4f98a2SYan Zheng } 16875d4f98a2SYan Zheng while (node) { 16885d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 16895d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 16905d4f98a2SYan Zheng if (inode) { 16915d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 16925d4f98a2SYan Zheng return inode; 16935d4f98a2SYan Zheng } 16945d4f98a2SYan Zheng 16954a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 16965d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 16975d4f98a2SYan Zheng goto again; 16985d4f98a2SYan Zheng 16995d4f98a2SYan Zheng node = rb_next(node); 17005d4f98a2SYan Zheng } 17015d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 17025d4f98a2SYan Zheng return NULL; 17035d4f98a2SYan Zheng } 17045d4f98a2SYan Zheng 17055d4f98a2SYan Zheng /* 17065d4f98a2SYan Zheng * get new location of data 17075d4f98a2SYan Zheng */ 17085d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 17095d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 17105d4f98a2SYan Zheng { 17115d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 17125d4f98a2SYan Zheng struct btrfs_path *path; 17135d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 17145d4f98a2SYan Zheng struct extent_buffer *leaf; 17155d4f98a2SYan Zheng int ret; 17165d4f98a2SYan Zheng 17175d4f98a2SYan Zheng path = btrfs_alloc_path(); 17185d4f98a2SYan Zheng if (!path) 17195d4f98a2SYan Zheng return -ENOMEM; 17205d4f98a2SYan Zheng 17215d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 1722f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 1723f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 17245d4f98a2SYan Zheng if (ret < 0) 17255d4f98a2SYan Zheng goto out; 17265d4f98a2SYan Zheng if (ret > 0) { 17275d4f98a2SYan Zheng ret = -ENOENT; 17285d4f98a2SYan Zheng goto out; 17295d4f98a2SYan Zheng } 17305d4f98a2SYan Zheng 17315d4f98a2SYan Zheng leaf = path->nodes[0]; 17325d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 17335d4f98a2SYan Zheng struct btrfs_file_extent_item); 17345d4f98a2SYan Zheng 17355d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 17365d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 17375d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 17385d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 17395d4f98a2SYan Zheng 17405d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 174183d4cfd4SJosef Bacik ret = -EINVAL; 17425d4f98a2SYan Zheng goto out; 17435d4f98a2SYan Zheng } 17445d4f98a2SYan Zheng 17455d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 17465d4f98a2SYan Zheng ret = 0; 17475d4f98a2SYan Zheng out: 17485d4f98a2SYan Zheng btrfs_free_path(path); 17495d4f98a2SYan Zheng return ret; 17505d4f98a2SYan Zheng } 17515d4f98a2SYan Zheng 17525d4f98a2SYan Zheng /* 17535d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 17545d4f98a2SYan Zheng * the new locations. 17555d4f98a2SYan Zheng */ 17563fd0a558SYan, Zheng static noinline_for_stack 17573fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 17585d4f98a2SYan Zheng struct reloc_control *rc, 17595d4f98a2SYan Zheng struct btrfs_root *root, 17603fd0a558SYan, Zheng struct extent_buffer *leaf) 17615d4f98a2SYan Zheng { 17620b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 17635d4f98a2SYan Zheng struct btrfs_key key; 17645d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 17655d4f98a2SYan Zheng struct inode *inode = NULL; 17665d4f98a2SYan Zheng u64 parent; 17675d4f98a2SYan Zheng u64 bytenr; 17683fd0a558SYan, Zheng u64 new_bytenr = 0; 17695d4f98a2SYan Zheng u64 num_bytes; 17705d4f98a2SYan Zheng u64 end; 17715d4f98a2SYan Zheng u32 nritems; 17725d4f98a2SYan Zheng u32 i; 177383d4cfd4SJosef Bacik int ret = 0; 17745d4f98a2SYan Zheng int first = 1; 17755d4f98a2SYan Zheng int dirty = 0; 17765d4f98a2SYan Zheng 17775d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 17785d4f98a2SYan Zheng return 0; 17795d4f98a2SYan Zheng 17805d4f98a2SYan Zheng /* reloc trees always use full backref */ 17815d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 17825d4f98a2SYan Zheng parent = leaf->start; 17835d4f98a2SYan Zheng else 17845d4f98a2SYan Zheng parent = 0; 17855d4f98a2SYan Zheng 17865d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 17875d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 178882fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 178982fa113fSQu Wenruo 17905d4f98a2SYan Zheng cond_resched(); 17915d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 17925d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 17935d4f98a2SYan Zheng continue; 17945d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 17955d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 17965d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 17975d4f98a2SYan Zheng continue; 17985d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 17995d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 18005d4f98a2SYan Zheng if (bytenr == 0) 18015d4f98a2SYan Zheng continue; 18029569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 18039569cc20SQu Wenruo rc->block_group->length)) 18045d4f98a2SYan Zheng continue; 18055d4f98a2SYan Zheng 18065d4f98a2SYan Zheng /* 18075d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 18085d4f98a2SYan Zheng * to complete and drop the extent cache 18095d4f98a2SYan Zheng */ 18105d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 18115d4f98a2SYan Zheng if (first) { 18125d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 18135d4f98a2SYan Zheng first = 0; 18144a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 18153fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 18165d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 18175d4f98a2SYan Zheng } 18184a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 18195d4f98a2SYan Zheng end = key.offset + 18205d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 18215d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 18220b246afaSJeff Mahoney fs_info->sectorsize)); 18230b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 18245d4f98a2SYan Zheng end--; 18255d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1826d0082371SJeff Mahoney key.offset, end); 18275d4f98a2SYan Zheng if (!ret) 18285d4f98a2SYan Zheng continue; 18295d4f98a2SYan Zheng 1830dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1831dcdbc059SNikolay Borisov key.offset, end, 1); 18325d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1833d0082371SJeff Mahoney key.offset, end); 18345d4f98a2SYan Zheng } 18355d4f98a2SYan Zheng } 18365d4f98a2SYan Zheng 18375d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 18385d4f98a2SYan Zheng bytenr, num_bytes); 183983d4cfd4SJosef Bacik if (ret) { 184083d4cfd4SJosef Bacik /* 184183d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 184283d4cfd4SJosef Bacik * in the file extent yet. 184383d4cfd4SJosef Bacik */ 184483d4cfd4SJosef Bacik break; 18453fd0a558SYan, Zheng } 18465d4f98a2SYan Zheng 18475d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 18485d4f98a2SYan Zheng dirty = 1; 18495d4f98a2SYan Zheng 18505d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 185182fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 185282fa113fSQu Wenruo num_bytes, parent); 185382fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 185482fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1855b06c4bf5SFilipe Manana key.objectid, key.offset); 185682fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 185783d4cfd4SJosef Bacik if (ret) { 185866642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 185983d4cfd4SJosef Bacik break; 186083d4cfd4SJosef Bacik } 18615d4f98a2SYan Zheng 1862ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1863ffd4bb2aSQu Wenruo num_bytes, parent); 1864ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1865ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1866b06c4bf5SFilipe Manana key.objectid, key.offset); 1867ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 186883d4cfd4SJosef Bacik if (ret) { 186966642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 187083d4cfd4SJosef Bacik break; 187183d4cfd4SJosef Bacik } 18725d4f98a2SYan Zheng } 18735d4f98a2SYan Zheng if (dirty) 18745d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 18753fd0a558SYan, Zheng if (inode) 18763fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 187783d4cfd4SJosef Bacik return ret; 18785d4f98a2SYan Zheng } 18795d4f98a2SYan Zheng 18805d4f98a2SYan Zheng static noinline_for_stack 18815d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 18825d4f98a2SYan Zheng struct btrfs_path *path, int level) 18835d4f98a2SYan Zheng { 18845d4f98a2SYan Zheng struct btrfs_disk_key key1; 18855d4f98a2SYan Zheng struct btrfs_disk_key key2; 18865d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 18875d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 18885d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 18895d4f98a2SYan Zheng } 18905d4f98a2SYan Zheng 18915d4f98a2SYan Zheng /* 18925d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 18935d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 18945d4f98a2SYan Zheng * reloc tree was create can be replaced. 18955d4f98a2SYan Zheng * 18965d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 18975d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 18985d4f98a2SYan Zheng * errors, a negative error number is returned. 18995d4f98a2SYan Zheng */ 19003fd0a558SYan, Zheng static noinline_for_stack 19013d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 19025d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 19035d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 19045d4f98a2SYan Zheng int lowest_level, int max_level) 19055d4f98a2SYan Zheng { 19060b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 19075d4f98a2SYan Zheng struct extent_buffer *eb; 19085d4f98a2SYan Zheng struct extent_buffer *parent; 190982fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 19105d4f98a2SYan Zheng struct btrfs_key key; 19115d4f98a2SYan Zheng u64 old_bytenr; 19125d4f98a2SYan Zheng u64 new_bytenr; 19135d4f98a2SYan Zheng u64 old_ptr_gen; 19145d4f98a2SYan Zheng u64 new_ptr_gen; 19155d4f98a2SYan Zheng u64 last_snapshot; 19165d4f98a2SYan Zheng u32 blocksize; 19173fd0a558SYan, Zheng int cow = 0; 19185d4f98a2SYan Zheng int level; 19195d4f98a2SYan Zheng int ret; 19205d4f98a2SYan Zheng int slot; 19215d4f98a2SYan Zheng 19225d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 19235d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 19245d4f98a2SYan Zheng 19255d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 19263fd0a558SYan, Zheng again: 19275d4f98a2SYan Zheng slot = path->slots[lowest_level]; 19285d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 19295d4f98a2SYan Zheng 19305d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 19318bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19325d4f98a2SYan Zheng level = btrfs_header_level(eb); 19335d4f98a2SYan Zheng 19345d4f98a2SYan Zheng if (level < lowest_level) { 19355d4f98a2SYan Zheng btrfs_tree_unlock(eb); 19365d4f98a2SYan Zheng free_extent_buffer(eb); 19375d4f98a2SYan Zheng return 0; 19385d4f98a2SYan Zheng } 19395d4f98a2SYan Zheng 19403fd0a558SYan, Zheng if (cow) { 19415d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 19425d4f98a2SYan Zheng BUG_ON(ret); 19433fd0a558SYan, Zheng } 19448bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19455d4f98a2SYan Zheng 19465d4f98a2SYan Zheng if (next_key) { 19475d4f98a2SYan Zheng next_key->objectid = (u64)-1; 19485d4f98a2SYan Zheng next_key->type = (u8)-1; 19495d4f98a2SYan Zheng next_key->offset = (u64)-1; 19505d4f98a2SYan Zheng } 19515d4f98a2SYan Zheng 19525d4f98a2SYan Zheng parent = eb; 19535d4f98a2SYan Zheng while (1) { 1954581c1760SQu Wenruo struct btrfs_key first_key; 1955581c1760SQu Wenruo 19565d4f98a2SYan Zheng level = btrfs_header_level(parent); 19575d4f98a2SYan Zheng BUG_ON(level < lowest_level); 19585d4f98a2SYan Zheng 19595d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 1960cbca7d59SFilipe Manana if (ret < 0) 1961cbca7d59SFilipe Manana break; 19625d4f98a2SYan Zheng if (ret && slot > 0) 19635d4f98a2SYan Zheng slot--; 19645d4f98a2SYan Zheng 19655d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 19665d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 19675d4f98a2SYan Zheng 19685d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 19690b246afaSJeff Mahoney blocksize = fs_info->nodesize; 19705d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 197117515f1bSQu Wenruo btrfs_node_key_to_cpu(parent, &first_key, slot); 19725d4f98a2SYan Zheng 19735d4f98a2SYan Zheng if (level <= max_level) { 19745d4f98a2SYan Zheng eb = path->nodes[level]; 19755d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 19765d4f98a2SYan Zheng path->slots[level]); 19775d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 19785d4f98a2SYan Zheng path->slots[level]); 19795d4f98a2SYan Zheng } else { 19805d4f98a2SYan Zheng new_bytenr = 0; 19815d4f98a2SYan Zheng new_ptr_gen = 0; 19825d4f98a2SYan Zheng } 19835d4f98a2SYan Zheng 1984fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 19855d4f98a2SYan Zheng ret = level; 19865d4f98a2SYan Zheng break; 19875d4f98a2SYan Zheng } 19885d4f98a2SYan Zheng 19895d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 19905d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 19913fd0a558SYan, Zheng if (level <= lowest_level) { 19925d4f98a2SYan Zheng ret = 0; 19935d4f98a2SYan Zheng break; 19945d4f98a2SYan Zheng } 19955d4f98a2SYan Zheng 1996581c1760SQu Wenruo eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen, 1997581c1760SQu Wenruo level - 1, &first_key); 199864c043deSLiu Bo if (IS_ERR(eb)) { 199964c043deSLiu Bo ret = PTR_ERR(eb); 2000264813acSLiu Bo break; 200164c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 200264c043deSLiu Bo ret = -EIO; 2003416bc658SJosef Bacik free_extent_buffer(eb); 2004379cde74SStefan Behrens break; 2005416bc658SJosef Bacik } 20065d4f98a2SYan Zheng btrfs_tree_lock(eb); 20073fd0a558SYan, Zheng if (cow) { 20085d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 20095d4f98a2SYan Zheng slot, &eb); 20105d4f98a2SYan Zheng BUG_ON(ret); 20115d4f98a2SYan Zheng } 20128bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 20135d4f98a2SYan Zheng 20145d4f98a2SYan Zheng btrfs_tree_unlock(parent); 20155d4f98a2SYan Zheng free_extent_buffer(parent); 20165d4f98a2SYan Zheng 20175d4f98a2SYan Zheng parent = eb; 20185d4f98a2SYan Zheng continue; 20195d4f98a2SYan Zheng } 20205d4f98a2SYan Zheng 20213fd0a558SYan, Zheng if (!cow) { 20223fd0a558SYan, Zheng btrfs_tree_unlock(parent); 20233fd0a558SYan, Zheng free_extent_buffer(parent); 20243fd0a558SYan, Zheng cow = 1; 20253fd0a558SYan, Zheng goto again; 20263fd0a558SYan, Zheng } 20273fd0a558SYan, Zheng 20285d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 20295d4f98a2SYan Zheng path->slots[level]); 2030b3b4aa74SDavid Sterba btrfs_release_path(path); 20315d4f98a2SYan Zheng 20325d4f98a2SYan Zheng path->lowest_level = level; 20335d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 20345d4f98a2SYan Zheng path->lowest_level = 0; 20355d4f98a2SYan Zheng BUG_ON(ret); 20365d4f98a2SYan Zheng 20375d4f98a2SYan Zheng /* 2038824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 2039824d8dffSQu Wenruo * 2040824d8dffSQu Wenruo * We must trace both trees. 2041824d8dffSQu Wenruo * 1) Tree reloc subtree 2042824d8dffSQu Wenruo * If not traced, we will leak data numbers 2043824d8dffSQu Wenruo * 2) Fs subtree 2044824d8dffSQu Wenruo * If not traced, we will double count old data 2045f616f5cdSQu Wenruo * 2046f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 2047f616f5cdSQu Wenruo * the swapped tree blocks. 2048f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 2049f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 2050824d8dffSQu Wenruo */ 2051370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 2052370a11b8SQu Wenruo rc->block_group, parent, slot, 2053370a11b8SQu Wenruo path->nodes[level], path->slots[level], 2054370a11b8SQu Wenruo last_snapshot); 2055370a11b8SQu Wenruo if (ret < 0) 2056370a11b8SQu Wenruo break; 2057824d8dffSQu Wenruo /* 20585d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 20595d4f98a2SYan Zheng */ 20605d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 20615d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 20625d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 20635d4f98a2SYan Zheng 20645d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 20655d4f98a2SYan Zheng path->slots[level], old_bytenr); 20665d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 20675d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 20685d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 20695d4f98a2SYan Zheng 207082fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 207182fa113fSQu Wenruo blocksize, path->nodes[level]->start); 207282fa113fSQu Wenruo ref.skip_qgroup = true; 207382fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 207482fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 20755d4f98a2SYan Zheng BUG_ON(ret); 207682fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 207782fa113fSQu Wenruo blocksize, 0); 207882fa113fSQu Wenruo ref.skip_qgroup = true; 207982fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 208082fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 20815d4f98a2SYan Zheng BUG_ON(ret); 20825d4f98a2SYan Zheng 2083ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 2084ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 2085ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 2086ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2087ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 20885d4f98a2SYan Zheng BUG_ON(ret); 20895d4f98a2SYan Zheng 2090ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 2091ffd4bb2aSQu Wenruo blocksize, 0); 2092ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 2093ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2094ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 20955d4f98a2SYan Zheng BUG_ON(ret); 20965d4f98a2SYan Zheng 20975d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 20985d4f98a2SYan Zheng 20995d4f98a2SYan Zheng ret = level; 21005d4f98a2SYan Zheng break; 21015d4f98a2SYan Zheng } 21025d4f98a2SYan Zheng btrfs_tree_unlock(parent); 21035d4f98a2SYan Zheng free_extent_buffer(parent); 21045d4f98a2SYan Zheng return ret; 21055d4f98a2SYan Zheng } 21065d4f98a2SYan Zheng 21075d4f98a2SYan Zheng /* 21085d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 21095d4f98a2SYan Zheng */ 21105d4f98a2SYan Zheng static noinline_for_stack 21115d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 21125d4f98a2SYan Zheng int *level) 21135d4f98a2SYan Zheng { 21145d4f98a2SYan Zheng struct extent_buffer *eb; 21155d4f98a2SYan Zheng int i; 21165d4f98a2SYan Zheng u64 last_snapshot; 21175d4f98a2SYan Zheng u32 nritems; 21185d4f98a2SYan Zheng 21195d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 21205d4f98a2SYan Zheng 21215d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 21225d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 21235d4f98a2SYan Zheng path->nodes[i] = NULL; 21245d4f98a2SYan Zheng } 21255d4f98a2SYan Zheng 21265d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 21275d4f98a2SYan Zheng eb = path->nodes[i]; 21285d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 21295d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 21305d4f98a2SYan Zheng path->slots[i]++; 21315d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 21325d4f98a2SYan Zheng last_snapshot) 21335d4f98a2SYan Zheng continue; 21345d4f98a2SYan Zheng 21355d4f98a2SYan Zheng *level = i; 21365d4f98a2SYan Zheng return 0; 21375d4f98a2SYan Zheng } 21385d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 21395d4f98a2SYan Zheng path->nodes[i] = NULL; 21405d4f98a2SYan Zheng } 21415d4f98a2SYan Zheng return 1; 21425d4f98a2SYan Zheng } 21435d4f98a2SYan Zheng 21445d4f98a2SYan Zheng /* 21455d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 21465d4f98a2SYan Zheng */ 21475d4f98a2SYan Zheng static noinline_for_stack 21485d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 21495d4f98a2SYan Zheng int *level) 21505d4f98a2SYan Zheng { 21512ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 21525d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 21535d4f98a2SYan Zheng int i; 21545d4f98a2SYan Zheng u64 bytenr; 21555d4f98a2SYan Zheng u64 ptr_gen = 0; 21565d4f98a2SYan Zheng u64 last_snapshot; 21575d4f98a2SYan Zheng u32 nritems; 21585d4f98a2SYan Zheng 21595d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 21605d4f98a2SYan Zheng 21615d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 2162581c1760SQu Wenruo struct btrfs_key first_key; 2163581c1760SQu Wenruo 21645d4f98a2SYan Zheng eb = path->nodes[i]; 21655d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 21665d4f98a2SYan Zheng while (path->slots[i] < nritems) { 21675d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 21685d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 21695d4f98a2SYan Zheng break; 21705d4f98a2SYan Zheng path->slots[i]++; 21715d4f98a2SYan Zheng } 21725d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 21735d4f98a2SYan Zheng if (i == *level) 21745d4f98a2SYan Zheng break; 21755d4f98a2SYan Zheng *level = i + 1; 21765d4f98a2SYan Zheng return 0; 21775d4f98a2SYan Zheng } 21785d4f98a2SYan Zheng if (i == 1) { 21795d4f98a2SYan Zheng *level = i; 21805d4f98a2SYan Zheng return 0; 21815d4f98a2SYan Zheng } 21825d4f98a2SYan Zheng 21835d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 2184581c1760SQu Wenruo btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]); 2185581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1, 2186581c1760SQu Wenruo &first_key); 218764c043deSLiu Bo if (IS_ERR(eb)) { 218864c043deSLiu Bo return PTR_ERR(eb); 218964c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2190416bc658SJosef Bacik free_extent_buffer(eb); 2191416bc658SJosef Bacik return -EIO; 2192416bc658SJosef Bacik } 21935d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 21945d4f98a2SYan Zheng path->nodes[i - 1] = eb; 21955d4f98a2SYan Zheng path->slots[i - 1] = 0; 21965d4f98a2SYan Zheng } 21975d4f98a2SYan Zheng return 1; 21985d4f98a2SYan Zheng } 21995d4f98a2SYan Zheng 22005d4f98a2SYan Zheng /* 22015d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 22025d4f98a2SYan Zheng * [min_key, max_key) 22035d4f98a2SYan Zheng */ 22045d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 22055d4f98a2SYan Zheng struct btrfs_key *min_key, 22065d4f98a2SYan Zheng struct btrfs_key *max_key) 22075d4f98a2SYan Zheng { 22080b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 22095d4f98a2SYan Zheng struct inode *inode = NULL; 22105d4f98a2SYan Zheng u64 objectid; 22115d4f98a2SYan Zheng u64 start, end; 221233345d01SLi Zefan u64 ino; 22135d4f98a2SYan Zheng 22145d4f98a2SYan Zheng objectid = min_key->objectid; 22155d4f98a2SYan Zheng while (1) { 22165d4f98a2SYan Zheng cond_resched(); 22175d4f98a2SYan Zheng iput(inode); 22185d4f98a2SYan Zheng 22195d4f98a2SYan Zheng if (objectid > max_key->objectid) 22205d4f98a2SYan Zheng break; 22215d4f98a2SYan Zheng 22225d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 22235d4f98a2SYan Zheng if (!inode) 22245d4f98a2SYan Zheng break; 22254a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 22265d4f98a2SYan Zheng 222733345d01SLi Zefan if (ino > max_key->objectid) { 22285d4f98a2SYan Zheng iput(inode); 22295d4f98a2SYan Zheng break; 22305d4f98a2SYan Zheng } 22315d4f98a2SYan Zheng 223233345d01SLi Zefan objectid = ino + 1; 22335d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 22345d4f98a2SYan Zheng continue; 22355d4f98a2SYan Zheng 223633345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 22375d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 22385d4f98a2SYan Zheng continue; 22395d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 22405d4f98a2SYan Zheng start = 0; 22415d4f98a2SYan Zheng else { 22425d4f98a2SYan Zheng start = min_key->offset; 22430b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 22445d4f98a2SYan Zheng } 22455d4f98a2SYan Zheng } else { 22465d4f98a2SYan Zheng start = 0; 22475d4f98a2SYan Zheng } 22485d4f98a2SYan Zheng 224933345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 22505d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 22515d4f98a2SYan Zheng continue; 22525d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 22535d4f98a2SYan Zheng end = (u64)-1; 22545d4f98a2SYan Zheng } else { 22555d4f98a2SYan Zheng if (max_key->offset == 0) 22565d4f98a2SYan Zheng continue; 22575d4f98a2SYan Zheng end = max_key->offset; 22580b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 22595d4f98a2SYan Zheng end--; 22605d4f98a2SYan Zheng } 22615d4f98a2SYan Zheng } else { 22625d4f98a2SYan Zheng end = (u64)-1; 22635d4f98a2SYan Zheng } 22645d4f98a2SYan Zheng 22655d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 2266d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 2267dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 2268d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 22695d4f98a2SYan Zheng } 22705d4f98a2SYan Zheng return 0; 22715d4f98a2SYan Zheng } 22725d4f98a2SYan Zheng 22735d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 22745d4f98a2SYan Zheng struct btrfs_key *key) 22755d4f98a2SYan Zheng 22765d4f98a2SYan Zheng { 22775d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 22785d4f98a2SYan Zheng if (!path->nodes[level]) 22795d4f98a2SYan Zheng break; 22805d4f98a2SYan Zheng if (path->slots[level] + 1 < 22815d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 22825d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 22835d4f98a2SYan Zheng path->slots[level] + 1); 22845d4f98a2SYan Zheng return 0; 22855d4f98a2SYan Zheng } 22865d4f98a2SYan Zheng level++; 22875d4f98a2SYan Zheng } 22885d4f98a2SYan Zheng return 1; 22895d4f98a2SYan Zheng } 22905d4f98a2SYan Zheng 22915d4f98a2SYan Zheng /* 2292d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 2293d2311e69SQu Wenruo */ 2294d2311e69SQu Wenruo static void insert_dirty_subvol(struct btrfs_trans_handle *trans, 2295d2311e69SQu Wenruo struct reloc_control *rc, 2296d2311e69SQu Wenruo struct btrfs_root *root) 2297d2311e69SQu Wenruo { 2298d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2299d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 2300d2311e69SQu Wenruo 2301d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 2302d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 2303d2311e69SQu Wenruo ASSERT(reloc_root); 2304d2311e69SQu Wenruo 2305d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 2306d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 2307d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 2308d2311e69SQu Wenruo reloc_root_item->drop_level = 0; 2309d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 2310d2311e69SQu Wenruo btrfs_update_reloc_root(trans, root); 2311d2311e69SQu Wenruo 2312d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 231300246528SJosef Bacik btrfs_grab_root(root); 2314d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 2315d2311e69SQu Wenruo } 2316d2311e69SQu Wenruo } 2317d2311e69SQu Wenruo 2318d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 2319d2311e69SQu Wenruo { 2320d2311e69SQu Wenruo struct btrfs_root *root; 2321d2311e69SQu Wenruo struct btrfs_root *next; 2322d2311e69SQu Wenruo int ret = 0; 232330d40577SQu Wenruo int ret2; 2324d2311e69SQu Wenruo 2325d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 2326d2311e69SQu Wenruo reloc_dirty_list) { 232730d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 232830d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 2329d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2330d2311e69SQu Wenruo 2331d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 2332d2311e69SQu Wenruo root->reloc_root = NULL; 23336282675eSQu Wenruo /* 23346282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 23356282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 23366282675eSQu Wenruo */ 23376282675eSQu Wenruo smp_wmb(); 23381fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 2339f28de8d8SJosef Bacik if (reloc_root) { 2340f44deb74SJosef Bacik /* 2341f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 2342f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 2343f44deb74SJosef Bacik * drop the ref ourselves. 2344f44deb74SJosef Bacik */ 2345f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 2346f44deb74SJosef Bacik if (ret2 < 0) { 2347f44deb74SJosef Bacik btrfs_put_root(reloc_root); 2348f44deb74SJosef Bacik if (!ret) 2349f28de8d8SJosef Bacik ret = ret2; 2350f28de8d8SJosef Bacik } 2351f44deb74SJosef Bacik } 235200246528SJosef Bacik btrfs_put_root(root); 235330d40577SQu Wenruo } else { 235430d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 23550078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 2356f44deb74SJosef Bacik if (ret2 < 0) { 2357f44deb74SJosef Bacik btrfs_put_root(root); 2358f44deb74SJosef Bacik if (!ret) 235930d40577SQu Wenruo ret = ret2; 236030d40577SQu Wenruo } 2361d2311e69SQu Wenruo } 2362f44deb74SJosef Bacik } 2363d2311e69SQu Wenruo return ret; 2364d2311e69SQu Wenruo } 2365d2311e69SQu Wenruo 2366d2311e69SQu Wenruo /* 23675d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 23685d4f98a2SYan Zheng * fs tree. 23695d4f98a2SYan Zheng */ 23705d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 23715d4f98a2SYan Zheng struct btrfs_root *root) 23725d4f98a2SYan Zheng { 23730b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 23745d4f98a2SYan Zheng struct btrfs_key key; 23755d4f98a2SYan Zheng struct btrfs_key next_key; 23769e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 23775d4f98a2SYan Zheng struct btrfs_root *reloc_root; 23785d4f98a2SYan Zheng struct btrfs_root_item *root_item; 23795d4f98a2SYan Zheng struct btrfs_path *path; 23803fd0a558SYan, Zheng struct extent_buffer *leaf; 23815d4f98a2SYan Zheng int level; 23825d4f98a2SYan Zheng int max_level; 23835d4f98a2SYan Zheng int replaced = 0; 23845d4f98a2SYan Zheng int ret; 23855d4f98a2SYan Zheng int err = 0; 23863fd0a558SYan, Zheng u32 min_reserved; 23875d4f98a2SYan Zheng 23885d4f98a2SYan Zheng path = btrfs_alloc_path(); 23895d4f98a2SYan Zheng if (!path) 23905d4f98a2SYan Zheng return -ENOMEM; 2391e4058b54SDavid Sterba path->reada = READA_FORWARD; 23925d4f98a2SYan Zheng 23935d4f98a2SYan Zheng reloc_root = root->reloc_root; 23945d4f98a2SYan Zheng root_item = &reloc_root->root_item; 23955d4f98a2SYan Zheng 23965d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 23975d4f98a2SYan Zheng level = btrfs_root_level(root_item); 239867439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 23995d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 24005d4f98a2SYan Zheng path->slots[level] = 0; 24015d4f98a2SYan Zheng } else { 24025d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 24035d4f98a2SYan Zheng 24045d4f98a2SYan Zheng level = root_item->drop_level; 24055d4f98a2SYan Zheng BUG_ON(level == 0); 24065d4f98a2SYan Zheng path->lowest_level = level; 24075d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 240833c66f43SYan Zheng path->lowest_level = 0; 24095d4f98a2SYan Zheng if (ret < 0) { 24105d4f98a2SYan Zheng btrfs_free_path(path); 24115d4f98a2SYan Zheng return ret; 24125d4f98a2SYan Zheng } 24135d4f98a2SYan Zheng 24145d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 24155d4f98a2SYan Zheng path->slots[level]); 24165d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 24175d4f98a2SYan Zheng 24185d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 24195d4f98a2SYan Zheng } 24205d4f98a2SYan Zheng 24210b246afaSJeff Mahoney min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 24225d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 24235d4f98a2SYan Zheng 24245d4f98a2SYan Zheng while (1) { 242508e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 242608e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 24273fd0a558SYan, Zheng if (ret) { 24289e6a0c52SJosef Bacik err = ret; 24299e6a0c52SJosef Bacik goto out; 24303fd0a558SYan, Zheng } 24319e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 24329e6a0c52SJosef Bacik if (IS_ERR(trans)) { 24339e6a0c52SJosef Bacik err = PTR_ERR(trans); 24349e6a0c52SJosef Bacik trans = NULL; 24359e6a0c52SJosef Bacik goto out; 24369e6a0c52SJosef Bacik } 24372abc726aSJosef Bacik 24382abc726aSJosef Bacik /* 24392abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 24402abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 24412abc726aSJosef Bacik * 24422abc726aSJosef Bacik * But that's ok, we started the trans handle on our 24432abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 24442abc726aSJosef Bacik * dirty list. At commit time we'll still call 24452abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 24462abc726aSJosef Bacik * appropriately. 24472abc726aSJosef Bacik */ 24482abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 24499e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 24503fd0a558SYan, Zheng 24513fd0a558SYan, Zheng replaced = 0; 24525d4f98a2SYan Zheng max_level = level; 24535d4f98a2SYan Zheng 24545d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 24555d4f98a2SYan Zheng if (ret < 0) { 24565d4f98a2SYan Zheng err = ret; 24575d4f98a2SYan Zheng goto out; 24585d4f98a2SYan Zheng } 24595d4f98a2SYan Zheng if (ret > 0) 24605d4f98a2SYan Zheng break; 24615d4f98a2SYan Zheng 24625d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 24635d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 24645d4f98a2SYan Zheng ret = 0; 24655d4f98a2SYan Zheng } else { 24663d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 24673fd0a558SYan, Zheng &next_key, level, max_level); 24685d4f98a2SYan Zheng } 24695d4f98a2SYan Zheng if (ret < 0) { 24705d4f98a2SYan Zheng err = ret; 24715d4f98a2SYan Zheng goto out; 24725d4f98a2SYan Zheng } 24735d4f98a2SYan Zheng 24745d4f98a2SYan Zheng if (ret > 0) { 24755d4f98a2SYan Zheng level = ret; 24765d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 24775d4f98a2SYan Zheng path->slots[level]); 24785d4f98a2SYan Zheng replaced = 1; 24795d4f98a2SYan Zheng } 24805d4f98a2SYan Zheng 24815d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 24825d4f98a2SYan Zheng if (ret > 0) 24835d4f98a2SYan Zheng break; 24845d4f98a2SYan Zheng 24855d4f98a2SYan Zheng BUG_ON(level == 0); 24865d4f98a2SYan Zheng /* 24875d4f98a2SYan Zheng * save the merging progress in the drop_progress. 24885d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 24895d4f98a2SYan Zheng */ 24905d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 24915d4f98a2SYan Zheng path->slots[level]); 24925d4f98a2SYan Zheng root_item->drop_level = level; 24935d4f98a2SYan Zheng 24943a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 24959e6a0c52SJosef Bacik trans = NULL; 24965d4f98a2SYan Zheng 24972ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 24985d4f98a2SYan Zheng 24995d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 25005d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 25015d4f98a2SYan Zheng } 25025d4f98a2SYan Zheng 25035d4f98a2SYan Zheng /* 25045d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 25055d4f98a2SYan Zheng * relocated and the block is tree root. 25065d4f98a2SYan Zheng */ 25075d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 25085d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 25095d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 25105d4f98a2SYan Zheng free_extent_buffer(leaf); 25115d4f98a2SYan Zheng if (ret < 0) 25125d4f98a2SYan Zheng err = ret; 25135d4f98a2SYan Zheng out: 25145d4f98a2SYan Zheng btrfs_free_path(path); 25155d4f98a2SYan Zheng 2516d2311e69SQu Wenruo if (err == 0) 2517d2311e69SQu Wenruo insert_dirty_subvol(trans, rc, root); 25185d4f98a2SYan Zheng 25199e6a0c52SJosef Bacik if (trans) 25203a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 25215d4f98a2SYan Zheng 25222ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 25235d4f98a2SYan Zheng 25245d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 25255d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 25265d4f98a2SYan Zheng 25275d4f98a2SYan Zheng return err; 25285d4f98a2SYan Zheng } 25295d4f98a2SYan Zheng 25303fd0a558SYan, Zheng static noinline_for_stack 25313fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 25325d4f98a2SYan Zheng { 25333fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 25340b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 25353fd0a558SYan, Zheng struct btrfs_root *reloc_root; 25365d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 25373fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 25383fd0a558SYan, Zheng u64 num_bytes = 0; 25393fd0a558SYan, Zheng int ret; 25403fd0a558SYan, Zheng 25410b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 25420b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 25433fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 25440b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 25457585717fSChris Mason 25463fd0a558SYan, Zheng again: 25473fd0a558SYan, Zheng if (!err) { 25483fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 254908e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 255008e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 25513fd0a558SYan, Zheng if (ret) 25523fd0a558SYan, Zheng err = ret; 25533fd0a558SYan, Zheng } 25543fd0a558SYan, Zheng 25557a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 25563612b495STsutomu Itoh if (IS_ERR(trans)) { 25573612b495STsutomu Itoh if (!err) 25582ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 255963f018beSNikolay Borisov num_bytes, NULL); 25603612b495STsutomu Itoh return PTR_ERR(trans); 25613612b495STsutomu Itoh } 25623fd0a558SYan, Zheng 25633fd0a558SYan, Zheng if (!err) { 25643fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 25653a45bb20SJeff Mahoney btrfs_end_transaction(trans); 25662ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 256763f018beSNikolay Borisov num_bytes, NULL); 25683fd0a558SYan, Zheng goto again; 25693fd0a558SYan, Zheng } 25703fd0a558SYan, Zheng } 25713fd0a558SYan, Zheng 25723fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 25733fd0a558SYan, Zheng 25743fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 25753fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 25763fd0a558SYan, Zheng struct btrfs_root, root_list); 25773fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 25783fd0a558SYan, Zheng 25790b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 25803fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 25813fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 25823fd0a558SYan, Zheng 25833fd0a558SYan, Zheng /* 25843fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 25853fd0a558SYan, Zheng * knows it should resumes merging 25863fd0a558SYan, Zheng */ 25873fd0a558SYan, Zheng if (!err) 25883fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 25893fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 25903fd0a558SYan, Zheng 25913fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 259200246528SJosef Bacik btrfs_put_root(root); 25933fd0a558SYan, Zheng } 25943fd0a558SYan, Zheng 25953fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 25963fd0a558SYan, Zheng 25973fd0a558SYan, Zheng if (!err) 25983a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 25993fd0a558SYan, Zheng else 26003a45bb20SJeff Mahoney btrfs_end_transaction(trans); 26013fd0a558SYan, Zheng return err; 26023fd0a558SYan, Zheng } 26033fd0a558SYan, Zheng 26043fd0a558SYan, Zheng static noinline_for_stack 2605aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 2606aca1bba6SLiu Bo { 2607aca1bba6SLiu Bo struct btrfs_root *reloc_root; 2608aca1bba6SLiu Bo 2609aca1bba6SLiu Bo while (!list_empty(list)) { 2610aca1bba6SLiu Bo reloc_root = list_entry(list->next, struct btrfs_root, 2611aca1bba6SLiu Bo root_list); 2612bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 2613aca1bba6SLiu Bo } 2614aca1bba6SLiu Bo } 2615aca1bba6SLiu Bo 2616aca1bba6SLiu Bo static noinline_for_stack 261794404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 26183fd0a558SYan, Zheng { 26190b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 26205d4f98a2SYan Zheng struct btrfs_root *root; 26215d4f98a2SYan Zheng struct btrfs_root *reloc_root; 26223fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 26233fd0a558SYan, Zheng int found = 0; 2624aca1bba6SLiu Bo int ret = 0; 26253fd0a558SYan, Zheng again: 26263fd0a558SYan, Zheng root = rc->extent_root; 26277585717fSChris Mason 26287585717fSChris Mason /* 26297585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 26307585717fSChris Mason * we have to make sure nobody is in the middle of 26317585717fSChris Mason * adding their roots to the list while we are 26327585717fSChris Mason * doing this splice 26337585717fSChris Mason */ 26340b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 26353fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 26360b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 26375d4f98a2SYan Zheng 26383fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 26393fd0a558SYan, Zheng found = 1; 26403fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 26413fd0a558SYan, Zheng struct btrfs_root, root_list); 26425d4f98a2SYan Zheng 26435d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 26440b246afaSJeff Mahoney root = read_fs_root(fs_info, 26455d4f98a2SYan Zheng reloc_root->root_key.offset); 26465d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 26475d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 26485d4f98a2SYan Zheng 26493fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 265000246528SJosef Bacik btrfs_put_root(root); 2651b37b39cdSJosef Bacik if (ret) { 265225e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 265325e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 265425e293c2SWang Shilong &reloc_roots); 2655aca1bba6SLiu Bo goto out; 2656b37b39cdSJosef Bacik } 26573fd0a558SYan, Zheng } else { 26583fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 265930d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 266030d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 266130d40577SQu Wenruo &rc->dirty_subvol_roots); 26623fd0a558SYan, Zheng } 26635d4f98a2SYan Zheng } 26645d4f98a2SYan Zheng 26653fd0a558SYan, Zheng if (found) { 26663fd0a558SYan, Zheng found = 0; 26673fd0a558SYan, Zheng goto again; 26685d4f98a2SYan Zheng } 2669aca1bba6SLiu Bo out: 2670aca1bba6SLiu Bo if (ret) { 26710b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 2672aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 2673aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 2674467bb1d2SWang Shilong 2675467bb1d2SWang Shilong /* new reloc root may be added */ 26760b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 2677467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 26780b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 2679467bb1d2SWang Shilong if (!list_empty(&reloc_roots)) 2680467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 2681aca1bba6SLiu Bo } 2682aca1bba6SLiu Bo 26837b7b7431SJosef Bacik /* 26847b7b7431SJosef Bacik * We used to have 26857b7b7431SJosef Bacik * 26867b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 26877b7b7431SJosef Bacik * 26887b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 26897b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 26907b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 26917b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 26927b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 26937b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 26947b7b7431SJosef Bacik * will be cleaned up on unmount. 26957b7b7431SJosef Bacik * 26967b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 26977b7b7431SJosef Bacik */ 26985d4f98a2SYan Zheng } 26995d4f98a2SYan Zheng 27005d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 27015d4f98a2SYan Zheng { 27025d4f98a2SYan Zheng struct tree_block *block; 27035d4f98a2SYan Zheng struct rb_node *rb_node; 27045d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 27055d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 27065d4f98a2SYan Zheng rb_erase(rb_node, blocks); 27075d4f98a2SYan Zheng kfree(block); 27085d4f98a2SYan Zheng } 27095d4f98a2SYan Zheng } 27105d4f98a2SYan Zheng 27115d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 27125d4f98a2SYan Zheng struct btrfs_root *reloc_root) 27135d4f98a2SYan Zheng { 27140b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 27155d4f98a2SYan Zheng struct btrfs_root *root; 2716442b1ac5SJosef Bacik int ret; 27175d4f98a2SYan Zheng 27185d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 27195d4f98a2SYan Zheng return 0; 27205d4f98a2SYan Zheng 27210b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 27225d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 27235d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 2724442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 272500246528SJosef Bacik btrfs_put_root(root); 27265d4f98a2SYan Zheng 2727442b1ac5SJosef Bacik return ret; 27285d4f98a2SYan Zheng } 27295d4f98a2SYan Zheng 27303fd0a558SYan, Zheng static noinline_for_stack 27313fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 27323fd0a558SYan, Zheng struct reloc_control *rc, 27335d4f98a2SYan Zheng struct backref_node *node, 2734dc4103f9SWang Shilong struct backref_edge *edges[]) 27355d4f98a2SYan Zheng { 27365d4f98a2SYan Zheng struct backref_node *next; 27375d4f98a2SYan Zheng struct btrfs_root *root; 27383fd0a558SYan, Zheng int index = 0; 27393fd0a558SYan, Zheng 27405d4f98a2SYan Zheng next = node; 27415d4f98a2SYan Zheng while (1) { 27425d4f98a2SYan Zheng cond_resched(); 27435d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 27445d4f98a2SYan Zheng root = next->root; 27453fd0a558SYan, Zheng BUG_ON(!root); 274627cdeb70SMiao Xie BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state)); 27475d4f98a2SYan Zheng 27485d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 27495d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 27505d4f98a2SYan Zheng break; 27515d4f98a2SYan Zheng } 27525d4f98a2SYan Zheng 27535d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 27543fd0a558SYan, Zheng root = root->reloc_root; 27553fd0a558SYan, Zheng 27563fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 27573fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 27583fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 27593fd0a558SYan, Zheng next->new_bytenr = root->node->start; 276000246528SJosef Bacik btrfs_put_root(next->root); 276100246528SJosef Bacik next->root = btrfs_grab_root(root); 27620b530bc5SJosef Bacik ASSERT(next->root); 27633fd0a558SYan, Zheng list_add_tail(&next->list, 27643fd0a558SYan, Zheng &rc->backref_cache.changed); 27659569cc20SQu Wenruo mark_block_processed(rc, next); 27665d4f98a2SYan Zheng break; 27675d4f98a2SYan Zheng } 27685d4f98a2SYan Zheng 27693fd0a558SYan, Zheng WARN_ON(1); 27705d4f98a2SYan Zheng root = NULL; 27715d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 27725d4f98a2SYan Zheng if (!next || next->level <= node->level) 27735d4f98a2SYan Zheng break; 27745d4f98a2SYan Zheng } 27753fd0a558SYan, Zheng if (!root) 27763fd0a558SYan, Zheng return NULL; 27775d4f98a2SYan Zheng 27783fd0a558SYan, Zheng next = node; 27793fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 27803fd0a558SYan, Zheng while (1) { 27813fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 27823fd0a558SYan, Zheng if (--index < 0) 27833fd0a558SYan, Zheng break; 27843fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 27853fd0a558SYan, Zheng } 27865d4f98a2SYan Zheng return root; 27875d4f98a2SYan Zheng } 27885d4f98a2SYan Zheng 27893fd0a558SYan, Zheng /* 27903fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 27913fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 27923fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 27933fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 27943fd0a558SYan, Zheng */ 27955d4f98a2SYan Zheng static noinline_for_stack 2796147d256eSZhaolei struct btrfs_root *select_one_root(struct backref_node *node) 27975d4f98a2SYan Zheng { 27983fd0a558SYan, Zheng struct backref_node *next; 27993fd0a558SYan, Zheng struct btrfs_root *root; 28003fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 28015d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 28023fd0a558SYan, Zheng int index = 0; 28033fd0a558SYan, Zheng 28043fd0a558SYan, Zheng next = node; 28053fd0a558SYan, Zheng while (1) { 28063fd0a558SYan, Zheng cond_resched(); 28073fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 28083fd0a558SYan, Zheng root = next->root; 28093fd0a558SYan, Zheng BUG_ON(!root); 28103fd0a558SYan, Zheng 281125985edcSLucas De Marchi /* no other choice for non-references counted tree */ 281227cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 28133fd0a558SYan, Zheng return root; 28143fd0a558SYan, Zheng 28153fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 28163fd0a558SYan, Zheng fs_root = root; 28173fd0a558SYan, Zheng 28183fd0a558SYan, Zheng if (next != node) 28193fd0a558SYan, Zheng return NULL; 28203fd0a558SYan, Zheng 28213fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 28223fd0a558SYan, Zheng if (!next || next->level <= node->level) 28233fd0a558SYan, Zheng break; 28243fd0a558SYan, Zheng } 28253fd0a558SYan, Zheng 28263fd0a558SYan, Zheng if (!fs_root) 28273fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 28283fd0a558SYan, Zheng return fs_root; 28295d4f98a2SYan Zheng } 28305d4f98a2SYan Zheng 28315d4f98a2SYan Zheng static noinline_for_stack 28323fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 28333fd0a558SYan, Zheng struct backref_node *node, int reserve) 28345d4f98a2SYan Zheng { 28350b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 28363fd0a558SYan, Zheng struct backref_node *next = node; 28373fd0a558SYan, Zheng struct backref_edge *edge; 28383fd0a558SYan, Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 28393fd0a558SYan, Zheng u64 num_bytes = 0; 28403fd0a558SYan, Zheng int index = 0; 28415d4f98a2SYan Zheng 28423fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 28433fd0a558SYan, Zheng 28443fd0a558SYan, Zheng while (next) { 28453fd0a558SYan, Zheng cond_resched(); 28465d4f98a2SYan Zheng while (1) { 28473fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 28485d4f98a2SYan Zheng break; 28495d4f98a2SYan Zheng 28500b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 28513fd0a558SYan, Zheng 28523fd0a558SYan, Zheng if (list_empty(&next->upper)) 28533fd0a558SYan, Zheng break; 28543fd0a558SYan, Zheng 28553fd0a558SYan, Zheng edge = list_entry(next->upper.next, 28563fd0a558SYan, Zheng struct backref_edge, list[LOWER]); 28573fd0a558SYan, Zheng edges[index++] = edge; 28583fd0a558SYan, Zheng next = edge->node[UPPER]; 28595d4f98a2SYan Zheng } 28603fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 28613fd0a558SYan, Zheng } 28623fd0a558SYan, Zheng return num_bytes; 28633fd0a558SYan, Zheng } 28643fd0a558SYan, Zheng 28653fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 28663fd0a558SYan, Zheng struct reloc_control *rc, 28673fd0a558SYan, Zheng struct backref_node *node) 28683fd0a558SYan, Zheng { 28693fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2870da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 28713fd0a558SYan, Zheng u64 num_bytes; 28723fd0a558SYan, Zheng int ret; 28730647bf56SWang Shilong u64 tmp; 28743fd0a558SYan, Zheng 28753fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 28763fd0a558SYan, Zheng 28773fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 28780647bf56SWang Shilong rc->reserved_bytes += num_bytes; 28798ca17f0fSJosef Bacik 28808ca17f0fSJosef Bacik /* 28818ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 28828ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 28838ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 28848ca17f0fSJosef Bacik */ 28850647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 28868ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 28873fd0a558SYan, Zheng if (ret) { 2888da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 28890647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 28900647bf56SWang Shilong tmp <<= 1; 28910647bf56SWang Shilong /* 28920647bf56SWang Shilong * only one thread can access block_rsv at this point, 28930647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 28940647bf56SWang Shilong * we expand more reservation size here to allow enough 289552042d8eSAndrea Gelmini * space for relocation and we will return earlier in 28960647bf56SWang Shilong * enospc case. 28970647bf56SWang Shilong */ 2898da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 28990647bf56SWang Shilong RELOCATION_RESERVED_NODES; 29008ca17f0fSJosef Bacik return -EAGAIN; 29013fd0a558SYan, Zheng } 29023fd0a558SYan, Zheng 29033fd0a558SYan, Zheng return 0; 29043fd0a558SYan, Zheng } 29053fd0a558SYan, Zheng 29065d4f98a2SYan Zheng /* 29075d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 29085d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 29095d4f98a2SYan Zheng * 29105d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 29115d4f98a2SYan Zheng * in that case this function just updates pointers. 29125d4f98a2SYan Zheng */ 29135d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 29143fd0a558SYan, Zheng struct reloc_control *rc, 29155d4f98a2SYan Zheng struct backref_node *node, 29165d4f98a2SYan Zheng struct btrfs_key *key, 29175d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 29185d4f98a2SYan Zheng { 29192ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 29205d4f98a2SYan Zheng struct backref_node *upper; 29215d4f98a2SYan Zheng struct backref_edge *edge; 29225d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 29235d4f98a2SYan Zheng struct btrfs_root *root; 29245d4f98a2SYan Zheng struct extent_buffer *eb; 29255d4f98a2SYan Zheng u32 blocksize; 29265d4f98a2SYan Zheng u64 bytenr; 29275d4f98a2SYan Zheng u64 generation; 29285d4f98a2SYan Zheng int slot; 29295d4f98a2SYan Zheng int ret; 29305d4f98a2SYan Zheng int err = 0; 29315d4f98a2SYan Zheng 29325d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 29335d4f98a2SYan Zheng 29345d4f98a2SYan Zheng path->lowest_level = node->level + 1; 29353fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 29365d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 2937581c1760SQu Wenruo struct btrfs_key first_key; 293882fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2939581c1760SQu Wenruo 29405d4f98a2SYan Zheng cond_resched(); 29415d4f98a2SYan Zheng 29425d4f98a2SYan Zheng upper = edge->node[UPPER]; 2943dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 29443fd0a558SYan, Zheng BUG_ON(!root); 29455d4f98a2SYan Zheng 29463fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 29473fd0a558SYan, Zheng if (!lowest) { 29483fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 29493fd0a558SYan, Zheng upper->level, &slot); 2950cbca7d59SFilipe Manana if (ret < 0) { 2951cbca7d59SFilipe Manana err = ret; 2952cbca7d59SFilipe Manana goto next; 2953cbca7d59SFilipe Manana } 29543fd0a558SYan, Zheng BUG_ON(ret); 29553fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 29563fd0a558SYan, Zheng if (node->eb->start == bytenr) 29573fd0a558SYan, Zheng goto next; 29583fd0a558SYan, Zheng } 29595d4f98a2SYan Zheng drop_node_buffer(upper); 29603fd0a558SYan, Zheng } 29615d4f98a2SYan Zheng 29625d4f98a2SYan Zheng if (!upper->eb) { 29635d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 29643561b9dbSLiu Bo if (ret) { 29653561b9dbSLiu Bo if (ret < 0) 29665d4f98a2SYan Zheng err = ret; 29673561b9dbSLiu Bo else 29683561b9dbSLiu Bo err = -ENOENT; 29693561b9dbSLiu Bo 29703561b9dbSLiu Bo btrfs_release_path(path); 29715d4f98a2SYan Zheng break; 29725d4f98a2SYan Zheng } 29735d4f98a2SYan Zheng 29743fd0a558SYan, Zheng if (!upper->eb) { 29753fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 29763fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 29773fd0a558SYan, Zheng } else { 29783fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 29793fd0a558SYan, Zheng } 29803fd0a558SYan, Zheng 29813fd0a558SYan, Zheng upper->locked = 1; 29823fd0a558SYan, Zheng path->locks[upper->level] = 0; 29833fd0a558SYan, Zheng 29845d4f98a2SYan Zheng slot = path->slots[upper->level]; 2985b3b4aa74SDavid Sterba btrfs_release_path(path); 29865d4f98a2SYan Zheng } else { 29875d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 29885d4f98a2SYan Zheng &slot); 2989cbca7d59SFilipe Manana if (ret < 0) { 2990cbca7d59SFilipe Manana err = ret; 2991cbca7d59SFilipe Manana goto next; 2992cbca7d59SFilipe Manana } 29935d4f98a2SYan Zheng BUG_ON(ret); 29945d4f98a2SYan Zheng } 29955d4f98a2SYan Zheng 29965d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 29973fd0a558SYan, Zheng if (lowest) { 29984547f4d8SLiu Bo if (bytenr != node->bytenr) { 29994547f4d8SLiu Bo btrfs_err(root->fs_info, 30004547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 30014547f4d8SLiu Bo bytenr, node->bytenr, slot, 30024547f4d8SLiu Bo upper->eb->start); 30034547f4d8SLiu Bo err = -EIO; 30044547f4d8SLiu Bo goto next; 30054547f4d8SLiu Bo } 30065d4f98a2SYan Zheng } else { 30073fd0a558SYan, Zheng if (node->eb->start == bytenr) 30083fd0a558SYan, Zheng goto next; 30095d4f98a2SYan Zheng } 30105d4f98a2SYan Zheng 3011da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 30125d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 3013581c1760SQu Wenruo btrfs_node_key_to_cpu(upper->eb, &first_key, slot); 3014581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, generation, 3015581c1760SQu Wenruo upper->level - 1, &first_key); 301664c043deSLiu Bo if (IS_ERR(eb)) { 301764c043deSLiu Bo err = PTR_ERR(eb); 301864c043deSLiu Bo goto next; 301964c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 3020416bc658SJosef Bacik free_extent_buffer(eb); 302197d9a8a4STsutomu Itoh err = -EIO; 302297d9a8a4STsutomu Itoh goto next; 302397d9a8a4STsutomu Itoh } 30245d4f98a2SYan Zheng btrfs_tree_lock(eb); 30258bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 30265d4f98a2SYan Zheng 30275d4f98a2SYan Zheng if (!node->eb) { 30285d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 30295d4f98a2SYan Zheng slot, &eb); 30303fd0a558SYan, Zheng btrfs_tree_unlock(eb); 30313fd0a558SYan, Zheng free_extent_buffer(eb); 30325d4f98a2SYan Zheng if (ret < 0) { 30335d4f98a2SYan Zheng err = ret; 30343fd0a558SYan, Zheng goto next; 30355d4f98a2SYan Zheng } 30363fd0a558SYan, Zheng BUG_ON(node->eb != eb); 30375d4f98a2SYan Zheng } else { 30385d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 30395d4f98a2SYan Zheng node->eb->start); 30405d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 30415d4f98a2SYan Zheng trans->transid); 30425d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 30435d4f98a2SYan Zheng 304482fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 30455d4f98a2SYan Zheng node->eb->start, blocksize, 304682fa113fSQu Wenruo upper->eb->start); 304782fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 304882fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 304982fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 305082fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 30515d4f98a2SYan Zheng BUG_ON(ret); 30525d4f98a2SYan Zheng 30535d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 30545d4f98a2SYan Zheng BUG_ON(ret); 30555d4f98a2SYan Zheng } 30563fd0a558SYan, Zheng next: 30573fd0a558SYan, Zheng if (!upper->pending) 30583fd0a558SYan, Zheng drop_node_buffer(upper); 30593fd0a558SYan, Zheng else 30603fd0a558SYan, Zheng unlock_node_buffer(upper); 30613fd0a558SYan, Zheng if (err) 30623fd0a558SYan, Zheng break; 30635d4f98a2SYan Zheng } 30643fd0a558SYan, Zheng 30653fd0a558SYan, Zheng if (!err && node->pending) { 30663fd0a558SYan, Zheng drop_node_buffer(node); 30673fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 30683fd0a558SYan, Zheng node->pending = 0; 30695d4f98a2SYan Zheng } 30703fd0a558SYan, Zheng 30715d4f98a2SYan Zheng path->lowest_level = 0; 30723fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 30735d4f98a2SYan Zheng return err; 30745d4f98a2SYan Zheng } 30755d4f98a2SYan Zheng 30765d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 30773fd0a558SYan, Zheng struct reloc_control *rc, 30785d4f98a2SYan Zheng struct backref_node *node, 30795d4f98a2SYan Zheng struct btrfs_path *path) 30805d4f98a2SYan Zheng { 30815d4f98a2SYan Zheng struct btrfs_key key; 30825d4f98a2SYan Zheng 30835d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 30843fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 30855d4f98a2SYan Zheng } 30865d4f98a2SYan Zheng 30875d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 30883fd0a558SYan, Zheng struct reloc_control *rc, 30893fd0a558SYan, Zheng struct btrfs_path *path, int err) 30905d4f98a2SYan Zheng { 30913fd0a558SYan, Zheng LIST_HEAD(list); 30923fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 30935d4f98a2SYan Zheng struct backref_node *node; 30945d4f98a2SYan Zheng int level; 30955d4f98a2SYan Zheng int ret; 30965d4f98a2SYan Zheng 30975d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 30985d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 30995d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 31003fd0a558SYan, Zheng struct backref_node, list); 31013fd0a558SYan, Zheng list_move_tail(&node->list, &list); 31023fd0a558SYan, Zheng BUG_ON(!node->pending); 31035d4f98a2SYan Zheng 31043fd0a558SYan, Zheng if (!err) { 31053fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 31065d4f98a2SYan Zheng if (ret < 0) 31075d4f98a2SYan Zheng err = ret; 31085d4f98a2SYan Zheng } 31095d4f98a2SYan Zheng } 31103fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 31113fd0a558SYan, Zheng } 31125d4f98a2SYan Zheng return err; 31135d4f98a2SYan Zheng } 31145d4f98a2SYan Zheng 31155d4f98a2SYan Zheng /* 31165d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 31175d4f98a2SYan Zheng * as processed. 31185d4f98a2SYan Zheng */ 31195d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 31205d4f98a2SYan Zheng struct backref_node *node) 31215d4f98a2SYan Zheng { 31225d4f98a2SYan Zheng struct backref_node *next = node; 31235d4f98a2SYan Zheng struct backref_edge *edge; 31245d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 31255d4f98a2SYan Zheng int index = 0; 31265d4f98a2SYan Zheng 31275d4f98a2SYan Zheng while (next) { 31285d4f98a2SYan Zheng cond_resched(); 31295d4f98a2SYan Zheng while (1) { 31305d4f98a2SYan Zheng if (next->processed) 31315d4f98a2SYan Zheng break; 31325d4f98a2SYan Zheng 31339569cc20SQu Wenruo mark_block_processed(rc, next); 31345d4f98a2SYan Zheng 31355d4f98a2SYan Zheng if (list_empty(&next->upper)) 31365d4f98a2SYan Zheng break; 31375d4f98a2SYan Zheng 31385d4f98a2SYan Zheng edge = list_entry(next->upper.next, 31395d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 31405d4f98a2SYan Zheng edges[index++] = edge; 31415d4f98a2SYan Zheng next = edge->node[UPPER]; 31425d4f98a2SYan Zheng } 31435d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 31445d4f98a2SYan Zheng } 31455d4f98a2SYan Zheng } 31465d4f98a2SYan Zheng 31477476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 31485d4f98a2SYan Zheng { 3149da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 31507476dfdaSDavid Sterba 31515d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 31529655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 31535d4f98a2SYan Zheng return 1; 31545d4f98a2SYan Zheng return 0; 31555d4f98a2SYan Zheng } 31565d4f98a2SYan Zheng 31572ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 31585d4f98a2SYan Zheng struct tree_block *block) 31595d4f98a2SYan Zheng { 31605d4f98a2SYan Zheng struct extent_buffer *eb; 31615d4f98a2SYan Zheng 3162581c1760SQu Wenruo eb = read_tree_block(fs_info, block->bytenr, block->key.offset, 3163581c1760SQu Wenruo block->level, NULL); 316464c043deSLiu Bo if (IS_ERR(eb)) { 316564c043deSLiu Bo return PTR_ERR(eb); 316664c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 3167416bc658SJosef Bacik free_extent_buffer(eb); 3168416bc658SJosef Bacik return -EIO; 3169416bc658SJosef Bacik } 31705d4f98a2SYan Zheng if (block->level == 0) 31715d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 31725d4f98a2SYan Zheng else 31735d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 31745d4f98a2SYan Zheng free_extent_buffer(eb); 31755d4f98a2SYan Zheng block->key_ready = 1; 31765d4f98a2SYan Zheng return 0; 31775d4f98a2SYan Zheng } 31785d4f98a2SYan Zheng 31795d4f98a2SYan Zheng /* 31805d4f98a2SYan Zheng * helper function to relocate a tree block 31815d4f98a2SYan Zheng */ 31825d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 31835d4f98a2SYan Zheng struct reloc_control *rc, 31845d4f98a2SYan Zheng struct backref_node *node, 31855d4f98a2SYan Zheng struct btrfs_key *key, 31865d4f98a2SYan Zheng struct btrfs_path *path) 31875d4f98a2SYan Zheng { 31885d4f98a2SYan Zheng struct btrfs_root *root; 31893fd0a558SYan, Zheng int ret = 0; 31905d4f98a2SYan Zheng 31913fd0a558SYan, Zheng if (!node) 31925d4f98a2SYan Zheng return 0; 31933fd0a558SYan, Zheng 31945f6b2e5cSJosef Bacik /* 31955f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 31965f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 31975f6b2e5cSJosef Bacik */ 31985f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 31995f6b2e5cSJosef Bacik if (ret) 32005f6b2e5cSJosef Bacik goto out; 32015f6b2e5cSJosef Bacik 32023fd0a558SYan, Zheng BUG_ON(node->processed); 3203147d256eSZhaolei root = select_one_root(node); 32043fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 32053fd0a558SYan, Zheng update_processed_blocks(rc, node); 32063fd0a558SYan, Zheng goto out; 32075d4f98a2SYan Zheng } 32085d4f98a2SYan Zheng 32093fd0a558SYan, Zheng if (root) { 321027cdeb70SMiao Xie if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) { 32113fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 32123fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 32133fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 32143fd0a558SYan, Zheng root = root->reloc_root; 32153fd0a558SYan, Zheng node->new_bytenr = root->node->start; 321600246528SJosef Bacik btrfs_put_root(node->root); 321700246528SJosef Bacik node->root = btrfs_grab_root(root); 32180b530bc5SJosef Bacik ASSERT(node->root); 32193fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 32203fd0a558SYan, Zheng } else { 32215d4f98a2SYan Zheng path->lowest_level = node->level; 32225d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 3223b3b4aa74SDavid Sterba btrfs_release_path(path); 32243fd0a558SYan, Zheng if (ret > 0) 32255d4f98a2SYan Zheng ret = 0; 32263fd0a558SYan, Zheng } 32273fd0a558SYan, Zheng if (!ret) 32283fd0a558SYan, Zheng update_processed_blocks(rc, node); 32293fd0a558SYan, Zheng } else { 32303fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 32313fd0a558SYan, Zheng } 32325d4f98a2SYan Zheng out: 32330647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 32343fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 32355d4f98a2SYan Zheng return ret; 32365d4f98a2SYan Zheng } 32375d4f98a2SYan Zheng 32385d4f98a2SYan Zheng /* 32395d4f98a2SYan Zheng * relocate a list of blocks 32405d4f98a2SYan Zheng */ 32415d4f98a2SYan Zheng static noinline_for_stack 32425d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 32435d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 32445d4f98a2SYan Zheng { 32452ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 32465d4f98a2SYan Zheng struct backref_node *node; 32475d4f98a2SYan Zheng struct btrfs_path *path; 32485d4f98a2SYan Zheng struct tree_block *block; 324998ff7b94SQu Wenruo struct tree_block *next; 32505d4f98a2SYan Zheng int ret; 32515d4f98a2SYan Zheng int err = 0; 32525d4f98a2SYan Zheng 32535d4f98a2SYan Zheng path = btrfs_alloc_path(); 3254e1a12670SLiu Bo if (!path) { 3255e1a12670SLiu Bo err = -ENOMEM; 325634c2b290SDavid Sterba goto out_free_blocks; 3257e1a12670SLiu Bo } 32585d4f98a2SYan Zheng 325998ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 326098ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 32615d4f98a2SYan Zheng if (!block->key_ready) 32622ff7e61eSJeff Mahoney readahead_tree_block(fs_info, block->bytenr); 32635d4f98a2SYan Zheng } 32645d4f98a2SYan Zheng 326598ff7b94SQu Wenruo /* Get first keys */ 326698ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 326734c2b290SDavid Sterba if (!block->key_ready) { 32682ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 326934c2b290SDavid Sterba if (err) 327034c2b290SDavid Sterba goto out_free_path; 327134c2b290SDavid Sterba } 32725d4f98a2SYan Zheng } 32735d4f98a2SYan Zheng 327498ff7b94SQu Wenruo /* Do tree relocation */ 327598ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 32763fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 32775d4f98a2SYan Zheng block->level, block->bytenr); 32785d4f98a2SYan Zheng if (IS_ERR(node)) { 32795d4f98a2SYan Zheng err = PTR_ERR(node); 32805d4f98a2SYan Zheng goto out; 32815d4f98a2SYan Zheng } 32825d4f98a2SYan Zheng 32835d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 32845d4f98a2SYan Zheng path); 32855d4f98a2SYan Zheng if (ret < 0) { 32865d4f98a2SYan Zheng err = ret; 328750dbbb71SJosef Bacik break; 32885d4f98a2SYan Zheng } 32895d4f98a2SYan Zheng } 32905d4f98a2SYan Zheng out: 32913fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 32925d4f98a2SYan Zheng 329334c2b290SDavid Sterba out_free_path: 32945d4f98a2SYan Zheng btrfs_free_path(path); 329534c2b290SDavid Sterba out_free_blocks: 3296e1a12670SLiu Bo free_block_list(blocks); 32975d4f98a2SYan Zheng return err; 32985d4f98a2SYan Zheng } 32995d4f98a2SYan Zheng 33005d4f98a2SYan Zheng static noinline_for_stack 3301efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 3302efa56464SYan, Zheng struct file_extent_cluster *cluster) 3303efa56464SYan, Zheng { 3304efa56464SYan, Zheng u64 alloc_hint = 0; 3305efa56464SYan, Zheng u64 start; 3306efa56464SYan, Zheng u64 end; 3307efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 3308efa56464SYan, Zheng u64 num_bytes; 3309efa56464SYan, Zheng int nr = 0; 3310efa56464SYan, Zheng int ret = 0; 3311dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 3312dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 331318513091SWang Xiaoguang u64 cur_offset; 3314364ecf36SQu Wenruo struct extent_changeset *data_reserved = NULL; 3315efa56464SYan, Zheng 3316efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 33175955102cSAl Viro inode_lock(inode); 3318efa56464SYan, Zheng 3319364ecf36SQu Wenruo ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start, 3320dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 3321efa56464SYan, Zheng if (ret) 3322efa56464SYan, Zheng goto out; 3323efa56464SYan, Zheng 332418513091SWang Xiaoguang cur_offset = prealloc_start; 3325efa56464SYan, Zheng while (nr < cluster->nr) { 3326efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 3327efa56464SYan, Zheng if (nr + 1 < cluster->nr) 3328efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 3329efa56464SYan, Zheng else 3330efa56464SYan, Zheng end = cluster->end - offset; 3331efa56464SYan, Zheng 3332d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 3333efa56464SYan, Zheng num_bytes = end + 1 - start; 333418513091SWang Xiaoguang if (cur_offset < start) 3335bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3336bc42bda2SQu Wenruo cur_offset, start - cur_offset); 3337efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 3338efa56464SYan, Zheng num_bytes, num_bytes, 3339efa56464SYan, Zheng end + 1, &alloc_hint); 334018513091SWang Xiaoguang cur_offset = end + 1; 3341d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 3342efa56464SYan, Zheng if (ret) 3343efa56464SYan, Zheng break; 3344efa56464SYan, Zheng nr++; 3345efa56464SYan, Zheng } 334618513091SWang Xiaoguang if (cur_offset < prealloc_end) 3347bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3348bc42bda2SQu Wenruo cur_offset, prealloc_end + 1 - cur_offset); 3349efa56464SYan, Zheng out: 33505955102cSAl Viro inode_unlock(inode); 3351364ecf36SQu Wenruo extent_changeset_free(data_reserved); 3352efa56464SYan, Zheng return ret; 3353efa56464SYan, Zheng } 3354efa56464SYan, Zheng 3355efa56464SYan, Zheng static noinline_for_stack 33560257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 33570257bb82SYan, Zheng u64 block_start) 33585d4f98a2SYan Zheng { 33595d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 33605d4f98a2SYan Zheng struct extent_map *em; 33610257bb82SYan, Zheng int ret = 0; 33625d4f98a2SYan Zheng 3363172ddd60SDavid Sterba em = alloc_extent_map(); 33640257bb82SYan, Zheng if (!em) 33650257bb82SYan, Zheng return -ENOMEM; 33660257bb82SYan, Zheng 33675d4f98a2SYan Zheng em->start = start; 33680257bb82SYan, Zheng em->len = end + 1 - start; 33690257bb82SYan, Zheng em->block_len = em->len; 33700257bb82SYan, Zheng em->block_start = block_start; 33715d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 33725d4f98a2SYan Zheng 3373d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 33745d4f98a2SYan Zheng while (1) { 3375890871beSChris Mason write_lock(&em_tree->lock); 337609a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 3377890871beSChris Mason write_unlock(&em_tree->lock); 33785d4f98a2SYan Zheng if (ret != -EEXIST) { 33795d4f98a2SYan Zheng free_extent_map(em); 33805d4f98a2SYan Zheng break; 33815d4f98a2SYan Zheng } 3382dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 33835d4f98a2SYan Zheng } 3384d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 33850257bb82SYan, Zheng return ret; 33860257bb82SYan, Zheng } 33875d4f98a2SYan Zheng 3388726a3421SQu Wenruo /* 3389726a3421SQu Wenruo * Allow error injection to test balance cancellation 3390726a3421SQu Wenruo */ 3391726a3421SQu Wenruo int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 3392726a3421SQu Wenruo { 3393726a3421SQu Wenruo return atomic_read(&fs_info->balance_cancel_req); 3394726a3421SQu Wenruo } 3395726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 3396726a3421SQu Wenruo 33970257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 33980257bb82SYan, Zheng struct file_extent_cluster *cluster) 33990257bb82SYan, Zheng { 34002ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 34010257bb82SYan, Zheng u64 page_start; 34020257bb82SYan, Zheng u64 page_end; 34030257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 34040257bb82SYan, Zheng unsigned long index; 34050257bb82SYan, Zheng unsigned long last_index; 34060257bb82SYan, Zheng struct page *page; 34070257bb82SYan, Zheng struct file_ra_state *ra; 34083b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 34090257bb82SYan, Zheng int nr = 0; 34100257bb82SYan, Zheng int ret = 0; 34110257bb82SYan, Zheng 34120257bb82SYan, Zheng if (!cluster->nr) 34130257bb82SYan, Zheng return 0; 34140257bb82SYan, Zheng 34150257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 34160257bb82SYan, Zheng if (!ra) 34170257bb82SYan, Zheng return -ENOMEM; 34180257bb82SYan, Zheng 3419efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 34200257bb82SYan, Zheng if (ret) 3421efa56464SYan, Zheng goto out; 34220257bb82SYan, Zheng 34230257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 34240257bb82SYan, Zheng 3425efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 3426efa56464SYan, Zheng cluster->end - offset, cluster->start); 3427efa56464SYan, Zheng if (ret) 3428efa56464SYan, Zheng goto out; 3429efa56464SYan, Zheng 343009cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 343109cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 34320257bb82SYan, Zheng while (index <= last_index) { 34339f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 34349f3db423SNikolay Borisov PAGE_SIZE); 3435efa56464SYan, Zheng if (ret) 3436efa56464SYan, Zheng goto out; 3437efa56464SYan, Zheng 34380257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 34390257bb82SYan, Zheng if (!page) { 34400257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 34410257bb82SYan, Zheng ra, NULL, index, 34420257bb82SYan, Zheng last_index + 1 - index); 3443a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 34443b16a4e3SJosef Bacik mask); 34450257bb82SYan, Zheng if (!page) { 3446691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 344743b18595SQu Wenruo PAGE_SIZE, true); 344844db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 34498702ba93SQu Wenruo PAGE_SIZE); 34500257bb82SYan, Zheng ret = -ENOMEM; 3451efa56464SYan, Zheng goto out; 34520257bb82SYan, Zheng } 34530257bb82SYan, Zheng } 34540257bb82SYan, Zheng 34550257bb82SYan, Zheng if (PageReadahead(page)) { 34560257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 34570257bb82SYan, Zheng ra, NULL, page, index, 34580257bb82SYan, Zheng last_index + 1 - index); 34590257bb82SYan, Zheng } 34600257bb82SYan, Zheng 34610257bb82SYan, Zheng if (!PageUptodate(page)) { 34620257bb82SYan, Zheng btrfs_readpage(NULL, page); 34630257bb82SYan, Zheng lock_page(page); 34640257bb82SYan, Zheng if (!PageUptodate(page)) { 34650257bb82SYan, Zheng unlock_page(page); 346609cbfeafSKirill A. Shutemov put_page(page); 3467691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 346843b18595SQu Wenruo PAGE_SIZE, true); 34698b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 34708702ba93SQu Wenruo PAGE_SIZE); 34710257bb82SYan, Zheng ret = -EIO; 3472efa56464SYan, Zheng goto out; 34730257bb82SYan, Zheng } 34740257bb82SYan, Zheng } 34750257bb82SYan, Zheng 34764eee4fa4SMiao Xie page_start = page_offset(page); 347709cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 34780257bb82SYan, Zheng 3479d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 34800257bb82SYan, Zheng 34810257bb82SYan, Zheng set_page_extent_mapped(page); 34820257bb82SYan, Zheng 34830257bb82SYan, Zheng if (nr < cluster->nr && 34840257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 34850257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 34860257bb82SYan, Zheng page_start, page_end, 3487ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 34880257bb82SYan, Zheng nr++; 34890257bb82SYan, Zheng } 34900257bb82SYan, Zheng 3491765f3cebSNikolay Borisov ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0, 3492330a5827SNikolay Borisov NULL); 3493765f3cebSNikolay Borisov if (ret) { 3494765f3cebSNikolay Borisov unlock_page(page); 3495765f3cebSNikolay Borisov put_page(page); 3496765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 349743b18595SQu Wenruo PAGE_SIZE, true); 3498765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 34998702ba93SQu Wenruo PAGE_SIZE); 3500765f3cebSNikolay Borisov 3501765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 3502765f3cebSNikolay Borisov page_start, page_end, 3503765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 3504765f3cebSNikolay Borisov goto out; 3505765f3cebSNikolay Borisov 3506765f3cebSNikolay Borisov } 35070257bb82SYan, Zheng set_page_dirty(page); 35080257bb82SYan, Zheng 35090257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 3510d0082371SJeff Mahoney page_start, page_end); 35110257bb82SYan, Zheng unlock_page(page); 351209cbfeafSKirill A. Shutemov put_page(page); 35130257bb82SYan, Zheng 35140257bb82SYan, Zheng index++; 35158702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 3516efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 35172ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 35187f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 35197f913c7cSQu Wenruo ret = -ECANCELED; 35207f913c7cSQu Wenruo goto out; 35217f913c7cSQu Wenruo } 35220257bb82SYan, Zheng } 35230257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 3524efa56464SYan, Zheng out: 35250257bb82SYan, Zheng kfree(ra); 35260257bb82SYan, Zheng return ret; 35270257bb82SYan, Zheng } 35280257bb82SYan, Zheng 35290257bb82SYan, Zheng static noinline_for_stack 35300257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 35310257bb82SYan, Zheng struct file_extent_cluster *cluster) 35320257bb82SYan, Zheng { 35330257bb82SYan, Zheng int ret; 35340257bb82SYan, Zheng 35350257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 35360257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 35370257bb82SYan, Zheng if (ret) 35380257bb82SYan, Zheng return ret; 35390257bb82SYan, Zheng cluster->nr = 0; 35400257bb82SYan, Zheng } 35410257bb82SYan, Zheng 35420257bb82SYan, Zheng if (!cluster->nr) 35430257bb82SYan, Zheng cluster->start = extent_key->objectid; 35440257bb82SYan, Zheng else 35450257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 35460257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 35470257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 35480257bb82SYan, Zheng cluster->nr++; 35490257bb82SYan, Zheng 35500257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 35510257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 35520257bb82SYan, Zheng if (ret) 35530257bb82SYan, Zheng return ret; 35540257bb82SYan, Zheng cluster->nr = 0; 35550257bb82SYan, Zheng } 35560257bb82SYan, Zheng return 0; 35575d4f98a2SYan Zheng } 35585d4f98a2SYan Zheng 35595d4f98a2SYan Zheng /* 35605d4f98a2SYan Zheng * helper to add a tree block to the list. 35615d4f98a2SYan Zheng * the major work is getting the generation and level of the block 35625d4f98a2SYan Zheng */ 35635d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 35645d4f98a2SYan Zheng struct btrfs_key *extent_key, 35655d4f98a2SYan Zheng struct btrfs_path *path, 35665d4f98a2SYan Zheng struct rb_root *blocks) 35675d4f98a2SYan Zheng { 35685d4f98a2SYan Zheng struct extent_buffer *eb; 35695d4f98a2SYan Zheng struct btrfs_extent_item *ei; 35705d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 35715d4f98a2SYan Zheng struct tree_block *block; 35725d4f98a2SYan Zheng struct rb_node *rb_node; 35735d4f98a2SYan Zheng u32 item_size; 35745d4f98a2SYan Zheng int level = -1; 35757fdf4b60SWang Shilong u64 generation; 35765d4f98a2SYan Zheng 35775d4f98a2SYan Zheng eb = path->nodes[0]; 35785d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 35795d4f98a2SYan Zheng 35803173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 35813173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 35825d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 35835d4f98a2SYan Zheng struct btrfs_extent_item); 35843173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 35855d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 35865d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 35875d4f98a2SYan Zheng } else { 35883173a18fSJosef Bacik level = (int)extent_key->offset; 35893173a18fSJosef Bacik } 35903173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 35916d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3592ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3593ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3594ba3c2b19SNikolay Borisov return -EINVAL; 35953173a18fSJosef Bacik } else { 35965d4f98a2SYan Zheng BUG(); 35975d4f98a2SYan Zheng } 35985d4f98a2SYan Zheng 3599b3b4aa74SDavid Sterba btrfs_release_path(path); 36005d4f98a2SYan Zheng 36015d4f98a2SYan Zheng BUG_ON(level == -1); 36025d4f98a2SYan Zheng 36035d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 36045d4f98a2SYan Zheng if (!block) 36055d4f98a2SYan Zheng return -ENOMEM; 36065d4f98a2SYan Zheng 36075d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3608da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 36095d4f98a2SYan Zheng block->key.offset = generation; 36105d4f98a2SYan Zheng block->level = level; 36115d4f98a2SYan Zheng block->key_ready = 0; 36125d4f98a2SYan Zheng 36135d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 361443c04fb1SJeff Mahoney if (rb_node) 361543c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, block->bytenr); 36165d4f98a2SYan Zheng 36175d4f98a2SYan Zheng return 0; 36185d4f98a2SYan Zheng } 36195d4f98a2SYan Zheng 36205d4f98a2SYan Zheng /* 36215d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 36225d4f98a2SYan Zheng */ 36235d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 36245d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 36255d4f98a2SYan Zheng struct rb_root *blocks) 36265d4f98a2SYan Zheng { 36270b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36285d4f98a2SYan Zheng struct btrfs_path *path; 36295d4f98a2SYan Zheng struct btrfs_key key; 36305d4f98a2SYan Zheng int ret; 36310b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 36325d4f98a2SYan Zheng 36337476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 36345d4f98a2SYan Zheng return 0; 36355d4f98a2SYan Zheng 36365d4f98a2SYan Zheng if (tree_search(blocks, bytenr)) 36375d4f98a2SYan Zheng return 0; 36385d4f98a2SYan Zheng 36395d4f98a2SYan Zheng path = btrfs_alloc_path(); 36405d4f98a2SYan Zheng if (!path) 36415d4f98a2SYan Zheng return -ENOMEM; 3642aee68ee5SJosef Bacik again: 36435d4f98a2SYan Zheng key.objectid = bytenr; 3644aee68ee5SJosef Bacik if (skinny) { 3645aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3646aee68ee5SJosef Bacik key.offset = (u64)-1; 3647aee68ee5SJosef Bacik } else { 36485d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 36495d4f98a2SYan Zheng key.offset = blocksize; 3650aee68ee5SJosef Bacik } 36515d4f98a2SYan Zheng 36525d4f98a2SYan Zheng path->search_commit_root = 1; 36535d4f98a2SYan Zheng path->skip_locking = 1; 36545d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 36555d4f98a2SYan Zheng if (ret < 0) 36565d4f98a2SYan Zheng goto out; 36575d4f98a2SYan Zheng 3658aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3659aee68ee5SJosef Bacik if (path->slots[0]) { 3660aee68ee5SJosef Bacik path->slots[0]--; 3661aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3662aee68ee5SJosef Bacik path->slots[0]); 36633173a18fSJosef Bacik if (key.objectid == bytenr && 3664aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3665aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3666aee68ee5SJosef Bacik key.offset == blocksize))) 36673173a18fSJosef Bacik ret = 0; 36683173a18fSJosef Bacik } 3669aee68ee5SJosef Bacik 3670aee68ee5SJosef Bacik if (ret) { 3671aee68ee5SJosef Bacik skinny = false; 3672aee68ee5SJosef Bacik btrfs_release_path(path); 3673aee68ee5SJosef Bacik goto again; 3674aee68ee5SJosef Bacik } 3675aee68ee5SJosef Bacik } 3676cdccee99SLiu Bo if (ret) { 3677cdccee99SLiu Bo ASSERT(ret == 1); 3678cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3679cdccee99SLiu Bo btrfs_err(fs_info, 3680cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3681cdccee99SLiu Bo bytenr); 3682cdccee99SLiu Bo WARN_ON(1); 3683cdccee99SLiu Bo ret = -EINVAL; 3684cdccee99SLiu Bo goto out; 3685cdccee99SLiu Bo } 36863173a18fSJosef Bacik 36875d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 36885d4f98a2SYan Zheng out: 36895d4f98a2SYan Zheng btrfs_free_path(path); 36905d4f98a2SYan Zheng return ret; 36915d4f98a2SYan Zheng } 36925d4f98a2SYan Zheng 36930af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 369432da5386SDavid Sterba struct btrfs_block_group *block_group, 36951bbc621eSChris Mason struct inode *inode, 36961bbc621eSChris Mason u64 ino) 36970af3d00bSJosef Bacik { 36980af3d00bSJosef Bacik struct btrfs_key key; 36990af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 37000af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 37010af3d00bSJosef Bacik int ret = 0; 37020af3d00bSJosef Bacik 37030af3d00bSJosef Bacik if (inode) 37040af3d00bSJosef Bacik goto truncate; 37050af3d00bSJosef Bacik 37060af3d00bSJosef Bacik key.objectid = ino; 37070af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 37080af3d00bSJosef Bacik key.offset = 0; 37090af3d00bSJosef Bacik 37104c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 37112e19f1f9SAl Viro if (IS_ERR(inode)) 37120af3d00bSJosef Bacik return -ENOENT; 37130af3d00bSJosef Bacik 37140af3d00bSJosef Bacik truncate: 37152ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 37167b61cd92SMiao Xie &fs_info->global_block_rsv); 37177b61cd92SMiao Xie if (ret) 37187b61cd92SMiao Xie goto out; 37197b61cd92SMiao Xie 37207a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 37210af3d00bSJosef Bacik if (IS_ERR(trans)) { 37223612b495STsutomu Itoh ret = PTR_ERR(trans); 37230af3d00bSJosef Bacik goto out; 37240af3d00bSJosef Bacik } 37250af3d00bSJosef Bacik 372677ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 37270af3d00bSJosef Bacik 37283a45bb20SJeff Mahoney btrfs_end_transaction(trans); 37292ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 37300af3d00bSJosef Bacik out: 37310af3d00bSJosef Bacik iput(inode); 37320af3d00bSJosef Bacik return ret; 37330af3d00bSJosef Bacik } 37340af3d00bSJosef Bacik 37355d4f98a2SYan Zheng /* 373619b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 373719b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 37385d4f98a2SYan Zheng */ 373919b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 374019b546d7SQu Wenruo struct btrfs_block_group *block_group, 374119b546d7SQu Wenruo u64 data_bytenr) 37425d4f98a2SYan Zheng { 374319b546d7SQu Wenruo u64 space_cache_ino; 374419b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 37455d4f98a2SYan Zheng struct btrfs_key key; 374619b546d7SQu Wenruo bool found = false; 374719b546d7SQu Wenruo int i; 37485d4f98a2SYan Zheng int ret; 37495d4f98a2SYan Zheng 375019b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 375119b546d7SQu Wenruo return 0; 37525d4f98a2SYan Zheng 375319b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 375419b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 375519b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 375619b546d7SQu Wenruo continue; 375719b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 375819b546d7SQu Wenruo if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG && 375919b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 376019b546d7SQu Wenruo found = true; 376119b546d7SQu Wenruo space_cache_ino = key.objectid; 376219b546d7SQu Wenruo break; 376319b546d7SQu Wenruo } 376419b546d7SQu Wenruo } 376519b546d7SQu Wenruo if (!found) 376619b546d7SQu Wenruo return -ENOENT; 376719b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 376819b546d7SQu Wenruo space_cache_ino); 37690af3d00bSJosef Bacik return ret; 37705d4f98a2SYan Zheng } 37715d4f98a2SYan Zheng 37725d4f98a2SYan Zheng /* 37732c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 37745d4f98a2SYan Zheng */ 37755d4f98a2SYan Zheng static noinline_for_stack 37765d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 37775d4f98a2SYan Zheng struct btrfs_key *extent_key, 37785d4f98a2SYan Zheng struct btrfs_path *path, 37795d4f98a2SYan Zheng struct rb_root *blocks) 37805d4f98a2SYan Zheng { 378119b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 378219b546d7SQu Wenruo struct ulist *leaves = NULL; 378319b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 378419b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 378519b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3786647f63bdSFilipe David Borba Manana int ret = 0; 37875d4f98a2SYan Zheng 3788b3b4aa74SDavid Sterba btrfs_release_path(path); 378919b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 379019b546d7SQu Wenruo 0, &leaves, NULL, true); 379119b546d7SQu Wenruo if (ret < 0) 379219b546d7SQu Wenruo return ret; 379319b546d7SQu Wenruo 379419b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 379519b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 379619b546d7SQu Wenruo struct extent_buffer *eb; 379719b546d7SQu Wenruo 379819b546d7SQu Wenruo eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL); 379919b546d7SQu Wenruo if (IS_ERR(eb)) { 380019b546d7SQu Wenruo ret = PTR_ERR(eb); 380119b546d7SQu Wenruo break; 380219b546d7SQu Wenruo } 380319b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 380419b546d7SQu Wenruo extent_key->objectid); 380519b546d7SQu Wenruo free_extent_buffer(eb); 380619b546d7SQu Wenruo if (ret < 0) 380719b546d7SQu Wenruo break; 380819b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 380919b546d7SQu Wenruo if (ret < 0) 381019b546d7SQu Wenruo break; 381119b546d7SQu Wenruo } 381219b546d7SQu Wenruo if (ret < 0) 38135d4f98a2SYan Zheng free_block_list(blocks); 381419b546d7SQu Wenruo ulist_free(leaves); 381519b546d7SQu Wenruo return ret; 38165d4f98a2SYan Zheng } 38175d4f98a2SYan Zheng 38185d4f98a2SYan Zheng /* 38192c016dc2SLiu Bo * helper to find next unprocessed extent 38205d4f98a2SYan Zheng */ 38215d4f98a2SYan Zheng static noinline_for_stack 3822147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 38233fd0a558SYan, Zheng struct btrfs_key *extent_key) 38245d4f98a2SYan Zheng { 38250b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 38265d4f98a2SYan Zheng struct btrfs_key key; 38275d4f98a2SYan Zheng struct extent_buffer *leaf; 38285d4f98a2SYan Zheng u64 start, end, last; 38295d4f98a2SYan Zheng int ret; 38305d4f98a2SYan Zheng 3831b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 38325d4f98a2SYan Zheng while (1) { 38335d4f98a2SYan Zheng cond_resched(); 38345d4f98a2SYan Zheng if (rc->search_start >= last) { 38355d4f98a2SYan Zheng ret = 1; 38365d4f98a2SYan Zheng break; 38375d4f98a2SYan Zheng } 38385d4f98a2SYan Zheng 38395d4f98a2SYan Zheng key.objectid = rc->search_start; 38405d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 38415d4f98a2SYan Zheng key.offset = 0; 38425d4f98a2SYan Zheng 38435d4f98a2SYan Zheng path->search_commit_root = 1; 38445d4f98a2SYan Zheng path->skip_locking = 1; 38455d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 38465d4f98a2SYan Zheng 0, 0); 38475d4f98a2SYan Zheng if (ret < 0) 38485d4f98a2SYan Zheng break; 38495d4f98a2SYan Zheng next: 38505d4f98a2SYan Zheng leaf = path->nodes[0]; 38515d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 38525d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 38535d4f98a2SYan Zheng if (ret != 0) 38545d4f98a2SYan Zheng break; 38555d4f98a2SYan Zheng leaf = path->nodes[0]; 38565d4f98a2SYan Zheng } 38575d4f98a2SYan Zheng 38585d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 38595d4f98a2SYan Zheng if (key.objectid >= last) { 38605d4f98a2SYan Zheng ret = 1; 38615d4f98a2SYan Zheng break; 38625d4f98a2SYan Zheng } 38635d4f98a2SYan Zheng 38643173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 38653173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 38663173a18fSJosef Bacik path->slots[0]++; 38673173a18fSJosef Bacik goto next; 38683173a18fSJosef Bacik } 38693173a18fSJosef Bacik 38703173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 38715d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 38725d4f98a2SYan Zheng path->slots[0]++; 38735d4f98a2SYan Zheng goto next; 38745d4f98a2SYan Zheng } 38755d4f98a2SYan Zheng 38763173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 38770b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 38783173a18fSJosef Bacik rc->search_start) { 38793173a18fSJosef Bacik path->slots[0]++; 38803173a18fSJosef Bacik goto next; 38813173a18fSJosef Bacik } 38823173a18fSJosef Bacik 38835d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 38845d4f98a2SYan Zheng key.objectid, &start, &end, 3885e6138876SJosef Bacik EXTENT_DIRTY, NULL); 38865d4f98a2SYan Zheng 38875d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3888b3b4aa74SDavid Sterba btrfs_release_path(path); 38895d4f98a2SYan Zheng rc->search_start = end + 1; 38905d4f98a2SYan Zheng } else { 38913173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 38925d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 38933173a18fSJosef Bacik else 38943173a18fSJosef Bacik rc->search_start = key.objectid + 38950b246afaSJeff Mahoney fs_info->nodesize; 38963fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 38975d4f98a2SYan Zheng return 0; 38985d4f98a2SYan Zheng } 38995d4f98a2SYan Zheng } 3900b3b4aa74SDavid Sterba btrfs_release_path(path); 39015d4f98a2SYan Zheng return ret; 39025d4f98a2SYan Zheng } 39035d4f98a2SYan Zheng 39045d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 39055d4f98a2SYan Zheng { 39065d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39077585717fSChris Mason 39087585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 39095d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 39107585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 39115d4f98a2SYan Zheng } 39125d4f98a2SYan Zheng 39135d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 39145d4f98a2SYan Zheng { 39155d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39167585717fSChris Mason 39177585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 39185d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 39197585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 39205d4f98a2SYan Zheng } 39215d4f98a2SYan Zheng 39225d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 39235d4f98a2SYan Zheng { 39245d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 39255d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 39265d4f98a2SYan Zheng return 1; 39275d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 39285d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 39295d4f98a2SYan Zheng return 1; 39305d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 39315d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 39325d4f98a2SYan Zheng return 1; 39335d4f98a2SYan Zheng return 0; 39345d4f98a2SYan Zheng } 39355d4f98a2SYan Zheng 39363fd0a558SYan, Zheng static noinline_for_stack 39373fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 39383fd0a558SYan, Zheng { 39393fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3940ac2fabacSJosef Bacik int ret; 39413fd0a558SYan, Zheng 39422ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 394366d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 39443fd0a558SYan, Zheng if (!rc->block_rsv) 39453fd0a558SYan, Zheng return -ENOMEM; 39463fd0a558SYan, Zheng 39473fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3948b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 39493fd0a558SYan, Zheng rc->extents_found = 0; 39503fd0a558SYan, Zheng rc->nodes_relocated = 0; 39513fd0a558SYan, Zheng rc->merging_rsv_size = 0; 39520647bf56SWang Shilong rc->reserved_bytes = 0; 3953da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 39540647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3955ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3956ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3957ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3958ac2fabacSJosef Bacik if (ret) 3959ac2fabacSJosef Bacik return ret; 39603fd0a558SYan, Zheng 39613fd0a558SYan, Zheng rc->create_reloc_tree = 1; 39623fd0a558SYan, Zheng set_reloc_control(rc); 39633fd0a558SYan, Zheng 39647a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 396528818947SLiu Bo if (IS_ERR(trans)) { 396628818947SLiu Bo unset_reloc_control(rc); 396728818947SLiu Bo /* 396828818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 396928818947SLiu Bo * cleanup. And callers are responsible to free the above 397028818947SLiu Bo * block rsv. 397128818947SLiu Bo */ 397228818947SLiu Bo return PTR_ERR(trans); 397328818947SLiu Bo } 39743a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 39753fd0a558SYan, Zheng return 0; 39763fd0a558SYan, Zheng } 397776dda93cSYan, Zheng 39785d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 39795d4f98a2SYan Zheng { 39802ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39815d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 39825d4f98a2SYan Zheng struct btrfs_key key; 39835d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 39845d4f98a2SYan Zheng struct btrfs_path *path; 39855d4f98a2SYan Zheng struct btrfs_extent_item *ei; 39865d4f98a2SYan Zheng u64 flags; 39875d4f98a2SYan Zheng u32 item_size; 39885d4f98a2SYan Zheng int ret; 39895d4f98a2SYan Zheng int err = 0; 3990c87f08caSChris Mason int progress = 0; 39915d4f98a2SYan Zheng 39925d4f98a2SYan Zheng path = btrfs_alloc_path(); 39933fd0a558SYan, Zheng if (!path) 39945d4f98a2SYan Zheng return -ENOMEM; 3995e4058b54SDavid Sterba path->reada = READA_FORWARD; 39963fd0a558SYan, Zheng 39973fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 39983fd0a558SYan, Zheng if (ret) { 39993fd0a558SYan, Zheng err = ret; 40003fd0a558SYan, Zheng goto out_free; 40012423fdfbSJiri Slaby } 40025d4f98a2SYan Zheng 40035d4f98a2SYan Zheng while (1) { 40040647bf56SWang Shilong rc->reserved_bytes = 0; 40050647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 40060647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 40070647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 40080647bf56SWang Shilong if (ret) { 40090647bf56SWang Shilong err = ret; 40100647bf56SWang Shilong break; 40110647bf56SWang Shilong } 4012c87f08caSChris Mason progress++; 4013a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 40140f788c58SLiu Bo if (IS_ERR(trans)) { 40150f788c58SLiu Bo err = PTR_ERR(trans); 40160f788c58SLiu Bo trans = NULL; 40170f788c58SLiu Bo break; 40180f788c58SLiu Bo } 4019c87f08caSChris Mason restart: 40203fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 40213a45bb20SJeff Mahoney btrfs_end_transaction(trans); 402242a657f5SPan Bian trans = NULL; 40233fd0a558SYan, Zheng continue; 40243fd0a558SYan, Zheng } 40253fd0a558SYan, Zheng 4026147d256eSZhaolei ret = find_next_extent(rc, path, &key); 40275d4f98a2SYan Zheng if (ret < 0) 40285d4f98a2SYan Zheng err = ret; 40295d4f98a2SYan Zheng if (ret != 0) 40305d4f98a2SYan Zheng break; 40315d4f98a2SYan Zheng 40325d4f98a2SYan Zheng rc->extents_found++; 40335d4f98a2SYan Zheng 40345d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 40355d4f98a2SYan Zheng struct btrfs_extent_item); 40363fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 40375d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 40385d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 40395d4f98a2SYan Zheng ret = check_extent_flags(flags); 40405d4f98a2SYan Zheng BUG_ON(ret); 40416d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 4042ba3c2b19SNikolay Borisov err = -EINVAL; 4043ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 4044ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 4045ba3c2b19SNikolay Borisov break; 40465d4f98a2SYan Zheng } else { 40475d4f98a2SYan Zheng BUG(); 40485d4f98a2SYan Zheng } 40495d4f98a2SYan Zheng 40505d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 40515d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 40525d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 40535d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 40545d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 40555d4f98a2SYan Zheng } else { 4056b3b4aa74SDavid Sterba btrfs_release_path(path); 40575d4f98a2SYan Zheng ret = 0; 40585d4f98a2SYan Zheng } 40595d4f98a2SYan Zheng if (ret < 0) { 40603fd0a558SYan, Zheng err = ret; 40615d4f98a2SYan Zheng break; 40625d4f98a2SYan Zheng } 40635d4f98a2SYan Zheng 40645d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 40655d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 40665d4f98a2SYan Zheng if (ret < 0) { 40673fd0a558SYan, Zheng if (ret != -EAGAIN) { 40685d4f98a2SYan Zheng err = ret; 40695d4f98a2SYan Zheng break; 40705d4f98a2SYan Zheng } 40713fd0a558SYan, Zheng rc->extents_found--; 40723fd0a558SYan, Zheng rc->search_start = key.objectid; 40733fd0a558SYan, Zheng } 40745d4f98a2SYan Zheng } 40755d4f98a2SYan Zheng 40763a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 40772ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 40783fd0a558SYan, Zheng trans = NULL; 40795d4f98a2SYan Zheng 40805d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 40815d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 40825d4f98a2SYan Zheng rc->found_file_extent = 1; 40830257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 40843fd0a558SYan, Zheng &key, &rc->cluster); 40855d4f98a2SYan Zheng if (ret < 0) { 40865d4f98a2SYan Zheng err = ret; 40875d4f98a2SYan Zheng break; 40885d4f98a2SYan Zheng } 40895d4f98a2SYan Zheng } 4090f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 4091f31ea088SQu Wenruo err = -ECANCELED; 4092f31ea088SQu Wenruo break; 4093f31ea088SQu Wenruo } 40945d4f98a2SYan Zheng } 4095c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 409643a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 40979689457bSShilong Wang if (ret == 1) { 4098c87f08caSChris Mason err = 0; 4099c87f08caSChris Mason progress = 0; 4100c87f08caSChris Mason goto restart; 4101c87f08caSChris Mason } 4102c87f08caSChris Mason } 41033fd0a558SYan, Zheng 4104b3b4aa74SDavid Sterba btrfs_release_path(path); 410591166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 41065d4f98a2SYan Zheng 41075d4f98a2SYan Zheng if (trans) { 41083a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 41092ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 41105d4f98a2SYan Zheng } 41115d4f98a2SYan Zheng 41120257bb82SYan, Zheng if (!err) { 41133fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 41143fd0a558SYan, Zheng &rc->cluster); 41150257bb82SYan, Zheng if (ret < 0) 41160257bb82SYan, Zheng err = ret; 41170257bb82SYan, Zheng } 41180257bb82SYan, Zheng 41193fd0a558SYan, Zheng rc->create_reloc_tree = 0; 41203fd0a558SYan, Zheng set_reloc_control(rc); 41210257bb82SYan, Zheng 41223fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 412363f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 41245d4f98a2SYan Zheng 41257f913c7cSQu Wenruo /* 41267f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 41277f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 41287f913c7cSQu Wenruo * 41297f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 41307f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 41317f913c7cSQu Wenruo * merge_reloc_roots() 41327f913c7cSQu Wenruo */ 41333fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 41345d4f98a2SYan Zheng 41355d4f98a2SYan Zheng merge_reloc_roots(rc); 41365d4f98a2SYan Zheng 41373fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 41385d4f98a2SYan Zheng unset_reloc_control(rc); 413963f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 41405d4f98a2SYan Zheng 41415d4f98a2SYan Zheng /* get rid of pinned extents */ 41427a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 414362b99540SQu Wenruo if (IS_ERR(trans)) { 41443612b495STsutomu Itoh err = PTR_ERR(trans); 414562b99540SQu Wenruo goto out_free; 414662b99540SQu Wenruo } 41473a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 41486217b0faSJosef Bacik out_free: 4149d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4150d2311e69SQu Wenruo if (ret < 0 && !err) 4151d2311e69SQu Wenruo err = ret; 41522ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 41533fd0a558SYan, Zheng btrfs_free_path(path); 41545d4f98a2SYan Zheng return err; 41555d4f98a2SYan Zheng } 41565d4f98a2SYan Zheng 41575d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 41580257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 41595d4f98a2SYan Zheng { 41605d4f98a2SYan Zheng struct btrfs_path *path; 41615d4f98a2SYan Zheng struct btrfs_inode_item *item; 41625d4f98a2SYan Zheng struct extent_buffer *leaf; 41635d4f98a2SYan Zheng int ret; 41645d4f98a2SYan Zheng 41655d4f98a2SYan Zheng path = btrfs_alloc_path(); 41665d4f98a2SYan Zheng if (!path) 41675d4f98a2SYan Zheng return -ENOMEM; 41685d4f98a2SYan Zheng 41695d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 41705d4f98a2SYan Zheng if (ret) 41715d4f98a2SYan Zheng goto out; 41725d4f98a2SYan Zheng 41735d4f98a2SYan Zheng leaf = path->nodes[0]; 41745d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 4175b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 41765d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 41770257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 41785d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 41793fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 41803fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 41815d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 41825d4f98a2SYan Zheng out: 41835d4f98a2SYan Zheng btrfs_free_path(path); 41845d4f98a2SYan Zheng return ret; 41855d4f98a2SYan Zheng } 41865d4f98a2SYan Zheng 41875d4f98a2SYan Zheng /* 41885d4f98a2SYan Zheng * helper to create inode for data relocation. 41895d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 41905d4f98a2SYan Zheng */ 41913fd0a558SYan, Zheng static noinline_for_stack 41923fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 419332da5386SDavid Sterba struct btrfs_block_group *group) 41945d4f98a2SYan Zheng { 41955d4f98a2SYan Zheng struct inode *inode = NULL; 41965d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 41975d4f98a2SYan Zheng struct btrfs_root *root; 41985d4f98a2SYan Zheng struct btrfs_key key; 41994624900dSZhaolei u64 objectid; 42005d4f98a2SYan Zheng int err = 0; 42015d4f98a2SYan Zheng 42025d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 42035d4f98a2SYan Zheng if (IS_ERR(root)) 42045d4f98a2SYan Zheng return ERR_CAST(root); 42055d4f98a2SYan Zheng 4206a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 420776deacf0SJosef Bacik if (IS_ERR(trans)) { 420800246528SJosef Bacik btrfs_put_root(root); 42093fd0a558SYan, Zheng return ERR_CAST(trans); 421076deacf0SJosef Bacik } 42115d4f98a2SYan Zheng 4212581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 42135d4f98a2SYan Zheng if (err) 42145d4f98a2SYan Zheng goto out; 42155d4f98a2SYan Zheng 42160257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 42175d4f98a2SYan Zheng BUG_ON(err); 42185d4f98a2SYan Zheng 42195d4f98a2SYan Zheng key.objectid = objectid; 42205d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 42215d4f98a2SYan Zheng key.offset = 0; 42224c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 42232e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 4224b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 42255d4f98a2SYan Zheng 422673f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 42275d4f98a2SYan Zheng out: 422800246528SJosef Bacik btrfs_put_root(root); 42293a45bb20SJeff Mahoney btrfs_end_transaction(trans); 42302ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 42315d4f98a2SYan Zheng if (err) { 42325d4f98a2SYan Zheng if (inode) 42335d4f98a2SYan Zheng iput(inode); 42345d4f98a2SYan Zheng inode = ERR_PTR(err); 42355d4f98a2SYan Zheng } 42365d4f98a2SYan Zheng return inode; 42375d4f98a2SYan Zheng } 42385d4f98a2SYan Zheng 4239c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 42403fd0a558SYan, Zheng { 42413fd0a558SYan, Zheng struct reloc_control *rc; 42423fd0a558SYan, Zheng 42433fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 42443fd0a558SYan, Zheng if (!rc) 42453fd0a558SYan, Zheng return NULL; 42463fd0a558SYan, Zheng 42473fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 4248d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 42492433bea5SQu Wenruo backref_cache_init(fs_info, &rc->backref_cache, 1); 42503fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 425143eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 425243eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 42533fd0a558SYan, Zheng return rc; 42543fd0a558SYan, Zheng } 42553fd0a558SYan, Zheng 42561a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 42571a0afa0eSJosef Bacik { 42581a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 42591a0afa0eSJosef Bacik 42601a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 42611a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 42621a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 42631a0afa0eSJosef Bacik kfree(node); 42641a0afa0eSJosef Bacik 42651a0afa0eSJosef Bacik kfree(rc); 42661a0afa0eSJosef Bacik } 42671a0afa0eSJosef Bacik 42685d4f98a2SYan Zheng /* 4269ebce0e01SAdam Borowski * Print the block group being relocated 4270ebce0e01SAdam Borowski */ 4271ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 427232da5386SDavid Sterba struct btrfs_block_group *block_group) 4273ebce0e01SAdam Borowski { 4274f89e09cfSAnand Jain char buf[128] = {'\0'}; 4275ebce0e01SAdam Borowski 4276f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 4277ebce0e01SAdam Borowski 4278ebce0e01SAdam Borowski btrfs_info(fs_info, 4279ebce0e01SAdam Borowski "relocating block group %llu flags %s", 4280b3470b5dSDavid Sterba block_group->start, buf); 4281ebce0e01SAdam Borowski } 4282ebce0e01SAdam Borowski 4283430640e3SQu Wenruo static const char *stage_to_string(int stage) 4284430640e3SQu Wenruo { 4285430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 4286430640e3SQu Wenruo return "move data extents"; 4287430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 4288430640e3SQu Wenruo return "update data pointers"; 4289430640e3SQu Wenruo return "unknown"; 4290430640e3SQu Wenruo } 4291430640e3SQu Wenruo 4292ebce0e01SAdam Borowski /* 42935d4f98a2SYan Zheng * function to relocate all extents in a block group. 42945d4f98a2SYan Zheng */ 42956bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 42965d4f98a2SYan Zheng { 429732da5386SDavid Sterba struct btrfs_block_group *bg; 42986bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 42995d4f98a2SYan Zheng struct reloc_control *rc; 43000af3d00bSJosef Bacik struct inode *inode; 43010af3d00bSJosef Bacik struct btrfs_path *path; 43025d4f98a2SYan Zheng int ret; 4303f0486c68SYan, Zheng int rw = 0; 43045d4f98a2SYan Zheng int err = 0; 43055d4f98a2SYan Zheng 4306eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 4307eede2bf3SOmar Sandoval if (!bg) 4308eede2bf3SOmar Sandoval return -ENOENT; 4309eede2bf3SOmar Sandoval 4310eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 4311eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 4312eede2bf3SOmar Sandoval return -ETXTBSY; 4313eede2bf3SOmar Sandoval } 4314eede2bf3SOmar Sandoval 4315c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 4316eede2bf3SOmar Sandoval if (!rc) { 4317eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 43185d4f98a2SYan Zheng return -ENOMEM; 4319eede2bf3SOmar Sandoval } 43205d4f98a2SYan Zheng 4321f0486c68SYan, Zheng rc->extent_root = extent_root; 4322eede2bf3SOmar Sandoval rc->block_group = bg; 43235d4f98a2SYan Zheng 4324b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 4325f0486c68SYan, Zheng if (ret) { 4326f0486c68SYan, Zheng err = ret; 4327f0486c68SYan, Zheng goto out; 4328f0486c68SYan, Zheng } 4329f0486c68SYan, Zheng rw = 1; 4330f0486c68SYan, Zheng 43310af3d00bSJosef Bacik path = btrfs_alloc_path(); 43320af3d00bSJosef Bacik if (!path) { 43330af3d00bSJosef Bacik err = -ENOMEM; 43340af3d00bSJosef Bacik goto out; 43350af3d00bSJosef Bacik } 43360af3d00bSJosef Bacik 43377949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 43380af3d00bSJosef Bacik btrfs_free_path(path); 43390af3d00bSJosef Bacik 43400af3d00bSJosef Bacik if (!IS_ERR(inode)) 43411bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 43420af3d00bSJosef Bacik else 43430af3d00bSJosef Bacik ret = PTR_ERR(inode); 43440af3d00bSJosef Bacik 43450af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 43460af3d00bSJosef Bacik err = ret; 43470af3d00bSJosef Bacik goto out; 43480af3d00bSJosef Bacik } 43490af3d00bSJosef Bacik 43505d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 43515d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 43525d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 43535d4f98a2SYan Zheng rc->data_inode = NULL; 43545d4f98a2SYan Zheng goto out; 43555d4f98a2SYan Zheng } 43565d4f98a2SYan Zheng 43570b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 43585d4f98a2SYan Zheng 43599cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 4360f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 43616374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 4362b3470b5dSDavid Sterba rc->block_group->start, 4363b3470b5dSDavid Sterba rc->block_group->length); 43645d4f98a2SYan Zheng 43655d4f98a2SYan Zheng while (1) { 4366430640e3SQu Wenruo int finishes_stage; 4367430640e3SQu Wenruo 436876dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 43695d4f98a2SYan Zheng ret = relocate_block_group(rc); 437076dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 4371ff612ba7SJosef Bacik if (ret < 0) 43725d4f98a2SYan Zheng err = ret; 4373ff612ba7SJosef Bacik 4374430640e3SQu Wenruo finishes_stage = rc->stage; 4375ff612ba7SJosef Bacik /* 4376ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 4377ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 4378ff612ba7SJosef Bacik * different block group we could end up hitting the 4379ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 4380ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 4381ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 4382ff612ba7SJosef Bacik * out of the loop if we hit an error. 4383ff612ba7SJosef Bacik */ 4384ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 4385ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 4386ff612ba7SJosef Bacik (u64)-1); 4387ff612ba7SJosef Bacik if (ret) 4388ff612ba7SJosef Bacik err = ret; 4389ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 4390ff612ba7SJosef Bacik 0, -1); 4391ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 43925d4f98a2SYan Zheng } 43935d4f98a2SYan Zheng 4394ff612ba7SJosef Bacik if (err < 0) 4395ff612ba7SJosef Bacik goto out; 4396ff612ba7SJosef Bacik 43975d4f98a2SYan Zheng if (rc->extents_found == 0) 43985d4f98a2SYan Zheng break; 43995d4f98a2SYan Zheng 4400430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 4401430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 44025d4f98a2SYan Zheng } 44035d4f98a2SYan Zheng 44045d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 44055d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 4406bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 44075d4f98a2SYan Zheng out: 4408f0486c68SYan, Zheng if (err && rw) 44092ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 44105d4f98a2SYan Zheng iput(rc->data_inode); 44115d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 44121a0afa0eSJosef Bacik free_reloc_control(rc); 44135d4f98a2SYan Zheng return err; 44145d4f98a2SYan Zheng } 44155d4f98a2SYan Zheng 441676dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 441776dda93cSYan, Zheng { 44180b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 441976dda93cSYan, Zheng struct btrfs_trans_handle *trans; 442079787eaaSJeff Mahoney int ret, err; 442176dda93cSYan, Zheng 44220b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 442379787eaaSJeff Mahoney if (IS_ERR(trans)) 442479787eaaSJeff Mahoney return PTR_ERR(trans); 442576dda93cSYan, Zheng 442676dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 442776dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 442876dda93cSYan, Zheng root->root_item.drop_level = 0; 442976dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 44300b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 443176dda93cSYan, Zheng &root->root_key, &root->root_item); 443276dda93cSYan, Zheng 44333a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 443479787eaaSJeff Mahoney if (err) 443579787eaaSJeff Mahoney return err; 443679787eaaSJeff Mahoney return ret; 443776dda93cSYan, Zheng } 443876dda93cSYan, Zheng 44395d4f98a2SYan Zheng /* 44405d4f98a2SYan Zheng * recover relocation interrupted by system crash. 44415d4f98a2SYan Zheng * 44425d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 44435d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 44445d4f98a2SYan Zheng */ 44455d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 44465d4f98a2SYan Zheng { 44470b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 44485d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 44495d4f98a2SYan Zheng struct btrfs_key key; 44505d4f98a2SYan Zheng struct btrfs_root *fs_root; 44515d4f98a2SYan Zheng struct btrfs_root *reloc_root; 44525d4f98a2SYan Zheng struct btrfs_path *path; 44535d4f98a2SYan Zheng struct extent_buffer *leaf; 44545d4f98a2SYan Zheng struct reloc_control *rc = NULL; 44555d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 44565d4f98a2SYan Zheng int ret; 44575d4f98a2SYan Zheng int err = 0; 44585d4f98a2SYan Zheng 44595d4f98a2SYan Zheng path = btrfs_alloc_path(); 44605d4f98a2SYan Zheng if (!path) 44615d4f98a2SYan Zheng return -ENOMEM; 4462e4058b54SDavid Sterba path->reada = READA_BACK; 44635d4f98a2SYan Zheng 44645d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 44655d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 44665d4f98a2SYan Zheng key.offset = (u64)-1; 44675d4f98a2SYan Zheng 44685d4f98a2SYan Zheng while (1) { 44690b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 44705d4f98a2SYan Zheng path, 0, 0); 44715d4f98a2SYan Zheng if (ret < 0) { 44725d4f98a2SYan Zheng err = ret; 44735d4f98a2SYan Zheng goto out; 44745d4f98a2SYan Zheng } 44755d4f98a2SYan Zheng if (ret > 0) { 44765d4f98a2SYan Zheng if (path->slots[0] == 0) 44775d4f98a2SYan Zheng break; 44785d4f98a2SYan Zheng path->slots[0]--; 44795d4f98a2SYan Zheng } 44805d4f98a2SYan Zheng leaf = path->nodes[0]; 44815d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4482b3b4aa74SDavid Sterba btrfs_release_path(path); 44835d4f98a2SYan Zheng 44845d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 44855d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 44865d4f98a2SYan Zheng break; 44875d4f98a2SYan Zheng 44883dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 44895d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 44905d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 44915d4f98a2SYan Zheng goto out; 44925d4f98a2SYan Zheng } 44935d4f98a2SYan Zheng 44943dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 44955d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 44965d4f98a2SYan Zheng 44975d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 44980b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, 44995d4f98a2SYan Zheng reloc_root->root_key.offset); 45005d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 450176dda93cSYan, Zheng ret = PTR_ERR(fs_root); 450276dda93cSYan, Zheng if (ret != -ENOENT) { 450376dda93cSYan, Zheng err = ret; 45045d4f98a2SYan Zheng goto out; 45055d4f98a2SYan Zheng } 450679787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 450779787eaaSJeff Mahoney if (ret < 0) { 450879787eaaSJeff Mahoney err = ret; 450979787eaaSJeff Mahoney goto out; 451079787eaaSJeff Mahoney } 4511932fd26dSJosef Bacik } else { 451200246528SJosef Bacik btrfs_put_root(fs_root); 451376dda93cSYan, Zheng } 45145d4f98a2SYan Zheng } 45155d4f98a2SYan Zheng 45165d4f98a2SYan Zheng if (key.offset == 0) 45175d4f98a2SYan Zheng break; 45185d4f98a2SYan Zheng 45195d4f98a2SYan Zheng key.offset--; 45205d4f98a2SYan Zheng } 4521b3b4aa74SDavid Sterba btrfs_release_path(path); 45225d4f98a2SYan Zheng 45235d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 45245d4f98a2SYan Zheng goto out; 45255d4f98a2SYan Zheng 4526c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 45275d4f98a2SYan Zheng if (!rc) { 45285d4f98a2SYan Zheng err = -ENOMEM; 45295d4f98a2SYan Zheng goto out; 45305d4f98a2SYan Zheng } 45315d4f98a2SYan Zheng 45320b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 45335d4f98a2SYan Zheng 45345d4f98a2SYan Zheng set_reloc_control(rc); 45355d4f98a2SYan Zheng 45367a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 45373612b495STsutomu Itoh if (IS_ERR(trans)) { 45383612b495STsutomu Itoh err = PTR_ERR(trans); 4539fb2d83eeSJosef Bacik goto out_unset; 45403612b495STsutomu Itoh } 45413fd0a558SYan, Zheng 45423fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 45433fd0a558SYan, Zheng 45445d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 45455d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 45465d4f98a2SYan Zheng struct btrfs_root, root_list); 45475d4f98a2SYan Zheng list_del(&reloc_root->root_list); 45485d4f98a2SYan Zheng 45495d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 45505d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 45515d4f98a2SYan Zheng &rc->reloc_roots); 45525d4f98a2SYan Zheng continue; 45535d4f98a2SYan Zheng } 45545d4f98a2SYan Zheng 45550b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, reloc_root->root_key.offset); 455679787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 455779787eaaSJeff Mahoney err = PTR_ERR(fs_root); 4558ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 45591402d17dSXiyu Yang btrfs_end_transaction(trans); 4560fb2d83eeSJosef Bacik goto out_unset; 456179787eaaSJeff Mahoney } 45625d4f98a2SYan Zheng 4563ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 456479787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 4565f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 456600246528SJosef Bacik btrfs_put_root(fs_root); 45675d4f98a2SYan Zheng } 45685d4f98a2SYan Zheng 45693a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 457079787eaaSJeff Mahoney if (err) 4571fb2d83eeSJosef Bacik goto out_unset; 45725d4f98a2SYan Zheng 45735d4f98a2SYan Zheng merge_reloc_roots(rc); 45745d4f98a2SYan Zheng 45755d4f98a2SYan Zheng unset_reloc_control(rc); 45765d4f98a2SYan Zheng 45777a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 457862b99540SQu Wenruo if (IS_ERR(trans)) { 45793612b495STsutomu Itoh err = PTR_ERR(trans); 45806217b0faSJosef Bacik goto out_clean; 458162b99540SQu Wenruo } 45823a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 45836217b0faSJosef Bacik out_clean: 4584d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4585d2311e69SQu Wenruo if (ret < 0 && !err) 4586d2311e69SQu Wenruo err = ret; 4587fb2d83eeSJosef Bacik out_unset: 4588fb2d83eeSJosef Bacik unset_reloc_control(rc); 45891a0afa0eSJosef Bacik free_reloc_control(rc); 45903612b495STsutomu Itoh out: 4591aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 4592aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4593aca1bba6SLiu Bo 45945d4f98a2SYan Zheng btrfs_free_path(path); 45955d4f98a2SYan Zheng 45965d4f98a2SYan Zheng if (err == 0) { 45975d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 45980b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 4599932fd26dSJosef Bacik if (IS_ERR(fs_root)) { 46005d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4601932fd26dSJosef Bacik } else { 460266b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 460300246528SJosef Bacik btrfs_put_root(fs_root); 4604932fd26dSJosef Bacik } 4605932fd26dSJosef Bacik } 46065d4f98a2SYan Zheng return err; 46075d4f98a2SYan Zheng } 46085d4f98a2SYan Zheng 46095d4f98a2SYan Zheng /* 46105d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 46115d4f98a2SYan Zheng * 46125d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 46135d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 46145d4f98a2SYan Zheng */ 46155d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 46165d4f98a2SYan Zheng { 46170b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 46185d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 46195d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 46205d4f98a2SYan Zheng int ret; 46215d4f98a2SYan Zheng u64 disk_bytenr; 46224577b014SJosef Bacik u64 new_bytenr; 46235d4f98a2SYan Zheng LIST_HEAD(list); 46245d4f98a2SYan Zheng 46255d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4626bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 46275d4f98a2SYan Zheng 46285d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 46290b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4630a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 463179787eaaSJeff Mahoney if (ret) 463279787eaaSJeff Mahoney goto out; 46335d4f98a2SYan Zheng 46345d4f98a2SYan Zheng while (!list_empty(&list)) { 46355d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 46365d4f98a2SYan Zheng list_del_init(&sums->list); 46375d4f98a2SYan Zheng 46384577b014SJosef Bacik /* 46394577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 46404577b014SJosef Bacik * We need to do this because we will read in entire prealloc 46414577b014SJosef Bacik * extents but we may have written to say the middle of the 46424577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 46434577b014SJosef Bacik * the right disk offset. 46444577b014SJosef Bacik * 46454577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 46464577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 46474577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 46484577b014SJosef Bacik * disk length. 46494577b014SJosef Bacik */ 4650bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 46514577b014SJosef Bacik sums->bytenr = new_bytenr; 46525d4f98a2SYan Zheng 4653f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 46545d4f98a2SYan Zheng } 465579787eaaSJeff Mahoney out: 46565d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4657411fc6bcSAndi Kleen return ret; 46585d4f98a2SYan Zheng } 46593fd0a558SYan, Zheng 466083d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 46613fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 46623fd0a558SYan, Zheng struct extent_buffer *cow) 46633fd0a558SYan, Zheng { 46640b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 46653fd0a558SYan, Zheng struct reloc_control *rc; 46663fd0a558SYan, Zheng struct backref_node *node; 46673fd0a558SYan, Zheng int first_cow = 0; 46683fd0a558SYan, Zheng int level; 466983d4cfd4SJosef Bacik int ret = 0; 46703fd0a558SYan, Zheng 46710b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 46723fd0a558SYan, Zheng if (!rc) 467383d4cfd4SJosef Bacik return 0; 46743fd0a558SYan, Zheng 46753fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 46763fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 46773fd0a558SYan, Zheng 46783fd0a558SYan, Zheng level = btrfs_header_level(buf); 46793fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 46803fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 46813fd0a558SYan, Zheng first_cow = 1; 46823fd0a558SYan, Zheng 46833fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 46843fd0a558SYan, Zheng rc->create_reloc_tree) { 46853fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 46863fd0a558SYan, Zheng 46873fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 46883fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 46893fd0a558SYan, Zheng node->new_bytenr != buf->start); 46903fd0a558SYan, Zheng 46913fd0a558SYan, Zheng drop_node_buffer(node); 469267439dadSDavid Sterba atomic_inc(&cow->refs); 46933fd0a558SYan, Zheng node->eb = cow; 46943fd0a558SYan, Zheng node->new_bytenr = cow->start; 46953fd0a558SYan, Zheng 46963fd0a558SYan, Zheng if (!node->pending) { 46973fd0a558SYan, Zheng list_move_tail(&node->list, 46983fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 46993fd0a558SYan, Zheng node->pending = 1; 47003fd0a558SYan, Zheng } 47013fd0a558SYan, Zheng 47023fd0a558SYan, Zheng if (first_cow) 47039569cc20SQu Wenruo mark_block_processed(rc, node); 47043fd0a558SYan, Zheng 47053fd0a558SYan, Zheng if (first_cow && level > 0) 47063fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 47073fd0a558SYan, Zheng } 47083fd0a558SYan, Zheng 470983d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 47103fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 471183d4cfd4SJosef Bacik return ret; 47123fd0a558SYan, Zheng } 47133fd0a558SYan, Zheng 47143fd0a558SYan, Zheng /* 47153fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 471601327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 47173fd0a558SYan, Zheng */ 4718147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 47193fd0a558SYan, Zheng u64 *bytes_to_reserve) 47203fd0a558SYan, Zheng { 472110995c04SQu Wenruo struct btrfs_root *root = pending->root; 472210995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 47233fd0a558SYan, Zheng 47246282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 47253fd0a558SYan, Zheng return; 47263fd0a558SYan, Zheng 47273fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 47283fd0a558SYan, Zheng return; 47293fd0a558SYan, Zheng 47303fd0a558SYan, Zheng root = root->reloc_root; 47313fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 47323fd0a558SYan, Zheng /* 47333fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 47343fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 47353fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 47363fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 47373fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 47383fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 47393fd0a558SYan, Zheng * tree may use more space than it frees. so we need 47403fd0a558SYan, Zheng * reserve extra space. 47413fd0a558SYan, Zheng */ 47423fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 47433fd0a558SYan, Zheng } 47443fd0a558SYan, Zheng 47453fd0a558SYan, Zheng /* 47463fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 47473fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4748f44deb74SJosef Bacik * 4749f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4750f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4751f44deb74SJosef Bacik * rc->reloc_roots. 47523fd0a558SYan, Zheng */ 475349b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 47543fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 47553fd0a558SYan, Zheng { 47563fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 47573fd0a558SYan, Zheng struct btrfs_root *reloc_root; 47583fd0a558SYan, Zheng struct btrfs_root *new_root; 475910995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 47603fd0a558SYan, Zheng int ret; 47613fd0a558SYan, Zheng 47626282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 476349b25e05SJeff Mahoney return 0; 47643fd0a558SYan, Zheng 47653fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 47663fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 47673fd0a558SYan, Zheng 47683fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 47693fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 47703fd0a558SYan, Zheng rc->block_rsv, 47713a584174SLu Fengqi rc->nodes_relocated, true); 477249b25e05SJeff Mahoney if (ret) 477349b25e05SJeff Mahoney return ret; 47743fd0a558SYan, Zheng } 47753fd0a558SYan, Zheng 47763fd0a558SYan, Zheng new_root = pending->snap; 47773fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 47783fd0a558SYan, Zheng new_root->root_key.objectid); 477949b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 478049b25e05SJeff Mahoney return PTR_ERR(reloc_root); 47813fd0a558SYan, Zheng 4782ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4783ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4784f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 47853fd0a558SYan, Zheng 478649b25e05SJeff Mahoney if (rc->create_reloc_tree) 47873fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 478849b25e05SJeff Mahoney return ret; 47893fd0a558SYan, Zheng } 4790