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 18133a0f1f7SQu Wenruo static void backref_cache_init(struct btrfs_fs_info *fs_info, 182a26195a5SQu Wenruo struct btrfs_backref_cache *cache, int is_reloc) 1835d4f98a2SYan Zheng { 1845d4f98a2SYan Zheng int i; 1856bef4d31SEric Paris cache->rb_root = RB_ROOT; 1865d4f98a2SYan Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 1875d4f98a2SYan Zheng INIT_LIST_HEAD(&cache->pending[i]); 1883fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->changed); 1893fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->detached); 1903fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->leaves); 19184780289SQu Wenruo INIT_LIST_HEAD(&cache->pending_edge); 19284780289SQu Wenruo INIT_LIST_HEAD(&cache->useless_node); 19333a0f1f7SQu Wenruo cache->fs_info = fs_info; 1942433bea5SQu Wenruo cache->is_reloc = is_reloc; 1955d4f98a2SYan Zheng } 1965d4f98a2SYan Zheng 197a26195a5SQu Wenruo static void backref_cache_cleanup(struct btrfs_backref_cache *cache) 1985d4f98a2SYan Zheng { 199a26195a5SQu Wenruo struct btrfs_backref_node *node; 2003fd0a558SYan, Zheng int i; 2013fd0a558SYan, Zheng 2023fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 2033fd0a558SYan, Zheng node = list_entry(cache->detached.next, 204a26195a5SQu Wenruo struct btrfs_backref_node, list); 2053fd0a558SYan, Zheng remove_backref_node(cache, node); 2063fd0a558SYan, Zheng } 2073fd0a558SYan, Zheng 2083fd0a558SYan, Zheng while (!list_empty(&cache->leaves)) { 2093fd0a558SYan, Zheng node = list_entry(cache->leaves.next, 210a26195a5SQu Wenruo struct btrfs_backref_node, lower); 2113fd0a558SYan, Zheng remove_backref_node(cache, node); 2123fd0a558SYan, Zheng } 2133fd0a558SYan, Zheng 2143fd0a558SYan, Zheng cache->last_trans = 0; 2153fd0a558SYan, Zheng 2163fd0a558SYan, Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 217f4907095SLiu Bo ASSERT(list_empty(&cache->pending[i])); 21884780289SQu Wenruo ASSERT(list_empty(&cache->pending_edge)); 21984780289SQu Wenruo ASSERT(list_empty(&cache->useless_node)); 220f4907095SLiu Bo ASSERT(list_empty(&cache->changed)); 221f4907095SLiu Bo ASSERT(list_empty(&cache->detached)); 222f4907095SLiu Bo ASSERT(RB_EMPTY_ROOT(&cache->rb_root)); 223f4907095SLiu Bo ASSERT(!cache->nr_nodes); 224f4907095SLiu Bo ASSERT(!cache->nr_edges); 2253fd0a558SYan, Zheng } 2263fd0a558SYan, Zheng 227a26195a5SQu Wenruo static struct btrfs_backref_node *alloc_backref_node( 228a26195a5SQu Wenruo struct btrfs_backref_cache *cache, u64 bytenr, int level) 2293fd0a558SYan, Zheng { 230a26195a5SQu Wenruo struct btrfs_backref_node *node; 2313fd0a558SYan, Zheng 2320304f2d8SQu Wenruo ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL); 2333fd0a558SYan, Zheng node = kzalloc(sizeof(*node), GFP_NOFS); 2340304f2d8SQu Wenruo if (!node) 2350304f2d8SQu Wenruo return node; 2360304f2d8SQu Wenruo 2373fd0a558SYan, Zheng INIT_LIST_HEAD(&node->list); 2385d4f98a2SYan Zheng INIT_LIST_HEAD(&node->upper); 2395d4f98a2SYan Zheng INIT_LIST_HEAD(&node->lower); 2405d4f98a2SYan Zheng RB_CLEAR_NODE(&node->rb_node); 2413fd0a558SYan, Zheng cache->nr_nodes++; 2420304f2d8SQu Wenruo node->level = level; 2430304f2d8SQu Wenruo node->bytenr = bytenr; 2440304f2d8SQu Wenruo 2453fd0a558SYan, Zheng return node; 2463fd0a558SYan, Zheng } 2473fd0a558SYan, Zheng 248a26195a5SQu Wenruo static void free_backref_node(struct btrfs_backref_cache *cache, 249a26195a5SQu Wenruo struct btrfs_backref_node *node) 2503fd0a558SYan, Zheng { 2513fd0a558SYan, Zheng if (node) { 2523fd0a558SYan, Zheng cache->nr_nodes--; 25300246528SJosef Bacik btrfs_put_root(node->root); 2543fd0a558SYan, Zheng kfree(node); 2553fd0a558SYan, Zheng } 2563fd0a558SYan, Zheng } 2573fd0a558SYan, Zheng 258a26195a5SQu Wenruo static struct btrfs_backref_edge *alloc_backref_edge( 259a26195a5SQu Wenruo struct btrfs_backref_cache *cache) 2603fd0a558SYan, Zheng { 261a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2623fd0a558SYan, Zheng 2633fd0a558SYan, Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 2643fd0a558SYan, Zheng if (edge) 2653fd0a558SYan, Zheng cache->nr_edges++; 2663fd0a558SYan, Zheng return edge; 2673fd0a558SYan, Zheng } 2683fd0a558SYan, Zheng 2692a979612SQu Wenruo #define LINK_LOWER (1 << 0) 2702a979612SQu Wenruo #define LINK_UPPER (1 << 1) 271a26195a5SQu Wenruo static void link_backref_edge(struct btrfs_backref_edge *edge, 272a26195a5SQu Wenruo struct btrfs_backref_node *lower, 273a26195a5SQu Wenruo struct btrfs_backref_node *upper, 2742a979612SQu Wenruo int link_which) 2752a979612SQu Wenruo { 2762a979612SQu Wenruo ASSERT(upper && lower && upper->level == lower->level + 1); 2772a979612SQu Wenruo edge->node[LOWER] = lower; 2782a979612SQu Wenruo edge->node[UPPER] = upper; 2792a979612SQu Wenruo if (link_which & LINK_LOWER) 2802a979612SQu Wenruo list_add_tail(&edge->list[LOWER], &lower->upper); 2812a979612SQu Wenruo if (link_which & LINK_UPPER) 2822a979612SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 2832a979612SQu Wenruo } 2842a979612SQu Wenruo 285a26195a5SQu Wenruo static void free_backref_edge(struct btrfs_backref_cache *cache, 286a26195a5SQu Wenruo struct btrfs_backref_edge *edge) 2873fd0a558SYan, Zheng { 2883fd0a558SYan, Zheng if (edge) { 2893fd0a558SYan, Zheng cache->nr_edges--; 2903fd0a558SYan, Zheng kfree(edge); 2913fd0a558SYan, Zheng } 2925d4f98a2SYan Zheng } 2935d4f98a2SYan Zheng 29448a3b636SEric Sandeen static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr) 29543c04fb1SJeff Mahoney { 29643c04fb1SJeff Mahoney 29743c04fb1SJeff Mahoney struct btrfs_fs_info *fs_info = NULL; 298a26195a5SQu Wenruo struct btrfs_backref_node *bnode = rb_entry(rb_node, 299a26195a5SQu Wenruo struct btrfs_backref_node, rb_node); 30043c04fb1SJeff Mahoney if (bnode->root) 30143c04fb1SJeff Mahoney fs_info = bnode->root->fs_info; 3025d163e0eSJeff Mahoney btrfs_panic(fs_info, errno, 3035d163e0eSJeff Mahoney "Inconsistency in backref cache found at offset %llu", 3045d163e0eSJeff Mahoney bytenr); 30543c04fb1SJeff Mahoney } 30643c04fb1SJeff Mahoney 3075d4f98a2SYan Zheng /* 3085d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 3095d4f98a2SYan Zheng */ 310a26195a5SQu Wenruo static struct btrfs_backref_node *walk_up_backref( 311a26195a5SQu Wenruo struct btrfs_backref_node *node, 312a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 3135d4f98a2SYan Zheng { 314a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 3155d4f98a2SYan Zheng int idx = *index; 3165d4f98a2SYan Zheng 3175d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 3185d4f98a2SYan Zheng edge = list_entry(node->upper.next, 319a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 3205d4f98a2SYan Zheng edges[idx++] = edge; 3215d4f98a2SYan Zheng node = edge->node[UPPER]; 3225d4f98a2SYan Zheng } 3233fd0a558SYan, Zheng BUG_ON(node->detached); 3245d4f98a2SYan Zheng *index = idx; 3255d4f98a2SYan Zheng return node; 3265d4f98a2SYan Zheng } 3275d4f98a2SYan Zheng 3285d4f98a2SYan Zheng /* 3295d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 3305d4f98a2SYan Zheng */ 331a26195a5SQu Wenruo static struct btrfs_backref_node *walk_down_backref( 332a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 3335d4f98a2SYan Zheng { 334a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 335a26195a5SQu Wenruo struct btrfs_backref_node *lower; 3365d4f98a2SYan Zheng int idx = *index; 3375d4f98a2SYan Zheng 3385d4f98a2SYan Zheng while (idx > 0) { 3395d4f98a2SYan Zheng edge = edges[idx - 1]; 3405d4f98a2SYan Zheng lower = edge->node[LOWER]; 3415d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 3425d4f98a2SYan Zheng idx--; 3435d4f98a2SYan Zheng continue; 3445d4f98a2SYan Zheng } 3455d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 346a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 3475d4f98a2SYan Zheng edges[idx - 1] = edge; 3485d4f98a2SYan Zheng *index = idx; 3495d4f98a2SYan Zheng return edge->node[UPPER]; 3505d4f98a2SYan Zheng } 3515d4f98a2SYan Zheng *index = 0; 3525d4f98a2SYan Zheng return NULL; 3535d4f98a2SYan Zheng } 3545d4f98a2SYan Zheng 355a26195a5SQu Wenruo static void unlock_node_buffer(struct btrfs_backref_node *node) 3565d4f98a2SYan Zheng { 3575d4f98a2SYan Zheng if (node->locked) { 3585d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 3595d4f98a2SYan Zheng node->locked = 0; 3605d4f98a2SYan Zheng } 3613fd0a558SYan, Zheng } 3623fd0a558SYan, Zheng 363a26195a5SQu Wenruo static void drop_node_buffer(struct btrfs_backref_node *node) 3643fd0a558SYan, Zheng { 3653fd0a558SYan, Zheng if (node->eb) { 3663fd0a558SYan, Zheng unlock_node_buffer(node); 3675d4f98a2SYan Zheng free_extent_buffer(node->eb); 3685d4f98a2SYan Zheng node->eb = NULL; 3695d4f98a2SYan Zheng } 3705d4f98a2SYan Zheng } 3715d4f98a2SYan Zheng 372a26195a5SQu Wenruo static void drop_backref_node(struct btrfs_backref_cache *tree, 373a26195a5SQu Wenruo struct btrfs_backref_node *node) 3745d4f98a2SYan Zheng { 3755d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 3765d4f98a2SYan Zheng 3775d4f98a2SYan Zheng drop_node_buffer(node); 3783fd0a558SYan, Zheng list_del(&node->list); 3795d4f98a2SYan Zheng list_del(&node->lower); 3803fd0a558SYan, Zheng if (!RB_EMPTY_NODE(&node->rb_node)) 3815d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 3823fd0a558SYan, Zheng free_backref_node(tree, node); 3835d4f98a2SYan Zheng } 3845d4f98a2SYan Zheng 3855d4f98a2SYan Zheng /* 3865d4f98a2SYan Zheng * remove a backref node from the backref cache 3875d4f98a2SYan Zheng */ 388a26195a5SQu Wenruo static void remove_backref_node(struct btrfs_backref_cache *cache, 389a26195a5SQu Wenruo struct btrfs_backref_node *node) 3905d4f98a2SYan Zheng { 391a26195a5SQu Wenruo struct btrfs_backref_node *upper; 392a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 3935d4f98a2SYan Zheng 3945d4f98a2SYan Zheng if (!node) 3955d4f98a2SYan Zheng return; 3965d4f98a2SYan Zheng 3973fd0a558SYan, Zheng BUG_ON(!node->lowest && !node->detached); 3985d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 399a26195a5SQu Wenruo edge = list_entry(node->upper.next, struct btrfs_backref_edge, 4005d4f98a2SYan Zheng list[LOWER]); 4015d4f98a2SYan Zheng upper = edge->node[UPPER]; 4025d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 4035d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 4043fd0a558SYan, Zheng free_backref_edge(cache, edge); 4053fd0a558SYan, Zheng 4063fd0a558SYan, Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 4073fd0a558SYan, Zheng BUG_ON(!list_empty(&node->upper)); 4083fd0a558SYan, Zheng drop_backref_node(cache, node); 4093fd0a558SYan, Zheng node = upper; 4103fd0a558SYan, Zheng node->lowest = 1; 4113fd0a558SYan, Zheng continue; 4123fd0a558SYan, Zheng } 4135d4f98a2SYan Zheng /* 4143fd0a558SYan, Zheng * add the node to leaf node list if no other 4155d4f98a2SYan Zheng * child block cached. 4165d4f98a2SYan Zheng */ 4175d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 4183fd0a558SYan, Zheng list_add_tail(&upper->lower, &cache->leaves); 4195d4f98a2SYan Zheng upper->lowest = 1; 4205d4f98a2SYan Zheng } 4215d4f98a2SYan Zheng } 4223fd0a558SYan, Zheng 4235d4f98a2SYan Zheng drop_backref_node(cache, node); 4245d4f98a2SYan Zheng } 4255d4f98a2SYan Zheng 426a26195a5SQu Wenruo static void update_backref_node(struct btrfs_backref_cache *cache, 427a26195a5SQu Wenruo struct btrfs_backref_node *node, u64 bytenr) 4283fd0a558SYan, Zheng { 4293fd0a558SYan, Zheng struct rb_node *rb_node; 4303fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 4313fd0a558SYan, Zheng node->bytenr = bytenr; 432e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node); 43343c04fb1SJeff Mahoney if (rb_node) 43443c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, bytenr); 4353fd0a558SYan, Zheng } 4363fd0a558SYan, Zheng 4373fd0a558SYan, Zheng /* 4383fd0a558SYan, Zheng * update backref cache after a transaction commit 4393fd0a558SYan, Zheng */ 4403fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 441a26195a5SQu Wenruo struct btrfs_backref_cache *cache) 4423fd0a558SYan, Zheng { 443a26195a5SQu Wenruo struct btrfs_backref_node *node; 4443fd0a558SYan, Zheng int level = 0; 4453fd0a558SYan, Zheng 4463fd0a558SYan, Zheng if (cache->last_trans == 0) { 4473fd0a558SYan, Zheng cache->last_trans = trans->transid; 4483fd0a558SYan, Zheng return 0; 4493fd0a558SYan, Zheng } 4503fd0a558SYan, Zheng 4513fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 4523fd0a558SYan, Zheng return 0; 4533fd0a558SYan, Zheng 4543fd0a558SYan, Zheng /* 4553fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 4563fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 4573fd0a558SYan, Zheng * so the detached nodes are no longer useful. 4583fd0a558SYan, Zheng */ 4593fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 4603fd0a558SYan, Zheng node = list_entry(cache->detached.next, 461a26195a5SQu Wenruo struct btrfs_backref_node, list); 4623fd0a558SYan, Zheng remove_backref_node(cache, node); 4633fd0a558SYan, Zheng } 4643fd0a558SYan, Zheng 4653fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 4663fd0a558SYan, Zheng node = list_entry(cache->changed.next, 467a26195a5SQu Wenruo struct btrfs_backref_node, list); 4683fd0a558SYan, Zheng list_del_init(&node->list); 4693fd0a558SYan, Zheng BUG_ON(node->pending); 4703fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 4713fd0a558SYan, Zheng } 4723fd0a558SYan, Zheng 4733fd0a558SYan, Zheng /* 4743fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 4753fd0a558SYan, Zheng * errors during processing the pending nodes. 4763fd0a558SYan, Zheng */ 4773fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 4783fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 4793fd0a558SYan, Zheng BUG_ON(!node->pending); 4803fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 4813fd0a558SYan, Zheng continue; 4823fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 4833fd0a558SYan, Zheng } 4843fd0a558SYan, Zheng } 4853fd0a558SYan, Zheng 4863fd0a558SYan, Zheng cache->last_trans = 0; 4873fd0a558SYan, Zheng return 1; 4883fd0a558SYan, Zheng } 4893fd0a558SYan, Zheng 4906282675eSQu Wenruo static bool reloc_root_is_dead(struct btrfs_root *root) 4916282675eSQu Wenruo { 4926282675eSQu Wenruo /* 4936282675eSQu Wenruo * Pair with set_bit/clear_bit in clean_dirty_subvols and 4946282675eSQu Wenruo * btrfs_update_reloc_root. We need to see the updated bit before 4956282675eSQu Wenruo * trying to access reloc_root 4966282675eSQu Wenruo */ 4976282675eSQu Wenruo smp_rmb(); 4986282675eSQu Wenruo if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state)) 4996282675eSQu Wenruo return true; 5006282675eSQu Wenruo return false; 5016282675eSQu Wenruo } 5026282675eSQu Wenruo 5036282675eSQu Wenruo /* 5046282675eSQu Wenruo * Check if this subvolume tree has valid reloc tree. 5056282675eSQu Wenruo * 5066282675eSQu Wenruo * Reloc tree after swap is considered dead, thus not considered as valid. 5076282675eSQu Wenruo * This is enough for most callers, as they don't distinguish dead reloc root 5086282675eSQu Wenruo * from no reloc root. But should_ignore_root() below is a special case. 5096282675eSQu Wenruo */ 5106282675eSQu Wenruo static bool have_reloc_root(struct btrfs_root *root) 5116282675eSQu Wenruo { 5126282675eSQu Wenruo if (reloc_root_is_dead(root)) 5136282675eSQu Wenruo return false; 5146282675eSQu Wenruo if (!root->reloc_root) 5156282675eSQu Wenruo return false; 5166282675eSQu Wenruo return true; 5176282675eSQu Wenruo } 518f2a97a9dSDavid Sterba 5193fd0a558SYan, Zheng static int should_ignore_root(struct btrfs_root *root) 5203fd0a558SYan, Zheng { 5213fd0a558SYan, Zheng struct btrfs_root *reloc_root; 5223fd0a558SYan, Zheng 52327cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 5243fd0a558SYan, Zheng return 0; 5253fd0a558SYan, Zheng 5266282675eSQu Wenruo /* This root has been merged with its reloc tree, we can ignore it */ 5276282675eSQu Wenruo if (reloc_root_is_dead(root)) 5286282675eSQu Wenruo return 1; 5296282675eSQu Wenruo 5303fd0a558SYan, Zheng reloc_root = root->reloc_root; 5313fd0a558SYan, Zheng if (!reloc_root) 5323fd0a558SYan, Zheng return 0; 5333fd0a558SYan, Zheng 5344d4225fcSJosef Bacik if (btrfs_header_generation(reloc_root->commit_root) == 5354d4225fcSJosef Bacik root->fs_info->running_transaction->transid) 5363fd0a558SYan, Zheng return 0; 5373fd0a558SYan, Zheng /* 5383fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 5393fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 5403fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 5413fd0a558SYan, Zheng * relocation. 5423fd0a558SYan, Zheng */ 5433fd0a558SYan, Zheng return 1; 5443fd0a558SYan, Zheng } 5455d4f98a2SYan Zheng /* 5465d4f98a2SYan Zheng * find reloc tree by address of tree root 5475d4f98a2SYan Zheng */ 5482433bea5SQu Wenruo struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr) 5495d4f98a2SYan Zheng { 5502433bea5SQu Wenruo struct reloc_control *rc = fs_info->reloc_ctl; 5515d4f98a2SYan Zheng struct rb_node *rb_node; 5525d4f98a2SYan Zheng struct mapping_node *node; 5535d4f98a2SYan Zheng struct btrfs_root *root = NULL; 5545d4f98a2SYan Zheng 5552433bea5SQu Wenruo ASSERT(rc); 5565d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 557e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr); 5585d4f98a2SYan Zheng if (rb_node) { 5595d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 5605d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 5615d4f98a2SYan Zheng } 5625d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 56300246528SJosef Bacik return btrfs_grab_root(root); 5645d4f98a2SYan Zheng } 5655d4f98a2SYan Zheng 5665d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 5675d4f98a2SYan Zheng u64 root_objectid) 5685d4f98a2SYan Zheng { 5695d4f98a2SYan Zheng struct btrfs_key key; 5705d4f98a2SYan Zheng 5715d4f98a2SYan Zheng key.objectid = root_objectid; 5725d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 5735d4f98a2SYan Zheng key.offset = (u64)-1; 5745d4f98a2SYan Zheng 575bc44d7c4SJosef Bacik return btrfs_get_fs_root(fs_info, &key, false); 5765d4f98a2SYan Zheng } 5775d4f98a2SYan Zheng 5785d4f98a2SYan Zheng /* 5794007ea87SQu Wenruo * Handle direct tree backref 5804007ea87SQu Wenruo * 5814007ea87SQu Wenruo * Direct tree backref means, the backref item shows its parent bytenr 5824007ea87SQu Wenruo * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined). 5834007ea87SQu Wenruo * 5844007ea87SQu Wenruo * @ref_key: The converted backref key. 5854007ea87SQu Wenruo * For keyed backref, it's the item key. 5864007ea87SQu Wenruo * For inlined backref, objectid is the bytenr, 5874007ea87SQu Wenruo * type is btrfs_inline_ref_type, offset is 5884007ea87SQu Wenruo * btrfs_inline_ref_offset. 5894007ea87SQu Wenruo */ 590a26195a5SQu Wenruo static int handle_direct_tree_backref(struct btrfs_backref_cache *cache, 5914007ea87SQu Wenruo struct btrfs_key *ref_key, 592a26195a5SQu Wenruo struct btrfs_backref_node *cur) 5934007ea87SQu Wenruo { 594a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 595a26195a5SQu Wenruo struct btrfs_backref_node *upper; 5964007ea87SQu Wenruo struct rb_node *rb_node; 5974007ea87SQu Wenruo 5984007ea87SQu Wenruo ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY); 5994007ea87SQu Wenruo 6004007ea87SQu Wenruo /* Only reloc root uses backref pointing to itself */ 6014007ea87SQu Wenruo if (ref_key->objectid == ref_key->offset) { 6024007ea87SQu Wenruo struct btrfs_root *root; 6034007ea87SQu Wenruo 6044007ea87SQu Wenruo cur->is_reloc_root = 1; 6054007ea87SQu Wenruo /* Only reloc backref cache cares about a specific root */ 6064007ea87SQu Wenruo if (cache->is_reloc) { 6074007ea87SQu Wenruo root = find_reloc_root(cache->fs_info, cur->bytenr); 6084007ea87SQu Wenruo if (WARN_ON(!root)) 6094007ea87SQu Wenruo return -ENOENT; 6104007ea87SQu Wenruo cur->root = root; 6114007ea87SQu Wenruo } else { 6124007ea87SQu Wenruo /* 6134007ea87SQu Wenruo * For generic purpose backref cache, reloc root node 6144007ea87SQu Wenruo * is useless. 6154007ea87SQu Wenruo */ 6164007ea87SQu Wenruo list_add(&cur->list, &cache->useless_node); 6174007ea87SQu Wenruo } 6184007ea87SQu Wenruo return 0; 6194007ea87SQu Wenruo } 6204007ea87SQu Wenruo 6214007ea87SQu Wenruo edge = alloc_backref_edge(cache); 6224007ea87SQu Wenruo if (!edge) 6234007ea87SQu Wenruo return -ENOMEM; 6244007ea87SQu Wenruo 625e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, ref_key->offset); 6264007ea87SQu Wenruo if (!rb_node) { 6274007ea87SQu Wenruo /* Parent node not yet cached */ 6280304f2d8SQu Wenruo upper = alloc_backref_node(cache, ref_key->offset, 6290304f2d8SQu Wenruo cur->level + 1); 6304007ea87SQu Wenruo if (!upper) { 6314007ea87SQu Wenruo free_backref_edge(cache, edge); 6324007ea87SQu Wenruo return -ENOMEM; 6334007ea87SQu Wenruo } 6344007ea87SQu Wenruo 6354007ea87SQu Wenruo /* 6364007ea87SQu Wenruo * Backrefs for the upper level block isn't cached, add the 6374007ea87SQu Wenruo * block to pending list 6384007ea87SQu Wenruo */ 6394007ea87SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 6404007ea87SQu Wenruo } else { 6414007ea87SQu Wenruo /* Parent node already cached */ 642a26195a5SQu Wenruo upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node); 6434007ea87SQu Wenruo ASSERT(upper->checked); 6444007ea87SQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 6454007ea87SQu Wenruo } 6462a979612SQu Wenruo link_backref_edge(edge, cur, upper, LINK_LOWER); 6474007ea87SQu Wenruo return 0; 6484007ea87SQu Wenruo } 6494007ea87SQu Wenruo 6504007ea87SQu Wenruo /* 6514d81ea8bSQu Wenruo * Handle indirect tree backref 6524d81ea8bSQu Wenruo * 6534d81ea8bSQu Wenruo * Indirect tree backref means, we only know which tree the node belongs to. 6544d81ea8bSQu Wenruo * We still need to do a tree search to find out the parents. This is for 6554d81ea8bSQu Wenruo * TREE_BLOCK_REF backref (keyed or inlined). 6564d81ea8bSQu Wenruo * 6574d81ea8bSQu Wenruo * @ref_key: The same as @ref_key in handle_direct_tree_backref() 6584d81ea8bSQu Wenruo * @tree_key: The first key of this tree block. 6594d81ea8bSQu Wenruo * @path: A clean (released) path, to avoid allocating path everytime 6604d81ea8bSQu Wenruo * the function get called. 6614d81ea8bSQu Wenruo */ 662a26195a5SQu Wenruo static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache, 6634d81ea8bSQu Wenruo struct btrfs_path *path, 6644d81ea8bSQu Wenruo struct btrfs_key *ref_key, 6654d81ea8bSQu Wenruo struct btrfs_key *tree_key, 666a26195a5SQu Wenruo struct btrfs_backref_node *cur) 6674d81ea8bSQu Wenruo { 6684d81ea8bSQu Wenruo struct btrfs_fs_info *fs_info = cache->fs_info; 669a26195a5SQu Wenruo struct btrfs_backref_node *upper; 670a26195a5SQu Wenruo struct btrfs_backref_node *lower; 671a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 6724d81ea8bSQu Wenruo struct extent_buffer *eb; 6734d81ea8bSQu Wenruo struct btrfs_root *root; 6744d81ea8bSQu Wenruo struct rb_node *rb_node; 6754d81ea8bSQu Wenruo int level; 6764d81ea8bSQu Wenruo bool need_check = true; 6774d81ea8bSQu Wenruo int ret; 6784d81ea8bSQu Wenruo 6794d81ea8bSQu Wenruo root = read_fs_root(fs_info, ref_key->offset); 6804d81ea8bSQu Wenruo if (IS_ERR(root)) 6814d81ea8bSQu Wenruo return PTR_ERR(root); 6824d81ea8bSQu Wenruo if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 6834d81ea8bSQu Wenruo cur->cowonly = 1; 6844d81ea8bSQu Wenruo 6854d81ea8bSQu Wenruo if (btrfs_root_level(&root->root_item) == cur->level) { 6864d81ea8bSQu Wenruo /* Tree root */ 6874d81ea8bSQu Wenruo ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr); 6884d81ea8bSQu Wenruo if (should_ignore_root(root)) { 6894d81ea8bSQu Wenruo btrfs_put_root(root); 6904d81ea8bSQu Wenruo list_add(&cur->list, &cache->useless_node); 6914d81ea8bSQu Wenruo } else { 6924d81ea8bSQu Wenruo cur->root = root; 6934d81ea8bSQu Wenruo } 6944d81ea8bSQu Wenruo return 0; 6954d81ea8bSQu Wenruo } 6964d81ea8bSQu Wenruo 6974d81ea8bSQu Wenruo level = cur->level + 1; 6984d81ea8bSQu Wenruo 6994d81ea8bSQu Wenruo /* Search the tree to find parent blocks referring to the block */ 7004d81ea8bSQu Wenruo path->search_commit_root = 1; 7014d81ea8bSQu Wenruo path->skip_locking = 1; 7024d81ea8bSQu Wenruo path->lowest_level = level; 7034d81ea8bSQu Wenruo ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0); 7044d81ea8bSQu Wenruo path->lowest_level = 0; 7054d81ea8bSQu Wenruo if (ret < 0) { 7064d81ea8bSQu Wenruo btrfs_put_root(root); 7074d81ea8bSQu Wenruo return ret; 7084d81ea8bSQu Wenruo } 7094d81ea8bSQu Wenruo if (ret > 0 && path->slots[level] > 0) 7104d81ea8bSQu Wenruo path->slots[level]--; 7114d81ea8bSQu Wenruo 7124d81ea8bSQu Wenruo eb = path->nodes[level]; 7134d81ea8bSQu Wenruo if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) { 7144d81ea8bSQu Wenruo btrfs_err(fs_info, 7154d81ea8bSQu Wenruo "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)", 7164d81ea8bSQu Wenruo cur->bytenr, level - 1, root->root_key.objectid, 7174d81ea8bSQu Wenruo tree_key->objectid, tree_key->type, tree_key->offset); 7184d81ea8bSQu Wenruo btrfs_put_root(root); 7194d81ea8bSQu Wenruo ret = -ENOENT; 7204d81ea8bSQu Wenruo goto out; 7214d81ea8bSQu Wenruo } 7224d81ea8bSQu Wenruo lower = cur; 7234d81ea8bSQu Wenruo 7244d81ea8bSQu Wenruo /* Add all nodes and edges in the path */ 7254d81ea8bSQu Wenruo for (; level < BTRFS_MAX_LEVEL; level++) { 7264d81ea8bSQu Wenruo if (!path->nodes[level]) { 7274d81ea8bSQu Wenruo ASSERT(btrfs_root_bytenr(&root->root_item) == 7284d81ea8bSQu Wenruo lower->bytenr); 7294d81ea8bSQu Wenruo if (should_ignore_root(root)) { 7304d81ea8bSQu Wenruo btrfs_put_root(root); 7314d81ea8bSQu Wenruo list_add(&lower->list, &cache->useless_node); 7324d81ea8bSQu Wenruo } else { 7334d81ea8bSQu Wenruo lower->root = root; 7344d81ea8bSQu Wenruo } 7354d81ea8bSQu Wenruo break; 7364d81ea8bSQu Wenruo } 7374d81ea8bSQu Wenruo 7384d81ea8bSQu Wenruo edge = alloc_backref_edge(cache); 7394d81ea8bSQu Wenruo if (!edge) { 7404d81ea8bSQu Wenruo btrfs_put_root(root); 7414d81ea8bSQu Wenruo ret = -ENOMEM; 7424d81ea8bSQu Wenruo goto out; 7434d81ea8bSQu Wenruo } 7444d81ea8bSQu Wenruo 7454d81ea8bSQu Wenruo eb = path->nodes[level]; 746e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, eb->start); 7474d81ea8bSQu Wenruo if (!rb_node) { 7480304f2d8SQu Wenruo upper = alloc_backref_node(cache, eb->start, 7490304f2d8SQu Wenruo lower->level + 1); 7504d81ea8bSQu Wenruo if (!upper) { 7514d81ea8bSQu Wenruo btrfs_put_root(root); 7524d81ea8bSQu Wenruo free_backref_edge(cache, edge); 7534d81ea8bSQu Wenruo ret = -ENOMEM; 7544d81ea8bSQu Wenruo goto out; 7554d81ea8bSQu Wenruo } 7564d81ea8bSQu Wenruo upper->owner = btrfs_header_owner(eb); 7574d81ea8bSQu Wenruo if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 7584d81ea8bSQu Wenruo upper->cowonly = 1; 7594d81ea8bSQu Wenruo 7604d81ea8bSQu Wenruo /* 7614d81ea8bSQu Wenruo * If we know the block isn't shared we can avoid 7624d81ea8bSQu Wenruo * checking its backrefs. 7634d81ea8bSQu Wenruo */ 7644d81ea8bSQu Wenruo if (btrfs_block_can_be_shared(root, eb)) 7654d81ea8bSQu Wenruo upper->checked = 0; 7664d81ea8bSQu Wenruo else 7674d81ea8bSQu Wenruo upper->checked = 1; 7684d81ea8bSQu Wenruo 7694d81ea8bSQu Wenruo /* 7704d81ea8bSQu Wenruo * Add the block to pending list if we need to check its 7714d81ea8bSQu Wenruo * backrefs, we only do this once while walking up a 7724d81ea8bSQu Wenruo * tree as we will catch anything else later on. 7734d81ea8bSQu Wenruo */ 7744d81ea8bSQu Wenruo if (!upper->checked && need_check) { 7754d81ea8bSQu Wenruo need_check = false; 7764d81ea8bSQu Wenruo list_add_tail(&edge->list[UPPER], 7774d81ea8bSQu Wenruo &cache->pending_edge); 7784d81ea8bSQu Wenruo } else { 7794d81ea8bSQu Wenruo if (upper->checked) 7804d81ea8bSQu Wenruo need_check = true; 7814d81ea8bSQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 7824d81ea8bSQu Wenruo } 7834d81ea8bSQu Wenruo } else { 784a26195a5SQu Wenruo upper = rb_entry(rb_node, struct btrfs_backref_node, 785a26195a5SQu Wenruo rb_node); 7864d81ea8bSQu Wenruo ASSERT(upper->checked); 7874d81ea8bSQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 7884d81ea8bSQu Wenruo if (!upper->owner) 7894d81ea8bSQu Wenruo upper->owner = btrfs_header_owner(eb); 7904d81ea8bSQu Wenruo } 7912a979612SQu Wenruo link_backref_edge(edge, lower, upper, LINK_LOWER); 7924d81ea8bSQu Wenruo 7934d81ea8bSQu Wenruo if (rb_node) { 7944d81ea8bSQu Wenruo btrfs_put_root(root); 7954d81ea8bSQu Wenruo break; 7964d81ea8bSQu Wenruo } 7974d81ea8bSQu Wenruo lower = upper; 7984d81ea8bSQu Wenruo upper = NULL; 7994d81ea8bSQu Wenruo } 8004d81ea8bSQu Wenruo out: 8014d81ea8bSQu Wenruo btrfs_release_path(path); 8024d81ea8bSQu Wenruo return ret; 8034d81ea8bSQu Wenruo } 8044d81ea8bSQu Wenruo 805a26195a5SQu Wenruo static int handle_one_tree_block(struct btrfs_backref_cache *cache, 806e7d571c7SQu Wenruo struct btrfs_path *path, 807e7d571c7SQu Wenruo struct btrfs_backref_iter *iter, 8085d4f98a2SYan Zheng struct btrfs_key *node_key, 809a26195a5SQu Wenruo struct btrfs_backref_node *cur) 8105d4f98a2SYan Zheng { 811e7d571c7SQu Wenruo struct btrfs_fs_info *fs_info = cache->fs_info; 812a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 813a26195a5SQu Wenruo struct btrfs_backref_node *exist; 8145d4f98a2SYan Zheng int ret; 8155d4f98a2SYan Zheng 81671f572a9SQu Wenruo ret = btrfs_backref_iter_start(iter, cur->bytenr); 817e7d571c7SQu Wenruo if (ret < 0) 818e7d571c7SQu Wenruo return ret; 81971f572a9SQu Wenruo /* 82071f572a9SQu Wenruo * We skip the first btrfs_tree_block_info, as we don't use the key 82171f572a9SQu Wenruo * stored in it, but fetch it from the tree block 82271f572a9SQu Wenruo */ 82371f572a9SQu Wenruo if (btrfs_backref_has_tree_block_info(iter)) { 82471f572a9SQu Wenruo ret = btrfs_backref_iter_next(iter); 825e7d571c7SQu Wenruo if (ret < 0) 82671f572a9SQu Wenruo goto out; 82771f572a9SQu Wenruo /* No extra backref? This means the tree block is corrupted */ 82871f572a9SQu Wenruo if (ret > 0) { 829e7d571c7SQu Wenruo ret = -EUCLEAN; 83071f572a9SQu Wenruo goto out; 83171f572a9SQu Wenruo } 83271f572a9SQu Wenruo } 8335d4f98a2SYan Zheng WARN_ON(cur->checked); 8345d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 8355d4f98a2SYan Zheng /* 83670f23fd6SJustin P. Mattock * the backref was added previously when processing 8375d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 8385d4f98a2SYan Zheng */ 83975bfb9afSJosef Bacik ASSERT(list_is_singular(&cur->upper)); 840a26195a5SQu Wenruo edge = list_entry(cur->upper.next, struct btrfs_backref_edge, 8415d4f98a2SYan Zheng list[LOWER]); 84275bfb9afSJosef Bacik ASSERT(list_empty(&edge->list[UPPER])); 8435d4f98a2SYan Zheng exist = edge->node[UPPER]; 8445d4f98a2SYan Zheng /* 8455d4f98a2SYan Zheng * add the upper level block to pending list if we need 8465d4f98a2SYan Zheng * check its backrefs 8475d4f98a2SYan Zheng */ 8485d4f98a2SYan Zheng if (!exist->checked) 84984780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 8505d4f98a2SYan Zheng } else { 8515d4f98a2SYan Zheng exist = NULL; 8525d4f98a2SYan Zheng } 8535d4f98a2SYan Zheng 85471f572a9SQu Wenruo for (; ret == 0; ret = btrfs_backref_iter_next(iter)) { 85571f572a9SQu Wenruo struct extent_buffer *eb; 85671f572a9SQu Wenruo struct btrfs_key key; 8573de28d57SLiu Bo int type; 85871f572a9SQu Wenruo 85971f572a9SQu Wenruo cond_resched(); 86071f572a9SQu Wenruo eb = btrfs_backref_get_eb(iter); 86171f572a9SQu Wenruo 86271f572a9SQu Wenruo key.objectid = iter->bytenr; 86371f572a9SQu Wenruo if (btrfs_backref_iter_is_inline_ref(iter)) { 86471f572a9SQu Wenruo struct btrfs_extent_inline_ref *iref; 86571f572a9SQu Wenruo 86671f572a9SQu Wenruo /* update key for inline back ref */ 86771f572a9SQu Wenruo iref = (struct btrfs_extent_inline_ref *) 86871f572a9SQu Wenruo ((unsigned long)iter->cur_ptr); 8693de28d57SLiu Bo type = btrfs_get_extent_inline_ref_type(eb, iref, 8703de28d57SLiu Bo BTRFS_REF_TYPE_BLOCK); 8713de28d57SLiu Bo if (type == BTRFS_REF_TYPE_INVALID) { 872e7d571c7SQu Wenruo ret = -EUCLEAN; 8733de28d57SLiu Bo goto out; 8743de28d57SLiu Bo } 8753de28d57SLiu Bo key.type = type; 8765d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 87771f572a9SQu Wenruo } else { 87871f572a9SQu Wenruo key.type = iter->cur_key.type; 87971f572a9SQu Wenruo key.offset = iter->cur_key.offset; 8805d4f98a2SYan Zheng } 8815d4f98a2SYan Zheng 882fa6ac715SQu Wenruo /* 883fa6ac715SQu Wenruo * Parent node found and matches current inline ref, no need to 884fa6ac715SQu Wenruo * rebuild this node for this inline ref. 885fa6ac715SQu Wenruo */ 8865d4f98a2SYan Zheng if (exist && 8875d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 8885d4f98a2SYan Zheng exist->owner == key.offset) || 8895d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 8905d4f98a2SYan Zheng exist->bytenr == key.offset))) { 8915d4f98a2SYan Zheng exist = NULL; 89271f572a9SQu Wenruo continue; 8935d4f98a2SYan Zheng } 8945d4f98a2SYan Zheng 895fa6ac715SQu Wenruo /* SHARED_BLOCK_REF means key.offset is the parent bytenr */ 8965d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 8974007ea87SQu Wenruo ret = handle_direct_tree_backref(cache, &key, cur); 898e7d571c7SQu Wenruo if (ret < 0) 8992433bea5SQu Wenruo goto out; 90071f572a9SQu Wenruo continue; 9016d8ff4e4SDavid Sterba } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) { 902e7d571c7SQu Wenruo ret = -EINVAL; 903e7d571c7SQu Wenruo btrfs_print_v0_err(fs_info); 904e7d571c7SQu Wenruo btrfs_handle_fs_error(fs_info, ret, NULL); 905ba3c2b19SNikolay Borisov goto out; 9065d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 90771f572a9SQu Wenruo continue; 9085d4f98a2SYan Zheng } 9095d4f98a2SYan Zheng 910fa6ac715SQu Wenruo /* 911fa6ac715SQu Wenruo * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset 912fa6ac715SQu Wenruo * means the root objectid. We need to search the tree to get 913fa6ac715SQu Wenruo * its parent bytenr. 914fa6ac715SQu Wenruo */ 9154d81ea8bSQu Wenruo ret = handle_indirect_tree_backref(cache, path, &key, node_key, 9164d81ea8bSQu Wenruo cur); 917e7d571c7SQu Wenruo if (ret < 0) 91871f572a9SQu Wenruo goto out; 9195d4f98a2SYan Zheng } 92071f572a9SQu Wenruo ret = 0; 9215d4f98a2SYan Zheng cur->checked = 1; 9225d4f98a2SYan Zheng WARN_ON(exist); 923e7d571c7SQu Wenruo out: 924e7d571c7SQu Wenruo btrfs_backref_iter_release(iter); 925e7d571c7SQu Wenruo return ret; 926e7d571c7SQu Wenruo } 9275d4f98a2SYan Zheng 928e7d571c7SQu Wenruo /* 9291f872924SQu Wenruo * In handle_one_tree_backref(), we have only linked the lower node to the edge, 9301f872924SQu Wenruo * but the upper node hasn't been linked to the edge. 931a26195a5SQu Wenruo * This means we can only iterate through btrfs_backref_node::upper to reach 932a26195a5SQu Wenruo * parent edges, but not through btrfs_backref_node::lower to reach children 933a26195a5SQu Wenruo * edges. 9341f872924SQu Wenruo * 935a26195a5SQu Wenruo * This function will finish the btrfs_backref_node::lower to related edges, 936a26195a5SQu Wenruo * so that backref cache can be bi-directionally iterated. 9371f872924SQu Wenruo * 9381f872924SQu Wenruo * Also, this will add the nodes to backref cache for the next run. 9391f872924SQu Wenruo */ 940a26195a5SQu Wenruo static int finish_upper_links(struct btrfs_backref_cache *cache, 941a26195a5SQu Wenruo struct btrfs_backref_node *start) 9421f872924SQu Wenruo { 9431f872924SQu Wenruo struct list_head *useless_node = &cache->useless_node; 944a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 9451f872924SQu Wenruo struct rb_node *rb_node; 9461f872924SQu Wenruo LIST_HEAD(pending_edge); 9471f872924SQu Wenruo 9481f872924SQu Wenruo ASSERT(start->checked); 9491f872924SQu Wenruo 9501f872924SQu Wenruo /* Insert this node to cache if it's not COW-only */ 9511f872924SQu Wenruo if (!start->cowonly) { 952e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, start->bytenr, 9531f872924SQu Wenruo &start->rb_node); 9541f872924SQu Wenruo if (rb_node) 9551f872924SQu Wenruo backref_tree_panic(rb_node, -EEXIST, start->bytenr); 9561f872924SQu Wenruo list_add_tail(&start->lower, &cache->leaves); 9571f872924SQu Wenruo } 9581f872924SQu Wenruo 9591f872924SQu Wenruo /* 9601f872924SQu Wenruo * Use breadth first search to iterate all related edges. 9611f872924SQu Wenruo * 9621f872924SQu Wenruo * The starting points are all the edges of this node 9631f872924SQu Wenruo */ 9641f872924SQu Wenruo list_for_each_entry(edge, &start->upper, list[LOWER]) 9651f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &pending_edge); 9661f872924SQu Wenruo 9671f872924SQu Wenruo while (!list_empty(&pending_edge)) { 968a26195a5SQu Wenruo struct btrfs_backref_node *upper; 969a26195a5SQu Wenruo struct btrfs_backref_node *lower; 9701f872924SQu Wenruo struct rb_node *rb_node; 9711f872924SQu Wenruo 972a26195a5SQu Wenruo edge = list_first_entry(&pending_edge, 973a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 9741f872924SQu Wenruo list_del_init(&edge->list[UPPER]); 9751f872924SQu Wenruo upper = edge->node[UPPER]; 9761f872924SQu Wenruo lower = edge->node[LOWER]; 9771f872924SQu Wenruo 9781f872924SQu Wenruo /* Parent is detached, no need to keep any edges */ 9791f872924SQu Wenruo if (upper->detached) { 9801f872924SQu Wenruo list_del(&edge->list[LOWER]); 9811f872924SQu Wenruo free_backref_edge(cache, edge); 9821f872924SQu Wenruo 9831f872924SQu Wenruo /* Lower node is orphan, queue for cleanup */ 9841f872924SQu Wenruo if (list_empty(&lower->upper)) 9851f872924SQu Wenruo list_add(&lower->list, useless_node); 9861f872924SQu Wenruo continue; 9871f872924SQu Wenruo } 9881f872924SQu Wenruo 9891f872924SQu Wenruo /* 9901f872924SQu Wenruo * All new nodes added in current build_backref_tree() haven't 9911f872924SQu Wenruo * been linked to the cache rb tree. 9921f872924SQu Wenruo * So if we have upper->rb_node populated, this means a cache 9931f872924SQu Wenruo * hit. We only need to link the edge, as @upper and all its 9941f872924SQu Wenruo * parent have already been linked. 9951f872924SQu Wenruo */ 9961f872924SQu Wenruo if (!RB_EMPTY_NODE(&upper->rb_node)) { 9971f872924SQu Wenruo if (upper->lowest) { 9981f872924SQu Wenruo list_del_init(&upper->lower); 9991f872924SQu Wenruo upper->lowest = 0; 10001f872924SQu Wenruo } 10011f872924SQu Wenruo 10021f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 10031f872924SQu Wenruo continue; 10041f872924SQu Wenruo } 10051f872924SQu Wenruo 10061f872924SQu Wenruo /* Sanity check, we shouldn't have any unchecked nodes */ 10071f872924SQu Wenruo if (!upper->checked) { 10081f872924SQu Wenruo ASSERT(0); 10091f872924SQu Wenruo return -EUCLEAN; 10101f872924SQu Wenruo } 10111f872924SQu Wenruo 10121f872924SQu Wenruo /* Sanity check, COW-only node has non-COW-only parent */ 10131f872924SQu Wenruo if (start->cowonly != upper->cowonly) { 10141f872924SQu Wenruo ASSERT(0); 10151f872924SQu Wenruo return -EUCLEAN; 10161f872924SQu Wenruo } 10171f872924SQu Wenruo 10181f872924SQu Wenruo /* Only cache non-COW-only (subvolume trees) tree blocks */ 10191f872924SQu Wenruo if (!upper->cowonly) { 1020e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr, 10211f872924SQu Wenruo &upper->rb_node); 10221f872924SQu Wenruo if (rb_node) { 10231f872924SQu Wenruo backref_tree_panic(rb_node, -EEXIST, 10241f872924SQu Wenruo upper->bytenr); 10251f872924SQu Wenruo return -EUCLEAN; 10261f872924SQu Wenruo } 10271f872924SQu Wenruo } 10281f872924SQu Wenruo 10291f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 10301f872924SQu Wenruo 10311f872924SQu Wenruo /* 10321f872924SQu Wenruo * Also queue all the parent edges of this uncached node to 10331f872924SQu Wenruo * finish the upper linkage 10341f872924SQu Wenruo */ 10351f872924SQu Wenruo list_for_each_entry(edge, &upper->upper, list[LOWER]) 10361f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &pending_edge); 10371f872924SQu Wenruo } 10381f872924SQu Wenruo return 0; 10391f872924SQu Wenruo } 10401f872924SQu Wenruo 10411f872924SQu Wenruo /* 104229db137bSQu Wenruo * For useless nodes, do two major clean ups: 104329db137bSQu Wenruo * 104429db137bSQu Wenruo * - Cleanup the children edges and nodes 104529db137bSQu Wenruo * If child node is also orphan (no parent) during cleanup, then the child 104629db137bSQu Wenruo * node will also be cleaned up. 104729db137bSQu Wenruo * 104829db137bSQu Wenruo * - Freeing up leaves (level 0), keeps nodes detached 104929db137bSQu Wenruo * For nodes, the node is still cached as "detached" 105029db137bSQu Wenruo * 105129db137bSQu Wenruo * Return false if @node is not in the @useless_nodes list. 105229db137bSQu Wenruo * Return true if @node is in the @useless_nodes list. 105329db137bSQu Wenruo */ 105429db137bSQu Wenruo static bool handle_useless_nodes(struct reloc_control *rc, 1055a26195a5SQu Wenruo struct btrfs_backref_node *node) 105629db137bSQu Wenruo { 1057a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 105829db137bSQu Wenruo struct list_head *useless_node = &cache->useless_node; 105929db137bSQu Wenruo bool ret = false; 106029db137bSQu Wenruo 106129db137bSQu Wenruo while (!list_empty(useless_node)) { 1062a26195a5SQu Wenruo struct btrfs_backref_node *cur; 106329db137bSQu Wenruo 1064a26195a5SQu Wenruo cur = list_first_entry(useless_node, struct btrfs_backref_node, 106529db137bSQu Wenruo list); 106629db137bSQu Wenruo list_del_init(&cur->list); 106729db137bSQu Wenruo 106829db137bSQu Wenruo /* Only tree root nodes can be added to @useless_nodes */ 106929db137bSQu Wenruo ASSERT(list_empty(&cur->upper)); 107029db137bSQu Wenruo 107129db137bSQu Wenruo if (cur == node) 107229db137bSQu Wenruo ret = true; 107329db137bSQu Wenruo 107429db137bSQu Wenruo /* The node is the lowest node */ 107529db137bSQu Wenruo if (cur->lowest) { 107629db137bSQu Wenruo list_del_init(&cur->lower); 107729db137bSQu Wenruo cur->lowest = 0; 107829db137bSQu Wenruo } 107929db137bSQu Wenruo 108029db137bSQu Wenruo /* Cleanup the lower edges */ 108129db137bSQu Wenruo while (!list_empty(&cur->lower)) { 1082a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1083a26195a5SQu Wenruo struct btrfs_backref_node *lower; 108429db137bSQu Wenruo 108529db137bSQu Wenruo edge = list_entry(cur->lower.next, 1086a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 108729db137bSQu Wenruo list_del(&edge->list[UPPER]); 108829db137bSQu Wenruo list_del(&edge->list[LOWER]); 108929db137bSQu Wenruo lower = edge->node[LOWER]; 109029db137bSQu Wenruo free_backref_edge(cache, edge); 109129db137bSQu Wenruo 109229db137bSQu Wenruo /* Child node is also orphan, queue for cleanup */ 109329db137bSQu Wenruo if (list_empty(&lower->upper)) 109429db137bSQu Wenruo list_add(&lower->list, useless_node); 109529db137bSQu Wenruo } 109629db137bSQu Wenruo /* Mark this block processed for relocation */ 109729db137bSQu Wenruo mark_block_processed(rc, cur); 109829db137bSQu Wenruo 109929db137bSQu Wenruo /* 110029db137bSQu Wenruo * Backref nodes for tree leaves are deleted from the cache. 110129db137bSQu Wenruo * Backref nodes for upper level tree blocks are left in the 110229db137bSQu Wenruo * cache to avoid unnecessary backref lookup. 110329db137bSQu Wenruo */ 110429db137bSQu Wenruo if (cur->level > 0) { 110529db137bSQu Wenruo list_add(&cur->list, &cache->detached); 110629db137bSQu Wenruo cur->detached = 1; 110729db137bSQu Wenruo } else { 110829db137bSQu Wenruo rb_erase(&cur->rb_node, &cache->rb_root); 110929db137bSQu Wenruo free_backref_node(cache, cur); 111029db137bSQu Wenruo } 111129db137bSQu Wenruo } 111229db137bSQu Wenruo return ret; 111329db137bSQu Wenruo } 111429db137bSQu Wenruo 111529db137bSQu Wenruo /* 1116e7d571c7SQu Wenruo * Build backref tree for a given tree block. Root of the backref tree 1117e7d571c7SQu Wenruo * corresponds the tree block, leaves of the backref tree correspond roots of 1118e7d571c7SQu Wenruo * b-trees that reference the tree block. 1119e7d571c7SQu Wenruo * 1120e7d571c7SQu Wenruo * The basic idea of this function is check backrefs of a given block to find 1121e7d571c7SQu Wenruo * upper level blocks that reference the block, and then check backrefs of 1122e7d571c7SQu Wenruo * these upper level blocks recursively. The recursion stops when tree root is 1123e7d571c7SQu Wenruo * reached or backrefs for the block is cached. 1124e7d571c7SQu Wenruo * 1125e7d571c7SQu Wenruo * NOTE: if we find that backrefs for a block are cached, we know backrefs for 1126e7d571c7SQu Wenruo * all upper level blocks that directly/indirectly reference the block are also 1127e7d571c7SQu Wenruo * cached. 1128e7d571c7SQu Wenruo */ 1129a26195a5SQu Wenruo static noinline_for_stack struct btrfs_backref_node *build_backref_tree( 1130e7d571c7SQu Wenruo struct reloc_control *rc, struct btrfs_key *node_key, 1131e7d571c7SQu Wenruo int level, u64 bytenr) 1132e7d571c7SQu Wenruo { 1133e7d571c7SQu Wenruo struct btrfs_backref_iter *iter; 1134a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 1135e7d571c7SQu Wenruo /* For searching parent of TREE_BLOCK_REF */ 1136e7d571c7SQu Wenruo struct btrfs_path *path; 1137a26195a5SQu Wenruo struct btrfs_backref_node *cur; 1138a26195a5SQu Wenruo struct btrfs_backref_node *upper; 1139a26195a5SQu Wenruo struct btrfs_backref_node *lower; 1140a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 1141a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1142e7d571c7SQu Wenruo int ret; 1143e7d571c7SQu Wenruo int err = 0; 1144e7d571c7SQu Wenruo 1145e7d571c7SQu Wenruo iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS); 1146e7d571c7SQu Wenruo if (!iter) 1147e7d571c7SQu Wenruo return ERR_PTR(-ENOMEM); 1148e7d571c7SQu Wenruo path = btrfs_alloc_path(); 1149e7d571c7SQu Wenruo if (!path) { 1150e7d571c7SQu Wenruo err = -ENOMEM; 1151e7d571c7SQu Wenruo goto out; 1152e7d571c7SQu Wenruo } 1153e7d571c7SQu Wenruo 1154e7d571c7SQu Wenruo node = alloc_backref_node(cache, bytenr, level); 1155e7d571c7SQu Wenruo if (!node) { 1156e7d571c7SQu Wenruo err = -ENOMEM; 1157e7d571c7SQu Wenruo goto out; 1158e7d571c7SQu Wenruo } 1159e7d571c7SQu Wenruo 1160e7d571c7SQu Wenruo node->lowest = 1; 1161e7d571c7SQu Wenruo cur = node; 1162e7d571c7SQu Wenruo 1163e7d571c7SQu Wenruo /* Breadth-first search to build backref cache */ 1164e7d571c7SQu Wenruo do { 1165e7d571c7SQu Wenruo ret = handle_one_tree_block(cache, path, iter, node_key, cur); 1166e7d571c7SQu Wenruo if (ret < 0) { 1167e7d571c7SQu Wenruo err = ret; 1168e7d571c7SQu Wenruo goto out; 1169e7d571c7SQu Wenruo } 1170e7d571c7SQu Wenruo edge = list_first_entry_or_null(&cache->pending_edge, 1171a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 1172e7d571c7SQu Wenruo /* 1173e7d571c7SQu Wenruo * The pending list isn't empty, take the first block to 1174e7d571c7SQu Wenruo * process 1175e7d571c7SQu Wenruo */ 1176e7d571c7SQu Wenruo if (edge) { 11775d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 11785d4f98a2SYan Zheng cur = edge->node[UPPER]; 11795d4f98a2SYan Zheng } 1180e7d571c7SQu Wenruo } while (edge); 11815d4f98a2SYan Zheng 11821f872924SQu Wenruo /* Finish the upper linkage of newly added edges/nodes */ 11831f872924SQu Wenruo ret = finish_upper_links(cache, node); 11841f872924SQu Wenruo if (ret < 0) { 11851f872924SQu Wenruo err = ret; 118675bfb9afSJosef Bacik goto out; 118775bfb9afSJosef Bacik } 118875bfb9afSJosef Bacik 118929db137bSQu Wenruo if (handle_useless_nodes(rc, node)) 11903fd0a558SYan, Zheng node = NULL; 11915d4f98a2SYan Zheng out: 119271f572a9SQu Wenruo btrfs_backref_iter_free(iter); 119371f572a9SQu Wenruo btrfs_free_path(path); 11945d4f98a2SYan Zheng if (err) { 119584780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 119684780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 1197a26195a5SQu Wenruo struct btrfs_backref_node, list); 119875bfb9afSJosef Bacik list_del_init(&lower->list); 11993fd0a558SYan, Zheng } 120084780289SQu Wenruo while (!list_empty(&cache->pending_edge)) { 120184780289SQu Wenruo edge = list_first_entry(&cache->pending_edge, 1202a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 120375bfb9afSJosef Bacik list_del(&edge->list[UPPER]); 12043fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 120575bfb9afSJosef Bacik lower = edge->node[LOWER]; 12065d4f98a2SYan Zheng upper = edge->node[UPPER]; 12073fd0a558SYan, Zheng free_backref_edge(cache, edge); 120875bfb9afSJosef Bacik 120975bfb9afSJosef Bacik /* 121075bfb9afSJosef Bacik * Lower is no longer linked to any upper backref nodes 121175bfb9afSJosef Bacik * and isn't in the cache, we can free it ourselves. 121275bfb9afSJosef Bacik */ 121375bfb9afSJosef Bacik if (list_empty(&lower->upper) && 121475bfb9afSJosef Bacik RB_EMPTY_NODE(&lower->rb_node)) 121584780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 121675bfb9afSJosef Bacik 121775bfb9afSJosef Bacik if (!RB_EMPTY_NODE(&upper->rb_node)) 121875bfb9afSJosef Bacik continue; 121975bfb9afSJosef Bacik 122001327610SNicholas D Steeves /* Add this guy's upper edges to the list to process */ 122175bfb9afSJosef Bacik list_for_each_entry(edge, &upper->upper, list[LOWER]) 122284780289SQu Wenruo list_add_tail(&edge->list[UPPER], 122384780289SQu Wenruo &cache->pending_edge); 122475bfb9afSJosef Bacik if (list_empty(&upper->upper)) 122584780289SQu Wenruo list_add(&upper->list, &cache->useless_node); 122675bfb9afSJosef Bacik } 122775bfb9afSJosef Bacik 122884780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 122984780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 1230a26195a5SQu Wenruo struct btrfs_backref_node, list); 123175bfb9afSJosef Bacik list_del_init(&lower->list); 12320fd8c3daSLiu Bo if (lower == node) 12330fd8c3daSLiu Bo node = NULL; 123475bfb9afSJosef Bacik free_backref_node(cache, lower); 12355d4f98a2SYan Zheng } 12360fd8c3daSLiu Bo 12378e19c973SJosef Bacik remove_backref_node(cache, node); 123884780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 123984780289SQu Wenruo list_empty(&cache->pending_edge)); 12405d4f98a2SYan Zheng return ERR_PTR(err); 12415d4f98a2SYan Zheng } 124275bfb9afSJosef Bacik ASSERT(!node || !node->detached); 124384780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 124484780289SQu Wenruo list_empty(&cache->pending_edge)); 12455d4f98a2SYan Zheng return node; 12465d4f98a2SYan Zheng } 12475d4f98a2SYan Zheng 12485d4f98a2SYan Zheng /* 12493fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 12503fd0a558SYan, Zheng * the backref node is created by cloning backref node that 12513fd0a558SYan, Zheng * corresponds to root of source tree 12523fd0a558SYan, Zheng */ 12533fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 12543fd0a558SYan, Zheng struct reloc_control *rc, 12553fd0a558SYan, Zheng struct btrfs_root *src, 12563fd0a558SYan, Zheng struct btrfs_root *dest) 12573fd0a558SYan, Zheng { 12583fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 1259a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 1260a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 1261a26195a5SQu Wenruo struct btrfs_backref_node *new_node; 1262a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1263a26195a5SQu Wenruo struct btrfs_backref_edge *new_edge; 12643fd0a558SYan, Zheng struct rb_node *rb_node; 12653fd0a558SYan, Zheng 12663fd0a558SYan, Zheng if (cache->last_trans > 0) 12673fd0a558SYan, Zheng update_backref_cache(trans, cache); 12683fd0a558SYan, Zheng 1269e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start); 12703fd0a558SYan, Zheng if (rb_node) { 1271a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, rb_node); 12723fd0a558SYan, Zheng if (node->detached) 12733fd0a558SYan, Zheng node = NULL; 12743fd0a558SYan, Zheng else 12753fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 12763fd0a558SYan, Zheng } 12773fd0a558SYan, Zheng 12783fd0a558SYan, Zheng if (!node) { 1279e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, 12803fd0a558SYan, Zheng reloc_root->commit_root->start); 12813fd0a558SYan, Zheng if (rb_node) { 1282a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, 12833fd0a558SYan, Zheng rb_node); 12843fd0a558SYan, Zheng BUG_ON(node->detached); 12853fd0a558SYan, Zheng } 12863fd0a558SYan, Zheng } 12873fd0a558SYan, Zheng 12883fd0a558SYan, Zheng if (!node) 12893fd0a558SYan, Zheng return 0; 12903fd0a558SYan, Zheng 12910304f2d8SQu Wenruo new_node = alloc_backref_node(cache, dest->node->start, node->level); 12923fd0a558SYan, Zheng if (!new_node) 12933fd0a558SYan, Zheng return -ENOMEM; 12943fd0a558SYan, Zheng 12953fd0a558SYan, Zheng new_node->lowest = node->lowest; 12966848ad64SYan, Zheng new_node->checked = 1; 129700246528SJosef Bacik new_node->root = btrfs_grab_root(dest); 12980b530bc5SJosef Bacik ASSERT(new_node->root); 12993fd0a558SYan, Zheng 13003fd0a558SYan, Zheng if (!node->lowest) { 13013fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 13023fd0a558SYan, Zheng new_edge = alloc_backref_edge(cache); 13033fd0a558SYan, Zheng if (!new_edge) 13043fd0a558SYan, Zheng goto fail; 13053fd0a558SYan, Zheng 13062a979612SQu Wenruo link_backref_edge(new_edge, edge->node[LOWER], new_node, 13072a979612SQu Wenruo LINK_UPPER); 13083fd0a558SYan, Zheng } 130976b9e23dSMiao Xie } else { 131076b9e23dSMiao Xie list_add_tail(&new_node->lower, &cache->leaves); 13113fd0a558SYan, Zheng } 13123fd0a558SYan, Zheng 1313e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr, 13143fd0a558SYan, Zheng &new_node->rb_node); 131543c04fb1SJeff Mahoney if (rb_node) 131643c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, new_node->bytenr); 13173fd0a558SYan, Zheng 13183fd0a558SYan, Zheng if (!new_node->lowest) { 13193fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 13203fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 13213fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 13223fd0a558SYan, Zheng } 13233fd0a558SYan, Zheng } 13243fd0a558SYan, Zheng return 0; 13253fd0a558SYan, Zheng fail: 13263fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 13273fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 1328a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 13293fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 13303fd0a558SYan, Zheng free_backref_edge(cache, new_edge); 13313fd0a558SYan, Zheng } 13323fd0a558SYan, Zheng free_backref_node(cache, new_node); 13333fd0a558SYan, Zheng return -ENOMEM; 13343fd0a558SYan, Zheng } 13353fd0a558SYan, Zheng 13363fd0a558SYan, Zheng /* 13375d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 13385d4f98a2SYan Zheng */ 1339ffd7b339SJeff Mahoney static int __must_check __add_reloc_root(struct btrfs_root *root) 13405d4f98a2SYan Zheng { 13410b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 13425d4f98a2SYan Zheng struct rb_node *rb_node; 13435d4f98a2SYan Zheng struct mapping_node *node; 13440b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 13455d4f98a2SYan Zheng 13465d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 1347ffd7b339SJeff Mahoney if (!node) 1348ffd7b339SJeff Mahoney return -ENOMEM; 13495d4f98a2SYan Zheng 1350ea287ab1SJosef Bacik node->bytenr = root->commit_root->start; 13515d4f98a2SYan Zheng node->data = root; 13525d4f98a2SYan Zheng 13535d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1354e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 13555d4f98a2SYan Zheng node->bytenr, &node->rb_node); 13565d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 1357ffd7b339SJeff Mahoney if (rb_node) { 13580b246afaSJeff Mahoney btrfs_panic(fs_info, -EEXIST, 13595d163e0eSJeff Mahoney "Duplicate root found for start=%llu while inserting into relocation tree", 13605d163e0eSJeff Mahoney node->bytenr); 1361ffd7b339SJeff Mahoney } 13625d4f98a2SYan Zheng 13635d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 13645d4f98a2SYan Zheng return 0; 13655d4f98a2SYan Zheng } 13665d4f98a2SYan Zheng 13675d4f98a2SYan Zheng /* 1368c974c464SWang Shilong * helper to delete the 'address of tree root -> reloc tree' 13695d4f98a2SYan Zheng * mapping 13705d4f98a2SYan Zheng */ 1371c974c464SWang Shilong static void __del_reloc_root(struct btrfs_root *root) 13725d4f98a2SYan Zheng { 13730b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 13745d4f98a2SYan Zheng struct rb_node *rb_node; 13755d4f98a2SYan Zheng struct mapping_node *node = NULL; 13760b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1377f44deb74SJosef Bacik bool put_ref = false; 13785d4f98a2SYan Zheng 137965c6e82bSQu Wenruo if (rc && root->node) { 13805d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1381e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 1382ea287ab1SJosef Bacik root->commit_root->start); 1383c974c464SWang Shilong if (rb_node) { 1384c974c464SWang Shilong node = rb_entry(rb_node, struct mapping_node, rb_node); 1385c974c464SWang Shilong rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 1386ea287ab1SJosef Bacik RB_CLEAR_NODE(&node->rb_node); 1387c974c464SWang Shilong } 1388c974c464SWang Shilong spin_unlock(&rc->reloc_root_tree.lock); 1389c974c464SWang Shilong if (!node) 1390c974c464SWang Shilong return; 1391c974c464SWang Shilong BUG_ON((struct btrfs_root *)node->data != root); 1392389305b2SQu Wenruo } 1393c974c464SWang Shilong 1394f44deb74SJosef Bacik /* 1395f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 1396f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 1397f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 1398f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 1399f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 1400f44deb74SJosef Bacik * messing with the list is in charge of the reference. 1401f44deb74SJosef Bacik */ 14020b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 1403f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 1404f44deb74SJosef Bacik put_ref = true; 1405c974c464SWang Shilong list_del_init(&root->root_list); 1406f44deb74SJosef Bacik } 14070b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 1408f44deb74SJosef Bacik if (put_ref) 1409f44deb74SJosef Bacik btrfs_put_root(root); 1410c974c464SWang Shilong kfree(node); 1411c974c464SWang Shilong } 1412c974c464SWang Shilong 1413c974c464SWang Shilong /* 1414c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 1415c974c464SWang Shilong * mapping 1416c974c464SWang Shilong */ 1417ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 1418c974c464SWang Shilong { 14190b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 1420c974c464SWang Shilong struct rb_node *rb_node; 1421c974c464SWang Shilong struct mapping_node *node = NULL; 14220b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1423c974c464SWang Shilong 1424c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 1425e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 1426ea287ab1SJosef Bacik root->commit_root->start); 14275d4f98a2SYan Zheng if (rb_node) { 14285d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 14295d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 14305d4f98a2SYan Zheng } 14315d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 14325d4f98a2SYan Zheng 14338f71f3e0SLiu Bo if (!node) 14348f71f3e0SLiu Bo return 0; 14355d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 14365d4f98a2SYan Zheng 14375d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1438ea287ab1SJosef Bacik node->bytenr = root->node->start; 1439e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 14405d4f98a2SYan Zheng node->bytenr, &node->rb_node); 14415d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 144243c04fb1SJeff Mahoney if (rb_node) 144343c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 14445d4f98a2SYan Zheng return 0; 14455d4f98a2SYan Zheng } 14465d4f98a2SYan Zheng 14473fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 14483fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 14495d4f98a2SYan Zheng { 14500b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14515d4f98a2SYan Zheng struct btrfs_root *reloc_root; 14525d4f98a2SYan Zheng struct extent_buffer *eb; 14535d4f98a2SYan Zheng struct btrfs_root_item *root_item; 14545d4f98a2SYan Zheng struct btrfs_key root_key; 14555d4f98a2SYan Zheng int ret; 14565d4f98a2SYan Zheng 14575d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 14585d4f98a2SYan Zheng BUG_ON(!root_item); 14595d4f98a2SYan Zheng 14605d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 14615d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 14623fd0a558SYan, Zheng root_key.offset = objectid; 14635d4f98a2SYan Zheng 14643fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 1465054570a1SFilipe Manana u64 commit_root_gen; 1466054570a1SFilipe Manana 14673fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 14685d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 14695d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 14705d4f98a2SYan Zheng BUG_ON(ret); 1471054570a1SFilipe Manana /* 1472054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 1473054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 1474054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 1475054570a1SFilipe Manana * either inside the critical section of a transaction commit 1476054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 1477054570a1SFilipe Manana * it's created before the transaction commit is started. 1478054570a1SFilipe Manana */ 1479054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 1480054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 14813fd0a558SYan, Zheng } else { 14823fd0a558SYan, Zheng /* 14833fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 14843fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 14853fd0a558SYan, Zheng * modified after it was created have RELOC flag 14863fd0a558SYan, Zheng * set in their headers. so it's OK to not update 14873fd0a558SYan, Zheng * the 'last_snapshot'. 14883fd0a558SYan, Zheng */ 14893fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 14903fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 14913fd0a558SYan, Zheng BUG_ON(ret); 14923fd0a558SYan, Zheng } 14933fd0a558SYan, Zheng 14945d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 14955d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 14965d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 14975d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 14983fd0a558SYan, Zheng 14993fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 15003fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 15013fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 15023fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 15035d4f98a2SYan Zheng root_item->drop_level = 0; 15043fd0a558SYan, Zheng } 15055d4f98a2SYan Zheng 15065d4f98a2SYan Zheng btrfs_tree_unlock(eb); 15075d4f98a2SYan Zheng free_extent_buffer(eb); 15085d4f98a2SYan Zheng 15090b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 15105d4f98a2SYan Zheng &root_key, root_item); 15115d4f98a2SYan Zheng BUG_ON(ret); 15125d4f98a2SYan Zheng kfree(root_item); 15135d4f98a2SYan Zheng 15143dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 15155d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 15163dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 15175d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 15183fd0a558SYan, Zheng return reloc_root; 15193fd0a558SYan, Zheng } 15203fd0a558SYan, Zheng 15213fd0a558SYan, Zheng /* 15223fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 15233fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 1524f44deb74SJosef Bacik * 1525f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 1526f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 15273fd0a558SYan, Zheng */ 15283fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 15293fd0a558SYan, Zheng struct btrfs_root *root) 15303fd0a558SYan, Zheng { 15310b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15323fd0a558SYan, Zheng struct btrfs_root *reloc_root; 15330b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 153420dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 15353fd0a558SYan, Zheng int clear_rsv = 0; 1536ffd7b339SJeff Mahoney int ret; 15373fd0a558SYan, Zheng 1538aec7db3bSJosef Bacik if (!rc) 15392abc726aSJosef Bacik return 0; 15402abc726aSJosef Bacik 15411fac4a54SQu Wenruo /* 15421fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 15431fac4a54SQu Wenruo * create/update the dead reloc tree 15441fac4a54SQu Wenruo */ 15456282675eSQu Wenruo if (reloc_root_is_dead(root)) 15461fac4a54SQu Wenruo return 0; 15471fac4a54SQu Wenruo 1548aec7db3bSJosef Bacik /* 1549aec7db3bSJosef Bacik * This is subtle but important. We do not do 1550aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 1551aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 1552aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 1553aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 1554aec7db3bSJosef Bacik * in. 1555aec7db3bSJosef Bacik */ 15563fd0a558SYan, Zheng if (root->reloc_root) { 15573fd0a558SYan, Zheng reloc_root = root->reloc_root; 15583fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 15593fd0a558SYan, Zheng return 0; 15603fd0a558SYan, Zheng } 15613fd0a558SYan, Zheng 1562aec7db3bSJosef Bacik /* 1563aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 1564aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 1565aec7db3bSJosef Bacik */ 1566aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 1567aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 1568aec7db3bSJosef Bacik return 0; 1569aec7db3bSJosef Bacik 157020dd2cbfSMiao Xie if (!trans->reloc_reserved) { 157120dd2cbfSMiao Xie rsv = trans->block_rsv; 15723fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 15733fd0a558SYan, Zheng clear_rsv = 1; 15743fd0a558SYan, Zheng } 15753fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 15763fd0a558SYan, Zheng if (clear_rsv) 157720dd2cbfSMiao Xie trans->block_rsv = rsv; 15785d4f98a2SYan Zheng 1579ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 1580ffd7b339SJeff Mahoney BUG_ON(ret < 0); 1581f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 15825d4f98a2SYan Zheng return 0; 15835d4f98a2SYan Zheng } 15845d4f98a2SYan Zheng 15855d4f98a2SYan Zheng /* 15865d4f98a2SYan Zheng * update root item of reloc tree 15875d4f98a2SYan Zheng */ 15885d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 15895d4f98a2SYan Zheng struct btrfs_root *root) 15905d4f98a2SYan Zheng { 15910b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15925d4f98a2SYan Zheng struct btrfs_root *reloc_root; 15935d4f98a2SYan Zheng struct btrfs_root_item *root_item; 15945d4f98a2SYan Zheng int ret; 15955d4f98a2SYan Zheng 15966282675eSQu Wenruo if (!have_reloc_root(root)) 15977585717fSChris Mason goto out; 15985d4f98a2SYan Zheng 15995d4f98a2SYan Zheng reloc_root = root->reloc_root; 16005d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16015d4f98a2SYan Zheng 1602f44deb74SJosef Bacik /* 1603f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 1604f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 1605f44deb74SJosef Bacik * hold it while we update the reloc root. 1606f44deb74SJosef Bacik */ 1607f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 1608f44deb74SJosef Bacik 1609d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 16100b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 16113fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 1612d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 16136282675eSQu Wenruo /* 16146282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 16156282675eSQu Wenruo * have_reloc_root will not touch it from now on. 16166282675eSQu Wenruo */ 16176282675eSQu Wenruo smp_wmb(); 1618c974c464SWang Shilong __del_reloc_root(reloc_root); 16195d4f98a2SYan Zheng } 16205d4f98a2SYan Zheng 16215d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 1622ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 16235d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 16245d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 16255d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 16265d4f98a2SYan Zheng } 16275d4f98a2SYan Zheng 16280b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 16295d4f98a2SYan Zheng &reloc_root->root_key, root_item); 16305d4f98a2SYan Zheng BUG_ON(ret); 1631f44deb74SJosef Bacik btrfs_put_root(reloc_root); 16327585717fSChris Mason out: 16335d4f98a2SYan Zheng return 0; 16345d4f98a2SYan Zheng } 16355d4f98a2SYan Zheng 16365d4f98a2SYan Zheng /* 16375d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 16385d4f98a2SYan Zheng * in a subvolume 16395d4f98a2SYan Zheng */ 16405d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 16415d4f98a2SYan Zheng { 16425d4f98a2SYan Zheng struct rb_node *node; 16435d4f98a2SYan Zheng struct rb_node *prev; 16445d4f98a2SYan Zheng struct btrfs_inode *entry; 16455d4f98a2SYan Zheng struct inode *inode; 16465d4f98a2SYan Zheng 16475d4f98a2SYan Zheng spin_lock(&root->inode_lock); 16485d4f98a2SYan Zheng again: 16495d4f98a2SYan Zheng node = root->inode_tree.rb_node; 16505d4f98a2SYan Zheng prev = NULL; 16515d4f98a2SYan Zheng while (node) { 16525d4f98a2SYan Zheng prev = node; 16535d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 16545d4f98a2SYan Zheng 16554a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 16565d4f98a2SYan Zheng node = node->rb_left; 16574a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 16585d4f98a2SYan Zheng node = node->rb_right; 16595d4f98a2SYan Zheng else 16605d4f98a2SYan Zheng break; 16615d4f98a2SYan Zheng } 16625d4f98a2SYan Zheng if (!node) { 16635d4f98a2SYan Zheng while (prev) { 16645d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 16654a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 16665d4f98a2SYan Zheng node = prev; 16675d4f98a2SYan Zheng break; 16685d4f98a2SYan Zheng } 16695d4f98a2SYan Zheng prev = rb_next(prev); 16705d4f98a2SYan Zheng } 16715d4f98a2SYan Zheng } 16725d4f98a2SYan Zheng while (node) { 16735d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 16745d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 16755d4f98a2SYan Zheng if (inode) { 16765d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 16775d4f98a2SYan Zheng return inode; 16785d4f98a2SYan Zheng } 16795d4f98a2SYan Zheng 16804a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 16815d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 16825d4f98a2SYan Zheng goto again; 16835d4f98a2SYan Zheng 16845d4f98a2SYan Zheng node = rb_next(node); 16855d4f98a2SYan Zheng } 16865d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 16875d4f98a2SYan Zheng return NULL; 16885d4f98a2SYan Zheng } 16895d4f98a2SYan Zheng 16905d4f98a2SYan Zheng /* 16915d4f98a2SYan Zheng * get new location of data 16925d4f98a2SYan Zheng */ 16935d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 16945d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 16955d4f98a2SYan Zheng { 16965d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 16975d4f98a2SYan Zheng struct btrfs_path *path; 16985d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 16995d4f98a2SYan Zheng struct extent_buffer *leaf; 17005d4f98a2SYan Zheng int ret; 17015d4f98a2SYan Zheng 17025d4f98a2SYan Zheng path = btrfs_alloc_path(); 17035d4f98a2SYan Zheng if (!path) 17045d4f98a2SYan Zheng return -ENOMEM; 17055d4f98a2SYan Zheng 17065d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 1707f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 1708f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 17095d4f98a2SYan Zheng if (ret < 0) 17105d4f98a2SYan Zheng goto out; 17115d4f98a2SYan Zheng if (ret > 0) { 17125d4f98a2SYan Zheng ret = -ENOENT; 17135d4f98a2SYan Zheng goto out; 17145d4f98a2SYan Zheng } 17155d4f98a2SYan Zheng 17165d4f98a2SYan Zheng leaf = path->nodes[0]; 17175d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 17185d4f98a2SYan Zheng struct btrfs_file_extent_item); 17195d4f98a2SYan Zheng 17205d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 17215d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 17225d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 17235d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 17245d4f98a2SYan Zheng 17255d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 172683d4cfd4SJosef Bacik ret = -EINVAL; 17275d4f98a2SYan Zheng goto out; 17285d4f98a2SYan Zheng } 17295d4f98a2SYan Zheng 17305d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 17315d4f98a2SYan Zheng ret = 0; 17325d4f98a2SYan Zheng out: 17335d4f98a2SYan Zheng btrfs_free_path(path); 17345d4f98a2SYan Zheng return ret; 17355d4f98a2SYan Zheng } 17365d4f98a2SYan Zheng 17375d4f98a2SYan Zheng /* 17385d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 17395d4f98a2SYan Zheng * the new locations. 17405d4f98a2SYan Zheng */ 17413fd0a558SYan, Zheng static noinline_for_stack 17423fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 17435d4f98a2SYan Zheng struct reloc_control *rc, 17445d4f98a2SYan Zheng struct btrfs_root *root, 17453fd0a558SYan, Zheng struct extent_buffer *leaf) 17465d4f98a2SYan Zheng { 17470b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 17485d4f98a2SYan Zheng struct btrfs_key key; 17495d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 17505d4f98a2SYan Zheng struct inode *inode = NULL; 17515d4f98a2SYan Zheng u64 parent; 17525d4f98a2SYan Zheng u64 bytenr; 17533fd0a558SYan, Zheng u64 new_bytenr = 0; 17545d4f98a2SYan Zheng u64 num_bytes; 17555d4f98a2SYan Zheng u64 end; 17565d4f98a2SYan Zheng u32 nritems; 17575d4f98a2SYan Zheng u32 i; 175883d4cfd4SJosef Bacik int ret = 0; 17595d4f98a2SYan Zheng int first = 1; 17605d4f98a2SYan Zheng int dirty = 0; 17615d4f98a2SYan Zheng 17625d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 17635d4f98a2SYan Zheng return 0; 17645d4f98a2SYan Zheng 17655d4f98a2SYan Zheng /* reloc trees always use full backref */ 17665d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 17675d4f98a2SYan Zheng parent = leaf->start; 17685d4f98a2SYan Zheng else 17695d4f98a2SYan Zheng parent = 0; 17705d4f98a2SYan Zheng 17715d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 17725d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 177382fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 177482fa113fSQu Wenruo 17755d4f98a2SYan Zheng cond_resched(); 17765d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 17775d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 17785d4f98a2SYan Zheng continue; 17795d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 17805d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 17815d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 17825d4f98a2SYan Zheng continue; 17835d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 17845d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 17855d4f98a2SYan Zheng if (bytenr == 0) 17865d4f98a2SYan Zheng continue; 17879569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 17889569cc20SQu Wenruo rc->block_group->length)) 17895d4f98a2SYan Zheng continue; 17905d4f98a2SYan Zheng 17915d4f98a2SYan Zheng /* 17925d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 17935d4f98a2SYan Zheng * to complete and drop the extent cache 17945d4f98a2SYan Zheng */ 17955d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 17965d4f98a2SYan Zheng if (first) { 17975d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 17985d4f98a2SYan Zheng first = 0; 17994a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 18003fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 18015d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 18025d4f98a2SYan Zheng } 18034a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 18045d4f98a2SYan Zheng end = key.offset + 18055d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 18065d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 18070b246afaSJeff Mahoney fs_info->sectorsize)); 18080b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 18095d4f98a2SYan Zheng end--; 18105d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1811d0082371SJeff Mahoney key.offset, end); 18125d4f98a2SYan Zheng if (!ret) 18135d4f98a2SYan Zheng continue; 18145d4f98a2SYan Zheng 1815dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1816dcdbc059SNikolay Borisov key.offset, end, 1); 18175d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1818d0082371SJeff Mahoney key.offset, end); 18195d4f98a2SYan Zheng } 18205d4f98a2SYan Zheng } 18215d4f98a2SYan Zheng 18225d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 18235d4f98a2SYan Zheng bytenr, num_bytes); 182483d4cfd4SJosef Bacik if (ret) { 182583d4cfd4SJosef Bacik /* 182683d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 182783d4cfd4SJosef Bacik * in the file extent yet. 182883d4cfd4SJosef Bacik */ 182983d4cfd4SJosef Bacik break; 18303fd0a558SYan, Zheng } 18315d4f98a2SYan Zheng 18325d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 18335d4f98a2SYan Zheng dirty = 1; 18345d4f98a2SYan Zheng 18355d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 183682fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 183782fa113fSQu Wenruo num_bytes, parent); 183882fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 183982fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1840b06c4bf5SFilipe Manana key.objectid, key.offset); 184182fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 184283d4cfd4SJosef Bacik if (ret) { 184366642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 184483d4cfd4SJosef Bacik break; 184583d4cfd4SJosef Bacik } 18465d4f98a2SYan Zheng 1847ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1848ffd4bb2aSQu Wenruo num_bytes, parent); 1849ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1850ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1851b06c4bf5SFilipe Manana key.objectid, key.offset); 1852ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 185383d4cfd4SJosef Bacik if (ret) { 185466642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 185583d4cfd4SJosef Bacik break; 185683d4cfd4SJosef Bacik } 18575d4f98a2SYan Zheng } 18585d4f98a2SYan Zheng if (dirty) 18595d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 18603fd0a558SYan, Zheng if (inode) 18613fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 186283d4cfd4SJosef Bacik return ret; 18635d4f98a2SYan Zheng } 18645d4f98a2SYan Zheng 18655d4f98a2SYan Zheng static noinline_for_stack 18665d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 18675d4f98a2SYan Zheng struct btrfs_path *path, int level) 18685d4f98a2SYan Zheng { 18695d4f98a2SYan Zheng struct btrfs_disk_key key1; 18705d4f98a2SYan Zheng struct btrfs_disk_key key2; 18715d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 18725d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 18735d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 18745d4f98a2SYan Zheng } 18755d4f98a2SYan Zheng 18765d4f98a2SYan Zheng /* 18775d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 18785d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 18795d4f98a2SYan Zheng * reloc tree was create can be replaced. 18805d4f98a2SYan Zheng * 18815d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 18825d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 18835d4f98a2SYan Zheng * errors, a negative error number is returned. 18845d4f98a2SYan Zheng */ 18853fd0a558SYan, Zheng static noinline_for_stack 18863d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 18875d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 18885d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 18895d4f98a2SYan Zheng int lowest_level, int max_level) 18905d4f98a2SYan Zheng { 18910b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 18925d4f98a2SYan Zheng struct extent_buffer *eb; 18935d4f98a2SYan Zheng struct extent_buffer *parent; 189482fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 18955d4f98a2SYan Zheng struct btrfs_key key; 18965d4f98a2SYan Zheng u64 old_bytenr; 18975d4f98a2SYan Zheng u64 new_bytenr; 18985d4f98a2SYan Zheng u64 old_ptr_gen; 18995d4f98a2SYan Zheng u64 new_ptr_gen; 19005d4f98a2SYan Zheng u64 last_snapshot; 19015d4f98a2SYan Zheng u32 blocksize; 19023fd0a558SYan, Zheng int cow = 0; 19035d4f98a2SYan Zheng int level; 19045d4f98a2SYan Zheng int ret; 19055d4f98a2SYan Zheng int slot; 19065d4f98a2SYan Zheng 19075d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 19085d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 19095d4f98a2SYan Zheng 19105d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 19113fd0a558SYan, Zheng again: 19125d4f98a2SYan Zheng slot = path->slots[lowest_level]; 19135d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 19145d4f98a2SYan Zheng 19155d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 19168bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19175d4f98a2SYan Zheng level = btrfs_header_level(eb); 19185d4f98a2SYan Zheng 19195d4f98a2SYan Zheng if (level < lowest_level) { 19205d4f98a2SYan Zheng btrfs_tree_unlock(eb); 19215d4f98a2SYan Zheng free_extent_buffer(eb); 19225d4f98a2SYan Zheng return 0; 19235d4f98a2SYan Zheng } 19245d4f98a2SYan Zheng 19253fd0a558SYan, Zheng if (cow) { 19265d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 19275d4f98a2SYan Zheng BUG_ON(ret); 19283fd0a558SYan, Zheng } 19298bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19305d4f98a2SYan Zheng 19315d4f98a2SYan Zheng if (next_key) { 19325d4f98a2SYan Zheng next_key->objectid = (u64)-1; 19335d4f98a2SYan Zheng next_key->type = (u8)-1; 19345d4f98a2SYan Zheng next_key->offset = (u64)-1; 19355d4f98a2SYan Zheng } 19365d4f98a2SYan Zheng 19375d4f98a2SYan Zheng parent = eb; 19385d4f98a2SYan Zheng while (1) { 1939581c1760SQu Wenruo struct btrfs_key first_key; 1940581c1760SQu Wenruo 19415d4f98a2SYan Zheng level = btrfs_header_level(parent); 19425d4f98a2SYan Zheng BUG_ON(level < lowest_level); 19435d4f98a2SYan Zheng 19445d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 1945cbca7d59SFilipe Manana if (ret < 0) 1946cbca7d59SFilipe Manana break; 19475d4f98a2SYan Zheng if (ret && slot > 0) 19485d4f98a2SYan Zheng slot--; 19495d4f98a2SYan Zheng 19505d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 19515d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 19525d4f98a2SYan Zheng 19535d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 19540b246afaSJeff Mahoney blocksize = fs_info->nodesize; 19555d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 195617515f1bSQu Wenruo btrfs_node_key_to_cpu(parent, &first_key, slot); 19575d4f98a2SYan Zheng 19585d4f98a2SYan Zheng if (level <= max_level) { 19595d4f98a2SYan Zheng eb = path->nodes[level]; 19605d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 19615d4f98a2SYan Zheng path->slots[level]); 19625d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 19635d4f98a2SYan Zheng path->slots[level]); 19645d4f98a2SYan Zheng } else { 19655d4f98a2SYan Zheng new_bytenr = 0; 19665d4f98a2SYan Zheng new_ptr_gen = 0; 19675d4f98a2SYan Zheng } 19685d4f98a2SYan Zheng 1969fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 19705d4f98a2SYan Zheng ret = level; 19715d4f98a2SYan Zheng break; 19725d4f98a2SYan Zheng } 19735d4f98a2SYan Zheng 19745d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 19755d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 19763fd0a558SYan, Zheng if (level <= lowest_level) { 19775d4f98a2SYan Zheng ret = 0; 19785d4f98a2SYan Zheng break; 19795d4f98a2SYan Zheng } 19805d4f98a2SYan Zheng 1981581c1760SQu Wenruo eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen, 1982581c1760SQu Wenruo level - 1, &first_key); 198364c043deSLiu Bo if (IS_ERR(eb)) { 198464c043deSLiu Bo ret = PTR_ERR(eb); 1985264813acSLiu Bo break; 198664c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 198764c043deSLiu Bo ret = -EIO; 1988416bc658SJosef Bacik free_extent_buffer(eb); 1989379cde74SStefan Behrens break; 1990416bc658SJosef Bacik } 19915d4f98a2SYan Zheng btrfs_tree_lock(eb); 19923fd0a558SYan, Zheng if (cow) { 19935d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 19945d4f98a2SYan Zheng slot, &eb); 19955d4f98a2SYan Zheng BUG_ON(ret); 19965d4f98a2SYan Zheng } 19978bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19985d4f98a2SYan Zheng 19995d4f98a2SYan Zheng btrfs_tree_unlock(parent); 20005d4f98a2SYan Zheng free_extent_buffer(parent); 20015d4f98a2SYan Zheng 20025d4f98a2SYan Zheng parent = eb; 20035d4f98a2SYan Zheng continue; 20045d4f98a2SYan Zheng } 20055d4f98a2SYan Zheng 20063fd0a558SYan, Zheng if (!cow) { 20073fd0a558SYan, Zheng btrfs_tree_unlock(parent); 20083fd0a558SYan, Zheng free_extent_buffer(parent); 20093fd0a558SYan, Zheng cow = 1; 20103fd0a558SYan, Zheng goto again; 20113fd0a558SYan, Zheng } 20123fd0a558SYan, Zheng 20135d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 20145d4f98a2SYan Zheng path->slots[level]); 2015b3b4aa74SDavid Sterba btrfs_release_path(path); 20165d4f98a2SYan Zheng 20175d4f98a2SYan Zheng path->lowest_level = level; 20185d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 20195d4f98a2SYan Zheng path->lowest_level = 0; 20205d4f98a2SYan Zheng BUG_ON(ret); 20215d4f98a2SYan Zheng 20225d4f98a2SYan Zheng /* 2023824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 2024824d8dffSQu Wenruo * 2025824d8dffSQu Wenruo * We must trace both trees. 2026824d8dffSQu Wenruo * 1) Tree reloc subtree 2027824d8dffSQu Wenruo * If not traced, we will leak data numbers 2028824d8dffSQu Wenruo * 2) Fs subtree 2029824d8dffSQu Wenruo * If not traced, we will double count old data 2030f616f5cdSQu Wenruo * 2031f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 2032f616f5cdSQu Wenruo * the swapped tree blocks. 2033f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 2034f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 2035824d8dffSQu Wenruo */ 2036370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 2037370a11b8SQu Wenruo rc->block_group, parent, slot, 2038370a11b8SQu Wenruo path->nodes[level], path->slots[level], 2039370a11b8SQu Wenruo last_snapshot); 2040370a11b8SQu Wenruo if (ret < 0) 2041370a11b8SQu Wenruo break; 2042824d8dffSQu Wenruo /* 20435d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 20445d4f98a2SYan Zheng */ 20455d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 20465d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 20475d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 20485d4f98a2SYan Zheng 20495d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 20505d4f98a2SYan Zheng path->slots[level], old_bytenr); 20515d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 20525d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 20535d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 20545d4f98a2SYan Zheng 205582fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 205682fa113fSQu Wenruo blocksize, path->nodes[level]->start); 205782fa113fSQu Wenruo ref.skip_qgroup = true; 205882fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 205982fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 20605d4f98a2SYan Zheng BUG_ON(ret); 206182fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 206282fa113fSQu Wenruo blocksize, 0); 206382fa113fSQu Wenruo ref.skip_qgroup = true; 206482fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 206582fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 20665d4f98a2SYan Zheng BUG_ON(ret); 20675d4f98a2SYan Zheng 2068ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 2069ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 2070ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 2071ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2072ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 20735d4f98a2SYan Zheng BUG_ON(ret); 20745d4f98a2SYan Zheng 2075ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 2076ffd4bb2aSQu Wenruo blocksize, 0); 2077ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 2078ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2079ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 20805d4f98a2SYan Zheng BUG_ON(ret); 20815d4f98a2SYan Zheng 20825d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 20835d4f98a2SYan Zheng 20845d4f98a2SYan Zheng ret = level; 20855d4f98a2SYan Zheng break; 20865d4f98a2SYan Zheng } 20875d4f98a2SYan Zheng btrfs_tree_unlock(parent); 20885d4f98a2SYan Zheng free_extent_buffer(parent); 20895d4f98a2SYan Zheng return ret; 20905d4f98a2SYan Zheng } 20915d4f98a2SYan Zheng 20925d4f98a2SYan Zheng /* 20935d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 20945d4f98a2SYan Zheng */ 20955d4f98a2SYan Zheng static noinline_for_stack 20965d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 20975d4f98a2SYan Zheng int *level) 20985d4f98a2SYan Zheng { 20995d4f98a2SYan Zheng struct extent_buffer *eb; 21005d4f98a2SYan Zheng int i; 21015d4f98a2SYan Zheng u64 last_snapshot; 21025d4f98a2SYan Zheng u32 nritems; 21035d4f98a2SYan Zheng 21045d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 21055d4f98a2SYan Zheng 21065d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 21075d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 21085d4f98a2SYan Zheng path->nodes[i] = NULL; 21095d4f98a2SYan Zheng } 21105d4f98a2SYan Zheng 21115d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 21125d4f98a2SYan Zheng eb = path->nodes[i]; 21135d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 21145d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 21155d4f98a2SYan Zheng path->slots[i]++; 21165d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 21175d4f98a2SYan Zheng last_snapshot) 21185d4f98a2SYan Zheng continue; 21195d4f98a2SYan Zheng 21205d4f98a2SYan Zheng *level = i; 21215d4f98a2SYan Zheng return 0; 21225d4f98a2SYan Zheng } 21235d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 21245d4f98a2SYan Zheng path->nodes[i] = NULL; 21255d4f98a2SYan Zheng } 21265d4f98a2SYan Zheng return 1; 21275d4f98a2SYan Zheng } 21285d4f98a2SYan Zheng 21295d4f98a2SYan Zheng /* 21305d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 21315d4f98a2SYan Zheng */ 21325d4f98a2SYan Zheng static noinline_for_stack 21335d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 21345d4f98a2SYan Zheng int *level) 21355d4f98a2SYan Zheng { 21362ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 21375d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 21385d4f98a2SYan Zheng int i; 21395d4f98a2SYan Zheng u64 bytenr; 21405d4f98a2SYan Zheng u64 ptr_gen = 0; 21415d4f98a2SYan Zheng u64 last_snapshot; 21425d4f98a2SYan Zheng u32 nritems; 21435d4f98a2SYan Zheng 21445d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 21455d4f98a2SYan Zheng 21465d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 2147581c1760SQu Wenruo struct btrfs_key first_key; 2148581c1760SQu Wenruo 21495d4f98a2SYan Zheng eb = path->nodes[i]; 21505d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 21515d4f98a2SYan Zheng while (path->slots[i] < nritems) { 21525d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 21535d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 21545d4f98a2SYan Zheng break; 21555d4f98a2SYan Zheng path->slots[i]++; 21565d4f98a2SYan Zheng } 21575d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 21585d4f98a2SYan Zheng if (i == *level) 21595d4f98a2SYan Zheng break; 21605d4f98a2SYan Zheng *level = i + 1; 21615d4f98a2SYan Zheng return 0; 21625d4f98a2SYan Zheng } 21635d4f98a2SYan Zheng if (i == 1) { 21645d4f98a2SYan Zheng *level = i; 21655d4f98a2SYan Zheng return 0; 21665d4f98a2SYan Zheng } 21675d4f98a2SYan Zheng 21685d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 2169581c1760SQu Wenruo btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]); 2170581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1, 2171581c1760SQu Wenruo &first_key); 217264c043deSLiu Bo if (IS_ERR(eb)) { 217364c043deSLiu Bo return PTR_ERR(eb); 217464c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2175416bc658SJosef Bacik free_extent_buffer(eb); 2176416bc658SJosef Bacik return -EIO; 2177416bc658SJosef Bacik } 21785d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 21795d4f98a2SYan Zheng path->nodes[i - 1] = eb; 21805d4f98a2SYan Zheng path->slots[i - 1] = 0; 21815d4f98a2SYan Zheng } 21825d4f98a2SYan Zheng return 1; 21835d4f98a2SYan Zheng } 21845d4f98a2SYan Zheng 21855d4f98a2SYan Zheng /* 21865d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 21875d4f98a2SYan Zheng * [min_key, max_key) 21885d4f98a2SYan Zheng */ 21895d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 21905d4f98a2SYan Zheng struct btrfs_key *min_key, 21915d4f98a2SYan Zheng struct btrfs_key *max_key) 21925d4f98a2SYan Zheng { 21930b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 21945d4f98a2SYan Zheng struct inode *inode = NULL; 21955d4f98a2SYan Zheng u64 objectid; 21965d4f98a2SYan Zheng u64 start, end; 219733345d01SLi Zefan u64 ino; 21985d4f98a2SYan Zheng 21995d4f98a2SYan Zheng objectid = min_key->objectid; 22005d4f98a2SYan Zheng while (1) { 22015d4f98a2SYan Zheng cond_resched(); 22025d4f98a2SYan Zheng iput(inode); 22035d4f98a2SYan Zheng 22045d4f98a2SYan Zheng if (objectid > max_key->objectid) 22055d4f98a2SYan Zheng break; 22065d4f98a2SYan Zheng 22075d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 22085d4f98a2SYan Zheng if (!inode) 22095d4f98a2SYan Zheng break; 22104a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 22115d4f98a2SYan Zheng 221233345d01SLi Zefan if (ino > max_key->objectid) { 22135d4f98a2SYan Zheng iput(inode); 22145d4f98a2SYan Zheng break; 22155d4f98a2SYan Zheng } 22165d4f98a2SYan Zheng 221733345d01SLi Zefan objectid = ino + 1; 22185d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 22195d4f98a2SYan Zheng continue; 22205d4f98a2SYan Zheng 222133345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 22225d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 22235d4f98a2SYan Zheng continue; 22245d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 22255d4f98a2SYan Zheng start = 0; 22265d4f98a2SYan Zheng else { 22275d4f98a2SYan Zheng start = min_key->offset; 22280b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 22295d4f98a2SYan Zheng } 22305d4f98a2SYan Zheng } else { 22315d4f98a2SYan Zheng start = 0; 22325d4f98a2SYan Zheng } 22335d4f98a2SYan Zheng 223433345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 22355d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 22365d4f98a2SYan Zheng continue; 22375d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 22385d4f98a2SYan Zheng end = (u64)-1; 22395d4f98a2SYan Zheng } else { 22405d4f98a2SYan Zheng if (max_key->offset == 0) 22415d4f98a2SYan Zheng continue; 22425d4f98a2SYan Zheng end = max_key->offset; 22430b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 22445d4f98a2SYan Zheng end--; 22455d4f98a2SYan Zheng } 22465d4f98a2SYan Zheng } else { 22475d4f98a2SYan Zheng end = (u64)-1; 22485d4f98a2SYan Zheng } 22495d4f98a2SYan Zheng 22505d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 2251d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 2252dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 2253d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 22545d4f98a2SYan Zheng } 22555d4f98a2SYan Zheng return 0; 22565d4f98a2SYan Zheng } 22575d4f98a2SYan Zheng 22585d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 22595d4f98a2SYan Zheng struct btrfs_key *key) 22605d4f98a2SYan Zheng 22615d4f98a2SYan Zheng { 22625d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 22635d4f98a2SYan Zheng if (!path->nodes[level]) 22645d4f98a2SYan Zheng break; 22655d4f98a2SYan Zheng if (path->slots[level] + 1 < 22665d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 22675d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 22685d4f98a2SYan Zheng path->slots[level] + 1); 22695d4f98a2SYan Zheng return 0; 22705d4f98a2SYan Zheng } 22715d4f98a2SYan Zheng level++; 22725d4f98a2SYan Zheng } 22735d4f98a2SYan Zheng return 1; 22745d4f98a2SYan Zheng } 22755d4f98a2SYan Zheng 22765d4f98a2SYan Zheng /* 2277d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 2278d2311e69SQu Wenruo */ 2279d2311e69SQu Wenruo static void insert_dirty_subvol(struct btrfs_trans_handle *trans, 2280d2311e69SQu Wenruo struct reloc_control *rc, 2281d2311e69SQu Wenruo struct btrfs_root *root) 2282d2311e69SQu Wenruo { 2283d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2284d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 2285d2311e69SQu Wenruo 2286d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 2287d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 2288d2311e69SQu Wenruo ASSERT(reloc_root); 2289d2311e69SQu Wenruo 2290d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 2291d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 2292d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 2293d2311e69SQu Wenruo reloc_root_item->drop_level = 0; 2294d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 2295d2311e69SQu Wenruo btrfs_update_reloc_root(trans, root); 2296d2311e69SQu Wenruo 2297d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 229800246528SJosef Bacik btrfs_grab_root(root); 2299d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 2300d2311e69SQu Wenruo } 2301d2311e69SQu Wenruo } 2302d2311e69SQu Wenruo 2303d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 2304d2311e69SQu Wenruo { 2305d2311e69SQu Wenruo struct btrfs_root *root; 2306d2311e69SQu Wenruo struct btrfs_root *next; 2307d2311e69SQu Wenruo int ret = 0; 230830d40577SQu Wenruo int ret2; 2309d2311e69SQu Wenruo 2310d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 2311d2311e69SQu Wenruo reloc_dirty_list) { 231230d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 231330d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 2314d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2315d2311e69SQu Wenruo 2316d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 2317d2311e69SQu Wenruo root->reloc_root = NULL; 23186282675eSQu Wenruo /* 23196282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 23206282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 23216282675eSQu Wenruo */ 23226282675eSQu Wenruo smp_wmb(); 23231fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 2324f28de8d8SJosef Bacik if (reloc_root) { 2325f44deb74SJosef Bacik /* 2326f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 2327f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 2328f44deb74SJosef Bacik * drop the ref ourselves. 2329f44deb74SJosef Bacik */ 2330f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 2331f44deb74SJosef Bacik if (ret2 < 0) { 2332f44deb74SJosef Bacik btrfs_put_root(reloc_root); 2333f44deb74SJosef Bacik if (!ret) 2334f28de8d8SJosef Bacik ret = ret2; 2335f28de8d8SJosef Bacik } 2336f44deb74SJosef Bacik } 233700246528SJosef Bacik btrfs_put_root(root); 233830d40577SQu Wenruo } else { 233930d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 23400078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 2341f44deb74SJosef Bacik if (ret2 < 0) { 2342f44deb74SJosef Bacik btrfs_put_root(root); 2343f44deb74SJosef Bacik if (!ret) 234430d40577SQu Wenruo ret = ret2; 234530d40577SQu Wenruo } 2346d2311e69SQu Wenruo } 2347f44deb74SJosef Bacik } 2348d2311e69SQu Wenruo return ret; 2349d2311e69SQu Wenruo } 2350d2311e69SQu Wenruo 2351d2311e69SQu Wenruo /* 23525d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 23535d4f98a2SYan Zheng * fs tree. 23545d4f98a2SYan Zheng */ 23555d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 23565d4f98a2SYan Zheng struct btrfs_root *root) 23575d4f98a2SYan Zheng { 23580b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 23595d4f98a2SYan Zheng struct btrfs_key key; 23605d4f98a2SYan Zheng struct btrfs_key next_key; 23619e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 23625d4f98a2SYan Zheng struct btrfs_root *reloc_root; 23635d4f98a2SYan Zheng struct btrfs_root_item *root_item; 23645d4f98a2SYan Zheng struct btrfs_path *path; 23653fd0a558SYan, Zheng struct extent_buffer *leaf; 23665d4f98a2SYan Zheng int level; 23675d4f98a2SYan Zheng int max_level; 23685d4f98a2SYan Zheng int replaced = 0; 23695d4f98a2SYan Zheng int ret; 23705d4f98a2SYan Zheng int err = 0; 23713fd0a558SYan, Zheng u32 min_reserved; 23725d4f98a2SYan Zheng 23735d4f98a2SYan Zheng path = btrfs_alloc_path(); 23745d4f98a2SYan Zheng if (!path) 23755d4f98a2SYan Zheng return -ENOMEM; 2376e4058b54SDavid Sterba path->reada = READA_FORWARD; 23775d4f98a2SYan Zheng 23785d4f98a2SYan Zheng reloc_root = root->reloc_root; 23795d4f98a2SYan Zheng root_item = &reloc_root->root_item; 23805d4f98a2SYan Zheng 23815d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 23825d4f98a2SYan Zheng level = btrfs_root_level(root_item); 238367439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 23845d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 23855d4f98a2SYan Zheng path->slots[level] = 0; 23865d4f98a2SYan Zheng } else { 23875d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 23885d4f98a2SYan Zheng 23895d4f98a2SYan Zheng level = root_item->drop_level; 23905d4f98a2SYan Zheng BUG_ON(level == 0); 23915d4f98a2SYan Zheng path->lowest_level = level; 23925d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 239333c66f43SYan Zheng path->lowest_level = 0; 23945d4f98a2SYan Zheng if (ret < 0) { 23955d4f98a2SYan Zheng btrfs_free_path(path); 23965d4f98a2SYan Zheng return ret; 23975d4f98a2SYan Zheng } 23985d4f98a2SYan Zheng 23995d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 24005d4f98a2SYan Zheng path->slots[level]); 24015d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 24025d4f98a2SYan Zheng 24035d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 24045d4f98a2SYan Zheng } 24055d4f98a2SYan Zheng 24060b246afaSJeff Mahoney min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 24075d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 24085d4f98a2SYan Zheng 24095d4f98a2SYan Zheng while (1) { 241008e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 241108e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 24123fd0a558SYan, Zheng if (ret) { 24139e6a0c52SJosef Bacik err = ret; 24149e6a0c52SJosef Bacik goto out; 24153fd0a558SYan, Zheng } 24169e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 24179e6a0c52SJosef Bacik if (IS_ERR(trans)) { 24189e6a0c52SJosef Bacik err = PTR_ERR(trans); 24199e6a0c52SJosef Bacik trans = NULL; 24209e6a0c52SJosef Bacik goto out; 24219e6a0c52SJosef Bacik } 24222abc726aSJosef Bacik 24232abc726aSJosef Bacik /* 24242abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 24252abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 24262abc726aSJosef Bacik * 24272abc726aSJosef Bacik * But that's ok, we started the trans handle on our 24282abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 24292abc726aSJosef Bacik * dirty list. At commit time we'll still call 24302abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 24312abc726aSJosef Bacik * appropriately. 24322abc726aSJosef Bacik */ 24332abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 24349e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 24353fd0a558SYan, Zheng 24363fd0a558SYan, Zheng replaced = 0; 24375d4f98a2SYan Zheng max_level = level; 24385d4f98a2SYan Zheng 24395d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 24405d4f98a2SYan Zheng if (ret < 0) { 24415d4f98a2SYan Zheng err = ret; 24425d4f98a2SYan Zheng goto out; 24435d4f98a2SYan Zheng } 24445d4f98a2SYan Zheng if (ret > 0) 24455d4f98a2SYan Zheng break; 24465d4f98a2SYan Zheng 24475d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 24485d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 24495d4f98a2SYan Zheng ret = 0; 24505d4f98a2SYan Zheng } else { 24513d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 24523fd0a558SYan, Zheng &next_key, level, max_level); 24535d4f98a2SYan Zheng } 24545d4f98a2SYan Zheng if (ret < 0) { 24555d4f98a2SYan Zheng err = ret; 24565d4f98a2SYan Zheng goto out; 24575d4f98a2SYan Zheng } 24585d4f98a2SYan Zheng 24595d4f98a2SYan Zheng if (ret > 0) { 24605d4f98a2SYan Zheng level = ret; 24615d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 24625d4f98a2SYan Zheng path->slots[level]); 24635d4f98a2SYan Zheng replaced = 1; 24645d4f98a2SYan Zheng } 24655d4f98a2SYan Zheng 24665d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 24675d4f98a2SYan Zheng if (ret > 0) 24685d4f98a2SYan Zheng break; 24695d4f98a2SYan Zheng 24705d4f98a2SYan Zheng BUG_ON(level == 0); 24715d4f98a2SYan Zheng /* 24725d4f98a2SYan Zheng * save the merging progress in the drop_progress. 24735d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 24745d4f98a2SYan Zheng */ 24755d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 24765d4f98a2SYan Zheng path->slots[level]); 24775d4f98a2SYan Zheng root_item->drop_level = level; 24785d4f98a2SYan Zheng 24793a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 24809e6a0c52SJosef Bacik trans = NULL; 24815d4f98a2SYan Zheng 24822ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 24835d4f98a2SYan Zheng 24845d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 24855d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 24865d4f98a2SYan Zheng } 24875d4f98a2SYan Zheng 24885d4f98a2SYan Zheng /* 24895d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 24905d4f98a2SYan Zheng * relocated and the block is tree root. 24915d4f98a2SYan Zheng */ 24925d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 24935d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 24945d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 24955d4f98a2SYan Zheng free_extent_buffer(leaf); 24965d4f98a2SYan Zheng if (ret < 0) 24975d4f98a2SYan Zheng err = ret; 24985d4f98a2SYan Zheng out: 24995d4f98a2SYan Zheng btrfs_free_path(path); 25005d4f98a2SYan Zheng 2501d2311e69SQu Wenruo if (err == 0) 2502d2311e69SQu Wenruo insert_dirty_subvol(trans, rc, root); 25035d4f98a2SYan Zheng 25049e6a0c52SJosef Bacik if (trans) 25053a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 25065d4f98a2SYan Zheng 25072ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 25085d4f98a2SYan Zheng 25095d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 25105d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 25115d4f98a2SYan Zheng 25125d4f98a2SYan Zheng return err; 25135d4f98a2SYan Zheng } 25145d4f98a2SYan Zheng 25153fd0a558SYan, Zheng static noinline_for_stack 25163fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 25175d4f98a2SYan Zheng { 25183fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 25190b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 25203fd0a558SYan, Zheng struct btrfs_root *reloc_root; 25215d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 25223fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 25233fd0a558SYan, Zheng u64 num_bytes = 0; 25243fd0a558SYan, Zheng int ret; 25253fd0a558SYan, Zheng 25260b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 25270b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 25283fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 25290b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 25307585717fSChris Mason 25313fd0a558SYan, Zheng again: 25323fd0a558SYan, Zheng if (!err) { 25333fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 253408e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 253508e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 25363fd0a558SYan, Zheng if (ret) 25373fd0a558SYan, Zheng err = ret; 25383fd0a558SYan, Zheng } 25393fd0a558SYan, Zheng 25407a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 25413612b495STsutomu Itoh if (IS_ERR(trans)) { 25423612b495STsutomu Itoh if (!err) 25432ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 254463f018beSNikolay Borisov num_bytes, NULL); 25453612b495STsutomu Itoh return PTR_ERR(trans); 25463612b495STsutomu Itoh } 25473fd0a558SYan, Zheng 25483fd0a558SYan, Zheng if (!err) { 25493fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 25503a45bb20SJeff Mahoney btrfs_end_transaction(trans); 25512ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 255263f018beSNikolay Borisov num_bytes, NULL); 25533fd0a558SYan, Zheng goto again; 25543fd0a558SYan, Zheng } 25553fd0a558SYan, Zheng } 25563fd0a558SYan, Zheng 25573fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 25583fd0a558SYan, Zheng 25593fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 25603fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 25613fd0a558SYan, Zheng struct btrfs_root, root_list); 25623fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 25633fd0a558SYan, Zheng 25640b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 25653fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 25663fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 25673fd0a558SYan, Zheng 25683fd0a558SYan, Zheng /* 25693fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 25703fd0a558SYan, Zheng * knows it should resumes merging 25713fd0a558SYan, Zheng */ 25723fd0a558SYan, Zheng if (!err) 25733fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 25743fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 25753fd0a558SYan, Zheng 25763fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 257700246528SJosef Bacik btrfs_put_root(root); 25783fd0a558SYan, Zheng } 25793fd0a558SYan, Zheng 25803fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 25813fd0a558SYan, Zheng 25823fd0a558SYan, Zheng if (!err) 25833a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 25843fd0a558SYan, Zheng else 25853a45bb20SJeff Mahoney btrfs_end_transaction(trans); 25863fd0a558SYan, Zheng return err; 25873fd0a558SYan, Zheng } 25883fd0a558SYan, Zheng 25893fd0a558SYan, Zheng static noinline_for_stack 2590aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 2591aca1bba6SLiu Bo { 2592aca1bba6SLiu Bo struct btrfs_root *reloc_root; 2593aca1bba6SLiu Bo 2594aca1bba6SLiu Bo while (!list_empty(list)) { 2595aca1bba6SLiu Bo reloc_root = list_entry(list->next, struct btrfs_root, 2596aca1bba6SLiu Bo root_list); 2597bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 2598aca1bba6SLiu Bo } 2599aca1bba6SLiu Bo } 2600aca1bba6SLiu Bo 2601aca1bba6SLiu Bo static noinline_for_stack 260294404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 26033fd0a558SYan, Zheng { 26040b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 26055d4f98a2SYan Zheng struct btrfs_root *root; 26065d4f98a2SYan Zheng struct btrfs_root *reloc_root; 26073fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 26083fd0a558SYan, Zheng int found = 0; 2609aca1bba6SLiu Bo int ret = 0; 26103fd0a558SYan, Zheng again: 26113fd0a558SYan, Zheng root = rc->extent_root; 26127585717fSChris Mason 26137585717fSChris Mason /* 26147585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 26157585717fSChris Mason * we have to make sure nobody is in the middle of 26167585717fSChris Mason * adding their roots to the list while we are 26177585717fSChris Mason * doing this splice 26187585717fSChris Mason */ 26190b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 26203fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 26210b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 26225d4f98a2SYan Zheng 26233fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 26243fd0a558SYan, Zheng found = 1; 26253fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 26263fd0a558SYan, Zheng struct btrfs_root, root_list); 26275d4f98a2SYan Zheng 26285d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 26290b246afaSJeff Mahoney root = read_fs_root(fs_info, 26305d4f98a2SYan Zheng reloc_root->root_key.offset); 26315d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 26325d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 26335d4f98a2SYan Zheng 26343fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 263500246528SJosef Bacik btrfs_put_root(root); 2636b37b39cdSJosef Bacik if (ret) { 263725e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 263825e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 263925e293c2SWang Shilong &reloc_roots); 2640aca1bba6SLiu Bo goto out; 2641b37b39cdSJosef Bacik } 26423fd0a558SYan, Zheng } else { 26433fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 264430d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 264530d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 264630d40577SQu Wenruo &rc->dirty_subvol_roots); 26473fd0a558SYan, Zheng } 26485d4f98a2SYan Zheng } 26495d4f98a2SYan Zheng 26503fd0a558SYan, Zheng if (found) { 26513fd0a558SYan, Zheng found = 0; 26523fd0a558SYan, Zheng goto again; 26535d4f98a2SYan Zheng } 2654aca1bba6SLiu Bo out: 2655aca1bba6SLiu Bo if (ret) { 26560b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 2657aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 2658aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 2659467bb1d2SWang Shilong 2660467bb1d2SWang Shilong /* new reloc root may be added */ 26610b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 2662467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 26630b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 2664467bb1d2SWang Shilong if (!list_empty(&reloc_roots)) 2665467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 2666aca1bba6SLiu Bo } 2667aca1bba6SLiu Bo 26687b7b7431SJosef Bacik /* 26697b7b7431SJosef Bacik * We used to have 26707b7b7431SJosef Bacik * 26717b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 26727b7b7431SJosef Bacik * 26737b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 26747b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 26757b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 26767b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 26777b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 26787b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 26797b7b7431SJosef Bacik * will be cleaned up on unmount. 26807b7b7431SJosef Bacik * 26817b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 26827b7b7431SJosef Bacik */ 26835d4f98a2SYan Zheng } 26845d4f98a2SYan Zheng 26855d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 26865d4f98a2SYan Zheng { 26875d4f98a2SYan Zheng struct tree_block *block; 26885d4f98a2SYan Zheng struct rb_node *rb_node; 26895d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 26905d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 26915d4f98a2SYan Zheng rb_erase(rb_node, blocks); 26925d4f98a2SYan Zheng kfree(block); 26935d4f98a2SYan Zheng } 26945d4f98a2SYan Zheng } 26955d4f98a2SYan Zheng 26965d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 26975d4f98a2SYan Zheng struct btrfs_root *reloc_root) 26985d4f98a2SYan Zheng { 26990b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 27005d4f98a2SYan Zheng struct btrfs_root *root; 2701442b1ac5SJosef Bacik int ret; 27025d4f98a2SYan Zheng 27035d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 27045d4f98a2SYan Zheng return 0; 27055d4f98a2SYan Zheng 27060b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 27075d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 27085d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 2709442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 271000246528SJosef Bacik btrfs_put_root(root); 27115d4f98a2SYan Zheng 2712442b1ac5SJosef Bacik return ret; 27135d4f98a2SYan Zheng } 27145d4f98a2SYan Zheng 27153fd0a558SYan, Zheng static noinline_for_stack 27163fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 27173fd0a558SYan, Zheng struct reloc_control *rc, 2718a26195a5SQu Wenruo struct btrfs_backref_node *node, 2719a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 27205d4f98a2SYan Zheng { 2721a26195a5SQu Wenruo struct btrfs_backref_node *next; 27225d4f98a2SYan Zheng struct btrfs_root *root; 27233fd0a558SYan, Zheng int index = 0; 27243fd0a558SYan, Zheng 27255d4f98a2SYan Zheng next = node; 27265d4f98a2SYan Zheng while (1) { 27275d4f98a2SYan Zheng cond_resched(); 27285d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 27295d4f98a2SYan Zheng root = next->root; 27303fd0a558SYan, Zheng BUG_ON(!root); 273127cdeb70SMiao Xie BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state)); 27325d4f98a2SYan Zheng 27335d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 27345d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 27355d4f98a2SYan Zheng break; 27365d4f98a2SYan Zheng } 27375d4f98a2SYan Zheng 27385d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 27393fd0a558SYan, Zheng root = root->reloc_root; 27403fd0a558SYan, Zheng 27413fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 27423fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 27433fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 27443fd0a558SYan, Zheng next->new_bytenr = root->node->start; 274500246528SJosef Bacik btrfs_put_root(next->root); 274600246528SJosef Bacik next->root = btrfs_grab_root(root); 27470b530bc5SJosef Bacik ASSERT(next->root); 27483fd0a558SYan, Zheng list_add_tail(&next->list, 27493fd0a558SYan, Zheng &rc->backref_cache.changed); 27509569cc20SQu Wenruo mark_block_processed(rc, next); 27515d4f98a2SYan Zheng break; 27525d4f98a2SYan Zheng } 27535d4f98a2SYan Zheng 27543fd0a558SYan, Zheng WARN_ON(1); 27555d4f98a2SYan Zheng root = NULL; 27565d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 27575d4f98a2SYan Zheng if (!next || next->level <= node->level) 27585d4f98a2SYan Zheng break; 27595d4f98a2SYan Zheng } 27603fd0a558SYan, Zheng if (!root) 27613fd0a558SYan, Zheng return NULL; 27625d4f98a2SYan Zheng 27633fd0a558SYan, Zheng next = node; 27643fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 27653fd0a558SYan, Zheng while (1) { 27663fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 27673fd0a558SYan, Zheng if (--index < 0) 27683fd0a558SYan, Zheng break; 27693fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 27703fd0a558SYan, Zheng } 27715d4f98a2SYan Zheng return root; 27725d4f98a2SYan Zheng } 27735d4f98a2SYan Zheng 27743fd0a558SYan, Zheng /* 27753fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 27763fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 27773fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 27783fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 27793fd0a558SYan, Zheng */ 27805d4f98a2SYan Zheng static noinline_for_stack 2781a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 27825d4f98a2SYan Zheng { 2783a26195a5SQu Wenruo struct btrfs_backref_node *next; 27843fd0a558SYan, Zheng struct btrfs_root *root; 27853fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2786a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 27873fd0a558SYan, Zheng int index = 0; 27883fd0a558SYan, Zheng 27893fd0a558SYan, Zheng next = node; 27903fd0a558SYan, Zheng while (1) { 27913fd0a558SYan, Zheng cond_resched(); 27923fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 27933fd0a558SYan, Zheng root = next->root; 27943fd0a558SYan, Zheng BUG_ON(!root); 27953fd0a558SYan, Zheng 279625985edcSLucas De Marchi /* no other choice for non-references counted tree */ 279727cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 27983fd0a558SYan, Zheng return root; 27993fd0a558SYan, Zheng 28003fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 28013fd0a558SYan, Zheng fs_root = root; 28023fd0a558SYan, Zheng 28033fd0a558SYan, Zheng if (next != node) 28043fd0a558SYan, Zheng return NULL; 28053fd0a558SYan, Zheng 28063fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 28073fd0a558SYan, Zheng if (!next || next->level <= node->level) 28083fd0a558SYan, Zheng break; 28093fd0a558SYan, Zheng } 28103fd0a558SYan, Zheng 28113fd0a558SYan, Zheng if (!fs_root) 28123fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 28133fd0a558SYan, Zheng return fs_root; 28145d4f98a2SYan Zheng } 28155d4f98a2SYan Zheng 28165d4f98a2SYan Zheng static noinline_for_stack 28173fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2818a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 28195d4f98a2SYan Zheng { 28200b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2821a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2822a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2823a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 28243fd0a558SYan, Zheng u64 num_bytes = 0; 28253fd0a558SYan, Zheng int index = 0; 28265d4f98a2SYan Zheng 28273fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 28283fd0a558SYan, Zheng 28293fd0a558SYan, Zheng while (next) { 28303fd0a558SYan, Zheng cond_resched(); 28315d4f98a2SYan Zheng while (1) { 28323fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 28335d4f98a2SYan Zheng break; 28345d4f98a2SYan Zheng 28350b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 28363fd0a558SYan, Zheng 28373fd0a558SYan, Zheng if (list_empty(&next->upper)) 28383fd0a558SYan, Zheng break; 28393fd0a558SYan, Zheng 28403fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2841a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 28423fd0a558SYan, Zheng edges[index++] = edge; 28433fd0a558SYan, Zheng next = edge->node[UPPER]; 28445d4f98a2SYan Zheng } 28453fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 28463fd0a558SYan, Zheng } 28473fd0a558SYan, Zheng return num_bytes; 28483fd0a558SYan, Zheng } 28493fd0a558SYan, Zheng 28503fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 28513fd0a558SYan, Zheng struct reloc_control *rc, 2852a26195a5SQu Wenruo struct btrfs_backref_node *node) 28533fd0a558SYan, Zheng { 28543fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2855da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 28563fd0a558SYan, Zheng u64 num_bytes; 28573fd0a558SYan, Zheng int ret; 28580647bf56SWang Shilong u64 tmp; 28593fd0a558SYan, Zheng 28603fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 28613fd0a558SYan, Zheng 28623fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 28630647bf56SWang Shilong rc->reserved_bytes += num_bytes; 28648ca17f0fSJosef Bacik 28658ca17f0fSJosef Bacik /* 28668ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 28678ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 28688ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 28698ca17f0fSJosef Bacik */ 28700647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 28718ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 28723fd0a558SYan, Zheng if (ret) { 2873da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 28740647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 28750647bf56SWang Shilong tmp <<= 1; 28760647bf56SWang Shilong /* 28770647bf56SWang Shilong * only one thread can access block_rsv at this point, 28780647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 28790647bf56SWang Shilong * we expand more reservation size here to allow enough 288052042d8eSAndrea Gelmini * space for relocation and we will return earlier in 28810647bf56SWang Shilong * enospc case. 28820647bf56SWang Shilong */ 2883da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 28840647bf56SWang Shilong RELOCATION_RESERVED_NODES; 28858ca17f0fSJosef Bacik return -EAGAIN; 28863fd0a558SYan, Zheng } 28873fd0a558SYan, Zheng 28883fd0a558SYan, Zheng return 0; 28893fd0a558SYan, Zheng } 28903fd0a558SYan, Zheng 28915d4f98a2SYan Zheng /* 28925d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 28935d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 28945d4f98a2SYan Zheng * 28955d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 28965d4f98a2SYan Zheng * in that case this function just updates pointers. 28975d4f98a2SYan Zheng */ 28985d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 28993fd0a558SYan, Zheng struct reloc_control *rc, 2900a26195a5SQu Wenruo struct btrfs_backref_node *node, 29015d4f98a2SYan Zheng struct btrfs_key *key, 29025d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 29035d4f98a2SYan Zheng { 29042ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2905a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2906a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2907a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 29085d4f98a2SYan Zheng struct btrfs_root *root; 29095d4f98a2SYan Zheng struct extent_buffer *eb; 29105d4f98a2SYan Zheng u32 blocksize; 29115d4f98a2SYan Zheng u64 bytenr; 29125d4f98a2SYan Zheng u64 generation; 29135d4f98a2SYan Zheng int slot; 29145d4f98a2SYan Zheng int ret; 29155d4f98a2SYan Zheng int err = 0; 29165d4f98a2SYan Zheng 29175d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 29185d4f98a2SYan Zheng 29195d4f98a2SYan Zheng path->lowest_level = node->level + 1; 29203fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 29215d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 2922581c1760SQu Wenruo struct btrfs_key first_key; 292382fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2924581c1760SQu Wenruo 29255d4f98a2SYan Zheng cond_resched(); 29265d4f98a2SYan Zheng 29275d4f98a2SYan Zheng upper = edge->node[UPPER]; 2928dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 29293fd0a558SYan, Zheng BUG_ON(!root); 29305d4f98a2SYan Zheng 29313fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 29323fd0a558SYan, Zheng if (!lowest) { 29333fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 29343fd0a558SYan, Zheng upper->level, &slot); 2935cbca7d59SFilipe Manana if (ret < 0) { 2936cbca7d59SFilipe Manana err = ret; 2937cbca7d59SFilipe Manana goto next; 2938cbca7d59SFilipe Manana } 29393fd0a558SYan, Zheng BUG_ON(ret); 29403fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 29413fd0a558SYan, Zheng if (node->eb->start == bytenr) 29423fd0a558SYan, Zheng goto next; 29433fd0a558SYan, Zheng } 29445d4f98a2SYan Zheng drop_node_buffer(upper); 29453fd0a558SYan, Zheng } 29465d4f98a2SYan Zheng 29475d4f98a2SYan Zheng if (!upper->eb) { 29485d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 29493561b9dbSLiu Bo if (ret) { 29503561b9dbSLiu Bo if (ret < 0) 29515d4f98a2SYan Zheng err = ret; 29523561b9dbSLiu Bo else 29533561b9dbSLiu Bo err = -ENOENT; 29543561b9dbSLiu Bo 29553561b9dbSLiu Bo btrfs_release_path(path); 29565d4f98a2SYan Zheng break; 29575d4f98a2SYan Zheng } 29585d4f98a2SYan Zheng 29593fd0a558SYan, Zheng if (!upper->eb) { 29603fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 29613fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 29623fd0a558SYan, Zheng } else { 29633fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 29643fd0a558SYan, Zheng } 29653fd0a558SYan, Zheng 29663fd0a558SYan, Zheng upper->locked = 1; 29673fd0a558SYan, Zheng path->locks[upper->level] = 0; 29683fd0a558SYan, Zheng 29695d4f98a2SYan Zheng slot = path->slots[upper->level]; 2970b3b4aa74SDavid Sterba btrfs_release_path(path); 29715d4f98a2SYan Zheng } else { 29725d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 29735d4f98a2SYan Zheng &slot); 2974cbca7d59SFilipe Manana if (ret < 0) { 2975cbca7d59SFilipe Manana err = ret; 2976cbca7d59SFilipe Manana goto next; 2977cbca7d59SFilipe Manana } 29785d4f98a2SYan Zheng BUG_ON(ret); 29795d4f98a2SYan Zheng } 29805d4f98a2SYan Zheng 29815d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 29823fd0a558SYan, Zheng if (lowest) { 29834547f4d8SLiu Bo if (bytenr != node->bytenr) { 29844547f4d8SLiu Bo btrfs_err(root->fs_info, 29854547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 29864547f4d8SLiu Bo bytenr, node->bytenr, slot, 29874547f4d8SLiu Bo upper->eb->start); 29884547f4d8SLiu Bo err = -EIO; 29894547f4d8SLiu Bo goto next; 29904547f4d8SLiu Bo } 29915d4f98a2SYan Zheng } else { 29923fd0a558SYan, Zheng if (node->eb->start == bytenr) 29933fd0a558SYan, Zheng goto next; 29945d4f98a2SYan Zheng } 29955d4f98a2SYan Zheng 2996da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 29975d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 2998581c1760SQu Wenruo btrfs_node_key_to_cpu(upper->eb, &first_key, slot); 2999581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, generation, 3000581c1760SQu Wenruo upper->level - 1, &first_key); 300164c043deSLiu Bo if (IS_ERR(eb)) { 300264c043deSLiu Bo err = PTR_ERR(eb); 300364c043deSLiu Bo goto next; 300464c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 3005416bc658SJosef Bacik free_extent_buffer(eb); 300697d9a8a4STsutomu Itoh err = -EIO; 300797d9a8a4STsutomu Itoh goto next; 300897d9a8a4STsutomu Itoh } 30095d4f98a2SYan Zheng btrfs_tree_lock(eb); 30108bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 30115d4f98a2SYan Zheng 30125d4f98a2SYan Zheng if (!node->eb) { 30135d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 30145d4f98a2SYan Zheng slot, &eb); 30153fd0a558SYan, Zheng btrfs_tree_unlock(eb); 30163fd0a558SYan, Zheng free_extent_buffer(eb); 30175d4f98a2SYan Zheng if (ret < 0) { 30185d4f98a2SYan Zheng err = ret; 30193fd0a558SYan, Zheng goto next; 30205d4f98a2SYan Zheng } 30213fd0a558SYan, Zheng BUG_ON(node->eb != eb); 30225d4f98a2SYan Zheng } else { 30235d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 30245d4f98a2SYan Zheng node->eb->start); 30255d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 30265d4f98a2SYan Zheng trans->transid); 30275d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 30285d4f98a2SYan Zheng 302982fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 30305d4f98a2SYan Zheng node->eb->start, blocksize, 303182fa113fSQu Wenruo upper->eb->start); 303282fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 303382fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 303482fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 303582fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 30365d4f98a2SYan Zheng BUG_ON(ret); 30375d4f98a2SYan Zheng 30385d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 30395d4f98a2SYan Zheng BUG_ON(ret); 30405d4f98a2SYan Zheng } 30413fd0a558SYan, Zheng next: 30423fd0a558SYan, Zheng if (!upper->pending) 30433fd0a558SYan, Zheng drop_node_buffer(upper); 30443fd0a558SYan, Zheng else 30453fd0a558SYan, Zheng unlock_node_buffer(upper); 30463fd0a558SYan, Zheng if (err) 30473fd0a558SYan, Zheng break; 30485d4f98a2SYan Zheng } 30493fd0a558SYan, Zheng 30503fd0a558SYan, Zheng if (!err && node->pending) { 30513fd0a558SYan, Zheng drop_node_buffer(node); 30523fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 30533fd0a558SYan, Zheng node->pending = 0; 30545d4f98a2SYan Zheng } 30553fd0a558SYan, Zheng 30565d4f98a2SYan Zheng path->lowest_level = 0; 30573fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 30585d4f98a2SYan Zheng return err; 30595d4f98a2SYan Zheng } 30605d4f98a2SYan Zheng 30615d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 30623fd0a558SYan, Zheng struct reloc_control *rc, 3063a26195a5SQu Wenruo struct btrfs_backref_node *node, 30645d4f98a2SYan Zheng struct btrfs_path *path) 30655d4f98a2SYan Zheng { 30665d4f98a2SYan Zheng struct btrfs_key key; 30675d4f98a2SYan Zheng 30685d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 30693fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 30705d4f98a2SYan Zheng } 30715d4f98a2SYan Zheng 30725d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 30733fd0a558SYan, Zheng struct reloc_control *rc, 30743fd0a558SYan, Zheng struct btrfs_path *path, int err) 30755d4f98a2SYan Zheng { 30763fd0a558SYan, Zheng LIST_HEAD(list); 3077a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 3078a26195a5SQu Wenruo struct btrfs_backref_node *node; 30795d4f98a2SYan Zheng int level; 30805d4f98a2SYan Zheng int ret; 30815d4f98a2SYan Zheng 30825d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 30835d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 30845d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 3085a26195a5SQu Wenruo struct btrfs_backref_node, list); 30863fd0a558SYan, Zheng list_move_tail(&node->list, &list); 30873fd0a558SYan, Zheng BUG_ON(!node->pending); 30885d4f98a2SYan Zheng 30893fd0a558SYan, Zheng if (!err) { 30903fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 30915d4f98a2SYan Zheng if (ret < 0) 30925d4f98a2SYan Zheng err = ret; 30935d4f98a2SYan Zheng } 30945d4f98a2SYan Zheng } 30953fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 30963fd0a558SYan, Zheng } 30975d4f98a2SYan Zheng return err; 30985d4f98a2SYan Zheng } 30995d4f98a2SYan Zheng 31005d4f98a2SYan Zheng /* 31015d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 31025d4f98a2SYan Zheng * as processed. 31035d4f98a2SYan Zheng */ 31045d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 3105a26195a5SQu Wenruo struct btrfs_backref_node *node) 31065d4f98a2SYan Zheng { 3107a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 3108a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 3109a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 31105d4f98a2SYan Zheng int index = 0; 31115d4f98a2SYan Zheng 31125d4f98a2SYan Zheng while (next) { 31135d4f98a2SYan Zheng cond_resched(); 31145d4f98a2SYan Zheng while (1) { 31155d4f98a2SYan Zheng if (next->processed) 31165d4f98a2SYan Zheng break; 31175d4f98a2SYan Zheng 31189569cc20SQu Wenruo mark_block_processed(rc, next); 31195d4f98a2SYan Zheng 31205d4f98a2SYan Zheng if (list_empty(&next->upper)) 31215d4f98a2SYan Zheng break; 31225d4f98a2SYan Zheng 31235d4f98a2SYan Zheng edge = list_entry(next->upper.next, 3124a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 31255d4f98a2SYan Zheng edges[index++] = edge; 31265d4f98a2SYan Zheng next = edge->node[UPPER]; 31275d4f98a2SYan Zheng } 31285d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 31295d4f98a2SYan Zheng } 31305d4f98a2SYan Zheng } 31315d4f98a2SYan Zheng 31327476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 31335d4f98a2SYan Zheng { 3134da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 31357476dfdaSDavid Sterba 31365d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 31379655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 31385d4f98a2SYan Zheng return 1; 31395d4f98a2SYan Zheng return 0; 31405d4f98a2SYan Zheng } 31415d4f98a2SYan Zheng 31422ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 31435d4f98a2SYan Zheng struct tree_block *block) 31445d4f98a2SYan Zheng { 31455d4f98a2SYan Zheng struct extent_buffer *eb; 31465d4f98a2SYan Zheng 3147581c1760SQu Wenruo eb = read_tree_block(fs_info, block->bytenr, block->key.offset, 3148581c1760SQu Wenruo block->level, NULL); 314964c043deSLiu Bo if (IS_ERR(eb)) { 315064c043deSLiu Bo return PTR_ERR(eb); 315164c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 3152416bc658SJosef Bacik free_extent_buffer(eb); 3153416bc658SJosef Bacik return -EIO; 3154416bc658SJosef Bacik } 31555d4f98a2SYan Zheng if (block->level == 0) 31565d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 31575d4f98a2SYan Zheng else 31585d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 31595d4f98a2SYan Zheng free_extent_buffer(eb); 31605d4f98a2SYan Zheng block->key_ready = 1; 31615d4f98a2SYan Zheng return 0; 31625d4f98a2SYan Zheng } 31635d4f98a2SYan Zheng 31645d4f98a2SYan Zheng /* 31655d4f98a2SYan Zheng * helper function to relocate a tree block 31665d4f98a2SYan Zheng */ 31675d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 31685d4f98a2SYan Zheng struct reloc_control *rc, 3169a26195a5SQu Wenruo struct btrfs_backref_node *node, 31705d4f98a2SYan Zheng struct btrfs_key *key, 31715d4f98a2SYan Zheng struct btrfs_path *path) 31725d4f98a2SYan Zheng { 31735d4f98a2SYan Zheng struct btrfs_root *root; 31743fd0a558SYan, Zheng int ret = 0; 31755d4f98a2SYan Zheng 31763fd0a558SYan, Zheng if (!node) 31775d4f98a2SYan Zheng return 0; 31783fd0a558SYan, Zheng 31795f6b2e5cSJosef Bacik /* 31805f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 31815f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 31825f6b2e5cSJosef Bacik */ 31835f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 31845f6b2e5cSJosef Bacik if (ret) 31855f6b2e5cSJosef Bacik goto out; 31865f6b2e5cSJosef Bacik 31873fd0a558SYan, Zheng BUG_ON(node->processed); 3188147d256eSZhaolei root = select_one_root(node); 31893fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 31903fd0a558SYan, Zheng update_processed_blocks(rc, node); 31913fd0a558SYan, Zheng goto out; 31925d4f98a2SYan Zheng } 31935d4f98a2SYan Zheng 31943fd0a558SYan, Zheng if (root) { 319527cdeb70SMiao Xie if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) { 31963fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 31973fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 31983fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 31993fd0a558SYan, Zheng root = root->reloc_root; 32003fd0a558SYan, Zheng node->new_bytenr = root->node->start; 320100246528SJosef Bacik btrfs_put_root(node->root); 320200246528SJosef Bacik node->root = btrfs_grab_root(root); 32030b530bc5SJosef Bacik ASSERT(node->root); 32043fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 32053fd0a558SYan, Zheng } else { 32065d4f98a2SYan Zheng path->lowest_level = node->level; 32075d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 3208b3b4aa74SDavid Sterba btrfs_release_path(path); 32093fd0a558SYan, Zheng if (ret > 0) 32105d4f98a2SYan Zheng ret = 0; 32113fd0a558SYan, Zheng } 32123fd0a558SYan, Zheng if (!ret) 32133fd0a558SYan, Zheng update_processed_blocks(rc, node); 32143fd0a558SYan, Zheng } else { 32153fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 32163fd0a558SYan, Zheng } 32175d4f98a2SYan Zheng out: 32180647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 32193fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 32205d4f98a2SYan Zheng return ret; 32215d4f98a2SYan Zheng } 32225d4f98a2SYan Zheng 32235d4f98a2SYan Zheng /* 32245d4f98a2SYan Zheng * relocate a list of blocks 32255d4f98a2SYan Zheng */ 32265d4f98a2SYan Zheng static noinline_for_stack 32275d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 32285d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 32295d4f98a2SYan Zheng { 32302ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 3231a26195a5SQu Wenruo struct btrfs_backref_node *node; 32325d4f98a2SYan Zheng struct btrfs_path *path; 32335d4f98a2SYan Zheng struct tree_block *block; 323498ff7b94SQu Wenruo struct tree_block *next; 32355d4f98a2SYan Zheng int ret; 32365d4f98a2SYan Zheng int err = 0; 32375d4f98a2SYan Zheng 32385d4f98a2SYan Zheng path = btrfs_alloc_path(); 3239e1a12670SLiu Bo if (!path) { 3240e1a12670SLiu Bo err = -ENOMEM; 324134c2b290SDavid Sterba goto out_free_blocks; 3242e1a12670SLiu Bo } 32435d4f98a2SYan Zheng 324498ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 324598ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 32465d4f98a2SYan Zheng if (!block->key_ready) 32472ff7e61eSJeff Mahoney readahead_tree_block(fs_info, block->bytenr); 32485d4f98a2SYan Zheng } 32495d4f98a2SYan Zheng 325098ff7b94SQu Wenruo /* Get first keys */ 325198ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 325234c2b290SDavid Sterba if (!block->key_ready) { 32532ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 325434c2b290SDavid Sterba if (err) 325534c2b290SDavid Sterba goto out_free_path; 325634c2b290SDavid Sterba } 32575d4f98a2SYan Zheng } 32585d4f98a2SYan Zheng 325998ff7b94SQu Wenruo /* Do tree relocation */ 326098ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 32613fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 32625d4f98a2SYan Zheng block->level, block->bytenr); 32635d4f98a2SYan Zheng if (IS_ERR(node)) { 32645d4f98a2SYan Zheng err = PTR_ERR(node); 32655d4f98a2SYan Zheng goto out; 32665d4f98a2SYan Zheng } 32675d4f98a2SYan Zheng 32685d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 32695d4f98a2SYan Zheng path); 32705d4f98a2SYan Zheng if (ret < 0) { 32715d4f98a2SYan Zheng err = ret; 327250dbbb71SJosef Bacik break; 32735d4f98a2SYan Zheng } 32745d4f98a2SYan Zheng } 32755d4f98a2SYan Zheng out: 32763fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 32775d4f98a2SYan Zheng 327834c2b290SDavid Sterba out_free_path: 32795d4f98a2SYan Zheng btrfs_free_path(path); 328034c2b290SDavid Sterba out_free_blocks: 3281e1a12670SLiu Bo free_block_list(blocks); 32825d4f98a2SYan Zheng return err; 32835d4f98a2SYan Zheng } 32845d4f98a2SYan Zheng 32855d4f98a2SYan Zheng static noinline_for_stack 3286efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 3287efa56464SYan, Zheng struct file_extent_cluster *cluster) 3288efa56464SYan, Zheng { 3289efa56464SYan, Zheng u64 alloc_hint = 0; 3290efa56464SYan, Zheng u64 start; 3291efa56464SYan, Zheng u64 end; 3292efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 3293efa56464SYan, Zheng u64 num_bytes; 3294efa56464SYan, Zheng int nr = 0; 3295efa56464SYan, Zheng int ret = 0; 3296dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 3297dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 329818513091SWang Xiaoguang u64 cur_offset; 3299364ecf36SQu Wenruo struct extent_changeset *data_reserved = NULL; 3300efa56464SYan, Zheng 3301efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 33025955102cSAl Viro inode_lock(inode); 3303efa56464SYan, Zheng 3304364ecf36SQu Wenruo ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start, 3305dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 3306efa56464SYan, Zheng if (ret) 3307efa56464SYan, Zheng goto out; 3308efa56464SYan, Zheng 330918513091SWang Xiaoguang cur_offset = prealloc_start; 3310efa56464SYan, Zheng while (nr < cluster->nr) { 3311efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 3312efa56464SYan, Zheng if (nr + 1 < cluster->nr) 3313efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 3314efa56464SYan, Zheng else 3315efa56464SYan, Zheng end = cluster->end - offset; 3316efa56464SYan, Zheng 3317d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 3318efa56464SYan, Zheng num_bytes = end + 1 - start; 331918513091SWang Xiaoguang if (cur_offset < start) 3320bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3321bc42bda2SQu Wenruo cur_offset, start - cur_offset); 3322efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 3323efa56464SYan, Zheng num_bytes, num_bytes, 3324efa56464SYan, Zheng end + 1, &alloc_hint); 332518513091SWang Xiaoguang cur_offset = end + 1; 3326d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 3327efa56464SYan, Zheng if (ret) 3328efa56464SYan, Zheng break; 3329efa56464SYan, Zheng nr++; 3330efa56464SYan, Zheng } 333118513091SWang Xiaoguang if (cur_offset < prealloc_end) 3332bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3333bc42bda2SQu Wenruo cur_offset, prealloc_end + 1 - cur_offset); 3334efa56464SYan, Zheng out: 33355955102cSAl Viro inode_unlock(inode); 3336364ecf36SQu Wenruo extent_changeset_free(data_reserved); 3337efa56464SYan, Zheng return ret; 3338efa56464SYan, Zheng } 3339efa56464SYan, Zheng 3340efa56464SYan, Zheng static noinline_for_stack 33410257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 33420257bb82SYan, Zheng u64 block_start) 33435d4f98a2SYan Zheng { 33445d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 33455d4f98a2SYan Zheng struct extent_map *em; 33460257bb82SYan, Zheng int ret = 0; 33475d4f98a2SYan Zheng 3348172ddd60SDavid Sterba em = alloc_extent_map(); 33490257bb82SYan, Zheng if (!em) 33500257bb82SYan, Zheng return -ENOMEM; 33510257bb82SYan, Zheng 33525d4f98a2SYan Zheng em->start = start; 33530257bb82SYan, Zheng em->len = end + 1 - start; 33540257bb82SYan, Zheng em->block_len = em->len; 33550257bb82SYan, Zheng em->block_start = block_start; 33565d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 33575d4f98a2SYan Zheng 3358d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 33595d4f98a2SYan Zheng while (1) { 3360890871beSChris Mason write_lock(&em_tree->lock); 336109a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 3362890871beSChris Mason write_unlock(&em_tree->lock); 33635d4f98a2SYan Zheng if (ret != -EEXIST) { 33645d4f98a2SYan Zheng free_extent_map(em); 33655d4f98a2SYan Zheng break; 33665d4f98a2SYan Zheng } 3367dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 33685d4f98a2SYan Zheng } 3369d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 33700257bb82SYan, Zheng return ret; 33710257bb82SYan, Zheng } 33725d4f98a2SYan Zheng 3373726a3421SQu Wenruo /* 3374726a3421SQu Wenruo * Allow error injection to test balance cancellation 3375726a3421SQu Wenruo */ 3376726a3421SQu Wenruo int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 3377726a3421SQu Wenruo { 3378726a3421SQu Wenruo return atomic_read(&fs_info->balance_cancel_req); 3379726a3421SQu Wenruo } 3380726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 3381726a3421SQu Wenruo 33820257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 33830257bb82SYan, Zheng struct file_extent_cluster *cluster) 33840257bb82SYan, Zheng { 33852ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 33860257bb82SYan, Zheng u64 page_start; 33870257bb82SYan, Zheng u64 page_end; 33880257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 33890257bb82SYan, Zheng unsigned long index; 33900257bb82SYan, Zheng unsigned long last_index; 33910257bb82SYan, Zheng struct page *page; 33920257bb82SYan, Zheng struct file_ra_state *ra; 33933b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 33940257bb82SYan, Zheng int nr = 0; 33950257bb82SYan, Zheng int ret = 0; 33960257bb82SYan, Zheng 33970257bb82SYan, Zheng if (!cluster->nr) 33980257bb82SYan, Zheng return 0; 33990257bb82SYan, Zheng 34000257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 34010257bb82SYan, Zheng if (!ra) 34020257bb82SYan, Zheng return -ENOMEM; 34030257bb82SYan, Zheng 3404efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 34050257bb82SYan, Zheng if (ret) 3406efa56464SYan, Zheng goto out; 34070257bb82SYan, Zheng 34080257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 34090257bb82SYan, Zheng 3410efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 3411efa56464SYan, Zheng cluster->end - offset, cluster->start); 3412efa56464SYan, Zheng if (ret) 3413efa56464SYan, Zheng goto out; 3414efa56464SYan, Zheng 341509cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 341609cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 34170257bb82SYan, Zheng while (index <= last_index) { 34189f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 34199f3db423SNikolay Borisov PAGE_SIZE); 3420efa56464SYan, Zheng if (ret) 3421efa56464SYan, Zheng goto out; 3422efa56464SYan, Zheng 34230257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 34240257bb82SYan, Zheng if (!page) { 34250257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 34260257bb82SYan, Zheng ra, NULL, index, 34270257bb82SYan, Zheng last_index + 1 - index); 3428a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 34293b16a4e3SJosef Bacik mask); 34300257bb82SYan, Zheng if (!page) { 3431691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 343243b18595SQu Wenruo PAGE_SIZE, true); 343344db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 34348702ba93SQu Wenruo PAGE_SIZE); 34350257bb82SYan, Zheng ret = -ENOMEM; 3436efa56464SYan, Zheng goto out; 34370257bb82SYan, Zheng } 34380257bb82SYan, Zheng } 34390257bb82SYan, Zheng 34400257bb82SYan, Zheng if (PageReadahead(page)) { 34410257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 34420257bb82SYan, Zheng ra, NULL, page, index, 34430257bb82SYan, Zheng last_index + 1 - index); 34440257bb82SYan, Zheng } 34450257bb82SYan, Zheng 34460257bb82SYan, Zheng if (!PageUptodate(page)) { 34470257bb82SYan, Zheng btrfs_readpage(NULL, page); 34480257bb82SYan, Zheng lock_page(page); 34490257bb82SYan, Zheng if (!PageUptodate(page)) { 34500257bb82SYan, Zheng unlock_page(page); 345109cbfeafSKirill A. Shutemov put_page(page); 3452691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 345343b18595SQu Wenruo PAGE_SIZE, true); 34548b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 34558702ba93SQu Wenruo PAGE_SIZE); 34560257bb82SYan, Zheng ret = -EIO; 3457efa56464SYan, Zheng goto out; 34580257bb82SYan, Zheng } 34590257bb82SYan, Zheng } 34600257bb82SYan, Zheng 34614eee4fa4SMiao Xie page_start = page_offset(page); 346209cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 34630257bb82SYan, Zheng 3464d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 34650257bb82SYan, Zheng 34660257bb82SYan, Zheng set_page_extent_mapped(page); 34670257bb82SYan, Zheng 34680257bb82SYan, Zheng if (nr < cluster->nr && 34690257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 34700257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 34710257bb82SYan, Zheng page_start, page_end, 3472ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 34730257bb82SYan, Zheng nr++; 34740257bb82SYan, Zheng } 34750257bb82SYan, Zheng 3476765f3cebSNikolay Borisov ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0, 3477330a5827SNikolay Borisov NULL); 3478765f3cebSNikolay Borisov if (ret) { 3479765f3cebSNikolay Borisov unlock_page(page); 3480765f3cebSNikolay Borisov put_page(page); 3481765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 348243b18595SQu Wenruo PAGE_SIZE, true); 3483765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 34848702ba93SQu Wenruo PAGE_SIZE); 3485765f3cebSNikolay Borisov 3486765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 3487765f3cebSNikolay Borisov page_start, page_end, 3488765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 3489765f3cebSNikolay Borisov goto out; 3490765f3cebSNikolay Borisov 3491765f3cebSNikolay Borisov } 34920257bb82SYan, Zheng set_page_dirty(page); 34930257bb82SYan, Zheng 34940257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 3495d0082371SJeff Mahoney page_start, page_end); 34960257bb82SYan, Zheng unlock_page(page); 349709cbfeafSKirill A. Shutemov put_page(page); 34980257bb82SYan, Zheng 34990257bb82SYan, Zheng index++; 35008702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 3501efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 35022ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 35037f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 35047f913c7cSQu Wenruo ret = -ECANCELED; 35057f913c7cSQu Wenruo goto out; 35067f913c7cSQu Wenruo } 35070257bb82SYan, Zheng } 35080257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 3509efa56464SYan, Zheng out: 35100257bb82SYan, Zheng kfree(ra); 35110257bb82SYan, Zheng return ret; 35120257bb82SYan, Zheng } 35130257bb82SYan, Zheng 35140257bb82SYan, Zheng static noinline_for_stack 35150257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 35160257bb82SYan, Zheng struct file_extent_cluster *cluster) 35170257bb82SYan, Zheng { 35180257bb82SYan, Zheng int ret; 35190257bb82SYan, Zheng 35200257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 35210257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 35220257bb82SYan, Zheng if (ret) 35230257bb82SYan, Zheng return ret; 35240257bb82SYan, Zheng cluster->nr = 0; 35250257bb82SYan, Zheng } 35260257bb82SYan, Zheng 35270257bb82SYan, Zheng if (!cluster->nr) 35280257bb82SYan, Zheng cluster->start = extent_key->objectid; 35290257bb82SYan, Zheng else 35300257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 35310257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 35320257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 35330257bb82SYan, Zheng cluster->nr++; 35340257bb82SYan, Zheng 35350257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 35360257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 35370257bb82SYan, Zheng if (ret) 35380257bb82SYan, Zheng return ret; 35390257bb82SYan, Zheng cluster->nr = 0; 35400257bb82SYan, Zheng } 35410257bb82SYan, Zheng return 0; 35425d4f98a2SYan Zheng } 35435d4f98a2SYan Zheng 35445d4f98a2SYan Zheng /* 35455d4f98a2SYan Zheng * helper to add a tree block to the list. 35465d4f98a2SYan Zheng * the major work is getting the generation and level of the block 35475d4f98a2SYan Zheng */ 35485d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 35495d4f98a2SYan Zheng struct btrfs_key *extent_key, 35505d4f98a2SYan Zheng struct btrfs_path *path, 35515d4f98a2SYan Zheng struct rb_root *blocks) 35525d4f98a2SYan Zheng { 35535d4f98a2SYan Zheng struct extent_buffer *eb; 35545d4f98a2SYan Zheng struct btrfs_extent_item *ei; 35555d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 35565d4f98a2SYan Zheng struct tree_block *block; 35575d4f98a2SYan Zheng struct rb_node *rb_node; 35585d4f98a2SYan Zheng u32 item_size; 35595d4f98a2SYan Zheng int level = -1; 35607fdf4b60SWang Shilong u64 generation; 35615d4f98a2SYan Zheng 35625d4f98a2SYan Zheng eb = path->nodes[0]; 35635d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 35645d4f98a2SYan Zheng 35653173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 35663173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 35675d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 35685d4f98a2SYan Zheng struct btrfs_extent_item); 35693173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 35705d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 35715d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 35725d4f98a2SYan Zheng } else { 35733173a18fSJosef Bacik level = (int)extent_key->offset; 35743173a18fSJosef Bacik } 35753173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 35766d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3577ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3578ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3579ba3c2b19SNikolay Borisov return -EINVAL; 35803173a18fSJosef Bacik } else { 35815d4f98a2SYan Zheng BUG(); 35825d4f98a2SYan Zheng } 35835d4f98a2SYan Zheng 3584b3b4aa74SDavid Sterba btrfs_release_path(path); 35855d4f98a2SYan Zheng 35865d4f98a2SYan Zheng BUG_ON(level == -1); 35875d4f98a2SYan Zheng 35885d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 35895d4f98a2SYan Zheng if (!block) 35905d4f98a2SYan Zheng return -ENOMEM; 35915d4f98a2SYan Zheng 35925d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3593da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 35945d4f98a2SYan Zheng block->key.offset = generation; 35955d4f98a2SYan Zheng block->level = level; 35965d4f98a2SYan Zheng block->key_ready = 0; 35975d4f98a2SYan Zheng 3598e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 359943c04fb1SJeff Mahoney if (rb_node) 360043c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, block->bytenr); 36015d4f98a2SYan Zheng 36025d4f98a2SYan Zheng return 0; 36035d4f98a2SYan Zheng } 36045d4f98a2SYan Zheng 36055d4f98a2SYan Zheng /* 36065d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 36075d4f98a2SYan Zheng */ 36085d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 36095d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 36105d4f98a2SYan Zheng struct rb_root *blocks) 36115d4f98a2SYan Zheng { 36120b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36135d4f98a2SYan Zheng struct btrfs_path *path; 36145d4f98a2SYan Zheng struct btrfs_key key; 36155d4f98a2SYan Zheng int ret; 36160b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 36175d4f98a2SYan Zheng 36187476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 36195d4f98a2SYan Zheng return 0; 36205d4f98a2SYan Zheng 3621e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 36225d4f98a2SYan Zheng return 0; 36235d4f98a2SYan Zheng 36245d4f98a2SYan Zheng path = btrfs_alloc_path(); 36255d4f98a2SYan Zheng if (!path) 36265d4f98a2SYan Zheng return -ENOMEM; 3627aee68ee5SJosef Bacik again: 36285d4f98a2SYan Zheng key.objectid = bytenr; 3629aee68ee5SJosef Bacik if (skinny) { 3630aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3631aee68ee5SJosef Bacik key.offset = (u64)-1; 3632aee68ee5SJosef Bacik } else { 36335d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 36345d4f98a2SYan Zheng key.offset = blocksize; 3635aee68ee5SJosef Bacik } 36365d4f98a2SYan Zheng 36375d4f98a2SYan Zheng path->search_commit_root = 1; 36385d4f98a2SYan Zheng path->skip_locking = 1; 36395d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 36405d4f98a2SYan Zheng if (ret < 0) 36415d4f98a2SYan Zheng goto out; 36425d4f98a2SYan Zheng 3643aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3644aee68ee5SJosef Bacik if (path->slots[0]) { 3645aee68ee5SJosef Bacik path->slots[0]--; 3646aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3647aee68ee5SJosef Bacik path->slots[0]); 36483173a18fSJosef Bacik if (key.objectid == bytenr && 3649aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3650aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3651aee68ee5SJosef Bacik key.offset == blocksize))) 36523173a18fSJosef Bacik ret = 0; 36533173a18fSJosef Bacik } 3654aee68ee5SJosef Bacik 3655aee68ee5SJosef Bacik if (ret) { 3656aee68ee5SJosef Bacik skinny = false; 3657aee68ee5SJosef Bacik btrfs_release_path(path); 3658aee68ee5SJosef Bacik goto again; 3659aee68ee5SJosef Bacik } 3660aee68ee5SJosef Bacik } 3661cdccee99SLiu Bo if (ret) { 3662cdccee99SLiu Bo ASSERT(ret == 1); 3663cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3664cdccee99SLiu Bo btrfs_err(fs_info, 3665cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3666cdccee99SLiu Bo bytenr); 3667cdccee99SLiu Bo WARN_ON(1); 3668cdccee99SLiu Bo ret = -EINVAL; 3669cdccee99SLiu Bo goto out; 3670cdccee99SLiu Bo } 36713173a18fSJosef Bacik 36725d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 36735d4f98a2SYan Zheng out: 36745d4f98a2SYan Zheng btrfs_free_path(path); 36755d4f98a2SYan Zheng return ret; 36765d4f98a2SYan Zheng } 36775d4f98a2SYan Zheng 36780af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 367932da5386SDavid Sterba struct btrfs_block_group *block_group, 36801bbc621eSChris Mason struct inode *inode, 36811bbc621eSChris Mason u64 ino) 36820af3d00bSJosef Bacik { 36830af3d00bSJosef Bacik struct btrfs_key key; 36840af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 36850af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 36860af3d00bSJosef Bacik int ret = 0; 36870af3d00bSJosef Bacik 36880af3d00bSJosef Bacik if (inode) 36890af3d00bSJosef Bacik goto truncate; 36900af3d00bSJosef Bacik 36910af3d00bSJosef Bacik key.objectid = ino; 36920af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 36930af3d00bSJosef Bacik key.offset = 0; 36940af3d00bSJosef Bacik 36954c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 36962e19f1f9SAl Viro if (IS_ERR(inode)) 36970af3d00bSJosef Bacik return -ENOENT; 36980af3d00bSJosef Bacik 36990af3d00bSJosef Bacik truncate: 37002ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 37017b61cd92SMiao Xie &fs_info->global_block_rsv); 37027b61cd92SMiao Xie if (ret) 37037b61cd92SMiao Xie goto out; 37047b61cd92SMiao Xie 37057a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 37060af3d00bSJosef Bacik if (IS_ERR(trans)) { 37073612b495STsutomu Itoh ret = PTR_ERR(trans); 37080af3d00bSJosef Bacik goto out; 37090af3d00bSJosef Bacik } 37100af3d00bSJosef Bacik 371177ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 37120af3d00bSJosef Bacik 37133a45bb20SJeff Mahoney btrfs_end_transaction(trans); 37142ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 37150af3d00bSJosef Bacik out: 37160af3d00bSJosef Bacik iput(inode); 37170af3d00bSJosef Bacik return ret; 37180af3d00bSJosef Bacik } 37190af3d00bSJosef Bacik 37205d4f98a2SYan Zheng /* 372119b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 372219b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 37235d4f98a2SYan Zheng */ 372419b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 372519b546d7SQu Wenruo struct btrfs_block_group *block_group, 372619b546d7SQu Wenruo u64 data_bytenr) 37275d4f98a2SYan Zheng { 372819b546d7SQu Wenruo u64 space_cache_ino; 372919b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 37305d4f98a2SYan Zheng struct btrfs_key key; 373119b546d7SQu Wenruo bool found = false; 373219b546d7SQu Wenruo int i; 37335d4f98a2SYan Zheng int ret; 37345d4f98a2SYan Zheng 373519b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 373619b546d7SQu Wenruo return 0; 37375d4f98a2SYan Zheng 373819b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 373919b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 374019b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 374119b546d7SQu Wenruo continue; 374219b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 374319b546d7SQu Wenruo if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG && 374419b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 374519b546d7SQu Wenruo found = true; 374619b546d7SQu Wenruo space_cache_ino = key.objectid; 374719b546d7SQu Wenruo break; 374819b546d7SQu Wenruo } 374919b546d7SQu Wenruo } 375019b546d7SQu Wenruo if (!found) 375119b546d7SQu Wenruo return -ENOENT; 375219b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 375319b546d7SQu Wenruo space_cache_ino); 37540af3d00bSJosef Bacik return ret; 37555d4f98a2SYan Zheng } 37565d4f98a2SYan Zheng 37575d4f98a2SYan Zheng /* 37582c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 37595d4f98a2SYan Zheng */ 37605d4f98a2SYan Zheng static noinline_for_stack 37615d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 37625d4f98a2SYan Zheng struct btrfs_key *extent_key, 37635d4f98a2SYan Zheng struct btrfs_path *path, 37645d4f98a2SYan Zheng struct rb_root *blocks) 37655d4f98a2SYan Zheng { 376619b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 376719b546d7SQu Wenruo struct ulist *leaves = NULL; 376819b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 376919b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 377019b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3771647f63bdSFilipe David Borba Manana int ret = 0; 37725d4f98a2SYan Zheng 3773b3b4aa74SDavid Sterba btrfs_release_path(path); 377419b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 377519b546d7SQu Wenruo 0, &leaves, NULL, true); 377619b546d7SQu Wenruo if (ret < 0) 377719b546d7SQu Wenruo return ret; 377819b546d7SQu Wenruo 377919b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 378019b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 378119b546d7SQu Wenruo struct extent_buffer *eb; 378219b546d7SQu Wenruo 378319b546d7SQu Wenruo eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL); 378419b546d7SQu Wenruo if (IS_ERR(eb)) { 378519b546d7SQu Wenruo ret = PTR_ERR(eb); 378619b546d7SQu Wenruo break; 378719b546d7SQu Wenruo } 378819b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 378919b546d7SQu Wenruo extent_key->objectid); 379019b546d7SQu Wenruo free_extent_buffer(eb); 379119b546d7SQu Wenruo if (ret < 0) 379219b546d7SQu Wenruo break; 379319b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 379419b546d7SQu Wenruo if (ret < 0) 379519b546d7SQu Wenruo break; 379619b546d7SQu Wenruo } 379719b546d7SQu Wenruo if (ret < 0) 37985d4f98a2SYan Zheng free_block_list(blocks); 379919b546d7SQu Wenruo ulist_free(leaves); 380019b546d7SQu Wenruo return ret; 38015d4f98a2SYan Zheng } 38025d4f98a2SYan Zheng 38035d4f98a2SYan Zheng /* 38042c016dc2SLiu Bo * helper to find next unprocessed extent 38055d4f98a2SYan Zheng */ 38065d4f98a2SYan Zheng static noinline_for_stack 3807147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 38083fd0a558SYan, Zheng struct btrfs_key *extent_key) 38095d4f98a2SYan Zheng { 38100b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 38115d4f98a2SYan Zheng struct btrfs_key key; 38125d4f98a2SYan Zheng struct extent_buffer *leaf; 38135d4f98a2SYan Zheng u64 start, end, last; 38145d4f98a2SYan Zheng int ret; 38155d4f98a2SYan Zheng 3816b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 38175d4f98a2SYan Zheng while (1) { 38185d4f98a2SYan Zheng cond_resched(); 38195d4f98a2SYan Zheng if (rc->search_start >= last) { 38205d4f98a2SYan Zheng ret = 1; 38215d4f98a2SYan Zheng break; 38225d4f98a2SYan Zheng } 38235d4f98a2SYan Zheng 38245d4f98a2SYan Zheng key.objectid = rc->search_start; 38255d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 38265d4f98a2SYan Zheng key.offset = 0; 38275d4f98a2SYan Zheng 38285d4f98a2SYan Zheng path->search_commit_root = 1; 38295d4f98a2SYan Zheng path->skip_locking = 1; 38305d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 38315d4f98a2SYan Zheng 0, 0); 38325d4f98a2SYan Zheng if (ret < 0) 38335d4f98a2SYan Zheng break; 38345d4f98a2SYan Zheng next: 38355d4f98a2SYan Zheng leaf = path->nodes[0]; 38365d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 38375d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 38385d4f98a2SYan Zheng if (ret != 0) 38395d4f98a2SYan Zheng break; 38405d4f98a2SYan Zheng leaf = path->nodes[0]; 38415d4f98a2SYan Zheng } 38425d4f98a2SYan Zheng 38435d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 38445d4f98a2SYan Zheng if (key.objectid >= last) { 38455d4f98a2SYan Zheng ret = 1; 38465d4f98a2SYan Zheng break; 38475d4f98a2SYan Zheng } 38485d4f98a2SYan Zheng 38493173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 38503173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 38513173a18fSJosef Bacik path->slots[0]++; 38523173a18fSJosef Bacik goto next; 38533173a18fSJosef Bacik } 38543173a18fSJosef Bacik 38553173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 38565d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 38575d4f98a2SYan Zheng path->slots[0]++; 38585d4f98a2SYan Zheng goto next; 38595d4f98a2SYan Zheng } 38605d4f98a2SYan Zheng 38613173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 38620b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 38633173a18fSJosef Bacik rc->search_start) { 38643173a18fSJosef Bacik path->slots[0]++; 38653173a18fSJosef Bacik goto next; 38663173a18fSJosef Bacik } 38673173a18fSJosef Bacik 38685d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 38695d4f98a2SYan Zheng key.objectid, &start, &end, 3870e6138876SJosef Bacik EXTENT_DIRTY, NULL); 38715d4f98a2SYan Zheng 38725d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3873b3b4aa74SDavid Sterba btrfs_release_path(path); 38745d4f98a2SYan Zheng rc->search_start = end + 1; 38755d4f98a2SYan Zheng } else { 38763173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 38775d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 38783173a18fSJosef Bacik else 38793173a18fSJosef Bacik rc->search_start = key.objectid + 38800b246afaSJeff Mahoney fs_info->nodesize; 38813fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 38825d4f98a2SYan Zheng return 0; 38835d4f98a2SYan Zheng } 38845d4f98a2SYan Zheng } 3885b3b4aa74SDavid Sterba btrfs_release_path(path); 38865d4f98a2SYan Zheng return ret; 38875d4f98a2SYan Zheng } 38885d4f98a2SYan Zheng 38895d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 38905d4f98a2SYan Zheng { 38915d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 38927585717fSChris Mason 38937585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 38945d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 38957585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 38965d4f98a2SYan Zheng } 38975d4f98a2SYan Zheng 38985d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 38995d4f98a2SYan Zheng { 39005d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39017585717fSChris Mason 39027585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 39035d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 39047585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 39055d4f98a2SYan Zheng } 39065d4f98a2SYan Zheng 39075d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 39085d4f98a2SYan Zheng { 39095d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 39105d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 39115d4f98a2SYan Zheng return 1; 39125d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 39135d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 39145d4f98a2SYan Zheng return 1; 39155d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 39165d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 39175d4f98a2SYan Zheng return 1; 39185d4f98a2SYan Zheng return 0; 39195d4f98a2SYan Zheng } 39205d4f98a2SYan Zheng 39213fd0a558SYan, Zheng static noinline_for_stack 39223fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 39233fd0a558SYan, Zheng { 39243fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3925ac2fabacSJosef Bacik int ret; 39263fd0a558SYan, Zheng 39272ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 392866d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 39293fd0a558SYan, Zheng if (!rc->block_rsv) 39303fd0a558SYan, Zheng return -ENOMEM; 39313fd0a558SYan, Zheng 39323fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3933b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 39343fd0a558SYan, Zheng rc->extents_found = 0; 39353fd0a558SYan, Zheng rc->nodes_relocated = 0; 39363fd0a558SYan, Zheng rc->merging_rsv_size = 0; 39370647bf56SWang Shilong rc->reserved_bytes = 0; 3938da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 39390647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3940ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3941ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3942ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3943ac2fabacSJosef Bacik if (ret) 3944ac2fabacSJosef Bacik return ret; 39453fd0a558SYan, Zheng 39463fd0a558SYan, Zheng rc->create_reloc_tree = 1; 39473fd0a558SYan, Zheng set_reloc_control(rc); 39483fd0a558SYan, Zheng 39497a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 395028818947SLiu Bo if (IS_ERR(trans)) { 395128818947SLiu Bo unset_reloc_control(rc); 395228818947SLiu Bo /* 395328818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 395428818947SLiu Bo * cleanup. And callers are responsible to free the above 395528818947SLiu Bo * block rsv. 395628818947SLiu Bo */ 395728818947SLiu Bo return PTR_ERR(trans); 395828818947SLiu Bo } 39593a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 39603fd0a558SYan, Zheng return 0; 39613fd0a558SYan, Zheng } 396276dda93cSYan, Zheng 39635d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 39645d4f98a2SYan Zheng { 39652ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39665d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 39675d4f98a2SYan Zheng struct btrfs_key key; 39685d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 39695d4f98a2SYan Zheng struct btrfs_path *path; 39705d4f98a2SYan Zheng struct btrfs_extent_item *ei; 39715d4f98a2SYan Zheng u64 flags; 39725d4f98a2SYan Zheng u32 item_size; 39735d4f98a2SYan Zheng int ret; 39745d4f98a2SYan Zheng int err = 0; 3975c87f08caSChris Mason int progress = 0; 39765d4f98a2SYan Zheng 39775d4f98a2SYan Zheng path = btrfs_alloc_path(); 39783fd0a558SYan, Zheng if (!path) 39795d4f98a2SYan Zheng return -ENOMEM; 3980e4058b54SDavid Sterba path->reada = READA_FORWARD; 39813fd0a558SYan, Zheng 39823fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 39833fd0a558SYan, Zheng if (ret) { 39843fd0a558SYan, Zheng err = ret; 39853fd0a558SYan, Zheng goto out_free; 39862423fdfbSJiri Slaby } 39875d4f98a2SYan Zheng 39885d4f98a2SYan Zheng while (1) { 39890647bf56SWang Shilong rc->reserved_bytes = 0; 39900647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 39910647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 39920647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 39930647bf56SWang Shilong if (ret) { 39940647bf56SWang Shilong err = ret; 39950647bf56SWang Shilong break; 39960647bf56SWang Shilong } 3997c87f08caSChris Mason progress++; 3998a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 39990f788c58SLiu Bo if (IS_ERR(trans)) { 40000f788c58SLiu Bo err = PTR_ERR(trans); 40010f788c58SLiu Bo trans = NULL; 40020f788c58SLiu Bo break; 40030f788c58SLiu Bo } 4004c87f08caSChris Mason restart: 40053fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 40063a45bb20SJeff Mahoney btrfs_end_transaction(trans); 400742a657f5SPan Bian trans = NULL; 40083fd0a558SYan, Zheng continue; 40093fd0a558SYan, Zheng } 40103fd0a558SYan, Zheng 4011147d256eSZhaolei ret = find_next_extent(rc, path, &key); 40125d4f98a2SYan Zheng if (ret < 0) 40135d4f98a2SYan Zheng err = ret; 40145d4f98a2SYan Zheng if (ret != 0) 40155d4f98a2SYan Zheng break; 40165d4f98a2SYan Zheng 40175d4f98a2SYan Zheng rc->extents_found++; 40185d4f98a2SYan Zheng 40195d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 40205d4f98a2SYan Zheng struct btrfs_extent_item); 40213fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 40225d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 40235d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 40245d4f98a2SYan Zheng ret = check_extent_flags(flags); 40255d4f98a2SYan Zheng BUG_ON(ret); 40266d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 4027ba3c2b19SNikolay Borisov err = -EINVAL; 4028ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 4029ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 4030ba3c2b19SNikolay Borisov break; 40315d4f98a2SYan Zheng } else { 40325d4f98a2SYan Zheng BUG(); 40335d4f98a2SYan Zheng } 40345d4f98a2SYan Zheng 40355d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 40365d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 40375d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 40385d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 40395d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 40405d4f98a2SYan Zheng } else { 4041b3b4aa74SDavid Sterba btrfs_release_path(path); 40425d4f98a2SYan Zheng ret = 0; 40435d4f98a2SYan Zheng } 40445d4f98a2SYan Zheng if (ret < 0) { 40453fd0a558SYan, Zheng err = ret; 40465d4f98a2SYan Zheng break; 40475d4f98a2SYan Zheng } 40485d4f98a2SYan Zheng 40495d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 40505d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 40515d4f98a2SYan Zheng if (ret < 0) { 40523fd0a558SYan, Zheng if (ret != -EAGAIN) { 40535d4f98a2SYan Zheng err = ret; 40545d4f98a2SYan Zheng break; 40555d4f98a2SYan Zheng } 40563fd0a558SYan, Zheng rc->extents_found--; 40573fd0a558SYan, Zheng rc->search_start = key.objectid; 40583fd0a558SYan, Zheng } 40595d4f98a2SYan Zheng } 40605d4f98a2SYan Zheng 40613a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 40622ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 40633fd0a558SYan, Zheng trans = NULL; 40645d4f98a2SYan Zheng 40655d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 40665d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 40675d4f98a2SYan Zheng rc->found_file_extent = 1; 40680257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 40693fd0a558SYan, Zheng &key, &rc->cluster); 40705d4f98a2SYan Zheng if (ret < 0) { 40715d4f98a2SYan Zheng err = ret; 40725d4f98a2SYan Zheng break; 40735d4f98a2SYan Zheng } 40745d4f98a2SYan Zheng } 4075f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 4076f31ea088SQu Wenruo err = -ECANCELED; 4077f31ea088SQu Wenruo break; 4078f31ea088SQu Wenruo } 40795d4f98a2SYan Zheng } 4080c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 408143a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 40829689457bSShilong Wang if (ret == 1) { 4083c87f08caSChris Mason err = 0; 4084c87f08caSChris Mason progress = 0; 4085c87f08caSChris Mason goto restart; 4086c87f08caSChris Mason } 4087c87f08caSChris Mason } 40883fd0a558SYan, Zheng 4089b3b4aa74SDavid Sterba btrfs_release_path(path); 409091166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 40915d4f98a2SYan Zheng 40925d4f98a2SYan Zheng if (trans) { 40933a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 40942ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 40955d4f98a2SYan Zheng } 40965d4f98a2SYan Zheng 40970257bb82SYan, Zheng if (!err) { 40983fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 40993fd0a558SYan, Zheng &rc->cluster); 41000257bb82SYan, Zheng if (ret < 0) 41010257bb82SYan, Zheng err = ret; 41020257bb82SYan, Zheng } 41030257bb82SYan, Zheng 41043fd0a558SYan, Zheng rc->create_reloc_tree = 0; 41053fd0a558SYan, Zheng set_reloc_control(rc); 41060257bb82SYan, Zheng 41073fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 410863f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 41095d4f98a2SYan Zheng 41107f913c7cSQu Wenruo /* 41117f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 41127f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 41137f913c7cSQu Wenruo * 41147f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 41157f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 41167f913c7cSQu Wenruo * merge_reloc_roots() 41177f913c7cSQu Wenruo */ 41183fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 41195d4f98a2SYan Zheng 41205d4f98a2SYan Zheng merge_reloc_roots(rc); 41215d4f98a2SYan Zheng 41223fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 41235d4f98a2SYan Zheng unset_reloc_control(rc); 412463f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 41255d4f98a2SYan Zheng 41265d4f98a2SYan Zheng /* get rid of pinned extents */ 41277a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 412862b99540SQu Wenruo if (IS_ERR(trans)) { 41293612b495STsutomu Itoh err = PTR_ERR(trans); 413062b99540SQu Wenruo goto out_free; 413162b99540SQu Wenruo } 41323a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 41336217b0faSJosef Bacik out_free: 4134d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4135d2311e69SQu Wenruo if (ret < 0 && !err) 4136d2311e69SQu Wenruo err = ret; 41372ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 41383fd0a558SYan, Zheng btrfs_free_path(path); 41395d4f98a2SYan Zheng return err; 41405d4f98a2SYan Zheng } 41415d4f98a2SYan Zheng 41425d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 41430257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 41445d4f98a2SYan Zheng { 41455d4f98a2SYan Zheng struct btrfs_path *path; 41465d4f98a2SYan Zheng struct btrfs_inode_item *item; 41475d4f98a2SYan Zheng struct extent_buffer *leaf; 41485d4f98a2SYan Zheng int ret; 41495d4f98a2SYan Zheng 41505d4f98a2SYan Zheng path = btrfs_alloc_path(); 41515d4f98a2SYan Zheng if (!path) 41525d4f98a2SYan Zheng return -ENOMEM; 41535d4f98a2SYan Zheng 41545d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 41555d4f98a2SYan Zheng if (ret) 41565d4f98a2SYan Zheng goto out; 41575d4f98a2SYan Zheng 41585d4f98a2SYan Zheng leaf = path->nodes[0]; 41595d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 4160b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 41615d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 41620257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 41635d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 41643fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 41653fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 41665d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 41675d4f98a2SYan Zheng out: 41685d4f98a2SYan Zheng btrfs_free_path(path); 41695d4f98a2SYan Zheng return ret; 41705d4f98a2SYan Zheng } 41715d4f98a2SYan Zheng 41725d4f98a2SYan Zheng /* 41735d4f98a2SYan Zheng * helper to create inode for data relocation. 41745d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 41755d4f98a2SYan Zheng */ 41763fd0a558SYan, Zheng static noinline_for_stack 41773fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 417832da5386SDavid Sterba struct btrfs_block_group *group) 41795d4f98a2SYan Zheng { 41805d4f98a2SYan Zheng struct inode *inode = NULL; 41815d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 41825d4f98a2SYan Zheng struct btrfs_root *root; 41835d4f98a2SYan Zheng struct btrfs_key key; 41844624900dSZhaolei u64 objectid; 41855d4f98a2SYan Zheng int err = 0; 41865d4f98a2SYan Zheng 41875d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 41885d4f98a2SYan Zheng if (IS_ERR(root)) 41895d4f98a2SYan Zheng return ERR_CAST(root); 41905d4f98a2SYan Zheng 4191a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 419276deacf0SJosef Bacik if (IS_ERR(trans)) { 419300246528SJosef Bacik btrfs_put_root(root); 41943fd0a558SYan, Zheng return ERR_CAST(trans); 419576deacf0SJosef Bacik } 41965d4f98a2SYan Zheng 4197581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 41985d4f98a2SYan Zheng if (err) 41995d4f98a2SYan Zheng goto out; 42005d4f98a2SYan Zheng 42010257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 42025d4f98a2SYan Zheng BUG_ON(err); 42035d4f98a2SYan Zheng 42045d4f98a2SYan Zheng key.objectid = objectid; 42055d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 42065d4f98a2SYan Zheng key.offset = 0; 42074c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 42082e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 4209b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 42105d4f98a2SYan Zheng 421173f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 42125d4f98a2SYan Zheng out: 421300246528SJosef Bacik btrfs_put_root(root); 42143a45bb20SJeff Mahoney btrfs_end_transaction(trans); 42152ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 42165d4f98a2SYan Zheng if (err) { 42175d4f98a2SYan Zheng if (inode) 42185d4f98a2SYan Zheng iput(inode); 42195d4f98a2SYan Zheng inode = ERR_PTR(err); 42205d4f98a2SYan Zheng } 42215d4f98a2SYan Zheng return inode; 42225d4f98a2SYan Zheng } 42235d4f98a2SYan Zheng 4224c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 42253fd0a558SYan, Zheng { 42263fd0a558SYan, Zheng struct reloc_control *rc; 42273fd0a558SYan, Zheng 42283fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 42293fd0a558SYan, Zheng if (!rc) 42303fd0a558SYan, Zheng return NULL; 42313fd0a558SYan, Zheng 42323fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 4233d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 42342433bea5SQu Wenruo backref_cache_init(fs_info, &rc->backref_cache, 1); 42353fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 423643eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 423743eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 42383fd0a558SYan, Zheng return rc; 42393fd0a558SYan, Zheng } 42403fd0a558SYan, Zheng 42411a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 42421a0afa0eSJosef Bacik { 42431a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 42441a0afa0eSJosef Bacik 42451a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 42461a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 42471a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 42481a0afa0eSJosef Bacik kfree(node); 42491a0afa0eSJosef Bacik 42501a0afa0eSJosef Bacik kfree(rc); 42511a0afa0eSJosef Bacik } 42521a0afa0eSJosef Bacik 42535d4f98a2SYan Zheng /* 4254ebce0e01SAdam Borowski * Print the block group being relocated 4255ebce0e01SAdam Borowski */ 4256ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 425732da5386SDavid Sterba struct btrfs_block_group *block_group) 4258ebce0e01SAdam Borowski { 4259f89e09cfSAnand Jain char buf[128] = {'\0'}; 4260ebce0e01SAdam Borowski 4261f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 4262ebce0e01SAdam Borowski 4263ebce0e01SAdam Borowski btrfs_info(fs_info, 4264ebce0e01SAdam Borowski "relocating block group %llu flags %s", 4265b3470b5dSDavid Sterba block_group->start, buf); 4266ebce0e01SAdam Borowski } 4267ebce0e01SAdam Borowski 4268430640e3SQu Wenruo static const char *stage_to_string(int stage) 4269430640e3SQu Wenruo { 4270430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 4271430640e3SQu Wenruo return "move data extents"; 4272430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 4273430640e3SQu Wenruo return "update data pointers"; 4274430640e3SQu Wenruo return "unknown"; 4275430640e3SQu Wenruo } 4276430640e3SQu Wenruo 4277ebce0e01SAdam Borowski /* 42785d4f98a2SYan Zheng * function to relocate all extents in a block group. 42795d4f98a2SYan Zheng */ 42806bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 42815d4f98a2SYan Zheng { 428232da5386SDavid Sterba struct btrfs_block_group *bg; 42836bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 42845d4f98a2SYan Zheng struct reloc_control *rc; 42850af3d00bSJosef Bacik struct inode *inode; 42860af3d00bSJosef Bacik struct btrfs_path *path; 42875d4f98a2SYan Zheng int ret; 4288f0486c68SYan, Zheng int rw = 0; 42895d4f98a2SYan Zheng int err = 0; 42905d4f98a2SYan Zheng 4291eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 4292eede2bf3SOmar Sandoval if (!bg) 4293eede2bf3SOmar Sandoval return -ENOENT; 4294eede2bf3SOmar Sandoval 4295eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 4296eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 4297eede2bf3SOmar Sandoval return -ETXTBSY; 4298eede2bf3SOmar Sandoval } 4299eede2bf3SOmar Sandoval 4300c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 4301eede2bf3SOmar Sandoval if (!rc) { 4302eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 43035d4f98a2SYan Zheng return -ENOMEM; 4304eede2bf3SOmar Sandoval } 43055d4f98a2SYan Zheng 4306f0486c68SYan, Zheng rc->extent_root = extent_root; 4307eede2bf3SOmar Sandoval rc->block_group = bg; 43085d4f98a2SYan Zheng 4309b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 4310f0486c68SYan, Zheng if (ret) { 4311f0486c68SYan, Zheng err = ret; 4312f0486c68SYan, Zheng goto out; 4313f0486c68SYan, Zheng } 4314f0486c68SYan, Zheng rw = 1; 4315f0486c68SYan, Zheng 43160af3d00bSJosef Bacik path = btrfs_alloc_path(); 43170af3d00bSJosef Bacik if (!path) { 43180af3d00bSJosef Bacik err = -ENOMEM; 43190af3d00bSJosef Bacik goto out; 43200af3d00bSJosef Bacik } 43210af3d00bSJosef Bacik 43227949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 43230af3d00bSJosef Bacik btrfs_free_path(path); 43240af3d00bSJosef Bacik 43250af3d00bSJosef Bacik if (!IS_ERR(inode)) 43261bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 43270af3d00bSJosef Bacik else 43280af3d00bSJosef Bacik ret = PTR_ERR(inode); 43290af3d00bSJosef Bacik 43300af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 43310af3d00bSJosef Bacik err = ret; 43320af3d00bSJosef Bacik goto out; 43330af3d00bSJosef Bacik } 43340af3d00bSJosef Bacik 43355d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 43365d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 43375d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 43385d4f98a2SYan Zheng rc->data_inode = NULL; 43395d4f98a2SYan Zheng goto out; 43405d4f98a2SYan Zheng } 43415d4f98a2SYan Zheng 43420b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 43435d4f98a2SYan Zheng 43449cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 4345f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 43466374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 4347b3470b5dSDavid Sterba rc->block_group->start, 4348b3470b5dSDavid Sterba rc->block_group->length); 43495d4f98a2SYan Zheng 43505d4f98a2SYan Zheng while (1) { 4351430640e3SQu Wenruo int finishes_stage; 4352430640e3SQu Wenruo 435376dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 43545d4f98a2SYan Zheng ret = relocate_block_group(rc); 435576dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 4356ff612ba7SJosef Bacik if (ret < 0) 43575d4f98a2SYan Zheng err = ret; 4358ff612ba7SJosef Bacik 4359430640e3SQu Wenruo finishes_stage = rc->stage; 4360ff612ba7SJosef Bacik /* 4361ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 4362ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 4363ff612ba7SJosef Bacik * different block group we could end up hitting the 4364ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 4365ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 4366ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 4367ff612ba7SJosef Bacik * out of the loop if we hit an error. 4368ff612ba7SJosef Bacik */ 4369ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 4370ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 4371ff612ba7SJosef Bacik (u64)-1); 4372ff612ba7SJosef Bacik if (ret) 4373ff612ba7SJosef Bacik err = ret; 4374ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 4375ff612ba7SJosef Bacik 0, -1); 4376ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 43775d4f98a2SYan Zheng } 43785d4f98a2SYan Zheng 4379ff612ba7SJosef Bacik if (err < 0) 4380ff612ba7SJosef Bacik goto out; 4381ff612ba7SJosef Bacik 43825d4f98a2SYan Zheng if (rc->extents_found == 0) 43835d4f98a2SYan Zheng break; 43845d4f98a2SYan Zheng 4385430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 4386430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 43875d4f98a2SYan Zheng } 43885d4f98a2SYan Zheng 43895d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 43905d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 4391bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 43925d4f98a2SYan Zheng out: 4393f0486c68SYan, Zheng if (err && rw) 43942ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 43955d4f98a2SYan Zheng iput(rc->data_inode); 43965d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 43971a0afa0eSJosef Bacik free_reloc_control(rc); 43985d4f98a2SYan Zheng return err; 43995d4f98a2SYan Zheng } 44005d4f98a2SYan Zheng 440176dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 440276dda93cSYan, Zheng { 44030b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 440476dda93cSYan, Zheng struct btrfs_trans_handle *trans; 440579787eaaSJeff Mahoney int ret, err; 440676dda93cSYan, Zheng 44070b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 440879787eaaSJeff Mahoney if (IS_ERR(trans)) 440979787eaaSJeff Mahoney return PTR_ERR(trans); 441076dda93cSYan, Zheng 441176dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 441276dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 441376dda93cSYan, Zheng root->root_item.drop_level = 0; 441476dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 44150b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 441676dda93cSYan, Zheng &root->root_key, &root->root_item); 441776dda93cSYan, Zheng 44183a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 441979787eaaSJeff Mahoney if (err) 442079787eaaSJeff Mahoney return err; 442179787eaaSJeff Mahoney return ret; 442276dda93cSYan, Zheng } 442376dda93cSYan, Zheng 44245d4f98a2SYan Zheng /* 44255d4f98a2SYan Zheng * recover relocation interrupted by system crash. 44265d4f98a2SYan Zheng * 44275d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 44285d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 44295d4f98a2SYan Zheng */ 44305d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 44315d4f98a2SYan Zheng { 44320b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 44335d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 44345d4f98a2SYan Zheng struct btrfs_key key; 44355d4f98a2SYan Zheng struct btrfs_root *fs_root; 44365d4f98a2SYan Zheng struct btrfs_root *reloc_root; 44375d4f98a2SYan Zheng struct btrfs_path *path; 44385d4f98a2SYan Zheng struct extent_buffer *leaf; 44395d4f98a2SYan Zheng struct reloc_control *rc = NULL; 44405d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 44415d4f98a2SYan Zheng int ret; 44425d4f98a2SYan Zheng int err = 0; 44435d4f98a2SYan Zheng 44445d4f98a2SYan Zheng path = btrfs_alloc_path(); 44455d4f98a2SYan Zheng if (!path) 44465d4f98a2SYan Zheng return -ENOMEM; 4447e4058b54SDavid Sterba path->reada = READA_BACK; 44485d4f98a2SYan Zheng 44495d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 44505d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 44515d4f98a2SYan Zheng key.offset = (u64)-1; 44525d4f98a2SYan Zheng 44535d4f98a2SYan Zheng while (1) { 44540b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 44555d4f98a2SYan Zheng path, 0, 0); 44565d4f98a2SYan Zheng if (ret < 0) { 44575d4f98a2SYan Zheng err = ret; 44585d4f98a2SYan Zheng goto out; 44595d4f98a2SYan Zheng } 44605d4f98a2SYan Zheng if (ret > 0) { 44615d4f98a2SYan Zheng if (path->slots[0] == 0) 44625d4f98a2SYan Zheng break; 44635d4f98a2SYan Zheng path->slots[0]--; 44645d4f98a2SYan Zheng } 44655d4f98a2SYan Zheng leaf = path->nodes[0]; 44665d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4467b3b4aa74SDavid Sterba btrfs_release_path(path); 44685d4f98a2SYan Zheng 44695d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 44705d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 44715d4f98a2SYan Zheng break; 44725d4f98a2SYan Zheng 44733dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 44745d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 44755d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 44765d4f98a2SYan Zheng goto out; 44775d4f98a2SYan Zheng } 44785d4f98a2SYan Zheng 44793dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 44805d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 44815d4f98a2SYan Zheng 44825d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 44830b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, 44845d4f98a2SYan Zheng reloc_root->root_key.offset); 44855d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 448676dda93cSYan, Zheng ret = PTR_ERR(fs_root); 448776dda93cSYan, Zheng if (ret != -ENOENT) { 448876dda93cSYan, Zheng err = ret; 44895d4f98a2SYan Zheng goto out; 44905d4f98a2SYan Zheng } 449179787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 449279787eaaSJeff Mahoney if (ret < 0) { 449379787eaaSJeff Mahoney err = ret; 449479787eaaSJeff Mahoney goto out; 449579787eaaSJeff Mahoney } 4496932fd26dSJosef Bacik } else { 449700246528SJosef Bacik btrfs_put_root(fs_root); 449876dda93cSYan, Zheng } 44995d4f98a2SYan Zheng } 45005d4f98a2SYan Zheng 45015d4f98a2SYan Zheng if (key.offset == 0) 45025d4f98a2SYan Zheng break; 45035d4f98a2SYan Zheng 45045d4f98a2SYan Zheng key.offset--; 45055d4f98a2SYan Zheng } 4506b3b4aa74SDavid Sterba btrfs_release_path(path); 45075d4f98a2SYan Zheng 45085d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 45095d4f98a2SYan Zheng goto out; 45105d4f98a2SYan Zheng 4511c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 45125d4f98a2SYan Zheng if (!rc) { 45135d4f98a2SYan Zheng err = -ENOMEM; 45145d4f98a2SYan Zheng goto out; 45155d4f98a2SYan Zheng } 45165d4f98a2SYan Zheng 45170b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 45185d4f98a2SYan Zheng 45195d4f98a2SYan Zheng set_reloc_control(rc); 45205d4f98a2SYan Zheng 45217a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 45223612b495STsutomu Itoh if (IS_ERR(trans)) { 45233612b495STsutomu Itoh err = PTR_ERR(trans); 4524fb2d83eeSJosef Bacik goto out_unset; 45253612b495STsutomu Itoh } 45263fd0a558SYan, Zheng 45273fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 45283fd0a558SYan, Zheng 45295d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 45305d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 45315d4f98a2SYan Zheng struct btrfs_root, root_list); 45325d4f98a2SYan Zheng list_del(&reloc_root->root_list); 45335d4f98a2SYan Zheng 45345d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 45355d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 45365d4f98a2SYan Zheng &rc->reloc_roots); 45375d4f98a2SYan Zheng continue; 45385d4f98a2SYan Zheng } 45395d4f98a2SYan Zheng 45400b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, reloc_root->root_key.offset); 454179787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 454279787eaaSJeff Mahoney err = PTR_ERR(fs_root); 4543ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 45441402d17dSXiyu Yang btrfs_end_transaction(trans); 4545fb2d83eeSJosef Bacik goto out_unset; 454679787eaaSJeff Mahoney } 45475d4f98a2SYan Zheng 4548ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 454979787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 4550f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 455100246528SJosef Bacik btrfs_put_root(fs_root); 45525d4f98a2SYan Zheng } 45535d4f98a2SYan Zheng 45543a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 455579787eaaSJeff Mahoney if (err) 4556fb2d83eeSJosef Bacik goto out_unset; 45575d4f98a2SYan Zheng 45585d4f98a2SYan Zheng merge_reloc_roots(rc); 45595d4f98a2SYan Zheng 45605d4f98a2SYan Zheng unset_reloc_control(rc); 45615d4f98a2SYan Zheng 45627a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 456362b99540SQu Wenruo if (IS_ERR(trans)) { 45643612b495STsutomu Itoh err = PTR_ERR(trans); 45656217b0faSJosef Bacik goto out_clean; 456662b99540SQu Wenruo } 45673a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 45686217b0faSJosef Bacik out_clean: 4569d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4570d2311e69SQu Wenruo if (ret < 0 && !err) 4571d2311e69SQu Wenruo err = ret; 4572fb2d83eeSJosef Bacik out_unset: 4573fb2d83eeSJosef Bacik unset_reloc_control(rc); 45741a0afa0eSJosef Bacik free_reloc_control(rc); 45753612b495STsutomu Itoh out: 4576aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 4577aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4578aca1bba6SLiu Bo 45795d4f98a2SYan Zheng btrfs_free_path(path); 45805d4f98a2SYan Zheng 45815d4f98a2SYan Zheng if (err == 0) { 45825d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 45830b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 4584932fd26dSJosef Bacik if (IS_ERR(fs_root)) { 45855d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4586932fd26dSJosef Bacik } else { 458766b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 458800246528SJosef Bacik btrfs_put_root(fs_root); 4589932fd26dSJosef Bacik } 4590932fd26dSJosef Bacik } 45915d4f98a2SYan Zheng return err; 45925d4f98a2SYan Zheng } 45935d4f98a2SYan Zheng 45945d4f98a2SYan Zheng /* 45955d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 45965d4f98a2SYan Zheng * 45975d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 45985d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 45995d4f98a2SYan Zheng */ 46005d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 46015d4f98a2SYan Zheng { 46020b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 46035d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 46045d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 46055d4f98a2SYan Zheng int ret; 46065d4f98a2SYan Zheng u64 disk_bytenr; 46074577b014SJosef Bacik u64 new_bytenr; 46085d4f98a2SYan Zheng LIST_HEAD(list); 46095d4f98a2SYan Zheng 46105d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4611bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 46125d4f98a2SYan Zheng 46135d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 46140b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4615a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 461679787eaaSJeff Mahoney if (ret) 461779787eaaSJeff Mahoney goto out; 46185d4f98a2SYan Zheng 46195d4f98a2SYan Zheng while (!list_empty(&list)) { 46205d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 46215d4f98a2SYan Zheng list_del_init(&sums->list); 46225d4f98a2SYan Zheng 46234577b014SJosef Bacik /* 46244577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 46254577b014SJosef Bacik * We need to do this because we will read in entire prealloc 46264577b014SJosef Bacik * extents but we may have written to say the middle of the 46274577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 46284577b014SJosef Bacik * the right disk offset. 46294577b014SJosef Bacik * 46304577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 46314577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 46324577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 46334577b014SJosef Bacik * disk length. 46344577b014SJosef Bacik */ 4635bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 46364577b014SJosef Bacik sums->bytenr = new_bytenr; 46375d4f98a2SYan Zheng 4638f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 46395d4f98a2SYan Zheng } 464079787eaaSJeff Mahoney out: 46415d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4642411fc6bcSAndi Kleen return ret; 46435d4f98a2SYan Zheng } 46443fd0a558SYan, Zheng 464583d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 46463fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 46473fd0a558SYan, Zheng struct extent_buffer *cow) 46483fd0a558SYan, Zheng { 46490b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 46503fd0a558SYan, Zheng struct reloc_control *rc; 4651a26195a5SQu Wenruo struct btrfs_backref_node *node; 46523fd0a558SYan, Zheng int first_cow = 0; 46533fd0a558SYan, Zheng int level; 465483d4cfd4SJosef Bacik int ret = 0; 46553fd0a558SYan, Zheng 46560b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 46573fd0a558SYan, Zheng if (!rc) 465883d4cfd4SJosef Bacik return 0; 46593fd0a558SYan, Zheng 46603fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 46613fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 46623fd0a558SYan, Zheng 46633fd0a558SYan, Zheng level = btrfs_header_level(buf); 46643fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 46653fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 46663fd0a558SYan, Zheng first_cow = 1; 46673fd0a558SYan, Zheng 46683fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 46693fd0a558SYan, Zheng rc->create_reloc_tree) { 46703fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 46713fd0a558SYan, Zheng 46723fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 46733fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 46743fd0a558SYan, Zheng node->new_bytenr != buf->start); 46753fd0a558SYan, Zheng 46763fd0a558SYan, Zheng drop_node_buffer(node); 467767439dadSDavid Sterba atomic_inc(&cow->refs); 46783fd0a558SYan, Zheng node->eb = cow; 46793fd0a558SYan, Zheng node->new_bytenr = cow->start; 46803fd0a558SYan, Zheng 46813fd0a558SYan, Zheng if (!node->pending) { 46823fd0a558SYan, Zheng list_move_tail(&node->list, 46833fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 46843fd0a558SYan, Zheng node->pending = 1; 46853fd0a558SYan, Zheng } 46863fd0a558SYan, Zheng 46873fd0a558SYan, Zheng if (first_cow) 46889569cc20SQu Wenruo mark_block_processed(rc, node); 46893fd0a558SYan, Zheng 46903fd0a558SYan, Zheng if (first_cow && level > 0) 46913fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 46923fd0a558SYan, Zheng } 46933fd0a558SYan, Zheng 469483d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 46953fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 469683d4cfd4SJosef Bacik return ret; 46973fd0a558SYan, Zheng } 46983fd0a558SYan, Zheng 46993fd0a558SYan, Zheng /* 47003fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 470101327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 47023fd0a558SYan, Zheng */ 4703147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 47043fd0a558SYan, Zheng u64 *bytes_to_reserve) 47053fd0a558SYan, Zheng { 470610995c04SQu Wenruo struct btrfs_root *root = pending->root; 470710995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 47083fd0a558SYan, Zheng 47096282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 47103fd0a558SYan, Zheng return; 47113fd0a558SYan, Zheng 47123fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 47133fd0a558SYan, Zheng return; 47143fd0a558SYan, Zheng 47153fd0a558SYan, Zheng root = root->reloc_root; 47163fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 47173fd0a558SYan, Zheng /* 47183fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 47193fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 47203fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 47213fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 47223fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 47233fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 47243fd0a558SYan, Zheng * tree may use more space than it frees. so we need 47253fd0a558SYan, Zheng * reserve extra space. 47263fd0a558SYan, Zheng */ 47273fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 47283fd0a558SYan, Zheng } 47293fd0a558SYan, Zheng 47303fd0a558SYan, Zheng /* 47313fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 47323fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4733f44deb74SJosef Bacik * 4734f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4735f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4736f44deb74SJosef Bacik * rc->reloc_roots. 47373fd0a558SYan, Zheng */ 473849b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 47393fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 47403fd0a558SYan, Zheng { 47413fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 47423fd0a558SYan, Zheng struct btrfs_root *reloc_root; 47433fd0a558SYan, Zheng struct btrfs_root *new_root; 474410995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 47453fd0a558SYan, Zheng int ret; 47463fd0a558SYan, Zheng 47476282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 474849b25e05SJeff Mahoney return 0; 47493fd0a558SYan, Zheng 47503fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 47513fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 47523fd0a558SYan, Zheng 47533fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 47543fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 47553fd0a558SYan, Zheng rc->block_rsv, 47563a584174SLu Fengqi rc->nodes_relocated, true); 475749b25e05SJeff Mahoney if (ret) 475849b25e05SJeff Mahoney return ret; 47593fd0a558SYan, Zheng } 47603fd0a558SYan, Zheng 47613fd0a558SYan, Zheng new_root = pending->snap; 47623fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 47633fd0a558SYan, Zheng new_root->root_key.objectid); 476449b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 476549b25e05SJeff Mahoney return PTR_ERR(reloc_root); 47663fd0a558SYan, Zheng 4767ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4768ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4769f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 47703fd0a558SYan, Zheng 477149b25e05SJeff Mahoney if (rc->create_reloc_tree) 47723fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 477349b25e05SJeff Mahoney return ret; 47743fd0a558SYan, Zheng } 4775