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; 952a979612SQu Wenruo 962a979612SQu Wenruo /* List of upper level edges, which link this node to its parents */ 975d4f98a2SYan Zheng struct list_head upper; 982a979612SQu Wenruo /* List of lower level edges, which link this node to its children */ 995d4f98a2SYan Zheng struct list_head lower; 1002a979612SQu Wenruo 1015d4f98a2SYan Zheng /* NULL if this node is not tree root */ 1025d4f98a2SYan Zheng struct btrfs_root *root; 1035d4f98a2SYan Zheng /* extent buffer got by COW the block */ 1045d4f98a2SYan Zheng struct extent_buffer *eb; 1055d4f98a2SYan Zheng /* level of tree block */ 1065d4f98a2SYan Zheng unsigned int level:8; 1073fd0a558SYan, Zheng /* is the block in non-reference counted tree */ 1083fd0a558SYan, Zheng unsigned int cowonly:1; 1093fd0a558SYan, Zheng /* 1 if no child node in the cache */ 1105d4f98a2SYan Zheng unsigned int lowest:1; 1115d4f98a2SYan Zheng /* is the extent buffer locked */ 1125d4f98a2SYan Zheng unsigned int locked:1; 1135d4f98a2SYan Zheng /* has the block been processed */ 1145d4f98a2SYan Zheng unsigned int processed:1; 1155d4f98a2SYan Zheng /* have backrefs of this block been checked */ 1165d4f98a2SYan Zheng unsigned int checked:1; 1173fd0a558SYan, Zheng /* 1183fd0a558SYan, Zheng * 1 if corresponding block has been cowed but some upper 1193fd0a558SYan, Zheng * level block pointers may not point to the new location 1203fd0a558SYan, Zheng */ 1213fd0a558SYan, Zheng unsigned int pending:1; 1223fd0a558SYan, Zheng /* 1233fd0a558SYan, Zheng * 1 if the backref node isn't connected to any other 1243fd0a558SYan, Zheng * backref node. 1253fd0a558SYan, Zheng */ 1263fd0a558SYan, Zheng unsigned int detached:1; 1272433bea5SQu Wenruo 1282433bea5SQu Wenruo /* 1292433bea5SQu Wenruo * For generic purpose backref cache, where we only care if it's a reloc 1302433bea5SQu Wenruo * root, doesn't care the source subvolid. 1312433bea5SQu Wenruo */ 1322433bea5SQu Wenruo unsigned int is_reloc_root:1; 1335d4f98a2SYan Zheng }; 1345d4f98a2SYan Zheng 1355d4f98a2SYan Zheng #define LOWER 0 1365d4f98a2SYan Zheng #define UPPER 1 1370647bf56SWang Shilong #define RELOCATION_RESERVED_NODES 256 1382a979612SQu Wenruo /* 1392a979612SQu Wenruo * present an edge connecting upper and lower backref nodes. 1402a979612SQu Wenruo */ 1412a979612SQu Wenruo struct backref_edge { 1422a979612SQu Wenruo /* 1432a979612SQu Wenruo * list[LOWER] is linked to backref_node::upper of lower level node, 1442a979612SQu Wenruo * and list[UPPER] is linked to backref_node::lower of upper level node. 1452a979612SQu Wenruo * 1462a979612SQu Wenruo * Also, build_backref_tree() uses list[UPPER] for pending edges, before 1472a979612SQu Wenruo * linking list[UPPER] to its upper level nodes. 1482a979612SQu Wenruo */ 1492a979612SQu Wenruo struct list_head list[2]; 1502a979612SQu Wenruo 1512a979612SQu Wenruo /* Two related nodes */ 1522a979612SQu Wenruo struct backref_node *node[2]; 1532a979612SQu Wenruo }; 1542a979612SQu Wenruo 1555d4f98a2SYan Zheng 1565d4f98a2SYan Zheng struct backref_cache { 1575d4f98a2SYan Zheng /* red black tree of all backref nodes in the cache */ 1585d4f98a2SYan Zheng struct rb_root rb_root; 1593fd0a558SYan, Zheng /* for passing backref nodes to btrfs_reloc_cow_block */ 1603fd0a558SYan, Zheng struct backref_node *path[BTRFS_MAX_LEVEL]; 1613fd0a558SYan, Zheng /* 1623fd0a558SYan, Zheng * list of blocks that have been cowed but some block 1633fd0a558SYan, Zheng * pointers in upper level blocks may not reflect the 1643fd0a558SYan, Zheng * new location 1653fd0a558SYan, Zheng */ 1665d4f98a2SYan Zheng struct list_head pending[BTRFS_MAX_LEVEL]; 1673fd0a558SYan, Zheng /* list of backref nodes with no child node */ 1683fd0a558SYan, Zheng struct list_head leaves; 1693fd0a558SYan, Zheng /* list of blocks that have been cowed in current transaction */ 1703fd0a558SYan, Zheng struct list_head changed; 1713fd0a558SYan, Zheng /* list of detached backref node. */ 1723fd0a558SYan, Zheng struct list_head detached; 1733fd0a558SYan, Zheng 1743fd0a558SYan, Zheng u64 last_trans; 1753fd0a558SYan, Zheng 1763fd0a558SYan, Zheng int nr_nodes; 1773fd0a558SYan, Zheng int nr_edges; 17884780289SQu Wenruo 17984780289SQu Wenruo /* The list of unchecked backref edges during backref cache build */ 18084780289SQu Wenruo struct list_head pending_edge; 18184780289SQu Wenruo 18284780289SQu Wenruo /* The list of useless backref nodes during backref cache build */ 18384780289SQu Wenruo struct list_head useless_node; 18433a0f1f7SQu Wenruo 18533a0f1f7SQu Wenruo struct btrfs_fs_info *fs_info; 1862433bea5SQu Wenruo 1872433bea5SQu Wenruo /* 1882433bea5SQu Wenruo * Whether this cache is for relocation 1892433bea5SQu Wenruo * 1902433bea5SQu Wenruo * Reloction backref cache require more info for reloc root compared 1912433bea5SQu Wenruo * to generic backref cache. 1922433bea5SQu Wenruo */ 1932433bea5SQu Wenruo unsigned int is_reloc; 1945d4f98a2SYan Zheng }; 1955d4f98a2SYan Zheng 1965d4f98a2SYan Zheng /* 1975d4f98a2SYan Zheng * map address of tree root to tree 1985d4f98a2SYan Zheng */ 1995d4f98a2SYan Zheng struct mapping_node { 2005d4f98a2SYan Zheng struct rb_node rb_node; 2015d4f98a2SYan Zheng u64 bytenr; 2025d4f98a2SYan Zheng void *data; 2035d4f98a2SYan Zheng }; 2045d4f98a2SYan Zheng 2055d4f98a2SYan Zheng struct mapping_tree { 2065d4f98a2SYan Zheng struct rb_root rb_root; 2075d4f98a2SYan Zheng spinlock_t lock; 2085d4f98a2SYan Zheng }; 2095d4f98a2SYan Zheng 2105d4f98a2SYan Zheng /* 2115d4f98a2SYan Zheng * present a tree block to process 2125d4f98a2SYan Zheng */ 2135d4f98a2SYan Zheng struct tree_block { 2145d4f98a2SYan Zheng struct rb_node rb_node; 2155d4f98a2SYan Zheng u64 bytenr; 2165d4f98a2SYan Zheng struct btrfs_key key; 2175d4f98a2SYan Zheng unsigned int level:8; 2185d4f98a2SYan Zheng unsigned int key_ready:1; 2195d4f98a2SYan Zheng }; 2205d4f98a2SYan Zheng 2210257bb82SYan, Zheng #define MAX_EXTENTS 128 2220257bb82SYan, Zheng 2230257bb82SYan, Zheng struct file_extent_cluster { 2240257bb82SYan, Zheng u64 start; 2250257bb82SYan, Zheng u64 end; 2260257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 2270257bb82SYan, Zheng unsigned int nr; 2280257bb82SYan, Zheng }; 2290257bb82SYan, Zheng 2305d4f98a2SYan Zheng struct reloc_control { 2315d4f98a2SYan Zheng /* block group to relocate */ 23232da5386SDavid Sterba struct btrfs_block_group *block_group; 2335d4f98a2SYan Zheng /* extent tree */ 2345d4f98a2SYan Zheng struct btrfs_root *extent_root; 2355d4f98a2SYan Zheng /* inode for moving data */ 2365d4f98a2SYan Zheng struct inode *data_inode; 2373fd0a558SYan, Zheng 2383fd0a558SYan, Zheng struct btrfs_block_rsv *block_rsv; 2393fd0a558SYan, Zheng 2403fd0a558SYan, Zheng struct backref_cache backref_cache; 2413fd0a558SYan, Zheng 2423fd0a558SYan, Zheng struct file_extent_cluster cluster; 2435d4f98a2SYan Zheng /* tree blocks have been processed */ 2445d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 2455d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 2465d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 2475d4f98a2SYan Zheng /* list of reloc trees */ 2485d4f98a2SYan Zheng struct list_head reloc_roots; 249d2311e69SQu Wenruo /* list of subvolume trees that get relocated */ 250d2311e69SQu Wenruo struct list_head dirty_subvol_roots; 2513fd0a558SYan, Zheng /* size of metadata reservation for merging reloc trees */ 2523fd0a558SYan, Zheng u64 merging_rsv_size; 2533fd0a558SYan, Zheng /* size of relocated tree nodes */ 2543fd0a558SYan, Zheng u64 nodes_relocated; 2550647bf56SWang Shilong /* reserved size for block group relocation*/ 2560647bf56SWang Shilong u64 reserved_bytes; 2573fd0a558SYan, Zheng 2585d4f98a2SYan Zheng u64 search_start; 2595d4f98a2SYan Zheng u64 extents_found; 2603fd0a558SYan, Zheng 2613fd0a558SYan, Zheng unsigned int stage:8; 2623fd0a558SYan, Zheng unsigned int create_reloc_tree:1; 2633fd0a558SYan, Zheng unsigned int merge_reloc_tree:1; 2645d4f98a2SYan Zheng unsigned int found_file_extent:1; 2655d4f98a2SYan Zheng }; 2665d4f98a2SYan Zheng 2675d4f98a2SYan Zheng /* stages of data relocation */ 2685d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 2695d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 2705d4f98a2SYan Zheng 2713fd0a558SYan, Zheng static void remove_backref_node(struct backref_cache *cache, 2723fd0a558SYan, Zheng struct backref_node *node); 2739569cc20SQu Wenruo 2749569cc20SQu Wenruo static void mark_block_processed(struct reloc_control *rc, 2759569cc20SQu Wenruo struct backref_node *node) 2769569cc20SQu Wenruo { 2779569cc20SQu Wenruo u32 blocksize; 2789569cc20SQu Wenruo 2799569cc20SQu Wenruo if (node->level == 0 || 2809569cc20SQu Wenruo in_range(node->bytenr, rc->block_group->start, 2819569cc20SQu Wenruo rc->block_group->length)) { 2829569cc20SQu Wenruo blocksize = rc->extent_root->fs_info->nodesize; 2839569cc20SQu Wenruo set_extent_bits(&rc->processed_blocks, node->bytenr, 2849569cc20SQu Wenruo node->bytenr + blocksize - 1, EXTENT_DIRTY); 2859569cc20SQu Wenruo } 2869569cc20SQu Wenruo node->processed = 1; 2879569cc20SQu Wenruo } 2889569cc20SQu Wenruo 2895d4f98a2SYan Zheng 2905d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 2915d4f98a2SYan Zheng { 2926bef4d31SEric Paris tree->rb_root = RB_ROOT; 2935d4f98a2SYan Zheng spin_lock_init(&tree->lock); 2945d4f98a2SYan Zheng } 2955d4f98a2SYan Zheng 29633a0f1f7SQu Wenruo static void backref_cache_init(struct btrfs_fs_info *fs_info, 2972433bea5SQu Wenruo struct backref_cache *cache, int is_reloc) 2985d4f98a2SYan Zheng { 2995d4f98a2SYan Zheng int i; 3006bef4d31SEric Paris cache->rb_root = RB_ROOT; 3015d4f98a2SYan Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 3025d4f98a2SYan Zheng INIT_LIST_HEAD(&cache->pending[i]); 3033fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->changed); 3043fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->detached); 3053fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->leaves); 30684780289SQu Wenruo INIT_LIST_HEAD(&cache->pending_edge); 30784780289SQu Wenruo INIT_LIST_HEAD(&cache->useless_node); 30833a0f1f7SQu Wenruo cache->fs_info = fs_info; 3092433bea5SQu Wenruo cache->is_reloc = is_reloc; 3105d4f98a2SYan Zheng } 3115d4f98a2SYan Zheng 3123fd0a558SYan, Zheng static void backref_cache_cleanup(struct backref_cache *cache) 3135d4f98a2SYan Zheng { 3143fd0a558SYan, Zheng struct backref_node *node; 3153fd0a558SYan, Zheng int i; 3163fd0a558SYan, Zheng 3173fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 3183fd0a558SYan, Zheng node = list_entry(cache->detached.next, 3193fd0a558SYan, Zheng struct backref_node, list); 3203fd0a558SYan, Zheng remove_backref_node(cache, node); 3213fd0a558SYan, Zheng } 3223fd0a558SYan, Zheng 3233fd0a558SYan, Zheng while (!list_empty(&cache->leaves)) { 3243fd0a558SYan, Zheng node = list_entry(cache->leaves.next, 3253fd0a558SYan, Zheng struct backref_node, lower); 3263fd0a558SYan, Zheng remove_backref_node(cache, node); 3273fd0a558SYan, Zheng } 3283fd0a558SYan, Zheng 3293fd0a558SYan, Zheng cache->last_trans = 0; 3303fd0a558SYan, Zheng 3313fd0a558SYan, Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 332f4907095SLiu Bo ASSERT(list_empty(&cache->pending[i])); 33384780289SQu Wenruo ASSERT(list_empty(&cache->pending_edge)); 33484780289SQu Wenruo ASSERT(list_empty(&cache->useless_node)); 335f4907095SLiu Bo ASSERT(list_empty(&cache->changed)); 336f4907095SLiu Bo ASSERT(list_empty(&cache->detached)); 337f4907095SLiu Bo ASSERT(RB_EMPTY_ROOT(&cache->rb_root)); 338f4907095SLiu Bo ASSERT(!cache->nr_nodes); 339f4907095SLiu Bo ASSERT(!cache->nr_edges); 3403fd0a558SYan, Zheng } 3413fd0a558SYan, Zheng 3420304f2d8SQu Wenruo static struct backref_node *alloc_backref_node(struct backref_cache *cache, 3430304f2d8SQu Wenruo u64 bytenr, int level) 3443fd0a558SYan, Zheng { 3453fd0a558SYan, Zheng struct backref_node *node; 3463fd0a558SYan, Zheng 3470304f2d8SQu Wenruo ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL); 3483fd0a558SYan, Zheng node = kzalloc(sizeof(*node), GFP_NOFS); 3490304f2d8SQu Wenruo if (!node) 3500304f2d8SQu Wenruo return node; 3510304f2d8SQu Wenruo 3523fd0a558SYan, Zheng INIT_LIST_HEAD(&node->list); 3535d4f98a2SYan Zheng INIT_LIST_HEAD(&node->upper); 3545d4f98a2SYan Zheng INIT_LIST_HEAD(&node->lower); 3555d4f98a2SYan Zheng RB_CLEAR_NODE(&node->rb_node); 3563fd0a558SYan, Zheng cache->nr_nodes++; 3570304f2d8SQu Wenruo node->level = level; 3580304f2d8SQu Wenruo node->bytenr = bytenr; 3590304f2d8SQu Wenruo 3603fd0a558SYan, Zheng return node; 3613fd0a558SYan, Zheng } 3623fd0a558SYan, Zheng 3633fd0a558SYan, Zheng static void free_backref_node(struct backref_cache *cache, 3643fd0a558SYan, Zheng struct backref_node *node) 3653fd0a558SYan, Zheng { 3663fd0a558SYan, Zheng if (node) { 3673fd0a558SYan, Zheng cache->nr_nodes--; 36800246528SJosef Bacik btrfs_put_root(node->root); 3693fd0a558SYan, Zheng kfree(node); 3703fd0a558SYan, Zheng } 3713fd0a558SYan, Zheng } 3723fd0a558SYan, Zheng 3733fd0a558SYan, Zheng static struct backref_edge *alloc_backref_edge(struct backref_cache *cache) 3743fd0a558SYan, Zheng { 3753fd0a558SYan, Zheng struct backref_edge *edge; 3763fd0a558SYan, Zheng 3773fd0a558SYan, Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 3783fd0a558SYan, Zheng if (edge) 3793fd0a558SYan, Zheng cache->nr_edges++; 3803fd0a558SYan, Zheng return edge; 3813fd0a558SYan, Zheng } 3823fd0a558SYan, Zheng 3832a979612SQu Wenruo #define LINK_LOWER (1 << 0) 3842a979612SQu Wenruo #define LINK_UPPER (1 << 1) 3852a979612SQu Wenruo static void link_backref_edge(struct backref_edge *edge, 3862a979612SQu Wenruo struct backref_node *lower, 3872a979612SQu Wenruo struct backref_node *upper, 3882a979612SQu Wenruo int link_which) 3892a979612SQu Wenruo { 3902a979612SQu Wenruo ASSERT(upper && lower && upper->level == lower->level + 1); 3912a979612SQu Wenruo edge->node[LOWER] = lower; 3922a979612SQu Wenruo edge->node[UPPER] = upper; 3932a979612SQu Wenruo if (link_which & LINK_LOWER) 3942a979612SQu Wenruo list_add_tail(&edge->list[LOWER], &lower->upper); 3952a979612SQu Wenruo if (link_which & LINK_UPPER) 3962a979612SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 3972a979612SQu Wenruo } 3982a979612SQu Wenruo 3993fd0a558SYan, Zheng static void free_backref_edge(struct backref_cache *cache, 4003fd0a558SYan, Zheng struct backref_edge *edge) 4013fd0a558SYan, Zheng { 4023fd0a558SYan, Zheng if (edge) { 4033fd0a558SYan, Zheng cache->nr_edges--; 4043fd0a558SYan, Zheng kfree(edge); 4053fd0a558SYan, Zheng } 4065d4f98a2SYan Zheng } 4075d4f98a2SYan Zheng 4085d4f98a2SYan Zheng static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr, 4095d4f98a2SYan Zheng struct rb_node *node) 4105d4f98a2SYan Zheng { 4115d4f98a2SYan Zheng struct rb_node **p = &root->rb_node; 4125d4f98a2SYan Zheng struct rb_node *parent = NULL; 4135d4f98a2SYan Zheng struct tree_entry *entry; 4145d4f98a2SYan Zheng 4155d4f98a2SYan Zheng while (*p) { 4165d4f98a2SYan Zheng parent = *p; 4175d4f98a2SYan Zheng entry = rb_entry(parent, struct tree_entry, rb_node); 4185d4f98a2SYan Zheng 4195d4f98a2SYan Zheng if (bytenr < entry->bytenr) 4205d4f98a2SYan Zheng p = &(*p)->rb_left; 4215d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 4225d4f98a2SYan Zheng p = &(*p)->rb_right; 4235d4f98a2SYan Zheng else 4245d4f98a2SYan Zheng return parent; 4255d4f98a2SYan Zheng } 4265d4f98a2SYan Zheng 4275d4f98a2SYan Zheng rb_link_node(node, parent, p); 4285d4f98a2SYan Zheng rb_insert_color(node, root); 4295d4f98a2SYan Zheng return NULL; 4305d4f98a2SYan Zheng } 4315d4f98a2SYan Zheng 4325d4f98a2SYan Zheng static struct rb_node *tree_search(struct rb_root *root, u64 bytenr) 4335d4f98a2SYan Zheng { 4345d4f98a2SYan Zheng struct rb_node *n = root->rb_node; 4355d4f98a2SYan Zheng struct tree_entry *entry; 4365d4f98a2SYan Zheng 4375d4f98a2SYan Zheng while (n) { 4385d4f98a2SYan Zheng entry = rb_entry(n, struct tree_entry, rb_node); 4395d4f98a2SYan Zheng 4405d4f98a2SYan Zheng if (bytenr < entry->bytenr) 4415d4f98a2SYan Zheng n = n->rb_left; 4425d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 4435d4f98a2SYan Zheng n = n->rb_right; 4445d4f98a2SYan Zheng else 4455d4f98a2SYan Zheng return n; 4465d4f98a2SYan Zheng } 4475d4f98a2SYan Zheng return NULL; 4485d4f98a2SYan Zheng } 4495d4f98a2SYan Zheng 45048a3b636SEric Sandeen static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr) 45143c04fb1SJeff Mahoney { 45243c04fb1SJeff Mahoney 45343c04fb1SJeff Mahoney struct btrfs_fs_info *fs_info = NULL; 45443c04fb1SJeff Mahoney struct backref_node *bnode = rb_entry(rb_node, struct backref_node, 45543c04fb1SJeff Mahoney rb_node); 45643c04fb1SJeff Mahoney if (bnode->root) 45743c04fb1SJeff Mahoney fs_info = bnode->root->fs_info; 4585d163e0eSJeff Mahoney btrfs_panic(fs_info, errno, 4595d163e0eSJeff Mahoney "Inconsistency in backref cache found at offset %llu", 4605d163e0eSJeff Mahoney bytenr); 46143c04fb1SJeff Mahoney } 46243c04fb1SJeff Mahoney 4635d4f98a2SYan Zheng /* 4645d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 4655d4f98a2SYan Zheng */ 4665d4f98a2SYan Zheng static struct backref_node *walk_up_backref(struct backref_node *node, 4675d4f98a2SYan Zheng struct backref_edge *edges[], 4685d4f98a2SYan Zheng int *index) 4695d4f98a2SYan Zheng { 4705d4f98a2SYan Zheng struct backref_edge *edge; 4715d4f98a2SYan Zheng int idx = *index; 4725d4f98a2SYan Zheng 4735d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 4745d4f98a2SYan Zheng edge = list_entry(node->upper.next, 4755d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 4765d4f98a2SYan Zheng edges[idx++] = edge; 4775d4f98a2SYan Zheng node = edge->node[UPPER]; 4785d4f98a2SYan Zheng } 4793fd0a558SYan, Zheng BUG_ON(node->detached); 4805d4f98a2SYan Zheng *index = idx; 4815d4f98a2SYan Zheng return node; 4825d4f98a2SYan Zheng } 4835d4f98a2SYan Zheng 4845d4f98a2SYan Zheng /* 4855d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 4865d4f98a2SYan Zheng */ 4875d4f98a2SYan Zheng static struct backref_node *walk_down_backref(struct backref_edge *edges[], 4885d4f98a2SYan Zheng int *index) 4895d4f98a2SYan Zheng { 4905d4f98a2SYan Zheng struct backref_edge *edge; 4915d4f98a2SYan Zheng struct backref_node *lower; 4925d4f98a2SYan Zheng int idx = *index; 4935d4f98a2SYan Zheng 4945d4f98a2SYan Zheng while (idx > 0) { 4955d4f98a2SYan Zheng edge = edges[idx - 1]; 4965d4f98a2SYan Zheng lower = edge->node[LOWER]; 4975d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 4985d4f98a2SYan Zheng idx--; 4995d4f98a2SYan Zheng continue; 5005d4f98a2SYan Zheng } 5015d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 5025d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 5035d4f98a2SYan Zheng edges[idx - 1] = edge; 5045d4f98a2SYan Zheng *index = idx; 5055d4f98a2SYan Zheng return edge->node[UPPER]; 5065d4f98a2SYan Zheng } 5075d4f98a2SYan Zheng *index = 0; 5085d4f98a2SYan Zheng return NULL; 5095d4f98a2SYan Zheng } 5105d4f98a2SYan Zheng 5113fd0a558SYan, Zheng static void unlock_node_buffer(struct backref_node *node) 5125d4f98a2SYan Zheng { 5135d4f98a2SYan Zheng if (node->locked) { 5145d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 5155d4f98a2SYan Zheng node->locked = 0; 5165d4f98a2SYan Zheng } 5173fd0a558SYan, Zheng } 5183fd0a558SYan, Zheng 5193fd0a558SYan, Zheng static void drop_node_buffer(struct backref_node *node) 5203fd0a558SYan, Zheng { 5213fd0a558SYan, Zheng if (node->eb) { 5223fd0a558SYan, Zheng unlock_node_buffer(node); 5235d4f98a2SYan Zheng free_extent_buffer(node->eb); 5245d4f98a2SYan Zheng node->eb = NULL; 5255d4f98a2SYan Zheng } 5265d4f98a2SYan Zheng } 5275d4f98a2SYan Zheng 5285d4f98a2SYan Zheng static void drop_backref_node(struct backref_cache *tree, 5295d4f98a2SYan Zheng struct backref_node *node) 5305d4f98a2SYan Zheng { 5315d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 5325d4f98a2SYan Zheng 5335d4f98a2SYan Zheng drop_node_buffer(node); 5343fd0a558SYan, Zheng list_del(&node->list); 5355d4f98a2SYan Zheng list_del(&node->lower); 5363fd0a558SYan, Zheng if (!RB_EMPTY_NODE(&node->rb_node)) 5375d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 5383fd0a558SYan, Zheng free_backref_node(tree, node); 5395d4f98a2SYan Zheng } 5405d4f98a2SYan Zheng 5415d4f98a2SYan Zheng /* 5425d4f98a2SYan Zheng * remove a backref node from the backref cache 5435d4f98a2SYan Zheng */ 5445d4f98a2SYan Zheng static void remove_backref_node(struct backref_cache *cache, 5455d4f98a2SYan Zheng struct backref_node *node) 5465d4f98a2SYan Zheng { 5475d4f98a2SYan Zheng struct backref_node *upper; 5485d4f98a2SYan Zheng struct backref_edge *edge; 5495d4f98a2SYan Zheng 5505d4f98a2SYan Zheng if (!node) 5515d4f98a2SYan Zheng return; 5525d4f98a2SYan Zheng 5533fd0a558SYan, Zheng BUG_ON(!node->lowest && !node->detached); 5545d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 5555d4f98a2SYan Zheng edge = list_entry(node->upper.next, struct backref_edge, 5565d4f98a2SYan Zheng list[LOWER]); 5575d4f98a2SYan Zheng upper = edge->node[UPPER]; 5585d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 5595d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 5603fd0a558SYan, Zheng free_backref_edge(cache, edge); 5613fd0a558SYan, Zheng 5623fd0a558SYan, Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 5633fd0a558SYan, Zheng BUG_ON(!list_empty(&node->upper)); 5643fd0a558SYan, Zheng drop_backref_node(cache, node); 5653fd0a558SYan, Zheng node = upper; 5663fd0a558SYan, Zheng node->lowest = 1; 5673fd0a558SYan, Zheng continue; 5683fd0a558SYan, Zheng } 5695d4f98a2SYan Zheng /* 5703fd0a558SYan, Zheng * add the node to leaf node list if no other 5715d4f98a2SYan Zheng * child block cached. 5725d4f98a2SYan Zheng */ 5735d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 5743fd0a558SYan, Zheng list_add_tail(&upper->lower, &cache->leaves); 5755d4f98a2SYan Zheng upper->lowest = 1; 5765d4f98a2SYan Zheng } 5775d4f98a2SYan Zheng } 5783fd0a558SYan, Zheng 5795d4f98a2SYan Zheng drop_backref_node(cache, node); 5805d4f98a2SYan Zheng } 5815d4f98a2SYan Zheng 5823fd0a558SYan, Zheng static void update_backref_node(struct backref_cache *cache, 5833fd0a558SYan, Zheng struct backref_node *node, u64 bytenr) 5843fd0a558SYan, Zheng { 5853fd0a558SYan, Zheng struct rb_node *rb_node; 5863fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 5873fd0a558SYan, Zheng node->bytenr = bytenr; 5883fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 58943c04fb1SJeff Mahoney if (rb_node) 59043c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, bytenr); 5913fd0a558SYan, Zheng } 5923fd0a558SYan, Zheng 5933fd0a558SYan, Zheng /* 5943fd0a558SYan, Zheng * update backref cache after a transaction commit 5953fd0a558SYan, Zheng */ 5963fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 5973fd0a558SYan, Zheng struct backref_cache *cache) 5983fd0a558SYan, Zheng { 5993fd0a558SYan, Zheng struct backref_node *node; 6003fd0a558SYan, Zheng int level = 0; 6013fd0a558SYan, Zheng 6023fd0a558SYan, Zheng if (cache->last_trans == 0) { 6033fd0a558SYan, Zheng cache->last_trans = trans->transid; 6043fd0a558SYan, Zheng return 0; 6053fd0a558SYan, Zheng } 6063fd0a558SYan, Zheng 6073fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 6083fd0a558SYan, Zheng return 0; 6093fd0a558SYan, Zheng 6103fd0a558SYan, Zheng /* 6113fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 6123fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 6133fd0a558SYan, Zheng * so the detached nodes are no longer useful. 6143fd0a558SYan, Zheng */ 6153fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 6163fd0a558SYan, Zheng node = list_entry(cache->detached.next, 6173fd0a558SYan, Zheng struct backref_node, list); 6183fd0a558SYan, Zheng remove_backref_node(cache, node); 6193fd0a558SYan, Zheng } 6203fd0a558SYan, Zheng 6213fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 6223fd0a558SYan, Zheng node = list_entry(cache->changed.next, 6233fd0a558SYan, Zheng struct backref_node, list); 6243fd0a558SYan, Zheng list_del_init(&node->list); 6253fd0a558SYan, Zheng BUG_ON(node->pending); 6263fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 6273fd0a558SYan, Zheng } 6283fd0a558SYan, Zheng 6293fd0a558SYan, Zheng /* 6303fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 6313fd0a558SYan, Zheng * errors during processing the pending nodes. 6323fd0a558SYan, Zheng */ 6333fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 6343fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 6353fd0a558SYan, Zheng BUG_ON(!node->pending); 6363fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 6373fd0a558SYan, Zheng continue; 6383fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 6393fd0a558SYan, Zheng } 6403fd0a558SYan, Zheng } 6413fd0a558SYan, Zheng 6423fd0a558SYan, Zheng cache->last_trans = 0; 6433fd0a558SYan, Zheng return 1; 6443fd0a558SYan, Zheng } 6453fd0a558SYan, Zheng 6466282675eSQu Wenruo static bool reloc_root_is_dead(struct btrfs_root *root) 6476282675eSQu Wenruo { 6486282675eSQu Wenruo /* 6496282675eSQu Wenruo * Pair with set_bit/clear_bit in clean_dirty_subvols and 6506282675eSQu Wenruo * btrfs_update_reloc_root. We need to see the updated bit before 6516282675eSQu Wenruo * trying to access reloc_root 6526282675eSQu Wenruo */ 6536282675eSQu Wenruo smp_rmb(); 6546282675eSQu Wenruo if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state)) 6556282675eSQu Wenruo return true; 6566282675eSQu Wenruo return false; 6576282675eSQu Wenruo } 6586282675eSQu Wenruo 6596282675eSQu Wenruo /* 6606282675eSQu Wenruo * Check if this subvolume tree has valid reloc tree. 6616282675eSQu Wenruo * 6626282675eSQu Wenruo * Reloc tree after swap is considered dead, thus not considered as valid. 6636282675eSQu Wenruo * This is enough for most callers, as they don't distinguish dead reloc root 6646282675eSQu Wenruo * from no reloc root. But should_ignore_root() below is a special case. 6656282675eSQu Wenruo */ 6666282675eSQu Wenruo static bool have_reloc_root(struct btrfs_root *root) 6676282675eSQu Wenruo { 6686282675eSQu Wenruo if (reloc_root_is_dead(root)) 6696282675eSQu Wenruo return false; 6706282675eSQu Wenruo if (!root->reloc_root) 6716282675eSQu Wenruo return false; 6726282675eSQu Wenruo return true; 6736282675eSQu Wenruo } 674f2a97a9dSDavid Sterba 6753fd0a558SYan, Zheng static int should_ignore_root(struct btrfs_root *root) 6763fd0a558SYan, Zheng { 6773fd0a558SYan, Zheng struct btrfs_root *reloc_root; 6783fd0a558SYan, Zheng 67927cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 6803fd0a558SYan, Zheng return 0; 6813fd0a558SYan, Zheng 6826282675eSQu Wenruo /* This root has been merged with its reloc tree, we can ignore it */ 6836282675eSQu Wenruo if (reloc_root_is_dead(root)) 6846282675eSQu Wenruo return 1; 6856282675eSQu Wenruo 6863fd0a558SYan, Zheng reloc_root = root->reloc_root; 6873fd0a558SYan, Zheng if (!reloc_root) 6883fd0a558SYan, Zheng return 0; 6893fd0a558SYan, Zheng 6904d4225fcSJosef Bacik if (btrfs_header_generation(reloc_root->commit_root) == 6914d4225fcSJosef Bacik root->fs_info->running_transaction->transid) 6923fd0a558SYan, Zheng return 0; 6933fd0a558SYan, Zheng /* 6943fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 6953fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 6963fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 6973fd0a558SYan, Zheng * relocation. 6983fd0a558SYan, Zheng */ 6993fd0a558SYan, Zheng return 1; 7003fd0a558SYan, Zheng } 7015d4f98a2SYan Zheng /* 7025d4f98a2SYan Zheng * find reloc tree by address of tree root 7035d4f98a2SYan Zheng */ 7042433bea5SQu Wenruo struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr) 7055d4f98a2SYan Zheng { 7062433bea5SQu Wenruo struct reloc_control *rc = fs_info->reloc_ctl; 7075d4f98a2SYan Zheng struct rb_node *rb_node; 7085d4f98a2SYan Zheng struct mapping_node *node; 7095d4f98a2SYan Zheng struct btrfs_root *root = NULL; 7105d4f98a2SYan Zheng 7112433bea5SQu Wenruo ASSERT(rc); 7125d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 7135d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr); 7145d4f98a2SYan Zheng if (rb_node) { 7155d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 7165d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 7175d4f98a2SYan Zheng } 7185d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 71900246528SJosef Bacik return btrfs_grab_root(root); 7205d4f98a2SYan Zheng } 7215d4f98a2SYan Zheng 7225d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 7235d4f98a2SYan Zheng u64 root_objectid) 7245d4f98a2SYan Zheng { 7255d4f98a2SYan Zheng struct btrfs_key key; 7265d4f98a2SYan Zheng 7275d4f98a2SYan Zheng key.objectid = root_objectid; 7285d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 7295d4f98a2SYan Zheng key.offset = (u64)-1; 7305d4f98a2SYan Zheng 731bc44d7c4SJosef Bacik return btrfs_get_fs_root(fs_info, &key, false); 7325d4f98a2SYan Zheng } 7335d4f98a2SYan Zheng 7345d4f98a2SYan Zheng /* 7354007ea87SQu Wenruo * Handle direct tree backref 7364007ea87SQu Wenruo * 7374007ea87SQu Wenruo * Direct tree backref means, the backref item shows its parent bytenr 7384007ea87SQu Wenruo * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined). 7394007ea87SQu Wenruo * 7404007ea87SQu Wenruo * @ref_key: The converted backref key. 7414007ea87SQu Wenruo * For keyed backref, it's the item key. 7424007ea87SQu Wenruo * For inlined backref, objectid is the bytenr, 7434007ea87SQu Wenruo * type is btrfs_inline_ref_type, offset is 7444007ea87SQu Wenruo * btrfs_inline_ref_offset. 7454007ea87SQu Wenruo */ 7464007ea87SQu Wenruo static int handle_direct_tree_backref(struct backref_cache *cache, 7474007ea87SQu Wenruo struct btrfs_key *ref_key, 7484007ea87SQu Wenruo struct backref_node *cur) 7494007ea87SQu Wenruo { 7504007ea87SQu Wenruo struct backref_edge *edge; 7514007ea87SQu Wenruo struct backref_node *upper; 7524007ea87SQu Wenruo struct rb_node *rb_node; 7534007ea87SQu Wenruo 7544007ea87SQu Wenruo ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY); 7554007ea87SQu Wenruo 7564007ea87SQu Wenruo /* Only reloc root uses backref pointing to itself */ 7574007ea87SQu Wenruo if (ref_key->objectid == ref_key->offset) { 7584007ea87SQu Wenruo struct btrfs_root *root; 7594007ea87SQu Wenruo 7604007ea87SQu Wenruo cur->is_reloc_root = 1; 7614007ea87SQu Wenruo /* Only reloc backref cache cares about a specific root */ 7624007ea87SQu Wenruo if (cache->is_reloc) { 7634007ea87SQu Wenruo root = find_reloc_root(cache->fs_info, cur->bytenr); 7644007ea87SQu Wenruo if (WARN_ON(!root)) 7654007ea87SQu Wenruo return -ENOENT; 7664007ea87SQu Wenruo cur->root = root; 7674007ea87SQu Wenruo } else { 7684007ea87SQu Wenruo /* 7694007ea87SQu Wenruo * For generic purpose backref cache, reloc root node 7704007ea87SQu Wenruo * is useless. 7714007ea87SQu Wenruo */ 7724007ea87SQu Wenruo list_add(&cur->list, &cache->useless_node); 7734007ea87SQu Wenruo } 7744007ea87SQu Wenruo return 0; 7754007ea87SQu Wenruo } 7764007ea87SQu Wenruo 7774007ea87SQu Wenruo edge = alloc_backref_edge(cache); 7784007ea87SQu Wenruo if (!edge) 7794007ea87SQu Wenruo return -ENOMEM; 7804007ea87SQu Wenruo 7814007ea87SQu Wenruo rb_node = tree_search(&cache->rb_root, ref_key->offset); 7824007ea87SQu Wenruo if (!rb_node) { 7834007ea87SQu Wenruo /* Parent node not yet cached */ 7840304f2d8SQu Wenruo upper = alloc_backref_node(cache, ref_key->offset, 7850304f2d8SQu Wenruo cur->level + 1); 7864007ea87SQu Wenruo if (!upper) { 7874007ea87SQu Wenruo free_backref_edge(cache, edge); 7884007ea87SQu Wenruo return -ENOMEM; 7894007ea87SQu Wenruo } 7904007ea87SQu Wenruo 7914007ea87SQu Wenruo /* 7924007ea87SQu Wenruo * Backrefs for the upper level block isn't cached, add the 7934007ea87SQu Wenruo * block to pending list 7944007ea87SQu Wenruo */ 7954007ea87SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 7964007ea87SQu Wenruo } else { 7974007ea87SQu Wenruo /* Parent node already cached */ 7984007ea87SQu Wenruo upper = rb_entry(rb_node, struct backref_node, rb_node); 7994007ea87SQu Wenruo ASSERT(upper->checked); 8004007ea87SQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 8014007ea87SQu Wenruo } 8022a979612SQu Wenruo link_backref_edge(edge, cur, upper, LINK_LOWER); 8034007ea87SQu Wenruo return 0; 8044007ea87SQu Wenruo } 8054007ea87SQu Wenruo 8064007ea87SQu Wenruo /* 8074d81ea8bSQu Wenruo * Handle indirect tree backref 8084d81ea8bSQu Wenruo * 8094d81ea8bSQu Wenruo * Indirect tree backref means, we only know which tree the node belongs to. 8104d81ea8bSQu Wenruo * We still need to do a tree search to find out the parents. This is for 8114d81ea8bSQu Wenruo * TREE_BLOCK_REF backref (keyed or inlined). 8124d81ea8bSQu Wenruo * 8134d81ea8bSQu Wenruo * @ref_key: The same as @ref_key in handle_direct_tree_backref() 8144d81ea8bSQu Wenruo * @tree_key: The first key of this tree block. 8154d81ea8bSQu Wenruo * @path: A clean (released) path, to avoid allocating path everytime 8164d81ea8bSQu Wenruo * the function get called. 8174d81ea8bSQu Wenruo */ 8184d81ea8bSQu Wenruo static int handle_indirect_tree_backref(struct backref_cache *cache, 8194d81ea8bSQu Wenruo struct btrfs_path *path, 8204d81ea8bSQu Wenruo struct btrfs_key *ref_key, 8214d81ea8bSQu Wenruo struct btrfs_key *tree_key, 8224d81ea8bSQu Wenruo struct backref_node *cur) 8234d81ea8bSQu Wenruo { 8244d81ea8bSQu Wenruo struct btrfs_fs_info *fs_info = cache->fs_info; 8254d81ea8bSQu Wenruo struct backref_node *upper; 8264d81ea8bSQu Wenruo struct backref_node *lower; 8274d81ea8bSQu Wenruo struct backref_edge *edge; 8284d81ea8bSQu Wenruo struct extent_buffer *eb; 8294d81ea8bSQu Wenruo struct btrfs_root *root; 8304d81ea8bSQu Wenruo struct rb_node *rb_node; 8314d81ea8bSQu Wenruo int level; 8324d81ea8bSQu Wenruo bool need_check = true; 8334d81ea8bSQu Wenruo int ret; 8344d81ea8bSQu Wenruo 8354d81ea8bSQu Wenruo root = read_fs_root(fs_info, ref_key->offset); 8364d81ea8bSQu Wenruo if (IS_ERR(root)) 8374d81ea8bSQu Wenruo return PTR_ERR(root); 8384d81ea8bSQu Wenruo if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 8394d81ea8bSQu Wenruo cur->cowonly = 1; 8404d81ea8bSQu Wenruo 8414d81ea8bSQu Wenruo if (btrfs_root_level(&root->root_item) == cur->level) { 8424d81ea8bSQu Wenruo /* Tree root */ 8434d81ea8bSQu Wenruo ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr); 8444d81ea8bSQu Wenruo if (should_ignore_root(root)) { 8454d81ea8bSQu Wenruo btrfs_put_root(root); 8464d81ea8bSQu Wenruo list_add(&cur->list, &cache->useless_node); 8474d81ea8bSQu Wenruo } else { 8484d81ea8bSQu Wenruo cur->root = root; 8494d81ea8bSQu Wenruo } 8504d81ea8bSQu Wenruo return 0; 8514d81ea8bSQu Wenruo } 8524d81ea8bSQu Wenruo 8534d81ea8bSQu Wenruo level = cur->level + 1; 8544d81ea8bSQu Wenruo 8554d81ea8bSQu Wenruo /* Search the tree to find parent blocks referring to the block */ 8564d81ea8bSQu Wenruo path->search_commit_root = 1; 8574d81ea8bSQu Wenruo path->skip_locking = 1; 8584d81ea8bSQu Wenruo path->lowest_level = level; 8594d81ea8bSQu Wenruo ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0); 8604d81ea8bSQu Wenruo path->lowest_level = 0; 8614d81ea8bSQu Wenruo if (ret < 0) { 8624d81ea8bSQu Wenruo btrfs_put_root(root); 8634d81ea8bSQu Wenruo return ret; 8644d81ea8bSQu Wenruo } 8654d81ea8bSQu Wenruo if (ret > 0 && path->slots[level] > 0) 8664d81ea8bSQu Wenruo path->slots[level]--; 8674d81ea8bSQu Wenruo 8684d81ea8bSQu Wenruo eb = path->nodes[level]; 8694d81ea8bSQu Wenruo if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) { 8704d81ea8bSQu Wenruo btrfs_err(fs_info, 8714d81ea8bSQu Wenruo "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)", 8724d81ea8bSQu Wenruo cur->bytenr, level - 1, root->root_key.objectid, 8734d81ea8bSQu Wenruo tree_key->objectid, tree_key->type, tree_key->offset); 8744d81ea8bSQu Wenruo btrfs_put_root(root); 8754d81ea8bSQu Wenruo ret = -ENOENT; 8764d81ea8bSQu Wenruo goto out; 8774d81ea8bSQu Wenruo } 8784d81ea8bSQu Wenruo lower = cur; 8794d81ea8bSQu Wenruo 8804d81ea8bSQu Wenruo /* Add all nodes and edges in the path */ 8814d81ea8bSQu Wenruo for (; level < BTRFS_MAX_LEVEL; level++) { 8824d81ea8bSQu Wenruo if (!path->nodes[level]) { 8834d81ea8bSQu Wenruo ASSERT(btrfs_root_bytenr(&root->root_item) == 8844d81ea8bSQu Wenruo lower->bytenr); 8854d81ea8bSQu Wenruo if (should_ignore_root(root)) { 8864d81ea8bSQu Wenruo btrfs_put_root(root); 8874d81ea8bSQu Wenruo list_add(&lower->list, &cache->useless_node); 8884d81ea8bSQu Wenruo } else { 8894d81ea8bSQu Wenruo lower->root = root; 8904d81ea8bSQu Wenruo } 8914d81ea8bSQu Wenruo break; 8924d81ea8bSQu Wenruo } 8934d81ea8bSQu Wenruo 8944d81ea8bSQu Wenruo edge = alloc_backref_edge(cache); 8954d81ea8bSQu Wenruo if (!edge) { 8964d81ea8bSQu Wenruo btrfs_put_root(root); 8974d81ea8bSQu Wenruo ret = -ENOMEM; 8984d81ea8bSQu Wenruo goto out; 8994d81ea8bSQu Wenruo } 9004d81ea8bSQu Wenruo 9014d81ea8bSQu Wenruo eb = path->nodes[level]; 9024d81ea8bSQu Wenruo rb_node = tree_search(&cache->rb_root, eb->start); 9034d81ea8bSQu Wenruo if (!rb_node) { 9040304f2d8SQu Wenruo upper = alloc_backref_node(cache, eb->start, 9050304f2d8SQu Wenruo lower->level + 1); 9064d81ea8bSQu Wenruo if (!upper) { 9074d81ea8bSQu Wenruo btrfs_put_root(root); 9084d81ea8bSQu Wenruo free_backref_edge(cache, edge); 9094d81ea8bSQu Wenruo ret = -ENOMEM; 9104d81ea8bSQu Wenruo goto out; 9114d81ea8bSQu Wenruo } 9124d81ea8bSQu Wenruo upper->owner = btrfs_header_owner(eb); 9134d81ea8bSQu Wenruo if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 9144d81ea8bSQu Wenruo upper->cowonly = 1; 9154d81ea8bSQu Wenruo 9164d81ea8bSQu Wenruo /* 9174d81ea8bSQu Wenruo * If we know the block isn't shared we can avoid 9184d81ea8bSQu Wenruo * checking its backrefs. 9194d81ea8bSQu Wenruo */ 9204d81ea8bSQu Wenruo if (btrfs_block_can_be_shared(root, eb)) 9214d81ea8bSQu Wenruo upper->checked = 0; 9224d81ea8bSQu Wenruo else 9234d81ea8bSQu Wenruo upper->checked = 1; 9244d81ea8bSQu Wenruo 9254d81ea8bSQu Wenruo /* 9264d81ea8bSQu Wenruo * Add the block to pending list if we need to check its 9274d81ea8bSQu Wenruo * backrefs, we only do this once while walking up a 9284d81ea8bSQu Wenruo * tree as we will catch anything else later on. 9294d81ea8bSQu Wenruo */ 9304d81ea8bSQu Wenruo if (!upper->checked && need_check) { 9314d81ea8bSQu Wenruo need_check = false; 9324d81ea8bSQu Wenruo list_add_tail(&edge->list[UPPER], 9334d81ea8bSQu Wenruo &cache->pending_edge); 9344d81ea8bSQu Wenruo } else { 9354d81ea8bSQu Wenruo if (upper->checked) 9364d81ea8bSQu Wenruo need_check = true; 9374d81ea8bSQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 9384d81ea8bSQu Wenruo } 9394d81ea8bSQu Wenruo } else { 9404d81ea8bSQu Wenruo upper = rb_entry(rb_node, struct backref_node, rb_node); 9414d81ea8bSQu Wenruo ASSERT(upper->checked); 9424d81ea8bSQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 9434d81ea8bSQu Wenruo if (!upper->owner) 9444d81ea8bSQu Wenruo upper->owner = btrfs_header_owner(eb); 9454d81ea8bSQu Wenruo } 9462a979612SQu Wenruo link_backref_edge(edge, lower, upper, LINK_LOWER); 9474d81ea8bSQu Wenruo 9484d81ea8bSQu Wenruo if (rb_node) { 9494d81ea8bSQu Wenruo btrfs_put_root(root); 9504d81ea8bSQu Wenruo break; 9514d81ea8bSQu Wenruo } 9524d81ea8bSQu Wenruo lower = upper; 9534d81ea8bSQu Wenruo upper = NULL; 9544d81ea8bSQu Wenruo } 9554d81ea8bSQu Wenruo out: 9564d81ea8bSQu Wenruo btrfs_release_path(path); 9574d81ea8bSQu Wenruo return ret; 9584d81ea8bSQu Wenruo } 9594d81ea8bSQu Wenruo 960e7d571c7SQu Wenruo static int handle_one_tree_block(struct backref_cache *cache, 961e7d571c7SQu Wenruo struct btrfs_path *path, 962e7d571c7SQu Wenruo struct btrfs_backref_iter *iter, 9635d4f98a2SYan Zheng struct btrfs_key *node_key, 964e7d571c7SQu Wenruo struct backref_node *cur) 9655d4f98a2SYan Zheng { 966e7d571c7SQu Wenruo struct btrfs_fs_info *fs_info = cache->fs_info; 9675d4f98a2SYan Zheng struct backref_edge *edge; 968e7d571c7SQu Wenruo struct backref_node *exist; 9695d4f98a2SYan Zheng int ret; 9705d4f98a2SYan Zheng 97171f572a9SQu Wenruo ret = btrfs_backref_iter_start(iter, cur->bytenr); 972e7d571c7SQu Wenruo if (ret < 0) 973e7d571c7SQu Wenruo return ret; 97471f572a9SQu Wenruo /* 97571f572a9SQu Wenruo * We skip the first btrfs_tree_block_info, as we don't use the key 97671f572a9SQu Wenruo * stored in it, but fetch it from the tree block 97771f572a9SQu Wenruo */ 97871f572a9SQu Wenruo if (btrfs_backref_has_tree_block_info(iter)) { 97971f572a9SQu Wenruo ret = btrfs_backref_iter_next(iter); 980e7d571c7SQu Wenruo if (ret < 0) 98171f572a9SQu Wenruo goto out; 98271f572a9SQu Wenruo /* No extra backref? This means the tree block is corrupted */ 98371f572a9SQu Wenruo if (ret > 0) { 984e7d571c7SQu Wenruo ret = -EUCLEAN; 98571f572a9SQu Wenruo goto out; 98671f572a9SQu Wenruo } 98771f572a9SQu Wenruo } 9885d4f98a2SYan Zheng WARN_ON(cur->checked); 9895d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 9905d4f98a2SYan Zheng /* 99170f23fd6SJustin P. Mattock * the backref was added previously when processing 9925d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 9935d4f98a2SYan Zheng */ 99475bfb9afSJosef Bacik ASSERT(list_is_singular(&cur->upper)); 9955d4f98a2SYan Zheng edge = list_entry(cur->upper.next, struct backref_edge, 9965d4f98a2SYan Zheng list[LOWER]); 99775bfb9afSJosef Bacik ASSERT(list_empty(&edge->list[UPPER])); 9985d4f98a2SYan Zheng exist = edge->node[UPPER]; 9995d4f98a2SYan Zheng /* 10005d4f98a2SYan Zheng * add the upper level block to pending list if we need 10015d4f98a2SYan Zheng * check its backrefs 10025d4f98a2SYan Zheng */ 10035d4f98a2SYan Zheng if (!exist->checked) 100484780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 10055d4f98a2SYan Zheng } else { 10065d4f98a2SYan Zheng exist = NULL; 10075d4f98a2SYan Zheng } 10085d4f98a2SYan Zheng 100971f572a9SQu Wenruo for (; ret == 0; ret = btrfs_backref_iter_next(iter)) { 101071f572a9SQu Wenruo struct extent_buffer *eb; 101171f572a9SQu Wenruo struct btrfs_key key; 10123de28d57SLiu Bo int type; 101371f572a9SQu Wenruo 101471f572a9SQu Wenruo cond_resched(); 101571f572a9SQu Wenruo eb = btrfs_backref_get_eb(iter); 101671f572a9SQu Wenruo 101771f572a9SQu Wenruo key.objectid = iter->bytenr; 101871f572a9SQu Wenruo if (btrfs_backref_iter_is_inline_ref(iter)) { 101971f572a9SQu Wenruo struct btrfs_extent_inline_ref *iref; 102071f572a9SQu Wenruo 102171f572a9SQu Wenruo /* update key for inline back ref */ 102271f572a9SQu Wenruo iref = (struct btrfs_extent_inline_ref *) 102371f572a9SQu Wenruo ((unsigned long)iter->cur_ptr); 10243de28d57SLiu Bo type = btrfs_get_extent_inline_ref_type(eb, iref, 10253de28d57SLiu Bo BTRFS_REF_TYPE_BLOCK); 10263de28d57SLiu Bo if (type == BTRFS_REF_TYPE_INVALID) { 1027e7d571c7SQu Wenruo ret = -EUCLEAN; 10283de28d57SLiu Bo goto out; 10293de28d57SLiu Bo } 10303de28d57SLiu Bo key.type = type; 10315d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 103271f572a9SQu Wenruo } else { 103371f572a9SQu Wenruo key.type = iter->cur_key.type; 103471f572a9SQu Wenruo key.offset = iter->cur_key.offset; 10355d4f98a2SYan Zheng } 10365d4f98a2SYan Zheng 1037fa6ac715SQu Wenruo /* 1038fa6ac715SQu Wenruo * Parent node found and matches current inline ref, no need to 1039fa6ac715SQu Wenruo * rebuild this node for this inline ref. 1040fa6ac715SQu Wenruo */ 10415d4f98a2SYan Zheng if (exist && 10425d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 10435d4f98a2SYan Zheng exist->owner == key.offset) || 10445d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 10455d4f98a2SYan Zheng exist->bytenr == key.offset))) { 10465d4f98a2SYan Zheng exist = NULL; 104771f572a9SQu Wenruo continue; 10485d4f98a2SYan Zheng } 10495d4f98a2SYan Zheng 1050fa6ac715SQu Wenruo /* SHARED_BLOCK_REF means key.offset is the parent bytenr */ 10515d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 10524007ea87SQu Wenruo ret = handle_direct_tree_backref(cache, &key, cur); 1053e7d571c7SQu Wenruo if (ret < 0) 10542433bea5SQu Wenruo goto out; 105571f572a9SQu Wenruo continue; 10566d8ff4e4SDavid Sterba } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) { 1057e7d571c7SQu Wenruo ret = -EINVAL; 1058e7d571c7SQu Wenruo btrfs_print_v0_err(fs_info); 1059e7d571c7SQu Wenruo btrfs_handle_fs_error(fs_info, ret, NULL); 1060ba3c2b19SNikolay Borisov goto out; 10615d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 106271f572a9SQu Wenruo continue; 10635d4f98a2SYan Zheng } 10645d4f98a2SYan Zheng 1065fa6ac715SQu Wenruo /* 1066fa6ac715SQu Wenruo * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset 1067fa6ac715SQu Wenruo * means the root objectid. We need to search the tree to get 1068fa6ac715SQu Wenruo * its parent bytenr. 1069fa6ac715SQu Wenruo */ 10704d81ea8bSQu Wenruo ret = handle_indirect_tree_backref(cache, path, &key, node_key, 10714d81ea8bSQu Wenruo cur); 1072e7d571c7SQu Wenruo if (ret < 0) 107371f572a9SQu Wenruo goto out; 10745d4f98a2SYan Zheng } 107571f572a9SQu Wenruo ret = 0; 10765d4f98a2SYan Zheng cur->checked = 1; 10775d4f98a2SYan Zheng WARN_ON(exist); 1078e7d571c7SQu Wenruo out: 1079e7d571c7SQu Wenruo btrfs_backref_iter_release(iter); 1080e7d571c7SQu Wenruo return ret; 1081e7d571c7SQu Wenruo } 10825d4f98a2SYan Zheng 1083e7d571c7SQu Wenruo /* 1084e7d571c7SQu Wenruo * Build backref tree for a given tree block. Root of the backref tree 1085e7d571c7SQu Wenruo * corresponds the tree block, leaves of the backref tree correspond roots of 1086e7d571c7SQu Wenruo * b-trees that reference the tree block. 1087e7d571c7SQu Wenruo * 1088e7d571c7SQu Wenruo * The basic idea of this function is check backrefs of a given block to find 1089e7d571c7SQu Wenruo * upper level blocks that reference the block, and then check backrefs of 1090e7d571c7SQu Wenruo * these upper level blocks recursively. The recursion stops when tree root is 1091e7d571c7SQu Wenruo * reached or backrefs for the block is cached. 1092e7d571c7SQu Wenruo * 1093e7d571c7SQu Wenruo * NOTE: if we find that backrefs for a block are cached, we know backrefs for 1094e7d571c7SQu Wenruo * all upper level blocks that directly/indirectly reference the block are also 1095e7d571c7SQu Wenruo * cached. 1096e7d571c7SQu Wenruo */ 1097e7d571c7SQu Wenruo static noinline_for_stack struct backref_node *build_backref_tree( 1098e7d571c7SQu Wenruo struct reloc_control *rc, struct btrfs_key *node_key, 1099e7d571c7SQu Wenruo int level, u64 bytenr) 1100e7d571c7SQu Wenruo { 1101e7d571c7SQu Wenruo struct btrfs_backref_iter *iter; 1102e7d571c7SQu Wenruo struct backref_cache *cache = &rc->backref_cache; 1103e7d571c7SQu Wenruo /* For searching parent of TREE_BLOCK_REF */ 1104e7d571c7SQu Wenruo struct btrfs_path *path; 1105e7d571c7SQu Wenruo struct backref_node *cur; 1106e7d571c7SQu Wenruo struct backref_node *upper; 1107e7d571c7SQu Wenruo struct backref_node *lower; 1108e7d571c7SQu Wenruo struct backref_node *node = NULL; 1109e7d571c7SQu Wenruo struct backref_edge *edge; 1110e7d571c7SQu Wenruo struct rb_node *rb_node; 1111e7d571c7SQu Wenruo int cowonly; 1112e7d571c7SQu Wenruo int ret; 1113e7d571c7SQu Wenruo int err = 0; 1114e7d571c7SQu Wenruo 1115e7d571c7SQu Wenruo iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS); 1116e7d571c7SQu Wenruo if (!iter) 1117e7d571c7SQu Wenruo return ERR_PTR(-ENOMEM); 1118e7d571c7SQu Wenruo path = btrfs_alloc_path(); 1119e7d571c7SQu Wenruo if (!path) { 1120e7d571c7SQu Wenruo err = -ENOMEM; 1121e7d571c7SQu Wenruo goto out; 1122e7d571c7SQu Wenruo } 1123e7d571c7SQu Wenruo 1124e7d571c7SQu Wenruo node = alloc_backref_node(cache, bytenr, level); 1125e7d571c7SQu Wenruo if (!node) { 1126e7d571c7SQu Wenruo err = -ENOMEM; 1127e7d571c7SQu Wenruo goto out; 1128e7d571c7SQu Wenruo } 1129e7d571c7SQu Wenruo 1130e7d571c7SQu Wenruo node->lowest = 1; 1131e7d571c7SQu Wenruo cur = node; 1132e7d571c7SQu Wenruo 1133e7d571c7SQu Wenruo /* Breadth-first search to build backref cache */ 1134e7d571c7SQu Wenruo do { 1135e7d571c7SQu Wenruo ret = handle_one_tree_block(cache, path, iter, node_key, cur); 1136e7d571c7SQu Wenruo if (ret < 0) { 1137e7d571c7SQu Wenruo err = ret; 1138e7d571c7SQu Wenruo goto out; 1139e7d571c7SQu Wenruo } 1140e7d571c7SQu Wenruo edge = list_first_entry_or_null(&cache->pending_edge, 114184780289SQu Wenruo struct backref_edge, list[UPPER]); 1142e7d571c7SQu Wenruo /* 1143e7d571c7SQu Wenruo * The pending list isn't empty, take the first block to 1144e7d571c7SQu Wenruo * process 1145e7d571c7SQu Wenruo */ 1146e7d571c7SQu Wenruo if (edge) { 11475d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 11485d4f98a2SYan Zheng cur = edge->node[UPPER]; 11495d4f98a2SYan Zheng } 1150e7d571c7SQu Wenruo } while (edge); 11515d4f98a2SYan Zheng 11525d4f98a2SYan Zheng /* 11535d4f98a2SYan Zheng * everything goes well, connect backref nodes and insert backref nodes 11545d4f98a2SYan Zheng * into the cache. 11555d4f98a2SYan Zheng */ 115675bfb9afSJosef Bacik ASSERT(node->checked); 11573fd0a558SYan, Zheng cowonly = node->cowonly; 11583fd0a558SYan, Zheng if (!cowonly) { 11593fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, 11603fd0a558SYan, Zheng &node->rb_node); 116143c04fb1SJeff Mahoney if (rb_node) 116243c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 11633fd0a558SYan, Zheng list_add_tail(&node->lower, &cache->leaves); 11643fd0a558SYan, Zheng } 11655d4f98a2SYan Zheng 11665d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) 116784780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 11685d4f98a2SYan Zheng 116984780289SQu Wenruo while (!list_empty(&cache->pending_edge)) { 117084780289SQu Wenruo edge = list_first_entry(&cache->pending_edge, 117184780289SQu Wenruo struct backref_edge, list[UPPER]); 11725d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 11735d4f98a2SYan Zheng upper = edge->node[UPPER]; 11743fd0a558SYan, Zheng if (upper->detached) { 11753fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 11763fd0a558SYan, Zheng lower = edge->node[LOWER]; 11773fd0a558SYan, Zheng free_backref_edge(cache, edge); 11783fd0a558SYan, Zheng if (list_empty(&lower->upper)) 117984780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 11803fd0a558SYan, Zheng continue; 11813fd0a558SYan, Zheng } 11825d4f98a2SYan Zheng 11835d4f98a2SYan Zheng if (!RB_EMPTY_NODE(&upper->rb_node)) { 11845d4f98a2SYan Zheng if (upper->lowest) { 11855d4f98a2SYan Zheng list_del_init(&upper->lower); 11865d4f98a2SYan Zheng upper->lowest = 0; 11875d4f98a2SYan Zheng } 11885d4f98a2SYan Zheng 11895d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 11905d4f98a2SYan Zheng continue; 11915d4f98a2SYan Zheng } 11925d4f98a2SYan Zheng 119375bfb9afSJosef Bacik if (!upper->checked) { 119475bfb9afSJosef Bacik /* 119575bfb9afSJosef Bacik * Still want to blow up for developers since this is a 119675bfb9afSJosef Bacik * logic bug. 119775bfb9afSJosef Bacik */ 119875bfb9afSJosef Bacik ASSERT(0); 119975bfb9afSJosef Bacik err = -EINVAL; 120075bfb9afSJosef Bacik goto out; 120175bfb9afSJosef Bacik } 120275bfb9afSJosef Bacik if (cowonly != upper->cowonly) { 120375bfb9afSJosef Bacik ASSERT(0); 120475bfb9afSJosef Bacik err = -EINVAL; 120575bfb9afSJosef Bacik goto out; 120675bfb9afSJosef Bacik } 120775bfb9afSJosef Bacik 12083fd0a558SYan, Zheng if (!cowonly) { 12095d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, upper->bytenr, 12105d4f98a2SYan Zheng &upper->rb_node); 121143c04fb1SJeff Mahoney if (rb_node) 121243c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, 121343c04fb1SJeff Mahoney upper->bytenr); 12143fd0a558SYan, Zheng } 12155d4f98a2SYan Zheng 12165d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 12175d4f98a2SYan Zheng 12185d4f98a2SYan Zheng list_for_each_entry(edge, &upper->upper, list[LOWER]) 121984780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 12205d4f98a2SYan Zheng } 12213fd0a558SYan, Zheng /* 12223fd0a558SYan, Zheng * process useless backref nodes. backref nodes for tree leaves 12233fd0a558SYan, Zheng * are deleted from the cache. backref nodes for upper level 12243fd0a558SYan, Zheng * tree blocks are left in the cache to avoid unnecessary backref 12253fd0a558SYan, Zheng * lookup. 12263fd0a558SYan, Zheng */ 122784780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 122884780289SQu Wenruo upper = list_first_entry(&cache->useless_node, 122984780289SQu Wenruo struct backref_node, list); 12303fd0a558SYan, Zheng list_del_init(&upper->list); 123175bfb9afSJosef Bacik ASSERT(list_empty(&upper->upper)); 12323fd0a558SYan, Zheng if (upper == node) 12333fd0a558SYan, Zheng node = NULL; 12343fd0a558SYan, Zheng if (upper->lowest) { 12353fd0a558SYan, Zheng list_del_init(&upper->lower); 12363fd0a558SYan, Zheng upper->lowest = 0; 12373fd0a558SYan, Zheng } 12383fd0a558SYan, Zheng while (!list_empty(&upper->lower)) { 123984780289SQu Wenruo edge = list_first_entry(&upper->lower, 12403fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 12413fd0a558SYan, Zheng list_del(&edge->list[UPPER]); 12423fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 12433fd0a558SYan, Zheng lower = edge->node[LOWER]; 12443fd0a558SYan, Zheng free_backref_edge(cache, edge); 12453fd0a558SYan, Zheng 12463fd0a558SYan, Zheng if (list_empty(&lower->upper)) 124784780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 12483fd0a558SYan, Zheng } 12499569cc20SQu Wenruo mark_block_processed(rc, upper); 12503fd0a558SYan, Zheng if (upper->level > 0) { 12513fd0a558SYan, Zheng list_add(&upper->list, &cache->detached); 12523fd0a558SYan, Zheng upper->detached = 1; 12533fd0a558SYan, Zheng } else { 12543fd0a558SYan, Zheng rb_erase(&upper->rb_node, &cache->rb_root); 12553fd0a558SYan, Zheng free_backref_node(cache, upper); 12563fd0a558SYan, Zheng } 12573fd0a558SYan, Zheng } 12585d4f98a2SYan Zheng out: 125971f572a9SQu Wenruo btrfs_backref_iter_free(iter); 126071f572a9SQu Wenruo btrfs_free_path(path); 12615d4f98a2SYan Zheng if (err) { 126284780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 126384780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 126475bfb9afSJosef Bacik struct backref_node, list); 126575bfb9afSJosef Bacik list_del_init(&lower->list); 12663fd0a558SYan, Zheng } 126784780289SQu Wenruo while (!list_empty(&cache->pending_edge)) { 126884780289SQu Wenruo edge = list_first_entry(&cache->pending_edge, 126984780289SQu Wenruo struct backref_edge, list[UPPER]); 127075bfb9afSJosef Bacik list_del(&edge->list[UPPER]); 12713fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 127275bfb9afSJosef Bacik lower = edge->node[LOWER]; 12735d4f98a2SYan Zheng upper = edge->node[UPPER]; 12743fd0a558SYan, Zheng free_backref_edge(cache, edge); 127575bfb9afSJosef Bacik 127675bfb9afSJosef Bacik /* 127775bfb9afSJosef Bacik * Lower is no longer linked to any upper backref nodes 127875bfb9afSJosef Bacik * and isn't in the cache, we can free it ourselves. 127975bfb9afSJosef Bacik */ 128075bfb9afSJosef Bacik if (list_empty(&lower->upper) && 128175bfb9afSJosef Bacik RB_EMPTY_NODE(&lower->rb_node)) 128284780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 128375bfb9afSJosef Bacik 128475bfb9afSJosef Bacik if (!RB_EMPTY_NODE(&upper->rb_node)) 128575bfb9afSJosef Bacik continue; 128675bfb9afSJosef Bacik 128701327610SNicholas D Steeves /* Add this guy's upper edges to the list to process */ 128875bfb9afSJosef Bacik list_for_each_entry(edge, &upper->upper, list[LOWER]) 128984780289SQu Wenruo list_add_tail(&edge->list[UPPER], 129084780289SQu Wenruo &cache->pending_edge); 129175bfb9afSJosef Bacik if (list_empty(&upper->upper)) 129284780289SQu Wenruo list_add(&upper->list, &cache->useless_node); 129375bfb9afSJosef Bacik } 129475bfb9afSJosef Bacik 129584780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 129684780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 129775bfb9afSJosef Bacik struct backref_node, list); 129875bfb9afSJosef Bacik list_del_init(&lower->list); 12990fd8c3daSLiu Bo if (lower == node) 13000fd8c3daSLiu Bo node = NULL; 130175bfb9afSJosef Bacik free_backref_node(cache, lower); 13025d4f98a2SYan Zheng } 13030fd8c3daSLiu Bo 13048e19c973SJosef Bacik remove_backref_node(cache, node); 130584780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 130684780289SQu Wenruo list_empty(&cache->pending_edge)); 13075d4f98a2SYan Zheng return ERR_PTR(err); 13085d4f98a2SYan Zheng } 130975bfb9afSJosef Bacik ASSERT(!node || !node->detached); 131084780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 131184780289SQu Wenruo list_empty(&cache->pending_edge)); 13125d4f98a2SYan Zheng return node; 13135d4f98a2SYan Zheng } 13145d4f98a2SYan Zheng 13155d4f98a2SYan Zheng /* 13163fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 13173fd0a558SYan, Zheng * the backref node is created by cloning backref node that 13183fd0a558SYan, Zheng * corresponds to root of source tree 13193fd0a558SYan, Zheng */ 13203fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 13213fd0a558SYan, Zheng struct reloc_control *rc, 13223fd0a558SYan, Zheng struct btrfs_root *src, 13233fd0a558SYan, Zheng struct btrfs_root *dest) 13243fd0a558SYan, Zheng { 13253fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 13263fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 13273fd0a558SYan, Zheng struct backref_node *node = NULL; 13283fd0a558SYan, Zheng struct backref_node *new_node; 13293fd0a558SYan, Zheng struct backref_edge *edge; 13303fd0a558SYan, Zheng struct backref_edge *new_edge; 13313fd0a558SYan, Zheng struct rb_node *rb_node; 13323fd0a558SYan, Zheng 13333fd0a558SYan, Zheng if (cache->last_trans > 0) 13343fd0a558SYan, Zheng update_backref_cache(trans, cache); 13353fd0a558SYan, Zheng 13363fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, src->commit_root->start); 13373fd0a558SYan, Zheng if (rb_node) { 13383fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, rb_node); 13393fd0a558SYan, Zheng if (node->detached) 13403fd0a558SYan, Zheng node = NULL; 13413fd0a558SYan, Zheng else 13423fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 13433fd0a558SYan, Zheng } 13443fd0a558SYan, Zheng 13453fd0a558SYan, Zheng if (!node) { 13463fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, 13473fd0a558SYan, Zheng reloc_root->commit_root->start); 13483fd0a558SYan, Zheng if (rb_node) { 13493fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, 13503fd0a558SYan, Zheng rb_node); 13513fd0a558SYan, Zheng BUG_ON(node->detached); 13523fd0a558SYan, Zheng } 13533fd0a558SYan, Zheng } 13543fd0a558SYan, Zheng 13553fd0a558SYan, Zheng if (!node) 13563fd0a558SYan, Zheng return 0; 13573fd0a558SYan, Zheng 13580304f2d8SQu Wenruo new_node = alloc_backref_node(cache, dest->node->start, node->level); 13593fd0a558SYan, Zheng if (!new_node) 13603fd0a558SYan, Zheng return -ENOMEM; 13613fd0a558SYan, Zheng 13623fd0a558SYan, Zheng new_node->lowest = node->lowest; 13636848ad64SYan, Zheng new_node->checked = 1; 136400246528SJosef Bacik new_node->root = btrfs_grab_root(dest); 13650b530bc5SJosef Bacik ASSERT(new_node->root); 13663fd0a558SYan, Zheng 13673fd0a558SYan, Zheng if (!node->lowest) { 13683fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 13693fd0a558SYan, Zheng new_edge = alloc_backref_edge(cache); 13703fd0a558SYan, Zheng if (!new_edge) 13713fd0a558SYan, Zheng goto fail; 13723fd0a558SYan, Zheng 13732a979612SQu Wenruo link_backref_edge(new_edge, edge->node[LOWER], new_node, 13742a979612SQu Wenruo LINK_UPPER); 13753fd0a558SYan, Zheng } 137676b9e23dSMiao Xie } else { 137776b9e23dSMiao Xie list_add_tail(&new_node->lower, &cache->leaves); 13783fd0a558SYan, Zheng } 13793fd0a558SYan, Zheng 13803fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, new_node->bytenr, 13813fd0a558SYan, Zheng &new_node->rb_node); 138243c04fb1SJeff Mahoney if (rb_node) 138343c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, new_node->bytenr); 13843fd0a558SYan, Zheng 13853fd0a558SYan, Zheng if (!new_node->lowest) { 13863fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 13873fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 13883fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 13893fd0a558SYan, Zheng } 13903fd0a558SYan, Zheng } 13913fd0a558SYan, Zheng return 0; 13923fd0a558SYan, Zheng fail: 13933fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 13943fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 13953fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 13963fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 13973fd0a558SYan, Zheng free_backref_edge(cache, new_edge); 13983fd0a558SYan, Zheng } 13993fd0a558SYan, Zheng free_backref_node(cache, new_node); 14003fd0a558SYan, Zheng return -ENOMEM; 14013fd0a558SYan, Zheng } 14023fd0a558SYan, Zheng 14033fd0a558SYan, Zheng /* 14045d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 14055d4f98a2SYan Zheng */ 1406ffd7b339SJeff Mahoney static int __must_check __add_reloc_root(struct btrfs_root *root) 14075d4f98a2SYan Zheng { 14080b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14095d4f98a2SYan Zheng struct rb_node *rb_node; 14105d4f98a2SYan Zheng struct mapping_node *node; 14110b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 14125d4f98a2SYan Zheng 14135d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 1414ffd7b339SJeff Mahoney if (!node) 1415ffd7b339SJeff Mahoney return -ENOMEM; 14165d4f98a2SYan Zheng 1417ea287ab1SJosef Bacik node->bytenr = root->commit_root->start; 14185d4f98a2SYan Zheng node->data = root; 14195d4f98a2SYan Zheng 14205d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 14215d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 14225d4f98a2SYan Zheng node->bytenr, &node->rb_node); 14235d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 1424ffd7b339SJeff Mahoney if (rb_node) { 14250b246afaSJeff Mahoney btrfs_panic(fs_info, -EEXIST, 14265d163e0eSJeff Mahoney "Duplicate root found for start=%llu while inserting into relocation tree", 14275d163e0eSJeff Mahoney node->bytenr); 1428ffd7b339SJeff Mahoney } 14295d4f98a2SYan Zheng 14305d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 14315d4f98a2SYan Zheng return 0; 14325d4f98a2SYan Zheng } 14335d4f98a2SYan Zheng 14345d4f98a2SYan Zheng /* 1435c974c464SWang Shilong * helper to delete the 'address of tree root -> reloc tree' 14365d4f98a2SYan Zheng * mapping 14375d4f98a2SYan Zheng */ 1438c974c464SWang Shilong static void __del_reloc_root(struct btrfs_root *root) 14395d4f98a2SYan Zheng { 14400b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14415d4f98a2SYan Zheng struct rb_node *rb_node; 14425d4f98a2SYan Zheng struct mapping_node *node = NULL; 14430b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1444f44deb74SJosef Bacik bool put_ref = false; 14455d4f98a2SYan Zheng 144665c6e82bSQu Wenruo if (rc && root->node) { 14475d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 14485d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, 1449ea287ab1SJosef Bacik root->commit_root->start); 1450c974c464SWang Shilong if (rb_node) { 1451c974c464SWang Shilong node = rb_entry(rb_node, struct mapping_node, rb_node); 1452c974c464SWang Shilong rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 1453ea287ab1SJosef Bacik RB_CLEAR_NODE(&node->rb_node); 1454c974c464SWang Shilong } 1455c974c464SWang Shilong spin_unlock(&rc->reloc_root_tree.lock); 1456c974c464SWang Shilong if (!node) 1457c974c464SWang Shilong return; 1458c974c464SWang Shilong BUG_ON((struct btrfs_root *)node->data != root); 1459389305b2SQu Wenruo } 1460c974c464SWang Shilong 1461f44deb74SJosef Bacik /* 1462f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 1463f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 1464f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 1465f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 1466f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 1467f44deb74SJosef Bacik * messing with the list is in charge of the reference. 1468f44deb74SJosef Bacik */ 14690b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 1470f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 1471f44deb74SJosef Bacik put_ref = true; 1472c974c464SWang Shilong list_del_init(&root->root_list); 1473f44deb74SJosef Bacik } 14740b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 1475f44deb74SJosef Bacik if (put_ref) 1476f44deb74SJosef Bacik btrfs_put_root(root); 1477c974c464SWang Shilong kfree(node); 1478c974c464SWang Shilong } 1479c974c464SWang Shilong 1480c974c464SWang Shilong /* 1481c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 1482c974c464SWang Shilong * mapping 1483c974c464SWang Shilong */ 1484ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 1485c974c464SWang Shilong { 14860b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 1487c974c464SWang Shilong struct rb_node *rb_node; 1488c974c464SWang Shilong struct mapping_node *node = NULL; 14890b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1490c974c464SWang Shilong 1491c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 1492c974c464SWang Shilong rb_node = tree_search(&rc->reloc_root_tree.rb_root, 1493ea287ab1SJosef Bacik root->commit_root->start); 14945d4f98a2SYan Zheng if (rb_node) { 14955d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 14965d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 14975d4f98a2SYan Zheng } 14985d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 14995d4f98a2SYan Zheng 15008f71f3e0SLiu Bo if (!node) 15018f71f3e0SLiu Bo return 0; 15025d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 15035d4f98a2SYan Zheng 15045d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1505ea287ab1SJosef Bacik node->bytenr = root->node->start; 15065d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 15075d4f98a2SYan Zheng node->bytenr, &node->rb_node); 15085d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 150943c04fb1SJeff Mahoney if (rb_node) 151043c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 15115d4f98a2SYan Zheng return 0; 15125d4f98a2SYan Zheng } 15135d4f98a2SYan Zheng 15143fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 15153fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 15165d4f98a2SYan Zheng { 15170b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15185d4f98a2SYan Zheng struct btrfs_root *reloc_root; 15195d4f98a2SYan Zheng struct extent_buffer *eb; 15205d4f98a2SYan Zheng struct btrfs_root_item *root_item; 15215d4f98a2SYan Zheng struct btrfs_key root_key; 15225d4f98a2SYan Zheng int ret; 15235d4f98a2SYan Zheng 15245d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 15255d4f98a2SYan Zheng BUG_ON(!root_item); 15265d4f98a2SYan Zheng 15275d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 15285d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 15293fd0a558SYan, Zheng root_key.offset = objectid; 15305d4f98a2SYan Zheng 15313fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 1532054570a1SFilipe Manana u64 commit_root_gen; 1533054570a1SFilipe Manana 15343fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 15355d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 15365d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 15375d4f98a2SYan Zheng BUG_ON(ret); 1538054570a1SFilipe Manana /* 1539054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 1540054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 1541054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 1542054570a1SFilipe Manana * either inside the critical section of a transaction commit 1543054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 1544054570a1SFilipe Manana * it's created before the transaction commit is started. 1545054570a1SFilipe Manana */ 1546054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 1547054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 15483fd0a558SYan, Zheng } else { 15493fd0a558SYan, Zheng /* 15503fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 15513fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 15523fd0a558SYan, Zheng * modified after it was created have RELOC flag 15533fd0a558SYan, Zheng * set in their headers. so it's OK to not update 15543fd0a558SYan, Zheng * the 'last_snapshot'. 15553fd0a558SYan, Zheng */ 15563fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 15573fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 15583fd0a558SYan, Zheng BUG_ON(ret); 15593fd0a558SYan, Zheng } 15603fd0a558SYan, Zheng 15615d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 15625d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 15635d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 15645d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 15653fd0a558SYan, Zheng 15663fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 15673fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 15683fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 15693fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 15705d4f98a2SYan Zheng root_item->drop_level = 0; 15713fd0a558SYan, Zheng } 15725d4f98a2SYan Zheng 15735d4f98a2SYan Zheng btrfs_tree_unlock(eb); 15745d4f98a2SYan Zheng free_extent_buffer(eb); 15755d4f98a2SYan Zheng 15760b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 15775d4f98a2SYan Zheng &root_key, root_item); 15785d4f98a2SYan Zheng BUG_ON(ret); 15795d4f98a2SYan Zheng kfree(root_item); 15805d4f98a2SYan Zheng 15813dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 15825d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 15833dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 15845d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 15853fd0a558SYan, Zheng return reloc_root; 15863fd0a558SYan, Zheng } 15873fd0a558SYan, Zheng 15883fd0a558SYan, Zheng /* 15893fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 15903fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 1591f44deb74SJosef Bacik * 1592f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 1593f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 15943fd0a558SYan, Zheng */ 15953fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 15963fd0a558SYan, Zheng struct btrfs_root *root) 15973fd0a558SYan, Zheng { 15980b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15993fd0a558SYan, Zheng struct btrfs_root *reloc_root; 16000b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 160120dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 16023fd0a558SYan, Zheng int clear_rsv = 0; 1603ffd7b339SJeff Mahoney int ret; 16043fd0a558SYan, Zheng 1605aec7db3bSJosef Bacik if (!rc) 16062abc726aSJosef Bacik return 0; 16072abc726aSJosef Bacik 16081fac4a54SQu Wenruo /* 16091fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 16101fac4a54SQu Wenruo * create/update the dead reloc tree 16111fac4a54SQu Wenruo */ 16126282675eSQu Wenruo if (reloc_root_is_dead(root)) 16131fac4a54SQu Wenruo return 0; 16141fac4a54SQu Wenruo 1615aec7db3bSJosef Bacik /* 1616aec7db3bSJosef Bacik * This is subtle but important. We do not do 1617aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 1618aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 1619aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 1620aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 1621aec7db3bSJosef Bacik * in. 1622aec7db3bSJosef Bacik */ 16233fd0a558SYan, Zheng if (root->reloc_root) { 16243fd0a558SYan, Zheng reloc_root = root->reloc_root; 16253fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 16263fd0a558SYan, Zheng return 0; 16273fd0a558SYan, Zheng } 16283fd0a558SYan, Zheng 1629aec7db3bSJosef Bacik /* 1630aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 1631aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 1632aec7db3bSJosef Bacik */ 1633aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 1634aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 1635aec7db3bSJosef Bacik return 0; 1636aec7db3bSJosef Bacik 163720dd2cbfSMiao Xie if (!trans->reloc_reserved) { 163820dd2cbfSMiao Xie rsv = trans->block_rsv; 16393fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 16403fd0a558SYan, Zheng clear_rsv = 1; 16413fd0a558SYan, Zheng } 16423fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 16433fd0a558SYan, Zheng if (clear_rsv) 164420dd2cbfSMiao Xie trans->block_rsv = rsv; 16455d4f98a2SYan Zheng 1646ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 1647ffd7b339SJeff Mahoney BUG_ON(ret < 0); 1648f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 16495d4f98a2SYan Zheng return 0; 16505d4f98a2SYan Zheng } 16515d4f98a2SYan Zheng 16525d4f98a2SYan Zheng /* 16535d4f98a2SYan Zheng * update root item of reloc tree 16545d4f98a2SYan Zheng */ 16555d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 16565d4f98a2SYan Zheng struct btrfs_root *root) 16575d4f98a2SYan Zheng { 16580b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 16595d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16605d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16615d4f98a2SYan Zheng int ret; 16625d4f98a2SYan Zheng 16636282675eSQu Wenruo if (!have_reloc_root(root)) 16647585717fSChris Mason goto out; 16655d4f98a2SYan Zheng 16665d4f98a2SYan Zheng reloc_root = root->reloc_root; 16675d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16685d4f98a2SYan Zheng 1669f44deb74SJosef Bacik /* 1670f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 1671f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 1672f44deb74SJosef Bacik * hold it while we update the reloc root. 1673f44deb74SJosef Bacik */ 1674f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 1675f44deb74SJosef Bacik 1676d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 16770b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 16783fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 1679d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 16806282675eSQu Wenruo /* 16816282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 16826282675eSQu Wenruo * have_reloc_root will not touch it from now on. 16836282675eSQu Wenruo */ 16846282675eSQu Wenruo smp_wmb(); 1685c974c464SWang Shilong __del_reloc_root(reloc_root); 16865d4f98a2SYan Zheng } 16875d4f98a2SYan Zheng 16885d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 1689ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 16905d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 16915d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 16925d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 16935d4f98a2SYan Zheng } 16945d4f98a2SYan Zheng 16950b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 16965d4f98a2SYan Zheng &reloc_root->root_key, root_item); 16975d4f98a2SYan Zheng BUG_ON(ret); 1698f44deb74SJosef Bacik btrfs_put_root(reloc_root); 16997585717fSChris Mason out: 17005d4f98a2SYan Zheng return 0; 17015d4f98a2SYan Zheng } 17025d4f98a2SYan Zheng 17035d4f98a2SYan Zheng /* 17045d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 17055d4f98a2SYan Zheng * in a subvolume 17065d4f98a2SYan Zheng */ 17075d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 17085d4f98a2SYan Zheng { 17095d4f98a2SYan Zheng struct rb_node *node; 17105d4f98a2SYan Zheng struct rb_node *prev; 17115d4f98a2SYan Zheng struct btrfs_inode *entry; 17125d4f98a2SYan Zheng struct inode *inode; 17135d4f98a2SYan Zheng 17145d4f98a2SYan Zheng spin_lock(&root->inode_lock); 17155d4f98a2SYan Zheng again: 17165d4f98a2SYan Zheng node = root->inode_tree.rb_node; 17175d4f98a2SYan Zheng prev = NULL; 17185d4f98a2SYan Zheng while (node) { 17195d4f98a2SYan Zheng prev = node; 17205d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 17215d4f98a2SYan Zheng 17224a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 17235d4f98a2SYan Zheng node = node->rb_left; 17244a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 17255d4f98a2SYan Zheng node = node->rb_right; 17265d4f98a2SYan Zheng else 17275d4f98a2SYan Zheng break; 17285d4f98a2SYan Zheng } 17295d4f98a2SYan Zheng if (!node) { 17305d4f98a2SYan Zheng while (prev) { 17315d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 17324a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 17335d4f98a2SYan Zheng node = prev; 17345d4f98a2SYan Zheng break; 17355d4f98a2SYan Zheng } 17365d4f98a2SYan Zheng prev = rb_next(prev); 17375d4f98a2SYan Zheng } 17385d4f98a2SYan Zheng } 17395d4f98a2SYan Zheng while (node) { 17405d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 17415d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 17425d4f98a2SYan Zheng if (inode) { 17435d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 17445d4f98a2SYan Zheng return inode; 17455d4f98a2SYan Zheng } 17465d4f98a2SYan Zheng 17474a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 17485d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 17495d4f98a2SYan Zheng goto again; 17505d4f98a2SYan Zheng 17515d4f98a2SYan Zheng node = rb_next(node); 17525d4f98a2SYan Zheng } 17535d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 17545d4f98a2SYan Zheng return NULL; 17555d4f98a2SYan Zheng } 17565d4f98a2SYan Zheng 17575d4f98a2SYan Zheng /* 17585d4f98a2SYan Zheng * get new location of data 17595d4f98a2SYan Zheng */ 17605d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 17615d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 17625d4f98a2SYan Zheng { 17635d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 17645d4f98a2SYan Zheng struct btrfs_path *path; 17655d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 17665d4f98a2SYan Zheng struct extent_buffer *leaf; 17675d4f98a2SYan Zheng int ret; 17685d4f98a2SYan Zheng 17695d4f98a2SYan Zheng path = btrfs_alloc_path(); 17705d4f98a2SYan Zheng if (!path) 17715d4f98a2SYan Zheng return -ENOMEM; 17725d4f98a2SYan Zheng 17735d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 1774f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 1775f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 17765d4f98a2SYan Zheng if (ret < 0) 17775d4f98a2SYan Zheng goto out; 17785d4f98a2SYan Zheng if (ret > 0) { 17795d4f98a2SYan Zheng ret = -ENOENT; 17805d4f98a2SYan Zheng goto out; 17815d4f98a2SYan Zheng } 17825d4f98a2SYan Zheng 17835d4f98a2SYan Zheng leaf = path->nodes[0]; 17845d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 17855d4f98a2SYan Zheng struct btrfs_file_extent_item); 17865d4f98a2SYan Zheng 17875d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 17885d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 17895d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 17905d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 17915d4f98a2SYan Zheng 17925d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 179383d4cfd4SJosef Bacik ret = -EINVAL; 17945d4f98a2SYan Zheng goto out; 17955d4f98a2SYan Zheng } 17965d4f98a2SYan Zheng 17975d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 17985d4f98a2SYan Zheng ret = 0; 17995d4f98a2SYan Zheng out: 18005d4f98a2SYan Zheng btrfs_free_path(path); 18015d4f98a2SYan Zheng return ret; 18025d4f98a2SYan Zheng } 18035d4f98a2SYan Zheng 18045d4f98a2SYan Zheng /* 18055d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 18065d4f98a2SYan Zheng * the new locations. 18075d4f98a2SYan Zheng */ 18083fd0a558SYan, Zheng static noinline_for_stack 18093fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 18105d4f98a2SYan Zheng struct reloc_control *rc, 18115d4f98a2SYan Zheng struct btrfs_root *root, 18123fd0a558SYan, Zheng struct extent_buffer *leaf) 18135d4f98a2SYan Zheng { 18140b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 18155d4f98a2SYan Zheng struct btrfs_key key; 18165d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 18175d4f98a2SYan Zheng struct inode *inode = NULL; 18185d4f98a2SYan Zheng u64 parent; 18195d4f98a2SYan Zheng u64 bytenr; 18203fd0a558SYan, Zheng u64 new_bytenr = 0; 18215d4f98a2SYan Zheng u64 num_bytes; 18225d4f98a2SYan Zheng u64 end; 18235d4f98a2SYan Zheng u32 nritems; 18245d4f98a2SYan Zheng u32 i; 182583d4cfd4SJosef Bacik int ret = 0; 18265d4f98a2SYan Zheng int first = 1; 18275d4f98a2SYan Zheng int dirty = 0; 18285d4f98a2SYan Zheng 18295d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 18305d4f98a2SYan Zheng return 0; 18315d4f98a2SYan Zheng 18325d4f98a2SYan Zheng /* reloc trees always use full backref */ 18335d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 18345d4f98a2SYan Zheng parent = leaf->start; 18355d4f98a2SYan Zheng else 18365d4f98a2SYan Zheng parent = 0; 18375d4f98a2SYan Zheng 18385d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 18395d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 184082fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 184182fa113fSQu Wenruo 18425d4f98a2SYan Zheng cond_resched(); 18435d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 18445d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 18455d4f98a2SYan Zheng continue; 18465d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 18475d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 18485d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 18495d4f98a2SYan Zheng continue; 18505d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 18515d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 18525d4f98a2SYan Zheng if (bytenr == 0) 18535d4f98a2SYan Zheng continue; 18549569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 18559569cc20SQu Wenruo rc->block_group->length)) 18565d4f98a2SYan Zheng continue; 18575d4f98a2SYan Zheng 18585d4f98a2SYan Zheng /* 18595d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 18605d4f98a2SYan Zheng * to complete and drop the extent cache 18615d4f98a2SYan Zheng */ 18625d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 18635d4f98a2SYan Zheng if (first) { 18645d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 18655d4f98a2SYan Zheng first = 0; 18664a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 18673fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 18685d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 18695d4f98a2SYan Zheng } 18704a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 18715d4f98a2SYan Zheng end = key.offset + 18725d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 18735d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 18740b246afaSJeff Mahoney fs_info->sectorsize)); 18750b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 18765d4f98a2SYan Zheng end--; 18775d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1878d0082371SJeff Mahoney key.offset, end); 18795d4f98a2SYan Zheng if (!ret) 18805d4f98a2SYan Zheng continue; 18815d4f98a2SYan Zheng 1882dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1883dcdbc059SNikolay Borisov key.offset, end, 1); 18845d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1885d0082371SJeff Mahoney key.offset, end); 18865d4f98a2SYan Zheng } 18875d4f98a2SYan Zheng } 18885d4f98a2SYan Zheng 18895d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 18905d4f98a2SYan Zheng bytenr, num_bytes); 189183d4cfd4SJosef Bacik if (ret) { 189283d4cfd4SJosef Bacik /* 189383d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 189483d4cfd4SJosef Bacik * in the file extent yet. 189583d4cfd4SJosef Bacik */ 189683d4cfd4SJosef Bacik break; 18973fd0a558SYan, Zheng } 18985d4f98a2SYan Zheng 18995d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 19005d4f98a2SYan Zheng dirty = 1; 19015d4f98a2SYan Zheng 19025d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 190382fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 190482fa113fSQu Wenruo num_bytes, parent); 190582fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 190682fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1907b06c4bf5SFilipe Manana key.objectid, key.offset); 190882fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 190983d4cfd4SJosef Bacik if (ret) { 191066642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 191183d4cfd4SJosef Bacik break; 191283d4cfd4SJosef Bacik } 19135d4f98a2SYan Zheng 1914ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1915ffd4bb2aSQu Wenruo num_bytes, parent); 1916ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1917ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1918b06c4bf5SFilipe Manana key.objectid, key.offset); 1919ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 192083d4cfd4SJosef Bacik if (ret) { 192166642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 192283d4cfd4SJosef Bacik break; 192383d4cfd4SJosef Bacik } 19245d4f98a2SYan Zheng } 19255d4f98a2SYan Zheng if (dirty) 19265d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 19273fd0a558SYan, Zheng if (inode) 19283fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 192983d4cfd4SJosef Bacik return ret; 19305d4f98a2SYan Zheng } 19315d4f98a2SYan Zheng 19325d4f98a2SYan Zheng static noinline_for_stack 19335d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 19345d4f98a2SYan Zheng struct btrfs_path *path, int level) 19355d4f98a2SYan Zheng { 19365d4f98a2SYan Zheng struct btrfs_disk_key key1; 19375d4f98a2SYan Zheng struct btrfs_disk_key key2; 19385d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 19395d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 19405d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 19415d4f98a2SYan Zheng } 19425d4f98a2SYan Zheng 19435d4f98a2SYan Zheng /* 19445d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 19455d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 19465d4f98a2SYan Zheng * reloc tree was create can be replaced. 19475d4f98a2SYan Zheng * 19485d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 19495d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 19505d4f98a2SYan Zheng * errors, a negative error number is returned. 19515d4f98a2SYan Zheng */ 19523fd0a558SYan, Zheng static noinline_for_stack 19533d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 19545d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 19555d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 19565d4f98a2SYan Zheng int lowest_level, int max_level) 19575d4f98a2SYan Zheng { 19580b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 19595d4f98a2SYan Zheng struct extent_buffer *eb; 19605d4f98a2SYan Zheng struct extent_buffer *parent; 196182fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 19625d4f98a2SYan Zheng struct btrfs_key key; 19635d4f98a2SYan Zheng u64 old_bytenr; 19645d4f98a2SYan Zheng u64 new_bytenr; 19655d4f98a2SYan Zheng u64 old_ptr_gen; 19665d4f98a2SYan Zheng u64 new_ptr_gen; 19675d4f98a2SYan Zheng u64 last_snapshot; 19685d4f98a2SYan Zheng u32 blocksize; 19693fd0a558SYan, Zheng int cow = 0; 19705d4f98a2SYan Zheng int level; 19715d4f98a2SYan Zheng int ret; 19725d4f98a2SYan Zheng int slot; 19735d4f98a2SYan Zheng 19745d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 19755d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 19765d4f98a2SYan Zheng 19775d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 19783fd0a558SYan, Zheng again: 19795d4f98a2SYan Zheng slot = path->slots[lowest_level]; 19805d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 19815d4f98a2SYan Zheng 19825d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 19838bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19845d4f98a2SYan Zheng level = btrfs_header_level(eb); 19855d4f98a2SYan Zheng 19865d4f98a2SYan Zheng if (level < lowest_level) { 19875d4f98a2SYan Zheng btrfs_tree_unlock(eb); 19885d4f98a2SYan Zheng free_extent_buffer(eb); 19895d4f98a2SYan Zheng return 0; 19905d4f98a2SYan Zheng } 19915d4f98a2SYan Zheng 19923fd0a558SYan, Zheng if (cow) { 19935d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 19945d4f98a2SYan Zheng BUG_ON(ret); 19953fd0a558SYan, Zheng } 19968bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19975d4f98a2SYan Zheng 19985d4f98a2SYan Zheng if (next_key) { 19995d4f98a2SYan Zheng next_key->objectid = (u64)-1; 20005d4f98a2SYan Zheng next_key->type = (u8)-1; 20015d4f98a2SYan Zheng next_key->offset = (u64)-1; 20025d4f98a2SYan Zheng } 20035d4f98a2SYan Zheng 20045d4f98a2SYan Zheng parent = eb; 20055d4f98a2SYan Zheng while (1) { 2006581c1760SQu Wenruo struct btrfs_key first_key; 2007581c1760SQu Wenruo 20085d4f98a2SYan Zheng level = btrfs_header_level(parent); 20095d4f98a2SYan Zheng BUG_ON(level < lowest_level); 20105d4f98a2SYan Zheng 20115d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 2012cbca7d59SFilipe Manana if (ret < 0) 2013cbca7d59SFilipe Manana break; 20145d4f98a2SYan Zheng if (ret && slot > 0) 20155d4f98a2SYan Zheng slot--; 20165d4f98a2SYan Zheng 20175d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 20185d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 20195d4f98a2SYan Zheng 20205d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 20210b246afaSJeff Mahoney blocksize = fs_info->nodesize; 20225d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 202317515f1bSQu Wenruo btrfs_node_key_to_cpu(parent, &first_key, slot); 20245d4f98a2SYan Zheng 20255d4f98a2SYan Zheng if (level <= max_level) { 20265d4f98a2SYan Zheng eb = path->nodes[level]; 20275d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 20285d4f98a2SYan Zheng path->slots[level]); 20295d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 20305d4f98a2SYan Zheng path->slots[level]); 20315d4f98a2SYan Zheng } else { 20325d4f98a2SYan Zheng new_bytenr = 0; 20335d4f98a2SYan Zheng new_ptr_gen = 0; 20345d4f98a2SYan Zheng } 20355d4f98a2SYan Zheng 2036fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 20375d4f98a2SYan Zheng ret = level; 20385d4f98a2SYan Zheng break; 20395d4f98a2SYan Zheng } 20405d4f98a2SYan Zheng 20415d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 20425d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 20433fd0a558SYan, Zheng if (level <= lowest_level) { 20445d4f98a2SYan Zheng ret = 0; 20455d4f98a2SYan Zheng break; 20465d4f98a2SYan Zheng } 20475d4f98a2SYan Zheng 2048581c1760SQu Wenruo eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen, 2049581c1760SQu Wenruo level - 1, &first_key); 205064c043deSLiu Bo if (IS_ERR(eb)) { 205164c043deSLiu Bo ret = PTR_ERR(eb); 2052264813acSLiu Bo break; 205364c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 205464c043deSLiu Bo ret = -EIO; 2055416bc658SJosef Bacik free_extent_buffer(eb); 2056379cde74SStefan Behrens break; 2057416bc658SJosef Bacik } 20585d4f98a2SYan Zheng btrfs_tree_lock(eb); 20593fd0a558SYan, Zheng if (cow) { 20605d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 20615d4f98a2SYan Zheng slot, &eb); 20625d4f98a2SYan Zheng BUG_ON(ret); 20635d4f98a2SYan Zheng } 20648bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 20655d4f98a2SYan Zheng 20665d4f98a2SYan Zheng btrfs_tree_unlock(parent); 20675d4f98a2SYan Zheng free_extent_buffer(parent); 20685d4f98a2SYan Zheng 20695d4f98a2SYan Zheng parent = eb; 20705d4f98a2SYan Zheng continue; 20715d4f98a2SYan Zheng } 20725d4f98a2SYan Zheng 20733fd0a558SYan, Zheng if (!cow) { 20743fd0a558SYan, Zheng btrfs_tree_unlock(parent); 20753fd0a558SYan, Zheng free_extent_buffer(parent); 20763fd0a558SYan, Zheng cow = 1; 20773fd0a558SYan, Zheng goto again; 20783fd0a558SYan, Zheng } 20793fd0a558SYan, Zheng 20805d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 20815d4f98a2SYan Zheng path->slots[level]); 2082b3b4aa74SDavid Sterba btrfs_release_path(path); 20835d4f98a2SYan Zheng 20845d4f98a2SYan Zheng path->lowest_level = level; 20855d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 20865d4f98a2SYan Zheng path->lowest_level = 0; 20875d4f98a2SYan Zheng BUG_ON(ret); 20885d4f98a2SYan Zheng 20895d4f98a2SYan Zheng /* 2090824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 2091824d8dffSQu Wenruo * 2092824d8dffSQu Wenruo * We must trace both trees. 2093824d8dffSQu Wenruo * 1) Tree reloc subtree 2094824d8dffSQu Wenruo * If not traced, we will leak data numbers 2095824d8dffSQu Wenruo * 2) Fs subtree 2096824d8dffSQu Wenruo * If not traced, we will double count old data 2097f616f5cdSQu Wenruo * 2098f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 2099f616f5cdSQu Wenruo * the swapped tree blocks. 2100f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 2101f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 2102824d8dffSQu Wenruo */ 2103370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 2104370a11b8SQu Wenruo rc->block_group, parent, slot, 2105370a11b8SQu Wenruo path->nodes[level], path->slots[level], 2106370a11b8SQu Wenruo last_snapshot); 2107370a11b8SQu Wenruo if (ret < 0) 2108370a11b8SQu Wenruo break; 2109824d8dffSQu Wenruo /* 21105d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 21115d4f98a2SYan Zheng */ 21125d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 21135d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 21145d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 21155d4f98a2SYan Zheng 21165d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 21175d4f98a2SYan Zheng path->slots[level], old_bytenr); 21185d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 21195d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 21205d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 21215d4f98a2SYan Zheng 212282fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 212382fa113fSQu Wenruo blocksize, path->nodes[level]->start); 212482fa113fSQu Wenruo ref.skip_qgroup = true; 212582fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 212682fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 21275d4f98a2SYan Zheng BUG_ON(ret); 212882fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 212982fa113fSQu Wenruo blocksize, 0); 213082fa113fSQu Wenruo ref.skip_qgroup = true; 213182fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 213282fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 21335d4f98a2SYan Zheng BUG_ON(ret); 21345d4f98a2SYan Zheng 2135ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 2136ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 2137ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 2138ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2139ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 21405d4f98a2SYan Zheng BUG_ON(ret); 21415d4f98a2SYan Zheng 2142ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 2143ffd4bb2aSQu Wenruo blocksize, 0); 2144ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 2145ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2146ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 21475d4f98a2SYan Zheng BUG_ON(ret); 21485d4f98a2SYan Zheng 21495d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 21505d4f98a2SYan Zheng 21515d4f98a2SYan Zheng ret = level; 21525d4f98a2SYan Zheng break; 21535d4f98a2SYan Zheng } 21545d4f98a2SYan Zheng btrfs_tree_unlock(parent); 21555d4f98a2SYan Zheng free_extent_buffer(parent); 21565d4f98a2SYan Zheng return ret; 21575d4f98a2SYan Zheng } 21585d4f98a2SYan Zheng 21595d4f98a2SYan Zheng /* 21605d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 21615d4f98a2SYan Zheng */ 21625d4f98a2SYan Zheng static noinline_for_stack 21635d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 21645d4f98a2SYan Zheng int *level) 21655d4f98a2SYan Zheng { 21665d4f98a2SYan Zheng struct extent_buffer *eb; 21675d4f98a2SYan Zheng int i; 21685d4f98a2SYan Zheng u64 last_snapshot; 21695d4f98a2SYan Zheng u32 nritems; 21705d4f98a2SYan Zheng 21715d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 21725d4f98a2SYan Zheng 21735d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 21745d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 21755d4f98a2SYan Zheng path->nodes[i] = NULL; 21765d4f98a2SYan Zheng } 21775d4f98a2SYan Zheng 21785d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 21795d4f98a2SYan Zheng eb = path->nodes[i]; 21805d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 21815d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 21825d4f98a2SYan Zheng path->slots[i]++; 21835d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 21845d4f98a2SYan Zheng last_snapshot) 21855d4f98a2SYan Zheng continue; 21865d4f98a2SYan Zheng 21875d4f98a2SYan Zheng *level = i; 21885d4f98a2SYan Zheng return 0; 21895d4f98a2SYan Zheng } 21905d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 21915d4f98a2SYan Zheng path->nodes[i] = NULL; 21925d4f98a2SYan Zheng } 21935d4f98a2SYan Zheng return 1; 21945d4f98a2SYan Zheng } 21955d4f98a2SYan Zheng 21965d4f98a2SYan Zheng /* 21975d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 21985d4f98a2SYan Zheng */ 21995d4f98a2SYan Zheng static noinline_for_stack 22005d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 22015d4f98a2SYan Zheng int *level) 22025d4f98a2SYan Zheng { 22032ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 22045d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 22055d4f98a2SYan Zheng int i; 22065d4f98a2SYan Zheng u64 bytenr; 22075d4f98a2SYan Zheng u64 ptr_gen = 0; 22085d4f98a2SYan Zheng u64 last_snapshot; 22095d4f98a2SYan Zheng u32 nritems; 22105d4f98a2SYan Zheng 22115d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 22125d4f98a2SYan Zheng 22135d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 2214581c1760SQu Wenruo struct btrfs_key first_key; 2215581c1760SQu Wenruo 22165d4f98a2SYan Zheng eb = path->nodes[i]; 22175d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 22185d4f98a2SYan Zheng while (path->slots[i] < nritems) { 22195d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 22205d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 22215d4f98a2SYan Zheng break; 22225d4f98a2SYan Zheng path->slots[i]++; 22235d4f98a2SYan Zheng } 22245d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 22255d4f98a2SYan Zheng if (i == *level) 22265d4f98a2SYan Zheng break; 22275d4f98a2SYan Zheng *level = i + 1; 22285d4f98a2SYan Zheng return 0; 22295d4f98a2SYan Zheng } 22305d4f98a2SYan Zheng if (i == 1) { 22315d4f98a2SYan Zheng *level = i; 22325d4f98a2SYan Zheng return 0; 22335d4f98a2SYan Zheng } 22345d4f98a2SYan Zheng 22355d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 2236581c1760SQu Wenruo btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]); 2237581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1, 2238581c1760SQu Wenruo &first_key); 223964c043deSLiu Bo if (IS_ERR(eb)) { 224064c043deSLiu Bo return PTR_ERR(eb); 224164c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2242416bc658SJosef Bacik free_extent_buffer(eb); 2243416bc658SJosef Bacik return -EIO; 2244416bc658SJosef Bacik } 22455d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 22465d4f98a2SYan Zheng path->nodes[i - 1] = eb; 22475d4f98a2SYan Zheng path->slots[i - 1] = 0; 22485d4f98a2SYan Zheng } 22495d4f98a2SYan Zheng return 1; 22505d4f98a2SYan Zheng } 22515d4f98a2SYan Zheng 22525d4f98a2SYan Zheng /* 22535d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 22545d4f98a2SYan Zheng * [min_key, max_key) 22555d4f98a2SYan Zheng */ 22565d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 22575d4f98a2SYan Zheng struct btrfs_key *min_key, 22585d4f98a2SYan Zheng struct btrfs_key *max_key) 22595d4f98a2SYan Zheng { 22600b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 22615d4f98a2SYan Zheng struct inode *inode = NULL; 22625d4f98a2SYan Zheng u64 objectid; 22635d4f98a2SYan Zheng u64 start, end; 226433345d01SLi Zefan u64 ino; 22655d4f98a2SYan Zheng 22665d4f98a2SYan Zheng objectid = min_key->objectid; 22675d4f98a2SYan Zheng while (1) { 22685d4f98a2SYan Zheng cond_resched(); 22695d4f98a2SYan Zheng iput(inode); 22705d4f98a2SYan Zheng 22715d4f98a2SYan Zheng if (objectid > max_key->objectid) 22725d4f98a2SYan Zheng break; 22735d4f98a2SYan Zheng 22745d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 22755d4f98a2SYan Zheng if (!inode) 22765d4f98a2SYan Zheng break; 22774a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 22785d4f98a2SYan Zheng 227933345d01SLi Zefan if (ino > max_key->objectid) { 22805d4f98a2SYan Zheng iput(inode); 22815d4f98a2SYan Zheng break; 22825d4f98a2SYan Zheng } 22835d4f98a2SYan Zheng 228433345d01SLi Zefan objectid = ino + 1; 22855d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 22865d4f98a2SYan Zheng continue; 22875d4f98a2SYan Zheng 228833345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 22895d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 22905d4f98a2SYan Zheng continue; 22915d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 22925d4f98a2SYan Zheng start = 0; 22935d4f98a2SYan Zheng else { 22945d4f98a2SYan Zheng start = min_key->offset; 22950b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 22965d4f98a2SYan Zheng } 22975d4f98a2SYan Zheng } else { 22985d4f98a2SYan Zheng start = 0; 22995d4f98a2SYan Zheng } 23005d4f98a2SYan Zheng 230133345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 23025d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 23035d4f98a2SYan Zheng continue; 23045d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 23055d4f98a2SYan Zheng end = (u64)-1; 23065d4f98a2SYan Zheng } else { 23075d4f98a2SYan Zheng if (max_key->offset == 0) 23085d4f98a2SYan Zheng continue; 23095d4f98a2SYan Zheng end = max_key->offset; 23100b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 23115d4f98a2SYan Zheng end--; 23125d4f98a2SYan Zheng } 23135d4f98a2SYan Zheng } else { 23145d4f98a2SYan Zheng end = (u64)-1; 23155d4f98a2SYan Zheng } 23165d4f98a2SYan Zheng 23175d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 2318d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 2319dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 2320d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 23215d4f98a2SYan Zheng } 23225d4f98a2SYan Zheng return 0; 23235d4f98a2SYan Zheng } 23245d4f98a2SYan Zheng 23255d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 23265d4f98a2SYan Zheng struct btrfs_key *key) 23275d4f98a2SYan Zheng 23285d4f98a2SYan Zheng { 23295d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 23305d4f98a2SYan Zheng if (!path->nodes[level]) 23315d4f98a2SYan Zheng break; 23325d4f98a2SYan Zheng if (path->slots[level] + 1 < 23335d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 23345d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 23355d4f98a2SYan Zheng path->slots[level] + 1); 23365d4f98a2SYan Zheng return 0; 23375d4f98a2SYan Zheng } 23385d4f98a2SYan Zheng level++; 23395d4f98a2SYan Zheng } 23405d4f98a2SYan Zheng return 1; 23415d4f98a2SYan Zheng } 23425d4f98a2SYan Zheng 23435d4f98a2SYan Zheng /* 2344d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 2345d2311e69SQu Wenruo */ 2346d2311e69SQu Wenruo static void insert_dirty_subvol(struct btrfs_trans_handle *trans, 2347d2311e69SQu Wenruo struct reloc_control *rc, 2348d2311e69SQu Wenruo struct btrfs_root *root) 2349d2311e69SQu Wenruo { 2350d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2351d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 2352d2311e69SQu Wenruo 2353d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 2354d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 2355d2311e69SQu Wenruo ASSERT(reloc_root); 2356d2311e69SQu Wenruo 2357d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 2358d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 2359d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 2360d2311e69SQu Wenruo reloc_root_item->drop_level = 0; 2361d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 2362d2311e69SQu Wenruo btrfs_update_reloc_root(trans, root); 2363d2311e69SQu Wenruo 2364d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 236500246528SJosef Bacik btrfs_grab_root(root); 2366d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 2367d2311e69SQu Wenruo } 2368d2311e69SQu Wenruo } 2369d2311e69SQu Wenruo 2370d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 2371d2311e69SQu Wenruo { 2372d2311e69SQu Wenruo struct btrfs_root *root; 2373d2311e69SQu Wenruo struct btrfs_root *next; 2374d2311e69SQu Wenruo int ret = 0; 237530d40577SQu Wenruo int ret2; 2376d2311e69SQu Wenruo 2377d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 2378d2311e69SQu Wenruo reloc_dirty_list) { 237930d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 238030d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 2381d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2382d2311e69SQu Wenruo 2383d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 2384d2311e69SQu Wenruo root->reloc_root = NULL; 23856282675eSQu Wenruo /* 23866282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 23876282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 23886282675eSQu Wenruo */ 23896282675eSQu Wenruo smp_wmb(); 23901fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 2391f28de8d8SJosef Bacik if (reloc_root) { 2392f44deb74SJosef Bacik /* 2393f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 2394f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 2395f44deb74SJosef Bacik * drop the ref ourselves. 2396f44deb74SJosef Bacik */ 2397f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 2398f44deb74SJosef Bacik if (ret2 < 0) { 2399f44deb74SJosef Bacik btrfs_put_root(reloc_root); 2400f44deb74SJosef Bacik if (!ret) 2401f28de8d8SJosef Bacik ret = ret2; 2402f28de8d8SJosef Bacik } 2403f44deb74SJosef Bacik } 240400246528SJosef Bacik btrfs_put_root(root); 240530d40577SQu Wenruo } else { 240630d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 24070078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 2408f44deb74SJosef Bacik if (ret2 < 0) { 2409f44deb74SJosef Bacik btrfs_put_root(root); 2410f44deb74SJosef Bacik if (!ret) 241130d40577SQu Wenruo ret = ret2; 241230d40577SQu Wenruo } 2413d2311e69SQu Wenruo } 2414f44deb74SJosef Bacik } 2415d2311e69SQu Wenruo return ret; 2416d2311e69SQu Wenruo } 2417d2311e69SQu Wenruo 2418d2311e69SQu Wenruo /* 24195d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 24205d4f98a2SYan Zheng * fs tree. 24215d4f98a2SYan Zheng */ 24225d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 24235d4f98a2SYan Zheng struct btrfs_root *root) 24245d4f98a2SYan Zheng { 24250b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 24265d4f98a2SYan Zheng struct btrfs_key key; 24275d4f98a2SYan Zheng struct btrfs_key next_key; 24289e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 24295d4f98a2SYan Zheng struct btrfs_root *reloc_root; 24305d4f98a2SYan Zheng struct btrfs_root_item *root_item; 24315d4f98a2SYan Zheng struct btrfs_path *path; 24323fd0a558SYan, Zheng struct extent_buffer *leaf; 24335d4f98a2SYan Zheng int level; 24345d4f98a2SYan Zheng int max_level; 24355d4f98a2SYan Zheng int replaced = 0; 24365d4f98a2SYan Zheng int ret; 24375d4f98a2SYan Zheng int err = 0; 24383fd0a558SYan, Zheng u32 min_reserved; 24395d4f98a2SYan Zheng 24405d4f98a2SYan Zheng path = btrfs_alloc_path(); 24415d4f98a2SYan Zheng if (!path) 24425d4f98a2SYan Zheng return -ENOMEM; 2443e4058b54SDavid Sterba path->reada = READA_FORWARD; 24445d4f98a2SYan Zheng 24455d4f98a2SYan Zheng reloc_root = root->reloc_root; 24465d4f98a2SYan Zheng root_item = &reloc_root->root_item; 24475d4f98a2SYan Zheng 24485d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 24495d4f98a2SYan Zheng level = btrfs_root_level(root_item); 245067439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 24515d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 24525d4f98a2SYan Zheng path->slots[level] = 0; 24535d4f98a2SYan Zheng } else { 24545d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 24555d4f98a2SYan Zheng 24565d4f98a2SYan Zheng level = root_item->drop_level; 24575d4f98a2SYan Zheng BUG_ON(level == 0); 24585d4f98a2SYan Zheng path->lowest_level = level; 24595d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 246033c66f43SYan Zheng path->lowest_level = 0; 24615d4f98a2SYan Zheng if (ret < 0) { 24625d4f98a2SYan Zheng btrfs_free_path(path); 24635d4f98a2SYan Zheng return ret; 24645d4f98a2SYan Zheng } 24655d4f98a2SYan Zheng 24665d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 24675d4f98a2SYan Zheng path->slots[level]); 24685d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 24695d4f98a2SYan Zheng 24705d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 24715d4f98a2SYan Zheng } 24725d4f98a2SYan Zheng 24730b246afaSJeff Mahoney min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 24745d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 24755d4f98a2SYan Zheng 24765d4f98a2SYan Zheng while (1) { 247708e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 247808e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 24793fd0a558SYan, Zheng if (ret) { 24809e6a0c52SJosef Bacik err = ret; 24819e6a0c52SJosef Bacik goto out; 24823fd0a558SYan, Zheng } 24839e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 24849e6a0c52SJosef Bacik if (IS_ERR(trans)) { 24859e6a0c52SJosef Bacik err = PTR_ERR(trans); 24869e6a0c52SJosef Bacik trans = NULL; 24879e6a0c52SJosef Bacik goto out; 24889e6a0c52SJosef Bacik } 24892abc726aSJosef Bacik 24902abc726aSJosef Bacik /* 24912abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 24922abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 24932abc726aSJosef Bacik * 24942abc726aSJosef Bacik * But that's ok, we started the trans handle on our 24952abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 24962abc726aSJosef Bacik * dirty list. At commit time we'll still call 24972abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 24982abc726aSJosef Bacik * appropriately. 24992abc726aSJosef Bacik */ 25002abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 25019e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 25023fd0a558SYan, Zheng 25033fd0a558SYan, Zheng replaced = 0; 25045d4f98a2SYan Zheng max_level = level; 25055d4f98a2SYan Zheng 25065d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 25075d4f98a2SYan Zheng if (ret < 0) { 25085d4f98a2SYan Zheng err = ret; 25095d4f98a2SYan Zheng goto out; 25105d4f98a2SYan Zheng } 25115d4f98a2SYan Zheng if (ret > 0) 25125d4f98a2SYan Zheng break; 25135d4f98a2SYan Zheng 25145d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 25155d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 25165d4f98a2SYan Zheng ret = 0; 25175d4f98a2SYan Zheng } else { 25183d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 25193fd0a558SYan, Zheng &next_key, level, max_level); 25205d4f98a2SYan Zheng } 25215d4f98a2SYan Zheng if (ret < 0) { 25225d4f98a2SYan Zheng err = ret; 25235d4f98a2SYan Zheng goto out; 25245d4f98a2SYan Zheng } 25255d4f98a2SYan Zheng 25265d4f98a2SYan Zheng if (ret > 0) { 25275d4f98a2SYan Zheng level = ret; 25285d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 25295d4f98a2SYan Zheng path->slots[level]); 25305d4f98a2SYan Zheng replaced = 1; 25315d4f98a2SYan Zheng } 25325d4f98a2SYan Zheng 25335d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 25345d4f98a2SYan Zheng if (ret > 0) 25355d4f98a2SYan Zheng break; 25365d4f98a2SYan Zheng 25375d4f98a2SYan Zheng BUG_ON(level == 0); 25385d4f98a2SYan Zheng /* 25395d4f98a2SYan Zheng * save the merging progress in the drop_progress. 25405d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 25415d4f98a2SYan Zheng */ 25425d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 25435d4f98a2SYan Zheng path->slots[level]); 25445d4f98a2SYan Zheng root_item->drop_level = level; 25455d4f98a2SYan Zheng 25463a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 25479e6a0c52SJosef Bacik trans = NULL; 25485d4f98a2SYan Zheng 25492ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 25505d4f98a2SYan Zheng 25515d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 25525d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 25535d4f98a2SYan Zheng } 25545d4f98a2SYan Zheng 25555d4f98a2SYan Zheng /* 25565d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 25575d4f98a2SYan Zheng * relocated and the block is tree root. 25585d4f98a2SYan Zheng */ 25595d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 25605d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 25615d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 25625d4f98a2SYan Zheng free_extent_buffer(leaf); 25635d4f98a2SYan Zheng if (ret < 0) 25645d4f98a2SYan Zheng err = ret; 25655d4f98a2SYan Zheng out: 25665d4f98a2SYan Zheng btrfs_free_path(path); 25675d4f98a2SYan Zheng 2568d2311e69SQu Wenruo if (err == 0) 2569d2311e69SQu Wenruo insert_dirty_subvol(trans, rc, root); 25705d4f98a2SYan Zheng 25719e6a0c52SJosef Bacik if (trans) 25723a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 25735d4f98a2SYan Zheng 25742ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 25755d4f98a2SYan Zheng 25765d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 25775d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 25785d4f98a2SYan Zheng 25795d4f98a2SYan Zheng return err; 25805d4f98a2SYan Zheng } 25815d4f98a2SYan Zheng 25823fd0a558SYan, Zheng static noinline_for_stack 25833fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 25845d4f98a2SYan Zheng { 25853fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 25860b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 25873fd0a558SYan, Zheng struct btrfs_root *reloc_root; 25885d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 25893fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 25903fd0a558SYan, Zheng u64 num_bytes = 0; 25913fd0a558SYan, Zheng int ret; 25923fd0a558SYan, Zheng 25930b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 25940b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 25953fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 25960b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 25977585717fSChris Mason 25983fd0a558SYan, Zheng again: 25993fd0a558SYan, Zheng if (!err) { 26003fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 260108e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 260208e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 26033fd0a558SYan, Zheng if (ret) 26043fd0a558SYan, Zheng err = ret; 26053fd0a558SYan, Zheng } 26063fd0a558SYan, Zheng 26077a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 26083612b495STsutomu Itoh if (IS_ERR(trans)) { 26093612b495STsutomu Itoh if (!err) 26102ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 261163f018beSNikolay Borisov num_bytes, NULL); 26123612b495STsutomu Itoh return PTR_ERR(trans); 26133612b495STsutomu Itoh } 26143fd0a558SYan, Zheng 26153fd0a558SYan, Zheng if (!err) { 26163fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 26173a45bb20SJeff Mahoney btrfs_end_transaction(trans); 26182ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 261963f018beSNikolay Borisov num_bytes, NULL); 26203fd0a558SYan, Zheng goto again; 26213fd0a558SYan, Zheng } 26223fd0a558SYan, Zheng } 26233fd0a558SYan, Zheng 26243fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 26253fd0a558SYan, Zheng 26263fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 26273fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 26283fd0a558SYan, Zheng struct btrfs_root, root_list); 26293fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 26303fd0a558SYan, Zheng 26310b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 26323fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 26333fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 26343fd0a558SYan, Zheng 26353fd0a558SYan, Zheng /* 26363fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 26373fd0a558SYan, Zheng * knows it should resumes merging 26383fd0a558SYan, Zheng */ 26393fd0a558SYan, Zheng if (!err) 26403fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 26413fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 26423fd0a558SYan, Zheng 26433fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 264400246528SJosef Bacik btrfs_put_root(root); 26453fd0a558SYan, Zheng } 26463fd0a558SYan, Zheng 26473fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 26483fd0a558SYan, Zheng 26493fd0a558SYan, Zheng if (!err) 26503a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 26513fd0a558SYan, Zheng else 26523a45bb20SJeff Mahoney btrfs_end_transaction(trans); 26533fd0a558SYan, Zheng return err; 26543fd0a558SYan, Zheng } 26553fd0a558SYan, Zheng 26563fd0a558SYan, Zheng static noinline_for_stack 2657aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 2658aca1bba6SLiu Bo { 2659aca1bba6SLiu Bo struct btrfs_root *reloc_root; 2660aca1bba6SLiu Bo 2661aca1bba6SLiu Bo while (!list_empty(list)) { 2662aca1bba6SLiu Bo reloc_root = list_entry(list->next, struct btrfs_root, 2663aca1bba6SLiu Bo root_list); 2664bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 2665aca1bba6SLiu Bo } 2666aca1bba6SLiu Bo } 2667aca1bba6SLiu Bo 2668aca1bba6SLiu Bo static noinline_for_stack 266994404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 26703fd0a558SYan, Zheng { 26710b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 26725d4f98a2SYan Zheng struct btrfs_root *root; 26735d4f98a2SYan Zheng struct btrfs_root *reloc_root; 26743fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 26753fd0a558SYan, Zheng int found = 0; 2676aca1bba6SLiu Bo int ret = 0; 26773fd0a558SYan, Zheng again: 26783fd0a558SYan, Zheng root = rc->extent_root; 26797585717fSChris Mason 26807585717fSChris Mason /* 26817585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 26827585717fSChris Mason * we have to make sure nobody is in the middle of 26837585717fSChris Mason * adding their roots to the list while we are 26847585717fSChris Mason * doing this splice 26857585717fSChris Mason */ 26860b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 26873fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 26880b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 26895d4f98a2SYan Zheng 26903fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 26913fd0a558SYan, Zheng found = 1; 26923fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 26933fd0a558SYan, Zheng struct btrfs_root, root_list); 26945d4f98a2SYan Zheng 26955d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 26960b246afaSJeff Mahoney root = read_fs_root(fs_info, 26975d4f98a2SYan Zheng reloc_root->root_key.offset); 26985d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 26995d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 27005d4f98a2SYan Zheng 27013fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 270200246528SJosef Bacik btrfs_put_root(root); 2703b37b39cdSJosef Bacik if (ret) { 270425e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 270525e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 270625e293c2SWang Shilong &reloc_roots); 2707aca1bba6SLiu Bo goto out; 2708b37b39cdSJosef Bacik } 27093fd0a558SYan, Zheng } else { 27103fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 271130d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 271230d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 271330d40577SQu Wenruo &rc->dirty_subvol_roots); 27143fd0a558SYan, Zheng } 27155d4f98a2SYan Zheng } 27165d4f98a2SYan Zheng 27173fd0a558SYan, Zheng if (found) { 27183fd0a558SYan, Zheng found = 0; 27193fd0a558SYan, Zheng goto again; 27205d4f98a2SYan Zheng } 2721aca1bba6SLiu Bo out: 2722aca1bba6SLiu Bo if (ret) { 27230b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 2724aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 2725aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 2726467bb1d2SWang Shilong 2727467bb1d2SWang Shilong /* new reloc root may be added */ 27280b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 2729467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 27300b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 2731467bb1d2SWang Shilong if (!list_empty(&reloc_roots)) 2732467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 2733aca1bba6SLiu Bo } 2734aca1bba6SLiu Bo 27357b7b7431SJosef Bacik /* 27367b7b7431SJosef Bacik * We used to have 27377b7b7431SJosef Bacik * 27387b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 27397b7b7431SJosef Bacik * 27407b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 27417b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 27427b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 27437b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 27447b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 27457b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 27467b7b7431SJosef Bacik * will be cleaned up on unmount. 27477b7b7431SJosef Bacik * 27487b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 27497b7b7431SJosef Bacik */ 27505d4f98a2SYan Zheng } 27515d4f98a2SYan Zheng 27525d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 27535d4f98a2SYan Zheng { 27545d4f98a2SYan Zheng struct tree_block *block; 27555d4f98a2SYan Zheng struct rb_node *rb_node; 27565d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 27575d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 27585d4f98a2SYan Zheng rb_erase(rb_node, blocks); 27595d4f98a2SYan Zheng kfree(block); 27605d4f98a2SYan Zheng } 27615d4f98a2SYan Zheng } 27625d4f98a2SYan Zheng 27635d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 27645d4f98a2SYan Zheng struct btrfs_root *reloc_root) 27655d4f98a2SYan Zheng { 27660b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 27675d4f98a2SYan Zheng struct btrfs_root *root; 2768442b1ac5SJosef Bacik int ret; 27695d4f98a2SYan Zheng 27705d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 27715d4f98a2SYan Zheng return 0; 27725d4f98a2SYan Zheng 27730b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 27745d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 27755d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 2776442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 277700246528SJosef Bacik btrfs_put_root(root); 27785d4f98a2SYan Zheng 2779442b1ac5SJosef Bacik return ret; 27805d4f98a2SYan Zheng } 27815d4f98a2SYan Zheng 27823fd0a558SYan, Zheng static noinline_for_stack 27833fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 27843fd0a558SYan, Zheng struct reloc_control *rc, 27855d4f98a2SYan Zheng struct backref_node *node, 2786dc4103f9SWang Shilong struct backref_edge *edges[]) 27875d4f98a2SYan Zheng { 27885d4f98a2SYan Zheng struct backref_node *next; 27895d4f98a2SYan Zheng struct btrfs_root *root; 27903fd0a558SYan, Zheng int index = 0; 27913fd0a558SYan, Zheng 27925d4f98a2SYan Zheng next = node; 27935d4f98a2SYan Zheng while (1) { 27945d4f98a2SYan Zheng cond_resched(); 27955d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 27965d4f98a2SYan Zheng root = next->root; 27973fd0a558SYan, Zheng BUG_ON(!root); 279827cdeb70SMiao Xie BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state)); 27995d4f98a2SYan Zheng 28005d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 28015d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 28025d4f98a2SYan Zheng break; 28035d4f98a2SYan Zheng } 28045d4f98a2SYan Zheng 28055d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 28063fd0a558SYan, Zheng root = root->reloc_root; 28073fd0a558SYan, Zheng 28083fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 28093fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 28103fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 28113fd0a558SYan, Zheng next->new_bytenr = root->node->start; 281200246528SJosef Bacik btrfs_put_root(next->root); 281300246528SJosef Bacik next->root = btrfs_grab_root(root); 28140b530bc5SJosef Bacik ASSERT(next->root); 28153fd0a558SYan, Zheng list_add_tail(&next->list, 28163fd0a558SYan, Zheng &rc->backref_cache.changed); 28179569cc20SQu Wenruo mark_block_processed(rc, next); 28185d4f98a2SYan Zheng break; 28195d4f98a2SYan Zheng } 28205d4f98a2SYan Zheng 28213fd0a558SYan, Zheng WARN_ON(1); 28225d4f98a2SYan Zheng root = NULL; 28235d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 28245d4f98a2SYan Zheng if (!next || next->level <= node->level) 28255d4f98a2SYan Zheng break; 28265d4f98a2SYan Zheng } 28273fd0a558SYan, Zheng if (!root) 28283fd0a558SYan, Zheng return NULL; 28295d4f98a2SYan Zheng 28303fd0a558SYan, Zheng next = node; 28313fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 28323fd0a558SYan, Zheng while (1) { 28333fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 28343fd0a558SYan, Zheng if (--index < 0) 28353fd0a558SYan, Zheng break; 28363fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 28373fd0a558SYan, Zheng } 28385d4f98a2SYan Zheng return root; 28395d4f98a2SYan Zheng } 28405d4f98a2SYan Zheng 28413fd0a558SYan, Zheng /* 28423fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 28433fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 28443fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 28453fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 28463fd0a558SYan, Zheng */ 28475d4f98a2SYan Zheng static noinline_for_stack 2848147d256eSZhaolei struct btrfs_root *select_one_root(struct backref_node *node) 28495d4f98a2SYan Zheng { 28503fd0a558SYan, Zheng struct backref_node *next; 28513fd0a558SYan, Zheng struct btrfs_root *root; 28523fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 28535d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 28543fd0a558SYan, Zheng int index = 0; 28553fd0a558SYan, Zheng 28563fd0a558SYan, Zheng next = node; 28573fd0a558SYan, Zheng while (1) { 28583fd0a558SYan, Zheng cond_resched(); 28593fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 28603fd0a558SYan, Zheng root = next->root; 28613fd0a558SYan, Zheng BUG_ON(!root); 28623fd0a558SYan, Zheng 286325985edcSLucas De Marchi /* no other choice for non-references counted tree */ 286427cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 28653fd0a558SYan, Zheng return root; 28663fd0a558SYan, Zheng 28673fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 28683fd0a558SYan, Zheng fs_root = root; 28693fd0a558SYan, Zheng 28703fd0a558SYan, Zheng if (next != node) 28713fd0a558SYan, Zheng return NULL; 28723fd0a558SYan, Zheng 28733fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 28743fd0a558SYan, Zheng if (!next || next->level <= node->level) 28753fd0a558SYan, Zheng break; 28763fd0a558SYan, Zheng } 28773fd0a558SYan, Zheng 28783fd0a558SYan, Zheng if (!fs_root) 28793fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 28803fd0a558SYan, Zheng return fs_root; 28815d4f98a2SYan Zheng } 28825d4f98a2SYan Zheng 28835d4f98a2SYan Zheng static noinline_for_stack 28843fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 28853fd0a558SYan, Zheng struct backref_node *node, int reserve) 28865d4f98a2SYan Zheng { 28870b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 28883fd0a558SYan, Zheng struct backref_node *next = node; 28893fd0a558SYan, Zheng struct backref_edge *edge; 28903fd0a558SYan, Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 28913fd0a558SYan, Zheng u64 num_bytes = 0; 28923fd0a558SYan, Zheng int index = 0; 28935d4f98a2SYan Zheng 28943fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 28953fd0a558SYan, Zheng 28963fd0a558SYan, Zheng while (next) { 28973fd0a558SYan, Zheng cond_resched(); 28985d4f98a2SYan Zheng while (1) { 28993fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 29005d4f98a2SYan Zheng break; 29015d4f98a2SYan Zheng 29020b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 29033fd0a558SYan, Zheng 29043fd0a558SYan, Zheng if (list_empty(&next->upper)) 29053fd0a558SYan, Zheng break; 29063fd0a558SYan, Zheng 29073fd0a558SYan, Zheng edge = list_entry(next->upper.next, 29083fd0a558SYan, Zheng struct backref_edge, list[LOWER]); 29093fd0a558SYan, Zheng edges[index++] = edge; 29103fd0a558SYan, Zheng next = edge->node[UPPER]; 29115d4f98a2SYan Zheng } 29123fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 29133fd0a558SYan, Zheng } 29143fd0a558SYan, Zheng return num_bytes; 29153fd0a558SYan, Zheng } 29163fd0a558SYan, Zheng 29173fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 29183fd0a558SYan, Zheng struct reloc_control *rc, 29193fd0a558SYan, Zheng struct backref_node *node) 29203fd0a558SYan, Zheng { 29213fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2922da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 29233fd0a558SYan, Zheng u64 num_bytes; 29243fd0a558SYan, Zheng int ret; 29250647bf56SWang Shilong u64 tmp; 29263fd0a558SYan, Zheng 29273fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 29283fd0a558SYan, Zheng 29293fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 29300647bf56SWang Shilong rc->reserved_bytes += num_bytes; 29318ca17f0fSJosef Bacik 29328ca17f0fSJosef Bacik /* 29338ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 29348ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 29358ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 29368ca17f0fSJosef Bacik */ 29370647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 29388ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 29393fd0a558SYan, Zheng if (ret) { 2940da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 29410647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 29420647bf56SWang Shilong tmp <<= 1; 29430647bf56SWang Shilong /* 29440647bf56SWang Shilong * only one thread can access block_rsv at this point, 29450647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 29460647bf56SWang Shilong * we expand more reservation size here to allow enough 294752042d8eSAndrea Gelmini * space for relocation and we will return earlier in 29480647bf56SWang Shilong * enospc case. 29490647bf56SWang Shilong */ 2950da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 29510647bf56SWang Shilong RELOCATION_RESERVED_NODES; 29528ca17f0fSJosef Bacik return -EAGAIN; 29533fd0a558SYan, Zheng } 29543fd0a558SYan, Zheng 29553fd0a558SYan, Zheng return 0; 29563fd0a558SYan, Zheng } 29573fd0a558SYan, Zheng 29585d4f98a2SYan Zheng /* 29595d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 29605d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 29615d4f98a2SYan Zheng * 29625d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 29635d4f98a2SYan Zheng * in that case this function just updates pointers. 29645d4f98a2SYan Zheng */ 29655d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 29663fd0a558SYan, Zheng struct reloc_control *rc, 29675d4f98a2SYan Zheng struct backref_node *node, 29685d4f98a2SYan Zheng struct btrfs_key *key, 29695d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 29705d4f98a2SYan Zheng { 29712ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 29725d4f98a2SYan Zheng struct backref_node *upper; 29735d4f98a2SYan Zheng struct backref_edge *edge; 29745d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 29755d4f98a2SYan Zheng struct btrfs_root *root; 29765d4f98a2SYan Zheng struct extent_buffer *eb; 29775d4f98a2SYan Zheng u32 blocksize; 29785d4f98a2SYan Zheng u64 bytenr; 29795d4f98a2SYan Zheng u64 generation; 29805d4f98a2SYan Zheng int slot; 29815d4f98a2SYan Zheng int ret; 29825d4f98a2SYan Zheng int err = 0; 29835d4f98a2SYan Zheng 29845d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 29855d4f98a2SYan Zheng 29865d4f98a2SYan Zheng path->lowest_level = node->level + 1; 29873fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 29885d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 2989581c1760SQu Wenruo struct btrfs_key first_key; 299082fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2991581c1760SQu Wenruo 29925d4f98a2SYan Zheng cond_resched(); 29935d4f98a2SYan Zheng 29945d4f98a2SYan Zheng upper = edge->node[UPPER]; 2995dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 29963fd0a558SYan, Zheng BUG_ON(!root); 29975d4f98a2SYan Zheng 29983fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 29993fd0a558SYan, Zheng if (!lowest) { 30003fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 30013fd0a558SYan, Zheng upper->level, &slot); 3002cbca7d59SFilipe Manana if (ret < 0) { 3003cbca7d59SFilipe Manana err = ret; 3004cbca7d59SFilipe Manana goto next; 3005cbca7d59SFilipe Manana } 30063fd0a558SYan, Zheng BUG_ON(ret); 30073fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 30083fd0a558SYan, Zheng if (node->eb->start == bytenr) 30093fd0a558SYan, Zheng goto next; 30103fd0a558SYan, Zheng } 30115d4f98a2SYan Zheng drop_node_buffer(upper); 30123fd0a558SYan, Zheng } 30135d4f98a2SYan Zheng 30145d4f98a2SYan Zheng if (!upper->eb) { 30155d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 30163561b9dbSLiu Bo if (ret) { 30173561b9dbSLiu Bo if (ret < 0) 30185d4f98a2SYan Zheng err = ret; 30193561b9dbSLiu Bo else 30203561b9dbSLiu Bo err = -ENOENT; 30213561b9dbSLiu Bo 30223561b9dbSLiu Bo btrfs_release_path(path); 30235d4f98a2SYan Zheng break; 30245d4f98a2SYan Zheng } 30255d4f98a2SYan Zheng 30263fd0a558SYan, Zheng if (!upper->eb) { 30273fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 30283fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 30293fd0a558SYan, Zheng } else { 30303fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 30313fd0a558SYan, Zheng } 30323fd0a558SYan, Zheng 30333fd0a558SYan, Zheng upper->locked = 1; 30343fd0a558SYan, Zheng path->locks[upper->level] = 0; 30353fd0a558SYan, Zheng 30365d4f98a2SYan Zheng slot = path->slots[upper->level]; 3037b3b4aa74SDavid Sterba btrfs_release_path(path); 30385d4f98a2SYan Zheng } else { 30395d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 30405d4f98a2SYan Zheng &slot); 3041cbca7d59SFilipe Manana if (ret < 0) { 3042cbca7d59SFilipe Manana err = ret; 3043cbca7d59SFilipe Manana goto next; 3044cbca7d59SFilipe Manana } 30455d4f98a2SYan Zheng BUG_ON(ret); 30465d4f98a2SYan Zheng } 30475d4f98a2SYan Zheng 30485d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 30493fd0a558SYan, Zheng if (lowest) { 30504547f4d8SLiu Bo if (bytenr != node->bytenr) { 30514547f4d8SLiu Bo btrfs_err(root->fs_info, 30524547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 30534547f4d8SLiu Bo bytenr, node->bytenr, slot, 30544547f4d8SLiu Bo upper->eb->start); 30554547f4d8SLiu Bo err = -EIO; 30564547f4d8SLiu Bo goto next; 30574547f4d8SLiu Bo } 30585d4f98a2SYan Zheng } else { 30593fd0a558SYan, Zheng if (node->eb->start == bytenr) 30603fd0a558SYan, Zheng goto next; 30615d4f98a2SYan Zheng } 30625d4f98a2SYan Zheng 3063da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 30645d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 3065581c1760SQu Wenruo btrfs_node_key_to_cpu(upper->eb, &first_key, slot); 3066581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, generation, 3067581c1760SQu Wenruo upper->level - 1, &first_key); 306864c043deSLiu Bo if (IS_ERR(eb)) { 306964c043deSLiu Bo err = PTR_ERR(eb); 307064c043deSLiu Bo goto next; 307164c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 3072416bc658SJosef Bacik free_extent_buffer(eb); 307397d9a8a4STsutomu Itoh err = -EIO; 307497d9a8a4STsutomu Itoh goto next; 307597d9a8a4STsutomu Itoh } 30765d4f98a2SYan Zheng btrfs_tree_lock(eb); 30778bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 30785d4f98a2SYan Zheng 30795d4f98a2SYan Zheng if (!node->eb) { 30805d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 30815d4f98a2SYan Zheng slot, &eb); 30823fd0a558SYan, Zheng btrfs_tree_unlock(eb); 30833fd0a558SYan, Zheng free_extent_buffer(eb); 30845d4f98a2SYan Zheng if (ret < 0) { 30855d4f98a2SYan Zheng err = ret; 30863fd0a558SYan, Zheng goto next; 30875d4f98a2SYan Zheng } 30883fd0a558SYan, Zheng BUG_ON(node->eb != eb); 30895d4f98a2SYan Zheng } else { 30905d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 30915d4f98a2SYan Zheng node->eb->start); 30925d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 30935d4f98a2SYan Zheng trans->transid); 30945d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 30955d4f98a2SYan Zheng 309682fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 30975d4f98a2SYan Zheng node->eb->start, blocksize, 309882fa113fSQu Wenruo upper->eb->start); 309982fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 310082fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 310182fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 310282fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 31035d4f98a2SYan Zheng BUG_ON(ret); 31045d4f98a2SYan Zheng 31055d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 31065d4f98a2SYan Zheng BUG_ON(ret); 31075d4f98a2SYan Zheng } 31083fd0a558SYan, Zheng next: 31093fd0a558SYan, Zheng if (!upper->pending) 31103fd0a558SYan, Zheng drop_node_buffer(upper); 31113fd0a558SYan, Zheng else 31123fd0a558SYan, Zheng unlock_node_buffer(upper); 31133fd0a558SYan, Zheng if (err) 31143fd0a558SYan, Zheng break; 31155d4f98a2SYan Zheng } 31163fd0a558SYan, Zheng 31173fd0a558SYan, Zheng if (!err && node->pending) { 31183fd0a558SYan, Zheng drop_node_buffer(node); 31193fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 31203fd0a558SYan, Zheng node->pending = 0; 31215d4f98a2SYan Zheng } 31223fd0a558SYan, Zheng 31235d4f98a2SYan Zheng path->lowest_level = 0; 31243fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 31255d4f98a2SYan Zheng return err; 31265d4f98a2SYan Zheng } 31275d4f98a2SYan Zheng 31285d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 31293fd0a558SYan, Zheng struct reloc_control *rc, 31305d4f98a2SYan Zheng struct backref_node *node, 31315d4f98a2SYan Zheng struct btrfs_path *path) 31325d4f98a2SYan Zheng { 31335d4f98a2SYan Zheng struct btrfs_key key; 31345d4f98a2SYan Zheng 31355d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 31363fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 31375d4f98a2SYan Zheng } 31385d4f98a2SYan Zheng 31395d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 31403fd0a558SYan, Zheng struct reloc_control *rc, 31413fd0a558SYan, Zheng struct btrfs_path *path, int err) 31425d4f98a2SYan Zheng { 31433fd0a558SYan, Zheng LIST_HEAD(list); 31443fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 31455d4f98a2SYan Zheng struct backref_node *node; 31465d4f98a2SYan Zheng int level; 31475d4f98a2SYan Zheng int ret; 31485d4f98a2SYan Zheng 31495d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 31505d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 31515d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 31523fd0a558SYan, Zheng struct backref_node, list); 31533fd0a558SYan, Zheng list_move_tail(&node->list, &list); 31543fd0a558SYan, Zheng BUG_ON(!node->pending); 31555d4f98a2SYan Zheng 31563fd0a558SYan, Zheng if (!err) { 31573fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 31585d4f98a2SYan Zheng if (ret < 0) 31595d4f98a2SYan Zheng err = ret; 31605d4f98a2SYan Zheng } 31615d4f98a2SYan Zheng } 31623fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 31633fd0a558SYan, Zheng } 31645d4f98a2SYan Zheng return err; 31655d4f98a2SYan Zheng } 31665d4f98a2SYan Zheng 31675d4f98a2SYan Zheng /* 31685d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 31695d4f98a2SYan Zheng * as processed. 31705d4f98a2SYan Zheng */ 31715d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 31725d4f98a2SYan Zheng struct backref_node *node) 31735d4f98a2SYan Zheng { 31745d4f98a2SYan Zheng struct backref_node *next = node; 31755d4f98a2SYan Zheng struct backref_edge *edge; 31765d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 31775d4f98a2SYan Zheng int index = 0; 31785d4f98a2SYan Zheng 31795d4f98a2SYan Zheng while (next) { 31805d4f98a2SYan Zheng cond_resched(); 31815d4f98a2SYan Zheng while (1) { 31825d4f98a2SYan Zheng if (next->processed) 31835d4f98a2SYan Zheng break; 31845d4f98a2SYan Zheng 31859569cc20SQu Wenruo mark_block_processed(rc, next); 31865d4f98a2SYan Zheng 31875d4f98a2SYan Zheng if (list_empty(&next->upper)) 31885d4f98a2SYan Zheng break; 31895d4f98a2SYan Zheng 31905d4f98a2SYan Zheng edge = list_entry(next->upper.next, 31915d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 31925d4f98a2SYan Zheng edges[index++] = edge; 31935d4f98a2SYan Zheng next = edge->node[UPPER]; 31945d4f98a2SYan Zheng } 31955d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 31965d4f98a2SYan Zheng } 31975d4f98a2SYan Zheng } 31985d4f98a2SYan Zheng 31997476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 32005d4f98a2SYan Zheng { 3201da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 32027476dfdaSDavid Sterba 32035d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 32049655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 32055d4f98a2SYan Zheng return 1; 32065d4f98a2SYan Zheng return 0; 32075d4f98a2SYan Zheng } 32085d4f98a2SYan Zheng 32092ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 32105d4f98a2SYan Zheng struct tree_block *block) 32115d4f98a2SYan Zheng { 32125d4f98a2SYan Zheng struct extent_buffer *eb; 32135d4f98a2SYan Zheng 3214581c1760SQu Wenruo eb = read_tree_block(fs_info, block->bytenr, block->key.offset, 3215581c1760SQu Wenruo block->level, NULL); 321664c043deSLiu Bo if (IS_ERR(eb)) { 321764c043deSLiu Bo return PTR_ERR(eb); 321864c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 3219416bc658SJosef Bacik free_extent_buffer(eb); 3220416bc658SJosef Bacik return -EIO; 3221416bc658SJosef Bacik } 32225d4f98a2SYan Zheng if (block->level == 0) 32235d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 32245d4f98a2SYan Zheng else 32255d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 32265d4f98a2SYan Zheng free_extent_buffer(eb); 32275d4f98a2SYan Zheng block->key_ready = 1; 32285d4f98a2SYan Zheng return 0; 32295d4f98a2SYan Zheng } 32305d4f98a2SYan Zheng 32315d4f98a2SYan Zheng /* 32325d4f98a2SYan Zheng * helper function to relocate a tree block 32335d4f98a2SYan Zheng */ 32345d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 32355d4f98a2SYan Zheng struct reloc_control *rc, 32365d4f98a2SYan Zheng struct backref_node *node, 32375d4f98a2SYan Zheng struct btrfs_key *key, 32385d4f98a2SYan Zheng struct btrfs_path *path) 32395d4f98a2SYan Zheng { 32405d4f98a2SYan Zheng struct btrfs_root *root; 32413fd0a558SYan, Zheng int ret = 0; 32425d4f98a2SYan Zheng 32433fd0a558SYan, Zheng if (!node) 32445d4f98a2SYan Zheng return 0; 32453fd0a558SYan, Zheng 32465f6b2e5cSJosef Bacik /* 32475f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 32485f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 32495f6b2e5cSJosef Bacik */ 32505f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 32515f6b2e5cSJosef Bacik if (ret) 32525f6b2e5cSJosef Bacik goto out; 32535f6b2e5cSJosef Bacik 32543fd0a558SYan, Zheng BUG_ON(node->processed); 3255147d256eSZhaolei root = select_one_root(node); 32563fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 32573fd0a558SYan, Zheng update_processed_blocks(rc, node); 32583fd0a558SYan, Zheng goto out; 32595d4f98a2SYan Zheng } 32605d4f98a2SYan Zheng 32613fd0a558SYan, Zheng if (root) { 326227cdeb70SMiao Xie if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) { 32633fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 32643fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 32653fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 32663fd0a558SYan, Zheng root = root->reloc_root; 32673fd0a558SYan, Zheng node->new_bytenr = root->node->start; 326800246528SJosef Bacik btrfs_put_root(node->root); 326900246528SJosef Bacik node->root = btrfs_grab_root(root); 32700b530bc5SJosef Bacik ASSERT(node->root); 32713fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 32723fd0a558SYan, Zheng } else { 32735d4f98a2SYan Zheng path->lowest_level = node->level; 32745d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 3275b3b4aa74SDavid Sterba btrfs_release_path(path); 32763fd0a558SYan, Zheng if (ret > 0) 32775d4f98a2SYan Zheng ret = 0; 32783fd0a558SYan, Zheng } 32793fd0a558SYan, Zheng if (!ret) 32803fd0a558SYan, Zheng update_processed_blocks(rc, node); 32813fd0a558SYan, Zheng } else { 32823fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 32833fd0a558SYan, Zheng } 32845d4f98a2SYan Zheng out: 32850647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 32863fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 32875d4f98a2SYan Zheng return ret; 32885d4f98a2SYan Zheng } 32895d4f98a2SYan Zheng 32905d4f98a2SYan Zheng /* 32915d4f98a2SYan Zheng * relocate a list of blocks 32925d4f98a2SYan Zheng */ 32935d4f98a2SYan Zheng static noinline_for_stack 32945d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 32955d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 32965d4f98a2SYan Zheng { 32972ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 32985d4f98a2SYan Zheng struct backref_node *node; 32995d4f98a2SYan Zheng struct btrfs_path *path; 33005d4f98a2SYan Zheng struct tree_block *block; 330198ff7b94SQu Wenruo struct tree_block *next; 33025d4f98a2SYan Zheng int ret; 33035d4f98a2SYan Zheng int err = 0; 33045d4f98a2SYan Zheng 33055d4f98a2SYan Zheng path = btrfs_alloc_path(); 3306e1a12670SLiu Bo if (!path) { 3307e1a12670SLiu Bo err = -ENOMEM; 330834c2b290SDavid Sterba goto out_free_blocks; 3309e1a12670SLiu Bo } 33105d4f98a2SYan Zheng 331198ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 331298ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 33135d4f98a2SYan Zheng if (!block->key_ready) 33142ff7e61eSJeff Mahoney readahead_tree_block(fs_info, block->bytenr); 33155d4f98a2SYan Zheng } 33165d4f98a2SYan Zheng 331798ff7b94SQu Wenruo /* Get first keys */ 331898ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 331934c2b290SDavid Sterba if (!block->key_ready) { 33202ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 332134c2b290SDavid Sterba if (err) 332234c2b290SDavid Sterba goto out_free_path; 332334c2b290SDavid Sterba } 33245d4f98a2SYan Zheng } 33255d4f98a2SYan Zheng 332698ff7b94SQu Wenruo /* Do tree relocation */ 332798ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 33283fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 33295d4f98a2SYan Zheng block->level, block->bytenr); 33305d4f98a2SYan Zheng if (IS_ERR(node)) { 33315d4f98a2SYan Zheng err = PTR_ERR(node); 33325d4f98a2SYan Zheng goto out; 33335d4f98a2SYan Zheng } 33345d4f98a2SYan Zheng 33355d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 33365d4f98a2SYan Zheng path); 33375d4f98a2SYan Zheng if (ret < 0) { 33385d4f98a2SYan Zheng err = ret; 333950dbbb71SJosef Bacik break; 33405d4f98a2SYan Zheng } 33415d4f98a2SYan Zheng } 33425d4f98a2SYan Zheng out: 33433fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 33445d4f98a2SYan Zheng 334534c2b290SDavid Sterba out_free_path: 33465d4f98a2SYan Zheng btrfs_free_path(path); 334734c2b290SDavid Sterba out_free_blocks: 3348e1a12670SLiu Bo free_block_list(blocks); 33495d4f98a2SYan Zheng return err; 33505d4f98a2SYan Zheng } 33515d4f98a2SYan Zheng 33525d4f98a2SYan Zheng static noinline_for_stack 3353efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 3354efa56464SYan, Zheng struct file_extent_cluster *cluster) 3355efa56464SYan, Zheng { 3356efa56464SYan, Zheng u64 alloc_hint = 0; 3357efa56464SYan, Zheng u64 start; 3358efa56464SYan, Zheng u64 end; 3359efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 3360efa56464SYan, Zheng u64 num_bytes; 3361efa56464SYan, Zheng int nr = 0; 3362efa56464SYan, Zheng int ret = 0; 3363dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 3364dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 336518513091SWang Xiaoguang u64 cur_offset; 3366364ecf36SQu Wenruo struct extent_changeset *data_reserved = NULL; 3367efa56464SYan, Zheng 3368efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 33695955102cSAl Viro inode_lock(inode); 3370efa56464SYan, Zheng 3371364ecf36SQu Wenruo ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start, 3372dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 3373efa56464SYan, Zheng if (ret) 3374efa56464SYan, Zheng goto out; 3375efa56464SYan, Zheng 337618513091SWang Xiaoguang cur_offset = prealloc_start; 3377efa56464SYan, Zheng while (nr < cluster->nr) { 3378efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 3379efa56464SYan, Zheng if (nr + 1 < cluster->nr) 3380efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 3381efa56464SYan, Zheng else 3382efa56464SYan, Zheng end = cluster->end - offset; 3383efa56464SYan, Zheng 3384d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 3385efa56464SYan, Zheng num_bytes = end + 1 - start; 338618513091SWang Xiaoguang if (cur_offset < start) 3387bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3388bc42bda2SQu Wenruo cur_offset, start - cur_offset); 3389efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 3390efa56464SYan, Zheng num_bytes, num_bytes, 3391efa56464SYan, Zheng end + 1, &alloc_hint); 339218513091SWang Xiaoguang cur_offset = end + 1; 3393d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 3394efa56464SYan, Zheng if (ret) 3395efa56464SYan, Zheng break; 3396efa56464SYan, Zheng nr++; 3397efa56464SYan, Zheng } 339818513091SWang Xiaoguang if (cur_offset < prealloc_end) 3399bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3400bc42bda2SQu Wenruo cur_offset, prealloc_end + 1 - cur_offset); 3401efa56464SYan, Zheng out: 34025955102cSAl Viro inode_unlock(inode); 3403364ecf36SQu Wenruo extent_changeset_free(data_reserved); 3404efa56464SYan, Zheng return ret; 3405efa56464SYan, Zheng } 3406efa56464SYan, Zheng 3407efa56464SYan, Zheng static noinline_for_stack 34080257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 34090257bb82SYan, Zheng u64 block_start) 34105d4f98a2SYan Zheng { 34115d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 34125d4f98a2SYan Zheng struct extent_map *em; 34130257bb82SYan, Zheng int ret = 0; 34145d4f98a2SYan Zheng 3415172ddd60SDavid Sterba em = alloc_extent_map(); 34160257bb82SYan, Zheng if (!em) 34170257bb82SYan, Zheng return -ENOMEM; 34180257bb82SYan, Zheng 34195d4f98a2SYan Zheng em->start = start; 34200257bb82SYan, Zheng em->len = end + 1 - start; 34210257bb82SYan, Zheng em->block_len = em->len; 34220257bb82SYan, Zheng em->block_start = block_start; 34235d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 34245d4f98a2SYan Zheng 3425d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 34265d4f98a2SYan Zheng while (1) { 3427890871beSChris Mason write_lock(&em_tree->lock); 342809a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 3429890871beSChris Mason write_unlock(&em_tree->lock); 34305d4f98a2SYan Zheng if (ret != -EEXIST) { 34315d4f98a2SYan Zheng free_extent_map(em); 34325d4f98a2SYan Zheng break; 34335d4f98a2SYan Zheng } 3434dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 34355d4f98a2SYan Zheng } 3436d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 34370257bb82SYan, Zheng return ret; 34380257bb82SYan, Zheng } 34395d4f98a2SYan Zheng 3440726a3421SQu Wenruo /* 3441726a3421SQu Wenruo * Allow error injection to test balance cancellation 3442726a3421SQu Wenruo */ 3443726a3421SQu Wenruo int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 3444726a3421SQu Wenruo { 3445726a3421SQu Wenruo return atomic_read(&fs_info->balance_cancel_req); 3446726a3421SQu Wenruo } 3447726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 3448726a3421SQu Wenruo 34490257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 34500257bb82SYan, Zheng struct file_extent_cluster *cluster) 34510257bb82SYan, Zheng { 34522ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 34530257bb82SYan, Zheng u64 page_start; 34540257bb82SYan, Zheng u64 page_end; 34550257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 34560257bb82SYan, Zheng unsigned long index; 34570257bb82SYan, Zheng unsigned long last_index; 34580257bb82SYan, Zheng struct page *page; 34590257bb82SYan, Zheng struct file_ra_state *ra; 34603b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 34610257bb82SYan, Zheng int nr = 0; 34620257bb82SYan, Zheng int ret = 0; 34630257bb82SYan, Zheng 34640257bb82SYan, Zheng if (!cluster->nr) 34650257bb82SYan, Zheng return 0; 34660257bb82SYan, Zheng 34670257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 34680257bb82SYan, Zheng if (!ra) 34690257bb82SYan, Zheng return -ENOMEM; 34700257bb82SYan, Zheng 3471efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 34720257bb82SYan, Zheng if (ret) 3473efa56464SYan, Zheng goto out; 34740257bb82SYan, Zheng 34750257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 34760257bb82SYan, Zheng 3477efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 3478efa56464SYan, Zheng cluster->end - offset, cluster->start); 3479efa56464SYan, Zheng if (ret) 3480efa56464SYan, Zheng goto out; 3481efa56464SYan, Zheng 348209cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 348309cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 34840257bb82SYan, Zheng while (index <= last_index) { 34859f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 34869f3db423SNikolay Borisov PAGE_SIZE); 3487efa56464SYan, Zheng if (ret) 3488efa56464SYan, Zheng goto out; 3489efa56464SYan, Zheng 34900257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 34910257bb82SYan, Zheng if (!page) { 34920257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 34930257bb82SYan, Zheng ra, NULL, index, 34940257bb82SYan, Zheng last_index + 1 - index); 3495a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 34963b16a4e3SJosef Bacik mask); 34970257bb82SYan, Zheng if (!page) { 3498691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 349943b18595SQu Wenruo PAGE_SIZE, true); 350044db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 35018702ba93SQu Wenruo PAGE_SIZE); 35020257bb82SYan, Zheng ret = -ENOMEM; 3503efa56464SYan, Zheng goto out; 35040257bb82SYan, Zheng } 35050257bb82SYan, Zheng } 35060257bb82SYan, Zheng 35070257bb82SYan, Zheng if (PageReadahead(page)) { 35080257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 35090257bb82SYan, Zheng ra, NULL, page, index, 35100257bb82SYan, Zheng last_index + 1 - index); 35110257bb82SYan, Zheng } 35120257bb82SYan, Zheng 35130257bb82SYan, Zheng if (!PageUptodate(page)) { 35140257bb82SYan, Zheng btrfs_readpage(NULL, page); 35150257bb82SYan, Zheng lock_page(page); 35160257bb82SYan, Zheng if (!PageUptodate(page)) { 35170257bb82SYan, Zheng unlock_page(page); 351809cbfeafSKirill A. Shutemov put_page(page); 3519691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 352043b18595SQu Wenruo PAGE_SIZE, true); 35218b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 35228702ba93SQu Wenruo PAGE_SIZE); 35230257bb82SYan, Zheng ret = -EIO; 3524efa56464SYan, Zheng goto out; 35250257bb82SYan, Zheng } 35260257bb82SYan, Zheng } 35270257bb82SYan, Zheng 35284eee4fa4SMiao Xie page_start = page_offset(page); 352909cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 35300257bb82SYan, Zheng 3531d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 35320257bb82SYan, Zheng 35330257bb82SYan, Zheng set_page_extent_mapped(page); 35340257bb82SYan, Zheng 35350257bb82SYan, Zheng if (nr < cluster->nr && 35360257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 35370257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 35380257bb82SYan, Zheng page_start, page_end, 3539ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 35400257bb82SYan, Zheng nr++; 35410257bb82SYan, Zheng } 35420257bb82SYan, Zheng 3543765f3cebSNikolay Borisov ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0, 3544330a5827SNikolay Borisov NULL); 3545765f3cebSNikolay Borisov if (ret) { 3546765f3cebSNikolay Borisov unlock_page(page); 3547765f3cebSNikolay Borisov put_page(page); 3548765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 354943b18595SQu Wenruo PAGE_SIZE, true); 3550765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 35518702ba93SQu Wenruo PAGE_SIZE); 3552765f3cebSNikolay Borisov 3553765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 3554765f3cebSNikolay Borisov page_start, page_end, 3555765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 3556765f3cebSNikolay Borisov goto out; 3557765f3cebSNikolay Borisov 3558765f3cebSNikolay Borisov } 35590257bb82SYan, Zheng set_page_dirty(page); 35600257bb82SYan, Zheng 35610257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 3562d0082371SJeff Mahoney page_start, page_end); 35630257bb82SYan, Zheng unlock_page(page); 356409cbfeafSKirill A. Shutemov put_page(page); 35650257bb82SYan, Zheng 35660257bb82SYan, Zheng index++; 35678702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 3568efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 35692ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 35707f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 35717f913c7cSQu Wenruo ret = -ECANCELED; 35727f913c7cSQu Wenruo goto out; 35737f913c7cSQu Wenruo } 35740257bb82SYan, Zheng } 35750257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 3576efa56464SYan, Zheng out: 35770257bb82SYan, Zheng kfree(ra); 35780257bb82SYan, Zheng return ret; 35790257bb82SYan, Zheng } 35800257bb82SYan, Zheng 35810257bb82SYan, Zheng static noinline_for_stack 35820257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 35830257bb82SYan, Zheng struct file_extent_cluster *cluster) 35840257bb82SYan, Zheng { 35850257bb82SYan, Zheng int ret; 35860257bb82SYan, Zheng 35870257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 35880257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 35890257bb82SYan, Zheng if (ret) 35900257bb82SYan, Zheng return ret; 35910257bb82SYan, Zheng cluster->nr = 0; 35920257bb82SYan, Zheng } 35930257bb82SYan, Zheng 35940257bb82SYan, Zheng if (!cluster->nr) 35950257bb82SYan, Zheng cluster->start = extent_key->objectid; 35960257bb82SYan, Zheng else 35970257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 35980257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 35990257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 36000257bb82SYan, Zheng cluster->nr++; 36010257bb82SYan, Zheng 36020257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 36030257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 36040257bb82SYan, Zheng if (ret) 36050257bb82SYan, Zheng return ret; 36060257bb82SYan, Zheng cluster->nr = 0; 36070257bb82SYan, Zheng } 36080257bb82SYan, Zheng return 0; 36095d4f98a2SYan Zheng } 36105d4f98a2SYan Zheng 36115d4f98a2SYan Zheng /* 36125d4f98a2SYan Zheng * helper to add a tree block to the list. 36135d4f98a2SYan Zheng * the major work is getting the generation and level of the block 36145d4f98a2SYan Zheng */ 36155d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 36165d4f98a2SYan Zheng struct btrfs_key *extent_key, 36175d4f98a2SYan Zheng struct btrfs_path *path, 36185d4f98a2SYan Zheng struct rb_root *blocks) 36195d4f98a2SYan Zheng { 36205d4f98a2SYan Zheng struct extent_buffer *eb; 36215d4f98a2SYan Zheng struct btrfs_extent_item *ei; 36225d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 36235d4f98a2SYan Zheng struct tree_block *block; 36245d4f98a2SYan Zheng struct rb_node *rb_node; 36255d4f98a2SYan Zheng u32 item_size; 36265d4f98a2SYan Zheng int level = -1; 36277fdf4b60SWang Shilong u64 generation; 36285d4f98a2SYan Zheng 36295d4f98a2SYan Zheng eb = path->nodes[0]; 36305d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 36315d4f98a2SYan Zheng 36323173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 36333173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 36345d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 36355d4f98a2SYan Zheng struct btrfs_extent_item); 36363173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 36375d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 36385d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 36395d4f98a2SYan Zheng } else { 36403173a18fSJosef Bacik level = (int)extent_key->offset; 36413173a18fSJosef Bacik } 36423173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 36436d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3644ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3645ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3646ba3c2b19SNikolay Borisov return -EINVAL; 36473173a18fSJosef Bacik } else { 36485d4f98a2SYan Zheng BUG(); 36495d4f98a2SYan Zheng } 36505d4f98a2SYan Zheng 3651b3b4aa74SDavid Sterba btrfs_release_path(path); 36525d4f98a2SYan Zheng 36535d4f98a2SYan Zheng BUG_ON(level == -1); 36545d4f98a2SYan Zheng 36555d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 36565d4f98a2SYan Zheng if (!block) 36575d4f98a2SYan Zheng return -ENOMEM; 36585d4f98a2SYan Zheng 36595d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3660da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 36615d4f98a2SYan Zheng block->key.offset = generation; 36625d4f98a2SYan Zheng block->level = level; 36635d4f98a2SYan Zheng block->key_ready = 0; 36645d4f98a2SYan Zheng 36655d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 366643c04fb1SJeff Mahoney if (rb_node) 366743c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, block->bytenr); 36685d4f98a2SYan Zheng 36695d4f98a2SYan Zheng return 0; 36705d4f98a2SYan Zheng } 36715d4f98a2SYan Zheng 36725d4f98a2SYan Zheng /* 36735d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 36745d4f98a2SYan Zheng */ 36755d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 36765d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 36775d4f98a2SYan Zheng struct rb_root *blocks) 36785d4f98a2SYan Zheng { 36790b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36805d4f98a2SYan Zheng struct btrfs_path *path; 36815d4f98a2SYan Zheng struct btrfs_key key; 36825d4f98a2SYan Zheng int ret; 36830b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 36845d4f98a2SYan Zheng 36857476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 36865d4f98a2SYan Zheng return 0; 36875d4f98a2SYan Zheng 36885d4f98a2SYan Zheng if (tree_search(blocks, bytenr)) 36895d4f98a2SYan Zheng return 0; 36905d4f98a2SYan Zheng 36915d4f98a2SYan Zheng path = btrfs_alloc_path(); 36925d4f98a2SYan Zheng if (!path) 36935d4f98a2SYan Zheng return -ENOMEM; 3694aee68ee5SJosef Bacik again: 36955d4f98a2SYan Zheng key.objectid = bytenr; 3696aee68ee5SJosef Bacik if (skinny) { 3697aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3698aee68ee5SJosef Bacik key.offset = (u64)-1; 3699aee68ee5SJosef Bacik } else { 37005d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 37015d4f98a2SYan Zheng key.offset = blocksize; 3702aee68ee5SJosef Bacik } 37035d4f98a2SYan Zheng 37045d4f98a2SYan Zheng path->search_commit_root = 1; 37055d4f98a2SYan Zheng path->skip_locking = 1; 37065d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 37075d4f98a2SYan Zheng if (ret < 0) 37085d4f98a2SYan Zheng goto out; 37095d4f98a2SYan Zheng 3710aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3711aee68ee5SJosef Bacik if (path->slots[0]) { 3712aee68ee5SJosef Bacik path->slots[0]--; 3713aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3714aee68ee5SJosef Bacik path->slots[0]); 37153173a18fSJosef Bacik if (key.objectid == bytenr && 3716aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3717aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3718aee68ee5SJosef Bacik key.offset == blocksize))) 37193173a18fSJosef Bacik ret = 0; 37203173a18fSJosef Bacik } 3721aee68ee5SJosef Bacik 3722aee68ee5SJosef Bacik if (ret) { 3723aee68ee5SJosef Bacik skinny = false; 3724aee68ee5SJosef Bacik btrfs_release_path(path); 3725aee68ee5SJosef Bacik goto again; 3726aee68ee5SJosef Bacik } 3727aee68ee5SJosef Bacik } 3728cdccee99SLiu Bo if (ret) { 3729cdccee99SLiu Bo ASSERT(ret == 1); 3730cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3731cdccee99SLiu Bo btrfs_err(fs_info, 3732cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3733cdccee99SLiu Bo bytenr); 3734cdccee99SLiu Bo WARN_ON(1); 3735cdccee99SLiu Bo ret = -EINVAL; 3736cdccee99SLiu Bo goto out; 3737cdccee99SLiu Bo } 37383173a18fSJosef Bacik 37395d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 37405d4f98a2SYan Zheng out: 37415d4f98a2SYan Zheng btrfs_free_path(path); 37425d4f98a2SYan Zheng return ret; 37435d4f98a2SYan Zheng } 37445d4f98a2SYan Zheng 37450af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 374632da5386SDavid Sterba struct btrfs_block_group *block_group, 37471bbc621eSChris Mason struct inode *inode, 37481bbc621eSChris Mason u64 ino) 37490af3d00bSJosef Bacik { 37500af3d00bSJosef Bacik struct btrfs_key key; 37510af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 37520af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 37530af3d00bSJosef Bacik int ret = 0; 37540af3d00bSJosef Bacik 37550af3d00bSJosef Bacik if (inode) 37560af3d00bSJosef Bacik goto truncate; 37570af3d00bSJosef Bacik 37580af3d00bSJosef Bacik key.objectid = ino; 37590af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 37600af3d00bSJosef Bacik key.offset = 0; 37610af3d00bSJosef Bacik 37624c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 37632e19f1f9SAl Viro if (IS_ERR(inode)) 37640af3d00bSJosef Bacik return -ENOENT; 37650af3d00bSJosef Bacik 37660af3d00bSJosef Bacik truncate: 37672ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 37687b61cd92SMiao Xie &fs_info->global_block_rsv); 37697b61cd92SMiao Xie if (ret) 37707b61cd92SMiao Xie goto out; 37717b61cd92SMiao Xie 37727a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 37730af3d00bSJosef Bacik if (IS_ERR(trans)) { 37743612b495STsutomu Itoh ret = PTR_ERR(trans); 37750af3d00bSJosef Bacik goto out; 37760af3d00bSJosef Bacik } 37770af3d00bSJosef Bacik 377877ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 37790af3d00bSJosef Bacik 37803a45bb20SJeff Mahoney btrfs_end_transaction(trans); 37812ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 37820af3d00bSJosef Bacik out: 37830af3d00bSJosef Bacik iput(inode); 37840af3d00bSJosef Bacik return ret; 37850af3d00bSJosef Bacik } 37860af3d00bSJosef Bacik 37875d4f98a2SYan Zheng /* 378819b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 378919b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 37905d4f98a2SYan Zheng */ 379119b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 379219b546d7SQu Wenruo struct btrfs_block_group *block_group, 379319b546d7SQu Wenruo u64 data_bytenr) 37945d4f98a2SYan Zheng { 379519b546d7SQu Wenruo u64 space_cache_ino; 379619b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 37975d4f98a2SYan Zheng struct btrfs_key key; 379819b546d7SQu Wenruo bool found = false; 379919b546d7SQu Wenruo int i; 38005d4f98a2SYan Zheng int ret; 38015d4f98a2SYan Zheng 380219b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 380319b546d7SQu Wenruo return 0; 38045d4f98a2SYan Zheng 380519b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 380619b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 380719b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 380819b546d7SQu Wenruo continue; 380919b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 381019b546d7SQu Wenruo if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG && 381119b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 381219b546d7SQu Wenruo found = true; 381319b546d7SQu Wenruo space_cache_ino = key.objectid; 381419b546d7SQu Wenruo break; 381519b546d7SQu Wenruo } 381619b546d7SQu Wenruo } 381719b546d7SQu Wenruo if (!found) 381819b546d7SQu Wenruo return -ENOENT; 381919b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 382019b546d7SQu Wenruo space_cache_ino); 38210af3d00bSJosef Bacik return ret; 38225d4f98a2SYan Zheng } 38235d4f98a2SYan Zheng 38245d4f98a2SYan Zheng /* 38252c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 38265d4f98a2SYan Zheng */ 38275d4f98a2SYan Zheng static noinline_for_stack 38285d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 38295d4f98a2SYan Zheng struct btrfs_key *extent_key, 38305d4f98a2SYan Zheng struct btrfs_path *path, 38315d4f98a2SYan Zheng struct rb_root *blocks) 38325d4f98a2SYan Zheng { 383319b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 383419b546d7SQu Wenruo struct ulist *leaves = NULL; 383519b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 383619b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 383719b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3838647f63bdSFilipe David Borba Manana int ret = 0; 38395d4f98a2SYan Zheng 3840b3b4aa74SDavid Sterba btrfs_release_path(path); 384119b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 384219b546d7SQu Wenruo 0, &leaves, NULL, true); 384319b546d7SQu Wenruo if (ret < 0) 384419b546d7SQu Wenruo return ret; 384519b546d7SQu Wenruo 384619b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 384719b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 384819b546d7SQu Wenruo struct extent_buffer *eb; 384919b546d7SQu Wenruo 385019b546d7SQu Wenruo eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL); 385119b546d7SQu Wenruo if (IS_ERR(eb)) { 385219b546d7SQu Wenruo ret = PTR_ERR(eb); 385319b546d7SQu Wenruo break; 385419b546d7SQu Wenruo } 385519b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 385619b546d7SQu Wenruo extent_key->objectid); 385719b546d7SQu Wenruo free_extent_buffer(eb); 385819b546d7SQu Wenruo if (ret < 0) 385919b546d7SQu Wenruo break; 386019b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 386119b546d7SQu Wenruo if (ret < 0) 386219b546d7SQu Wenruo break; 386319b546d7SQu Wenruo } 386419b546d7SQu Wenruo if (ret < 0) 38655d4f98a2SYan Zheng free_block_list(blocks); 386619b546d7SQu Wenruo ulist_free(leaves); 386719b546d7SQu Wenruo return ret; 38685d4f98a2SYan Zheng } 38695d4f98a2SYan Zheng 38705d4f98a2SYan Zheng /* 38712c016dc2SLiu Bo * helper to find next unprocessed extent 38725d4f98a2SYan Zheng */ 38735d4f98a2SYan Zheng static noinline_for_stack 3874147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 38753fd0a558SYan, Zheng struct btrfs_key *extent_key) 38765d4f98a2SYan Zheng { 38770b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 38785d4f98a2SYan Zheng struct btrfs_key key; 38795d4f98a2SYan Zheng struct extent_buffer *leaf; 38805d4f98a2SYan Zheng u64 start, end, last; 38815d4f98a2SYan Zheng int ret; 38825d4f98a2SYan Zheng 3883b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 38845d4f98a2SYan Zheng while (1) { 38855d4f98a2SYan Zheng cond_resched(); 38865d4f98a2SYan Zheng if (rc->search_start >= last) { 38875d4f98a2SYan Zheng ret = 1; 38885d4f98a2SYan Zheng break; 38895d4f98a2SYan Zheng } 38905d4f98a2SYan Zheng 38915d4f98a2SYan Zheng key.objectid = rc->search_start; 38925d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 38935d4f98a2SYan Zheng key.offset = 0; 38945d4f98a2SYan Zheng 38955d4f98a2SYan Zheng path->search_commit_root = 1; 38965d4f98a2SYan Zheng path->skip_locking = 1; 38975d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 38985d4f98a2SYan Zheng 0, 0); 38995d4f98a2SYan Zheng if (ret < 0) 39005d4f98a2SYan Zheng break; 39015d4f98a2SYan Zheng next: 39025d4f98a2SYan Zheng leaf = path->nodes[0]; 39035d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 39045d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 39055d4f98a2SYan Zheng if (ret != 0) 39065d4f98a2SYan Zheng break; 39075d4f98a2SYan Zheng leaf = path->nodes[0]; 39085d4f98a2SYan Zheng } 39095d4f98a2SYan Zheng 39105d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 39115d4f98a2SYan Zheng if (key.objectid >= last) { 39125d4f98a2SYan Zheng ret = 1; 39135d4f98a2SYan Zheng break; 39145d4f98a2SYan Zheng } 39155d4f98a2SYan Zheng 39163173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 39173173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 39183173a18fSJosef Bacik path->slots[0]++; 39193173a18fSJosef Bacik goto next; 39203173a18fSJosef Bacik } 39213173a18fSJosef Bacik 39223173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 39235d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 39245d4f98a2SYan Zheng path->slots[0]++; 39255d4f98a2SYan Zheng goto next; 39265d4f98a2SYan Zheng } 39275d4f98a2SYan Zheng 39283173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 39290b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 39303173a18fSJosef Bacik rc->search_start) { 39313173a18fSJosef Bacik path->slots[0]++; 39323173a18fSJosef Bacik goto next; 39333173a18fSJosef Bacik } 39343173a18fSJosef Bacik 39355d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 39365d4f98a2SYan Zheng key.objectid, &start, &end, 3937e6138876SJosef Bacik EXTENT_DIRTY, NULL); 39385d4f98a2SYan Zheng 39395d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3940b3b4aa74SDavid Sterba btrfs_release_path(path); 39415d4f98a2SYan Zheng rc->search_start = end + 1; 39425d4f98a2SYan Zheng } else { 39433173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 39445d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 39453173a18fSJosef Bacik else 39463173a18fSJosef Bacik rc->search_start = key.objectid + 39470b246afaSJeff Mahoney fs_info->nodesize; 39483fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 39495d4f98a2SYan Zheng return 0; 39505d4f98a2SYan Zheng } 39515d4f98a2SYan Zheng } 3952b3b4aa74SDavid Sterba btrfs_release_path(path); 39535d4f98a2SYan Zheng return ret; 39545d4f98a2SYan Zheng } 39555d4f98a2SYan Zheng 39565d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 39575d4f98a2SYan Zheng { 39585d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39597585717fSChris Mason 39607585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 39615d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 39627585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 39635d4f98a2SYan Zheng } 39645d4f98a2SYan Zheng 39655d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 39665d4f98a2SYan Zheng { 39675d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39687585717fSChris Mason 39697585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 39705d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 39717585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 39725d4f98a2SYan Zheng } 39735d4f98a2SYan Zheng 39745d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 39755d4f98a2SYan Zheng { 39765d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 39775d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 39785d4f98a2SYan Zheng return 1; 39795d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 39805d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 39815d4f98a2SYan Zheng return 1; 39825d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 39835d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 39845d4f98a2SYan Zheng return 1; 39855d4f98a2SYan Zheng return 0; 39865d4f98a2SYan Zheng } 39875d4f98a2SYan Zheng 39883fd0a558SYan, Zheng static noinline_for_stack 39893fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 39903fd0a558SYan, Zheng { 39913fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3992ac2fabacSJosef Bacik int ret; 39933fd0a558SYan, Zheng 39942ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 399566d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 39963fd0a558SYan, Zheng if (!rc->block_rsv) 39973fd0a558SYan, Zheng return -ENOMEM; 39983fd0a558SYan, Zheng 39993fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 4000b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 40013fd0a558SYan, Zheng rc->extents_found = 0; 40023fd0a558SYan, Zheng rc->nodes_relocated = 0; 40033fd0a558SYan, Zheng rc->merging_rsv_size = 0; 40040647bf56SWang Shilong rc->reserved_bytes = 0; 4005da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 40060647bf56SWang Shilong RELOCATION_RESERVED_NODES; 4007ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 4008ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 4009ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 4010ac2fabacSJosef Bacik if (ret) 4011ac2fabacSJosef Bacik return ret; 40123fd0a558SYan, Zheng 40133fd0a558SYan, Zheng rc->create_reloc_tree = 1; 40143fd0a558SYan, Zheng set_reloc_control(rc); 40153fd0a558SYan, Zheng 40167a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 401728818947SLiu Bo if (IS_ERR(trans)) { 401828818947SLiu Bo unset_reloc_control(rc); 401928818947SLiu Bo /* 402028818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 402128818947SLiu Bo * cleanup. And callers are responsible to free the above 402228818947SLiu Bo * block rsv. 402328818947SLiu Bo */ 402428818947SLiu Bo return PTR_ERR(trans); 402528818947SLiu Bo } 40263a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 40273fd0a558SYan, Zheng return 0; 40283fd0a558SYan, Zheng } 402976dda93cSYan, Zheng 40305d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 40315d4f98a2SYan Zheng { 40322ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 40335d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 40345d4f98a2SYan Zheng struct btrfs_key key; 40355d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 40365d4f98a2SYan Zheng struct btrfs_path *path; 40375d4f98a2SYan Zheng struct btrfs_extent_item *ei; 40385d4f98a2SYan Zheng u64 flags; 40395d4f98a2SYan Zheng u32 item_size; 40405d4f98a2SYan Zheng int ret; 40415d4f98a2SYan Zheng int err = 0; 4042c87f08caSChris Mason int progress = 0; 40435d4f98a2SYan Zheng 40445d4f98a2SYan Zheng path = btrfs_alloc_path(); 40453fd0a558SYan, Zheng if (!path) 40465d4f98a2SYan Zheng return -ENOMEM; 4047e4058b54SDavid Sterba path->reada = READA_FORWARD; 40483fd0a558SYan, Zheng 40493fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 40503fd0a558SYan, Zheng if (ret) { 40513fd0a558SYan, Zheng err = ret; 40523fd0a558SYan, Zheng goto out_free; 40532423fdfbSJiri Slaby } 40545d4f98a2SYan Zheng 40555d4f98a2SYan Zheng while (1) { 40560647bf56SWang Shilong rc->reserved_bytes = 0; 40570647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 40580647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 40590647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 40600647bf56SWang Shilong if (ret) { 40610647bf56SWang Shilong err = ret; 40620647bf56SWang Shilong break; 40630647bf56SWang Shilong } 4064c87f08caSChris Mason progress++; 4065a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 40660f788c58SLiu Bo if (IS_ERR(trans)) { 40670f788c58SLiu Bo err = PTR_ERR(trans); 40680f788c58SLiu Bo trans = NULL; 40690f788c58SLiu Bo break; 40700f788c58SLiu Bo } 4071c87f08caSChris Mason restart: 40723fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 40733a45bb20SJeff Mahoney btrfs_end_transaction(trans); 407442a657f5SPan Bian trans = NULL; 40753fd0a558SYan, Zheng continue; 40763fd0a558SYan, Zheng } 40773fd0a558SYan, Zheng 4078147d256eSZhaolei ret = find_next_extent(rc, path, &key); 40795d4f98a2SYan Zheng if (ret < 0) 40805d4f98a2SYan Zheng err = ret; 40815d4f98a2SYan Zheng if (ret != 0) 40825d4f98a2SYan Zheng break; 40835d4f98a2SYan Zheng 40845d4f98a2SYan Zheng rc->extents_found++; 40855d4f98a2SYan Zheng 40865d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 40875d4f98a2SYan Zheng struct btrfs_extent_item); 40883fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 40895d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 40905d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 40915d4f98a2SYan Zheng ret = check_extent_flags(flags); 40925d4f98a2SYan Zheng BUG_ON(ret); 40936d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 4094ba3c2b19SNikolay Borisov err = -EINVAL; 4095ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 4096ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 4097ba3c2b19SNikolay Borisov break; 40985d4f98a2SYan Zheng } else { 40995d4f98a2SYan Zheng BUG(); 41005d4f98a2SYan Zheng } 41015d4f98a2SYan Zheng 41025d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 41035d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 41045d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 41055d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 41065d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 41075d4f98a2SYan Zheng } else { 4108b3b4aa74SDavid Sterba btrfs_release_path(path); 41095d4f98a2SYan Zheng ret = 0; 41105d4f98a2SYan Zheng } 41115d4f98a2SYan Zheng if (ret < 0) { 41123fd0a558SYan, Zheng err = ret; 41135d4f98a2SYan Zheng break; 41145d4f98a2SYan Zheng } 41155d4f98a2SYan Zheng 41165d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 41175d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 41185d4f98a2SYan Zheng if (ret < 0) { 41193fd0a558SYan, Zheng if (ret != -EAGAIN) { 41205d4f98a2SYan Zheng err = ret; 41215d4f98a2SYan Zheng break; 41225d4f98a2SYan Zheng } 41233fd0a558SYan, Zheng rc->extents_found--; 41243fd0a558SYan, Zheng rc->search_start = key.objectid; 41253fd0a558SYan, Zheng } 41265d4f98a2SYan Zheng } 41275d4f98a2SYan Zheng 41283a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 41292ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 41303fd0a558SYan, Zheng trans = NULL; 41315d4f98a2SYan Zheng 41325d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 41335d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 41345d4f98a2SYan Zheng rc->found_file_extent = 1; 41350257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 41363fd0a558SYan, Zheng &key, &rc->cluster); 41375d4f98a2SYan Zheng if (ret < 0) { 41385d4f98a2SYan Zheng err = ret; 41395d4f98a2SYan Zheng break; 41405d4f98a2SYan Zheng } 41415d4f98a2SYan Zheng } 4142f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 4143f31ea088SQu Wenruo err = -ECANCELED; 4144f31ea088SQu Wenruo break; 4145f31ea088SQu Wenruo } 41465d4f98a2SYan Zheng } 4147c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 414843a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 41499689457bSShilong Wang if (ret == 1) { 4150c87f08caSChris Mason err = 0; 4151c87f08caSChris Mason progress = 0; 4152c87f08caSChris Mason goto restart; 4153c87f08caSChris Mason } 4154c87f08caSChris Mason } 41553fd0a558SYan, Zheng 4156b3b4aa74SDavid Sterba btrfs_release_path(path); 415791166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 41585d4f98a2SYan Zheng 41595d4f98a2SYan Zheng if (trans) { 41603a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 41612ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 41625d4f98a2SYan Zheng } 41635d4f98a2SYan Zheng 41640257bb82SYan, Zheng if (!err) { 41653fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 41663fd0a558SYan, Zheng &rc->cluster); 41670257bb82SYan, Zheng if (ret < 0) 41680257bb82SYan, Zheng err = ret; 41690257bb82SYan, Zheng } 41700257bb82SYan, Zheng 41713fd0a558SYan, Zheng rc->create_reloc_tree = 0; 41723fd0a558SYan, Zheng set_reloc_control(rc); 41730257bb82SYan, Zheng 41743fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 417563f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 41765d4f98a2SYan Zheng 41777f913c7cSQu Wenruo /* 41787f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 41797f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 41807f913c7cSQu Wenruo * 41817f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 41827f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 41837f913c7cSQu Wenruo * merge_reloc_roots() 41847f913c7cSQu Wenruo */ 41853fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 41865d4f98a2SYan Zheng 41875d4f98a2SYan Zheng merge_reloc_roots(rc); 41885d4f98a2SYan Zheng 41893fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 41905d4f98a2SYan Zheng unset_reloc_control(rc); 419163f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 41925d4f98a2SYan Zheng 41935d4f98a2SYan Zheng /* get rid of pinned extents */ 41947a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 419562b99540SQu Wenruo if (IS_ERR(trans)) { 41963612b495STsutomu Itoh err = PTR_ERR(trans); 419762b99540SQu Wenruo goto out_free; 419862b99540SQu Wenruo } 41993a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 42006217b0faSJosef Bacik out_free: 4201d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4202d2311e69SQu Wenruo if (ret < 0 && !err) 4203d2311e69SQu Wenruo err = ret; 42042ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 42053fd0a558SYan, Zheng btrfs_free_path(path); 42065d4f98a2SYan Zheng return err; 42075d4f98a2SYan Zheng } 42085d4f98a2SYan Zheng 42095d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 42100257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 42115d4f98a2SYan Zheng { 42125d4f98a2SYan Zheng struct btrfs_path *path; 42135d4f98a2SYan Zheng struct btrfs_inode_item *item; 42145d4f98a2SYan Zheng struct extent_buffer *leaf; 42155d4f98a2SYan Zheng int ret; 42165d4f98a2SYan Zheng 42175d4f98a2SYan Zheng path = btrfs_alloc_path(); 42185d4f98a2SYan Zheng if (!path) 42195d4f98a2SYan Zheng return -ENOMEM; 42205d4f98a2SYan Zheng 42215d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 42225d4f98a2SYan Zheng if (ret) 42235d4f98a2SYan Zheng goto out; 42245d4f98a2SYan Zheng 42255d4f98a2SYan Zheng leaf = path->nodes[0]; 42265d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 4227b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 42285d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 42290257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 42305d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 42313fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 42323fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 42335d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 42345d4f98a2SYan Zheng out: 42355d4f98a2SYan Zheng btrfs_free_path(path); 42365d4f98a2SYan Zheng return ret; 42375d4f98a2SYan Zheng } 42385d4f98a2SYan Zheng 42395d4f98a2SYan Zheng /* 42405d4f98a2SYan Zheng * helper to create inode for data relocation. 42415d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 42425d4f98a2SYan Zheng */ 42433fd0a558SYan, Zheng static noinline_for_stack 42443fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 424532da5386SDavid Sterba struct btrfs_block_group *group) 42465d4f98a2SYan Zheng { 42475d4f98a2SYan Zheng struct inode *inode = NULL; 42485d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 42495d4f98a2SYan Zheng struct btrfs_root *root; 42505d4f98a2SYan Zheng struct btrfs_key key; 42514624900dSZhaolei u64 objectid; 42525d4f98a2SYan Zheng int err = 0; 42535d4f98a2SYan Zheng 42545d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 42555d4f98a2SYan Zheng if (IS_ERR(root)) 42565d4f98a2SYan Zheng return ERR_CAST(root); 42575d4f98a2SYan Zheng 4258a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 425976deacf0SJosef Bacik if (IS_ERR(trans)) { 426000246528SJosef Bacik btrfs_put_root(root); 42613fd0a558SYan, Zheng return ERR_CAST(trans); 426276deacf0SJosef Bacik } 42635d4f98a2SYan Zheng 4264581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 42655d4f98a2SYan Zheng if (err) 42665d4f98a2SYan Zheng goto out; 42675d4f98a2SYan Zheng 42680257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 42695d4f98a2SYan Zheng BUG_ON(err); 42705d4f98a2SYan Zheng 42715d4f98a2SYan Zheng key.objectid = objectid; 42725d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 42735d4f98a2SYan Zheng key.offset = 0; 42744c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 42752e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 4276b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 42775d4f98a2SYan Zheng 427873f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 42795d4f98a2SYan Zheng out: 428000246528SJosef Bacik btrfs_put_root(root); 42813a45bb20SJeff Mahoney btrfs_end_transaction(trans); 42822ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 42835d4f98a2SYan Zheng if (err) { 42845d4f98a2SYan Zheng if (inode) 42855d4f98a2SYan Zheng iput(inode); 42865d4f98a2SYan Zheng inode = ERR_PTR(err); 42875d4f98a2SYan Zheng } 42885d4f98a2SYan Zheng return inode; 42895d4f98a2SYan Zheng } 42905d4f98a2SYan Zheng 4291c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 42923fd0a558SYan, Zheng { 42933fd0a558SYan, Zheng struct reloc_control *rc; 42943fd0a558SYan, Zheng 42953fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 42963fd0a558SYan, Zheng if (!rc) 42973fd0a558SYan, Zheng return NULL; 42983fd0a558SYan, Zheng 42993fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 4300d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 43012433bea5SQu Wenruo backref_cache_init(fs_info, &rc->backref_cache, 1); 43023fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 430343eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 430443eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 43053fd0a558SYan, Zheng return rc; 43063fd0a558SYan, Zheng } 43073fd0a558SYan, Zheng 43081a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 43091a0afa0eSJosef Bacik { 43101a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 43111a0afa0eSJosef Bacik 43121a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 43131a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 43141a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 43151a0afa0eSJosef Bacik kfree(node); 43161a0afa0eSJosef Bacik 43171a0afa0eSJosef Bacik kfree(rc); 43181a0afa0eSJosef Bacik } 43191a0afa0eSJosef Bacik 43205d4f98a2SYan Zheng /* 4321ebce0e01SAdam Borowski * Print the block group being relocated 4322ebce0e01SAdam Borowski */ 4323ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 432432da5386SDavid Sterba struct btrfs_block_group *block_group) 4325ebce0e01SAdam Borowski { 4326f89e09cfSAnand Jain char buf[128] = {'\0'}; 4327ebce0e01SAdam Borowski 4328f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 4329ebce0e01SAdam Borowski 4330ebce0e01SAdam Borowski btrfs_info(fs_info, 4331ebce0e01SAdam Borowski "relocating block group %llu flags %s", 4332b3470b5dSDavid Sterba block_group->start, buf); 4333ebce0e01SAdam Borowski } 4334ebce0e01SAdam Borowski 4335430640e3SQu Wenruo static const char *stage_to_string(int stage) 4336430640e3SQu Wenruo { 4337430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 4338430640e3SQu Wenruo return "move data extents"; 4339430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 4340430640e3SQu Wenruo return "update data pointers"; 4341430640e3SQu Wenruo return "unknown"; 4342430640e3SQu Wenruo } 4343430640e3SQu Wenruo 4344ebce0e01SAdam Borowski /* 43455d4f98a2SYan Zheng * function to relocate all extents in a block group. 43465d4f98a2SYan Zheng */ 43476bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 43485d4f98a2SYan Zheng { 434932da5386SDavid Sterba struct btrfs_block_group *bg; 43506bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 43515d4f98a2SYan Zheng struct reloc_control *rc; 43520af3d00bSJosef Bacik struct inode *inode; 43530af3d00bSJosef Bacik struct btrfs_path *path; 43545d4f98a2SYan Zheng int ret; 4355f0486c68SYan, Zheng int rw = 0; 43565d4f98a2SYan Zheng int err = 0; 43575d4f98a2SYan Zheng 4358eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 4359eede2bf3SOmar Sandoval if (!bg) 4360eede2bf3SOmar Sandoval return -ENOENT; 4361eede2bf3SOmar Sandoval 4362eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 4363eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 4364eede2bf3SOmar Sandoval return -ETXTBSY; 4365eede2bf3SOmar Sandoval } 4366eede2bf3SOmar Sandoval 4367c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 4368eede2bf3SOmar Sandoval if (!rc) { 4369eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 43705d4f98a2SYan Zheng return -ENOMEM; 4371eede2bf3SOmar Sandoval } 43725d4f98a2SYan Zheng 4373f0486c68SYan, Zheng rc->extent_root = extent_root; 4374eede2bf3SOmar Sandoval rc->block_group = bg; 43755d4f98a2SYan Zheng 4376b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 4377f0486c68SYan, Zheng if (ret) { 4378f0486c68SYan, Zheng err = ret; 4379f0486c68SYan, Zheng goto out; 4380f0486c68SYan, Zheng } 4381f0486c68SYan, Zheng rw = 1; 4382f0486c68SYan, Zheng 43830af3d00bSJosef Bacik path = btrfs_alloc_path(); 43840af3d00bSJosef Bacik if (!path) { 43850af3d00bSJosef Bacik err = -ENOMEM; 43860af3d00bSJosef Bacik goto out; 43870af3d00bSJosef Bacik } 43880af3d00bSJosef Bacik 43897949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 43900af3d00bSJosef Bacik btrfs_free_path(path); 43910af3d00bSJosef Bacik 43920af3d00bSJosef Bacik if (!IS_ERR(inode)) 43931bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 43940af3d00bSJosef Bacik else 43950af3d00bSJosef Bacik ret = PTR_ERR(inode); 43960af3d00bSJosef Bacik 43970af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 43980af3d00bSJosef Bacik err = ret; 43990af3d00bSJosef Bacik goto out; 44000af3d00bSJosef Bacik } 44010af3d00bSJosef Bacik 44025d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 44035d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 44045d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 44055d4f98a2SYan Zheng rc->data_inode = NULL; 44065d4f98a2SYan Zheng goto out; 44075d4f98a2SYan Zheng } 44085d4f98a2SYan Zheng 44090b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 44105d4f98a2SYan Zheng 44119cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 4412f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 44136374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 4414b3470b5dSDavid Sterba rc->block_group->start, 4415b3470b5dSDavid Sterba rc->block_group->length); 44165d4f98a2SYan Zheng 44175d4f98a2SYan Zheng while (1) { 4418430640e3SQu Wenruo int finishes_stage; 4419430640e3SQu Wenruo 442076dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 44215d4f98a2SYan Zheng ret = relocate_block_group(rc); 442276dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 4423ff612ba7SJosef Bacik if (ret < 0) 44245d4f98a2SYan Zheng err = ret; 4425ff612ba7SJosef Bacik 4426430640e3SQu Wenruo finishes_stage = rc->stage; 4427ff612ba7SJosef Bacik /* 4428ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 4429ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 4430ff612ba7SJosef Bacik * different block group we could end up hitting the 4431ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 4432ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 4433ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 4434ff612ba7SJosef Bacik * out of the loop if we hit an error. 4435ff612ba7SJosef Bacik */ 4436ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 4437ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 4438ff612ba7SJosef Bacik (u64)-1); 4439ff612ba7SJosef Bacik if (ret) 4440ff612ba7SJosef Bacik err = ret; 4441ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 4442ff612ba7SJosef Bacik 0, -1); 4443ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 44445d4f98a2SYan Zheng } 44455d4f98a2SYan Zheng 4446ff612ba7SJosef Bacik if (err < 0) 4447ff612ba7SJosef Bacik goto out; 4448ff612ba7SJosef Bacik 44495d4f98a2SYan Zheng if (rc->extents_found == 0) 44505d4f98a2SYan Zheng break; 44515d4f98a2SYan Zheng 4452430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 4453430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 44545d4f98a2SYan Zheng } 44555d4f98a2SYan Zheng 44565d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 44575d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 4458bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 44595d4f98a2SYan Zheng out: 4460f0486c68SYan, Zheng if (err && rw) 44612ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 44625d4f98a2SYan Zheng iput(rc->data_inode); 44635d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 44641a0afa0eSJosef Bacik free_reloc_control(rc); 44655d4f98a2SYan Zheng return err; 44665d4f98a2SYan Zheng } 44675d4f98a2SYan Zheng 446876dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 446976dda93cSYan, Zheng { 44700b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 447176dda93cSYan, Zheng struct btrfs_trans_handle *trans; 447279787eaaSJeff Mahoney int ret, err; 447376dda93cSYan, Zheng 44740b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 447579787eaaSJeff Mahoney if (IS_ERR(trans)) 447679787eaaSJeff Mahoney return PTR_ERR(trans); 447776dda93cSYan, Zheng 447876dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 447976dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 448076dda93cSYan, Zheng root->root_item.drop_level = 0; 448176dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 44820b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 448376dda93cSYan, Zheng &root->root_key, &root->root_item); 448476dda93cSYan, Zheng 44853a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 448679787eaaSJeff Mahoney if (err) 448779787eaaSJeff Mahoney return err; 448879787eaaSJeff Mahoney return ret; 448976dda93cSYan, Zheng } 449076dda93cSYan, Zheng 44915d4f98a2SYan Zheng /* 44925d4f98a2SYan Zheng * recover relocation interrupted by system crash. 44935d4f98a2SYan Zheng * 44945d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 44955d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 44965d4f98a2SYan Zheng */ 44975d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 44985d4f98a2SYan Zheng { 44990b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 45005d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 45015d4f98a2SYan Zheng struct btrfs_key key; 45025d4f98a2SYan Zheng struct btrfs_root *fs_root; 45035d4f98a2SYan Zheng struct btrfs_root *reloc_root; 45045d4f98a2SYan Zheng struct btrfs_path *path; 45055d4f98a2SYan Zheng struct extent_buffer *leaf; 45065d4f98a2SYan Zheng struct reloc_control *rc = NULL; 45075d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 45085d4f98a2SYan Zheng int ret; 45095d4f98a2SYan Zheng int err = 0; 45105d4f98a2SYan Zheng 45115d4f98a2SYan Zheng path = btrfs_alloc_path(); 45125d4f98a2SYan Zheng if (!path) 45135d4f98a2SYan Zheng return -ENOMEM; 4514e4058b54SDavid Sterba path->reada = READA_BACK; 45155d4f98a2SYan Zheng 45165d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 45175d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 45185d4f98a2SYan Zheng key.offset = (u64)-1; 45195d4f98a2SYan Zheng 45205d4f98a2SYan Zheng while (1) { 45210b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 45225d4f98a2SYan Zheng path, 0, 0); 45235d4f98a2SYan Zheng if (ret < 0) { 45245d4f98a2SYan Zheng err = ret; 45255d4f98a2SYan Zheng goto out; 45265d4f98a2SYan Zheng } 45275d4f98a2SYan Zheng if (ret > 0) { 45285d4f98a2SYan Zheng if (path->slots[0] == 0) 45295d4f98a2SYan Zheng break; 45305d4f98a2SYan Zheng path->slots[0]--; 45315d4f98a2SYan Zheng } 45325d4f98a2SYan Zheng leaf = path->nodes[0]; 45335d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4534b3b4aa74SDavid Sterba btrfs_release_path(path); 45355d4f98a2SYan Zheng 45365d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 45375d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 45385d4f98a2SYan Zheng break; 45395d4f98a2SYan Zheng 45403dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 45415d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 45425d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 45435d4f98a2SYan Zheng goto out; 45445d4f98a2SYan Zheng } 45455d4f98a2SYan Zheng 45463dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 45475d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 45485d4f98a2SYan Zheng 45495d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 45500b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, 45515d4f98a2SYan Zheng reloc_root->root_key.offset); 45525d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 455376dda93cSYan, Zheng ret = PTR_ERR(fs_root); 455476dda93cSYan, Zheng if (ret != -ENOENT) { 455576dda93cSYan, Zheng err = ret; 45565d4f98a2SYan Zheng goto out; 45575d4f98a2SYan Zheng } 455879787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 455979787eaaSJeff Mahoney if (ret < 0) { 456079787eaaSJeff Mahoney err = ret; 456179787eaaSJeff Mahoney goto out; 456279787eaaSJeff Mahoney } 4563932fd26dSJosef Bacik } else { 456400246528SJosef Bacik btrfs_put_root(fs_root); 456576dda93cSYan, Zheng } 45665d4f98a2SYan Zheng } 45675d4f98a2SYan Zheng 45685d4f98a2SYan Zheng if (key.offset == 0) 45695d4f98a2SYan Zheng break; 45705d4f98a2SYan Zheng 45715d4f98a2SYan Zheng key.offset--; 45725d4f98a2SYan Zheng } 4573b3b4aa74SDavid Sterba btrfs_release_path(path); 45745d4f98a2SYan Zheng 45755d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 45765d4f98a2SYan Zheng goto out; 45775d4f98a2SYan Zheng 4578c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 45795d4f98a2SYan Zheng if (!rc) { 45805d4f98a2SYan Zheng err = -ENOMEM; 45815d4f98a2SYan Zheng goto out; 45825d4f98a2SYan Zheng } 45835d4f98a2SYan Zheng 45840b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 45855d4f98a2SYan Zheng 45865d4f98a2SYan Zheng set_reloc_control(rc); 45875d4f98a2SYan Zheng 45887a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 45893612b495STsutomu Itoh if (IS_ERR(trans)) { 45903612b495STsutomu Itoh err = PTR_ERR(trans); 4591fb2d83eeSJosef Bacik goto out_unset; 45923612b495STsutomu Itoh } 45933fd0a558SYan, Zheng 45943fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 45953fd0a558SYan, Zheng 45965d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 45975d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 45985d4f98a2SYan Zheng struct btrfs_root, root_list); 45995d4f98a2SYan Zheng list_del(&reloc_root->root_list); 46005d4f98a2SYan Zheng 46015d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 46025d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 46035d4f98a2SYan Zheng &rc->reloc_roots); 46045d4f98a2SYan Zheng continue; 46055d4f98a2SYan Zheng } 46065d4f98a2SYan Zheng 46070b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, reloc_root->root_key.offset); 460879787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 460979787eaaSJeff Mahoney err = PTR_ERR(fs_root); 4610ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 46111402d17dSXiyu Yang btrfs_end_transaction(trans); 4612fb2d83eeSJosef Bacik goto out_unset; 461379787eaaSJeff Mahoney } 46145d4f98a2SYan Zheng 4615ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 461679787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 4617f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 461800246528SJosef Bacik btrfs_put_root(fs_root); 46195d4f98a2SYan Zheng } 46205d4f98a2SYan Zheng 46213a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 462279787eaaSJeff Mahoney if (err) 4623fb2d83eeSJosef Bacik goto out_unset; 46245d4f98a2SYan Zheng 46255d4f98a2SYan Zheng merge_reloc_roots(rc); 46265d4f98a2SYan Zheng 46275d4f98a2SYan Zheng unset_reloc_control(rc); 46285d4f98a2SYan Zheng 46297a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 463062b99540SQu Wenruo if (IS_ERR(trans)) { 46313612b495STsutomu Itoh err = PTR_ERR(trans); 46326217b0faSJosef Bacik goto out_clean; 463362b99540SQu Wenruo } 46343a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 46356217b0faSJosef Bacik out_clean: 4636d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4637d2311e69SQu Wenruo if (ret < 0 && !err) 4638d2311e69SQu Wenruo err = ret; 4639fb2d83eeSJosef Bacik out_unset: 4640fb2d83eeSJosef Bacik unset_reloc_control(rc); 46411a0afa0eSJosef Bacik free_reloc_control(rc); 46423612b495STsutomu Itoh out: 4643aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 4644aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4645aca1bba6SLiu Bo 46465d4f98a2SYan Zheng btrfs_free_path(path); 46475d4f98a2SYan Zheng 46485d4f98a2SYan Zheng if (err == 0) { 46495d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 46500b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 4651932fd26dSJosef Bacik if (IS_ERR(fs_root)) { 46525d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4653932fd26dSJosef Bacik } else { 465466b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 465500246528SJosef Bacik btrfs_put_root(fs_root); 4656932fd26dSJosef Bacik } 4657932fd26dSJosef Bacik } 46585d4f98a2SYan Zheng return err; 46595d4f98a2SYan Zheng } 46605d4f98a2SYan Zheng 46615d4f98a2SYan Zheng /* 46625d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 46635d4f98a2SYan Zheng * 46645d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 46655d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 46665d4f98a2SYan Zheng */ 46675d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 46685d4f98a2SYan Zheng { 46690b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 46705d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 46715d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 46725d4f98a2SYan Zheng int ret; 46735d4f98a2SYan Zheng u64 disk_bytenr; 46744577b014SJosef Bacik u64 new_bytenr; 46755d4f98a2SYan Zheng LIST_HEAD(list); 46765d4f98a2SYan Zheng 46775d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4678bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 46795d4f98a2SYan Zheng 46805d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 46810b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4682a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 468379787eaaSJeff Mahoney if (ret) 468479787eaaSJeff Mahoney goto out; 46855d4f98a2SYan Zheng 46865d4f98a2SYan Zheng while (!list_empty(&list)) { 46875d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 46885d4f98a2SYan Zheng list_del_init(&sums->list); 46895d4f98a2SYan Zheng 46904577b014SJosef Bacik /* 46914577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 46924577b014SJosef Bacik * We need to do this because we will read in entire prealloc 46934577b014SJosef Bacik * extents but we may have written to say the middle of the 46944577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 46954577b014SJosef Bacik * the right disk offset. 46964577b014SJosef Bacik * 46974577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 46984577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 46994577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 47004577b014SJosef Bacik * disk length. 47014577b014SJosef Bacik */ 4702bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 47034577b014SJosef Bacik sums->bytenr = new_bytenr; 47045d4f98a2SYan Zheng 4705f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 47065d4f98a2SYan Zheng } 470779787eaaSJeff Mahoney out: 47085d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4709411fc6bcSAndi Kleen return ret; 47105d4f98a2SYan Zheng } 47113fd0a558SYan, Zheng 471283d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 47133fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 47143fd0a558SYan, Zheng struct extent_buffer *cow) 47153fd0a558SYan, Zheng { 47160b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 47173fd0a558SYan, Zheng struct reloc_control *rc; 47183fd0a558SYan, Zheng struct backref_node *node; 47193fd0a558SYan, Zheng int first_cow = 0; 47203fd0a558SYan, Zheng int level; 472183d4cfd4SJosef Bacik int ret = 0; 47223fd0a558SYan, Zheng 47230b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 47243fd0a558SYan, Zheng if (!rc) 472583d4cfd4SJosef Bacik return 0; 47263fd0a558SYan, Zheng 47273fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 47283fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 47293fd0a558SYan, Zheng 47303fd0a558SYan, Zheng level = btrfs_header_level(buf); 47313fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 47323fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 47333fd0a558SYan, Zheng first_cow = 1; 47343fd0a558SYan, Zheng 47353fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 47363fd0a558SYan, Zheng rc->create_reloc_tree) { 47373fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 47383fd0a558SYan, Zheng 47393fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 47403fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 47413fd0a558SYan, Zheng node->new_bytenr != buf->start); 47423fd0a558SYan, Zheng 47433fd0a558SYan, Zheng drop_node_buffer(node); 474467439dadSDavid Sterba atomic_inc(&cow->refs); 47453fd0a558SYan, Zheng node->eb = cow; 47463fd0a558SYan, Zheng node->new_bytenr = cow->start; 47473fd0a558SYan, Zheng 47483fd0a558SYan, Zheng if (!node->pending) { 47493fd0a558SYan, Zheng list_move_tail(&node->list, 47503fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 47513fd0a558SYan, Zheng node->pending = 1; 47523fd0a558SYan, Zheng } 47533fd0a558SYan, Zheng 47543fd0a558SYan, Zheng if (first_cow) 47559569cc20SQu Wenruo mark_block_processed(rc, node); 47563fd0a558SYan, Zheng 47573fd0a558SYan, Zheng if (first_cow && level > 0) 47583fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 47593fd0a558SYan, Zheng } 47603fd0a558SYan, Zheng 476183d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 47623fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 476383d4cfd4SJosef Bacik return ret; 47643fd0a558SYan, Zheng } 47653fd0a558SYan, Zheng 47663fd0a558SYan, Zheng /* 47673fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 476801327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 47693fd0a558SYan, Zheng */ 4770147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 47713fd0a558SYan, Zheng u64 *bytes_to_reserve) 47723fd0a558SYan, Zheng { 477310995c04SQu Wenruo struct btrfs_root *root = pending->root; 477410995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 47753fd0a558SYan, Zheng 47766282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 47773fd0a558SYan, Zheng return; 47783fd0a558SYan, Zheng 47793fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 47803fd0a558SYan, Zheng return; 47813fd0a558SYan, Zheng 47823fd0a558SYan, Zheng root = root->reloc_root; 47833fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 47843fd0a558SYan, Zheng /* 47853fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 47863fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 47873fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 47883fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 47893fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 47903fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 47913fd0a558SYan, Zheng * tree may use more space than it frees. so we need 47923fd0a558SYan, Zheng * reserve extra space. 47933fd0a558SYan, Zheng */ 47943fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 47953fd0a558SYan, Zheng } 47963fd0a558SYan, Zheng 47973fd0a558SYan, Zheng /* 47983fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 47993fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4800f44deb74SJosef Bacik * 4801f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4802f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4803f44deb74SJosef Bacik * rc->reloc_roots. 48043fd0a558SYan, Zheng */ 480549b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 48063fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 48073fd0a558SYan, Zheng { 48083fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 48093fd0a558SYan, Zheng struct btrfs_root *reloc_root; 48103fd0a558SYan, Zheng struct btrfs_root *new_root; 481110995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 48123fd0a558SYan, Zheng int ret; 48133fd0a558SYan, Zheng 48146282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 481549b25e05SJeff Mahoney return 0; 48163fd0a558SYan, Zheng 48173fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 48183fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 48193fd0a558SYan, Zheng 48203fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 48213fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 48223fd0a558SYan, Zheng rc->block_rsv, 48233a584174SLu Fengqi rc->nodes_relocated, true); 482449b25e05SJeff Mahoney if (ret) 482549b25e05SJeff Mahoney return ret; 48263fd0a558SYan, Zheng } 48273fd0a558SYan, Zheng 48283fd0a558SYan, Zheng new_root = pending->snap; 48293fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 48303fd0a558SYan, Zheng new_root->root_key.objectid); 483149b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 483249b25e05SJeff Mahoney return PTR_ERR(reloc_root); 48333fd0a558SYan, Zheng 4834ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4835ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4836f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 48373fd0a558SYan, Zheng 483849b25e05SJeff Mahoney if (rc->create_reloc_tree) 48393fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 484049b25e05SJeff Mahoney return ret; 48413fd0a558SYan, Zheng } 4842