1c1d7c514SDavid Sterba // SPDX-License-Identifier: GPL-2.0 25d4f98a2SYan Zheng /* 35d4f98a2SYan Zheng * Copyright (C) 2009 Oracle. All rights reserved. 45d4f98a2SYan Zheng */ 55d4f98a2SYan Zheng 65d4f98a2SYan Zheng #include <linux/sched.h> 75d4f98a2SYan Zheng #include <linux/pagemap.h> 85d4f98a2SYan Zheng #include <linux/writeback.h> 95d4f98a2SYan Zheng #include <linux/blkdev.h> 105d4f98a2SYan Zheng #include <linux/rbtree.h> 115a0e3ad6STejun Heo #include <linux/slab.h> 12726a3421SQu Wenruo #include <linux/error-injection.h> 135d4f98a2SYan Zheng #include "ctree.h" 145d4f98a2SYan Zheng #include "disk-io.h" 155d4f98a2SYan Zheng #include "transaction.h" 165d4f98a2SYan Zheng #include "volumes.h" 175d4f98a2SYan Zheng #include "locking.h" 185d4f98a2SYan Zheng #include "btrfs_inode.h" 195d4f98a2SYan Zheng #include "async-thread.h" 200af3d00bSJosef Bacik #include "free-space-cache.h" 21581bb050SLi Zefan #include "inode-map.h" 2262b99540SQu Wenruo #include "qgroup.h" 23cdccee99SLiu Bo #include "print-tree.h" 2486736342SJosef Bacik #include "delalloc-space.h" 25aac0023cSJosef Bacik #include "block-group.h" 2619b546d7SQu Wenruo #include "backref.h" 27e9a28dc5SQu Wenruo #include "misc.h" 285d4f98a2SYan Zheng 295d4f98a2SYan Zheng /* 300c891389SQu Wenruo * Relocation overview 310c891389SQu Wenruo * 320c891389SQu Wenruo * [What does relocation do] 330c891389SQu Wenruo * 340c891389SQu Wenruo * The objective of relocation is to relocate all extents of the target block 350c891389SQu Wenruo * group to other block groups. 360c891389SQu Wenruo * This is utilized by resize (shrink only), profile converting, compacting 370c891389SQu Wenruo * space, or balance routine to spread chunks over devices. 380c891389SQu Wenruo * 390c891389SQu Wenruo * Before | After 400c891389SQu Wenruo * ------------------------------------------------------------------ 410c891389SQu Wenruo * BG A: 10 data extents | BG A: deleted 420c891389SQu Wenruo * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated) 430c891389SQu Wenruo * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated) 440c891389SQu Wenruo * 450c891389SQu Wenruo * [How does relocation work] 460c891389SQu Wenruo * 470c891389SQu Wenruo * 1. Mark the target block group read-only 480c891389SQu Wenruo * New extents won't be allocated from the target block group. 490c891389SQu Wenruo * 500c891389SQu Wenruo * 2.1 Record each extent in the target block group 510c891389SQu Wenruo * To build a proper map of extents to be relocated. 520c891389SQu Wenruo * 530c891389SQu Wenruo * 2.2 Build data reloc tree and reloc trees 540c891389SQu Wenruo * Data reloc tree will contain an inode, recording all newly relocated 550c891389SQu Wenruo * data extents. 560c891389SQu Wenruo * There will be only one data reloc tree for one data block group. 570c891389SQu Wenruo * 580c891389SQu Wenruo * Reloc tree will be a special snapshot of its source tree, containing 590c891389SQu Wenruo * relocated tree blocks. 600c891389SQu Wenruo * Each tree referring to a tree block in target block group will get its 610c891389SQu Wenruo * reloc tree built. 620c891389SQu Wenruo * 630c891389SQu Wenruo * 2.3 Swap source tree with its corresponding reloc tree 640c891389SQu Wenruo * Each involved tree only refers to new extents after swap. 650c891389SQu Wenruo * 660c891389SQu Wenruo * 3. Cleanup reloc trees and data reloc tree. 670c891389SQu Wenruo * As old extents in the target block group are still referenced by reloc 680c891389SQu Wenruo * trees, we need to clean them up before really freeing the target block 690c891389SQu Wenruo * group. 700c891389SQu Wenruo * 710c891389SQu Wenruo * The main complexity is in steps 2.2 and 2.3. 720c891389SQu Wenruo * 730c891389SQu Wenruo * The entry point of relocation is relocate_block_group() function. 740c891389SQu Wenruo */ 750c891389SQu Wenruo 760647bf56SWang Shilong #define RELOCATION_RESERVED_NODES 256 772a979612SQu Wenruo /* 785d4f98a2SYan Zheng * map address of tree root to tree 795d4f98a2SYan Zheng */ 805d4f98a2SYan Zheng struct mapping_node { 81e9a28dc5SQu Wenruo struct { 825d4f98a2SYan Zheng struct rb_node rb_node; 835d4f98a2SYan Zheng u64 bytenr; 84e9a28dc5SQu Wenruo }; /* Use rb_simle_node for search/insert */ 855d4f98a2SYan Zheng void *data; 865d4f98a2SYan Zheng }; 875d4f98a2SYan Zheng 885d4f98a2SYan Zheng struct mapping_tree { 895d4f98a2SYan Zheng struct rb_root rb_root; 905d4f98a2SYan Zheng spinlock_t lock; 915d4f98a2SYan Zheng }; 925d4f98a2SYan Zheng 935d4f98a2SYan Zheng /* 945d4f98a2SYan Zheng * present a tree block to process 955d4f98a2SYan Zheng */ 965d4f98a2SYan Zheng struct tree_block { 97e9a28dc5SQu Wenruo struct { 985d4f98a2SYan Zheng struct rb_node rb_node; 995d4f98a2SYan Zheng u64 bytenr; 100e9a28dc5SQu Wenruo }; /* Use rb_simple_node for search/insert */ 1015d4f98a2SYan Zheng struct btrfs_key key; 1025d4f98a2SYan Zheng unsigned int level:8; 1035d4f98a2SYan Zheng unsigned int key_ready:1; 1045d4f98a2SYan Zheng }; 1055d4f98a2SYan Zheng 1060257bb82SYan, Zheng #define MAX_EXTENTS 128 1070257bb82SYan, Zheng 1080257bb82SYan, Zheng struct file_extent_cluster { 1090257bb82SYan, Zheng u64 start; 1100257bb82SYan, Zheng u64 end; 1110257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 1120257bb82SYan, Zheng unsigned int nr; 1130257bb82SYan, Zheng }; 1140257bb82SYan, Zheng 1155d4f98a2SYan Zheng struct reloc_control { 1165d4f98a2SYan Zheng /* block group to relocate */ 11732da5386SDavid Sterba struct btrfs_block_group *block_group; 1185d4f98a2SYan Zheng /* extent tree */ 1195d4f98a2SYan Zheng struct btrfs_root *extent_root; 1205d4f98a2SYan Zheng /* inode for moving data */ 1215d4f98a2SYan Zheng struct inode *data_inode; 1223fd0a558SYan, Zheng 1233fd0a558SYan, Zheng struct btrfs_block_rsv *block_rsv; 1243fd0a558SYan, Zheng 125a26195a5SQu Wenruo struct btrfs_backref_cache backref_cache; 1263fd0a558SYan, Zheng 1273fd0a558SYan, Zheng struct file_extent_cluster cluster; 1285d4f98a2SYan Zheng /* tree blocks have been processed */ 1295d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 1305d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 1315d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 1325d4f98a2SYan Zheng /* list of reloc trees */ 1335d4f98a2SYan Zheng struct list_head reloc_roots; 134d2311e69SQu Wenruo /* list of subvolume trees that get relocated */ 135d2311e69SQu Wenruo struct list_head dirty_subvol_roots; 1363fd0a558SYan, Zheng /* size of metadata reservation for merging reloc trees */ 1373fd0a558SYan, Zheng u64 merging_rsv_size; 1383fd0a558SYan, Zheng /* size of relocated tree nodes */ 1393fd0a558SYan, Zheng u64 nodes_relocated; 1400647bf56SWang Shilong /* reserved size for block group relocation*/ 1410647bf56SWang Shilong u64 reserved_bytes; 1423fd0a558SYan, Zheng 1435d4f98a2SYan Zheng u64 search_start; 1445d4f98a2SYan Zheng u64 extents_found; 1453fd0a558SYan, Zheng 1463fd0a558SYan, Zheng unsigned int stage:8; 1473fd0a558SYan, Zheng unsigned int create_reloc_tree:1; 1483fd0a558SYan, Zheng unsigned int merge_reloc_tree:1; 1495d4f98a2SYan Zheng unsigned int found_file_extent:1; 1505d4f98a2SYan Zheng }; 1515d4f98a2SYan Zheng 1525d4f98a2SYan Zheng /* stages of data relocation */ 1535d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 1545d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 1555d4f98a2SYan Zheng 156a26195a5SQu Wenruo static void remove_backref_node(struct btrfs_backref_cache *cache, 157a26195a5SQu Wenruo struct btrfs_backref_node *node); 1589569cc20SQu Wenruo 1599569cc20SQu Wenruo static void mark_block_processed(struct reloc_control *rc, 160a26195a5SQu Wenruo struct btrfs_backref_node *node) 1619569cc20SQu Wenruo { 1629569cc20SQu Wenruo u32 blocksize; 1639569cc20SQu Wenruo 1649569cc20SQu Wenruo if (node->level == 0 || 1659569cc20SQu Wenruo in_range(node->bytenr, rc->block_group->start, 1669569cc20SQu Wenruo rc->block_group->length)) { 1679569cc20SQu Wenruo blocksize = rc->extent_root->fs_info->nodesize; 1689569cc20SQu Wenruo set_extent_bits(&rc->processed_blocks, node->bytenr, 1699569cc20SQu Wenruo node->bytenr + blocksize - 1, EXTENT_DIRTY); 1709569cc20SQu Wenruo } 1719569cc20SQu Wenruo node->processed = 1; 1729569cc20SQu Wenruo } 1739569cc20SQu Wenruo 1745d4f98a2SYan Zheng 1755d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 1765d4f98a2SYan Zheng { 1776bef4d31SEric Paris tree->rb_root = RB_ROOT; 1785d4f98a2SYan Zheng spin_lock_init(&tree->lock); 1795d4f98a2SYan Zheng } 1805d4f98a2SYan Zheng 181a26195a5SQu Wenruo static void backref_cache_cleanup(struct btrfs_backref_cache *cache) 1825d4f98a2SYan Zheng { 183a26195a5SQu Wenruo struct btrfs_backref_node *node; 1843fd0a558SYan, Zheng int i; 1853fd0a558SYan, Zheng 1863fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 1873fd0a558SYan, Zheng node = list_entry(cache->detached.next, 188a26195a5SQu Wenruo struct btrfs_backref_node, list); 1893fd0a558SYan, Zheng remove_backref_node(cache, node); 1903fd0a558SYan, Zheng } 1913fd0a558SYan, Zheng 1923fd0a558SYan, Zheng while (!list_empty(&cache->leaves)) { 1933fd0a558SYan, Zheng node = list_entry(cache->leaves.next, 194a26195a5SQu Wenruo struct btrfs_backref_node, lower); 1953fd0a558SYan, Zheng remove_backref_node(cache, node); 1963fd0a558SYan, Zheng } 1973fd0a558SYan, Zheng 1983fd0a558SYan, Zheng cache->last_trans = 0; 1993fd0a558SYan, Zheng 2003fd0a558SYan, Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 201f4907095SLiu Bo ASSERT(list_empty(&cache->pending[i])); 20284780289SQu Wenruo ASSERT(list_empty(&cache->pending_edge)); 20384780289SQu Wenruo ASSERT(list_empty(&cache->useless_node)); 204f4907095SLiu Bo ASSERT(list_empty(&cache->changed)); 205f4907095SLiu Bo ASSERT(list_empty(&cache->detached)); 206f4907095SLiu Bo ASSERT(RB_EMPTY_ROOT(&cache->rb_root)); 207f4907095SLiu Bo ASSERT(!cache->nr_nodes); 208f4907095SLiu Bo ASSERT(!cache->nr_edges); 2093fd0a558SYan, Zheng } 2103fd0a558SYan, Zheng 211a26195a5SQu Wenruo static void free_backref_node(struct btrfs_backref_cache *cache, 212a26195a5SQu Wenruo struct btrfs_backref_node *node) 2133fd0a558SYan, Zheng { 2143fd0a558SYan, Zheng if (node) { 2153fd0a558SYan, Zheng cache->nr_nodes--; 21600246528SJosef Bacik btrfs_put_root(node->root); 2173fd0a558SYan, Zheng kfree(node); 2183fd0a558SYan, Zheng } 2193fd0a558SYan, Zheng } 2203fd0a558SYan, Zheng 2212a979612SQu Wenruo #define LINK_LOWER (1 << 0) 2222a979612SQu Wenruo #define LINK_UPPER (1 << 1) 223a26195a5SQu Wenruo static void link_backref_edge(struct btrfs_backref_edge *edge, 224a26195a5SQu Wenruo struct btrfs_backref_node *lower, 225a26195a5SQu Wenruo struct btrfs_backref_node *upper, 2262a979612SQu Wenruo int link_which) 2272a979612SQu Wenruo { 2282a979612SQu Wenruo ASSERT(upper && lower && upper->level == lower->level + 1); 2292a979612SQu Wenruo edge->node[LOWER] = lower; 2302a979612SQu Wenruo edge->node[UPPER] = upper; 2312a979612SQu Wenruo if (link_which & LINK_LOWER) 2322a979612SQu Wenruo list_add_tail(&edge->list[LOWER], &lower->upper); 2332a979612SQu Wenruo if (link_which & LINK_UPPER) 2342a979612SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 2352a979612SQu Wenruo } 2362a979612SQu Wenruo 237a26195a5SQu Wenruo static void free_backref_edge(struct btrfs_backref_cache *cache, 238a26195a5SQu Wenruo struct btrfs_backref_edge *edge) 2393fd0a558SYan, Zheng { 2403fd0a558SYan, Zheng if (edge) { 2413fd0a558SYan, Zheng cache->nr_edges--; 2423fd0a558SYan, Zheng kfree(edge); 2433fd0a558SYan, Zheng } 2445d4f98a2SYan Zheng } 2455d4f98a2SYan Zheng 24648a3b636SEric Sandeen static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr) 24743c04fb1SJeff Mahoney { 24843c04fb1SJeff Mahoney 24943c04fb1SJeff Mahoney struct btrfs_fs_info *fs_info = NULL; 250a26195a5SQu Wenruo struct btrfs_backref_node *bnode = rb_entry(rb_node, 251a26195a5SQu Wenruo struct btrfs_backref_node, rb_node); 25243c04fb1SJeff Mahoney if (bnode->root) 25343c04fb1SJeff Mahoney fs_info = bnode->root->fs_info; 2545d163e0eSJeff Mahoney btrfs_panic(fs_info, errno, 2555d163e0eSJeff Mahoney "Inconsistency in backref cache found at offset %llu", 2565d163e0eSJeff Mahoney bytenr); 25743c04fb1SJeff Mahoney } 25843c04fb1SJeff Mahoney 2595d4f98a2SYan Zheng /* 2605d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 2615d4f98a2SYan Zheng */ 262a26195a5SQu Wenruo static struct btrfs_backref_node *walk_up_backref( 263a26195a5SQu Wenruo struct btrfs_backref_node *node, 264a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 2655d4f98a2SYan Zheng { 266a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2675d4f98a2SYan Zheng int idx = *index; 2685d4f98a2SYan Zheng 2695d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 2705d4f98a2SYan Zheng edge = list_entry(node->upper.next, 271a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 2725d4f98a2SYan Zheng edges[idx++] = edge; 2735d4f98a2SYan Zheng node = edge->node[UPPER]; 2745d4f98a2SYan Zheng } 2753fd0a558SYan, Zheng BUG_ON(node->detached); 2765d4f98a2SYan Zheng *index = idx; 2775d4f98a2SYan Zheng return node; 2785d4f98a2SYan Zheng } 2795d4f98a2SYan Zheng 2805d4f98a2SYan Zheng /* 2815d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 2825d4f98a2SYan Zheng */ 283a26195a5SQu Wenruo static struct btrfs_backref_node *walk_down_backref( 284a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 2855d4f98a2SYan Zheng { 286a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 287a26195a5SQu Wenruo struct btrfs_backref_node *lower; 2885d4f98a2SYan Zheng int idx = *index; 2895d4f98a2SYan Zheng 2905d4f98a2SYan Zheng while (idx > 0) { 2915d4f98a2SYan Zheng edge = edges[idx - 1]; 2925d4f98a2SYan Zheng lower = edge->node[LOWER]; 2935d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 2945d4f98a2SYan Zheng idx--; 2955d4f98a2SYan Zheng continue; 2965d4f98a2SYan Zheng } 2975d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 298a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 2995d4f98a2SYan Zheng edges[idx - 1] = edge; 3005d4f98a2SYan Zheng *index = idx; 3015d4f98a2SYan Zheng return edge->node[UPPER]; 3025d4f98a2SYan Zheng } 3035d4f98a2SYan Zheng *index = 0; 3045d4f98a2SYan Zheng return NULL; 3055d4f98a2SYan Zheng } 3065d4f98a2SYan Zheng 307a26195a5SQu Wenruo static void unlock_node_buffer(struct btrfs_backref_node *node) 3085d4f98a2SYan Zheng { 3095d4f98a2SYan Zheng if (node->locked) { 3105d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 3115d4f98a2SYan Zheng node->locked = 0; 3125d4f98a2SYan Zheng } 3133fd0a558SYan, Zheng } 3143fd0a558SYan, Zheng 315a26195a5SQu Wenruo static void drop_node_buffer(struct btrfs_backref_node *node) 3163fd0a558SYan, Zheng { 3173fd0a558SYan, Zheng if (node->eb) { 3183fd0a558SYan, Zheng unlock_node_buffer(node); 3195d4f98a2SYan Zheng free_extent_buffer(node->eb); 3205d4f98a2SYan Zheng node->eb = NULL; 3215d4f98a2SYan Zheng } 3225d4f98a2SYan Zheng } 3235d4f98a2SYan Zheng 324a26195a5SQu Wenruo static void drop_backref_node(struct btrfs_backref_cache *tree, 325a26195a5SQu Wenruo struct btrfs_backref_node *node) 3265d4f98a2SYan Zheng { 3275d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 3285d4f98a2SYan Zheng 3295d4f98a2SYan Zheng drop_node_buffer(node); 3303fd0a558SYan, Zheng list_del(&node->list); 3315d4f98a2SYan Zheng list_del(&node->lower); 3323fd0a558SYan, Zheng if (!RB_EMPTY_NODE(&node->rb_node)) 3335d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 3343fd0a558SYan, Zheng free_backref_node(tree, node); 3355d4f98a2SYan Zheng } 3365d4f98a2SYan Zheng 3375d4f98a2SYan Zheng /* 3385d4f98a2SYan Zheng * remove a backref node from the backref cache 3395d4f98a2SYan Zheng */ 340a26195a5SQu Wenruo static void remove_backref_node(struct btrfs_backref_cache *cache, 341a26195a5SQu Wenruo struct btrfs_backref_node *node) 3425d4f98a2SYan Zheng { 343a26195a5SQu Wenruo struct btrfs_backref_node *upper; 344a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 3455d4f98a2SYan Zheng 3465d4f98a2SYan Zheng if (!node) 3475d4f98a2SYan Zheng return; 3485d4f98a2SYan Zheng 3493fd0a558SYan, Zheng BUG_ON(!node->lowest && !node->detached); 3505d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 351a26195a5SQu Wenruo edge = list_entry(node->upper.next, struct btrfs_backref_edge, 3525d4f98a2SYan Zheng list[LOWER]); 3535d4f98a2SYan Zheng upper = edge->node[UPPER]; 3545d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 3555d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 3563fd0a558SYan, Zheng free_backref_edge(cache, edge); 3573fd0a558SYan, Zheng 3583fd0a558SYan, Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 3593fd0a558SYan, Zheng BUG_ON(!list_empty(&node->upper)); 3603fd0a558SYan, Zheng drop_backref_node(cache, node); 3613fd0a558SYan, Zheng node = upper; 3623fd0a558SYan, Zheng node->lowest = 1; 3633fd0a558SYan, Zheng continue; 3643fd0a558SYan, Zheng } 3655d4f98a2SYan Zheng /* 3663fd0a558SYan, Zheng * add the node to leaf node list if no other 3675d4f98a2SYan Zheng * child block cached. 3685d4f98a2SYan Zheng */ 3695d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 3703fd0a558SYan, Zheng list_add_tail(&upper->lower, &cache->leaves); 3715d4f98a2SYan Zheng upper->lowest = 1; 3725d4f98a2SYan Zheng } 3735d4f98a2SYan Zheng } 3743fd0a558SYan, Zheng 3755d4f98a2SYan Zheng drop_backref_node(cache, node); 3765d4f98a2SYan Zheng } 3775d4f98a2SYan Zheng 378a26195a5SQu Wenruo static void update_backref_node(struct btrfs_backref_cache *cache, 379a26195a5SQu Wenruo struct btrfs_backref_node *node, u64 bytenr) 3803fd0a558SYan, Zheng { 3813fd0a558SYan, Zheng struct rb_node *rb_node; 3823fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 3833fd0a558SYan, Zheng node->bytenr = bytenr; 384e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node); 38543c04fb1SJeff Mahoney if (rb_node) 38643c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, bytenr); 3873fd0a558SYan, Zheng } 3883fd0a558SYan, Zheng 3893fd0a558SYan, Zheng /* 3903fd0a558SYan, Zheng * update backref cache after a transaction commit 3913fd0a558SYan, Zheng */ 3923fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 393a26195a5SQu Wenruo struct btrfs_backref_cache *cache) 3943fd0a558SYan, Zheng { 395a26195a5SQu Wenruo struct btrfs_backref_node *node; 3963fd0a558SYan, Zheng int level = 0; 3973fd0a558SYan, Zheng 3983fd0a558SYan, Zheng if (cache->last_trans == 0) { 3993fd0a558SYan, Zheng cache->last_trans = trans->transid; 4003fd0a558SYan, Zheng return 0; 4013fd0a558SYan, Zheng } 4023fd0a558SYan, Zheng 4033fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 4043fd0a558SYan, Zheng return 0; 4053fd0a558SYan, Zheng 4063fd0a558SYan, Zheng /* 4073fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 4083fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 4093fd0a558SYan, Zheng * so the detached nodes are no longer useful. 4103fd0a558SYan, Zheng */ 4113fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 4123fd0a558SYan, Zheng node = list_entry(cache->detached.next, 413a26195a5SQu Wenruo struct btrfs_backref_node, list); 4143fd0a558SYan, Zheng remove_backref_node(cache, node); 4153fd0a558SYan, Zheng } 4163fd0a558SYan, Zheng 4173fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 4183fd0a558SYan, Zheng node = list_entry(cache->changed.next, 419a26195a5SQu Wenruo struct btrfs_backref_node, list); 4203fd0a558SYan, Zheng list_del_init(&node->list); 4213fd0a558SYan, Zheng BUG_ON(node->pending); 4223fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 4233fd0a558SYan, Zheng } 4243fd0a558SYan, Zheng 4253fd0a558SYan, Zheng /* 4263fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 4273fd0a558SYan, Zheng * errors during processing the pending nodes. 4283fd0a558SYan, Zheng */ 4293fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 4303fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 4313fd0a558SYan, Zheng BUG_ON(!node->pending); 4323fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 4333fd0a558SYan, Zheng continue; 4343fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 4353fd0a558SYan, Zheng } 4363fd0a558SYan, Zheng } 4373fd0a558SYan, Zheng 4383fd0a558SYan, Zheng cache->last_trans = 0; 4393fd0a558SYan, Zheng return 1; 4403fd0a558SYan, Zheng } 4413fd0a558SYan, Zheng 4426282675eSQu Wenruo static bool reloc_root_is_dead(struct btrfs_root *root) 4436282675eSQu Wenruo { 4446282675eSQu Wenruo /* 4456282675eSQu Wenruo * Pair with set_bit/clear_bit in clean_dirty_subvols and 4466282675eSQu Wenruo * btrfs_update_reloc_root. We need to see the updated bit before 4476282675eSQu Wenruo * trying to access reloc_root 4486282675eSQu Wenruo */ 4496282675eSQu Wenruo smp_rmb(); 4506282675eSQu Wenruo if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state)) 4516282675eSQu Wenruo return true; 4526282675eSQu Wenruo return false; 4536282675eSQu Wenruo } 4546282675eSQu Wenruo 4556282675eSQu Wenruo /* 4566282675eSQu Wenruo * Check if this subvolume tree has valid reloc tree. 4576282675eSQu Wenruo * 4586282675eSQu Wenruo * Reloc tree after swap is considered dead, thus not considered as valid. 4596282675eSQu Wenruo * This is enough for most callers, as they don't distinguish dead reloc root 4606282675eSQu Wenruo * from no reloc root. But should_ignore_root() below is a special case. 4616282675eSQu Wenruo */ 4626282675eSQu Wenruo static bool have_reloc_root(struct btrfs_root *root) 4636282675eSQu Wenruo { 4646282675eSQu Wenruo if (reloc_root_is_dead(root)) 4656282675eSQu Wenruo return false; 4666282675eSQu Wenruo if (!root->reloc_root) 4676282675eSQu Wenruo return false; 4686282675eSQu Wenruo return true; 4696282675eSQu Wenruo } 470f2a97a9dSDavid Sterba 4713fd0a558SYan, Zheng static int should_ignore_root(struct btrfs_root *root) 4723fd0a558SYan, Zheng { 4733fd0a558SYan, Zheng struct btrfs_root *reloc_root; 4743fd0a558SYan, Zheng 47527cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 4763fd0a558SYan, Zheng return 0; 4773fd0a558SYan, Zheng 4786282675eSQu Wenruo /* This root has been merged with its reloc tree, we can ignore it */ 4796282675eSQu Wenruo if (reloc_root_is_dead(root)) 4806282675eSQu Wenruo return 1; 4816282675eSQu Wenruo 4823fd0a558SYan, Zheng reloc_root = root->reloc_root; 4833fd0a558SYan, Zheng if (!reloc_root) 4843fd0a558SYan, Zheng return 0; 4853fd0a558SYan, Zheng 4864d4225fcSJosef Bacik if (btrfs_header_generation(reloc_root->commit_root) == 4874d4225fcSJosef Bacik root->fs_info->running_transaction->transid) 4883fd0a558SYan, Zheng return 0; 4893fd0a558SYan, Zheng /* 4903fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 4913fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 4923fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 4933fd0a558SYan, Zheng * relocation. 4943fd0a558SYan, Zheng */ 4953fd0a558SYan, Zheng return 1; 4963fd0a558SYan, Zheng } 4975d4f98a2SYan Zheng /* 4985d4f98a2SYan Zheng * find reloc tree by address of tree root 4995d4f98a2SYan Zheng */ 5002433bea5SQu Wenruo struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr) 5015d4f98a2SYan Zheng { 5022433bea5SQu Wenruo struct reloc_control *rc = fs_info->reloc_ctl; 5035d4f98a2SYan Zheng struct rb_node *rb_node; 5045d4f98a2SYan Zheng struct mapping_node *node; 5055d4f98a2SYan Zheng struct btrfs_root *root = NULL; 5065d4f98a2SYan Zheng 5072433bea5SQu Wenruo ASSERT(rc); 5085d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 509e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr); 5105d4f98a2SYan Zheng if (rb_node) { 5115d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 5125d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 5135d4f98a2SYan Zheng } 5145d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 51500246528SJosef Bacik return btrfs_grab_root(root); 5165d4f98a2SYan Zheng } 5175d4f98a2SYan Zheng 5185d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 5195d4f98a2SYan Zheng u64 root_objectid) 5205d4f98a2SYan Zheng { 5215d4f98a2SYan Zheng struct btrfs_key key; 5225d4f98a2SYan Zheng 5235d4f98a2SYan Zheng key.objectid = root_objectid; 5245d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 5255d4f98a2SYan Zheng key.offset = (u64)-1; 5265d4f98a2SYan Zheng 527bc44d7c4SJosef Bacik return btrfs_get_fs_root(fs_info, &key, false); 5285d4f98a2SYan Zheng } 5295d4f98a2SYan Zheng 5305d4f98a2SYan Zheng /* 5314007ea87SQu Wenruo * Handle direct tree backref 5324007ea87SQu Wenruo * 5334007ea87SQu Wenruo * Direct tree backref means, the backref item shows its parent bytenr 5344007ea87SQu Wenruo * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined). 5354007ea87SQu Wenruo * 5364007ea87SQu Wenruo * @ref_key: The converted backref key. 5374007ea87SQu Wenruo * For keyed backref, it's the item key. 5384007ea87SQu Wenruo * For inlined backref, objectid is the bytenr, 5394007ea87SQu Wenruo * type is btrfs_inline_ref_type, offset is 5404007ea87SQu Wenruo * btrfs_inline_ref_offset. 5414007ea87SQu Wenruo */ 542a26195a5SQu Wenruo static int handle_direct_tree_backref(struct btrfs_backref_cache *cache, 5434007ea87SQu Wenruo struct btrfs_key *ref_key, 544a26195a5SQu Wenruo struct btrfs_backref_node *cur) 5454007ea87SQu Wenruo { 546a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 547a26195a5SQu Wenruo struct btrfs_backref_node *upper; 5484007ea87SQu Wenruo struct rb_node *rb_node; 5494007ea87SQu Wenruo 5504007ea87SQu Wenruo ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY); 5514007ea87SQu Wenruo 5524007ea87SQu Wenruo /* Only reloc root uses backref pointing to itself */ 5534007ea87SQu Wenruo if (ref_key->objectid == ref_key->offset) { 5544007ea87SQu Wenruo struct btrfs_root *root; 5554007ea87SQu Wenruo 5564007ea87SQu Wenruo cur->is_reloc_root = 1; 5574007ea87SQu Wenruo /* Only reloc backref cache cares about a specific root */ 5584007ea87SQu Wenruo if (cache->is_reloc) { 5594007ea87SQu Wenruo root = find_reloc_root(cache->fs_info, cur->bytenr); 5604007ea87SQu Wenruo if (WARN_ON(!root)) 5614007ea87SQu Wenruo return -ENOENT; 5624007ea87SQu Wenruo cur->root = root; 5634007ea87SQu Wenruo } else { 5644007ea87SQu Wenruo /* 5654007ea87SQu Wenruo * For generic purpose backref cache, reloc root node 5664007ea87SQu Wenruo * is useless. 5674007ea87SQu Wenruo */ 5684007ea87SQu Wenruo list_add(&cur->list, &cache->useless_node); 5694007ea87SQu Wenruo } 5704007ea87SQu Wenruo return 0; 5714007ea87SQu Wenruo } 5724007ea87SQu Wenruo 57347254d07SQu Wenruo edge = btrfs_backref_alloc_edge(cache); 5744007ea87SQu Wenruo if (!edge) 5754007ea87SQu Wenruo return -ENOMEM; 5764007ea87SQu Wenruo 577e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, ref_key->offset); 5784007ea87SQu Wenruo if (!rb_node) { 5794007ea87SQu Wenruo /* Parent node not yet cached */ 580b1818dabSQu Wenruo upper = btrfs_backref_alloc_node(cache, ref_key->offset, 5810304f2d8SQu Wenruo cur->level + 1); 5824007ea87SQu Wenruo if (!upper) { 5834007ea87SQu Wenruo free_backref_edge(cache, edge); 5844007ea87SQu Wenruo return -ENOMEM; 5854007ea87SQu Wenruo } 5864007ea87SQu Wenruo 5874007ea87SQu Wenruo /* 5884007ea87SQu Wenruo * Backrefs for the upper level block isn't cached, add the 5894007ea87SQu Wenruo * block to pending list 5904007ea87SQu Wenruo */ 5914007ea87SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 5924007ea87SQu Wenruo } else { 5934007ea87SQu Wenruo /* Parent node already cached */ 594a26195a5SQu Wenruo upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node); 5954007ea87SQu Wenruo ASSERT(upper->checked); 5964007ea87SQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 5974007ea87SQu Wenruo } 5982a979612SQu Wenruo link_backref_edge(edge, cur, upper, LINK_LOWER); 5994007ea87SQu Wenruo return 0; 6004007ea87SQu Wenruo } 6014007ea87SQu Wenruo 6024007ea87SQu Wenruo /* 6034d81ea8bSQu Wenruo * Handle indirect tree backref 6044d81ea8bSQu Wenruo * 6054d81ea8bSQu Wenruo * Indirect tree backref means, we only know which tree the node belongs to. 6064d81ea8bSQu Wenruo * We still need to do a tree search to find out the parents. This is for 6074d81ea8bSQu Wenruo * TREE_BLOCK_REF backref (keyed or inlined). 6084d81ea8bSQu Wenruo * 6094d81ea8bSQu Wenruo * @ref_key: The same as @ref_key in handle_direct_tree_backref() 6104d81ea8bSQu Wenruo * @tree_key: The first key of this tree block. 6114d81ea8bSQu Wenruo * @path: A clean (released) path, to avoid allocating path everytime 6124d81ea8bSQu Wenruo * the function get called. 6134d81ea8bSQu Wenruo */ 614a26195a5SQu Wenruo static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache, 6154d81ea8bSQu Wenruo struct btrfs_path *path, 6164d81ea8bSQu Wenruo struct btrfs_key *ref_key, 6174d81ea8bSQu Wenruo struct btrfs_key *tree_key, 618a26195a5SQu Wenruo struct btrfs_backref_node *cur) 6194d81ea8bSQu Wenruo { 6204d81ea8bSQu Wenruo struct btrfs_fs_info *fs_info = cache->fs_info; 621a26195a5SQu Wenruo struct btrfs_backref_node *upper; 622a26195a5SQu Wenruo struct btrfs_backref_node *lower; 623a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 6244d81ea8bSQu Wenruo struct extent_buffer *eb; 6254d81ea8bSQu Wenruo struct btrfs_root *root; 6264d81ea8bSQu Wenruo struct rb_node *rb_node; 6274d81ea8bSQu Wenruo int level; 6284d81ea8bSQu Wenruo bool need_check = true; 6294d81ea8bSQu Wenruo int ret; 6304d81ea8bSQu Wenruo 6314d81ea8bSQu Wenruo root = read_fs_root(fs_info, ref_key->offset); 6324d81ea8bSQu Wenruo if (IS_ERR(root)) 6334d81ea8bSQu Wenruo return PTR_ERR(root); 6344d81ea8bSQu Wenruo if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 6354d81ea8bSQu Wenruo cur->cowonly = 1; 6364d81ea8bSQu Wenruo 6374d81ea8bSQu Wenruo if (btrfs_root_level(&root->root_item) == cur->level) { 6384d81ea8bSQu Wenruo /* Tree root */ 6394d81ea8bSQu Wenruo ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr); 6404d81ea8bSQu Wenruo if (should_ignore_root(root)) { 6414d81ea8bSQu Wenruo btrfs_put_root(root); 6424d81ea8bSQu Wenruo list_add(&cur->list, &cache->useless_node); 6434d81ea8bSQu Wenruo } else { 6444d81ea8bSQu Wenruo cur->root = root; 6454d81ea8bSQu Wenruo } 6464d81ea8bSQu Wenruo return 0; 6474d81ea8bSQu Wenruo } 6484d81ea8bSQu Wenruo 6494d81ea8bSQu Wenruo level = cur->level + 1; 6504d81ea8bSQu Wenruo 6514d81ea8bSQu Wenruo /* Search the tree to find parent blocks referring to the block */ 6524d81ea8bSQu Wenruo path->search_commit_root = 1; 6534d81ea8bSQu Wenruo path->skip_locking = 1; 6544d81ea8bSQu Wenruo path->lowest_level = level; 6554d81ea8bSQu Wenruo ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0); 6564d81ea8bSQu Wenruo path->lowest_level = 0; 6574d81ea8bSQu Wenruo if (ret < 0) { 6584d81ea8bSQu Wenruo btrfs_put_root(root); 6594d81ea8bSQu Wenruo return ret; 6604d81ea8bSQu Wenruo } 6614d81ea8bSQu Wenruo if (ret > 0 && path->slots[level] > 0) 6624d81ea8bSQu Wenruo path->slots[level]--; 6634d81ea8bSQu Wenruo 6644d81ea8bSQu Wenruo eb = path->nodes[level]; 6654d81ea8bSQu Wenruo if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) { 6664d81ea8bSQu Wenruo btrfs_err(fs_info, 6674d81ea8bSQu Wenruo "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)", 6684d81ea8bSQu Wenruo cur->bytenr, level - 1, root->root_key.objectid, 6694d81ea8bSQu Wenruo tree_key->objectid, tree_key->type, tree_key->offset); 6704d81ea8bSQu Wenruo btrfs_put_root(root); 6714d81ea8bSQu Wenruo ret = -ENOENT; 6724d81ea8bSQu Wenruo goto out; 6734d81ea8bSQu Wenruo } 6744d81ea8bSQu Wenruo lower = cur; 6754d81ea8bSQu Wenruo 6764d81ea8bSQu Wenruo /* Add all nodes and edges in the path */ 6774d81ea8bSQu Wenruo for (; level < BTRFS_MAX_LEVEL; level++) { 6784d81ea8bSQu Wenruo if (!path->nodes[level]) { 6794d81ea8bSQu Wenruo ASSERT(btrfs_root_bytenr(&root->root_item) == 6804d81ea8bSQu Wenruo lower->bytenr); 6814d81ea8bSQu Wenruo if (should_ignore_root(root)) { 6824d81ea8bSQu Wenruo btrfs_put_root(root); 6834d81ea8bSQu Wenruo list_add(&lower->list, &cache->useless_node); 6844d81ea8bSQu Wenruo } else { 6854d81ea8bSQu Wenruo lower->root = root; 6864d81ea8bSQu Wenruo } 6874d81ea8bSQu Wenruo break; 6884d81ea8bSQu Wenruo } 6894d81ea8bSQu Wenruo 69047254d07SQu Wenruo edge = btrfs_backref_alloc_edge(cache); 6914d81ea8bSQu Wenruo if (!edge) { 6924d81ea8bSQu Wenruo btrfs_put_root(root); 6934d81ea8bSQu Wenruo ret = -ENOMEM; 6944d81ea8bSQu Wenruo goto out; 6954d81ea8bSQu Wenruo } 6964d81ea8bSQu Wenruo 6974d81ea8bSQu Wenruo eb = path->nodes[level]; 698e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, eb->start); 6994d81ea8bSQu Wenruo if (!rb_node) { 700b1818dabSQu Wenruo upper = btrfs_backref_alloc_node(cache, eb->start, 7010304f2d8SQu Wenruo lower->level + 1); 7024d81ea8bSQu Wenruo if (!upper) { 7034d81ea8bSQu Wenruo btrfs_put_root(root); 7044d81ea8bSQu Wenruo free_backref_edge(cache, edge); 7054d81ea8bSQu Wenruo ret = -ENOMEM; 7064d81ea8bSQu Wenruo goto out; 7074d81ea8bSQu Wenruo } 7084d81ea8bSQu Wenruo upper->owner = btrfs_header_owner(eb); 7094d81ea8bSQu Wenruo if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 7104d81ea8bSQu Wenruo upper->cowonly = 1; 7114d81ea8bSQu Wenruo 7124d81ea8bSQu Wenruo /* 7134d81ea8bSQu Wenruo * If we know the block isn't shared we can avoid 7144d81ea8bSQu Wenruo * checking its backrefs. 7154d81ea8bSQu Wenruo */ 7164d81ea8bSQu Wenruo if (btrfs_block_can_be_shared(root, eb)) 7174d81ea8bSQu Wenruo upper->checked = 0; 7184d81ea8bSQu Wenruo else 7194d81ea8bSQu Wenruo upper->checked = 1; 7204d81ea8bSQu Wenruo 7214d81ea8bSQu Wenruo /* 7224d81ea8bSQu Wenruo * Add the block to pending list if we need to check its 7234d81ea8bSQu Wenruo * backrefs, we only do this once while walking up a 7244d81ea8bSQu Wenruo * tree as we will catch anything else later on. 7254d81ea8bSQu Wenruo */ 7264d81ea8bSQu Wenruo if (!upper->checked && need_check) { 7274d81ea8bSQu Wenruo need_check = false; 7284d81ea8bSQu Wenruo list_add_tail(&edge->list[UPPER], 7294d81ea8bSQu Wenruo &cache->pending_edge); 7304d81ea8bSQu Wenruo } else { 7314d81ea8bSQu Wenruo if (upper->checked) 7324d81ea8bSQu Wenruo need_check = true; 7334d81ea8bSQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 7344d81ea8bSQu Wenruo } 7354d81ea8bSQu Wenruo } else { 736a26195a5SQu Wenruo upper = rb_entry(rb_node, struct btrfs_backref_node, 737a26195a5SQu Wenruo rb_node); 7384d81ea8bSQu Wenruo ASSERT(upper->checked); 7394d81ea8bSQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 7404d81ea8bSQu Wenruo if (!upper->owner) 7414d81ea8bSQu Wenruo upper->owner = btrfs_header_owner(eb); 7424d81ea8bSQu Wenruo } 7432a979612SQu Wenruo link_backref_edge(edge, lower, upper, LINK_LOWER); 7444d81ea8bSQu Wenruo 7454d81ea8bSQu Wenruo if (rb_node) { 7464d81ea8bSQu Wenruo btrfs_put_root(root); 7474d81ea8bSQu Wenruo break; 7484d81ea8bSQu Wenruo } 7494d81ea8bSQu Wenruo lower = upper; 7504d81ea8bSQu Wenruo upper = NULL; 7514d81ea8bSQu Wenruo } 7524d81ea8bSQu Wenruo out: 7534d81ea8bSQu Wenruo btrfs_release_path(path); 7544d81ea8bSQu Wenruo return ret; 7554d81ea8bSQu Wenruo } 7564d81ea8bSQu Wenruo 757a26195a5SQu Wenruo static int handle_one_tree_block(struct btrfs_backref_cache *cache, 758e7d571c7SQu Wenruo struct btrfs_path *path, 759e7d571c7SQu Wenruo struct btrfs_backref_iter *iter, 7605d4f98a2SYan Zheng struct btrfs_key *node_key, 761a26195a5SQu Wenruo struct btrfs_backref_node *cur) 7625d4f98a2SYan Zheng { 763e7d571c7SQu Wenruo struct btrfs_fs_info *fs_info = cache->fs_info; 764a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 765a26195a5SQu Wenruo struct btrfs_backref_node *exist; 7665d4f98a2SYan Zheng int ret; 7675d4f98a2SYan Zheng 76871f572a9SQu Wenruo ret = btrfs_backref_iter_start(iter, cur->bytenr); 769e7d571c7SQu Wenruo if (ret < 0) 770e7d571c7SQu Wenruo return ret; 77171f572a9SQu Wenruo /* 77271f572a9SQu Wenruo * We skip the first btrfs_tree_block_info, as we don't use the key 77371f572a9SQu Wenruo * stored in it, but fetch it from the tree block 77471f572a9SQu Wenruo */ 77571f572a9SQu Wenruo if (btrfs_backref_has_tree_block_info(iter)) { 77671f572a9SQu Wenruo ret = btrfs_backref_iter_next(iter); 777e7d571c7SQu Wenruo if (ret < 0) 77871f572a9SQu Wenruo goto out; 77971f572a9SQu Wenruo /* No extra backref? This means the tree block is corrupted */ 78071f572a9SQu Wenruo if (ret > 0) { 781e7d571c7SQu Wenruo ret = -EUCLEAN; 78271f572a9SQu Wenruo goto out; 78371f572a9SQu Wenruo } 78471f572a9SQu Wenruo } 7855d4f98a2SYan Zheng WARN_ON(cur->checked); 7865d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 7875d4f98a2SYan Zheng /* 78870f23fd6SJustin P. Mattock * the backref was added previously when processing 7895d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 7905d4f98a2SYan Zheng */ 79175bfb9afSJosef Bacik ASSERT(list_is_singular(&cur->upper)); 792a26195a5SQu Wenruo edge = list_entry(cur->upper.next, struct btrfs_backref_edge, 7935d4f98a2SYan Zheng list[LOWER]); 79475bfb9afSJosef Bacik ASSERT(list_empty(&edge->list[UPPER])); 7955d4f98a2SYan Zheng exist = edge->node[UPPER]; 7965d4f98a2SYan Zheng /* 7975d4f98a2SYan Zheng * add the upper level block to pending list if we need 7985d4f98a2SYan Zheng * check its backrefs 7995d4f98a2SYan Zheng */ 8005d4f98a2SYan Zheng if (!exist->checked) 80184780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 8025d4f98a2SYan Zheng } else { 8035d4f98a2SYan Zheng exist = NULL; 8045d4f98a2SYan Zheng } 8055d4f98a2SYan Zheng 80671f572a9SQu Wenruo for (; ret == 0; ret = btrfs_backref_iter_next(iter)) { 80771f572a9SQu Wenruo struct extent_buffer *eb; 80871f572a9SQu Wenruo struct btrfs_key key; 8093de28d57SLiu Bo int type; 81071f572a9SQu Wenruo 81171f572a9SQu Wenruo cond_resched(); 81271f572a9SQu Wenruo eb = btrfs_backref_get_eb(iter); 81371f572a9SQu Wenruo 81471f572a9SQu Wenruo key.objectid = iter->bytenr; 81571f572a9SQu Wenruo if (btrfs_backref_iter_is_inline_ref(iter)) { 81671f572a9SQu Wenruo struct btrfs_extent_inline_ref *iref; 81771f572a9SQu Wenruo 81871f572a9SQu Wenruo /* update key for inline back ref */ 81971f572a9SQu Wenruo iref = (struct btrfs_extent_inline_ref *) 82071f572a9SQu Wenruo ((unsigned long)iter->cur_ptr); 8213de28d57SLiu Bo type = btrfs_get_extent_inline_ref_type(eb, iref, 8223de28d57SLiu Bo BTRFS_REF_TYPE_BLOCK); 8233de28d57SLiu Bo if (type == BTRFS_REF_TYPE_INVALID) { 824e7d571c7SQu Wenruo ret = -EUCLEAN; 8253de28d57SLiu Bo goto out; 8263de28d57SLiu Bo } 8273de28d57SLiu Bo key.type = type; 8285d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 82971f572a9SQu Wenruo } else { 83071f572a9SQu Wenruo key.type = iter->cur_key.type; 83171f572a9SQu Wenruo key.offset = iter->cur_key.offset; 8325d4f98a2SYan Zheng } 8335d4f98a2SYan Zheng 834fa6ac715SQu Wenruo /* 835fa6ac715SQu Wenruo * Parent node found and matches current inline ref, no need to 836fa6ac715SQu Wenruo * rebuild this node for this inline ref. 837fa6ac715SQu Wenruo */ 8385d4f98a2SYan Zheng if (exist && 8395d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 8405d4f98a2SYan Zheng exist->owner == key.offset) || 8415d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 8425d4f98a2SYan Zheng exist->bytenr == key.offset))) { 8435d4f98a2SYan Zheng exist = NULL; 84471f572a9SQu Wenruo continue; 8455d4f98a2SYan Zheng } 8465d4f98a2SYan Zheng 847fa6ac715SQu Wenruo /* SHARED_BLOCK_REF means key.offset is the parent bytenr */ 8485d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 8494007ea87SQu Wenruo ret = handle_direct_tree_backref(cache, &key, cur); 850e7d571c7SQu Wenruo if (ret < 0) 8512433bea5SQu Wenruo goto out; 85271f572a9SQu Wenruo continue; 8536d8ff4e4SDavid Sterba } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) { 854e7d571c7SQu Wenruo ret = -EINVAL; 855e7d571c7SQu Wenruo btrfs_print_v0_err(fs_info); 856e7d571c7SQu Wenruo btrfs_handle_fs_error(fs_info, ret, NULL); 857ba3c2b19SNikolay Borisov goto out; 8585d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 85971f572a9SQu Wenruo continue; 8605d4f98a2SYan Zheng } 8615d4f98a2SYan Zheng 862fa6ac715SQu Wenruo /* 863fa6ac715SQu Wenruo * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset 864fa6ac715SQu Wenruo * means the root objectid. We need to search the tree to get 865fa6ac715SQu Wenruo * its parent bytenr. 866fa6ac715SQu Wenruo */ 8674d81ea8bSQu Wenruo ret = handle_indirect_tree_backref(cache, path, &key, node_key, 8684d81ea8bSQu Wenruo cur); 869e7d571c7SQu Wenruo if (ret < 0) 87071f572a9SQu Wenruo goto out; 8715d4f98a2SYan Zheng } 87271f572a9SQu Wenruo ret = 0; 8735d4f98a2SYan Zheng cur->checked = 1; 8745d4f98a2SYan Zheng WARN_ON(exist); 875e7d571c7SQu Wenruo out: 876e7d571c7SQu Wenruo btrfs_backref_iter_release(iter); 877e7d571c7SQu Wenruo return ret; 878e7d571c7SQu Wenruo } 8795d4f98a2SYan Zheng 880e7d571c7SQu Wenruo /* 8811f872924SQu Wenruo * In handle_one_tree_backref(), we have only linked the lower node to the edge, 8821f872924SQu Wenruo * but the upper node hasn't been linked to the edge. 883a26195a5SQu Wenruo * This means we can only iterate through btrfs_backref_node::upper to reach 884a26195a5SQu Wenruo * parent edges, but not through btrfs_backref_node::lower to reach children 885a26195a5SQu Wenruo * edges. 8861f872924SQu Wenruo * 887a26195a5SQu Wenruo * This function will finish the btrfs_backref_node::lower to related edges, 888a26195a5SQu Wenruo * so that backref cache can be bi-directionally iterated. 8891f872924SQu Wenruo * 8901f872924SQu Wenruo * Also, this will add the nodes to backref cache for the next run. 8911f872924SQu Wenruo */ 892a26195a5SQu Wenruo static int finish_upper_links(struct btrfs_backref_cache *cache, 893a26195a5SQu Wenruo struct btrfs_backref_node *start) 8941f872924SQu Wenruo { 8951f872924SQu Wenruo struct list_head *useless_node = &cache->useless_node; 896a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 8971f872924SQu Wenruo struct rb_node *rb_node; 8981f872924SQu Wenruo LIST_HEAD(pending_edge); 8991f872924SQu Wenruo 9001f872924SQu Wenruo ASSERT(start->checked); 9011f872924SQu Wenruo 9021f872924SQu Wenruo /* Insert this node to cache if it's not COW-only */ 9031f872924SQu Wenruo if (!start->cowonly) { 904e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, start->bytenr, 9051f872924SQu Wenruo &start->rb_node); 9061f872924SQu Wenruo if (rb_node) 9071f872924SQu Wenruo backref_tree_panic(rb_node, -EEXIST, start->bytenr); 9081f872924SQu Wenruo list_add_tail(&start->lower, &cache->leaves); 9091f872924SQu Wenruo } 9101f872924SQu Wenruo 9111f872924SQu Wenruo /* 9121f872924SQu Wenruo * Use breadth first search to iterate all related edges. 9131f872924SQu Wenruo * 9141f872924SQu Wenruo * The starting points are all the edges of this node 9151f872924SQu Wenruo */ 9161f872924SQu Wenruo list_for_each_entry(edge, &start->upper, list[LOWER]) 9171f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &pending_edge); 9181f872924SQu Wenruo 9191f872924SQu Wenruo while (!list_empty(&pending_edge)) { 920a26195a5SQu Wenruo struct btrfs_backref_node *upper; 921a26195a5SQu Wenruo struct btrfs_backref_node *lower; 9221f872924SQu Wenruo struct rb_node *rb_node; 9231f872924SQu Wenruo 924a26195a5SQu Wenruo edge = list_first_entry(&pending_edge, 925a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 9261f872924SQu Wenruo list_del_init(&edge->list[UPPER]); 9271f872924SQu Wenruo upper = edge->node[UPPER]; 9281f872924SQu Wenruo lower = edge->node[LOWER]; 9291f872924SQu Wenruo 9301f872924SQu Wenruo /* Parent is detached, no need to keep any edges */ 9311f872924SQu Wenruo if (upper->detached) { 9321f872924SQu Wenruo list_del(&edge->list[LOWER]); 9331f872924SQu Wenruo free_backref_edge(cache, edge); 9341f872924SQu Wenruo 9351f872924SQu Wenruo /* Lower node is orphan, queue for cleanup */ 9361f872924SQu Wenruo if (list_empty(&lower->upper)) 9371f872924SQu Wenruo list_add(&lower->list, useless_node); 9381f872924SQu Wenruo continue; 9391f872924SQu Wenruo } 9401f872924SQu Wenruo 9411f872924SQu Wenruo /* 9421f872924SQu Wenruo * All new nodes added in current build_backref_tree() haven't 9431f872924SQu Wenruo * been linked to the cache rb tree. 9441f872924SQu Wenruo * So if we have upper->rb_node populated, this means a cache 9451f872924SQu Wenruo * hit. We only need to link the edge, as @upper and all its 9461f872924SQu Wenruo * parent have already been linked. 9471f872924SQu Wenruo */ 9481f872924SQu Wenruo if (!RB_EMPTY_NODE(&upper->rb_node)) { 9491f872924SQu Wenruo if (upper->lowest) { 9501f872924SQu Wenruo list_del_init(&upper->lower); 9511f872924SQu Wenruo upper->lowest = 0; 9521f872924SQu Wenruo } 9531f872924SQu Wenruo 9541f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 9551f872924SQu Wenruo continue; 9561f872924SQu Wenruo } 9571f872924SQu Wenruo 9581f872924SQu Wenruo /* Sanity check, we shouldn't have any unchecked nodes */ 9591f872924SQu Wenruo if (!upper->checked) { 9601f872924SQu Wenruo ASSERT(0); 9611f872924SQu Wenruo return -EUCLEAN; 9621f872924SQu Wenruo } 9631f872924SQu Wenruo 9641f872924SQu Wenruo /* Sanity check, COW-only node has non-COW-only parent */ 9651f872924SQu Wenruo if (start->cowonly != upper->cowonly) { 9661f872924SQu Wenruo ASSERT(0); 9671f872924SQu Wenruo return -EUCLEAN; 9681f872924SQu Wenruo } 9691f872924SQu Wenruo 9701f872924SQu Wenruo /* Only cache non-COW-only (subvolume trees) tree blocks */ 9711f872924SQu Wenruo if (!upper->cowonly) { 972e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr, 9731f872924SQu Wenruo &upper->rb_node); 9741f872924SQu Wenruo if (rb_node) { 9751f872924SQu Wenruo backref_tree_panic(rb_node, -EEXIST, 9761f872924SQu Wenruo upper->bytenr); 9771f872924SQu Wenruo return -EUCLEAN; 9781f872924SQu Wenruo } 9791f872924SQu Wenruo } 9801f872924SQu Wenruo 9811f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 9821f872924SQu Wenruo 9831f872924SQu Wenruo /* 9841f872924SQu Wenruo * Also queue all the parent edges of this uncached node to 9851f872924SQu Wenruo * finish the upper linkage 9861f872924SQu Wenruo */ 9871f872924SQu Wenruo list_for_each_entry(edge, &upper->upper, list[LOWER]) 9881f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &pending_edge); 9891f872924SQu Wenruo } 9901f872924SQu Wenruo return 0; 9911f872924SQu Wenruo } 9921f872924SQu Wenruo 9931f872924SQu Wenruo /* 99429db137bSQu Wenruo * For useless nodes, do two major clean ups: 99529db137bSQu Wenruo * 99629db137bSQu Wenruo * - Cleanup the children edges and nodes 99729db137bSQu Wenruo * If child node is also orphan (no parent) during cleanup, then the child 99829db137bSQu Wenruo * node will also be cleaned up. 99929db137bSQu Wenruo * 100029db137bSQu Wenruo * - Freeing up leaves (level 0), keeps nodes detached 100129db137bSQu Wenruo * For nodes, the node is still cached as "detached" 100229db137bSQu Wenruo * 100329db137bSQu Wenruo * Return false if @node is not in the @useless_nodes list. 100429db137bSQu Wenruo * Return true if @node is in the @useless_nodes list. 100529db137bSQu Wenruo */ 100629db137bSQu Wenruo static bool handle_useless_nodes(struct reloc_control *rc, 1007a26195a5SQu Wenruo struct btrfs_backref_node *node) 100829db137bSQu Wenruo { 1009a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 101029db137bSQu Wenruo struct list_head *useless_node = &cache->useless_node; 101129db137bSQu Wenruo bool ret = false; 101229db137bSQu Wenruo 101329db137bSQu Wenruo while (!list_empty(useless_node)) { 1014a26195a5SQu Wenruo struct btrfs_backref_node *cur; 101529db137bSQu Wenruo 1016a26195a5SQu Wenruo cur = list_first_entry(useless_node, struct btrfs_backref_node, 101729db137bSQu Wenruo list); 101829db137bSQu Wenruo list_del_init(&cur->list); 101929db137bSQu Wenruo 102029db137bSQu Wenruo /* Only tree root nodes can be added to @useless_nodes */ 102129db137bSQu Wenruo ASSERT(list_empty(&cur->upper)); 102229db137bSQu Wenruo 102329db137bSQu Wenruo if (cur == node) 102429db137bSQu Wenruo ret = true; 102529db137bSQu Wenruo 102629db137bSQu Wenruo /* The node is the lowest node */ 102729db137bSQu Wenruo if (cur->lowest) { 102829db137bSQu Wenruo list_del_init(&cur->lower); 102929db137bSQu Wenruo cur->lowest = 0; 103029db137bSQu Wenruo } 103129db137bSQu Wenruo 103229db137bSQu Wenruo /* Cleanup the lower edges */ 103329db137bSQu Wenruo while (!list_empty(&cur->lower)) { 1034a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1035a26195a5SQu Wenruo struct btrfs_backref_node *lower; 103629db137bSQu Wenruo 103729db137bSQu Wenruo edge = list_entry(cur->lower.next, 1038a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 103929db137bSQu Wenruo list_del(&edge->list[UPPER]); 104029db137bSQu Wenruo list_del(&edge->list[LOWER]); 104129db137bSQu Wenruo lower = edge->node[LOWER]; 104229db137bSQu Wenruo free_backref_edge(cache, edge); 104329db137bSQu Wenruo 104429db137bSQu Wenruo /* Child node is also orphan, queue for cleanup */ 104529db137bSQu Wenruo if (list_empty(&lower->upper)) 104629db137bSQu Wenruo list_add(&lower->list, useless_node); 104729db137bSQu Wenruo } 104829db137bSQu Wenruo /* Mark this block processed for relocation */ 104929db137bSQu Wenruo mark_block_processed(rc, cur); 105029db137bSQu Wenruo 105129db137bSQu Wenruo /* 105229db137bSQu Wenruo * Backref nodes for tree leaves are deleted from the cache. 105329db137bSQu Wenruo * Backref nodes for upper level tree blocks are left in the 105429db137bSQu Wenruo * cache to avoid unnecessary backref lookup. 105529db137bSQu Wenruo */ 105629db137bSQu Wenruo if (cur->level > 0) { 105729db137bSQu Wenruo list_add(&cur->list, &cache->detached); 105829db137bSQu Wenruo cur->detached = 1; 105929db137bSQu Wenruo } else { 106029db137bSQu Wenruo rb_erase(&cur->rb_node, &cache->rb_root); 106129db137bSQu Wenruo free_backref_node(cache, cur); 106229db137bSQu Wenruo } 106329db137bSQu Wenruo } 106429db137bSQu Wenruo return ret; 106529db137bSQu Wenruo } 106629db137bSQu Wenruo 106729db137bSQu Wenruo /* 1068e7d571c7SQu Wenruo * Build backref tree for a given tree block. Root of the backref tree 1069e7d571c7SQu Wenruo * corresponds the tree block, leaves of the backref tree correspond roots of 1070e7d571c7SQu Wenruo * b-trees that reference the tree block. 1071e7d571c7SQu Wenruo * 1072e7d571c7SQu Wenruo * The basic idea of this function is check backrefs of a given block to find 1073e7d571c7SQu Wenruo * upper level blocks that reference the block, and then check backrefs of 1074e7d571c7SQu Wenruo * these upper level blocks recursively. The recursion stops when tree root is 1075e7d571c7SQu Wenruo * reached or backrefs for the block is cached. 1076e7d571c7SQu Wenruo * 1077e7d571c7SQu Wenruo * NOTE: if we find that backrefs for a block are cached, we know backrefs for 1078e7d571c7SQu Wenruo * all upper level blocks that directly/indirectly reference the block are also 1079e7d571c7SQu Wenruo * cached. 1080e7d571c7SQu Wenruo */ 1081a26195a5SQu Wenruo static noinline_for_stack struct btrfs_backref_node *build_backref_tree( 1082e7d571c7SQu Wenruo struct reloc_control *rc, struct btrfs_key *node_key, 1083e7d571c7SQu Wenruo int level, u64 bytenr) 1084e7d571c7SQu Wenruo { 1085e7d571c7SQu Wenruo struct btrfs_backref_iter *iter; 1086a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 1087e7d571c7SQu Wenruo /* For searching parent of TREE_BLOCK_REF */ 1088e7d571c7SQu Wenruo struct btrfs_path *path; 1089a26195a5SQu Wenruo struct btrfs_backref_node *cur; 1090a26195a5SQu Wenruo struct btrfs_backref_node *upper; 1091a26195a5SQu Wenruo struct btrfs_backref_node *lower; 1092a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 1093a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1094e7d571c7SQu Wenruo int ret; 1095e7d571c7SQu Wenruo int err = 0; 1096e7d571c7SQu Wenruo 1097e7d571c7SQu Wenruo iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS); 1098e7d571c7SQu Wenruo if (!iter) 1099e7d571c7SQu Wenruo return ERR_PTR(-ENOMEM); 1100e7d571c7SQu Wenruo path = btrfs_alloc_path(); 1101e7d571c7SQu Wenruo if (!path) { 1102e7d571c7SQu Wenruo err = -ENOMEM; 1103e7d571c7SQu Wenruo goto out; 1104e7d571c7SQu Wenruo } 1105e7d571c7SQu Wenruo 1106b1818dabSQu Wenruo node = btrfs_backref_alloc_node(cache, bytenr, level); 1107e7d571c7SQu Wenruo if (!node) { 1108e7d571c7SQu Wenruo err = -ENOMEM; 1109e7d571c7SQu Wenruo goto out; 1110e7d571c7SQu Wenruo } 1111e7d571c7SQu Wenruo 1112e7d571c7SQu Wenruo node->lowest = 1; 1113e7d571c7SQu Wenruo cur = node; 1114e7d571c7SQu Wenruo 1115e7d571c7SQu Wenruo /* Breadth-first search to build backref cache */ 1116e7d571c7SQu Wenruo do { 1117e7d571c7SQu Wenruo ret = handle_one_tree_block(cache, path, iter, node_key, cur); 1118e7d571c7SQu Wenruo if (ret < 0) { 1119e7d571c7SQu Wenruo err = ret; 1120e7d571c7SQu Wenruo goto out; 1121e7d571c7SQu Wenruo } 1122e7d571c7SQu Wenruo edge = list_first_entry_or_null(&cache->pending_edge, 1123a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 1124e7d571c7SQu Wenruo /* 1125e7d571c7SQu Wenruo * The pending list isn't empty, take the first block to 1126e7d571c7SQu Wenruo * process 1127e7d571c7SQu Wenruo */ 1128e7d571c7SQu Wenruo if (edge) { 11295d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 11305d4f98a2SYan Zheng cur = edge->node[UPPER]; 11315d4f98a2SYan Zheng } 1132e7d571c7SQu Wenruo } while (edge); 11335d4f98a2SYan Zheng 11341f872924SQu Wenruo /* Finish the upper linkage of newly added edges/nodes */ 11351f872924SQu Wenruo ret = finish_upper_links(cache, node); 11361f872924SQu Wenruo if (ret < 0) { 11371f872924SQu Wenruo err = ret; 113875bfb9afSJosef Bacik goto out; 113975bfb9afSJosef Bacik } 114075bfb9afSJosef Bacik 114129db137bSQu Wenruo if (handle_useless_nodes(rc, node)) 11423fd0a558SYan, Zheng node = NULL; 11435d4f98a2SYan Zheng out: 114471f572a9SQu Wenruo btrfs_backref_iter_free(iter); 114571f572a9SQu Wenruo btrfs_free_path(path); 11465d4f98a2SYan Zheng if (err) { 114784780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 114884780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 1149a26195a5SQu Wenruo struct btrfs_backref_node, list); 115075bfb9afSJosef Bacik list_del_init(&lower->list); 11513fd0a558SYan, Zheng } 115284780289SQu Wenruo while (!list_empty(&cache->pending_edge)) { 115384780289SQu Wenruo edge = list_first_entry(&cache->pending_edge, 1154a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 115575bfb9afSJosef Bacik list_del(&edge->list[UPPER]); 11563fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 115775bfb9afSJosef Bacik lower = edge->node[LOWER]; 11585d4f98a2SYan Zheng upper = edge->node[UPPER]; 11593fd0a558SYan, Zheng free_backref_edge(cache, edge); 116075bfb9afSJosef Bacik 116175bfb9afSJosef Bacik /* 116275bfb9afSJosef Bacik * Lower is no longer linked to any upper backref nodes 116375bfb9afSJosef Bacik * and isn't in the cache, we can free it ourselves. 116475bfb9afSJosef Bacik */ 116575bfb9afSJosef Bacik if (list_empty(&lower->upper) && 116675bfb9afSJosef Bacik RB_EMPTY_NODE(&lower->rb_node)) 116784780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 116875bfb9afSJosef Bacik 116975bfb9afSJosef Bacik if (!RB_EMPTY_NODE(&upper->rb_node)) 117075bfb9afSJosef Bacik continue; 117175bfb9afSJosef Bacik 117201327610SNicholas D Steeves /* Add this guy's upper edges to the list to process */ 117375bfb9afSJosef Bacik list_for_each_entry(edge, &upper->upper, list[LOWER]) 117484780289SQu Wenruo list_add_tail(&edge->list[UPPER], 117584780289SQu Wenruo &cache->pending_edge); 117675bfb9afSJosef Bacik if (list_empty(&upper->upper)) 117784780289SQu Wenruo list_add(&upper->list, &cache->useless_node); 117875bfb9afSJosef Bacik } 117975bfb9afSJosef Bacik 118084780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 118184780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 1182a26195a5SQu Wenruo struct btrfs_backref_node, list); 118375bfb9afSJosef Bacik list_del_init(&lower->list); 11840fd8c3daSLiu Bo if (lower == node) 11850fd8c3daSLiu Bo node = NULL; 118675bfb9afSJosef Bacik free_backref_node(cache, lower); 11875d4f98a2SYan Zheng } 11880fd8c3daSLiu Bo 11898e19c973SJosef Bacik remove_backref_node(cache, node); 119084780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 119184780289SQu Wenruo list_empty(&cache->pending_edge)); 11925d4f98a2SYan Zheng return ERR_PTR(err); 11935d4f98a2SYan Zheng } 119475bfb9afSJosef Bacik ASSERT(!node || !node->detached); 119584780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 119684780289SQu Wenruo list_empty(&cache->pending_edge)); 11975d4f98a2SYan Zheng return node; 11985d4f98a2SYan Zheng } 11995d4f98a2SYan Zheng 12005d4f98a2SYan Zheng /* 12013fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 12023fd0a558SYan, Zheng * the backref node is created by cloning backref node that 12033fd0a558SYan, Zheng * corresponds to root of source tree 12043fd0a558SYan, Zheng */ 12053fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 12063fd0a558SYan, Zheng struct reloc_control *rc, 12073fd0a558SYan, Zheng struct btrfs_root *src, 12083fd0a558SYan, Zheng struct btrfs_root *dest) 12093fd0a558SYan, Zheng { 12103fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 1211a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 1212a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 1213a26195a5SQu Wenruo struct btrfs_backref_node *new_node; 1214a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1215a26195a5SQu Wenruo struct btrfs_backref_edge *new_edge; 12163fd0a558SYan, Zheng struct rb_node *rb_node; 12173fd0a558SYan, Zheng 12183fd0a558SYan, Zheng if (cache->last_trans > 0) 12193fd0a558SYan, Zheng update_backref_cache(trans, cache); 12203fd0a558SYan, Zheng 1221e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start); 12223fd0a558SYan, Zheng if (rb_node) { 1223a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, rb_node); 12243fd0a558SYan, Zheng if (node->detached) 12253fd0a558SYan, Zheng node = NULL; 12263fd0a558SYan, Zheng else 12273fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 12283fd0a558SYan, Zheng } 12293fd0a558SYan, Zheng 12303fd0a558SYan, Zheng if (!node) { 1231e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, 12323fd0a558SYan, Zheng reloc_root->commit_root->start); 12333fd0a558SYan, Zheng if (rb_node) { 1234a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, 12353fd0a558SYan, Zheng rb_node); 12363fd0a558SYan, Zheng BUG_ON(node->detached); 12373fd0a558SYan, Zheng } 12383fd0a558SYan, Zheng } 12393fd0a558SYan, Zheng 12403fd0a558SYan, Zheng if (!node) 12413fd0a558SYan, Zheng return 0; 12423fd0a558SYan, Zheng 1243b1818dabSQu Wenruo new_node = btrfs_backref_alloc_node(cache, dest->node->start, 1244b1818dabSQu Wenruo node->level); 12453fd0a558SYan, Zheng if (!new_node) 12463fd0a558SYan, Zheng return -ENOMEM; 12473fd0a558SYan, Zheng 12483fd0a558SYan, Zheng new_node->lowest = node->lowest; 12496848ad64SYan, Zheng new_node->checked = 1; 125000246528SJosef Bacik new_node->root = btrfs_grab_root(dest); 12510b530bc5SJosef Bacik ASSERT(new_node->root); 12523fd0a558SYan, Zheng 12533fd0a558SYan, Zheng if (!node->lowest) { 12543fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 125547254d07SQu Wenruo new_edge = btrfs_backref_alloc_edge(cache); 12563fd0a558SYan, Zheng if (!new_edge) 12573fd0a558SYan, Zheng goto fail; 12583fd0a558SYan, Zheng 12592a979612SQu Wenruo link_backref_edge(new_edge, edge->node[LOWER], new_node, 12602a979612SQu Wenruo LINK_UPPER); 12613fd0a558SYan, Zheng } 126276b9e23dSMiao Xie } else { 126376b9e23dSMiao Xie list_add_tail(&new_node->lower, &cache->leaves); 12643fd0a558SYan, Zheng } 12653fd0a558SYan, Zheng 1266e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr, 12673fd0a558SYan, Zheng &new_node->rb_node); 126843c04fb1SJeff Mahoney if (rb_node) 126943c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, new_node->bytenr); 12703fd0a558SYan, Zheng 12713fd0a558SYan, Zheng if (!new_node->lowest) { 12723fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 12733fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 12743fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 12753fd0a558SYan, Zheng } 12763fd0a558SYan, Zheng } 12773fd0a558SYan, Zheng return 0; 12783fd0a558SYan, Zheng fail: 12793fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 12803fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 1281a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 12823fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 12833fd0a558SYan, Zheng free_backref_edge(cache, new_edge); 12843fd0a558SYan, Zheng } 12853fd0a558SYan, Zheng free_backref_node(cache, new_node); 12863fd0a558SYan, Zheng return -ENOMEM; 12873fd0a558SYan, Zheng } 12883fd0a558SYan, Zheng 12893fd0a558SYan, Zheng /* 12905d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 12915d4f98a2SYan Zheng */ 1292ffd7b339SJeff Mahoney static int __must_check __add_reloc_root(struct btrfs_root *root) 12935d4f98a2SYan Zheng { 12940b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 12955d4f98a2SYan Zheng struct rb_node *rb_node; 12965d4f98a2SYan Zheng struct mapping_node *node; 12970b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 12985d4f98a2SYan Zheng 12995d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 1300ffd7b339SJeff Mahoney if (!node) 1301ffd7b339SJeff Mahoney return -ENOMEM; 13025d4f98a2SYan Zheng 1303ea287ab1SJosef Bacik node->bytenr = root->commit_root->start; 13045d4f98a2SYan Zheng node->data = root; 13055d4f98a2SYan Zheng 13065d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1307e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 13085d4f98a2SYan Zheng node->bytenr, &node->rb_node); 13095d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 1310ffd7b339SJeff Mahoney if (rb_node) { 13110b246afaSJeff Mahoney btrfs_panic(fs_info, -EEXIST, 13125d163e0eSJeff Mahoney "Duplicate root found for start=%llu while inserting into relocation tree", 13135d163e0eSJeff Mahoney node->bytenr); 1314ffd7b339SJeff Mahoney } 13155d4f98a2SYan Zheng 13165d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 13175d4f98a2SYan Zheng return 0; 13185d4f98a2SYan Zheng } 13195d4f98a2SYan Zheng 13205d4f98a2SYan Zheng /* 1321c974c464SWang Shilong * helper to delete the 'address of tree root -> reloc tree' 13225d4f98a2SYan Zheng * mapping 13235d4f98a2SYan Zheng */ 1324c974c464SWang Shilong static void __del_reloc_root(struct btrfs_root *root) 13255d4f98a2SYan Zheng { 13260b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 13275d4f98a2SYan Zheng struct rb_node *rb_node; 13285d4f98a2SYan Zheng struct mapping_node *node = NULL; 13290b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1330f44deb74SJosef Bacik bool put_ref = false; 13315d4f98a2SYan Zheng 133265c6e82bSQu Wenruo if (rc && root->node) { 13335d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1334e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 1335ea287ab1SJosef Bacik root->commit_root->start); 1336c974c464SWang Shilong if (rb_node) { 1337c974c464SWang Shilong node = rb_entry(rb_node, struct mapping_node, rb_node); 1338c974c464SWang Shilong rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 1339ea287ab1SJosef Bacik RB_CLEAR_NODE(&node->rb_node); 1340c974c464SWang Shilong } 1341c974c464SWang Shilong spin_unlock(&rc->reloc_root_tree.lock); 1342c974c464SWang Shilong if (!node) 1343c974c464SWang Shilong return; 1344c974c464SWang Shilong BUG_ON((struct btrfs_root *)node->data != root); 1345389305b2SQu Wenruo } 1346c974c464SWang Shilong 1347f44deb74SJosef Bacik /* 1348f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 1349f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 1350f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 1351f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 1352f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 1353f44deb74SJosef Bacik * messing with the list is in charge of the reference. 1354f44deb74SJosef Bacik */ 13550b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 1356f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 1357f44deb74SJosef Bacik put_ref = true; 1358c974c464SWang Shilong list_del_init(&root->root_list); 1359f44deb74SJosef Bacik } 13600b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 1361f44deb74SJosef Bacik if (put_ref) 1362f44deb74SJosef Bacik btrfs_put_root(root); 1363c974c464SWang Shilong kfree(node); 1364c974c464SWang Shilong } 1365c974c464SWang Shilong 1366c974c464SWang Shilong /* 1367c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 1368c974c464SWang Shilong * mapping 1369c974c464SWang Shilong */ 1370ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 1371c974c464SWang Shilong { 13720b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 1373c974c464SWang Shilong struct rb_node *rb_node; 1374c974c464SWang Shilong struct mapping_node *node = NULL; 13750b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1376c974c464SWang Shilong 1377c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 1378e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 1379ea287ab1SJosef Bacik root->commit_root->start); 13805d4f98a2SYan Zheng if (rb_node) { 13815d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 13825d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 13835d4f98a2SYan Zheng } 13845d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 13855d4f98a2SYan Zheng 13868f71f3e0SLiu Bo if (!node) 13878f71f3e0SLiu Bo return 0; 13885d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 13895d4f98a2SYan Zheng 13905d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1391ea287ab1SJosef Bacik node->bytenr = root->node->start; 1392e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 13935d4f98a2SYan Zheng node->bytenr, &node->rb_node); 13945d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 139543c04fb1SJeff Mahoney if (rb_node) 139643c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 13975d4f98a2SYan Zheng return 0; 13985d4f98a2SYan Zheng } 13995d4f98a2SYan Zheng 14003fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 14013fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 14025d4f98a2SYan Zheng { 14030b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14045d4f98a2SYan Zheng struct btrfs_root *reloc_root; 14055d4f98a2SYan Zheng struct extent_buffer *eb; 14065d4f98a2SYan Zheng struct btrfs_root_item *root_item; 14075d4f98a2SYan Zheng struct btrfs_key root_key; 14085d4f98a2SYan Zheng int ret; 14095d4f98a2SYan Zheng 14105d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 14115d4f98a2SYan Zheng BUG_ON(!root_item); 14125d4f98a2SYan Zheng 14135d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 14145d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 14153fd0a558SYan, Zheng root_key.offset = objectid; 14165d4f98a2SYan Zheng 14173fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 1418054570a1SFilipe Manana u64 commit_root_gen; 1419054570a1SFilipe Manana 14203fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 14215d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 14225d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 14235d4f98a2SYan Zheng BUG_ON(ret); 1424054570a1SFilipe Manana /* 1425054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 1426054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 1427054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 1428054570a1SFilipe Manana * either inside the critical section of a transaction commit 1429054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 1430054570a1SFilipe Manana * it's created before the transaction commit is started. 1431054570a1SFilipe Manana */ 1432054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 1433054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 14343fd0a558SYan, Zheng } else { 14353fd0a558SYan, Zheng /* 14363fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 14373fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 14383fd0a558SYan, Zheng * modified after it was created have RELOC flag 14393fd0a558SYan, Zheng * set in their headers. so it's OK to not update 14403fd0a558SYan, Zheng * the 'last_snapshot'. 14413fd0a558SYan, Zheng */ 14423fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 14433fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 14443fd0a558SYan, Zheng BUG_ON(ret); 14453fd0a558SYan, Zheng } 14463fd0a558SYan, Zheng 14475d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 14485d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 14495d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 14505d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 14513fd0a558SYan, Zheng 14523fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 14533fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 14543fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 14553fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 14565d4f98a2SYan Zheng root_item->drop_level = 0; 14573fd0a558SYan, Zheng } 14585d4f98a2SYan Zheng 14595d4f98a2SYan Zheng btrfs_tree_unlock(eb); 14605d4f98a2SYan Zheng free_extent_buffer(eb); 14615d4f98a2SYan Zheng 14620b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 14635d4f98a2SYan Zheng &root_key, root_item); 14645d4f98a2SYan Zheng BUG_ON(ret); 14655d4f98a2SYan Zheng kfree(root_item); 14665d4f98a2SYan Zheng 14673dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 14685d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 14693dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 14705d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 14713fd0a558SYan, Zheng return reloc_root; 14723fd0a558SYan, Zheng } 14733fd0a558SYan, Zheng 14743fd0a558SYan, Zheng /* 14753fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 14763fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 1477f44deb74SJosef Bacik * 1478f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 1479f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 14803fd0a558SYan, Zheng */ 14813fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 14823fd0a558SYan, Zheng struct btrfs_root *root) 14833fd0a558SYan, Zheng { 14840b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14853fd0a558SYan, Zheng struct btrfs_root *reloc_root; 14860b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 148720dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 14883fd0a558SYan, Zheng int clear_rsv = 0; 1489ffd7b339SJeff Mahoney int ret; 14903fd0a558SYan, Zheng 1491aec7db3bSJosef Bacik if (!rc) 14922abc726aSJosef Bacik return 0; 14932abc726aSJosef Bacik 14941fac4a54SQu Wenruo /* 14951fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 14961fac4a54SQu Wenruo * create/update the dead reloc tree 14971fac4a54SQu Wenruo */ 14986282675eSQu Wenruo if (reloc_root_is_dead(root)) 14991fac4a54SQu Wenruo return 0; 15001fac4a54SQu Wenruo 1501aec7db3bSJosef Bacik /* 1502aec7db3bSJosef Bacik * This is subtle but important. We do not do 1503aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 1504aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 1505aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 1506aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 1507aec7db3bSJosef Bacik * in. 1508aec7db3bSJosef Bacik */ 15093fd0a558SYan, Zheng if (root->reloc_root) { 15103fd0a558SYan, Zheng reloc_root = root->reloc_root; 15113fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 15123fd0a558SYan, Zheng return 0; 15133fd0a558SYan, Zheng } 15143fd0a558SYan, Zheng 1515aec7db3bSJosef Bacik /* 1516aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 1517aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 1518aec7db3bSJosef Bacik */ 1519aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 1520aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 1521aec7db3bSJosef Bacik return 0; 1522aec7db3bSJosef Bacik 152320dd2cbfSMiao Xie if (!trans->reloc_reserved) { 152420dd2cbfSMiao Xie rsv = trans->block_rsv; 15253fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 15263fd0a558SYan, Zheng clear_rsv = 1; 15273fd0a558SYan, Zheng } 15283fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 15293fd0a558SYan, Zheng if (clear_rsv) 153020dd2cbfSMiao Xie trans->block_rsv = rsv; 15315d4f98a2SYan Zheng 1532ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 1533ffd7b339SJeff Mahoney BUG_ON(ret < 0); 1534f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 15355d4f98a2SYan Zheng return 0; 15365d4f98a2SYan Zheng } 15375d4f98a2SYan Zheng 15385d4f98a2SYan Zheng /* 15395d4f98a2SYan Zheng * update root item of reloc tree 15405d4f98a2SYan Zheng */ 15415d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 15425d4f98a2SYan Zheng struct btrfs_root *root) 15435d4f98a2SYan Zheng { 15440b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15455d4f98a2SYan Zheng struct btrfs_root *reloc_root; 15465d4f98a2SYan Zheng struct btrfs_root_item *root_item; 15475d4f98a2SYan Zheng int ret; 15485d4f98a2SYan Zheng 15496282675eSQu Wenruo if (!have_reloc_root(root)) 15507585717fSChris Mason goto out; 15515d4f98a2SYan Zheng 15525d4f98a2SYan Zheng reloc_root = root->reloc_root; 15535d4f98a2SYan Zheng root_item = &reloc_root->root_item; 15545d4f98a2SYan Zheng 1555f44deb74SJosef Bacik /* 1556f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 1557f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 1558f44deb74SJosef Bacik * hold it while we update the reloc root. 1559f44deb74SJosef Bacik */ 1560f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 1561f44deb74SJosef Bacik 1562d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 15630b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 15643fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 1565d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 15666282675eSQu Wenruo /* 15676282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 15686282675eSQu Wenruo * have_reloc_root will not touch it from now on. 15696282675eSQu Wenruo */ 15706282675eSQu Wenruo smp_wmb(); 1571c974c464SWang Shilong __del_reloc_root(reloc_root); 15725d4f98a2SYan Zheng } 15735d4f98a2SYan Zheng 15745d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 1575ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 15765d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 15775d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 15785d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 15795d4f98a2SYan Zheng } 15805d4f98a2SYan Zheng 15810b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 15825d4f98a2SYan Zheng &reloc_root->root_key, root_item); 15835d4f98a2SYan Zheng BUG_ON(ret); 1584f44deb74SJosef Bacik btrfs_put_root(reloc_root); 15857585717fSChris Mason out: 15865d4f98a2SYan Zheng return 0; 15875d4f98a2SYan Zheng } 15885d4f98a2SYan Zheng 15895d4f98a2SYan Zheng /* 15905d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 15915d4f98a2SYan Zheng * in a subvolume 15925d4f98a2SYan Zheng */ 15935d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 15945d4f98a2SYan Zheng { 15955d4f98a2SYan Zheng struct rb_node *node; 15965d4f98a2SYan Zheng struct rb_node *prev; 15975d4f98a2SYan Zheng struct btrfs_inode *entry; 15985d4f98a2SYan Zheng struct inode *inode; 15995d4f98a2SYan Zheng 16005d4f98a2SYan Zheng spin_lock(&root->inode_lock); 16015d4f98a2SYan Zheng again: 16025d4f98a2SYan Zheng node = root->inode_tree.rb_node; 16035d4f98a2SYan Zheng prev = NULL; 16045d4f98a2SYan Zheng while (node) { 16055d4f98a2SYan Zheng prev = node; 16065d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 16075d4f98a2SYan Zheng 16084a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 16095d4f98a2SYan Zheng node = node->rb_left; 16104a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 16115d4f98a2SYan Zheng node = node->rb_right; 16125d4f98a2SYan Zheng else 16135d4f98a2SYan Zheng break; 16145d4f98a2SYan Zheng } 16155d4f98a2SYan Zheng if (!node) { 16165d4f98a2SYan Zheng while (prev) { 16175d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 16184a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 16195d4f98a2SYan Zheng node = prev; 16205d4f98a2SYan Zheng break; 16215d4f98a2SYan Zheng } 16225d4f98a2SYan Zheng prev = rb_next(prev); 16235d4f98a2SYan Zheng } 16245d4f98a2SYan Zheng } 16255d4f98a2SYan Zheng while (node) { 16265d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 16275d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 16285d4f98a2SYan Zheng if (inode) { 16295d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 16305d4f98a2SYan Zheng return inode; 16315d4f98a2SYan Zheng } 16325d4f98a2SYan Zheng 16334a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 16345d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 16355d4f98a2SYan Zheng goto again; 16365d4f98a2SYan Zheng 16375d4f98a2SYan Zheng node = rb_next(node); 16385d4f98a2SYan Zheng } 16395d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 16405d4f98a2SYan Zheng return NULL; 16415d4f98a2SYan Zheng } 16425d4f98a2SYan Zheng 16435d4f98a2SYan Zheng /* 16445d4f98a2SYan Zheng * get new location of data 16455d4f98a2SYan Zheng */ 16465d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 16475d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 16485d4f98a2SYan Zheng { 16495d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 16505d4f98a2SYan Zheng struct btrfs_path *path; 16515d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 16525d4f98a2SYan Zheng struct extent_buffer *leaf; 16535d4f98a2SYan Zheng int ret; 16545d4f98a2SYan Zheng 16555d4f98a2SYan Zheng path = btrfs_alloc_path(); 16565d4f98a2SYan Zheng if (!path) 16575d4f98a2SYan Zheng return -ENOMEM; 16585d4f98a2SYan Zheng 16595d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 1660f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 1661f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 16625d4f98a2SYan Zheng if (ret < 0) 16635d4f98a2SYan Zheng goto out; 16645d4f98a2SYan Zheng if (ret > 0) { 16655d4f98a2SYan Zheng ret = -ENOENT; 16665d4f98a2SYan Zheng goto out; 16675d4f98a2SYan Zheng } 16685d4f98a2SYan Zheng 16695d4f98a2SYan Zheng leaf = path->nodes[0]; 16705d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 16715d4f98a2SYan Zheng struct btrfs_file_extent_item); 16725d4f98a2SYan Zheng 16735d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 16745d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 16755d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 16765d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 16775d4f98a2SYan Zheng 16785d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 167983d4cfd4SJosef Bacik ret = -EINVAL; 16805d4f98a2SYan Zheng goto out; 16815d4f98a2SYan Zheng } 16825d4f98a2SYan Zheng 16835d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 16845d4f98a2SYan Zheng ret = 0; 16855d4f98a2SYan Zheng out: 16865d4f98a2SYan Zheng btrfs_free_path(path); 16875d4f98a2SYan Zheng return ret; 16885d4f98a2SYan Zheng } 16895d4f98a2SYan Zheng 16905d4f98a2SYan Zheng /* 16915d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 16925d4f98a2SYan Zheng * the new locations. 16935d4f98a2SYan Zheng */ 16943fd0a558SYan, Zheng static noinline_for_stack 16953fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 16965d4f98a2SYan Zheng struct reloc_control *rc, 16975d4f98a2SYan Zheng struct btrfs_root *root, 16983fd0a558SYan, Zheng struct extent_buffer *leaf) 16995d4f98a2SYan Zheng { 17000b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 17015d4f98a2SYan Zheng struct btrfs_key key; 17025d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 17035d4f98a2SYan Zheng struct inode *inode = NULL; 17045d4f98a2SYan Zheng u64 parent; 17055d4f98a2SYan Zheng u64 bytenr; 17063fd0a558SYan, Zheng u64 new_bytenr = 0; 17075d4f98a2SYan Zheng u64 num_bytes; 17085d4f98a2SYan Zheng u64 end; 17095d4f98a2SYan Zheng u32 nritems; 17105d4f98a2SYan Zheng u32 i; 171183d4cfd4SJosef Bacik int ret = 0; 17125d4f98a2SYan Zheng int first = 1; 17135d4f98a2SYan Zheng int dirty = 0; 17145d4f98a2SYan Zheng 17155d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 17165d4f98a2SYan Zheng return 0; 17175d4f98a2SYan Zheng 17185d4f98a2SYan Zheng /* reloc trees always use full backref */ 17195d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 17205d4f98a2SYan Zheng parent = leaf->start; 17215d4f98a2SYan Zheng else 17225d4f98a2SYan Zheng parent = 0; 17235d4f98a2SYan Zheng 17245d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 17255d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 172682fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 172782fa113fSQu Wenruo 17285d4f98a2SYan Zheng cond_resched(); 17295d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 17305d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 17315d4f98a2SYan Zheng continue; 17325d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 17335d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 17345d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 17355d4f98a2SYan Zheng continue; 17365d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 17375d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 17385d4f98a2SYan Zheng if (bytenr == 0) 17395d4f98a2SYan Zheng continue; 17409569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 17419569cc20SQu Wenruo rc->block_group->length)) 17425d4f98a2SYan Zheng continue; 17435d4f98a2SYan Zheng 17445d4f98a2SYan Zheng /* 17455d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 17465d4f98a2SYan Zheng * to complete and drop the extent cache 17475d4f98a2SYan Zheng */ 17485d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 17495d4f98a2SYan Zheng if (first) { 17505d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 17515d4f98a2SYan Zheng first = 0; 17524a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 17533fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 17545d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 17555d4f98a2SYan Zheng } 17564a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 17575d4f98a2SYan Zheng end = key.offset + 17585d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 17595d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 17600b246afaSJeff Mahoney fs_info->sectorsize)); 17610b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 17625d4f98a2SYan Zheng end--; 17635d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1764d0082371SJeff Mahoney key.offset, end); 17655d4f98a2SYan Zheng if (!ret) 17665d4f98a2SYan Zheng continue; 17675d4f98a2SYan Zheng 1768dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1769dcdbc059SNikolay Borisov key.offset, end, 1); 17705d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1771d0082371SJeff Mahoney key.offset, end); 17725d4f98a2SYan Zheng } 17735d4f98a2SYan Zheng } 17745d4f98a2SYan Zheng 17755d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 17765d4f98a2SYan Zheng bytenr, num_bytes); 177783d4cfd4SJosef Bacik if (ret) { 177883d4cfd4SJosef Bacik /* 177983d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 178083d4cfd4SJosef Bacik * in the file extent yet. 178183d4cfd4SJosef Bacik */ 178283d4cfd4SJosef Bacik break; 17833fd0a558SYan, Zheng } 17845d4f98a2SYan Zheng 17855d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 17865d4f98a2SYan Zheng dirty = 1; 17875d4f98a2SYan Zheng 17885d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 178982fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 179082fa113fSQu Wenruo num_bytes, parent); 179182fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 179282fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1793b06c4bf5SFilipe Manana key.objectid, key.offset); 179482fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 179583d4cfd4SJosef Bacik if (ret) { 179666642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 179783d4cfd4SJosef Bacik break; 179883d4cfd4SJosef Bacik } 17995d4f98a2SYan Zheng 1800ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1801ffd4bb2aSQu Wenruo num_bytes, parent); 1802ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1803ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1804b06c4bf5SFilipe Manana key.objectid, key.offset); 1805ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 180683d4cfd4SJosef Bacik if (ret) { 180766642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 180883d4cfd4SJosef Bacik break; 180983d4cfd4SJosef Bacik } 18105d4f98a2SYan Zheng } 18115d4f98a2SYan Zheng if (dirty) 18125d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 18133fd0a558SYan, Zheng if (inode) 18143fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 181583d4cfd4SJosef Bacik return ret; 18165d4f98a2SYan Zheng } 18175d4f98a2SYan Zheng 18185d4f98a2SYan Zheng static noinline_for_stack 18195d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 18205d4f98a2SYan Zheng struct btrfs_path *path, int level) 18215d4f98a2SYan Zheng { 18225d4f98a2SYan Zheng struct btrfs_disk_key key1; 18235d4f98a2SYan Zheng struct btrfs_disk_key key2; 18245d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 18255d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 18265d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 18275d4f98a2SYan Zheng } 18285d4f98a2SYan Zheng 18295d4f98a2SYan Zheng /* 18305d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 18315d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 18325d4f98a2SYan Zheng * reloc tree was create can be replaced. 18335d4f98a2SYan Zheng * 18345d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 18355d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 18365d4f98a2SYan Zheng * errors, a negative error number is returned. 18375d4f98a2SYan Zheng */ 18383fd0a558SYan, Zheng static noinline_for_stack 18393d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 18405d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 18415d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 18425d4f98a2SYan Zheng int lowest_level, int max_level) 18435d4f98a2SYan Zheng { 18440b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 18455d4f98a2SYan Zheng struct extent_buffer *eb; 18465d4f98a2SYan Zheng struct extent_buffer *parent; 184782fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 18485d4f98a2SYan Zheng struct btrfs_key key; 18495d4f98a2SYan Zheng u64 old_bytenr; 18505d4f98a2SYan Zheng u64 new_bytenr; 18515d4f98a2SYan Zheng u64 old_ptr_gen; 18525d4f98a2SYan Zheng u64 new_ptr_gen; 18535d4f98a2SYan Zheng u64 last_snapshot; 18545d4f98a2SYan Zheng u32 blocksize; 18553fd0a558SYan, Zheng int cow = 0; 18565d4f98a2SYan Zheng int level; 18575d4f98a2SYan Zheng int ret; 18585d4f98a2SYan Zheng int slot; 18595d4f98a2SYan Zheng 18605d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 18615d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 18625d4f98a2SYan Zheng 18635d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 18643fd0a558SYan, Zheng again: 18655d4f98a2SYan Zheng slot = path->slots[lowest_level]; 18665d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 18675d4f98a2SYan Zheng 18685d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 18698bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 18705d4f98a2SYan Zheng level = btrfs_header_level(eb); 18715d4f98a2SYan Zheng 18725d4f98a2SYan Zheng if (level < lowest_level) { 18735d4f98a2SYan Zheng btrfs_tree_unlock(eb); 18745d4f98a2SYan Zheng free_extent_buffer(eb); 18755d4f98a2SYan Zheng return 0; 18765d4f98a2SYan Zheng } 18775d4f98a2SYan Zheng 18783fd0a558SYan, Zheng if (cow) { 18795d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 18805d4f98a2SYan Zheng BUG_ON(ret); 18813fd0a558SYan, Zheng } 18828bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 18835d4f98a2SYan Zheng 18845d4f98a2SYan Zheng if (next_key) { 18855d4f98a2SYan Zheng next_key->objectid = (u64)-1; 18865d4f98a2SYan Zheng next_key->type = (u8)-1; 18875d4f98a2SYan Zheng next_key->offset = (u64)-1; 18885d4f98a2SYan Zheng } 18895d4f98a2SYan Zheng 18905d4f98a2SYan Zheng parent = eb; 18915d4f98a2SYan Zheng while (1) { 1892581c1760SQu Wenruo struct btrfs_key first_key; 1893581c1760SQu Wenruo 18945d4f98a2SYan Zheng level = btrfs_header_level(parent); 18955d4f98a2SYan Zheng BUG_ON(level < lowest_level); 18965d4f98a2SYan Zheng 18975d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 1898cbca7d59SFilipe Manana if (ret < 0) 1899cbca7d59SFilipe Manana break; 19005d4f98a2SYan Zheng if (ret && slot > 0) 19015d4f98a2SYan Zheng slot--; 19025d4f98a2SYan Zheng 19035d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 19045d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 19055d4f98a2SYan Zheng 19065d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 19070b246afaSJeff Mahoney blocksize = fs_info->nodesize; 19085d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 190917515f1bSQu Wenruo btrfs_node_key_to_cpu(parent, &first_key, slot); 19105d4f98a2SYan Zheng 19115d4f98a2SYan Zheng if (level <= max_level) { 19125d4f98a2SYan Zheng eb = path->nodes[level]; 19135d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 19145d4f98a2SYan Zheng path->slots[level]); 19155d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 19165d4f98a2SYan Zheng path->slots[level]); 19175d4f98a2SYan Zheng } else { 19185d4f98a2SYan Zheng new_bytenr = 0; 19195d4f98a2SYan Zheng new_ptr_gen = 0; 19205d4f98a2SYan Zheng } 19215d4f98a2SYan Zheng 1922fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 19235d4f98a2SYan Zheng ret = level; 19245d4f98a2SYan Zheng break; 19255d4f98a2SYan Zheng } 19265d4f98a2SYan Zheng 19275d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 19285d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 19293fd0a558SYan, Zheng if (level <= lowest_level) { 19305d4f98a2SYan Zheng ret = 0; 19315d4f98a2SYan Zheng break; 19325d4f98a2SYan Zheng } 19335d4f98a2SYan Zheng 1934581c1760SQu Wenruo eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen, 1935581c1760SQu Wenruo level - 1, &first_key); 193664c043deSLiu Bo if (IS_ERR(eb)) { 193764c043deSLiu Bo ret = PTR_ERR(eb); 1938264813acSLiu Bo break; 193964c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 194064c043deSLiu Bo ret = -EIO; 1941416bc658SJosef Bacik free_extent_buffer(eb); 1942379cde74SStefan Behrens break; 1943416bc658SJosef Bacik } 19445d4f98a2SYan Zheng btrfs_tree_lock(eb); 19453fd0a558SYan, Zheng if (cow) { 19465d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 19475d4f98a2SYan Zheng slot, &eb); 19485d4f98a2SYan Zheng BUG_ON(ret); 19495d4f98a2SYan Zheng } 19508bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19515d4f98a2SYan Zheng 19525d4f98a2SYan Zheng btrfs_tree_unlock(parent); 19535d4f98a2SYan Zheng free_extent_buffer(parent); 19545d4f98a2SYan Zheng 19555d4f98a2SYan Zheng parent = eb; 19565d4f98a2SYan Zheng continue; 19575d4f98a2SYan Zheng } 19585d4f98a2SYan Zheng 19593fd0a558SYan, Zheng if (!cow) { 19603fd0a558SYan, Zheng btrfs_tree_unlock(parent); 19613fd0a558SYan, Zheng free_extent_buffer(parent); 19623fd0a558SYan, Zheng cow = 1; 19633fd0a558SYan, Zheng goto again; 19643fd0a558SYan, Zheng } 19653fd0a558SYan, Zheng 19665d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 19675d4f98a2SYan Zheng path->slots[level]); 1968b3b4aa74SDavid Sterba btrfs_release_path(path); 19695d4f98a2SYan Zheng 19705d4f98a2SYan Zheng path->lowest_level = level; 19715d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 19725d4f98a2SYan Zheng path->lowest_level = 0; 19735d4f98a2SYan Zheng BUG_ON(ret); 19745d4f98a2SYan Zheng 19755d4f98a2SYan Zheng /* 1976824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 1977824d8dffSQu Wenruo * 1978824d8dffSQu Wenruo * We must trace both trees. 1979824d8dffSQu Wenruo * 1) Tree reloc subtree 1980824d8dffSQu Wenruo * If not traced, we will leak data numbers 1981824d8dffSQu Wenruo * 2) Fs subtree 1982824d8dffSQu Wenruo * If not traced, we will double count old data 1983f616f5cdSQu Wenruo * 1984f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 1985f616f5cdSQu Wenruo * the swapped tree blocks. 1986f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 1987f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 1988824d8dffSQu Wenruo */ 1989370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 1990370a11b8SQu Wenruo rc->block_group, parent, slot, 1991370a11b8SQu Wenruo path->nodes[level], path->slots[level], 1992370a11b8SQu Wenruo last_snapshot); 1993370a11b8SQu Wenruo if (ret < 0) 1994370a11b8SQu Wenruo break; 1995824d8dffSQu Wenruo /* 19965d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 19975d4f98a2SYan Zheng */ 19985d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 19995d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 20005d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 20015d4f98a2SYan Zheng 20025d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 20035d4f98a2SYan Zheng path->slots[level], old_bytenr); 20045d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 20055d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 20065d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 20075d4f98a2SYan Zheng 200882fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 200982fa113fSQu Wenruo blocksize, path->nodes[level]->start); 201082fa113fSQu Wenruo ref.skip_qgroup = true; 201182fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 201282fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 20135d4f98a2SYan Zheng BUG_ON(ret); 201482fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 201582fa113fSQu Wenruo blocksize, 0); 201682fa113fSQu Wenruo ref.skip_qgroup = true; 201782fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 201882fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 20195d4f98a2SYan Zheng BUG_ON(ret); 20205d4f98a2SYan Zheng 2021ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 2022ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 2023ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 2024ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2025ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 20265d4f98a2SYan Zheng BUG_ON(ret); 20275d4f98a2SYan Zheng 2028ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 2029ffd4bb2aSQu Wenruo blocksize, 0); 2030ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 2031ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2032ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 20335d4f98a2SYan Zheng BUG_ON(ret); 20345d4f98a2SYan Zheng 20355d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 20365d4f98a2SYan Zheng 20375d4f98a2SYan Zheng ret = level; 20385d4f98a2SYan Zheng break; 20395d4f98a2SYan Zheng } 20405d4f98a2SYan Zheng btrfs_tree_unlock(parent); 20415d4f98a2SYan Zheng free_extent_buffer(parent); 20425d4f98a2SYan Zheng return ret; 20435d4f98a2SYan Zheng } 20445d4f98a2SYan Zheng 20455d4f98a2SYan Zheng /* 20465d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 20475d4f98a2SYan Zheng */ 20485d4f98a2SYan Zheng static noinline_for_stack 20495d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 20505d4f98a2SYan Zheng int *level) 20515d4f98a2SYan Zheng { 20525d4f98a2SYan Zheng struct extent_buffer *eb; 20535d4f98a2SYan Zheng int i; 20545d4f98a2SYan Zheng u64 last_snapshot; 20555d4f98a2SYan Zheng u32 nritems; 20565d4f98a2SYan Zheng 20575d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 20585d4f98a2SYan Zheng 20595d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 20605d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 20615d4f98a2SYan Zheng path->nodes[i] = NULL; 20625d4f98a2SYan Zheng } 20635d4f98a2SYan Zheng 20645d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 20655d4f98a2SYan Zheng eb = path->nodes[i]; 20665d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 20675d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 20685d4f98a2SYan Zheng path->slots[i]++; 20695d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 20705d4f98a2SYan Zheng last_snapshot) 20715d4f98a2SYan Zheng continue; 20725d4f98a2SYan Zheng 20735d4f98a2SYan Zheng *level = i; 20745d4f98a2SYan Zheng return 0; 20755d4f98a2SYan Zheng } 20765d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 20775d4f98a2SYan Zheng path->nodes[i] = NULL; 20785d4f98a2SYan Zheng } 20795d4f98a2SYan Zheng return 1; 20805d4f98a2SYan Zheng } 20815d4f98a2SYan Zheng 20825d4f98a2SYan Zheng /* 20835d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 20845d4f98a2SYan Zheng */ 20855d4f98a2SYan Zheng static noinline_for_stack 20865d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 20875d4f98a2SYan Zheng int *level) 20885d4f98a2SYan Zheng { 20892ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 20905d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 20915d4f98a2SYan Zheng int i; 20925d4f98a2SYan Zheng u64 bytenr; 20935d4f98a2SYan Zheng u64 ptr_gen = 0; 20945d4f98a2SYan Zheng u64 last_snapshot; 20955d4f98a2SYan Zheng u32 nritems; 20965d4f98a2SYan Zheng 20975d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 20985d4f98a2SYan Zheng 20995d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 2100581c1760SQu Wenruo struct btrfs_key first_key; 2101581c1760SQu Wenruo 21025d4f98a2SYan Zheng eb = path->nodes[i]; 21035d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 21045d4f98a2SYan Zheng while (path->slots[i] < nritems) { 21055d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 21065d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 21075d4f98a2SYan Zheng break; 21085d4f98a2SYan Zheng path->slots[i]++; 21095d4f98a2SYan Zheng } 21105d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 21115d4f98a2SYan Zheng if (i == *level) 21125d4f98a2SYan Zheng break; 21135d4f98a2SYan Zheng *level = i + 1; 21145d4f98a2SYan Zheng return 0; 21155d4f98a2SYan Zheng } 21165d4f98a2SYan Zheng if (i == 1) { 21175d4f98a2SYan Zheng *level = i; 21185d4f98a2SYan Zheng return 0; 21195d4f98a2SYan Zheng } 21205d4f98a2SYan Zheng 21215d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 2122581c1760SQu Wenruo btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]); 2123581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1, 2124581c1760SQu Wenruo &first_key); 212564c043deSLiu Bo if (IS_ERR(eb)) { 212664c043deSLiu Bo return PTR_ERR(eb); 212764c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2128416bc658SJosef Bacik free_extent_buffer(eb); 2129416bc658SJosef Bacik return -EIO; 2130416bc658SJosef Bacik } 21315d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 21325d4f98a2SYan Zheng path->nodes[i - 1] = eb; 21335d4f98a2SYan Zheng path->slots[i - 1] = 0; 21345d4f98a2SYan Zheng } 21355d4f98a2SYan Zheng return 1; 21365d4f98a2SYan Zheng } 21375d4f98a2SYan Zheng 21385d4f98a2SYan Zheng /* 21395d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 21405d4f98a2SYan Zheng * [min_key, max_key) 21415d4f98a2SYan Zheng */ 21425d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 21435d4f98a2SYan Zheng struct btrfs_key *min_key, 21445d4f98a2SYan Zheng struct btrfs_key *max_key) 21455d4f98a2SYan Zheng { 21460b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 21475d4f98a2SYan Zheng struct inode *inode = NULL; 21485d4f98a2SYan Zheng u64 objectid; 21495d4f98a2SYan Zheng u64 start, end; 215033345d01SLi Zefan u64 ino; 21515d4f98a2SYan Zheng 21525d4f98a2SYan Zheng objectid = min_key->objectid; 21535d4f98a2SYan Zheng while (1) { 21545d4f98a2SYan Zheng cond_resched(); 21555d4f98a2SYan Zheng iput(inode); 21565d4f98a2SYan Zheng 21575d4f98a2SYan Zheng if (objectid > max_key->objectid) 21585d4f98a2SYan Zheng break; 21595d4f98a2SYan Zheng 21605d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 21615d4f98a2SYan Zheng if (!inode) 21625d4f98a2SYan Zheng break; 21634a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 21645d4f98a2SYan Zheng 216533345d01SLi Zefan if (ino > max_key->objectid) { 21665d4f98a2SYan Zheng iput(inode); 21675d4f98a2SYan Zheng break; 21685d4f98a2SYan Zheng } 21695d4f98a2SYan Zheng 217033345d01SLi Zefan objectid = ino + 1; 21715d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 21725d4f98a2SYan Zheng continue; 21735d4f98a2SYan Zheng 217433345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 21755d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 21765d4f98a2SYan Zheng continue; 21775d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 21785d4f98a2SYan Zheng start = 0; 21795d4f98a2SYan Zheng else { 21805d4f98a2SYan Zheng start = min_key->offset; 21810b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 21825d4f98a2SYan Zheng } 21835d4f98a2SYan Zheng } else { 21845d4f98a2SYan Zheng start = 0; 21855d4f98a2SYan Zheng } 21865d4f98a2SYan Zheng 218733345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 21885d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 21895d4f98a2SYan Zheng continue; 21905d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 21915d4f98a2SYan Zheng end = (u64)-1; 21925d4f98a2SYan Zheng } else { 21935d4f98a2SYan Zheng if (max_key->offset == 0) 21945d4f98a2SYan Zheng continue; 21955d4f98a2SYan Zheng end = max_key->offset; 21960b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 21975d4f98a2SYan Zheng end--; 21985d4f98a2SYan Zheng } 21995d4f98a2SYan Zheng } else { 22005d4f98a2SYan Zheng end = (u64)-1; 22015d4f98a2SYan Zheng } 22025d4f98a2SYan Zheng 22035d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 2204d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 2205dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 2206d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 22075d4f98a2SYan Zheng } 22085d4f98a2SYan Zheng return 0; 22095d4f98a2SYan Zheng } 22105d4f98a2SYan Zheng 22115d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 22125d4f98a2SYan Zheng struct btrfs_key *key) 22135d4f98a2SYan Zheng 22145d4f98a2SYan Zheng { 22155d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 22165d4f98a2SYan Zheng if (!path->nodes[level]) 22175d4f98a2SYan Zheng break; 22185d4f98a2SYan Zheng if (path->slots[level] + 1 < 22195d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 22205d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 22215d4f98a2SYan Zheng path->slots[level] + 1); 22225d4f98a2SYan Zheng return 0; 22235d4f98a2SYan Zheng } 22245d4f98a2SYan Zheng level++; 22255d4f98a2SYan Zheng } 22265d4f98a2SYan Zheng return 1; 22275d4f98a2SYan Zheng } 22285d4f98a2SYan Zheng 22295d4f98a2SYan Zheng /* 2230d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 2231d2311e69SQu Wenruo */ 2232d2311e69SQu Wenruo static void insert_dirty_subvol(struct btrfs_trans_handle *trans, 2233d2311e69SQu Wenruo struct reloc_control *rc, 2234d2311e69SQu Wenruo struct btrfs_root *root) 2235d2311e69SQu Wenruo { 2236d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2237d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 2238d2311e69SQu Wenruo 2239d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 2240d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 2241d2311e69SQu Wenruo ASSERT(reloc_root); 2242d2311e69SQu Wenruo 2243d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 2244d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 2245d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 2246d2311e69SQu Wenruo reloc_root_item->drop_level = 0; 2247d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 2248d2311e69SQu Wenruo btrfs_update_reloc_root(trans, root); 2249d2311e69SQu Wenruo 2250d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 225100246528SJosef Bacik btrfs_grab_root(root); 2252d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 2253d2311e69SQu Wenruo } 2254d2311e69SQu Wenruo } 2255d2311e69SQu Wenruo 2256d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 2257d2311e69SQu Wenruo { 2258d2311e69SQu Wenruo struct btrfs_root *root; 2259d2311e69SQu Wenruo struct btrfs_root *next; 2260d2311e69SQu Wenruo int ret = 0; 226130d40577SQu Wenruo int ret2; 2262d2311e69SQu Wenruo 2263d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 2264d2311e69SQu Wenruo reloc_dirty_list) { 226530d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 226630d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 2267d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2268d2311e69SQu Wenruo 2269d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 2270d2311e69SQu Wenruo root->reloc_root = NULL; 22716282675eSQu Wenruo /* 22726282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 22736282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 22746282675eSQu Wenruo */ 22756282675eSQu Wenruo smp_wmb(); 22761fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 2277f28de8d8SJosef Bacik if (reloc_root) { 2278f44deb74SJosef Bacik /* 2279f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 2280f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 2281f44deb74SJosef Bacik * drop the ref ourselves. 2282f44deb74SJosef Bacik */ 2283f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 2284f44deb74SJosef Bacik if (ret2 < 0) { 2285f44deb74SJosef Bacik btrfs_put_root(reloc_root); 2286f44deb74SJosef Bacik if (!ret) 2287f28de8d8SJosef Bacik ret = ret2; 2288f28de8d8SJosef Bacik } 2289f44deb74SJosef Bacik } 229000246528SJosef Bacik btrfs_put_root(root); 229130d40577SQu Wenruo } else { 229230d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 22930078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 2294f44deb74SJosef Bacik if (ret2 < 0) { 2295f44deb74SJosef Bacik btrfs_put_root(root); 2296f44deb74SJosef Bacik if (!ret) 229730d40577SQu Wenruo ret = ret2; 229830d40577SQu Wenruo } 2299d2311e69SQu Wenruo } 2300f44deb74SJosef Bacik } 2301d2311e69SQu Wenruo return ret; 2302d2311e69SQu Wenruo } 2303d2311e69SQu Wenruo 2304d2311e69SQu Wenruo /* 23055d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 23065d4f98a2SYan Zheng * fs tree. 23075d4f98a2SYan Zheng */ 23085d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 23095d4f98a2SYan Zheng struct btrfs_root *root) 23105d4f98a2SYan Zheng { 23110b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 23125d4f98a2SYan Zheng struct btrfs_key key; 23135d4f98a2SYan Zheng struct btrfs_key next_key; 23149e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 23155d4f98a2SYan Zheng struct btrfs_root *reloc_root; 23165d4f98a2SYan Zheng struct btrfs_root_item *root_item; 23175d4f98a2SYan Zheng struct btrfs_path *path; 23183fd0a558SYan, Zheng struct extent_buffer *leaf; 23195d4f98a2SYan Zheng int level; 23205d4f98a2SYan Zheng int max_level; 23215d4f98a2SYan Zheng int replaced = 0; 23225d4f98a2SYan Zheng int ret; 23235d4f98a2SYan Zheng int err = 0; 23243fd0a558SYan, Zheng u32 min_reserved; 23255d4f98a2SYan Zheng 23265d4f98a2SYan Zheng path = btrfs_alloc_path(); 23275d4f98a2SYan Zheng if (!path) 23285d4f98a2SYan Zheng return -ENOMEM; 2329e4058b54SDavid Sterba path->reada = READA_FORWARD; 23305d4f98a2SYan Zheng 23315d4f98a2SYan Zheng reloc_root = root->reloc_root; 23325d4f98a2SYan Zheng root_item = &reloc_root->root_item; 23335d4f98a2SYan Zheng 23345d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 23355d4f98a2SYan Zheng level = btrfs_root_level(root_item); 233667439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 23375d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 23385d4f98a2SYan Zheng path->slots[level] = 0; 23395d4f98a2SYan Zheng } else { 23405d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 23415d4f98a2SYan Zheng 23425d4f98a2SYan Zheng level = root_item->drop_level; 23435d4f98a2SYan Zheng BUG_ON(level == 0); 23445d4f98a2SYan Zheng path->lowest_level = level; 23455d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 234633c66f43SYan Zheng path->lowest_level = 0; 23475d4f98a2SYan Zheng if (ret < 0) { 23485d4f98a2SYan Zheng btrfs_free_path(path); 23495d4f98a2SYan Zheng return ret; 23505d4f98a2SYan Zheng } 23515d4f98a2SYan Zheng 23525d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 23535d4f98a2SYan Zheng path->slots[level]); 23545d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 23555d4f98a2SYan Zheng 23565d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 23575d4f98a2SYan Zheng } 23585d4f98a2SYan Zheng 23590b246afaSJeff Mahoney min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 23605d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 23615d4f98a2SYan Zheng 23625d4f98a2SYan Zheng while (1) { 236308e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 236408e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 23653fd0a558SYan, Zheng if (ret) { 23669e6a0c52SJosef Bacik err = ret; 23679e6a0c52SJosef Bacik goto out; 23683fd0a558SYan, Zheng } 23699e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 23709e6a0c52SJosef Bacik if (IS_ERR(trans)) { 23719e6a0c52SJosef Bacik err = PTR_ERR(trans); 23729e6a0c52SJosef Bacik trans = NULL; 23739e6a0c52SJosef Bacik goto out; 23749e6a0c52SJosef Bacik } 23752abc726aSJosef Bacik 23762abc726aSJosef Bacik /* 23772abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 23782abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 23792abc726aSJosef Bacik * 23802abc726aSJosef Bacik * But that's ok, we started the trans handle on our 23812abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 23822abc726aSJosef Bacik * dirty list. At commit time we'll still call 23832abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 23842abc726aSJosef Bacik * appropriately. 23852abc726aSJosef Bacik */ 23862abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 23879e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 23883fd0a558SYan, Zheng 23893fd0a558SYan, Zheng replaced = 0; 23905d4f98a2SYan Zheng max_level = level; 23915d4f98a2SYan Zheng 23925d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 23935d4f98a2SYan Zheng if (ret < 0) { 23945d4f98a2SYan Zheng err = ret; 23955d4f98a2SYan Zheng goto out; 23965d4f98a2SYan Zheng } 23975d4f98a2SYan Zheng if (ret > 0) 23985d4f98a2SYan Zheng break; 23995d4f98a2SYan Zheng 24005d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 24015d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 24025d4f98a2SYan Zheng ret = 0; 24035d4f98a2SYan Zheng } else { 24043d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 24053fd0a558SYan, Zheng &next_key, level, max_level); 24065d4f98a2SYan Zheng } 24075d4f98a2SYan Zheng if (ret < 0) { 24085d4f98a2SYan Zheng err = ret; 24095d4f98a2SYan Zheng goto out; 24105d4f98a2SYan Zheng } 24115d4f98a2SYan Zheng 24125d4f98a2SYan Zheng if (ret > 0) { 24135d4f98a2SYan Zheng level = ret; 24145d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 24155d4f98a2SYan Zheng path->slots[level]); 24165d4f98a2SYan Zheng replaced = 1; 24175d4f98a2SYan Zheng } 24185d4f98a2SYan Zheng 24195d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 24205d4f98a2SYan Zheng if (ret > 0) 24215d4f98a2SYan Zheng break; 24225d4f98a2SYan Zheng 24235d4f98a2SYan Zheng BUG_ON(level == 0); 24245d4f98a2SYan Zheng /* 24255d4f98a2SYan Zheng * save the merging progress in the drop_progress. 24265d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 24275d4f98a2SYan Zheng */ 24285d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 24295d4f98a2SYan Zheng path->slots[level]); 24305d4f98a2SYan Zheng root_item->drop_level = level; 24315d4f98a2SYan Zheng 24323a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 24339e6a0c52SJosef Bacik trans = NULL; 24345d4f98a2SYan Zheng 24352ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 24365d4f98a2SYan Zheng 24375d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 24385d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 24395d4f98a2SYan Zheng } 24405d4f98a2SYan Zheng 24415d4f98a2SYan Zheng /* 24425d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 24435d4f98a2SYan Zheng * relocated and the block is tree root. 24445d4f98a2SYan Zheng */ 24455d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 24465d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 24475d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 24485d4f98a2SYan Zheng free_extent_buffer(leaf); 24495d4f98a2SYan Zheng if (ret < 0) 24505d4f98a2SYan Zheng err = ret; 24515d4f98a2SYan Zheng out: 24525d4f98a2SYan Zheng btrfs_free_path(path); 24535d4f98a2SYan Zheng 2454d2311e69SQu Wenruo if (err == 0) 2455d2311e69SQu Wenruo insert_dirty_subvol(trans, rc, root); 24565d4f98a2SYan Zheng 24579e6a0c52SJosef Bacik if (trans) 24583a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 24595d4f98a2SYan Zheng 24602ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 24615d4f98a2SYan Zheng 24625d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 24635d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 24645d4f98a2SYan Zheng 24655d4f98a2SYan Zheng return err; 24665d4f98a2SYan Zheng } 24675d4f98a2SYan Zheng 24683fd0a558SYan, Zheng static noinline_for_stack 24693fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 24705d4f98a2SYan Zheng { 24713fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 24720b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 24733fd0a558SYan, Zheng struct btrfs_root *reloc_root; 24745d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 24753fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 24763fd0a558SYan, Zheng u64 num_bytes = 0; 24773fd0a558SYan, Zheng int ret; 24783fd0a558SYan, Zheng 24790b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 24800b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 24813fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 24820b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 24837585717fSChris Mason 24843fd0a558SYan, Zheng again: 24853fd0a558SYan, Zheng if (!err) { 24863fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 248708e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 248808e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 24893fd0a558SYan, Zheng if (ret) 24903fd0a558SYan, Zheng err = ret; 24913fd0a558SYan, Zheng } 24923fd0a558SYan, Zheng 24937a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 24943612b495STsutomu Itoh if (IS_ERR(trans)) { 24953612b495STsutomu Itoh if (!err) 24962ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 249763f018beSNikolay Borisov num_bytes, NULL); 24983612b495STsutomu Itoh return PTR_ERR(trans); 24993612b495STsutomu Itoh } 25003fd0a558SYan, Zheng 25013fd0a558SYan, Zheng if (!err) { 25023fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 25033a45bb20SJeff Mahoney btrfs_end_transaction(trans); 25042ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 250563f018beSNikolay Borisov num_bytes, NULL); 25063fd0a558SYan, Zheng goto again; 25073fd0a558SYan, Zheng } 25083fd0a558SYan, Zheng } 25093fd0a558SYan, Zheng 25103fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 25113fd0a558SYan, Zheng 25123fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 25133fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 25143fd0a558SYan, Zheng struct btrfs_root, root_list); 25153fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 25163fd0a558SYan, Zheng 25170b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 25183fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 25193fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 25203fd0a558SYan, Zheng 25213fd0a558SYan, Zheng /* 25223fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 25233fd0a558SYan, Zheng * knows it should resumes merging 25243fd0a558SYan, Zheng */ 25253fd0a558SYan, Zheng if (!err) 25263fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 25273fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 25283fd0a558SYan, Zheng 25293fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 253000246528SJosef Bacik btrfs_put_root(root); 25313fd0a558SYan, Zheng } 25323fd0a558SYan, Zheng 25333fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 25343fd0a558SYan, Zheng 25353fd0a558SYan, Zheng if (!err) 25363a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 25373fd0a558SYan, Zheng else 25383a45bb20SJeff Mahoney btrfs_end_transaction(trans); 25393fd0a558SYan, Zheng return err; 25403fd0a558SYan, Zheng } 25413fd0a558SYan, Zheng 25423fd0a558SYan, Zheng static noinline_for_stack 2543aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 2544aca1bba6SLiu Bo { 2545aca1bba6SLiu Bo struct btrfs_root *reloc_root; 2546aca1bba6SLiu Bo 2547aca1bba6SLiu Bo while (!list_empty(list)) { 2548aca1bba6SLiu Bo reloc_root = list_entry(list->next, struct btrfs_root, 2549aca1bba6SLiu Bo root_list); 2550bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 2551aca1bba6SLiu Bo } 2552aca1bba6SLiu Bo } 2553aca1bba6SLiu Bo 2554aca1bba6SLiu Bo static noinline_for_stack 255594404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 25563fd0a558SYan, Zheng { 25570b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 25585d4f98a2SYan Zheng struct btrfs_root *root; 25595d4f98a2SYan Zheng struct btrfs_root *reloc_root; 25603fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 25613fd0a558SYan, Zheng int found = 0; 2562aca1bba6SLiu Bo int ret = 0; 25633fd0a558SYan, Zheng again: 25643fd0a558SYan, Zheng root = rc->extent_root; 25657585717fSChris Mason 25667585717fSChris Mason /* 25677585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 25687585717fSChris Mason * we have to make sure nobody is in the middle of 25697585717fSChris Mason * adding their roots to the list while we are 25707585717fSChris Mason * doing this splice 25717585717fSChris Mason */ 25720b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 25733fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 25740b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 25755d4f98a2SYan Zheng 25763fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 25773fd0a558SYan, Zheng found = 1; 25783fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 25793fd0a558SYan, Zheng struct btrfs_root, root_list); 25805d4f98a2SYan Zheng 25815d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 25820b246afaSJeff Mahoney root = read_fs_root(fs_info, 25835d4f98a2SYan Zheng reloc_root->root_key.offset); 25845d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 25855d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 25865d4f98a2SYan Zheng 25873fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 258800246528SJosef Bacik btrfs_put_root(root); 2589b37b39cdSJosef Bacik if (ret) { 259025e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 259125e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 259225e293c2SWang Shilong &reloc_roots); 2593aca1bba6SLiu Bo goto out; 2594b37b39cdSJosef Bacik } 25953fd0a558SYan, Zheng } else { 25963fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 259730d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 259830d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 259930d40577SQu Wenruo &rc->dirty_subvol_roots); 26003fd0a558SYan, Zheng } 26015d4f98a2SYan Zheng } 26025d4f98a2SYan Zheng 26033fd0a558SYan, Zheng if (found) { 26043fd0a558SYan, Zheng found = 0; 26053fd0a558SYan, Zheng goto again; 26065d4f98a2SYan Zheng } 2607aca1bba6SLiu Bo out: 2608aca1bba6SLiu Bo if (ret) { 26090b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 2610aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 2611aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 2612467bb1d2SWang Shilong 2613467bb1d2SWang Shilong /* new reloc root may be added */ 26140b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 2615467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 26160b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 2617467bb1d2SWang Shilong if (!list_empty(&reloc_roots)) 2618467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 2619aca1bba6SLiu Bo } 2620aca1bba6SLiu Bo 26217b7b7431SJosef Bacik /* 26227b7b7431SJosef Bacik * We used to have 26237b7b7431SJosef Bacik * 26247b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 26257b7b7431SJosef Bacik * 26267b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 26277b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 26287b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 26297b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 26307b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 26317b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 26327b7b7431SJosef Bacik * will be cleaned up on unmount. 26337b7b7431SJosef Bacik * 26347b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 26357b7b7431SJosef Bacik */ 26365d4f98a2SYan Zheng } 26375d4f98a2SYan Zheng 26385d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 26395d4f98a2SYan Zheng { 26405d4f98a2SYan Zheng struct tree_block *block; 26415d4f98a2SYan Zheng struct rb_node *rb_node; 26425d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 26435d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 26445d4f98a2SYan Zheng rb_erase(rb_node, blocks); 26455d4f98a2SYan Zheng kfree(block); 26465d4f98a2SYan Zheng } 26475d4f98a2SYan Zheng } 26485d4f98a2SYan Zheng 26495d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 26505d4f98a2SYan Zheng struct btrfs_root *reloc_root) 26515d4f98a2SYan Zheng { 26520b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 26535d4f98a2SYan Zheng struct btrfs_root *root; 2654442b1ac5SJosef Bacik int ret; 26555d4f98a2SYan Zheng 26565d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 26575d4f98a2SYan Zheng return 0; 26585d4f98a2SYan Zheng 26590b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 26605d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 26615d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 2662442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 266300246528SJosef Bacik btrfs_put_root(root); 26645d4f98a2SYan Zheng 2665442b1ac5SJosef Bacik return ret; 26665d4f98a2SYan Zheng } 26675d4f98a2SYan Zheng 26683fd0a558SYan, Zheng static noinline_for_stack 26693fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 26703fd0a558SYan, Zheng struct reloc_control *rc, 2671a26195a5SQu Wenruo struct btrfs_backref_node *node, 2672a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 26735d4f98a2SYan Zheng { 2674a26195a5SQu Wenruo struct btrfs_backref_node *next; 26755d4f98a2SYan Zheng struct btrfs_root *root; 26763fd0a558SYan, Zheng int index = 0; 26773fd0a558SYan, Zheng 26785d4f98a2SYan Zheng next = node; 26795d4f98a2SYan Zheng while (1) { 26805d4f98a2SYan Zheng cond_resched(); 26815d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 26825d4f98a2SYan Zheng root = next->root; 26833fd0a558SYan, Zheng BUG_ON(!root); 268427cdeb70SMiao Xie BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state)); 26855d4f98a2SYan Zheng 26865d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 26875d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 26885d4f98a2SYan Zheng break; 26895d4f98a2SYan Zheng } 26905d4f98a2SYan Zheng 26915d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 26923fd0a558SYan, Zheng root = root->reloc_root; 26933fd0a558SYan, Zheng 26943fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 26953fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 26963fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 26973fd0a558SYan, Zheng next->new_bytenr = root->node->start; 269800246528SJosef Bacik btrfs_put_root(next->root); 269900246528SJosef Bacik next->root = btrfs_grab_root(root); 27000b530bc5SJosef Bacik ASSERT(next->root); 27013fd0a558SYan, Zheng list_add_tail(&next->list, 27023fd0a558SYan, Zheng &rc->backref_cache.changed); 27039569cc20SQu Wenruo mark_block_processed(rc, next); 27045d4f98a2SYan Zheng break; 27055d4f98a2SYan Zheng } 27065d4f98a2SYan Zheng 27073fd0a558SYan, Zheng WARN_ON(1); 27085d4f98a2SYan Zheng root = NULL; 27095d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 27105d4f98a2SYan Zheng if (!next || next->level <= node->level) 27115d4f98a2SYan Zheng break; 27125d4f98a2SYan Zheng } 27133fd0a558SYan, Zheng if (!root) 27143fd0a558SYan, Zheng return NULL; 27155d4f98a2SYan Zheng 27163fd0a558SYan, Zheng next = node; 27173fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 27183fd0a558SYan, Zheng while (1) { 27193fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 27203fd0a558SYan, Zheng if (--index < 0) 27213fd0a558SYan, Zheng break; 27223fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 27233fd0a558SYan, Zheng } 27245d4f98a2SYan Zheng return root; 27255d4f98a2SYan Zheng } 27265d4f98a2SYan Zheng 27273fd0a558SYan, Zheng /* 27283fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 27293fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 27303fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 27313fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 27323fd0a558SYan, Zheng */ 27335d4f98a2SYan Zheng static noinline_for_stack 2734a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 27355d4f98a2SYan Zheng { 2736a26195a5SQu Wenruo struct btrfs_backref_node *next; 27373fd0a558SYan, Zheng struct btrfs_root *root; 27383fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2739a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 27403fd0a558SYan, Zheng int index = 0; 27413fd0a558SYan, Zheng 27423fd0a558SYan, Zheng next = node; 27433fd0a558SYan, Zheng while (1) { 27443fd0a558SYan, Zheng cond_resched(); 27453fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 27463fd0a558SYan, Zheng root = next->root; 27473fd0a558SYan, Zheng BUG_ON(!root); 27483fd0a558SYan, Zheng 274925985edcSLucas De Marchi /* no other choice for non-references counted tree */ 275027cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 27513fd0a558SYan, Zheng return root; 27523fd0a558SYan, Zheng 27533fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 27543fd0a558SYan, Zheng fs_root = root; 27553fd0a558SYan, Zheng 27563fd0a558SYan, Zheng if (next != node) 27573fd0a558SYan, Zheng return NULL; 27583fd0a558SYan, Zheng 27593fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 27603fd0a558SYan, Zheng if (!next || next->level <= node->level) 27613fd0a558SYan, Zheng break; 27623fd0a558SYan, Zheng } 27633fd0a558SYan, Zheng 27643fd0a558SYan, Zheng if (!fs_root) 27653fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 27663fd0a558SYan, Zheng return fs_root; 27675d4f98a2SYan Zheng } 27685d4f98a2SYan Zheng 27695d4f98a2SYan Zheng static noinline_for_stack 27703fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2771a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 27725d4f98a2SYan Zheng { 27730b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2774a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2775a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2776a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 27773fd0a558SYan, Zheng u64 num_bytes = 0; 27783fd0a558SYan, Zheng int index = 0; 27795d4f98a2SYan Zheng 27803fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 27813fd0a558SYan, Zheng 27823fd0a558SYan, Zheng while (next) { 27833fd0a558SYan, Zheng cond_resched(); 27845d4f98a2SYan Zheng while (1) { 27853fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 27865d4f98a2SYan Zheng break; 27875d4f98a2SYan Zheng 27880b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 27893fd0a558SYan, Zheng 27903fd0a558SYan, Zheng if (list_empty(&next->upper)) 27913fd0a558SYan, Zheng break; 27923fd0a558SYan, Zheng 27933fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2794a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 27953fd0a558SYan, Zheng edges[index++] = edge; 27963fd0a558SYan, Zheng next = edge->node[UPPER]; 27975d4f98a2SYan Zheng } 27983fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 27993fd0a558SYan, Zheng } 28003fd0a558SYan, Zheng return num_bytes; 28013fd0a558SYan, Zheng } 28023fd0a558SYan, Zheng 28033fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 28043fd0a558SYan, Zheng struct reloc_control *rc, 2805a26195a5SQu Wenruo struct btrfs_backref_node *node) 28063fd0a558SYan, Zheng { 28073fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2808da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 28093fd0a558SYan, Zheng u64 num_bytes; 28103fd0a558SYan, Zheng int ret; 28110647bf56SWang Shilong u64 tmp; 28123fd0a558SYan, Zheng 28133fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 28143fd0a558SYan, Zheng 28153fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 28160647bf56SWang Shilong rc->reserved_bytes += num_bytes; 28178ca17f0fSJosef Bacik 28188ca17f0fSJosef Bacik /* 28198ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 28208ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 28218ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 28228ca17f0fSJosef Bacik */ 28230647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 28248ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 28253fd0a558SYan, Zheng if (ret) { 2826da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 28270647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 28280647bf56SWang Shilong tmp <<= 1; 28290647bf56SWang Shilong /* 28300647bf56SWang Shilong * only one thread can access block_rsv at this point, 28310647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 28320647bf56SWang Shilong * we expand more reservation size here to allow enough 283352042d8eSAndrea Gelmini * space for relocation and we will return earlier in 28340647bf56SWang Shilong * enospc case. 28350647bf56SWang Shilong */ 2836da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 28370647bf56SWang Shilong RELOCATION_RESERVED_NODES; 28388ca17f0fSJosef Bacik return -EAGAIN; 28393fd0a558SYan, Zheng } 28403fd0a558SYan, Zheng 28413fd0a558SYan, Zheng return 0; 28423fd0a558SYan, Zheng } 28433fd0a558SYan, Zheng 28445d4f98a2SYan Zheng /* 28455d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 28465d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 28475d4f98a2SYan Zheng * 28485d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 28495d4f98a2SYan Zheng * in that case this function just updates pointers. 28505d4f98a2SYan Zheng */ 28515d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 28523fd0a558SYan, Zheng struct reloc_control *rc, 2853a26195a5SQu Wenruo struct btrfs_backref_node *node, 28545d4f98a2SYan Zheng struct btrfs_key *key, 28555d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 28565d4f98a2SYan Zheng { 28572ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2858a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2859a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2860a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 28615d4f98a2SYan Zheng struct btrfs_root *root; 28625d4f98a2SYan Zheng struct extent_buffer *eb; 28635d4f98a2SYan Zheng u32 blocksize; 28645d4f98a2SYan Zheng u64 bytenr; 28655d4f98a2SYan Zheng u64 generation; 28665d4f98a2SYan Zheng int slot; 28675d4f98a2SYan Zheng int ret; 28685d4f98a2SYan Zheng int err = 0; 28695d4f98a2SYan Zheng 28705d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 28715d4f98a2SYan Zheng 28725d4f98a2SYan Zheng path->lowest_level = node->level + 1; 28733fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 28745d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 2875581c1760SQu Wenruo struct btrfs_key first_key; 287682fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2877581c1760SQu Wenruo 28785d4f98a2SYan Zheng cond_resched(); 28795d4f98a2SYan Zheng 28805d4f98a2SYan Zheng upper = edge->node[UPPER]; 2881dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 28823fd0a558SYan, Zheng BUG_ON(!root); 28835d4f98a2SYan Zheng 28843fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 28853fd0a558SYan, Zheng if (!lowest) { 28863fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 28873fd0a558SYan, Zheng upper->level, &slot); 2888cbca7d59SFilipe Manana if (ret < 0) { 2889cbca7d59SFilipe Manana err = ret; 2890cbca7d59SFilipe Manana goto next; 2891cbca7d59SFilipe Manana } 28923fd0a558SYan, Zheng BUG_ON(ret); 28933fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 28943fd0a558SYan, Zheng if (node->eb->start == bytenr) 28953fd0a558SYan, Zheng goto next; 28963fd0a558SYan, Zheng } 28975d4f98a2SYan Zheng drop_node_buffer(upper); 28983fd0a558SYan, Zheng } 28995d4f98a2SYan Zheng 29005d4f98a2SYan Zheng if (!upper->eb) { 29015d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 29023561b9dbSLiu Bo if (ret) { 29033561b9dbSLiu Bo if (ret < 0) 29045d4f98a2SYan Zheng err = ret; 29053561b9dbSLiu Bo else 29063561b9dbSLiu Bo err = -ENOENT; 29073561b9dbSLiu Bo 29083561b9dbSLiu Bo btrfs_release_path(path); 29095d4f98a2SYan Zheng break; 29105d4f98a2SYan Zheng } 29115d4f98a2SYan Zheng 29123fd0a558SYan, Zheng if (!upper->eb) { 29133fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 29143fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 29153fd0a558SYan, Zheng } else { 29163fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 29173fd0a558SYan, Zheng } 29183fd0a558SYan, Zheng 29193fd0a558SYan, Zheng upper->locked = 1; 29203fd0a558SYan, Zheng path->locks[upper->level] = 0; 29213fd0a558SYan, Zheng 29225d4f98a2SYan Zheng slot = path->slots[upper->level]; 2923b3b4aa74SDavid Sterba btrfs_release_path(path); 29245d4f98a2SYan Zheng } else { 29255d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 29265d4f98a2SYan Zheng &slot); 2927cbca7d59SFilipe Manana if (ret < 0) { 2928cbca7d59SFilipe Manana err = ret; 2929cbca7d59SFilipe Manana goto next; 2930cbca7d59SFilipe Manana } 29315d4f98a2SYan Zheng BUG_ON(ret); 29325d4f98a2SYan Zheng } 29335d4f98a2SYan Zheng 29345d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 29353fd0a558SYan, Zheng if (lowest) { 29364547f4d8SLiu Bo if (bytenr != node->bytenr) { 29374547f4d8SLiu Bo btrfs_err(root->fs_info, 29384547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 29394547f4d8SLiu Bo bytenr, node->bytenr, slot, 29404547f4d8SLiu Bo upper->eb->start); 29414547f4d8SLiu Bo err = -EIO; 29424547f4d8SLiu Bo goto next; 29434547f4d8SLiu Bo } 29445d4f98a2SYan Zheng } else { 29453fd0a558SYan, Zheng if (node->eb->start == bytenr) 29463fd0a558SYan, Zheng goto next; 29475d4f98a2SYan Zheng } 29485d4f98a2SYan Zheng 2949da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 29505d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 2951581c1760SQu Wenruo btrfs_node_key_to_cpu(upper->eb, &first_key, slot); 2952581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, generation, 2953581c1760SQu Wenruo upper->level - 1, &first_key); 295464c043deSLiu Bo if (IS_ERR(eb)) { 295564c043deSLiu Bo err = PTR_ERR(eb); 295664c043deSLiu Bo goto next; 295764c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2958416bc658SJosef Bacik free_extent_buffer(eb); 295997d9a8a4STsutomu Itoh err = -EIO; 296097d9a8a4STsutomu Itoh goto next; 296197d9a8a4STsutomu Itoh } 29625d4f98a2SYan Zheng btrfs_tree_lock(eb); 29638bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 29645d4f98a2SYan Zheng 29655d4f98a2SYan Zheng if (!node->eb) { 29665d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 29675d4f98a2SYan Zheng slot, &eb); 29683fd0a558SYan, Zheng btrfs_tree_unlock(eb); 29693fd0a558SYan, Zheng free_extent_buffer(eb); 29705d4f98a2SYan Zheng if (ret < 0) { 29715d4f98a2SYan Zheng err = ret; 29723fd0a558SYan, Zheng goto next; 29735d4f98a2SYan Zheng } 29743fd0a558SYan, Zheng BUG_ON(node->eb != eb); 29755d4f98a2SYan Zheng } else { 29765d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 29775d4f98a2SYan Zheng node->eb->start); 29785d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 29795d4f98a2SYan Zheng trans->transid); 29805d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 29815d4f98a2SYan Zheng 298282fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 29835d4f98a2SYan Zheng node->eb->start, blocksize, 298482fa113fSQu Wenruo upper->eb->start); 298582fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 298682fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 298782fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 298882fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 29895d4f98a2SYan Zheng BUG_ON(ret); 29905d4f98a2SYan Zheng 29915d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 29925d4f98a2SYan Zheng BUG_ON(ret); 29935d4f98a2SYan Zheng } 29943fd0a558SYan, Zheng next: 29953fd0a558SYan, Zheng if (!upper->pending) 29963fd0a558SYan, Zheng drop_node_buffer(upper); 29973fd0a558SYan, Zheng else 29983fd0a558SYan, Zheng unlock_node_buffer(upper); 29993fd0a558SYan, Zheng if (err) 30003fd0a558SYan, Zheng break; 30015d4f98a2SYan Zheng } 30023fd0a558SYan, Zheng 30033fd0a558SYan, Zheng if (!err && node->pending) { 30043fd0a558SYan, Zheng drop_node_buffer(node); 30053fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 30063fd0a558SYan, Zheng node->pending = 0; 30075d4f98a2SYan Zheng } 30083fd0a558SYan, Zheng 30095d4f98a2SYan Zheng path->lowest_level = 0; 30103fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 30115d4f98a2SYan Zheng return err; 30125d4f98a2SYan Zheng } 30135d4f98a2SYan Zheng 30145d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 30153fd0a558SYan, Zheng struct reloc_control *rc, 3016a26195a5SQu Wenruo struct btrfs_backref_node *node, 30175d4f98a2SYan Zheng struct btrfs_path *path) 30185d4f98a2SYan Zheng { 30195d4f98a2SYan Zheng struct btrfs_key key; 30205d4f98a2SYan Zheng 30215d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 30223fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 30235d4f98a2SYan Zheng } 30245d4f98a2SYan Zheng 30255d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 30263fd0a558SYan, Zheng struct reloc_control *rc, 30273fd0a558SYan, Zheng struct btrfs_path *path, int err) 30285d4f98a2SYan Zheng { 30293fd0a558SYan, Zheng LIST_HEAD(list); 3030a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 3031a26195a5SQu Wenruo struct btrfs_backref_node *node; 30325d4f98a2SYan Zheng int level; 30335d4f98a2SYan Zheng int ret; 30345d4f98a2SYan Zheng 30355d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 30365d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 30375d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 3038a26195a5SQu Wenruo struct btrfs_backref_node, list); 30393fd0a558SYan, Zheng list_move_tail(&node->list, &list); 30403fd0a558SYan, Zheng BUG_ON(!node->pending); 30415d4f98a2SYan Zheng 30423fd0a558SYan, Zheng if (!err) { 30433fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 30445d4f98a2SYan Zheng if (ret < 0) 30455d4f98a2SYan Zheng err = ret; 30465d4f98a2SYan Zheng } 30475d4f98a2SYan Zheng } 30483fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 30493fd0a558SYan, Zheng } 30505d4f98a2SYan Zheng return err; 30515d4f98a2SYan Zheng } 30525d4f98a2SYan Zheng 30535d4f98a2SYan Zheng /* 30545d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 30555d4f98a2SYan Zheng * as processed. 30565d4f98a2SYan Zheng */ 30575d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 3058a26195a5SQu Wenruo struct btrfs_backref_node *node) 30595d4f98a2SYan Zheng { 3060a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 3061a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 3062a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 30635d4f98a2SYan Zheng int index = 0; 30645d4f98a2SYan Zheng 30655d4f98a2SYan Zheng while (next) { 30665d4f98a2SYan Zheng cond_resched(); 30675d4f98a2SYan Zheng while (1) { 30685d4f98a2SYan Zheng if (next->processed) 30695d4f98a2SYan Zheng break; 30705d4f98a2SYan Zheng 30719569cc20SQu Wenruo mark_block_processed(rc, next); 30725d4f98a2SYan Zheng 30735d4f98a2SYan Zheng if (list_empty(&next->upper)) 30745d4f98a2SYan Zheng break; 30755d4f98a2SYan Zheng 30765d4f98a2SYan Zheng edge = list_entry(next->upper.next, 3077a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 30785d4f98a2SYan Zheng edges[index++] = edge; 30795d4f98a2SYan Zheng next = edge->node[UPPER]; 30805d4f98a2SYan Zheng } 30815d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 30825d4f98a2SYan Zheng } 30835d4f98a2SYan Zheng } 30845d4f98a2SYan Zheng 30857476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 30865d4f98a2SYan Zheng { 3087da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 30887476dfdaSDavid Sterba 30895d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 30909655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 30915d4f98a2SYan Zheng return 1; 30925d4f98a2SYan Zheng return 0; 30935d4f98a2SYan Zheng } 30945d4f98a2SYan Zheng 30952ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 30965d4f98a2SYan Zheng struct tree_block *block) 30975d4f98a2SYan Zheng { 30985d4f98a2SYan Zheng struct extent_buffer *eb; 30995d4f98a2SYan Zheng 3100581c1760SQu Wenruo eb = read_tree_block(fs_info, block->bytenr, block->key.offset, 3101581c1760SQu Wenruo block->level, NULL); 310264c043deSLiu Bo if (IS_ERR(eb)) { 310364c043deSLiu Bo return PTR_ERR(eb); 310464c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 3105416bc658SJosef Bacik free_extent_buffer(eb); 3106416bc658SJosef Bacik return -EIO; 3107416bc658SJosef Bacik } 31085d4f98a2SYan Zheng if (block->level == 0) 31095d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 31105d4f98a2SYan Zheng else 31115d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 31125d4f98a2SYan Zheng free_extent_buffer(eb); 31135d4f98a2SYan Zheng block->key_ready = 1; 31145d4f98a2SYan Zheng return 0; 31155d4f98a2SYan Zheng } 31165d4f98a2SYan Zheng 31175d4f98a2SYan Zheng /* 31185d4f98a2SYan Zheng * helper function to relocate a tree block 31195d4f98a2SYan Zheng */ 31205d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 31215d4f98a2SYan Zheng struct reloc_control *rc, 3122a26195a5SQu Wenruo struct btrfs_backref_node *node, 31235d4f98a2SYan Zheng struct btrfs_key *key, 31245d4f98a2SYan Zheng struct btrfs_path *path) 31255d4f98a2SYan Zheng { 31265d4f98a2SYan Zheng struct btrfs_root *root; 31273fd0a558SYan, Zheng int ret = 0; 31285d4f98a2SYan Zheng 31293fd0a558SYan, Zheng if (!node) 31305d4f98a2SYan Zheng return 0; 31313fd0a558SYan, Zheng 31325f6b2e5cSJosef Bacik /* 31335f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 31345f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 31355f6b2e5cSJosef Bacik */ 31365f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 31375f6b2e5cSJosef Bacik if (ret) 31385f6b2e5cSJosef Bacik goto out; 31395f6b2e5cSJosef Bacik 31403fd0a558SYan, Zheng BUG_ON(node->processed); 3141147d256eSZhaolei root = select_one_root(node); 31423fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 31433fd0a558SYan, Zheng update_processed_blocks(rc, node); 31443fd0a558SYan, Zheng goto out; 31455d4f98a2SYan Zheng } 31465d4f98a2SYan Zheng 31473fd0a558SYan, Zheng if (root) { 314827cdeb70SMiao Xie if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) { 31493fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 31503fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 31513fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 31523fd0a558SYan, Zheng root = root->reloc_root; 31533fd0a558SYan, Zheng node->new_bytenr = root->node->start; 315400246528SJosef Bacik btrfs_put_root(node->root); 315500246528SJosef Bacik node->root = btrfs_grab_root(root); 31560b530bc5SJosef Bacik ASSERT(node->root); 31573fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 31583fd0a558SYan, Zheng } else { 31595d4f98a2SYan Zheng path->lowest_level = node->level; 31605d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 3161b3b4aa74SDavid Sterba btrfs_release_path(path); 31623fd0a558SYan, Zheng if (ret > 0) 31635d4f98a2SYan Zheng ret = 0; 31643fd0a558SYan, Zheng } 31653fd0a558SYan, Zheng if (!ret) 31663fd0a558SYan, Zheng update_processed_blocks(rc, node); 31673fd0a558SYan, Zheng } else { 31683fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 31693fd0a558SYan, Zheng } 31705d4f98a2SYan Zheng out: 31710647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 31723fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 31735d4f98a2SYan Zheng return ret; 31745d4f98a2SYan Zheng } 31755d4f98a2SYan Zheng 31765d4f98a2SYan Zheng /* 31775d4f98a2SYan Zheng * relocate a list of blocks 31785d4f98a2SYan Zheng */ 31795d4f98a2SYan Zheng static noinline_for_stack 31805d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 31815d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 31825d4f98a2SYan Zheng { 31832ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 3184a26195a5SQu Wenruo struct btrfs_backref_node *node; 31855d4f98a2SYan Zheng struct btrfs_path *path; 31865d4f98a2SYan Zheng struct tree_block *block; 318798ff7b94SQu Wenruo struct tree_block *next; 31885d4f98a2SYan Zheng int ret; 31895d4f98a2SYan Zheng int err = 0; 31905d4f98a2SYan Zheng 31915d4f98a2SYan Zheng path = btrfs_alloc_path(); 3192e1a12670SLiu Bo if (!path) { 3193e1a12670SLiu Bo err = -ENOMEM; 319434c2b290SDavid Sterba goto out_free_blocks; 3195e1a12670SLiu Bo } 31965d4f98a2SYan Zheng 319798ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 319898ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 31995d4f98a2SYan Zheng if (!block->key_ready) 32002ff7e61eSJeff Mahoney readahead_tree_block(fs_info, block->bytenr); 32015d4f98a2SYan Zheng } 32025d4f98a2SYan Zheng 320398ff7b94SQu Wenruo /* Get first keys */ 320498ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 320534c2b290SDavid Sterba if (!block->key_ready) { 32062ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 320734c2b290SDavid Sterba if (err) 320834c2b290SDavid Sterba goto out_free_path; 320934c2b290SDavid Sterba } 32105d4f98a2SYan Zheng } 32115d4f98a2SYan Zheng 321298ff7b94SQu Wenruo /* Do tree relocation */ 321398ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 32143fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 32155d4f98a2SYan Zheng block->level, block->bytenr); 32165d4f98a2SYan Zheng if (IS_ERR(node)) { 32175d4f98a2SYan Zheng err = PTR_ERR(node); 32185d4f98a2SYan Zheng goto out; 32195d4f98a2SYan Zheng } 32205d4f98a2SYan Zheng 32215d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 32225d4f98a2SYan Zheng path); 32235d4f98a2SYan Zheng if (ret < 0) { 32245d4f98a2SYan Zheng err = ret; 322550dbbb71SJosef Bacik break; 32265d4f98a2SYan Zheng } 32275d4f98a2SYan Zheng } 32285d4f98a2SYan Zheng out: 32293fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 32305d4f98a2SYan Zheng 323134c2b290SDavid Sterba out_free_path: 32325d4f98a2SYan Zheng btrfs_free_path(path); 323334c2b290SDavid Sterba out_free_blocks: 3234e1a12670SLiu Bo free_block_list(blocks); 32355d4f98a2SYan Zheng return err; 32365d4f98a2SYan Zheng } 32375d4f98a2SYan Zheng 32385d4f98a2SYan Zheng static noinline_for_stack 3239efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 3240efa56464SYan, Zheng struct file_extent_cluster *cluster) 3241efa56464SYan, Zheng { 3242efa56464SYan, Zheng u64 alloc_hint = 0; 3243efa56464SYan, Zheng u64 start; 3244efa56464SYan, Zheng u64 end; 3245efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 3246efa56464SYan, Zheng u64 num_bytes; 3247efa56464SYan, Zheng int nr = 0; 3248efa56464SYan, Zheng int ret = 0; 3249dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 3250dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 325118513091SWang Xiaoguang u64 cur_offset; 3252364ecf36SQu Wenruo struct extent_changeset *data_reserved = NULL; 3253efa56464SYan, Zheng 3254efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 32555955102cSAl Viro inode_lock(inode); 3256efa56464SYan, Zheng 3257364ecf36SQu Wenruo ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start, 3258dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 3259efa56464SYan, Zheng if (ret) 3260efa56464SYan, Zheng goto out; 3261efa56464SYan, Zheng 326218513091SWang Xiaoguang cur_offset = prealloc_start; 3263efa56464SYan, Zheng while (nr < cluster->nr) { 3264efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 3265efa56464SYan, Zheng if (nr + 1 < cluster->nr) 3266efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 3267efa56464SYan, Zheng else 3268efa56464SYan, Zheng end = cluster->end - offset; 3269efa56464SYan, Zheng 3270d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 3271efa56464SYan, Zheng num_bytes = end + 1 - start; 327218513091SWang Xiaoguang if (cur_offset < start) 3273bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3274bc42bda2SQu Wenruo cur_offset, start - cur_offset); 3275efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 3276efa56464SYan, Zheng num_bytes, num_bytes, 3277efa56464SYan, Zheng end + 1, &alloc_hint); 327818513091SWang Xiaoguang cur_offset = end + 1; 3279d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 3280efa56464SYan, Zheng if (ret) 3281efa56464SYan, Zheng break; 3282efa56464SYan, Zheng nr++; 3283efa56464SYan, Zheng } 328418513091SWang Xiaoguang if (cur_offset < prealloc_end) 3285bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3286bc42bda2SQu Wenruo cur_offset, prealloc_end + 1 - cur_offset); 3287efa56464SYan, Zheng out: 32885955102cSAl Viro inode_unlock(inode); 3289364ecf36SQu Wenruo extent_changeset_free(data_reserved); 3290efa56464SYan, Zheng return ret; 3291efa56464SYan, Zheng } 3292efa56464SYan, Zheng 3293efa56464SYan, Zheng static noinline_for_stack 32940257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 32950257bb82SYan, Zheng u64 block_start) 32965d4f98a2SYan Zheng { 32975d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 32985d4f98a2SYan Zheng struct extent_map *em; 32990257bb82SYan, Zheng int ret = 0; 33005d4f98a2SYan Zheng 3301172ddd60SDavid Sterba em = alloc_extent_map(); 33020257bb82SYan, Zheng if (!em) 33030257bb82SYan, Zheng return -ENOMEM; 33040257bb82SYan, Zheng 33055d4f98a2SYan Zheng em->start = start; 33060257bb82SYan, Zheng em->len = end + 1 - start; 33070257bb82SYan, Zheng em->block_len = em->len; 33080257bb82SYan, Zheng em->block_start = block_start; 33095d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 33105d4f98a2SYan Zheng 3311d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 33125d4f98a2SYan Zheng while (1) { 3313890871beSChris Mason write_lock(&em_tree->lock); 331409a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 3315890871beSChris Mason write_unlock(&em_tree->lock); 33165d4f98a2SYan Zheng if (ret != -EEXIST) { 33175d4f98a2SYan Zheng free_extent_map(em); 33185d4f98a2SYan Zheng break; 33195d4f98a2SYan Zheng } 3320dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 33215d4f98a2SYan Zheng } 3322d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 33230257bb82SYan, Zheng return ret; 33240257bb82SYan, Zheng } 33255d4f98a2SYan Zheng 3326726a3421SQu Wenruo /* 3327726a3421SQu Wenruo * Allow error injection to test balance cancellation 3328726a3421SQu Wenruo */ 3329726a3421SQu Wenruo int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 3330726a3421SQu Wenruo { 3331726a3421SQu Wenruo return atomic_read(&fs_info->balance_cancel_req); 3332726a3421SQu Wenruo } 3333726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 3334726a3421SQu Wenruo 33350257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 33360257bb82SYan, Zheng struct file_extent_cluster *cluster) 33370257bb82SYan, Zheng { 33382ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 33390257bb82SYan, Zheng u64 page_start; 33400257bb82SYan, Zheng u64 page_end; 33410257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 33420257bb82SYan, Zheng unsigned long index; 33430257bb82SYan, Zheng unsigned long last_index; 33440257bb82SYan, Zheng struct page *page; 33450257bb82SYan, Zheng struct file_ra_state *ra; 33463b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 33470257bb82SYan, Zheng int nr = 0; 33480257bb82SYan, Zheng int ret = 0; 33490257bb82SYan, Zheng 33500257bb82SYan, Zheng if (!cluster->nr) 33510257bb82SYan, Zheng return 0; 33520257bb82SYan, Zheng 33530257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 33540257bb82SYan, Zheng if (!ra) 33550257bb82SYan, Zheng return -ENOMEM; 33560257bb82SYan, Zheng 3357efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 33580257bb82SYan, Zheng if (ret) 3359efa56464SYan, Zheng goto out; 33600257bb82SYan, Zheng 33610257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 33620257bb82SYan, Zheng 3363efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 3364efa56464SYan, Zheng cluster->end - offset, cluster->start); 3365efa56464SYan, Zheng if (ret) 3366efa56464SYan, Zheng goto out; 3367efa56464SYan, Zheng 336809cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 336909cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 33700257bb82SYan, Zheng while (index <= last_index) { 33719f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 33729f3db423SNikolay Borisov PAGE_SIZE); 3373efa56464SYan, Zheng if (ret) 3374efa56464SYan, Zheng goto out; 3375efa56464SYan, Zheng 33760257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 33770257bb82SYan, Zheng if (!page) { 33780257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 33790257bb82SYan, Zheng ra, NULL, index, 33800257bb82SYan, Zheng last_index + 1 - index); 3381a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 33823b16a4e3SJosef Bacik mask); 33830257bb82SYan, Zheng if (!page) { 3384691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 338543b18595SQu Wenruo PAGE_SIZE, true); 338644db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 33878702ba93SQu Wenruo PAGE_SIZE); 33880257bb82SYan, Zheng ret = -ENOMEM; 3389efa56464SYan, Zheng goto out; 33900257bb82SYan, Zheng } 33910257bb82SYan, Zheng } 33920257bb82SYan, Zheng 33930257bb82SYan, Zheng if (PageReadahead(page)) { 33940257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 33950257bb82SYan, Zheng ra, NULL, page, index, 33960257bb82SYan, Zheng last_index + 1 - index); 33970257bb82SYan, Zheng } 33980257bb82SYan, Zheng 33990257bb82SYan, Zheng if (!PageUptodate(page)) { 34000257bb82SYan, Zheng btrfs_readpage(NULL, page); 34010257bb82SYan, Zheng lock_page(page); 34020257bb82SYan, Zheng if (!PageUptodate(page)) { 34030257bb82SYan, Zheng unlock_page(page); 340409cbfeafSKirill A. Shutemov put_page(page); 3405691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 340643b18595SQu Wenruo PAGE_SIZE, true); 34078b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 34088702ba93SQu Wenruo PAGE_SIZE); 34090257bb82SYan, Zheng ret = -EIO; 3410efa56464SYan, Zheng goto out; 34110257bb82SYan, Zheng } 34120257bb82SYan, Zheng } 34130257bb82SYan, Zheng 34144eee4fa4SMiao Xie page_start = page_offset(page); 341509cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 34160257bb82SYan, Zheng 3417d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 34180257bb82SYan, Zheng 34190257bb82SYan, Zheng set_page_extent_mapped(page); 34200257bb82SYan, Zheng 34210257bb82SYan, Zheng if (nr < cluster->nr && 34220257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 34230257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 34240257bb82SYan, Zheng page_start, page_end, 3425ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 34260257bb82SYan, Zheng nr++; 34270257bb82SYan, Zheng } 34280257bb82SYan, Zheng 3429765f3cebSNikolay Borisov ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0, 3430330a5827SNikolay Borisov NULL); 3431765f3cebSNikolay Borisov if (ret) { 3432765f3cebSNikolay Borisov unlock_page(page); 3433765f3cebSNikolay Borisov put_page(page); 3434765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 343543b18595SQu Wenruo PAGE_SIZE, true); 3436765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 34378702ba93SQu Wenruo PAGE_SIZE); 3438765f3cebSNikolay Borisov 3439765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 3440765f3cebSNikolay Borisov page_start, page_end, 3441765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 3442765f3cebSNikolay Borisov goto out; 3443765f3cebSNikolay Borisov 3444765f3cebSNikolay Borisov } 34450257bb82SYan, Zheng set_page_dirty(page); 34460257bb82SYan, Zheng 34470257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 3448d0082371SJeff Mahoney page_start, page_end); 34490257bb82SYan, Zheng unlock_page(page); 345009cbfeafSKirill A. Shutemov put_page(page); 34510257bb82SYan, Zheng 34520257bb82SYan, Zheng index++; 34538702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 3454efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 34552ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 34567f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 34577f913c7cSQu Wenruo ret = -ECANCELED; 34587f913c7cSQu Wenruo goto out; 34597f913c7cSQu Wenruo } 34600257bb82SYan, Zheng } 34610257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 3462efa56464SYan, Zheng out: 34630257bb82SYan, Zheng kfree(ra); 34640257bb82SYan, Zheng return ret; 34650257bb82SYan, Zheng } 34660257bb82SYan, Zheng 34670257bb82SYan, Zheng static noinline_for_stack 34680257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 34690257bb82SYan, Zheng struct file_extent_cluster *cluster) 34700257bb82SYan, Zheng { 34710257bb82SYan, Zheng int ret; 34720257bb82SYan, Zheng 34730257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 34740257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 34750257bb82SYan, Zheng if (ret) 34760257bb82SYan, Zheng return ret; 34770257bb82SYan, Zheng cluster->nr = 0; 34780257bb82SYan, Zheng } 34790257bb82SYan, Zheng 34800257bb82SYan, Zheng if (!cluster->nr) 34810257bb82SYan, Zheng cluster->start = extent_key->objectid; 34820257bb82SYan, Zheng else 34830257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 34840257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 34850257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 34860257bb82SYan, Zheng cluster->nr++; 34870257bb82SYan, Zheng 34880257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 34890257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 34900257bb82SYan, Zheng if (ret) 34910257bb82SYan, Zheng return ret; 34920257bb82SYan, Zheng cluster->nr = 0; 34930257bb82SYan, Zheng } 34940257bb82SYan, Zheng return 0; 34955d4f98a2SYan Zheng } 34965d4f98a2SYan Zheng 34975d4f98a2SYan Zheng /* 34985d4f98a2SYan Zheng * helper to add a tree block to the list. 34995d4f98a2SYan Zheng * the major work is getting the generation and level of the block 35005d4f98a2SYan Zheng */ 35015d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 35025d4f98a2SYan Zheng struct btrfs_key *extent_key, 35035d4f98a2SYan Zheng struct btrfs_path *path, 35045d4f98a2SYan Zheng struct rb_root *blocks) 35055d4f98a2SYan Zheng { 35065d4f98a2SYan Zheng struct extent_buffer *eb; 35075d4f98a2SYan Zheng struct btrfs_extent_item *ei; 35085d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 35095d4f98a2SYan Zheng struct tree_block *block; 35105d4f98a2SYan Zheng struct rb_node *rb_node; 35115d4f98a2SYan Zheng u32 item_size; 35125d4f98a2SYan Zheng int level = -1; 35137fdf4b60SWang Shilong u64 generation; 35145d4f98a2SYan Zheng 35155d4f98a2SYan Zheng eb = path->nodes[0]; 35165d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 35175d4f98a2SYan Zheng 35183173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 35193173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 35205d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 35215d4f98a2SYan Zheng struct btrfs_extent_item); 35223173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 35235d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 35245d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 35255d4f98a2SYan Zheng } else { 35263173a18fSJosef Bacik level = (int)extent_key->offset; 35273173a18fSJosef Bacik } 35283173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 35296d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3530ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3531ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3532ba3c2b19SNikolay Borisov return -EINVAL; 35333173a18fSJosef Bacik } else { 35345d4f98a2SYan Zheng BUG(); 35355d4f98a2SYan Zheng } 35365d4f98a2SYan Zheng 3537b3b4aa74SDavid Sterba btrfs_release_path(path); 35385d4f98a2SYan Zheng 35395d4f98a2SYan Zheng BUG_ON(level == -1); 35405d4f98a2SYan Zheng 35415d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 35425d4f98a2SYan Zheng if (!block) 35435d4f98a2SYan Zheng return -ENOMEM; 35445d4f98a2SYan Zheng 35455d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3546da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 35475d4f98a2SYan Zheng block->key.offset = generation; 35485d4f98a2SYan Zheng block->level = level; 35495d4f98a2SYan Zheng block->key_ready = 0; 35505d4f98a2SYan Zheng 3551e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 355243c04fb1SJeff Mahoney if (rb_node) 355343c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, block->bytenr); 35545d4f98a2SYan Zheng 35555d4f98a2SYan Zheng return 0; 35565d4f98a2SYan Zheng } 35575d4f98a2SYan Zheng 35585d4f98a2SYan Zheng /* 35595d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 35605d4f98a2SYan Zheng */ 35615d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 35625d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 35635d4f98a2SYan Zheng struct rb_root *blocks) 35645d4f98a2SYan Zheng { 35650b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 35665d4f98a2SYan Zheng struct btrfs_path *path; 35675d4f98a2SYan Zheng struct btrfs_key key; 35685d4f98a2SYan Zheng int ret; 35690b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 35705d4f98a2SYan Zheng 35717476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 35725d4f98a2SYan Zheng return 0; 35735d4f98a2SYan Zheng 3574e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 35755d4f98a2SYan Zheng return 0; 35765d4f98a2SYan Zheng 35775d4f98a2SYan Zheng path = btrfs_alloc_path(); 35785d4f98a2SYan Zheng if (!path) 35795d4f98a2SYan Zheng return -ENOMEM; 3580aee68ee5SJosef Bacik again: 35815d4f98a2SYan Zheng key.objectid = bytenr; 3582aee68ee5SJosef Bacik if (skinny) { 3583aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3584aee68ee5SJosef Bacik key.offset = (u64)-1; 3585aee68ee5SJosef Bacik } else { 35865d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 35875d4f98a2SYan Zheng key.offset = blocksize; 3588aee68ee5SJosef Bacik } 35895d4f98a2SYan Zheng 35905d4f98a2SYan Zheng path->search_commit_root = 1; 35915d4f98a2SYan Zheng path->skip_locking = 1; 35925d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 35935d4f98a2SYan Zheng if (ret < 0) 35945d4f98a2SYan Zheng goto out; 35955d4f98a2SYan Zheng 3596aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3597aee68ee5SJosef Bacik if (path->slots[0]) { 3598aee68ee5SJosef Bacik path->slots[0]--; 3599aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3600aee68ee5SJosef Bacik path->slots[0]); 36013173a18fSJosef Bacik if (key.objectid == bytenr && 3602aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3603aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3604aee68ee5SJosef Bacik key.offset == blocksize))) 36053173a18fSJosef Bacik ret = 0; 36063173a18fSJosef Bacik } 3607aee68ee5SJosef Bacik 3608aee68ee5SJosef Bacik if (ret) { 3609aee68ee5SJosef Bacik skinny = false; 3610aee68ee5SJosef Bacik btrfs_release_path(path); 3611aee68ee5SJosef Bacik goto again; 3612aee68ee5SJosef Bacik } 3613aee68ee5SJosef Bacik } 3614cdccee99SLiu Bo if (ret) { 3615cdccee99SLiu Bo ASSERT(ret == 1); 3616cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3617cdccee99SLiu Bo btrfs_err(fs_info, 3618cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3619cdccee99SLiu Bo bytenr); 3620cdccee99SLiu Bo WARN_ON(1); 3621cdccee99SLiu Bo ret = -EINVAL; 3622cdccee99SLiu Bo goto out; 3623cdccee99SLiu Bo } 36243173a18fSJosef Bacik 36255d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 36265d4f98a2SYan Zheng out: 36275d4f98a2SYan Zheng btrfs_free_path(path); 36285d4f98a2SYan Zheng return ret; 36295d4f98a2SYan Zheng } 36305d4f98a2SYan Zheng 36310af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 363232da5386SDavid Sterba struct btrfs_block_group *block_group, 36331bbc621eSChris Mason struct inode *inode, 36341bbc621eSChris Mason u64 ino) 36350af3d00bSJosef Bacik { 36360af3d00bSJosef Bacik struct btrfs_key key; 36370af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 36380af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 36390af3d00bSJosef Bacik int ret = 0; 36400af3d00bSJosef Bacik 36410af3d00bSJosef Bacik if (inode) 36420af3d00bSJosef Bacik goto truncate; 36430af3d00bSJosef Bacik 36440af3d00bSJosef Bacik key.objectid = ino; 36450af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 36460af3d00bSJosef Bacik key.offset = 0; 36470af3d00bSJosef Bacik 36484c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 36492e19f1f9SAl Viro if (IS_ERR(inode)) 36500af3d00bSJosef Bacik return -ENOENT; 36510af3d00bSJosef Bacik 36520af3d00bSJosef Bacik truncate: 36532ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 36547b61cd92SMiao Xie &fs_info->global_block_rsv); 36557b61cd92SMiao Xie if (ret) 36567b61cd92SMiao Xie goto out; 36577b61cd92SMiao Xie 36587a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 36590af3d00bSJosef Bacik if (IS_ERR(trans)) { 36603612b495STsutomu Itoh ret = PTR_ERR(trans); 36610af3d00bSJosef Bacik goto out; 36620af3d00bSJosef Bacik } 36630af3d00bSJosef Bacik 366477ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 36650af3d00bSJosef Bacik 36663a45bb20SJeff Mahoney btrfs_end_transaction(trans); 36672ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 36680af3d00bSJosef Bacik out: 36690af3d00bSJosef Bacik iput(inode); 36700af3d00bSJosef Bacik return ret; 36710af3d00bSJosef Bacik } 36720af3d00bSJosef Bacik 36735d4f98a2SYan Zheng /* 367419b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 367519b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 36765d4f98a2SYan Zheng */ 367719b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 367819b546d7SQu Wenruo struct btrfs_block_group *block_group, 367919b546d7SQu Wenruo u64 data_bytenr) 36805d4f98a2SYan Zheng { 368119b546d7SQu Wenruo u64 space_cache_ino; 368219b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 36835d4f98a2SYan Zheng struct btrfs_key key; 368419b546d7SQu Wenruo bool found = false; 368519b546d7SQu Wenruo int i; 36865d4f98a2SYan Zheng int ret; 36875d4f98a2SYan Zheng 368819b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 368919b546d7SQu Wenruo return 0; 36905d4f98a2SYan Zheng 369119b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 369219b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 369319b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 369419b546d7SQu Wenruo continue; 369519b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 369619b546d7SQu Wenruo if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG && 369719b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 369819b546d7SQu Wenruo found = true; 369919b546d7SQu Wenruo space_cache_ino = key.objectid; 370019b546d7SQu Wenruo break; 370119b546d7SQu Wenruo } 370219b546d7SQu Wenruo } 370319b546d7SQu Wenruo if (!found) 370419b546d7SQu Wenruo return -ENOENT; 370519b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 370619b546d7SQu Wenruo space_cache_ino); 37070af3d00bSJosef Bacik return ret; 37085d4f98a2SYan Zheng } 37095d4f98a2SYan Zheng 37105d4f98a2SYan Zheng /* 37112c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 37125d4f98a2SYan Zheng */ 37135d4f98a2SYan Zheng static noinline_for_stack 37145d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 37155d4f98a2SYan Zheng struct btrfs_key *extent_key, 37165d4f98a2SYan Zheng struct btrfs_path *path, 37175d4f98a2SYan Zheng struct rb_root *blocks) 37185d4f98a2SYan Zheng { 371919b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 372019b546d7SQu Wenruo struct ulist *leaves = NULL; 372119b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 372219b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 372319b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3724647f63bdSFilipe David Borba Manana int ret = 0; 37255d4f98a2SYan Zheng 3726b3b4aa74SDavid Sterba btrfs_release_path(path); 372719b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 372819b546d7SQu Wenruo 0, &leaves, NULL, true); 372919b546d7SQu Wenruo if (ret < 0) 373019b546d7SQu Wenruo return ret; 373119b546d7SQu Wenruo 373219b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 373319b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 373419b546d7SQu Wenruo struct extent_buffer *eb; 373519b546d7SQu Wenruo 373619b546d7SQu Wenruo eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL); 373719b546d7SQu Wenruo if (IS_ERR(eb)) { 373819b546d7SQu Wenruo ret = PTR_ERR(eb); 373919b546d7SQu Wenruo break; 374019b546d7SQu Wenruo } 374119b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 374219b546d7SQu Wenruo extent_key->objectid); 374319b546d7SQu Wenruo free_extent_buffer(eb); 374419b546d7SQu Wenruo if (ret < 0) 374519b546d7SQu Wenruo break; 374619b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 374719b546d7SQu Wenruo if (ret < 0) 374819b546d7SQu Wenruo break; 374919b546d7SQu Wenruo } 375019b546d7SQu Wenruo if (ret < 0) 37515d4f98a2SYan Zheng free_block_list(blocks); 375219b546d7SQu Wenruo ulist_free(leaves); 375319b546d7SQu Wenruo return ret; 37545d4f98a2SYan Zheng } 37555d4f98a2SYan Zheng 37565d4f98a2SYan Zheng /* 37572c016dc2SLiu Bo * helper to find next unprocessed extent 37585d4f98a2SYan Zheng */ 37595d4f98a2SYan Zheng static noinline_for_stack 3760147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 37613fd0a558SYan, Zheng struct btrfs_key *extent_key) 37625d4f98a2SYan Zheng { 37630b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 37645d4f98a2SYan Zheng struct btrfs_key key; 37655d4f98a2SYan Zheng struct extent_buffer *leaf; 37665d4f98a2SYan Zheng u64 start, end, last; 37675d4f98a2SYan Zheng int ret; 37685d4f98a2SYan Zheng 3769b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 37705d4f98a2SYan Zheng while (1) { 37715d4f98a2SYan Zheng cond_resched(); 37725d4f98a2SYan Zheng if (rc->search_start >= last) { 37735d4f98a2SYan Zheng ret = 1; 37745d4f98a2SYan Zheng break; 37755d4f98a2SYan Zheng } 37765d4f98a2SYan Zheng 37775d4f98a2SYan Zheng key.objectid = rc->search_start; 37785d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 37795d4f98a2SYan Zheng key.offset = 0; 37805d4f98a2SYan Zheng 37815d4f98a2SYan Zheng path->search_commit_root = 1; 37825d4f98a2SYan Zheng path->skip_locking = 1; 37835d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 37845d4f98a2SYan Zheng 0, 0); 37855d4f98a2SYan Zheng if (ret < 0) 37865d4f98a2SYan Zheng break; 37875d4f98a2SYan Zheng next: 37885d4f98a2SYan Zheng leaf = path->nodes[0]; 37895d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 37905d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 37915d4f98a2SYan Zheng if (ret != 0) 37925d4f98a2SYan Zheng break; 37935d4f98a2SYan Zheng leaf = path->nodes[0]; 37945d4f98a2SYan Zheng } 37955d4f98a2SYan Zheng 37965d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 37975d4f98a2SYan Zheng if (key.objectid >= last) { 37985d4f98a2SYan Zheng ret = 1; 37995d4f98a2SYan Zheng break; 38005d4f98a2SYan Zheng } 38015d4f98a2SYan Zheng 38023173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 38033173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 38043173a18fSJosef Bacik path->slots[0]++; 38053173a18fSJosef Bacik goto next; 38063173a18fSJosef Bacik } 38073173a18fSJosef Bacik 38083173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 38095d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 38105d4f98a2SYan Zheng path->slots[0]++; 38115d4f98a2SYan Zheng goto next; 38125d4f98a2SYan Zheng } 38135d4f98a2SYan Zheng 38143173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 38150b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 38163173a18fSJosef Bacik rc->search_start) { 38173173a18fSJosef Bacik path->slots[0]++; 38183173a18fSJosef Bacik goto next; 38193173a18fSJosef Bacik } 38203173a18fSJosef Bacik 38215d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 38225d4f98a2SYan Zheng key.objectid, &start, &end, 3823e6138876SJosef Bacik EXTENT_DIRTY, NULL); 38245d4f98a2SYan Zheng 38255d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3826b3b4aa74SDavid Sterba btrfs_release_path(path); 38275d4f98a2SYan Zheng rc->search_start = end + 1; 38285d4f98a2SYan Zheng } else { 38293173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 38305d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 38313173a18fSJosef Bacik else 38323173a18fSJosef Bacik rc->search_start = key.objectid + 38330b246afaSJeff Mahoney fs_info->nodesize; 38343fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 38355d4f98a2SYan Zheng return 0; 38365d4f98a2SYan Zheng } 38375d4f98a2SYan Zheng } 3838b3b4aa74SDavid Sterba btrfs_release_path(path); 38395d4f98a2SYan Zheng return ret; 38405d4f98a2SYan Zheng } 38415d4f98a2SYan Zheng 38425d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 38435d4f98a2SYan Zheng { 38445d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 38457585717fSChris Mason 38467585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 38475d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 38487585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 38495d4f98a2SYan Zheng } 38505d4f98a2SYan Zheng 38515d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 38525d4f98a2SYan Zheng { 38535d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 38547585717fSChris Mason 38557585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 38565d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 38577585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 38585d4f98a2SYan Zheng } 38595d4f98a2SYan Zheng 38605d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 38615d4f98a2SYan Zheng { 38625d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 38635d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 38645d4f98a2SYan Zheng return 1; 38655d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 38665d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 38675d4f98a2SYan Zheng return 1; 38685d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 38695d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 38705d4f98a2SYan Zheng return 1; 38715d4f98a2SYan Zheng return 0; 38725d4f98a2SYan Zheng } 38735d4f98a2SYan Zheng 38743fd0a558SYan, Zheng static noinline_for_stack 38753fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 38763fd0a558SYan, Zheng { 38773fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3878ac2fabacSJosef Bacik int ret; 38793fd0a558SYan, Zheng 38802ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 388166d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 38823fd0a558SYan, Zheng if (!rc->block_rsv) 38833fd0a558SYan, Zheng return -ENOMEM; 38843fd0a558SYan, Zheng 38853fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3886b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 38873fd0a558SYan, Zheng rc->extents_found = 0; 38883fd0a558SYan, Zheng rc->nodes_relocated = 0; 38893fd0a558SYan, Zheng rc->merging_rsv_size = 0; 38900647bf56SWang Shilong rc->reserved_bytes = 0; 3891da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 38920647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3893ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3894ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3895ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3896ac2fabacSJosef Bacik if (ret) 3897ac2fabacSJosef Bacik return ret; 38983fd0a558SYan, Zheng 38993fd0a558SYan, Zheng rc->create_reloc_tree = 1; 39003fd0a558SYan, Zheng set_reloc_control(rc); 39013fd0a558SYan, Zheng 39027a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 390328818947SLiu Bo if (IS_ERR(trans)) { 390428818947SLiu Bo unset_reloc_control(rc); 390528818947SLiu Bo /* 390628818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 390728818947SLiu Bo * cleanup. And callers are responsible to free the above 390828818947SLiu Bo * block rsv. 390928818947SLiu Bo */ 391028818947SLiu Bo return PTR_ERR(trans); 391128818947SLiu Bo } 39123a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 39133fd0a558SYan, Zheng return 0; 39143fd0a558SYan, Zheng } 391576dda93cSYan, Zheng 39165d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 39175d4f98a2SYan Zheng { 39182ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39195d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 39205d4f98a2SYan Zheng struct btrfs_key key; 39215d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 39225d4f98a2SYan Zheng struct btrfs_path *path; 39235d4f98a2SYan Zheng struct btrfs_extent_item *ei; 39245d4f98a2SYan Zheng u64 flags; 39255d4f98a2SYan Zheng u32 item_size; 39265d4f98a2SYan Zheng int ret; 39275d4f98a2SYan Zheng int err = 0; 3928c87f08caSChris Mason int progress = 0; 39295d4f98a2SYan Zheng 39305d4f98a2SYan Zheng path = btrfs_alloc_path(); 39313fd0a558SYan, Zheng if (!path) 39325d4f98a2SYan Zheng return -ENOMEM; 3933e4058b54SDavid Sterba path->reada = READA_FORWARD; 39343fd0a558SYan, Zheng 39353fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 39363fd0a558SYan, Zheng if (ret) { 39373fd0a558SYan, Zheng err = ret; 39383fd0a558SYan, Zheng goto out_free; 39392423fdfbSJiri Slaby } 39405d4f98a2SYan Zheng 39415d4f98a2SYan Zheng while (1) { 39420647bf56SWang Shilong rc->reserved_bytes = 0; 39430647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 39440647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 39450647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 39460647bf56SWang Shilong if (ret) { 39470647bf56SWang Shilong err = ret; 39480647bf56SWang Shilong break; 39490647bf56SWang Shilong } 3950c87f08caSChris Mason progress++; 3951a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 39520f788c58SLiu Bo if (IS_ERR(trans)) { 39530f788c58SLiu Bo err = PTR_ERR(trans); 39540f788c58SLiu Bo trans = NULL; 39550f788c58SLiu Bo break; 39560f788c58SLiu Bo } 3957c87f08caSChris Mason restart: 39583fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 39593a45bb20SJeff Mahoney btrfs_end_transaction(trans); 396042a657f5SPan Bian trans = NULL; 39613fd0a558SYan, Zheng continue; 39623fd0a558SYan, Zheng } 39633fd0a558SYan, Zheng 3964147d256eSZhaolei ret = find_next_extent(rc, path, &key); 39655d4f98a2SYan Zheng if (ret < 0) 39665d4f98a2SYan Zheng err = ret; 39675d4f98a2SYan Zheng if (ret != 0) 39685d4f98a2SYan Zheng break; 39695d4f98a2SYan Zheng 39705d4f98a2SYan Zheng rc->extents_found++; 39715d4f98a2SYan Zheng 39725d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 39735d4f98a2SYan Zheng struct btrfs_extent_item); 39743fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 39755d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 39765d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 39775d4f98a2SYan Zheng ret = check_extent_flags(flags); 39785d4f98a2SYan Zheng BUG_ON(ret); 39796d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3980ba3c2b19SNikolay Borisov err = -EINVAL; 3981ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 3982ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 3983ba3c2b19SNikolay Borisov break; 39845d4f98a2SYan Zheng } else { 39855d4f98a2SYan Zheng BUG(); 39865d4f98a2SYan Zheng } 39875d4f98a2SYan Zheng 39885d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 39895d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 39905d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 39915d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 39925d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 39935d4f98a2SYan Zheng } else { 3994b3b4aa74SDavid Sterba btrfs_release_path(path); 39955d4f98a2SYan Zheng ret = 0; 39965d4f98a2SYan Zheng } 39975d4f98a2SYan Zheng if (ret < 0) { 39983fd0a558SYan, Zheng err = ret; 39995d4f98a2SYan Zheng break; 40005d4f98a2SYan Zheng } 40015d4f98a2SYan Zheng 40025d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 40035d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 40045d4f98a2SYan Zheng if (ret < 0) { 40053fd0a558SYan, Zheng if (ret != -EAGAIN) { 40065d4f98a2SYan Zheng err = ret; 40075d4f98a2SYan Zheng break; 40085d4f98a2SYan Zheng } 40093fd0a558SYan, Zheng rc->extents_found--; 40103fd0a558SYan, Zheng rc->search_start = key.objectid; 40113fd0a558SYan, Zheng } 40125d4f98a2SYan Zheng } 40135d4f98a2SYan Zheng 40143a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 40152ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 40163fd0a558SYan, Zheng trans = NULL; 40175d4f98a2SYan Zheng 40185d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 40195d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 40205d4f98a2SYan Zheng rc->found_file_extent = 1; 40210257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 40223fd0a558SYan, Zheng &key, &rc->cluster); 40235d4f98a2SYan Zheng if (ret < 0) { 40245d4f98a2SYan Zheng err = ret; 40255d4f98a2SYan Zheng break; 40265d4f98a2SYan Zheng } 40275d4f98a2SYan Zheng } 4028f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 4029f31ea088SQu Wenruo err = -ECANCELED; 4030f31ea088SQu Wenruo break; 4031f31ea088SQu Wenruo } 40325d4f98a2SYan Zheng } 4033c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 403443a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 40359689457bSShilong Wang if (ret == 1) { 4036c87f08caSChris Mason err = 0; 4037c87f08caSChris Mason progress = 0; 4038c87f08caSChris Mason goto restart; 4039c87f08caSChris Mason } 4040c87f08caSChris Mason } 40413fd0a558SYan, Zheng 4042b3b4aa74SDavid Sterba btrfs_release_path(path); 404391166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 40445d4f98a2SYan Zheng 40455d4f98a2SYan Zheng if (trans) { 40463a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 40472ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 40485d4f98a2SYan Zheng } 40495d4f98a2SYan Zheng 40500257bb82SYan, Zheng if (!err) { 40513fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 40523fd0a558SYan, Zheng &rc->cluster); 40530257bb82SYan, Zheng if (ret < 0) 40540257bb82SYan, Zheng err = ret; 40550257bb82SYan, Zheng } 40560257bb82SYan, Zheng 40573fd0a558SYan, Zheng rc->create_reloc_tree = 0; 40583fd0a558SYan, Zheng set_reloc_control(rc); 40590257bb82SYan, Zheng 40603fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 406163f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 40625d4f98a2SYan Zheng 40637f913c7cSQu Wenruo /* 40647f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 40657f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 40667f913c7cSQu Wenruo * 40677f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 40687f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 40697f913c7cSQu Wenruo * merge_reloc_roots() 40707f913c7cSQu Wenruo */ 40713fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 40725d4f98a2SYan Zheng 40735d4f98a2SYan Zheng merge_reloc_roots(rc); 40745d4f98a2SYan Zheng 40753fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 40765d4f98a2SYan Zheng unset_reloc_control(rc); 407763f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 40785d4f98a2SYan Zheng 40795d4f98a2SYan Zheng /* get rid of pinned extents */ 40807a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 408162b99540SQu Wenruo if (IS_ERR(trans)) { 40823612b495STsutomu Itoh err = PTR_ERR(trans); 408362b99540SQu Wenruo goto out_free; 408462b99540SQu Wenruo } 40853a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 40866217b0faSJosef Bacik out_free: 4087d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4088d2311e69SQu Wenruo if (ret < 0 && !err) 4089d2311e69SQu Wenruo err = ret; 40902ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 40913fd0a558SYan, Zheng btrfs_free_path(path); 40925d4f98a2SYan Zheng return err; 40935d4f98a2SYan Zheng } 40945d4f98a2SYan Zheng 40955d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 40960257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 40975d4f98a2SYan Zheng { 40985d4f98a2SYan Zheng struct btrfs_path *path; 40995d4f98a2SYan Zheng struct btrfs_inode_item *item; 41005d4f98a2SYan Zheng struct extent_buffer *leaf; 41015d4f98a2SYan Zheng int ret; 41025d4f98a2SYan Zheng 41035d4f98a2SYan Zheng path = btrfs_alloc_path(); 41045d4f98a2SYan Zheng if (!path) 41055d4f98a2SYan Zheng return -ENOMEM; 41065d4f98a2SYan Zheng 41075d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 41085d4f98a2SYan Zheng if (ret) 41095d4f98a2SYan Zheng goto out; 41105d4f98a2SYan Zheng 41115d4f98a2SYan Zheng leaf = path->nodes[0]; 41125d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 4113b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 41145d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 41150257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 41165d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 41173fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 41183fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 41195d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 41205d4f98a2SYan Zheng out: 41215d4f98a2SYan Zheng btrfs_free_path(path); 41225d4f98a2SYan Zheng return ret; 41235d4f98a2SYan Zheng } 41245d4f98a2SYan Zheng 41255d4f98a2SYan Zheng /* 41265d4f98a2SYan Zheng * helper to create inode for data relocation. 41275d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 41285d4f98a2SYan Zheng */ 41293fd0a558SYan, Zheng static noinline_for_stack 41303fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 413132da5386SDavid Sterba struct btrfs_block_group *group) 41325d4f98a2SYan Zheng { 41335d4f98a2SYan Zheng struct inode *inode = NULL; 41345d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 41355d4f98a2SYan Zheng struct btrfs_root *root; 41365d4f98a2SYan Zheng struct btrfs_key key; 41374624900dSZhaolei u64 objectid; 41385d4f98a2SYan Zheng int err = 0; 41395d4f98a2SYan Zheng 41405d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 41415d4f98a2SYan Zheng if (IS_ERR(root)) 41425d4f98a2SYan Zheng return ERR_CAST(root); 41435d4f98a2SYan Zheng 4144a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 414576deacf0SJosef Bacik if (IS_ERR(trans)) { 414600246528SJosef Bacik btrfs_put_root(root); 41473fd0a558SYan, Zheng return ERR_CAST(trans); 414876deacf0SJosef Bacik } 41495d4f98a2SYan Zheng 4150581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 41515d4f98a2SYan Zheng if (err) 41525d4f98a2SYan Zheng goto out; 41535d4f98a2SYan Zheng 41540257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 41555d4f98a2SYan Zheng BUG_ON(err); 41565d4f98a2SYan Zheng 41575d4f98a2SYan Zheng key.objectid = objectid; 41585d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 41595d4f98a2SYan Zheng key.offset = 0; 41604c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 41612e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 4162b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 41635d4f98a2SYan Zheng 416473f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 41655d4f98a2SYan Zheng out: 416600246528SJosef Bacik btrfs_put_root(root); 41673a45bb20SJeff Mahoney btrfs_end_transaction(trans); 41682ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 41695d4f98a2SYan Zheng if (err) { 41705d4f98a2SYan Zheng if (inode) 41715d4f98a2SYan Zheng iput(inode); 41725d4f98a2SYan Zheng inode = ERR_PTR(err); 41735d4f98a2SYan Zheng } 41745d4f98a2SYan Zheng return inode; 41755d4f98a2SYan Zheng } 41765d4f98a2SYan Zheng 4177c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 41783fd0a558SYan, Zheng { 41793fd0a558SYan, Zheng struct reloc_control *rc; 41803fd0a558SYan, Zheng 41813fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 41823fd0a558SYan, Zheng if (!rc) 41833fd0a558SYan, Zheng return NULL; 41843fd0a558SYan, Zheng 41853fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 4186d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 4187584fb121SQu Wenruo btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1); 41883fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 418943eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 419043eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 41913fd0a558SYan, Zheng return rc; 41923fd0a558SYan, Zheng } 41933fd0a558SYan, Zheng 41941a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 41951a0afa0eSJosef Bacik { 41961a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 41971a0afa0eSJosef Bacik 41981a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 41991a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 42001a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 42011a0afa0eSJosef Bacik kfree(node); 42021a0afa0eSJosef Bacik 42031a0afa0eSJosef Bacik kfree(rc); 42041a0afa0eSJosef Bacik } 42051a0afa0eSJosef Bacik 42065d4f98a2SYan Zheng /* 4207ebce0e01SAdam Borowski * Print the block group being relocated 4208ebce0e01SAdam Borowski */ 4209ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 421032da5386SDavid Sterba struct btrfs_block_group *block_group) 4211ebce0e01SAdam Borowski { 4212f89e09cfSAnand Jain char buf[128] = {'\0'}; 4213ebce0e01SAdam Borowski 4214f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 4215ebce0e01SAdam Borowski 4216ebce0e01SAdam Borowski btrfs_info(fs_info, 4217ebce0e01SAdam Borowski "relocating block group %llu flags %s", 4218b3470b5dSDavid Sterba block_group->start, buf); 4219ebce0e01SAdam Borowski } 4220ebce0e01SAdam Borowski 4221430640e3SQu Wenruo static const char *stage_to_string(int stage) 4222430640e3SQu Wenruo { 4223430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 4224430640e3SQu Wenruo return "move data extents"; 4225430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 4226430640e3SQu Wenruo return "update data pointers"; 4227430640e3SQu Wenruo return "unknown"; 4228430640e3SQu Wenruo } 4229430640e3SQu Wenruo 4230ebce0e01SAdam Borowski /* 42315d4f98a2SYan Zheng * function to relocate all extents in a block group. 42325d4f98a2SYan Zheng */ 42336bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 42345d4f98a2SYan Zheng { 423532da5386SDavid Sterba struct btrfs_block_group *bg; 42366bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 42375d4f98a2SYan Zheng struct reloc_control *rc; 42380af3d00bSJosef Bacik struct inode *inode; 42390af3d00bSJosef Bacik struct btrfs_path *path; 42405d4f98a2SYan Zheng int ret; 4241f0486c68SYan, Zheng int rw = 0; 42425d4f98a2SYan Zheng int err = 0; 42435d4f98a2SYan Zheng 4244eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 4245eede2bf3SOmar Sandoval if (!bg) 4246eede2bf3SOmar Sandoval return -ENOENT; 4247eede2bf3SOmar Sandoval 4248eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 4249eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 4250eede2bf3SOmar Sandoval return -ETXTBSY; 4251eede2bf3SOmar Sandoval } 4252eede2bf3SOmar Sandoval 4253c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 4254eede2bf3SOmar Sandoval if (!rc) { 4255eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 42565d4f98a2SYan Zheng return -ENOMEM; 4257eede2bf3SOmar Sandoval } 42585d4f98a2SYan Zheng 4259f0486c68SYan, Zheng rc->extent_root = extent_root; 4260eede2bf3SOmar Sandoval rc->block_group = bg; 42615d4f98a2SYan Zheng 4262b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 4263f0486c68SYan, Zheng if (ret) { 4264f0486c68SYan, Zheng err = ret; 4265f0486c68SYan, Zheng goto out; 4266f0486c68SYan, Zheng } 4267f0486c68SYan, Zheng rw = 1; 4268f0486c68SYan, Zheng 42690af3d00bSJosef Bacik path = btrfs_alloc_path(); 42700af3d00bSJosef Bacik if (!path) { 42710af3d00bSJosef Bacik err = -ENOMEM; 42720af3d00bSJosef Bacik goto out; 42730af3d00bSJosef Bacik } 42740af3d00bSJosef Bacik 42757949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 42760af3d00bSJosef Bacik btrfs_free_path(path); 42770af3d00bSJosef Bacik 42780af3d00bSJosef Bacik if (!IS_ERR(inode)) 42791bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 42800af3d00bSJosef Bacik else 42810af3d00bSJosef Bacik ret = PTR_ERR(inode); 42820af3d00bSJosef Bacik 42830af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 42840af3d00bSJosef Bacik err = ret; 42850af3d00bSJosef Bacik goto out; 42860af3d00bSJosef Bacik } 42870af3d00bSJosef Bacik 42885d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 42895d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 42905d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 42915d4f98a2SYan Zheng rc->data_inode = NULL; 42925d4f98a2SYan Zheng goto out; 42935d4f98a2SYan Zheng } 42945d4f98a2SYan Zheng 42950b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 42965d4f98a2SYan Zheng 42979cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 4298f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 42996374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 4300b3470b5dSDavid Sterba rc->block_group->start, 4301b3470b5dSDavid Sterba rc->block_group->length); 43025d4f98a2SYan Zheng 43035d4f98a2SYan Zheng while (1) { 4304430640e3SQu Wenruo int finishes_stage; 4305430640e3SQu Wenruo 430676dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 43075d4f98a2SYan Zheng ret = relocate_block_group(rc); 430876dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 4309ff612ba7SJosef Bacik if (ret < 0) 43105d4f98a2SYan Zheng err = ret; 4311ff612ba7SJosef Bacik 4312430640e3SQu Wenruo finishes_stage = rc->stage; 4313ff612ba7SJosef Bacik /* 4314ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 4315ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 4316ff612ba7SJosef Bacik * different block group we could end up hitting the 4317ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 4318ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 4319ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 4320ff612ba7SJosef Bacik * out of the loop if we hit an error. 4321ff612ba7SJosef Bacik */ 4322ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 4323ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 4324ff612ba7SJosef Bacik (u64)-1); 4325ff612ba7SJosef Bacik if (ret) 4326ff612ba7SJosef Bacik err = ret; 4327ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 4328ff612ba7SJosef Bacik 0, -1); 4329ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 43305d4f98a2SYan Zheng } 43315d4f98a2SYan Zheng 4332ff612ba7SJosef Bacik if (err < 0) 4333ff612ba7SJosef Bacik goto out; 4334ff612ba7SJosef Bacik 43355d4f98a2SYan Zheng if (rc->extents_found == 0) 43365d4f98a2SYan Zheng break; 43375d4f98a2SYan Zheng 4338430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 4339430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 43405d4f98a2SYan Zheng } 43415d4f98a2SYan Zheng 43425d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 43435d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 4344bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 43455d4f98a2SYan Zheng out: 4346f0486c68SYan, Zheng if (err && rw) 43472ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 43485d4f98a2SYan Zheng iput(rc->data_inode); 43495d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 43501a0afa0eSJosef Bacik free_reloc_control(rc); 43515d4f98a2SYan Zheng return err; 43525d4f98a2SYan Zheng } 43535d4f98a2SYan Zheng 435476dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 435576dda93cSYan, Zheng { 43560b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 435776dda93cSYan, Zheng struct btrfs_trans_handle *trans; 435879787eaaSJeff Mahoney int ret, err; 435976dda93cSYan, Zheng 43600b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 436179787eaaSJeff Mahoney if (IS_ERR(trans)) 436279787eaaSJeff Mahoney return PTR_ERR(trans); 436376dda93cSYan, Zheng 436476dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 436576dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 436676dda93cSYan, Zheng root->root_item.drop_level = 0; 436776dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 43680b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 436976dda93cSYan, Zheng &root->root_key, &root->root_item); 437076dda93cSYan, Zheng 43713a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 437279787eaaSJeff Mahoney if (err) 437379787eaaSJeff Mahoney return err; 437479787eaaSJeff Mahoney return ret; 437576dda93cSYan, Zheng } 437676dda93cSYan, Zheng 43775d4f98a2SYan Zheng /* 43785d4f98a2SYan Zheng * recover relocation interrupted by system crash. 43795d4f98a2SYan Zheng * 43805d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 43815d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 43825d4f98a2SYan Zheng */ 43835d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 43845d4f98a2SYan Zheng { 43850b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 43865d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 43875d4f98a2SYan Zheng struct btrfs_key key; 43885d4f98a2SYan Zheng struct btrfs_root *fs_root; 43895d4f98a2SYan Zheng struct btrfs_root *reloc_root; 43905d4f98a2SYan Zheng struct btrfs_path *path; 43915d4f98a2SYan Zheng struct extent_buffer *leaf; 43925d4f98a2SYan Zheng struct reloc_control *rc = NULL; 43935d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 43945d4f98a2SYan Zheng int ret; 43955d4f98a2SYan Zheng int err = 0; 43965d4f98a2SYan Zheng 43975d4f98a2SYan Zheng path = btrfs_alloc_path(); 43985d4f98a2SYan Zheng if (!path) 43995d4f98a2SYan Zheng return -ENOMEM; 4400e4058b54SDavid Sterba path->reada = READA_BACK; 44015d4f98a2SYan Zheng 44025d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 44035d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 44045d4f98a2SYan Zheng key.offset = (u64)-1; 44055d4f98a2SYan Zheng 44065d4f98a2SYan Zheng while (1) { 44070b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 44085d4f98a2SYan Zheng path, 0, 0); 44095d4f98a2SYan Zheng if (ret < 0) { 44105d4f98a2SYan Zheng err = ret; 44115d4f98a2SYan Zheng goto out; 44125d4f98a2SYan Zheng } 44135d4f98a2SYan Zheng if (ret > 0) { 44145d4f98a2SYan Zheng if (path->slots[0] == 0) 44155d4f98a2SYan Zheng break; 44165d4f98a2SYan Zheng path->slots[0]--; 44175d4f98a2SYan Zheng } 44185d4f98a2SYan Zheng leaf = path->nodes[0]; 44195d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4420b3b4aa74SDavid Sterba btrfs_release_path(path); 44215d4f98a2SYan Zheng 44225d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 44235d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 44245d4f98a2SYan Zheng break; 44255d4f98a2SYan Zheng 44263dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 44275d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 44285d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 44295d4f98a2SYan Zheng goto out; 44305d4f98a2SYan Zheng } 44315d4f98a2SYan Zheng 44323dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 44335d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 44345d4f98a2SYan Zheng 44355d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 44360b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, 44375d4f98a2SYan Zheng reloc_root->root_key.offset); 44385d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 443976dda93cSYan, Zheng ret = PTR_ERR(fs_root); 444076dda93cSYan, Zheng if (ret != -ENOENT) { 444176dda93cSYan, Zheng err = ret; 44425d4f98a2SYan Zheng goto out; 44435d4f98a2SYan Zheng } 444479787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 444579787eaaSJeff Mahoney if (ret < 0) { 444679787eaaSJeff Mahoney err = ret; 444779787eaaSJeff Mahoney goto out; 444879787eaaSJeff Mahoney } 4449932fd26dSJosef Bacik } else { 445000246528SJosef Bacik btrfs_put_root(fs_root); 445176dda93cSYan, Zheng } 44525d4f98a2SYan Zheng } 44535d4f98a2SYan Zheng 44545d4f98a2SYan Zheng if (key.offset == 0) 44555d4f98a2SYan Zheng break; 44565d4f98a2SYan Zheng 44575d4f98a2SYan Zheng key.offset--; 44585d4f98a2SYan Zheng } 4459b3b4aa74SDavid Sterba btrfs_release_path(path); 44605d4f98a2SYan Zheng 44615d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 44625d4f98a2SYan Zheng goto out; 44635d4f98a2SYan Zheng 4464c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 44655d4f98a2SYan Zheng if (!rc) { 44665d4f98a2SYan Zheng err = -ENOMEM; 44675d4f98a2SYan Zheng goto out; 44685d4f98a2SYan Zheng } 44695d4f98a2SYan Zheng 44700b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 44715d4f98a2SYan Zheng 44725d4f98a2SYan Zheng set_reloc_control(rc); 44735d4f98a2SYan Zheng 44747a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 44753612b495STsutomu Itoh if (IS_ERR(trans)) { 44763612b495STsutomu Itoh err = PTR_ERR(trans); 4477fb2d83eeSJosef Bacik goto out_unset; 44783612b495STsutomu Itoh } 44793fd0a558SYan, Zheng 44803fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 44813fd0a558SYan, Zheng 44825d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 44835d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 44845d4f98a2SYan Zheng struct btrfs_root, root_list); 44855d4f98a2SYan Zheng list_del(&reloc_root->root_list); 44865d4f98a2SYan Zheng 44875d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 44885d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 44895d4f98a2SYan Zheng &rc->reloc_roots); 44905d4f98a2SYan Zheng continue; 44915d4f98a2SYan Zheng } 44925d4f98a2SYan Zheng 44930b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, reloc_root->root_key.offset); 449479787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 449579787eaaSJeff Mahoney err = PTR_ERR(fs_root); 4496ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 44971402d17dSXiyu Yang btrfs_end_transaction(trans); 4498fb2d83eeSJosef Bacik goto out_unset; 449979787eaaSJeff Mahoney } 45005d4f98a2SYan Zheng 4501ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 450279787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 4503f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 450400246528SJosef Bacik btrfs_put_root(fs_root); 45055d4f98a2SYan Zheng } 45065d4f98a2SYan Zheng 45073a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 450879787eaaSJeff Mahoney if (err) 4509fb2d83eeSJosef Bacik goto out_unset; 45105d4f98a2SYan Zheng 45115d4f98a2SYan Zheng merge_reloc_roots(rc); 45125d4f98a2SYan Zheng 45135d4f98a2SYan Zheng unset_reloc_control(rc); 45145d4f98a2SYan Zheng 45157a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 451662b99540SQu Wenruo if (IS_ERR(trans)) { 45173612b495STsutomu Itoh err = PTR_ERR(trans); 45186217b0faSJosef Bacik goto out_clean; 451962b99540SQu Wenruo } 45203a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 45216217b0faSJosef Bacik out_clean: 4522d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4523d2311e69SQu Wenruo if (ret < 0 && !err) 4524d2311e69SQu Wenruo err = ret; 4525fb2d83eeSJosef Bacik out_unset: 4526fb2d83eeSJosef Bacik unset_reloc_control(rc); 45271a0afa0eSJosef Bacik free_reloc_control(rc); 45283612b495STsutomu Itoh out: 4529aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 4530aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4531aca1bba6SLiu Bo 45325d4f98a2SYan Zheng btrfs_free_path(path); 45335d4f98a2SYan Zheng 45345d4f98a2SYan Zheng if (err == 0) { 45355d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 45360b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 4537932fd26dSJosef Bacik if (IS_ERR(fs_root)) { 45385d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4539932fd26dSJosef Bacik } else { 454066b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 454100246528SJosef Bacik btrfs_put_root(fs_root); 4542932fd26dSJosef Bacik } 4543932fd26dSJosef Bacik } 45445d4f98a2SYan Zheng return err; 45455d4f98a2SYan Zheng } 45465d4f98a2SYan Zheng 45475d4f98a2SYan Zheng /* 45485d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 45495d4f98a2SYan Zheng * 45505d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 45515d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 45525d4f98a2SYan Zheng */ 45535d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 45545d4f98a2SYan Zheng { 45550b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 45565d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 45575d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 45585d4f98a2SYan Zheng int ret; 45595d4f98a2SYan Zheng u64 disk_bytenr; 45604577b014SJosef Bacik u64 new_bytenr; 45615d4f98a2SYan Zheng LIST_HEAD(list); 45625d4f98a2SYan Zheng 45635d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4564bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 45655d4f98a2SYan Zheng 45665d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 45670b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4568a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 456979787eaaSJeff Mahoney if (ret) 457079787eaaSJeff Mahoney goto out; 45715d4f98a2SYan Zheng 45725d4f98a2SYan Zheng while (!list_empty(&list)) { 45735d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 45745d4f98a2SYan Zheng list_del_init(&sums->list); 45755d4f98a2SYan Zheng 45764577b014SJosef Bacik /* 45774577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 45784577b014SJosef Bacik * We need to do this because we will read in entire prealloc 45794577b014SJosef Bacik * extents but we may have written to say the middle of the 45804577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 45814577b014SJosef Bacik * the right disk offset. 45824577b014SJosef Bacik * 45834577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 45844577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 45854577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 45864577b014SJosef Bacik * disk length. 45874577b014SJosef Bacik */ 4588bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 45894577b014SJosef Bacik sums->bytenr = new_bytenr; 45905d4f98a2SYan Zheng 4591f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 45925d4f98a2SYan Zheng } 459379787eaaSJeff Mahoney out: 45945d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4595411fc6bcSAndi Kleen return ret; 45965d4f98a2SYan Zheng } 45973fd0a558SYan, Zheng 459883d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 45993fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 46003fd0a558SYan, Zheng struct extent_buffer *cow) 46013fd0a558SYan, Zheng { 46020b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 46033fd0a558SYan, Zheng struct reloc_control *rc; 4604a26195a5SQu Wenruo struct btrfs_backref_node *node; 46053fd0a558SYan, Zheng int first_cow = 0; 46063fd0a558SYan, Zheng int level; 460783d4cfd4SJosef Bacik int ret = 0; 46083fd0a558SYan, Zheng 46090b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 46103fd0a558SYan, Zheng if (!rc) 461183d4cfd4SJosef Bacik return 0; 46123fd0a558SYan, Zheng 46133fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 46143fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 46153fd0a558SYan, Zheng 46163fd0a558SYan, Zheng level = btrfs_header_level(buf); 46173fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 46183fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 46193fd0a558SYan, Zheng first_cow = 1; 46203fd0a558SYan, Zheng 46213fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 46223fd0a558SYan, Zheng rc->create_reloc_tree) { 46233fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 46243fd0a558SYan, Zheng 46253fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 46263fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 46273fd0a558SYan, Zheng node->new_bytenr != buf->start); 46283fd0a558SYan, Zheng 46293fd0a558SYan, Zheng drop_node_buffer(node); 463067439dadSDavid Sterba atomic_inc(&cow->refs); 46313fd0a558SYan, Zheng node->eb = cow; 46323fd0a558SYan, Zheng node->new_bytenr = cow->start; 46333fd0a558SYan, Zheng 46343fd0a558SYan, Zheng if (!node->pending) { 46353fd0a558SYan, Zheng list_move_tail(&node->list, 46363fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 46373fd0a558SYan, Zheng node->pending = 1; 46383fd0a558SYan, Zheng } 46393fd0a558SYan, Zheng 46403fd0a558SYan, Zheng if (first_cow) 46419569cc20SQu Wenruo mark_block_processed(rc, node); 46423fd0a558SYan, Zheng 46433fd0a558SYan, Zheng if (first_cow && level > 0) 46443fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 46453fd0a558SYan, Zheng } 46463fd0a558SYan, Zheng 464783d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 46483fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 464983d4cfd4SJosef Bacik return ret; 46503fd0a558SYan, Zheng } 46513fd0a558SYan, Zheng 46523fd0a558SYan, Zheng /* 46533fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 465401327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 46553fd0a558SYan, Zheng */ 4656147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 46573fd0a558SYan, Zheng u64 *bytes_to_reserve) 46583fd0a558SYan, Zheng { 465910995c04SQu Wenruo struct btrfs_root *root = pending->root; 466010995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 46613fd0a558SYan, Zheng 46626282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 46633fd0a558SYan, Zheng return; 46643fd0a558SYan, Zheng 46653fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 46663fd0a558SYan, Zheng return; 46673fd0a558SYan, Zheng 46683fd0a558SYan, Zheng root = root->reloc_root; 46693fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 46703fd0a558SYan, Zheng /* 46713fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 46723fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 46733fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 46743fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 46753fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 46763fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 46773fd0a558SYan, Zheng * tree may use more space than it frees. so we need 46783fd0a558SYan, Zheng * reserve extra space. 46793fd0a558SYan, Zheng */ 46803fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 46813fd0a558SYan, Zheng } 46823fd0a558SYan, Zheng 46833fd0a558SYan, Zheng /* 46843fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 46853fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4686f44deb74SJosef Bacik * 4687f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4688f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4689f44deb74SJosef Bacik * rc->reloc_roots. 46903fd0a558SYan, Zheng */ 469149b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 46923fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 46933fd0a558SYan, Zheng { 46943fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 46953fd0a558SYan, Zheng struct btrfs_root *reloc_root; 46963fd0a558SYan, Zheng struct btrfs_root *new_root; 469710995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 46983fd0a558SYan, Zheng int ret; 46993fd0a558SYan, Zheng 47006282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 470149b25e05SJeff Mahoney return 0; 47023fd0a558SYan, Zheng 47033fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 47043fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 47053fd0a558SYan, Zheng 47063fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 47073fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 47083fd0a558SYan, Zheng rc->block_rsv, 47093a584174SLu Fengqi rc->nodes_relocated, true); 471049b25e05SJeff Mahoney if (ret) 471149b25e05SJeff Mahoney return ret; 47123fd0a558SYan, Zheng } 47133fd0a558SYan, Zheng 47143fd0a558SYan, Zheng new_root = pending->snap; 47153fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 47163fd0a558SYan, Zheng new_root->root_key.objectid); 471749b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 471849b25e05SJeff Mahoney return PTR_ERR(reloc_root); 47193fd0a558SYan, Zheng 4720ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4721ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4722f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 47233fd0a558SYan, Zheng 472449b25e05SJeff Mahoney if (rc->create_reloc_tree) 47253fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 472649b25e05SJeff Mahoney return ret; 47273fd0a558SYan, Zheng } 4728