1c1d7c514SDavid Sterba // SPDX-License-Identifier: GPL-2.0 25d4f98a2SYan Zheng /* 35d4f98a2SYan Zheng * Copyright (C) 2009 Oracle. All rights reserved. 45d4f98a2SYan Zheng */ 55d4f98a2SYan Zheng 65d4f98a2SYan Zheng #include <linux/sched.h> 75d4f98a2SYan Zheng #include <linux/pagemap.h> 85d4f98a2SYan Zheng #include <linux/writeback.h> 95d4f98a2SYan Zheng #include <linux/blkdev.h> 105d4f98a2SYan Zheng #include <linux/rbtree.h> 115a0e3ad6STejun Heo #include <linux/slab.h> 12726a3421SQu Wenruo #include <linux/error-injection.h> 135d4f98a2SYan Zheng #include "ctree.h" 145d4f98a2SYan Zheng #include "disk-io.h" 155d4f98a2SYan Zheng #include "transaction.h" 165d4f98a2SYan Zheng #include "volumes.h" 175d4f98a2SYan Zheng #include "locking.h" 185d4f98a2SYan Zheng #include "btrfs_inode.h" 195d4f98a2SYan Zheng #include "async-thread.h" 200af3d00bSJosef Bacik #include "free-space-cache.h" 21581bb050SLi Zefan #include "inode-map.h" 2262b99540SQu Wenruo #include "qgroup.h" 23cdccee99SLiu Bo #include "print-tree.h" 2486736342SJosef Bacik #include "delalloc-space.h" 25aac0023cSJosef Bacik #include "block-group.h" 2619b546d7SQu Wenruo #include "backref.h" 27e9a28dc5SQu Wenruo #include "misc.h" 285d4f98a2SYan Zheng 295d4f98a2SYan Zheng /* 300c891389SQu Wenruo * Relocation overview 310c891389SQu Wenruo * 320c891389SQu Wenruo * [What does relocation do] 330c891389SQu Wenruo * 340c891389SQu Wenruo * The objective of relocation is to relocate all extents of the target block 350c891389SQu Wenruo * group to other block groups. 360c891389SQu Wenruo * This is utilized by resize (shrink only), profile converting, compacting 370c891389SQu Wenruo * space, or balance routine to spread chunks over devices. 380c891389SQu Wenruo * 390c891389SQu Wenruo * Before | After 400c891389SQu Wenruo * ------------------------------------------------------------------ 410c891389SQu Wenruo * BG A: 10 data extents | BG A: deleted 420c891389SQu Wenruo * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated) 430c891389SQu Wenruo * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated) 440c891389SQu Wenruo * 450c891389SQu Wenruo * [How does relocation work] 460c891389SQu Wenruo * 470c891389SQu Wenruo * 1. Mark the target block group read-only 480c891389SQu Wenruo * New extents won't be allocated from the target block group. 490c891389SQu Wenruo * 500c891389SQu Wenruo * 2.1 Record each extent in the target block group 510c891389SQu Wenruo * To build a proper map of extents to be relocated. 520c891389SQu Wenruo * 530c891389SQu Wenruo * 2.2 Build data reloc tree and reloc trees 540c891389SQu Wenruo * Data reloc tree will contain an inode, recording all newly relocated 550c891389SQu Wenruo * data extents. 560c891389SQu Wenruo * There will be only one data reloc tree for one data block group. 570c891389SQu Wenruo * 580c891389SQu Wenruo * Reloc tree will be a special snapshot of its source tree, containing 590c891389SQu Wenruo * relocated tree blocks. 600c891389SQu Wenruo * Each tree referring to a tree block in target block group will get its 610c891389SQu Wenruo * reloc tree built. 620c891389SQu Wenruo * 630c891389SQu Wenruo * 2.3 Swap source tree with its corresponding reloc tree 640c891389SQu Wenruo * Each involved tree only refers to new extents after swap. 650c891389SQu Wenruo * 660c891389SQu Wenruo * 3. Cleanup reloc trees and data reloc tree. 670c891389SQu Wenruo * As old extents in the target block group are still referenced by reloc 680c891389SQu Wenruo * trees, we need to clean them up before really freeing the target block 690c891389SQu Wenruo * group. 700c891389SQu Wenruo * 710c891389SQu Wenruo * The main complexity is in steps 2.2 and 2.3. 720c891389SQu Wenruo * 730c891389SQu Wenruo * The entry point of relocation is relocate_block_group() function. 740c891389SQu Wenruo */ 750c891389SQu Wenruo 760647bf56SWang Shilong #define RELOCATION_RESERVED_NODES 256 772a979612SQu Wenruo /* 785d4f98a2SYan Zheng * map address of tree root to tree 795d4f98a2SYan Zheng */ 805d4f98a2SYan Zheng struct mapping_node { 81e9a28dc5SQu Wenruo struct { 825d4f98a2SYan Zheng struct rb_node rb_node; 835d4f98a2SYan Zheng u64 bytenr; 84e9a28dc5SQu Wenruo }; /* Use rb_simle_node for search/insert */ 855d4f98a2SYan Zheng void *data; 865d4f98a2SYan Zheng }; 875d4f98a2SYan Zheng 885d4f98a2SYan Zheng struct mapping_tree { 895d4f98a2SYan Zheng struct rb_root rb_root; 905d4f98a2SYan Zheng spinlock_t lock; 915d4f98a2SYan Zheng }; 925d4f98a2SYan Zheng 935d4f98a2SYan Zheng /* 945d4f98a2SYan Zheng * present a tree block to process 955d4f98a2SYan Zheng */ 965d4f98a2SYan Zheng struct tree_block { 97e9a28dc5SQu Wenruo struct { 985d4f98a2SYan Zheng struct rb_node rb_node; 995d4f98a2SYan Zheng u64 bytenr; 100e9a28dc5SQu Wenruo }; /* Use rb_simple_node for search/insert */ 1015d4f98a2SYan Zheng struct btrfs_key key; 1025d4f98a2SYan Zheng unsigned int level:8; 1035d4f98a2SYan Zheng unsigned int key_ready:1; 1045d4f98a2SYan Zheng }; 1055d4f98a2SYan Zheng 1060257bb82SYan, Zheng #define MAX_EXTENTS 128 1070257bb82SYan, Zheng 1080257bb82SYan, Zheng struct file_extent_cluster { 1090257bb82SYan, Zheng u64 start; 1100257bb82SYan, Zheng u64 end; 1110257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 1120257bb82SYan, Zheng unsigned int nr; 1130257bb82SYan, Zheng }; 1140257bb82SYan, Zheng 1155d4f98a2SYan Zheng struct reloc_control { 1165d4f98a2SYan Zheng /* block group to relocate */ 11732da5386SDavid Sterba struct btrfs_block_group *block_group; 1185d4f98a2SYan Zheng /* extent tree */ 1195d4f98a2SYan Zheng struct btrfs_root *extent_root; 1205d4f98a2SYan Zheng /* inode for moving data */ 1215d4f98a2SYan Zheng struct inode *data_inode; 1223fd0a558SYan, Zheng 1233fd0a558SYan, Zheng struct btrfs_block_rsv *block_rsv; 1243fd0a558SYan, Zheng 125a26195a5SQu Wenruo struct btrfs_backref_cache backref_cache; 1263fd0a558SYan, Zheng 1273fd0a558SYan, Zheng struct file_extent_cluster cluster; 1285d4f98a2SYan Zheng /* tree blocks have been processed */ 1295d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 1305d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 1315d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 1325d4f98a2SYan Zheng /* list of reloc trees */ 1335d4f98a2SYan Zheng struct list_head reloc_roots; 134d2311e69SQu Wenruo /* list of subvolume trees that get relocated */ 135d2311e69SQu Wenruo struct list_head dirty_subvol_roots; 1363fd0a558SYan, Zheng /* size of metadata reservation for merging reloc trees */ 1373fd0a558SYan, Zheng u64 merging_rsv_size; 1383fd0a558SYan, Zheng /* size of relocated tree nodes */ 1393fd0a558SYan, Zheng u64 nodes_relocated; 1400647bf56SWang Shilong /* reserved size for block group relocation*/ 1410647bf56SWang Shilong u64 reserved_bytes; 1423fd0a558SYan, Zheng 1435d4f98a2SYan Zheng u64 search_start; 1445d4f98a2SYan Zheng u64 extents_found; 1453fd0a558SYan, Zheng 1463fd0a558SYan, Zheng unsigned int stage:8; 1473fd0a558SYan, Zheng unsigned int create_reloc_tree:1; 1483fd0a558SYan, Zheng unsigned int merge_reloc_tree:1; 1495d4f98a2SYan Zheng unsigned int found_file_extent:1; 1505d4f98a2SYan Zheng }; 1515d4f98a2SYan Zheng 1525d4f98a2SYan Zheng /* stages of data relocation */ 1535d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 1545d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 1555d4f98a2SYan Zheng 156a26195a5SQu Wenruo static void remove_backref_node(struct btrfs_backref_cache *cache, 157a26195a5SQu Wenruo struct btrfs_backref_node *node); 1589569cc20SQu Wenruo 1599569cc20SQu Wenruo static void mark_block_processed(struct reloc_control *rc, 160a26195a5SQu Wenruo struct btrfs_backref_node *node) 1619569cc20SQu Wenruo { 1629569cc20SQu Wenruo u32 blocksize; 1639569cc20SQu Wenruo 1649569cc20SQu Wenruo if (node->level == 0 || 1659569cc20SQu Wenruo in_range(node->bytenr, rc->block_group->start, 1669569cc20SQu Wenruo rc->block_group->length)) { 1679569cc20SQu Wenruo blocksize = rc->extent_root->fs_info->nodesize; 1689569cc20SQu Wenruo set_extent_bits(&rc->processed_blocks, node->bytenr, 1699569cc20SQu Wenruo node->bytenr + blocksize - 1, EXTENT_DIRTY); 1709569cc20SQu Wenruo } 1719569cc20SQu Wenruo node->processed = 1; 1729569cc20SQu Wenruo } 1739569cc20SQu Wenruo 1745d4f98a2SYan Zheng 1755d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 1765d4f98a2SYan Zheng { 1776bef4d31SEric Paris tree->rb_root = RB_ROOT; 1785d4f98a2SYan Zheng spin_lock_init(&tree->lock); 1795d4f98a2SYan Zheng } 1805d4f98a2SYan Zheng 181a26195a5SQu Wenruo static void backref_cache_cleanup(struct btrfs_backref_cache *cache) 1825d4f98a2SYan Zheng { 183a26195a5SQu Wenruo struct btrfs_backref_node *node; 1843fd0a558SYan, Zheng int i; 1853fd0a558SYan, Zheng 1863fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 1873fd0a558SYan, Zheng node = list_entry(cache->detached.next, 188a26195a5SQu Wenruo struct btrfs_backref_node, list); 1893fd0a558SYan, Zheng remove_backref_node(cache, node); 1903fd0a558SYan, Zheng } 1913fd0a558SYan, Zheng 1923fd0a558SYan, Zheng while (!list_empty(&cache->leaves)) { 1933fd0a558SYan, Zheng node = list_entry(cache->leaves.next, 194a26195a5SQu Wenruo struct btrfs_backref_node, lower); 1953fd0a558SYan, Zheng remove_backref_node(cache, node); 1963fd0a558SYan, Zheng } 1973fd0a558SYan, Zheng 1983fd0a558SYan, Zheng cache->last_trans = 0; 1993fd0a558SYan, Zheng 2003fd0a558SYan, Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 201f4907095SLiu Bo ASSERT(list_empty(&cache->pending[i])); 20284780289SQu Wenruo ASSERT(list_empty(&cache->pending_edge)); 20384780289SQu Wenruo ASSERT(list_empty(&cache->useless_node)); 204f4907095SLiu Bo ASSERT(list_empty(&cache->changed)); 205f4907095SLiu Bo ASSERT(list_empty(&cache->detached)); 206f4907095SLiu Bo ASSERT(RB_EMPTY_ROOT(&cache->rb_root)); 207f4907095SLiu Bo ASSERT(!cache->nr_nodes); 208f4907095SLiu Bo ASSERT(!cache->nr_edges); 2093fd0a558SYan, Zheng } 2103fd0a558SYan, Zheng 211a26195a5SQu Wenruo static struct btrfs_backref_node *alloc_backref_node( 212a26195a5SQu Wenruo struct btrfs_backref_cache *cache, u64 bytenr, int level) 2133fd0a558SYan, Zheng { 214a26195a5SQu Wenruo struct btrfs_backref_node *node; 2153fd0a558SYan, Zheng 2160304f2d8SQu Wenruo ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL); 2173fd0a558SYan, Zheng node = kzalloc(sizeof(*node), GFP_NOFS); 2180304f2d8SQu Wenruo if (!node) 2190304f2d8SQu Wenruo return node; 2200304f2d8SQu Wenruo 2213fd0a558SYan, Zheng INIT_LIST_HEAD(&node->list); 2225d4f98a2SYan Zheng INIT_LIST_HEAD(&node->upper); 2235d4f98a2SYan Zheng INIT_LIST_HEAD(&node->lower); 2245d4f98a2SYan Zheng RB_CLEAR_NODE(&node->rb_node); 2253fd0a558SYan, Zheng cache->nr_nodes++; 2260304f2d8SQu Wenruo node->level = level; 2270304f2d8SQu Wenruo node->bytenr = bytenr; 2280304f2d8SQu Wenruo 2293fd0a558SYan, Zheng return node; 2303fd0a558SYan, Zheng } 2313fd0a558SYan, Zheng 232a26195a5SQu Wenruo static void free_backref_node(struct btrfs_backref_cache *cache, 233a26195a5SQu Wenruo struct btrfs_backref_node *node) 2343fd0a558SYan, Zheng { 2353fd0a558SYan, Zheng if (node) { 2363fd0a558SYan, Zheng cache->nr_nodes--; 23700246528SJosef Bacik btrfs_put_root(node->root); 2383fd0a558SYan, Zheng kfree(node); 2393fd0a558SYan, Zheng } 2403fd0a558SYan, Zheng } 2413fd0a558SYan, Zheng 242a26195a5SQu Wenruo static struct btrfs_backref_edge *alloc_backref_edge( 243a26195a5SQu Wenruo struct btrfs_backref_cache *cache) 2443fd0a558SYan, Zheng { 245a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2463fd0a558SYan, Zheng 2473fd0a558SYan, Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 2483fd0a558SYan, Zheng if (edge) 2493fd0a558SYan, Zheng cache->nr_edges++; 2503fd0a558SYan, Zheng return edge; 2513fd0a558SYan, Zheng } 2523fd0a558SYan, Zheng 2532a979612SQu Wenruo #define LINK_LOWER (1 << 0) 2542a979612SQu Wenruo #define LINK_UPPER (1 << 1) 255a26195a5SQu Wenruo static void link_backref_edge(struct btrfs_backref_edge *edge, 256a26195a5SQu Wenruo struct btrfs_backref_node *lower, 257a26195a5SQu Wenruo struct btrfs_backref_node *upper, 2582a979612SQu Wenruo int link_which) 2592a979612SQu Wenruo { 2602a979612SQu Wenruo ASSERT(upper && lower && upper->level == lower->level + 1); 2612a979612SQu Wenruo edge->node[LOWER] = lower; 2622a979612SQu Wenruo edge->node[UPPER] = upper; 2632a979612SQu Wenruo if (link_which & LINK_LOWER) 2642a979612SQu Wenruo list_add_tail(&edge->list[LOWER], &lower->upper); 2652a979612SQu Wenruo if (link_which & LINK_UPPER) 2662a979612SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 2672a979612SQu Wenruo } 2682a979612SQu Wenruo 269a26195a5SQu Wenruo static void free_backref_edge(struct btrfs_backref_cache *cache, 270a26195a5SQu Wenruo struct btrfs_backref_edge *edge) 2713fd0a558SYan, Zheng { 2723fd0a558SYan, Zheng if (edge) { 2733fd0a558SYan, Zheng cache->nr_edges--; 2743fd0a558SYan, Zheng kfree(edge); 2753fd0a558SYan, Zheng } 2765d4f98a2SYan Zheng } 2775d4f98a2SYan Zheng 27848a3b636SEric Sandeen static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr) 27943c04fb1SJeff Mahoney { 28043c04fb1SJeff Mahoney 28143c04fb1SJeff Mahoney struct btrfs_fs_info *fs_info = NULL; 282a26195a5SQu Wenruo struct btrfs_backref_node *bnode = rb_entry(rb_node, 283a26195a5SQu Wenruo struct btrfs_backref_node, rb_node); 28443c04fb1SJeff Mahoney if (bnode->root) 28543c04fb1SJeff Mahoney fs_info = bnode->root->fs_info; 2865d163e0eSJeff Mahoney btrfs_panic(fs_info, errno, 2875d163e0eSJeff Mahoney "Inconsistency in backref cache found at offset %llu", 2885d163e0eSJeff Mahoney bytenr); 28943c04fb1SJeff Mahoney } 29043c04fb1SJeff Mahoney 2915d4f98a2SYan Zheng /* 2925d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 2935d4f98a2SYan Zheng */ 294a26195a5SQu Wenruo static struct btrfs_backref_node *walk_up_backref( 295a26195a5SQu Wenruo struct btrfs_backref_node *node, 296a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 2975d4f98a2SYan Zheng { 298a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2995d4f98a2SYan Zheng int idx = *index; 3005d4f98a2SYan Zheng 3015d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 3025d4f98a2SYan Zheng edge = list_entry(node->upper.next, 303a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 3045d4f98a2SYan Zheng edges[idx++] = edge; 3055d4f98a2SYan Zheng node = edge->node[UPPER]; 3065d4f98a2SYan Zheng } 3073fd0a558SYan, Zheng BUG_ON(node->detached); 3085d4f98a2SYan Zheng *index = idx; 3095d4f98a2SYan Zheng return node; 3105d4f98a2SYan Zheng } 3115d4f98a2SYan Zheng 3125d4f98a2SYan Zheng /* 3135d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 3145d4f98a2SYan Zheng */ 315a26195a5SQu Wenruo static struct btrfs_backref_node *walk_down_backref( 316a26195a5SQu Wenruo struct btrfs_backref_edge *edges[], int *index) 3175d4f98a2SYan Zheng { 318a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 319a26195a5SQu Wenruo struct btrfs_backref_node *lower; 3205d4f98a2SYan Zheng int idx = *index; 3215d4f98a2SYan Zheng 3225d4f98a2SYan Zheng while (idx > 0) { 3235d4f98a2SYan Zheng edge = edges[idx - 1]; 3245d4f98a2SYan Zheng lower = edge->node[LOWER]; 3255d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 3265d4f98a2SYan Zheng idx--; 3275d4f98a2SYan Zheng continue; 3285d4f98a2SYan Zheng } 3295d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 330a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 3315d4f98a2SYan Zheng edges[idx - 1] = edge; 3325d4f98a2SYan Zheng *index = idx; 3335d4f98a2SYan Zheng return edge->node[UPPER]; 3345d4f98a2SYan Zheng } 3355d4f98a2SYan Zheng *index = 0; 3365d4f98a2SYan Zheng return NULL; 3375d4f98a2SYan Zheng } 3385d4f98a2SYan Zheng 339a26195a5SQu Wenruo static void unlock_node_buffer(struct btrfs_backref_node *node) 3405d4f98a2SYan Zheng { 3415d4f98a2SYan Zheng if (node->locked) { 3425d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 3435d4f98a2SYan Zheng node->locked = 0; 3445d4f98a2SYan Zheng } 3453fd0a558SYan, Zheng } 3463fd0a558SYan, Zheng 347a26195a5SQu Wenruo static void drop_node_buffer(struct btrfs_backref_node *node) 3483fd0a558SYan, Zheng { 3493fd0a558SYan, Zheng if (node->eb) { 3503fd0a558SYan, Zheng unlock_node_buffer(node); 3515d4f98a2SYan Zheng free_extent_buffer(node->eb); 3525d4f98a2SYan Zheng node->eb = NULL; 3535d4f98a2SYan Zheng } 3545d4f98a2SYan Zheng } 3555d4f98a2SYan Zheng 356a26195a5SQu Wenruo static void drop_backref_node(struct btrfs_backref_cache *tree, 357a26195a5SQu Wenruo struct btrfs_backref_node *node) 3585d4f98a2SYan Zheng { 3595d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 3605d4f98a2SYan Zheng 3615d4f98a2SYan Zheng drop_node_buffer(node); 3623fd0a558SYan, Zheng list_del(&node->list); 3635d4f98a2SYan Zheng list_del(&node->lower); 3643fd0a558SYan, Zheng if (!RB_EMPTY_NODE(&node->rb_node)) 3655d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 3663fd0a558SYan, Zheng free_backref_node(tree, node); 3675d4f98a2SYan Zheng } 3685d4f98a2SYan Zheng 3695d4f98a2SYan Zheng /* 3705d4f98a2SYan Zheng * remove a backref node from the backref cache 3715d4f98a2SYan Zheng */ 372a26195a5SQu Wenruo static void remove_backref_node(struct btrfs_backref_cache *cache, 373a26195a5SQu Wenruo struct btrfs_backref_node *node) 3745d4f98a2SYan Zheng { 375a26195a5SQu Wenruo struct btrfs_backref_node *upper; 376a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 3775d4f98a2SYan Zheng 3785d4f98a2SYan Zheng if (!node) 3795d4f98a2SYan Zheng return; 3805d4f98a2SYan Zheng 3813fd0a558SYan, Zheng BUG_ON(!node->lowest && !node->detached); 3825d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 383a26195a5SQu Wenruo edge = list_entry(node->upper.next, struct btrfs_backref_edge, 3845d4f98a2SYan Zheng list[LOWER]); 3855d4f98a2SYan Zheng upper = edge->node[UPPER]; 3865d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 3875d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 3883fd0a558SYan, Zheng free_backref_edge(cache, edge); 3893fd0a558SYan, Zheng 3903fd0a558SYan, Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 3913fd0a558SYan, Zheng BUG_ON(!list_empty(&node->upper)); 3923fd0a558SYan, Zheng drop_backref_node(cache, node); 3933fd0a558SYan, Zheng node = upper; 3943fd0a558SYan, Zheng node->lowest = 1; 3953fd0a558SYan, Zheng continue; 3963fd0a558SYan, Zheng } 3975d4f98a2SYan Zheng /* 3983fd0a558SYan, Zheng * add the node to leaf node list if no other 3995d4f98a2SYan Zheng * child block cached. 4005d4f98a2SYan Zheng */ 4015d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 4023fd0a558SYan, Zheng list_add_tail(&upper->lower, &cache->leaves); 4035d4f98a2SYan Zheng upper->lowest = 1; 4045d4f98a2SYan Zheng } 4055d4f98a2SYan Zheng } 4063fd0a558SYan, Zheng 4075d4f98a2SYan Zheng drop_backref_node(cache, node); 4085d4f98a2SYan Zheng } 4095d4f98a2SYan Zheng 410a26195a5SQu Wenruo static void update_backref_node(struct btrfs_backref_cache *cache, 411a26195a5SQu Wenruo struct btrfs_backref_node *node, u64 bytenr) 4123fd0a558SYan, Zheng { 4133fd0a558SYan, Zheng struct rb_node *rb_node; 4143fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 4153fd0a558SYan, Zheng node->bytenr = bytenr; 416e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node); 41743c04fb1SJeff Mahoney if (rb_node) 41843c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, bytenr); 4193fd0a558SYan, Zheng } 4203fd0a558SYan, Zheng 4213fd0a558SYan, Zheng /* 4223fd0a558SYan, Zheng * update backref cache after a transaction commit 4233fd0a558SYan, Zheng */ 4243fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 425a26195a5SQu Wenruo struct btrfs_backref_cache *cache) 4263fd0a558SYan, Zheng { 427a26195a5SQu Wenruo struct btrfs_backref_node *node; 4283fd0a558SYan, Zheng int level = 0; 4293fd0a558SYan, Zheng 4303fd0a558SYan, Zheng if (cache->last_trans == 0) { 4313fd0a558SYan, Zheng cache->last_trans = trans->transid; 4323fd0a558SYan, Zheng return 0; 4333fd0a558SYan, Zheng } 4343fd0a558SYan, Zheng 4353fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 4363fd0a558SYan, Zheng return 0; 4373fd0a558SYan, Zheng 4383fd0a558SYan, Zheng /* 4393fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 4403fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 4413fd0a558SYan, Zheng * so the detached nodes are no longer useful. 4423fd0a558SYan, Zheng */ 4433fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 4443fd0a558SYan, Zheng node = list_entry(cache->detached.next, 445a26195a5SQu Wenruo struct btrfs_backref_node, list); 4463fd0a558SYan, Zheng remove_backref_node(cache, node); 4473fd0a558SYan, Zheng } 4483fd0a558SYan, Zheng 4493fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 4503fd0a558SYan, Zheng node = list_entry(cache->changed.next, 451a26195a5SQu Wenruo struct btrfs_backref_node, list); 4523fd0a558SYan, Zheng list_del_init(&node->list); 4533fd0a558SYan, Zheng BUG_ON(node->pending); 4543fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 4553fd0a558SYan, Zheng } 4563fd0a558SYan, Zheng 4573fd0a558SYan, Zheng /* 4583fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 4593fd0a558SYan, Zheng * errors during processing the pending nodes. 4603fd0a558SYan, Zheng */ 4613fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 4623fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 4633fd0a558SYan, Zheng BUG_ON(!node->pending); 4643fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 4653fd0a558SYan, Zheng continue; 4663fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 4673fd0a558SYan, Zheng } 4683fd0a558SYan, Zheng } 4693fd0a558SYan, Zheng 4703fd0a558SYan, Zheng cache->last_trans = 0; 4713fd0a558SYan, Zheng return 1; 4723fd0a558SYan, Zheng } 4733fd0a558SYan, Zheng 4746282675eSQu Wenruo static bool reloc_root_is_dead(struct btrfs_root *root) 4756282675eSQu Wenruo { 4766282675eSQu Wenruo /* 4776282675eSQu Wenruo * Pair with set_bit/clear_bit in clean_dirty_subvols and 4786282675eSQu Wenruo * btrfs_update_reloc_root. We need to see the updated bit before 4796282675eSQu Wenruo * trying to access reloc_root 4806282675eSQu Wenruo */ 4816282675eSQu Wenruo smp_rmb(); 4826282675eSQu Wenruo if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state)) 4836282675eSQu Wenruo return true; 4846282675eSQu Wenruo return false; 4856282675eSQu Wenruo } 4866282675eSQu Wenruo 4876282675eSQu Wenruo /* 4886282675eSQu Wenruo * Check if this subvolume tree has valid reloc tree. 4896282675eSQu Wenruo * 4906282675eSQu Wenruo * Reloc tree after swap is considered dead, thus not considered as valid. 4916282675eSQu Wenruo * This is enough for most callers, as they don't distinguish dead reloc root 4926282675eSQu Wenruo * from no reloc root. But should_ignore_root() below is a special case. 4936282675eSQu Wenruo */ 4946282675eSQu Wenruo static bool have_reloc_root(struct btrfs_root *root) 4956282675eSQu Wenruo { 4966282675eSQu Wenruo if (reloc_root_is_dead(root)) 4976282675eSQu Wenruo return false; 4986282675eSQu Wenruo if (!root->reloc_root) 4996282675eSQu Wenruo return false; 5006282675eSQu Wenruo return true; 5016282675eSQu Wenruo } 502f2a97a9dSDavid Sterba 5033fd0a558SYan, Zheng static int should_ignore_root(struct btrfs_root *root) 5043fd0a558SYan, Zheng { 5053fd0a558SYan, Zheng struct btrfs_root *reloc_root; 5063fd0a558SYan, Zheng 50727cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 5083fd0a558SYan, Zheng return 0; 5093fd0a558SYan, Zheng 5106282675eSQu Wenruo /* This root has been merged with its reloc tree, we can ignore it */ 5116282675eSQu Wenruo if (reloc_root_is_dead(root)) 5126282675eSQu Wenruo return 1; 5136282675eSQu Wenruo 5143fd0a558SYan, Zheng reloc_root = root->reloc_root; 5153fd0a558SYan, Zheng if (!reloc_root) 5163fd0a558SYan, Zheng return 0; 5173fd0a558SYan, Zheng 5184d4225fcSJosef Bacik if (btrfs_header_generation(reloc_root->commit_root) == 5194d4225fcSJosef Bacik root->fs_info->running_transaction->transid) 5203fd0a558SYan, Zheng return 0; 5213fd0a558SYan, Zheng /* 5223fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 5233fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 5243fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 5253fd0a558SYan, Zheng * relocation. 5263fd0a558SYan, Zheng */ 5273fd0a558SYan, Zheng return 1; 5283fd0a558SYan, Zheng } 5295d4f98a2SYan Zheng /* 5305d4f98a2SYan Zheng * find reloc tree by address of tree root 5315d4f98a2SYan Zheng */ 5322433bea5SQu Wenruo struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr) 5335d4f98a2SYan Zheng { 5342433bea5SQu Wenruo struct reloc_control *rc = fs_info->reloc_ctl; 5355d4f98a2SYan Zheng struct rb_node *rb_node; 5365d4f98a2SYan Zheng struct mapping_node *node; 5375d4f98a2SYan Zheng struct btrfs_root *root = NULL; 5385d4f98a2SYan Zheng 5392433bea5SQu Wenruo ASSERT(rc); 5405d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 541e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr); 5425d4f98a2SYan Zheng if (rb_node) { 5435d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 5445d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 5455d4f98a2SYan Zheng } 5465d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 54700246528SJosef Bacik return btrfs_grab_root(root); 5485d4f98a2SYan Zheng } 5495d4f98a2SYan Zheng 5505d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 5515d4f98a2SYan Zheng u64 root_objectid) 5525d4f98a2SYan Zheng { 5535d4f98a2SYan Zheng struct btrfs_key key; 5545d4f98a2SYan Zheng 5555d4f98a2SYan Zheng key.objectid = root_objectid; 5565d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 5575d4f98a2SYan Zheng key.offset = (u64)-1; 5585d4f98a2SYan Zheng 559bc44d7c4SJosef Bacik return btrfs_get_fs_root(fs_info, &key, false); 5605d4f98a2SYan Zheng } 5615d4f98a2SYan Zheng 5625d4f98a2SYan Zheng /* 5634007ea87SQu Wenruo * Handle direct tree backref 5644007ea87SQu Wenruo * 5654007ea87SQu Wenruo * Direct tree backref means, the backref item shows its parent bytenr 5664007ea87SQu Wenruo * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined). 5674007ea87SQu Wenruo * 5684007ea87SQu Wenruo * @ref_key: The converted backref key. 5694007ea87SQu Wenruo * For keyed backref, it's the item key. 5704007ea87SQu Wenruo * For inlined backref, objectid is the bytenr, 5714007ea87SQu Wenruo * type is btrfs_inline_ref_type, offset is 5724007ea87SQu Wenruo * btrfs_inline_ref_offset. 5734007ea87SQu Wenruo */ 574a26195a5SQu Wenruo static int handle_direct_tree_backref(struct btrfs_backref_cache *cache, 5754007ea87SQu Wenruo struct btrfs_key *ref_key, 576a26195a5SQu Wenruo struct btrfs_backref_node *cur) 5774007ea87SQu Wenruo { 578a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 579a26195a5SQu Wenruo struct btrfs_backref_node *upper; 5804007ea87SQu Wenruo struct rb_node *rb_node; 5814007ea87SQu Wenruo 5824007ea87SQu Wenruo ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY); 5834007ea87SQu Wenruo 5844007ea87SQu Wenruo /* Only reloc root uses backref pointing to itself */ 5854007ea87SQu Wenruo if (ref_key->objectid == ref_key->offset) { 5864007ea87SQu Wenruo struct btrfs_root *root; 5874007ea87SQu Wenruo 5884007ea87SQu Wenruo cur->is_reloc_root = 1; 5894007ea87SQu Wenruo /* Only reloc backref cache cares about a specific root */ 5904007ea87SQu Wenruo if (cache->is_reloc) { 5914007ea87SQu Wenruo root = find_reloc_root(cache->fs_info, cur->bytenr); 5924007ea87SQu Wenruo if (WARN_ON(!root)) 5934007ea87SQu Wenruo return -ENOENT; 5944007ea87SQu Wenruo cur->root = root; 5954007ea87SQu Wenruo } else { 5964007ea87SQu Wenruo /* 5974007ea87SQu Wenruo * For generic purpose backref cache, reloc root node 5984007ea87SQu Wenruo * is useless. 5994007ea87SQu Wenruo */ 6004007ea87SQu Wenruo list_add(&cur->list, &cache->useless_node); 6014007ea87SQu Wenruo } 6024007ea87SQu Wenruo return 0; 6034007ea87SQu Wenruo } 6044007ea87SQu Wenruo 6054007ea87SQu Wenruo edge = alloc_backref_edge(cache); 6064007ea87SQu Wenruo if (!edge) 6074007ea87SQu Wenruo return -ENOMEM; 6084007ea87SQu Wenruo 609e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, ref_key->offset); 6104007ea87SQu Wenruo if (!rb_node) { 6114007ea87SQu Wenruo /* Parent node not yet cached */ 6120304f2d8SQu Wenruo upper = alloc_backref_node(cache, ref_key->offset, 6130304f2d8SQu Wenruo cur->level + 1); 6144007ea87SQu Wenruo if (!upper) { 6154007ea87SQu Wenruo free_backref_edge(cache, edge); 6164007ea87SQu Wenruo return -ENOMEM; 6174007ea87SQu Wenruo } 6184007ea87SQu Wenruo 6194007ea87SQu Wenruo /* 6204007ea87SQu Wenruo * Backrefs for the upper level block isn't cached, add the 6214007ea87SQu Wenruo * block to pending list 6224007ea87SQu Wenruo */ 6234007ea87SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 6244007ea87SQu Wenruo } else { 6254007ea87SQu Wenruo /* Parent node already cached */ 626a26195a5SQu Wenruo upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node); 6274007ea87SQu Wenruo ASSERT(upper->checked); 6284007ea87SQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 6294007ea87SQu Wenruo } 6302a979612SQu Wenruo link_backref_edge(edge, cur, upper, LINK_LOWER); 6314007ea87SQu Wenruo return 0; 6324007ea87SQu Wenruo } 6334007ea87SQu Wenruo 6344007ea87SQu Wenruo /* 6354d81ea8bSQu Wenruo * Handle indirect tree backref 6364d81ea8bSQu Wenruo * 6374d81ea8bSQu Wenruo * Indirect tree backref means, we only know which tree the node belongs to. 6384d81ea8bSQu Wenruo * We still need to do a tree search to find out the parents. This is for 6394d81ea8bSQu Wenruo * TREE_BLOCK_REF backref (keyed or inlined). 6404d81ea8bSQu Wenruo * 6414d81ea8bSQu Wenruo * @ref_key: The same as @ref_key in handle_direct_tree_backref() 6424d81ea8bSQu Wenruo * @tree_key: The first key of this tree block. 6434d81ea8bSQu Wenruo * @path: A clean (released) path, to avoid allocating path everytime 6444d81ea8bSQu Wenruo * the function get called. 6454d81ea8bSQu Wenruo */ 646a26195a5SQu Wenruo static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache, 6474d81ea8bSQu Wenruo struct btrfs_path *path, 6484d81ea8bSQu Wenruo struct btrfs_key *ref_key, 6494d81ea8bSQu Wenruo struct btrfs_key *tree_key, 650a26195a5SQu Wenruo struct btrfs_backref_node *cur) 6514d81ea8bSQu Wenruo { 6524d81ea8bSQu Wenruo struct btrfs_fs_info *fs_info = cache->fs_info; 653a26195a5SQu Wenruo struct btrfs_backref_node *upper; 654a26195a5SQu Wenruo struct btrfs_backref_node *lower; 655a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 6564d81ea8bSQu Wenruo struct extent_buffer *eb; 6574d81ea8bSQu Wenruo struct btrfs_root *root; 6584d81ea8bSQu Wenruo struct rb_node *rb_node; 6594d81ea8bSQu Wenruo int level; 6604d81ea8bSQu Wenruo bool need_check = true; 6614d81ea8bSQu Wenruo int ret; 6624d81ea8bSQu Wenruo 6634d81ea8bSQu Wenruo root = read_fs_root(fs_info, ref_key->offset); 6644d81ea8bSQu Wenruo if (IS_ERR(root)) 6654d81ea8bSQu Wenruo return PTR_ERR(root); 6664d81ea8bSQu Wenruo if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 6674d81ea8bSQu Wenruo cur->cowonly = 1; 6684d81ea8bSQu Wenruo 6694d81ea8bSQu Wenruo if (btrfs_root_level(&root->root_item) == cur->level) { 6704d81ea8bSQu Wenruo /* Tree root */ 6714d81ea8bSQu Wenruo ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr); 6724d81ea8bSQu Wenruo if (should_ignore_root(root)) { 6734d81ea8bSQu Wenruo btrfs_put_root(root); 6744d81ea8bSQu Wenruo list_add(&cur->list, &cache->useless_node); 6754d81ea8bSQu Wenruo } else { 6764d81ea8bSQu Wenruo cur->root = root; 6774d81ea8bSQu Wenruo } 6784d81ea8bSQu Wenruo return 0; 6794d81ea8bSQu Wenruo } 6804d81ea8bSQu Wenruo 6814d81ea8bSQu Wenruo level = cur->level + 1; 6824d81ea8bSQu Wenruo 6834d81ea8bSQu Wenruo /* Search the tree to find parent blocks referring to the block */ 6844d81ea8bSQu Wenruo path->search_commit_root = 1; 6854d81ea8bSQu Wenruo path->skip_locking = 1; 6864d81ea8bSQu Wenruo path->lowest_level = level; 6874d81ea8bSQu Wenruo ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0); 6884d81ea8bSQu Wenruo path->lowest_level = 0; 6894d81ea8bSQu Wenruo if (ret < 0) { 6904d81ea8bSQu Wenruo btrfs_put_root(root); 6914d81ea8bSQu Wenruo return ret; 6924d81ea8bSQu Wenruo } 6934d81ea8bSQu Wenruo if (ret > 0 && path->slots[level] > 0) 6944d81ea8bSQu Wenruo path->slots[level]--; 6954d81ea8bSQu Wenruo 6964d81ea8bSQu Wenruo eb = path->nodes[level]; 6974d81ea8bSQu Wenruo if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) { 6984d81ea8bSQu Wenruo btrfs_err(fs_info, 6994d81ea8bSQu Wenruo "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)", 7004d81ea8bSQu Wenruo cur->bytenr, level - 1, root->root_key.objectid, 7014d81ea8bSQu Wenruo tree_key->objectid, tree_key->type, tree_key->offset); 7024d81ea8bSQu Wenruo btrfs_put_root(root); 7034d81ea8bSQu Wenruo ret = -ENOENT; 7044d81ea8bSQu Wenruo goto out; 7054d81ea8bSQu Wenruo } 7064d81ea8bSQu Wenruo lower = cur; 7074d81ea8bSQu Wenruo 7084d81ea8bSQu Wenruo /* Add all nodes and edges in the path */ 7094d81ea8bSQu Wenruo for (; level < BTRFS_MAX_LEVEL; level++) { 7104d81ea8bSQu Wenruo if (!path->nodes[level]) { 7114d81ea8bSQu Wenruo ASSERT(btrfs_root_bytenr(&root->root_item) == 7124d81ea8bSQu Wenruo lower->bytenr); 7134d81ea8bSQu Wenruo if (should_ignore_root(root)) { 7144d81ea8bSQu Wenruo btrfs_put_root(root); 7154d81ea8bSQu Wenruo list_add(&lower->list, &cache->useless_node); 7164d81ea8bSQu Wenruo } else { 7174d81ea8bSQu Wenruo lower->root = root; 7184d81ea8bSQu Wenruo } 7194d81ea8bSQu Wenruo break; 7204d81ea8bSQu Wenruo } 7214d81ea8bSQu Wenruo 7224d81ea8bSQu Wenruo edge = alloc_backref_edge(cache); 7234d81ea8bSQu Wenruo if (!edge) { 7244d81ea8bSQu Wenruo btrfs_put_root(root); 7254d81ea8bSQu Wenruo ret = -ENOMEM; 7264d81ea8bSQu Wenruo goto out; 7274d81ea8bSQu Wenruo } 7284d81ea8bSQu Wenruo 7294d81ea8bSQu Wenruo eb = path->nodes[level]; 730e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, eb->start); 7314d81ea8bSQu Wenruo if (!rb_node) { 7320304f2d8SQu Wenruo upper = alloc_backref_node(cache, eb->start, 7330304f2d8SQu Wenruo lower->level + 1); 7344d81ea8bSQu Wenruo if (!upper) { 7354d81ea8bSQu Wenruo btrfs_put_root(root); 7364d81ea8bSQu Wenruo free_backref_edge(cache, edge); 7374d81ea8bSQu Wenruo ret = -ENOMEM; 7384d81ea8bSQu Wenruo goto out; 7394d81ea8bSQu Wenruo } 7404d81ea8bSQu Wenruo upper->owner = btrfs_header_owner(eb); 7414d81ea8bSQu Wenruo if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 7424d81ea8bSQu Wenruo upper->cowonly = 1; 7434d81ea8bSQu Wenruo 7444d81ea8bSQu Wenruo /* 7454d81ea8bSQu Wenruo * If we know the block isn't shared we can avoid 7464d81ea8bSQu Wenruo * checking its backrefs. 7474d81ea8bSQu Wenruo */ 7484d81ea8bSQu Wenruo if (btrfs_block_can_be_shared(root, eb)) 7494d81ea8bSQu Wenruo upper->checked = 0; 7504d81ea8bSQu Wenruo else 7514d81ea8bSQu Wenruo upper->checked = 1; 7524d81ea8bSQu Wenruo 7534d81ea8bSQu Wenruo /* 7544d81ea8bSQu Wenruo * Add the block to pending list if we need to check its 7554d81ea8bSQu Wenruo * backrefs, we only do this once while walking up a 7564d81ea8bSQu Wenruo * tree as we will catch anything else later on. 7574d81ea8bSQu Wenruo */ 7584d81ea8bSQu Wenruo if (!upper->checked && need_check) { 7594d81ea8bSQu Wenruo need_check = false; 7604d81ea8bSQu Wenruo list_add_tail(&edge->list[UPPER], 7614d81ea8bSQu Wenruo &cache->pending_edge); 7624d81ea8bSQu Wenruo } else { 7634d81ea8bSQu Wenruo if (upper->checked) 7644d81ea8bSQu Wenruo need_check = true; 7654d81ea8bSQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 7664d81ea8bSQu Wenruo } 7674d81ea8bSQu Wenruo } else { 768a26195a5SQu Wenruo upper = rb_entry(rb_node, struct btrfs_backref_node, 769a26195a5SQu Wenruo rb_node); 7704d81ea8bSQu Wenruo ASSERT(upper->checked); 7714d81ea8bSQu Wenruo INIT_LIST_HEAD(&edge->list[UPPER]); 7724d81ea8bSQu Wenruo if (!upper->owner) 7734d81ea8bSQu Wenruo upper->owner = btrfs_header_owner(eb); 7744d81ea8bSQu Wenruo } 7752a979612SQu Wenruo link_backref_edge(edge, lower, upper, LINK_LOWER); 7764d81ea8bSQu Wenruo 7774d81ea8bSQu Wenruo if (rb_node) { 7784d81ea8bSQu Wenruo btrfs_put_root(root); 7794d81ea8bSQu Wenruo break; 7804d81ea8bSQu Wenruo } 7814d81ea8bSQu Wenruo lower = upper; 7824d81ea8bSQu Wenruo upper = NULL; 7834d81ea8bSQu Wenruo } 7844d81ea8bSQu Wenruo out: 7854d81ea8bSQu Wenruo btrfs_release_path(path); 7864d81ea8bSQu Wenruo return ret; 7874d81ea8bSQu Wenruo } 7884d81ea8bSQu Wenruo 789a26195a5SQu Wenruo static int handle_one_tree_block(struct btrfs_backref_cache *cache, 790e7d571c7SQu Wenruo struct btrfs_path *path, 791e7d571c7SQu Wenruo struct btrfs_backref_iter *iter, 7925d4f98a2SYan Zheng struct btrfs_key *node_key, 793a26195a5SQu Wenruo struct btrfs_backref_node *cur) 7945d4f98a2SYan Zheng { 795e7d571c7SQu Wenruo struct btrfs_fs_info *fs_info = cache->fs_info; 796a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 797a26195a5SQu Wenruo struct btrfs_backref_node *exist; 7985d4f98a2SYan Zheng int ret; 7995d4f98a2SYan Zheng 80071f572a9SQu Wenruo ret = btrfs_backref_iter_start(iter, cur->bytenr); 801e7d571c7SQu Wenruo if (ret < 0) 802e7d571c7SQu Wenruo return ret; 80371f572a9SQu Wenruo /* 80471f572a9SQu Wenruo * We skip the first btrfs_tree_block_info, as we don't use the key 80571f572a9SQu Wenruo * stored in it, but fetch it from the tree block 80671f572a9SQu Wenruo */ 80771f572a9SQu Wenruo if (btrfs_backref_has_tree_block_info(iter)) { 80871f572a9SQu Wenruo ret = btrfs_backref_iter_next(iter); 809e7d571c7SQu Wenruo if (ret < 0) 81071f572a9SQu Wenruo goto out; 81171f572a9SQu Wenruo /* No extra backref? This means the tree block is corrupted */ 81271f572a9SQu Wenruo if (ret > 0) { 813e7d571c7SQu Wenruo ret = -EUCLEAN; 81471f572a9SQu Wenruo goto out; 81571f572a9SQu Wenruo } 81671f572a9SQu Wenruo } 8175d4f98a2SYan Zheng WARN_ON(cur->checked); 8185d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 8195d4f98a2SYan Zheng /* 82070f23fd6SJustin P. Mattock * the backref was added previously when processing 8215d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 8225d4f98a2SYan Zheng */ 82375bfb9afSJosef Bacik ASSERT(list_is_singular(&cur->upper)); 824a26195a5SQu Wenruo edge = list_entry(cur->upper.next, struct btrfs_backref_edge, 8255d4f98a2SYan Zheng list[LOWER]); 82675bfb9afSJosef Bacik ASSERT(list_empty(&edge->list[UPPER])); 8275d4f98a2SYan Zheng exist = edge->node[UPPER]; 8285d4f98a2SYan Zheng /* 8295d4f98a2SYan Zheng * add the upper level block to pending list if we need 8305d4f98a2SYan Zheng * check its backrefs 8315d4f98a2SYan Zheng */ 8325d4f98a2SYan Zheng if (!exist->checked) 83384780289SQu Wenruo list_add_tail(&edge->list[UPPER], &cache->pending_edge); 8345d4f98a2SYan Zheng } else { 8355d4f98a2SYan Zheng exist = NULL; 8365d4f98a2SYan Zheng } 8375d4f98a2SYan Zheng 83871f572a9SQu Wenruo for (; ret == 0; ret = btrfs_backref_iter_next(iter)) { 83971f572a9SQu Wenruo struct extent_buffer *eb; 84071f572a9SQu Wenruo struct btrfs_key key; 8413de28d57SLiu Bo int type; 84271f572a9SQu Wenruo 84371f572a9SQu Wenruo cond_resched(); 84471f572a9SQu Wenruo eb = btrfs_backref_get_eb(iter); 84571f572a9SQu Wenruo 84671f572a9SQu Wenruo key.objectid = iter->bytenr; 84771f572a9SQu Wenruo if (btrfs_backref_iter_is_inline_ref(iter)) { 84871f572a9SQu Wenruo struct btrfs_extent_inline_ref *iref; 84971f572a9SQu Wenruo 85071f572a9SQu Wenruo /* update key for inline back ref */ 85171f572a9SQu Wenruo iref = (struct btrfs_extent_inline_ref *) 85271f572a9SQu Wenruo ((unsigned long)iter->cur_ptr); 8533de28d57SLiu Bo type = btrfs_get_extent_inline_ref_type(eb, iref, 8543de28d57SLiu Bo BTRFS_REF_TYPE_BLOCK); 8553de28d57SLiu Bo if (type == BTRFS_REF_TYPE_INVALID) { 856e7d571c7SQu Wenruo ret = -EUCLEAN; 8573de28d57SLiu Bo goto out; 8583de28d57SLiu Bo } 8593de28d57SLiu Bo key.type = type; 8605d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 86171f572a9SQu Wenruo } else { 86271f572a9SQu Wenruo key.type = iter->cur_key.type; 86371f572a9SQu Wenruo key.offset = iter->cur_key.offset; 8645d4f98a2SYan Zheng } 8655d4f98a2SYan Zheng 866fa6ac715SQu Wenruo /* 867fa6ac715SQu Wenruo * Parent node found and matches current inline ref, no need to 868fa6ac715SQu Wenruo * rebuild this node for this inline ref. 869fa6ac715SQu Wenruo */ 8705d4f98a2SYan Zheng if (exist && 8715d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 8725d4f98a2SYan Zheng exist->owner == key.offset) || 8735d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 8745d4f98a2SYan Zheng exist->bytenr == key.offset))) { 8755d4f98a2SYan Zheng exist = NULL; 87671f572a9SQu Wenruo continue; 8775d4f98a2SYan Zheng } 8785d4f98a2SYan Zheng 879fa6ac715SQu Wenruo /* SHARED_BLOCK_REF means key.offset is the parent bytenr */ 8805d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 8814007ea87SQu Wenruo ret = handle_direct_tree_backref(cache, &key, cur); 882e7d571c7SQu Wenruo if (ret < 0) 8832433bea5SQu Wenruo goto out; 88471f572a9SQu Wenruo continue; 8856d8ff4e4SDavid Sterba } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) { 886e7d571c7SQu Wenruo ret = -EINVAL; 887e7d571c7SQu Wenruo btrfs_print_v0_err(fs_info); 888e7d571c7SQu Wenruo btrfs_handle_fs_error(fs_info, ret, NULL); 889ba3c2b19SNikolay Borisov goto out; 8905d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 89171f572a9SQu Wenruo continue; 8925d4f98a2SYan Zheng } 8935d4f98a2SYan Zheng 894fa6ac715SQu Wenruo /* 895fa6ac715SQu Wenruo * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset 896fa6ac715SQu Wenruo * means the root objectid. We need to search the tree to get 897fa6ac715SQu Wenruo * its parent bytenr. 898fa6ac715SQu Wenruo */ 8994d81ea8bSQu Wenruo ret = handle_indirect_tree_backref(cache, path, &key, node_key, 9004d81ea8bSQu Wenruo cur); 901e7d571c7SQu Wenruo if (ret < 0) 90271f572a9SQu Wenruo goto out; 9035d4f98a2SYan Zheng } 90471f572a9SQu Wenruo ret = 0; 9055d4f98a2SYan Zheng cur->checked = 1; 9065d4f98a2SYan Zheng WARN_ON(exist); 907e7d571c7SQu Wenruo out: 908e7d571c7SQu Wenruo btrfs_backref_iter_release(iter); 909e7d571c7SQu Wenruo return ret; 910e7d571c7SQu Wenruo } 9115d4f98a2SYan Zheng 912e7d571c7SQu Wenruo /* 9131f872924SQu Wenruo * In handle_one_tree_backref(), we have only linked the lower node to the edge, 9141f872924SQu Wenruo * but the upper node hasn't been linked to the edge. 915a26195a5SQu Wenruo * This means we can only iterate through btrfs_backref_node::upper to reach 916a26195a5SQu Wenruo * parent edges, but not through btrfs_backref_node::lower to reach children 917a26195a5SQu Wenruo * edges. 9181f872924SQu Wenruo * 919a26195a5SQu Wenruo * This function will finish the btrfs_backref_node::lower to related edges, 920a26195a5SQu Wenruo * so that backref cache can be bi-directionally iterated. 9211f872924SQu Wenruo * 9221f872924SQu Wenruo * Also, this will add the nodes to backref cache for the next run. 9231f872924SQu Wenruo */ 924a26195a5SQu Wenruo static int finish_upper_links(struct btrfs_backref_cache *cache, 925a26195a5SQu Wenruo struct btrfs_backref_node *start) 9261f872924SQu Wenruo { 9271f872924SQu Wenruo struct list_head *useless_node = &cache->useless_node; 928a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 9291f872924SQu Wenruo struct rb_node *rb_node; 9301f872924SQu Wenruo LIST_HEAD(pending_edge); 9311f872924SQu Wenruo 9321f872924SQu Wenruo ASSERT(start->checked); 9331f872924SQu Wenruo 9341f872924SQu Wenruo /* Insert this node to cache if it's not COW-only */ 9351f872924SQu Wenruo if (!start->cowonly) { 936e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, start->bytenr, 9371f872924SQu Wenruo &start->rb_node); 9381f872924SQu Wenruo if (rb_node) 9391f872924SQu Wenruo backref_tree_panic(rb_node, -EEXIST, start->bytenr); 9401f872924SQu Wenruo list_add_tail(&start->lower, &cache->leaves); 9411f872924SQu Wenruo } 9421f872924SQu Wenruo 9431f872924SQu Wenruo /* 9441f872924SQu Wenruo * Use breadth first search to iterate all related edges. 9451f872924SQu Wenruo * 9461f872924SQu Wenruo * The starting points are all the edges of this node 9471f872924SQu Wenruo */ 9481f872924SQu Wenruo list_for_each_entry(edge, &start->upper, list[LOWER]) 9491f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &pending_edge); 9501f872924SQu Wenruo 9511f872924SQu Wenruo while (!list_empty(&pending_edge)) { 952a26195a5SQu Wenruo struct btrfs_backref_node *upper; 953a26195a5SQu Wenruo struct btrfs_backref_node *lower; 9541f872924SQu Wenruo struct rb_node *rb_node; 9551f872924SQu Wenruo 956a26195a5SQu Wenruo edge = list_first_entry(&pending_edge, 957a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 9581f872924SQu Wenruo list_del_init(&edge->list[UPPER]); 9591f872924SQu Wenruo upper = edge->node[UPPER]; 9601f872924SQu Wenruo lower = edge->node[LOWER]; 9611f872924SQu Wenruo 9621f872924SQu Wenruo /* Parent is detached, no need to keep any edges */ 9631f872924SQu Wenruo if (upper->detached) { 9641f872924SQu Wenruo list_del(&edge->list[LOWER]); 9651f872924SQu Wenruo free_backref_edge(cache, edge); 9661f872924SQu Wenruo 9671f872924SQu Wenruo /* Lower node is orphan, queue for cleanup */ 9681f872924SQu Wenruo if (list_empty(&lower->upper)) 9691f872924SQu Wenruo list_add(&lower->list, useless_node); 9701f872924SQu Wenruo continue; 9711f872924SQu Wenruo } 9721f872924SQu Wenruo 9731f872924SQu Wenruo /* 9741f872924SQu Wenruo * All new nodes added in current build_backref_tree() haven't 9751f872924SQu Wenruo * been linked to the cache rb tree. 9761f872924SQu Wenruo * So if we have upper->rb_node populated, this means a cache 9771f872924SQu Wenruo * hit. We only need to link the edge, as @upper and all its 9781f872924SQu Wenruo * parent have already been linked. 9791f872924SQu Wenruo */ 9801f872924SQu Wenruo if (!RB_EMPTY_NODE(&upper->rb_node)) { 9811f872924SQu Wenruo if (upper->lowest) { 9821f872924SQu Wenruo list_del_init(&upper->lower); 9831f872924SQu Wenruo upper->lowest = 0; 9841f872924SQu Wenruo } 9851f872924SQu Wenruo 9861f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 9871f872924SQu Wenruo continue; 9881f872924SQu Wenruo } 9891f872924SQu Wenruo 9901f872924SQu Wenruo /* Sanity check, we shouldn't have any unchecked nodes */ 9911f872924SQu Wenruo if (!upper->checked) { 9921f872924SQu Wenruo ASSERT(0); 9931f872924SQu Wenruo return -EUCLEAN; 9941f872924SQu Wenruo } 9951f872924SQu Wenruo 9961f872924SQu Wenruo /* Sanity check, COW-only node has non-COW-only parent */ 9971f872924SQu Wenruo if (start->cowonly != upper->cowonly) { 9981f872924SQu Wenruo ASSERT(0); 9991f872924SQu Wenruo return -EUCLEAN; 10001f872924SQu Wenruo } 10011f872924SQu Wenruo 10021f872924SQu Wenruo /* Only cache non-COW-only (subvolume trees) tree blocks */ 10031f872924SQu Wenruo if (!upper->cowonly) { 1004e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr, 10051f872924SQu Wenruo &upper->rb_node); 10061f872924SQu Wenruo if (rb_node) { 10071f872924SQu Wenruo backref_tree_panic(rb_node, -EEXIST, 10081f872924SQu Wenruo upper->bytenr); 10091f872924SQu Wenruo return -EUCLEAN; 10101f872924SQu Wenruo } 10111f872924SQu Wenruo } 10121f872924SQu Wenruo 10131f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &upper->lower); 10141f872924SQu Wenruo 10151f872924SQu Wenruo /* 10161f872924SQu Wenruo * Also queue all the parent edges of this uncached node to 10171f872924SQu Wenruo * finish the upper linkage 10181f872924SQu Wenruo */ 10191f872924SQu Wenruo list_for_each_entry(edge, &upper->upper, list[LOWER]) 10201f872924SQu Wenruo list_add_tail(&edge->list[UPPER], &pending_edge); 10211f872924SQu Wenruo } 10221f872924SQu Wenruo return 0; 10231f872924SQu Wenruo } 10241f872924SQu Wenruo 10251f872924SQu Wenruo /* 102629db137bSQu Wenruo * For useless nodes, do two major clean ups: 102729db137bSQu Wenruo * 102829db137bSQu Wenruo * - Cleanup the children edges and nodes 102929db137bSQu Wenruo * If child node is also orphan (no parent) during cleanup, then the child 103029db137bSQu Wenruo * node will also be cleaned up. 103129db137bSQu Wenruo * 103229db137bSQu Wenruo * - Freeing up leaves (level 0), keeps nodes detached 103329db137bSQu Wenruo * For nodes, the node is still cached as "detached" 103429db137bSQu Wenruo * 103529db137bSQu Wenruo * Return false if @node is not in the @useless_nodes list. 103629db137bSQu Wenruo * Return true if @node is in the @useless_nodes list. 103729db137bSQu Wenruo */ 103829db137bSQu Wenruo static bool handle_useless_nodes(struct reloc_control *rc, 1039a26195a5SQu Wenruo struct btrfs_backref_node *node) 104029db137bSQu Wenruo { 1041a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 104229db137bSQu Wenruo struct list_head *useless_node = &cache->useless_node; 104329db137bSQu Wenruo bool ret = false; 104429db137bSQu Wenruo 104529db137bSQu Wenruo while (!list_empty(useless_node)) { 1046a26195a5SQu Wenruo struct btrfs_backref_node *cur; 104729db137bSQu Wenruo 1048a26195a5SQu Wenruo cur = list_first_entry(useless_node, struct btrfs_backref_node, 104929db137bSQu Wenruo list); 105029db137bSQu Wenruo list_del_init(&cur->list); 105129db137bSQu Wenruo 105229db137bSQu Wenruo /* Only tree root nodes can be added to @useless_nodes */ 105329db137bSQu Wenruo ASSERT(list_empty(&cur->upper)); 105429db137bSQu Wenruo 105529db137bSQu Wenruo if (cur == node) 105629db137bSQu Wenruo ret = true; 105729db137bSQu Wenruo 105829db137bSQu Wenruo /* The node is the lowest node */ 105929db137bSQu Wenruo if (cur->lowest) { 106029db137bSQu Wenruo list_del_init(&cur->lower); 106129db137bSQu Wenruo cur->lowest = 0; 106229db137bSQu Wenruo } 106329db137bSQu Wenruo 106429db137bSQu Wenruo /* Cleanup the lower edges */ 106529db137bSQu Wenruo while (!list_empty(&cur->lower)) { 1066a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1067a26195a5SQu Wenruo struct btrfs_backref_node *lower; 106829db137bSQu Wenruo 106929db137bSQu Wenruo edge = list_entry(cur->lower.next, 1070a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 107129db137bSQu Wenruo list_del(&edge->list[UPPER]); 107229db137bSQu Wenruo list_del(&edge->list[LOWER]); 107329db137bSQu Wenruo lower = edge->node[LOWER]; 107429db137bSQu Wenruo free_backref_edge(cache, edge); 107529db137bSQu Wenruo 107629db137bSQu Wenruo /* Child node is also orphan, queue for cleanup */ 107729db137bSQu Wenruo if (list_empty(&lower->upper)) 107829db137bSQu Wenruo list_add(&lower->list, useless_node); 107929db137bSQu Wenruo } 108029db137bSQu Wenruo /* Mark this block processed for relocation */ 108129db137bSQu Wenruo mark_block_processed(rc, cur); 108229db137bSQu Wenruo 108329db137bSQu Wenruo /* 108429db137bSQu Wenruo * Backref nodes for tree leaves are deleted from the cache. 108529db137bSQu Wenruo * Backref nodes for upper level tree blocks are left in the 108629db137bSQu Wenruo * cache to avoid unnecessary backref lookup. 108729db137bSQu Wenruo */ 108829db137bSQu Wenruo if (cur->level > 0) { 108929db137bSQu Wenruo list_add(&cur->list, &cache->detached); 109029db137bSQu Wenruo cur->detached = 1; 109129db137bSQu Wenruo } else { 109229db137bSQu Wenruo rb_erase(&cur->rb_node, &cache->rb_root); 109329db137bSQu Wenruo free_backref_node(cache, cur); 109429db137bSQu Wenruo } 109529db137bSQu Wenruo } 109629db137bSQu Wenruo return ret; 109729db137bSQu Wenruo } 109829db137bSQu Wenruo 109929db137bSQu Wenruo /* 1100e7d571c7SQu Wenruo * Build backref tree for a given tree block. Root of the backref tree 1101e7d571c7SQu Wenruo * corresponds the tree block, leaves of the backref tree correspond roots of 1102e7d571c7SQu Wenruo * b-trees that reference the tree block. 1103e7d571c7SQu Wenruo * 1104e7d571c7SQu Wenruo * The basic idea of this function is check backrefs of a given block to find 1105e7d571c7SQu Wenruo * upper level blocks that reference the block, and then check backrefs of 1106e7d571c7SQu Wenruo * these upper level blocks recursively. The recursion stops when tree root is 1107e7d571c7SQu Wenruo * reached or backrefs for the block is cached. 1108e7d571c7SQu Wenruo * 1109e7d571c7SQu Wenruo * NOTE: if we find that backrefs for a block are cached, we know backrefs for 1110e7d571c7SQu Wenruo * all upper level blocks that directly/indirectly reference the block are also 1111e7d571c7SQu Wenruo * cached. 1112e7d571c7SQu Wenruo */ 1113a26195a5SQu Wenruo static noinline_for_stack struct btrfs_backref_node *build_backref_tree( 1114e7d571c7SQu Wenruo struct reloc_control *rc, struct btrfs_key *node_key, 1115e7d571c7SQu Wenruo int level, u64 bytenr) 1116e7d571c7SQu Wenruo { 1117e7d571c7SQu Wenruo struct btrfs_backref_iter *iter; 1118a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 1119e7d571c7SQu Wenruo /* For searching parent of TREE_BLOCK_REF */ 1120e7d571c7SQu Wenruo struct btrfs_path *path; 1121a26195a5SQu Wenruo struct btrfs_backref_node *cur; 1122a26195a5SQu Wenruo struct btrfs_backref_node *upper; 1123a26195a5SQu Wenruo struct btrfs_backref_node *lower; 1124a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 1125a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1126e7d571c7SQu Wenruo int ret; 1127e7d571c7SQu Wenruo int err = 0; 1128e7d571c7SQu Wenruo 1129e7d571c7SQu Wenruo iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS); 1130e7d571c7SQu Wenruo if (!iter) 1131e7d571c7SQu Wenruo return ERR_PTR(-ENOMEM); 1132e7d571c7SQu Wenruo path = btrfs_alloc_path(); 1133e7d571c7SQu Wenruo if (!path) { 1134e7d571c7SQu Wenruo err = -ENOMEM; 1135e7d571c7SQu Wenruo goto out; 1136e7d571c7SQu Wenruo } 1137e7d571c7SQu Wenruo 1138e7d571c7SQu Wenruo node = alloc_backref_node(cache, bytenr, level); 1139e7d571c7SQu Wenruo if (!node) { 1140e7d571c7SQu Wenruo err = -ENOMEM; 1141e7d571c7SQu Wenruo goto out; 1142e7d571c7SQu Wenruo } 1143e7d571c7SQu Wenruo 1144e7d571c7SQu Wenruo node->lowest = 1; 1145e7d571c7SQu Wenruo cur = node; 1146e7d571c7SQu Wenruo 1147e7d571c7SQu Wenruo /* Breadth-first search to build backref cache */ 1148e7d571c7SQu Wenruo do { 1149e7d571c7SQu Wenruo ret = handle_one_tree_block(cache, path, iter, node_key, cur); 1150e7d571c7SQu Wenruo if (ret < 0) { 1151e7d571c7SQu Wenruo err = ret; 1152e7d571c7SQu Wenruo goto out; 1153e7d571c7SQu Wenruo } 1154e7d571c7SQu Wenruo edge = list_first_entry_or_null(&cache->pending_edge, 1155a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 1156e7d571c7SQu Wenruo /* 1157e7d571c7SQu Wenruo * The pending list isn't empty, take the first block to 1158e7d571c7SQu Wenruo * process 1159e7d571c7SQu Wenruo */ 1160e7d571c7SQu Wenruo if (edge) { 11615d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 11625d4f98a2SYan Zheng cur = edge->node[UPPER]; 11635d4f98a2SYan Zheng } 1164e7d571c7SQu Wenruo } while (edge); 11655d4f98a2SYan Zheng 11661f872924SQu Wenruo /* Finish the upper linkage of newly added edges/nodes */ 11671f872924SQu Wenruo ret = finish_upper_links(cache, node); 11681f872924SQu Wenruo if (ret < 0) { 11691f872924SQu Wenruo err = ret; 117075bfb9afSJosef Bacik goto out; 117175bfb9afSJosef Bacik } 117275bfb9afSJosef Bacik 117329db137bSQu Wenruo if (handle_useless_nodes(rc, node)) 11743fd0a558SYan, Zheng node = NULL; 11755d4f98a2SYan Zheng out: 117671f572a9SQu Wenruo btrfs_backref_iter_free(iter); 117771f572a9SQu Wenruo btrfs_free_path(path); 11785d4f98a2SYan Zheng if (err) { 117984780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 118084780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 1181a26195a5SQu Wenruo struct btrfs_backref_node, list); 118275bfb9afSJosef Bacik list_del_init(&lower->list); 11833fd0a558SYan, Zheng } 118484780289SQu Wenruo while (!list_empty(&cache->pending_edge)) { 118584780289SQu Wenruo edge = list_first_entry(&cache->pending_edge, 1186a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 118775bfb9afSJosef Bacik list_del(&edge->list[UPPER]); 11883fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 118975bfb9afSJosef Bacik lower = edge->node[LOWER]; 11905d4f98a2SYan Zheng upper = edge->node[UPPER]; 11913fd0a558SYan, Zheng free_backref_edge(cache, edge); 119275bfb9afSJosef Bacik 119375bfb9afSJosef Bacik /* 119475bfb9afSJosef Bacik * Lower is no longer linked to any upper backref nodes 119575bfb9afSJosef Bacik * and isn't in the cache, we can free it ourselves. 119675bfb9afSJosef Bacik */ 119775bfb9afSJosef Bacik if (list_empty(&lower->upper) && 119875bfb9afSJosef Bacik RB_EMPTY_NODE(&lower->rb_node)) 119984780289SQu Wenruo list_add(&lower->list, &cache->useless_node); 120075bfb9afSJosef Bacik 120175bfb9afSJosef Bacik if (!RB_EMPTY_NODE(&upper->rb_node)) 120275bfb9afSJosef Bacik continue; 120375bfb9afSJosef Bacik 120401327610SNicholas D Steeves /* Add this guy's upper edges to the list to process */ 120575bfb9afSJosef Bacik list_for_each_entry(edge, &upper->upper, list[LOWER]) 120684780289SQu Wenruo list_add_tail(&edge->list[UPPER], 120784780289SQu Wenruo &cache->pending_edge); 120875bfb9afSJosef Bacik if (list_empty(&upper->upper)) 120984780289SQu Wenruo list_add(&upper->list, &cache->useless_node); 121075bfb9afSJosef Bacik } 121175bfb9afSJosef Bacik 121284780289SQu Wenruo while (!list_empty(&cache->useless_node)) { 121384780289SQu Wenruo lower = list_first_entry(&cache->useless_node, 1214a26195a5SQu Wenruo struct btrfs_backref_node, list); 121575bfb9afSJosef Bacik list_del_init(&lower->list); 12160fd8c3daSLiu Bo if (lower == node) 12170fd8c3daSLiu Bo node = NULL; 121875bfb9afSJosef Bacik free_backref_node(cache, lower); 12195d4f98a2SYan Zheng } 12200fd8c3daSLiu Bo 12218e19c973SJosef Bacik remove_backref_node(cache, node); 122284780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 122384780289SQu Wenruo list_empty(&cache->pending_edge)); 12245d4f98a2SYan Zheng return ERR_PTR(err); 12255d4f98a2SYan Zheng } 122675bfb9afSJosef Bacik ASSERT(!node || !node->detached); 122784780289SQu Wenruo ASSERT(list_empty(&cache->useless_node) && 122884780289SQu Wenruo list_empty(&cache->pending_edge)); 12295d4f98a2SYan Zheng return node; 12305d4f98a2SYan Zheng } 12315d4f98a2SYan Zheng 12325d4f98a2SYan Zheng /* 12333fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 12343fd0a558SYan, Zheng * the backref node is created by cloning backref node that 12353fd0a558SYan, Zheng * corresponds to root of source tree 12363fd0a558SYan, Zheng */ 12373fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 12383fd0a558SYan, Zheng struct reloc_control *rc, 12393fd0a558SYan, Zheng struct btrfs_root *src, 12403fd0a558SYan, Zheng struct btrfs_root *dest) 12413fd0a558SYan, Zheng { 12423fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 1243a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 1244a26195a5SQu Wenruo struct btrfs_backref_node *node = NULL; 1245a26195a5SQu Wenruo struct btrfs_backref_node *new_node; 1246a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 1247a26195a5SQu Wenruo struct btrfs_backref_edge *new_edge; 12483fd0a558SYan, Zheng struct rb_node *rb_node; 12493fd0a558SYan, Zheng 12503fd0a558SYan, Zheng if (cache->last_trans > 0) 12513fd0a558SYan, Zheng update_backref_cache(trans, cache); 12523fd0a558SYan, Zheng 1253e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start); 12543fd0a558SYan, Zheng if (rb_node) { 1255a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, rb_node); 12563fd0a558SYan, Zheng if (node->detached) 12573fd0a558SYan, Zheng node = NULL; 12583fd0a558SYan, Zheng else 12593fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 12603fd0a558SYan, Zheng } 12613fd0a558SYan, Zheng 12623fd0a558SYan, Zheng if (!node) { 1263e9a28dc5SQu Wenruo rb_node = rb_simple_search(&cache->rb_root, 12643fd0a558SYan, Zheng reloc_root->commit_root->start); 12653fd0a558SYan, Zheng if (rb_node) { 1266a26195a5SQu Wenruo node = rb_entry(rb_node, struct btrfs_backref_node, 12673fd0a558SYan, Zheng rb_node); 12683fd0a558SYan, Zheng BUG_ON(node->detached); 12693fd0a558SYan, Zheng } 12703fd0a558SYan, Zheng } 12713fd0a558SYan, Zheng 12723fd0a558SYan, Zheng if (!node) 12733fd0a558SYan, Zheng return 0; 12743fd0a558SYan, Zheng 12750304f2d8SQu Wenruo new_node = alloc_backref_node(cache, dest->node->start, node->level); 12763fd0a558SYan, Zheng if (!new_node) 12773fd0a558SYan, Zheng return -ENOMEM; 12783fd0a558SYan, Zheng 12793fd0a558SYan, Zheng new_node->lowest = node->lowest; 12806848ad64SYan, Zheng new_node->checked = 1; 128100246528SJosef Bacik new_node->root = btrfs_grab_root(dest); 12820b530bc5SJosef Bacik ASSERT(new_node->root); 12833fd0a558SYan, Zheng 12843fd0a558SYan, Zheng if (!node->lowest) { 12853fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 12863fd0a558SYan, Zheng new_edge = alloc_backref_edge(cache); 12873fd0a558SYan, Zheng if (!new_edge) 12883fd0a558SYan, Zheng goto fail; 12893fd0a558SYan, Zheng 12902a979612SQu Wenruo link_backref_edge(new_edge, edge->node[LOWER], new_node, 12912a979612SQu Wenruo LINK_UPPER); 12923fd0a558SYan, Zheng } 129376b9e23dSMiao Xie } else { 129476b9e23dSMiao Xie list_add_tail(&new_node->lower, &cache->leaves); 12953fd0a558SYan, Zheng } 12963fd0a558SYan, Zheng 1297e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr, 12983fd0a558SYan, Zheng &new_node->rb_node); 129943c04fb1SJeff Mahoney if (rb_node) 130043c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, new_node->bytenr); 13013fd0a558SYan, Zheng 13023fd0a558SYan, Zheng if (!new_node->lowest) { 13033fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 13043fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 13053fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 13063fd0a558SYan, Zheng } 13073fd0a558SYan, Zheng } 13083fd0a558SYan, Zheng return 0; 13093fd0a558SYan, Zheng fail: 13103fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 13113fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 1312a26195a5SQu Wenruo struct btrfs_backref_edge, list[UPPER]); 13133fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 13143fd0a558SYan, Zheng free_backref_edge(cache, new_edge); 13153fd0a558SYan, Zheng } 13163fd0a558SYan, Zheng free_backref_node(cache, new_node); 13173fd0a558SYan, Zheng return -ENOMEM; 13183fd0a558SYan, Zheng } 13193fd0a558SYan, Zheng 13203fd0a558SYan, Zheng /* 13215d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 13225d4f98a2SYan Zheng */ 1323ffd7b339SJeff Mahoney static int __must_check __add_reloc_root(struct btrfs_root *root) 13245d4f98a2SYan Zheng { 13250b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 13265d4f98a2SYan Zheng struct rb_node *rb_node; 13275d4f98a2SYan Zheng struct mapping_node *node; 13280b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 13295d4f98a2SYan Zheng 13305d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 1331ffd7b339SJeff Mahoney if (!node) 1332ffd7b339SJeff Mahoney return -ENOMEM; 13335d4f98a2SYan Zheng 1334ea287ab1SJosef Bacik node->bytenr = root->commit_root->start; 13355d4f98a2SYan Zheng node->data = root; 13365d4f98a2SYan Zheng 13375d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1338e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 13395d4f98a2SYan Zheng node->bytenr, &node->rb_node); 13405d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 1341ffd7b339SJeff Mahoney if (rb_node) { 13420b246afaSJeff Mahoney btrfs_panic(fs_info, -EEXIST, 13435d163e0eSJeff Mahoney "Duplicate root found for start=%llu while inserting into relocation tree", 13445d163e0eSJeff Mahoney node->bytenr); 1345ffd7b339SJeff Mahoney } 13465d4f98a2SYan Zheng 13475d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 13485d4f98a2SYan Zheng return 0; 13495d4f98a2SYan Zheng } 13505d4f98a2SYan Zheng 13515d4f98a2SYan Zheng /* 1352c974c464SWang Shilong * helper to delete the 'address of tree root -> reloc tree' 13535d4f98a2SYan Zheng * mapping 13545d4f98a2SYan Zheng */ 1355c974c464SWang Shilong static void __del_reloc_root(struct btrfs_root *root) 13565d4f98a2SYan Zheng { 13570b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 13585d4f98a2SYan Zheng struct rb_node *rb_node; 13595d4f98a2SYan Zheng struct mapping_node *node = NULL; 13600b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1361f44deb74SJosef Bacik bool put_ref = false; 13625d4f98a2SYan Zheng 136365c6e82bSQu Wenruo if (rc && root->node) { 13645d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1365e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 1366ea287ab1SJosef Bacik root->commit_root->start); 1367c974c464SWang Shilong if (rb_node) { 1368c974c464SWang Shilong node = rb_entry(rb_node, struct mapping_node, rb_node); 1369c974c464SWang Shilong rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 1370ea287ab1SJosef Bacik RB_CLEAR_NODE(&node->rb_node); 1371c974c464SWang Shilong } 1372c974c464SWang Shilong spin_unlock(&rc->reloc_root_tree.lock); 1373c974c464SWang Shilong if (!node) 1374c974c464SWang Shilong return; 1375c974c464SWang Shilong BUG_ON((struct btrfs_root *)node->data != root); 1376389305b2SQu Wenruo } 1377c974c464SWang Shilong 1378f44deb74SJosef Bacik /* 1379f44deb74SJosef Bacik * We only put the reloc root here if it's on the list. There's a lot 1380f44deb74SJosef Bacik * of places where the pattern is to splice the rc->reloc_roots, process 1381f44deb74SJosef Bacik * the reloc roots, and then add the reloc root back onto 1382f44deb74SJosef Bacik * rc->reloc_roots. If we call __del_reloc_root while it's off of the 1383f44deb74SJosef Bacik * list we don't want the reference being dropped, because the guy 1384f44deb74SJosef Bacik * messing with the list is in charge of the reference. 1385f44deb74SJosef Bacik */ 13860b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 1387f44deb74SJosef Bacik if (!list_empty(&root->root_list)) { 1388f44deb74SJosef Bacik put_ref = true; 1389c974c464SWang Shilong list_del_init(&root->root_list); 1390f44deb74SJosef Bacik } 13910b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 1392f44deb74SJosef Bacik if (put_ref) 1393f44deb74SJosef Bacik btrfs_put_root(root); 1394c974c464SWang Shilong kfree(node); 1395c974c464SWang Shilong } 1396c974c464SWang Shilong 1397c974c464SWang Shilong /* 1398c974c464SWang Shilong * helper to update the 'address of tree root -> reloc tree' 1399c974c464SWang Shilong * mapping 1400c974c464SWang Shilong */ 1401ea287ab1SJosef Bacik static int __update_reloc_root(struct btrfs_root *root) 1402c974c464SWang Shilong { 14030b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 1404c974c464SWang Shilong struct rb_node *rb_node; 1405c974c464SWang Shilong struct mapping_node *node = NULL; 14060b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 1407c974c464SWang Shilong 1408c974c464SWang Shilong spin_lock(&rc->reloc_root_tree.lock); 1409e9a28dc5SQu Wenruo rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, 1410ea287ab1SJosef Bacik root->commit_root->start); 14115d4f98a2SYan Zheng if (rb_node) { 14125d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 14135d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 14145d4f98a2SYan Zheng } 14155d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 14165d4f98a2SYan Zheng 14178f71f3e0SLiu Bo if (!node) 14188f71f3e0SLiu Bo return 0; 14195d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 14205d4f98a2SYan Zheng 14215d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 1422ea287ab1SJosef Bacik node->bytenr = root->node->start; 1423e9a28dc5SQu Wenruo rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root, 14245d4f98a2SYan Zheng node->bytenr, &node->rb_node); 14255d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 142643c04fb1SJeff Mahoney if (rb_node) 142743c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 14285d4f98a2SYan Zheng return 0; 14295d4f98a2SYan Zheng } 14305d4f98a2SYan Zheng 14313fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 14323fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 14335d4f98a2SYan Zheng { 14340b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 14355d4f98a2SYan Zheng struct btrfs_root *reloc_root; 14365d4f98a2SYan Zheng struct extent_buffer *eb; 14375d4f98a2SYan Zheng struct btrfs_root_item *root_item; 14385d4f98a2SYan Zheng struct btrfs_key root_key; 14395d4f98a2SYan Zheng int ret; 14405d4f98a2SYan Zheng 14415d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 14425d4f98a2SYan Zheng BUG_ON(!root_item); 14435d4f98a2SYan Zheng 14445d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 14455d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 14463fd0a558SYan, Zheng root_key.offset = objectid; 14475d4f98a2SYan Zheng 14483fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 1449054570a1SFilipe Manana u64 commit_root_gen; 1450054570a1SFilipe Manana 14513fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 14525d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 14535d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 14545d4f98a2SYan Zheng BUG_ON(ret); 1455054570a1SFilipe Manana /* 1456054570a1SFilipe Manana * Set the last_snapshot field to the generation of the commit 1457054570a1SFilipe Manana * root - like this ctree.c:btrfs_block_can_be_shared() behaves 1458054570a1SFilipe Manana * correctly (returns true) when the relocation root is created 1459054570a1SFilipe Manana * either inside the critical section of a transaction commit 1460054570a1SFilipe Manana * (through transaction.c:qgroup_account_snapshot()) and when 1461054570a1SFilipe Manana * it's created before the transaction commit is started. 1462054570a1SFilipe Manana */ 1463054570a1SFilipe Manana commit_root_gen = btrfs_header_generation(root->commit_root); 1464054570a1SFilipe Manana btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen); 14653fd0a558SYan, Zheng } else { 14663fd0a558SYan, Zheng /* 14673fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 14683fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 14693fd0a558SYan, Zheng * modified after it was created have RELOC flag 14703fd0a558SYan, Zheng * set in their headers. so it's OK to not update 14713fd0a558SYan, Zheng * the 'last_snapshot'. 14723fd0a558SYan, Zheng */ 14733fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 14743fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 14753fd0a558SYan, Zheng BUG_ON(ret); 14763fd0a558SYan, Zheng } 14773fd0a558SYan, Zheng 14785d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 14795d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 14805d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 14815d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 14823fd0a558SYan, Zheng 14833fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 14843fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 14853fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 14863fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 14875d4f98a2SYan Zheng root_item->drop_level = 0; 14883fd0a558SYan, Zheng } 14895d4f98a2SYan Zheng 14905d4f98a2SYan Zheng btrfs_tree_unlock(eb); 14915d4f98a2SYan Zheng free_extent_buffer(eb); 14925d4f98a2SYan Zheng 14930b246afaSJeff Mahoney ret = btrfs_insert_root(trans, fs_info->tree_root, 14945d4f98a2SYan Zheng &root_key, root_item); 14955d4f98a2SYan Zheng BUG_ON(ret); 14965d4f98a2SYan Zheng kfree(root_item); 14975d4f98a2SYan Zheng 14983dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key); 14995d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 15003dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 15015d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 15023fd0a558SYan, Zheng return reloc_root; 15033fd0a558SYan, Zheng } 15043fd0a558SYan, Zheng 15053fd0a558SYan, Zheng /* 15063fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 15073fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 1508f44deb74SJosef Bacik * 1509f44deb74SJosef Bacik * The reloc_root comes out of here with two references, one for 1510f44deb74SJosef Bacik * root->reloc_root, and another for being on the rc->reloc_roots list. 15113fd0a558SYan, Zheng */ 15123fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 15133fd0a558SYan, Zheng struct btrfs_root *root) 15143fd0a558SYan, Zheng { 15150b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15163fd0a558SYan, Zheng struct btrfs_root *reloc_root; 15170b246afaSJeff Mahoney struct reloc_control *rc = fs_info->reloc_ctl; 151820dd2cbfSMiao Xie struct btrfs_block_rsv *rsv; 15193fd0a558SYan, Zheng int clear_rsv = 0; 1520ffd7b339SJeff Mahoney int ret; 15213fd0a558SYan, Zheng 1522aec7db3bSJosef Bacik if (!rc) 15232abc726aSJosef Bacik return 0; 15242abc726aSJosef Bacik 15251fac4a54SQu Wenruo /* 15261fac4a54SQu Wenruo * The subvolume has reloc tree but the swap is finished, no need to 15271fac4a54SQu Wenruo * create/update the dead reloc tree 15281fac4a54SQu Wenruo */ 15296282675eSQu Wenruo if (reloc_root_is_dead(root)) 15301fac4a54SQu Wenruo return 0; 15311fac4a54SQu Wenruo 1532aec7db3bSJosef Bacik /* 1533aec7db3bSJosef Bacik * This is subtle but important. We do not do 1534aec7db3bSJosef Bacik * record_root_in_transaction for reloc roots, instead we record their 1535aec7db3bSJosef Bacik * corresponding fs root, and then here we update the last trans for the 1536aec7db3bSJosef Bacik * reloc root. This means that we have to do this for the entire life 1537aec7db3bSJosef Bacik * of the reloc root, regardless of which stage of the relocation we are 1538aec7db3bSJosef Bacik * in. 1539aec7db3bSJosef Bacik */ 15403fd0a558SYan, Zheng if (root->reloc_root) { 15413fd0a558SYan, Zheng reloc_root = root->reloc_root; 15423fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 15433fd0a558SYan, Zheng return 0; 15443fd0a558SYan, Zheng } 15453fd0a558SYan, Zheng 1546aec7db3bSJosef Bacik /* 1547aec7db3bSJosef Bacik * We are merging reloc roots, we do not need new reloc trees. Also 1548aec7db3bSJosef Bacik * reloc trees never need their own reloc tree. 1549aec7db3bSJosef Bacik */ 1550aec7db3bSJosef Bacik if (!rc->create_reloc_tree || 1551aec7db3bSJosef Bacik root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 1552aec7db3bSJosef Bacik return 0; 1553aec7db3bSJosef Bacik 155420dd2cbfSMiao Xie if (!trans->reloc_reserved) { 155520dd2cbfSMiao Xie rsv = trans->block_rsv; 15563fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 15573fd0a558SYan, Zheng clear_rsv = 1; 15583fd0a558SYan, Zheng } 15593fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 15603fd0a558SYan, Zheng if (clear_rsv) 156120dd2cbfSMiao Xie trans->block_rsv = rsv; 15625d4f98a2SYan Zheng 1563ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 1564ffd7b339SJeff Mahoney BUG_ON(ret < 0); 1565f44deb74SJosef Bacik root->reloc_root = btrfs_grab_root(reloc_root); 15665d4f98a2SYan Zheng return 0; 15675d4f98a2SYan Zheng } 15685d4f98a2SYan Zheng 15695d4f98a2SYan Zheng /* 15705d4f98a2SYan Zheng * update root item of reloc tree 15715d4f98a2SYan Zheng */ 15725d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 15735d4f98a2SYan Zheng struct btrfs_root *root) 15745d4f98a2SYan Zheng { 15750b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15765d4f98a2SYan Zheng struct btrfs_root *reloc_root; 15775d4f98a2SYan Zheng struct btrfs_root_item *root_item; 15785d4f98a2SYan Zheng int ret; 15795d4f98a2SYan Zheng 15806282675eSQu Wenruo if (!have_reloc_root(root)) 15817585717fSChris Mason goto out; 15825d4f98a2SYan Zheng 15835d4f98a2SYan Zheng reloc_root = root->reloc_root; 15845d4f98a2SYan Zheng root_item = &reloc_root->root_item; 15855d4f98a2SYan Zheng 1586f44deb74SJosef Bacik /* 1587f44deb74SJosef Bacik * We are probably ok here, but __del_reloc_root() will drop its ref of 1588f44deb74SJosef Bacik * the root. We have the ref for root->reloc_root, but just in case 1589f44deb74SJosef Bacik * hold it while we update the reloc root. 1590f44deb74SJosef Bacik */ 1591f44deb74SJosef Bacik btrfs_grab_root(reloc_root); 1592f44deb74SJosef Bacik 1593d2311e69SQu Wenruo /* root->reloc_root will stay until current relocation finished */ 15940b246afaSJeff Mahoney if (fs_info->reloc_ctl->merge_reloc_tree && 15953fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 1596d2311e69SQu Wenruo set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 15976282675eSQu Wenruo /* 15986282675eSQu Wenruo * Mark the tree as dead before we change reloc_root so 15996282675eSQu Wenruo * have_reloc_root will not touch it from now on. 16006282675eSQu Wenruo */ 16016282675eSQu Wenruo smp_wmb(); 1602c974c464SWang Shilong __del_reloc_root(reloc_root); 16035d4f98a2SYan Zheng } 16045d4f98a2SYan Zheng 16055d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 1606ea287ab1SJosef Bacik __update_reloc_root(reloc_root); 16075d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 16085d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 16095d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 16105d4f98a2SYan Zheng } 16115d4f98a2SYan Zheng 16120b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 16135d4f98a2SYan Zheng &reloc_root->root_key, root_item); 16145d4f98a2SYan Zheng BUG_ON(ret); 1615f44deb74SJosef Bacik btrfs_put_root(reloc_root); 16167585717fSChris Mason out: 16175d4f98a2SYan Zheng return 0; 16185d4f98a2SYan Zheng } 16195d4f98a2SYan Zheng 16205d4f98a2SYan Zheng /* 16215d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 16225d4f98a2SYan Zheng * in a subvolume 16235d4f98a2SYan Zheng */ 16245d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 16255d4f98a2SYan Zheng { 16265d4f98a2SYan Zheng struct rb_node *node; 16275d4f98a2SYan Zheng struct rb_node *prev; 16285d4f98a2SYan Zheng struct btrfs_inode *entry; 16295d4f98a2SYan Zheng struct inode *inode; 16305d4f98a2SYan Zheng 16315d4f98a2SYan Zheng spin_lock(&root->inode_lock); 16325d4f98a2SYan Zheng again: 16335d4f98a2SYan Zheng node = root->inode_tree.rb_node; 16345d4f98a2SYan Zheng prev = NULL; 16355d4f98a2SYan Zheng while (node) { 16365d4f98a2SYan Zheng prev = node; 16375d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 16385d4f98a2SYan Zheng 16394a0cc7caSNikolay Borisov if (objectid < btrfs_ino(entry)) 16405d4f98a2SYan Zheng node = node->rb_left; 16414a0cc7caSNikolay Borisov else if (objectid > btrfs_ino(entry)) 16425d4f98a2SYan Zheng node = node->rb_right; 16435d4f98a2SYan Zheng else 16445d4f98a2SYan Zheng break; 16455d4f98a2SYan Zheng } 16465d4f98a2SYan Zheng if (!node) { 16475d4f98a2SYan Zheng while (prev) { 16485d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 16494a0cc7caSNikolay Borisov if (objectid <= btrfs_ino(entry)) { 16505d4f98a2SYan Zheng node = prev; 16515d4f98a2SYan Zheng break; 16525d4f98a2SYan Zheng } 16535d4f98a2SYan Zheng prev = rb_next(prev); 16545d4f98a2SYan Zheng } 16555d4f98a2SYan Zheng } 16565d4f98a2SYan Zheng while (node) { 16575d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 16585d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 16595d4f98a2SYan Zheng if (inode) { 16605d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 16615d4f98a2SYan Zheng return inode; 16625d4f98a2SYan Zheng } 16635d4f98a2SYan Zheng 16644a0cc7caSNikolay Borisov objectid = btrfs_ino(entry) + 1; 16655d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 16665d4f98a2SYan Zheng goto again; 16675d4f98a2SYan Zheng 16685d4f98a2SYan Zheng node = rb_next(node); 16695d4f98a2SYan Zheng } 16705d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 16715d4f98a2SYan Zheng return NULL; 16725d4f98a2SYan Zheng } 16735d4f98a2SYan Zheng 16745d4f98a2SYan Zheng /* 16755d4f98a2SYan Zheng * get new location of data 16765d4f98a2SYan Zheng */ 16775d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 16785d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 16795d4f98a2SYan Zheng { 16805d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 16815d4f98a2SYan Zheng struct btrfs_path *path; 16825d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 16835d4f98a2SYan Zheng struct extent_buffer *leaf; 16845d4f98a2SYan Zheng int ret; 16855d4f98a2SYan Zheng 16865d4f98a2SYan Zheng path = btrfs_alloc_path(); 16875d4f98a2SYan Zheng if (!path) 16885d4f98a2SYan Zheng return -ENOMEM; 16895d4f98a2SYan Zheng 16905d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 1691f85b7379SDavid Sterba ret = btrfs_lookup_file_extent(NULL, root, path, 1692f85b7379SDavid Sterba btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0); 16935d4f98a2SYan Zheng if (ret < 0) 16945d4f98a2SYan Zheng goto out; 16955d4f98a2SYan Zheng if (ret > 0) { 16965d4f98a2SYan Zheng ret = -ENOENT; 16975d4f98a2SYan Zheng goto out; 16985d4f98a2SYan Zheng } 16995d4f98a2SYan Zheng 17005d4f98a2SYan Zheng leaf = path->nodes[0]; 17015d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 17025d4f98a2SYan Zheng struct btrfs_file_extent_item); 17035d4f98a2SYan Zheng 17045d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 17055d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 17065d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 17075d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 17085d4f98a2SYan Zheng 17095d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 171083d4cfd4SJosef Bacik ret = -EINVAL; 17115d4f98a2SYan Zheng goto out; 17125d4f98a2SYan Zheng } 17135d4f98a2SYan Zheng 17145d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 17155d4f98a2SYan Zheng ret = 0; 17165d4f98a2SYan Zheng out: 17175d4f98a2SYan Zheng btrfs_free_path(path); 17185d4f98a2SYan Zheng return ret; 17195d4f98a2SYan Zheng } 17205d4f98a2SYan Zheng 17215d4f98a2SYan Zheng /* 17225d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 17235d4f98a2SYan Zheng * the new locations. 17245d4f98a2SYan Zheng */ 17253fd0a558SYan, Zheng static noinline_for_stack 17263fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 17275d4f98a2SYan Zheng struct reloc_control *rc, 17285d4f98a2SYan Zheng struct btrfs_root *root, 17293fd0a558SYan, Zheng struct extent_buffer *leaf) 17305d4f98a2SYan Zheng { 17310b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 17325d4f98a2SYan Zheng struct btrfs_key key; 17335d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 17345d4f98a2SYan Zheng struct inode *inode = NULL; 17355d4f98a2SYan Zheng u64 parent; 17365d4f98a2SYan Zheng u64 bytenr; 17373fd0a558SYan, Zheng u64 new_bytenr = 0; 17385d4f98a2SYan Zheng u64 num_bytes; 17395d4f98a2SYan Zheng u64 end; 17405d4f98a2SYan Zheng u32 nritems; 17415d4f98a2SYan Zheng u32 i; 174283d4cfd4SJosef Bacik int ret = 0; 17435d4f98a2SYan Zheng int first = 1; 17445d4f98a2SYan Zheng int dirty = 0; 17455d4f98a2SYan Zheng 17465d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 17475d4f98a2SYan Zheng return 0; 17485d4f98a2SYan Zheng 17495d4f98a2SYan Zheng /* reloc trees always use full backref */ 17505d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 17515d4f98a2SYan Zheng parent = leaf->start; 17525d4f98a2SYan Zheng else 17535d4f98a2SYan Zheng parent = 0; 17545d4f98a2SYan Zheng 17555d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 17565d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 175782fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 175882fa113fSQu Wenruo 17595d4f98a2SYan Zheng cond_resched(); 17605d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 17615d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 17625d4f98a2SYan Zheng continue; 17635d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 17645d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 17655d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 17665d4f98a2SYan Zheng continue; 17675d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 17685d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 17695d4f98a2SYan Zheng if (bytenr == 0) 17705d4f98a2SYan Zheng continue; 17719569cc20SQu Wenruo if (!in_range(bytenr, rc->block_group->start, 17729569cc20SQu Wenruo rc->block_group->length)) 17735d4f98a2SYan Zheng continue; 17745d4f98a2SYan Zheng 17755d4f98a2SYan Zheng /* 17765d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 17775d4f98a2SYan Zheng * to complete and drop the extent cache 17785d4f98a2SYan Zheng */ 17795d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 17805d4f98a2SYan Zheng if (first) { 17815d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 17825d4f98a2SYan Zheng first = 0; 17834a0cc7caSNikolay Borisov } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) { 17843fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 17855d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 17865d4f98a2SYan Zheng } 17874a0cc7caSNikolay Borisov if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) { 17885d4f98a2SYan Zheng end = key.offset + 17895d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 17905d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 17910b246afaSJeff Mahoney fs_info->sectorsize)); 17920b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 17935d4f98a2SYan Zheng end--; 17945d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1795d0082371SJeff Mahoney key.offset, end); 17965d4f98a2SYan Zheng if (!ret) 17975d4f98a2SYan Zheng continue; 17985d4f98a2SYan Zheng 1799dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), 1800dcdbc059SNikolay Borisov key.offset, end, 1); 18015d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1802d0082371SJeff Mahoney key.offset, end); 18035d4f98a2SYan Zheng } 18045d4f98a2SYan Zheng } 18055d4f98a2SYan Zheng 18065d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 18075d4f98a2SYan Zheng bytenr, num_bytes); 180883d4cfd4SJosef Bacik if (ret) { 180983d4cfd4SJosef Bacik /* 181083d4cfd4SJosef Bacik * Don't have to abort since we've not changed anything 181183d4cfd4SJosef Bacik * in the file extent yet. 181283d4cfd4SJosef Bacik */ 181383d4cfd4SJosef Bacik break; 18143fd0a558SYan, Zheng } 18155d4f98a2SYan Zheng 18165d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 18175d4f98a2SYan Zheng dirty = 1; 18185d4f98a2SYan Zheng 18195d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 182082fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 182182fa113fSQu Wenruo num_bytes, parent); 182282fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 182382fa113fSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1824b06c4bf5SFilipe Manana key.objectid, key.offset); 182582fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 182683d4cfd4SJosef Bacik if (ret) { 182766642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 182883d4cfd4SJosef Bacik break; 182983d4cfd4SJosef Bacik } 18305d4f98a2SYan Zheng 1831ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr, 1832ffd4bb2aSQu Wenruo num_bytes, parent); 1833ffd4bb2aSQu Wenruo ref.real_root = root->root_key.objectid; 1834ffd4bb2aSQu Wenruo btrfs_init_data_ref(&ref, btrfs_header_owner(leaf), 1835b06c4bf5SFilipe Manana key.objectid, key.offset); 1836ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 183783d4cfd4SJosef Bacik if (ret) { 183866642832SJeff Mahoney btrfs_abort_transaction(trans, ret); 183983d4cfd4SJosef Bacik break; 184083d4cfd4SJosef Bacik } 18415d4f98a2SYan Zheng } 18425d4f98a2SYan Zheng if (dirty) 18435d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 18443fd0a558SYan, Zheng if (inode) 18453fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 184683d4cfd4SJosef Bacik return ret; 18475d4f98a2SYan Zheng } 18485d4f98a2SYan Zheng 18495d4f98a2SYan Zheng static noinline_for_stack 18505d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 18515d4f98a2SYan Zheng struct btrfs_path *path, int level) 18525d4f98a2SYan Zheng { 18535d4f98a2SYan Zheng struct btrfs_disk_key key1; 18545d4f98a2SYan Zheng struct btrfs_disk_key key2; 18555d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 18565d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 18575d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 18585d4f98a2SYan Zheng } 18595d4f98a2SYan Zheng 18605d4f98a2SYan Zheng /* 18615d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 18625d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 18635d4f98a2SYan Zheng * reloc tree was create can be replaced. 18645d4f98a2SYan Zheng * 18655d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 18665d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 18675d4f98a2SYan Zheng * errors, a negative error number is returned. 18685d4f98a2SYan Zheng */ 18693fd0a558SYan, Zheng static noinline_for_stack 18703d0174f7SQu Wenruo int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc, 18715d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 18725d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 18735d4f98a2SYan Zheng int lowest_level, int max_level) 18745d4f98a2SYan Zheng { 18750b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = dest->fs_info; 18765d4f98a2SYan Zheng struct extent_buffer *eb; 18775d4f98a2SYan Zheng struct extent_buffer *parent; 187882fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 18795d4f98a2SYan Zheng struct btrfs_key key; 18805d4f98a2SYan Zheng u64 old_bytenr; 18815d4f98a2SYan Zheng u64 new_bytenr; 18825d4f98a2SYan Zheng u64 old_ptr_gen; 18835d4f98a2SYan Zheng u64 new_ptr_gen; 18845d4f98a2SYan Zheng u64 last_snapshot; 18855d4f98a2SYan Zheng u32 blocksize; 18863fd0a558SYan, Zheng int cow = 0; 18875d4f98a2SYan Zheng int level; 18885d4f98a2SYan Zheng int ret; 18895d4f98a2SYan Zheng int slot; 18905d4f98a2SYan Zheng 18915d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 18925d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 18935d4f98a2SYan Zheng 18945d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 18953fd0a558SYan, Zheng again: 18965d4f98a2SYan Zheng slot = path->slots[lowest_level]; 18975d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 18985d4f98a2SYan Zheng 18995d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 19008bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19015d4f98a2SYan Zheng level = btrfs_header_level(eb); 19025d4f98a2SYan Zheng 19035d4f98a2SYan Zheng if (level < lowest_level) { 19045d4f98a2SYan Zheng btrfs_tree_unlock(eb); 19055d4f98a2SYan Zheng free_extent_buffer(eb); 19065d4f98a2SYan Zheng return 0; 19075d4f98a2SYan Zheng } 19085d4f98a2SYan Zheng 19093fd0a558SYan, Zheng if (cow) { 19105d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 19115d4f98a2SYan Zheng BUG_ON(ret); 19123fd0a558SYan, Zheng } 19138bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19145d4f98a2SYan Zheng 19155d4f98a2SYan Zheng if (next_key) { 19165d4f98a2SYan Zheng next_key->objectid = (u64)-1; 19175d4f98a2SYan Zheng next_key->type = (u8)-1; 19185d4f98a2SYan Zheng next_key->offset = (u64)-1; 19195d4f98a2SYan Zheng } 19205d4f98a2SYan Zheng 19215d4f98a2SYan Zheng parent = eb; 19225d4f98a2SYan Zheng while (1) { 1923581c1760SQu Wenruo struct btrfs_key first_key; 1924581c1760SQu Wenruo 19255d4f98a2SYan Zheng level = btrfs_header_level(parent); 19265d4f98a2SYan Zheng BUG_ON(level < lowest_level); 19275d4f98a2SYan Zheng 19285d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 1929cbca7d59SFilipe Manana if (ret < 0) 1930cbca7d59SFilipe Manana break; 19315d4f98a2SYan Zheng if (ret && slot > 0) 19325d4f98a2SYan Zheng slot--; 19335d4f98a2SYan Zheng 19345d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 19355d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 19365d4f98a2SYan Zheng 19375d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 19380b246afaSJeff Mahoney blocksize = fs_info->nodesize; 19395d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 194017515f1bSQu Wenruo btrfs_node_key_to_cpu(parent, &first_key, slot); 19415d4f98a2SYan Zheng 19425d4f98a2SYan Zheng if (level <= max_level) { 19435d4f98a2SYan Zheng eb = path->nodes[level]; 19445d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 19455d4f98a2SYan Zheng path->slots[level]); 19465d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 19475d4f98a2SYan Zheng path->slots[level]); 19485d4f98a2SYan Zheng } else { 19495d4f98a2SYan Zheng new_bytenr = 0; 19505d4f98a2SYan Zheng new_ptr_gen = 0; 19515d4f98a2SYan Zheng } 19525d4f98a2SYan Zheng 1953fae7f21cSDulshani Gunawardhana if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) { 19545d4f98a2SYan Zheng ret = level; 19555d4f98a2SYan Zheng break; 19565d4f98a2SYan Zheng } 19575d4f98a2SYan Zheng 19585d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 19595d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 19603fd0a558SYan, Zheng if (level <= lowest_level) { 19615d4f98a2SYan Zheng ret = 0; 19625d4f98a2SYan Zheng break; 19635d4f98a2SYan Zheng } 19645d4f98a2SYan Zheng 1965581c1760SQu Wenruo eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen, 1966581c1760SQu Wenruo level - 1, &first_key); 196764c043deSLiu Bo if (IS_ERR(eb)) { 196864c043deSLiu Bo ret = PTR_ERR(eb); 1969264813acSLiu Bo break; 197064c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 197164c043deSLiu Bo ret = -EIO; 1972416bc658SJosef Bacik free_extent_buffer(eb); 1973379cde74SStefan Behrens break; 1974416bc658SJosef Bacik } 19755d4f98a2SYan Zheng btrfs_tree_lock(eb); 19763fd0a558SYan, Zheng if (cow) { 19775d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 19785d4f98a2SYan Zheng slot, &eb); 19795d4f98a2SYan Zheng BUG_ON(ret); 19805d4f98a2SYan Zheng } 19818bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 19825d4f98a2SYan Zheng 19835d4f98a2SYan Zheng btrfs_tree_unlock(parent); 19845d4f98a2SYan Zheng free_extent_buffer(parent); 19855d4f98a2SYan Zheng 19865d4f98a2SYan Zheng parent = eb; 19875d4f98a2SYan Zheng continue; 19885d4f98a2SYan Zheng } 19895d4f98a2SYan Zheng 19903fd0a558SYan, Zheng if (!cow) { 19913fd0a558SYan, Zheng btrfs_tree_unlock(parent); 19923fd0a558SYan, Zheng free_extent_buffer(parent); 19933fd0a558SYan, Zheng cow = 1; 19943fd0a558SYan, Zheng goto again; 19953fd0a558SYan, Zheng } 19963fd0a558SYan, Zheng 19975d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 19985d4f98a2SYan Zheng path->slots[level]); 1999b3b4aa74SDavid Sterba btrfs_release_path(path); 20005d4f98a2SYan Zheng 20015d4f98a2SYan Zheng path->lowest_level = level; 20025d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 20035d4f98a2SYan Zheng path->lowest_level = 0; 20045d4f98a2SYan Zheng BUG_ON(ret); 20055d4f98a2SYan Zheng 20065d4f98a2SYan Zheng /* 2007824d8dffSQu Wenruo * Info qgroup to trace both subtrees. 2008824d8dffSQu Wenruo * 2009824d8dffSQu Wenruo * We must trace both trees. 2010824d8dffSQu Wenruo * 1) Tree reloc subtree 2011824d8dffSQu Wenruo * If not traced, we will leak data numbers 2012824d8dffSQu Wenruo * 2) Fs subtree 2013824d8dffSQu Wenruo * If not traced, we will double count old data 2014f616f5cdSQu Wenruo * 2015f616f5cdSQu Wenruo * We don't scan the subtree right now, but only record 2016f616f5cdSQu Wenruo * the swapped tree blocks. 2017f616f5cdSQu Wenruo * The real subtree rescan is delayed until we have new 2018f616f5cdSQu Wenruo * CoW on the subtree root node before transaction commit. 2019824d8dffSQu Wenruo */ 2020370a11b8SQu Wenruo ret = btrfs_qgroup_add_swapped_blocks(trans, dest, 2021370a11b8SQu Wenruo rc->block_group, parent, slot, 2022370a11b8SQu Wenruo path->nodes[level], path->slots[level], 2023370a11b8SQu Wenruo last_snapshot); 2024370a11b8SQu Wenruo if (ret < 0) 2025370a11b8SQu Wenruo break; 2026824d8dffSQu Wenruo /* 20275d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 20285d4f98a2SYan Zheng */ 20295d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 20305d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 20315d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 20325d4f98a2SYan Zheng 20335d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 20345d4f98a2SYan Zheng path->slots[level], old_bytenr); 20355d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 20365d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 20375d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 20385d4f98a2SYan Zheng 203982fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr, 204082fa113fSQu Wenruo blocksize, path->nodes[level]->start); 204182fa113fSQu Wenruo ref.skip_qgroup = true; 204282fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 204382fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 20445d4f98a2SYan Zheng BUG_ON(ret); 204582fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr, 204682fa113fSQu Wenruo blocksize, 0); 204782fa113fSQu Wenruo ref.skip_qgroup = true; 204882fa113fSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 204982fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 20505d4f98a2SYan Zheng BUG_ON(ret); 20515d4f98a2SYan Zheng 2052ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr, 2053ffd4bb2aSQu Wenruo blocksize, path->nodes[level]->start); 2054ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid); 2055ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2056ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 20575d4f98a2SYan Zheng BUG_ON(ret); 20585d4f98a2SYan Zheng 2059ffd4bb2aSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr, 2060ffd4bb2aSQu Wenruo blocksize, 0); 2061ffd4bb2aSQu Wenruo btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid); 2062ffd4bb2aSQu Wenruo ref.skip_qgroup = true; 2063ffd4bb2aSQu Wenruo ret = btrfs_free_extent(trans, &ref); 20645d4f98a2SYan Zheng BUG_ON(ret); 20655d4f98a2SYan Zheng 20665d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 20675d4f98a2SYan Zheng 20685d4f98a2SYan Zheng ret = level; 20695d4f98a2SYan Zheng break; 20705d4f98a2SYan Zheng } 20715d4f98a2SYan Zheng btrfs_tree_unlock(parent); 20725d4f98a2SYan Zheng free_extent_buffer(parent); 20735d4f98a2SYan Zheng return ret; 20745d4f98a2SYan Zheng } 20755d4f98a2SYan Zheng 20765d4f98a2SYan Zheng /* 20775d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 20785d4f98a2SYan Zheng */ 20795d4f98a2SYan Zheng static noinline_for_stack 20805d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 20815d4f98a2SYan Zheng int *level) 20825d4f98a2SYan Zheng { 20835d4f98a2SYan Zheng struct extent_buffer *eb; 20845d4f98a2SYan Zheng int i; 20855d4f98a2SYan Zheng u64 last_snapshot; 20865d4f98a2SYan Zheng u32 nritems; 20875d4f98a2SYan Zheng 20885d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 20895d4f98a2SYan Zheng 20905d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 20915d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 20925d4f98a2SYan Zheng path->nodes[i] = NULL; 20935d4f98a2SYan Zheng } 20945d4f98a2SYan Zheng 20955d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 20965d4f98a2SYan Zheng eb = path->nodes[i]; 20975d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 20985d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 20995d4f98a2SYan Zheng path->slots[i]++; 21005d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 21015d4f98a2SYan Zheng last_snapshot) 21025d4f98a2SYan Zheng continue; 21035d4f98a2SYan Zheng 21045d4f98a2SYan Zheng *level = i; 21055d4f98a2SYan Zheng return 0; 21065d4f98a2SYan Zheng } 21075d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 21085d4f98a2SYan Zheng path->nodes[i] = NULL; 21095d4f98a2SYan Zheng } 21105d4f98a2SYan Zheng return 1; 21115d4f98a2SYan Zheng } 21125d4f98a2SYan Zheng 21135d4f98a2SYan Zheng /* 21145d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 21155d4f98a2SYan Zheng */ 21165d4f98a2SYan Zheng static noinline_for_stack 21175d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 21185d4f98a2SYan Zheng int *level) 21195d4f98a2SYan Zheng { 21202ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 21215d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 21225d4f98a2SYan Zheng int i; 21235d4f98a2SYan Zheng u64 bytenr; 21245d4f98a2SYan Zheng u64 ptr_gen = 0; 21255d4f98a2SYan Zheng u64 last_snapshot; 21265d4f98a2SYan Zheng u32 nritems; 21275d4f98a2SYan Zheng 21285d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 21295d4f98a2SYan Zheng 21305d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 2131581c1760SQu Wenruo struct btrfs_key first_key; 2132581c1760SQu Wenruo 21335d4f98a2SYan Zheng eb = path->nodes[i]; 21345d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 21355d4f98a2SYan Zheng while (path->slots[i] < nritems) { 21365d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 21375d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 21385d4f98a2SYan Zheng break; 21395d4f98a2SYan Zheng path->slots[i]++; 21405d4f98a2SYan Zheng } 21415d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 21425d4f98a2SYan Zheng if (i == *level) 21435d4f98a2SYan Zheng break; 21445d4f98a2SYan Zheng *level = i + 1; 21455d4f98a2SYan Zheng return 0; 21465d4f98a2SYan Zheng } 21475d4f98a2SYan Zheng if (i == 1) { 21485d4f98a2SYan Zheng *level = i; 21495d4f98a2SYan Zheng return 0; 21505d4f98a2SYan Zheng } 21515d4f98a2SYan Zheng 21525d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 2153581c1760SQu Wenruo btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]); 2154581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1, 2155581c1760SQu Wenruo &first_key); 215664c043deSLiu Bo if (IS_ERR(eb)) { 215764c043deSLiu Bo return PTR_ERR(eb); 215864c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2159416bc658SJosef Bacik free_extent_buffer(eb); 2160416bc658SJosef Bacik return -EIO; 2161416bc658SJosef Bacik } 21625d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 21635d4f98a2SYan Zheng path->nodes[i - 1] = eb; 21645d4f98a2SYan Zheng path->slots[i - 1] = 0; 21655d4f98a2SYan Zheng } 21665d4f98a2SYan Zheng return 1; 21675d4f98a2SYan Zheng } 21685d4f98a2SYan Zheng 21695d4f98a2SYan Zheng /* 21705d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 21715d4f98a2SYan Zheng * [min_key, max_key) 21725d4f98a2SYan Zheng */ 21735d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 21745d4f98a2SYan Zheng struct btrfs_key *min_key, 21755d4f98a2SYan Zheng struct btrfs_key *max_key) 21765d4f98a2SYan Zheng { 21770b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 21785d4f98a2SYan Zheng struct inode *inode = NULL; 21795d4f98a2SYan Zheng u64 objectid; 21805d4f98a2SYan Zheng u64 start, end; 218133345d01SLi Zefan u64 ino; 21825d4f98a2SYan Zheng 21835d4f98a2SYan Zheng objectid = min_key->objectid; 21845d4f98a2SYan Zheng while (1) { 21855d4f98a2SYan Zheng cond_resched(); 21865d4f98a2SYan Zheng iput(inode); 21875d4f98a2SYan Zheng 21885d4f98a2SYan Zheng if (objectid > max_key->objectid) 21895d4f98a2SYan Zheng break; 21905d4f98a2SYan Zheng 21915d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 21925d4f98a2SYan Zheng if (!inode) 21935d4f98a2SYan Zheng break; 21944a0cc7caSNikolay Borisov ino = btrfs_ino(BTRFS_I(inode)); 21955d4f98a2SYan Zheng 219633345d01SLi Zefan if (ino > max_key->objectid) { 21975d4f98a2SYan Zheng iput(inode); 21985d4f98a2SYan Zheng break; 21995d4f98a2SYan Zheng } 22005d4f98a2SYan Zheng 220133345d01SLi Zefan objectid = ino + 1; 22025d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 22035d4f98a2SYan Zheng continue; 22045d4f98a2SYan Zheng 220533345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 22065d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 22075d4f98a2SYan Zheng continue; 22085d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 22095d4f98a2SYan Zheng start = 0; 22105d4f98a2SYan Zheng else { 22115d4f98a2SYan Zheng start = min_key->offset; 22120b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize)); 22135d4f98a2SYan Zheng } 22145d4f98a2SYan Zheng } else { 22155d4f98a2SYan Zheng start = 0; 22165d4f98a2SYan Zheng } 22175d4f98a2SYan Zheng 221833345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 22195d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 22205d4f98a2SYan Zheng continue; 22215d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 22225d4f98a2SYan Zheng end = (u64)-1; 22235d4f98a2SYan Zheng } else { 22245d4f98a2SYan Zheng if (max_key->offset == 0) 22255d4f98a2SYan Zheng continue; 22265d4f98a2SYan Zheng end = max_key->offset; 22270b246afaSJeff Mahoney WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize)); 22285d4f98a2SYan Zheng end--; 22295d4f98a2SYan Zheng } 22305d4f98a2SYan Zheng } else { 22315d4f98a2SYan Zheng end = (u64)-1; 22325d4f98a2SYan Zheng } 22335d4f98a2SYan Zheng 22345d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 2235d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 2236dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1); 2237d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 22385d4f98a2SYan Zheng } 22395d4f98a2SYan Zheng return 0; 22405d4f98a2SYan Zheng } 22415d4f98a2SYan Zheng 22425d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 22435d4f98a2SYan Zheng struct btrfs_key *key) 22445d4f98a2SYan Zheng 22455d4f98a2SYan Zheng { 22465d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 22475d4f98a2SYan Zheng if (!path->nodes[level]) 22485d4f98a2SYan Zheng break; 22495d4f98a2SYan Zheng if (path->slots[level] + 1 < 22505d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 22515d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 22525d4f98a2SYan Zheng path->slots[level] + 1); 22535d4f98a2SYan Zheng return 0; 22545d4f98a2SYan Zheng } 22555d4f98a2SYan Zheng level++; 22565d4f98a2SYan Zheng } 22575d4f98a2SYan Zheng return 1; 22585d4f98a2SYan Zheng } 22595d4f98a2SYan Zheng 22605d4f98a2SYan Zheng /* 2261d2311e69SQu Wenruo * Insert current subvolume into reloc_control::dirty_subvol_roots 2262d2311e69SQu Wenruo */ 2263d2311e69SQu Wenruo static void insert_dirty_subvol(struct btrfs_trans_handle *trans, 2264d2311e69SQu Wenruo struct reloc_control *rc, 2265d2311e69SQu Wenruo struct btrfs_root *root) 2266d2311e69SQu Wenruo { 2267d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2268d2311e69SQu Wenruo struct btrfs_root_item *reloc_root_item; 2269d2311e69SQu Wenruo 2270d2311e69SQu Wenruo /* @root must be a subvolume tree root with a valid reloc tree */ 2271d2311e69SQu Wenruo ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 2272d2311e69SQu Wenruo ASSERT(reloc_root); 2273d2311e69SQu Wenruo 2274d2311e69SQu Wenruo reloc_root_item = &reloc_root->root_item; 2275d2311e69SQu Wenruo memset(&reloc_root_item->drop_progress, 0, 2276d2311e69SQu Wenruo sizeof(reloc_root_item->drop_progress)); 2277d2311e69SQu Wenruo reloc_root_item->drop_level = 0; 2278d2311e69SQu Wenruo btrfs_set_root_refs(reloc_root_item, 0); 2279d2311e69SQu Wenruo btrfs_update_reloc_root(trans, root); 2280d2311e69SQu Wenruo 2281d2311e69SQu Wenruo if (list_empty(&root->reloc_dirty_list)) { 228200246528SJosef Bacik btrfs_grab_root(root); 2283d2311e69SQu Wenruo list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots); 2284d2311e69SQu Wenruo } 2285d2311e69SQu Wenruo } 2286d2311e69SQu Wenruo 2287d2311e69SQu Wenruo static int clean_dirty_subvols(struct reloc_control *rc) 2288d2311e69SQu Wenruo { 2289d2311e69SQu Wenruo struct btrfs_root *root; 2290d2311e69SQu Wenruo struct btrfs_root *next; 2291d2311e69SQu Wenruo int ret = 0; 229230d40577SQu Wenruo int ret2; 2293d2311e69SQu Wenruo 2294d2311e69SQu Wenruo list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots, 2295d2311e69SQu Wenruo reloc_dirty_list) { 229630d40577SQu Wenruo if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 229730d40577SQu Wenruo /* Merged subvolume, cleanup its reloc root */ 2298d2311e69SQu Wenruo struct btrfs_root *reloc_root = root->reloc_root; 2299d2311e69SQu Wenruo 2300d2311e69SQu Wenruo list_del_init(&root->reloc_dirty_list); 2301d2311e69SQu Wenruo root->reloc_root = NULL; 23026282675eSQu Wenruo /* 23036282675eSQu Wenruo * Need barrier to ensure clear_bit() only happens after 23046282675eSQu Wenruo * root->reloc_root = NULL. Pairs with have_reloc_root. 23056282675eSQu Wenruo */ 23066282675eSQu Wenruo smp_wmb(); 23071fac4a54SQu Wenruo clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state); 2308f28de8d8SJosef Bacik if (reloc_root) { 2309f44deb74SJosef Bacik /* 2310f44deb74SJosef Bacik * btrfs_drop_snapshot drops our ref we hold for 2311f44deb74SJosef Bacik * ->reloc_root. If it fails however we must 2312f44deb74SJosef Bacik * drop the ref ourselves. 2313f44deb74SJosef Bacik */ 2314f28de8d8SJosef Bacik ret2 = btrfs_drop_snapshot(reloc_root, 0, 1); 2315f44deb74SJosef Bacik if (ret2 < 0) { 2316f44deb74SJosef Bacik btrfs_put_root(reloc_root); 2317f44deb74SJosef Bacik if (!ret) 2318f28de8d8SJosef Bacik ret = ret2; 2319f28de8d8SJosef Bacik } 2320f44deb74SJosef Bacik } 232100246528SJosef Bacik btrfs_put_root(root); 232230d40577SQu Wenruo } else { 232330d40577SQu Wenruo /* Orphan reloc tree, just clean it up */ 23240078a9f9SNikolay Borisov ret2 = btrfs_drop_snapshot(root, 0, 1); 2325f44deb74SJosef Bacik if (ret2 < 0) { 2326f44deb74SJosef Bacik btrfs_put_root(root); 2327f44deb74SJosef Bacik if (!ret) 232830d40577SQu Wenruo ret = ret2; 232930d40577SQu Wenruo } 2330d2311e69SQu Wenruo } 2331f44deb74SJosef Bacik } 2332d2311e69SQu Wenruo return ret; 2333d2311e69SQu Wenruo } 2334d2311e69SQu Wenruo 2335d2311e69SQu Wenruo /* 23365d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 23375d4f98a2SYan Zheng * fs tree. 23385d4f98a2SYan Zheng */ 23395d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 23405d4f98a2SYan Zheng struct btrfs_root *root) 23415d4f98a2SYan Zheng { 23420b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 23435d4f98a2SYan Zheng struct btrfs_key key; 23445d4f98a2SYan Zheng struct btrfs_key next_key; 23459e6a0c52SJosef Bacik struct btrfs_trans_handle *trans = NULL; 23465d4f98a2SYan Zheng struct btrfs_root *reloc_root; 23475d4f98a2SYan Zheng struct btrfs_root_item *root_item; 23485d4f98a2SYan Zheng struct btrfs_path *path; 23493fd0a558SYan, Zheng struct extent_buffer *leaf; 23505d4f98a2SYan Zheng int level; 23515d4f98a2SYan Zheng int max_level; 23525d4f98a2SYan Zheng int replaced = 0; 23535d4f98a2SYan Zheng int ret; 23545d4f98a2SYan Zheng int err = 0; 23553fd0a558SYan, Zheng u32 min_reserved; 23565d4f98a2SYan Zheng 23575d4f98a2SYan Zheng path = btrfs_alloc_path(); 23585d4f98a2SYan Zheng if (!path) 23595d4f98a2SYan Zheng return -ENOMEM; 2360e4058b54SDavid Sterba path->reada = READA_FORWARD; 23615d4f98a2SYan Zheng 23625d4f98a2SYan Zheng reloc_root = root->reloc_root; 23635d4f98a2SYan Zheng root_item = &reloc_root->root_item; 23645d4f98a2SYan Zheng 23655d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 23665d4f98a2SYan Zheng level = btrfs_root_level(root_item); 236767439dadSDavid Sterba atomic_inc(&reloc_root->node->refs); 23685d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 23695d4f98a2SYan Zheng path->slots[level] = 0; 23705d4f98a2SYan Zheng } else { 23715d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 23725d4f98a2SYan Zheng 23735d4f98a2SYan Zheng level = root_item->drop_level; 23745d4f98a2SYan Zheng BUG_ON(level == 0); 23755d4f98a2SYan Zheng path->lowest_level = level; 23765d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 237733c66f43SYan Zheng path->lowest_level = 0; 23785d4f98a2SYan Zheng if (ret < 0) { 23795d4f98a2SYan Zheng btrfs_free_path(path); 23805d4f98a2SYan Zheng return ret; 23815d4f98a2SYan Zheng } 23825d4f98a2SYan Zheng 23835d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 23845d4f98a2SYan Zheng path->slots[level]); 23855d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 23865d4f98a2SYan Zheng 23875d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 23885d4f98a2SYan Zheng } 23895d4f98a2SYan Zheng 23900b246afaSJeff Mahoney min_reserved = fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 23915d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 23925d4f98a2SYan Zheng 23935d4f98a2SYan Zheng while (1) { 239408e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 239508e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 23963fd0a558SYan, Zheng if (ret) { 23979e6a0c52SJosef Bacik err = ret; 23989e6a0c52SJosef Bacik goto out; 23993fd0a558SYan, Zheng } 24009e6a0c52SJosef Bacik trans = btrfs_start_transaction(root, 0); 24019e6a0c52SJosef Bacik if (IS_ERR(trans)) { 24029e6a0c52SJosef Bacik err = PTR_ERR(trans); 24039e6a0c52SJosef Bacik trans = NULL; 24049e6a0c52SJosef Bacik goto out; 24059e6a0c52SJosef Bacik } 24062abc726aSJosef Bacik 24072abc726aSJosef Bacik /* 24082abc726aSJosef Bacik * At this point we no longer have a reloc_control, so we can't 24092abc726aSJosef Bacik * depend on btrfs_init_reloc_root to update our last_trans. 24102abc726aSJosef Bacik * 24112abc726aSJosef Bacik * But that's ok, we started the trans handle on our 24122abc726aSJosef Bacik * corresponding fs_root, which means it's been added to the 24132abc726aSJosef Bacik * dirty list. At commit time we'll still call 24142abc726aSJosef Bacik * btrfs_update_reloc_root() and update our root item 24152abc726aSJosef Bacik * appropriately. 24162abc726aSJosef Bacik */ 24172abc726aSJosef Bacik reloc_root->last_trans = trans->transid; 24189e6a0c52SJosef Bacik trans->block_rsv = rc->block_rsv; 24193fd0a558SYan, Zheng 24203fd0a558SYan, Zheng replaced = 0; 24215d4f98a2SYan Zheng max_level = level; 24225d4f98a2SYan Zheng 24235d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 24245d4f98a2SYan Zheng if (ret < 0) { 24255d4f98a2SYan Zheng err = ret; 24265d4f98a2SYan Zheng goto out; 24275d4f98a2SYan Zheng } 24285d4f98a2SYan Zheng if (ret > 0) 24295d4f98a2SYan Zheng break; 24305d4f98a2SYan Zheng 24315d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 24325d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 24335d4f98a2SYan Zheng ret = 0; 24345d4f98a2SYan Zheng } else { 24353d0174f7SQu Wenruo ret = replace_path(trans, rc, root, reloc_root, path, 24363fd0a558SYan, Zheng &next_key, level, max_level); 24375d4f98a2SYan Zheng } 24385d4f98a2SYan Zheng if (ret < 0) { 24395d4f98a2SYan Zheng err = ret; 24405d4f98a2SYan Zheng goto out; 24415d4f98a2SYan Zheng } 24425d4f98a2SYan Zheng 24435d4f98a2SYan Zheng if (ret > 0) { 24445d4f98a2SYan Zheng level = ret; 24455d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 24465d4f98a2SYan Zheng path->slots[level]); 24475d4f98a2SYan Zheng replaced = 1; 24485d4f98a2SYan Zheng } 24495d4f98a2SYan Zheng 24505d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 24515d4f98a2SYan Zheng if (ret > 0) 24525d4f98a2SYan Zheng break; 24535d4f98a2SYan Zheng 24545d4f98a2SYan Zheng BUG_ON(level == 0); 24555d4f98a2SYan Zheng /* 24565d4f98a2SYan Zheng * save the merging progress in the drop_progress. 24575d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 24585d4f98a2SYan Zheng */ 24595d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 24605d4f98a2SYan Zheng path->slots[level]); 24615d4f98a2SYan Zheng root_item->drop_level = level; 24625d4f98a2SYan Zheng 24633a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 24649e6a0c52SJosef Bacik trans = NULL; 24655d4f98a2SYan Zheng 24662ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 24675d4f98a2SYan Zheng 24685d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 24695d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 24705d4f98a2SYan Zheng } 24715d4f98a2SYan Zheng 24725d4f98a2SYan Zheng /* 24735d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 24745d4f98a2SYan Zheng * relocated and the block is tree root. 24755d4f98a2SYan Zheng */ 24765d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 24775d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 24785d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 24795d4f98a2SYan Zheng free_extent_buffer(leaf); 24805d4f98a2SYan Zheng if (ret < 0) 24815d4f98a2SYan Zheng err = ret; 24825d4f98a2SYan Zheng out: 24835d4f98a2SYan Zheng btrfs_free_path(path); 24845d4f98a2SYan Zheng 2485d2311e69SQu Wenruo if (err == 0) 2486d2311e69SQu Wenruo insert_dirty_subvol(trans, rc, root); 24875d4f98a2SYan Zheng 24889e6a0c52SJosef Bacik if (trans) 24893a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 24905d4f98a2SYan Zheng 24912ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 24925d4f98a2SYan Zheng 24935d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 24945d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 24955d4f98a2SYan Zheng 24965d4f98a2SYan Zheng return err; 24975d4f98a2SYan Zheng } 24985d4f98a2SYan Zheng 24993fd0a558SYan, Zheng static noinline_for_stack 25003fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 25015d4f98a2SYan Zheng { 25023fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 25030b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 25043fd0a558SYan, Zheng struct btrfs_root *reloc_root; 25055d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 25063fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 25073fd0a558SYan, Zheng u64 num_bytes = 0; 25083fd0a558SYan, Zheng int ret; 25093fd0a558SYan, Zheng 25100b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 25110b246afaSJeff Mahoney rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 25123fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 25130b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 25147585717fSChris Mason 25153fd0a558SYan, Zheng again: 25163fd0a558SYan, Zheng if (!err) { 25173fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 251808e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 251908e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 25203fd0a558SYan, Zheng if (ret) 25213fd0a558SYan, Zheng err = ret; 25223fd0a558SYan, Zheng } 25233fd0a558SYan, Zheng 25247a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 25253612b495STsutomu Itoh if (IS_ERR(trans)) { 25263612b495STsutomu Itoh if (!err) 25272ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 252863f018beSNikolay Borisov num_bytes, NULL); 25293612b495STsutomu Itoh return PTR_ERR(trans); 25303612b495STsutomu Itoh } 25313fd0a558SYan, Zheng 25323fd0a558SYan, Zheng if (!err) { 25333fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 25343a45bb20SJeff Mahoney btrfs_end_transaction(trans); 25352ff7e61eSJeff Mahoney btrfs_block_rsv_release(fs_info, rc->block_rsv, 253663f018beSNikolay Borisov num_bytes, NULL); 25373fd0a558SYan, Zheng goto again; 25383fd0a558SYan, Zheng } 25393fd0a558SYan, Zheng } 25403fd0a558SYan, Zheng 25413fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 25423fd0a558SYan, Zheng 25433fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 25443fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 25453fd0a558SYan, Zheng struct btrfs_root, root_list); 25463fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 25473fd0a558SYan, Zheng 25480b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 25493fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 25503fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 25513fd0a558SYan, Zheng 25523fd0a558SYan, Zheng /* 25533fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 25543fd0a558SYan, Zheng * knows it should resumes merging 25553fd0a558SYan, Zheng */ 25563fd0a558SYan, Zheng if (!err) 25573fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 25583fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 25593fd0a558SYan, Zheng 25603fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 256100246528SJosef Bacik btrfs_put_root(root); 25623fd0a558SYan, Zheng } 25633fd0a558SYan, Zheng 25643fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 25653fd0a558SYan, Zheng 25663fd0a558SYan, Zheng if (!err) 25673a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 25683fd0a558SYan, Zheng else 25693a45bb20SJeff Mahoney btrfs_end_transaction(trans); 25703fd0a558SYan, Zheng return err; 25713fd0a558SYan, Zheng } 25723fd0a558SYan, Zheng 25733fd0a558SYan, Zheng static noinline_for_stack 2574aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 2575aca1bba6SLiu Bo { 2576aca1bba6SLiu Bo struct btrfs_root *reloc_root; 2577aca1bba6SLiu Bo 2578aca1bba6SLiu Bo while (!list_empty(list)) { 2579aca1bba6SLiu Bo reloc_root = list_entry(list->next, struct btrfs_root, 2580aca1bba6SLiu Bo root_list); 2581bb166d72SNaohiro Aota __del_reloc_root(reloc_root); 2582aca1bba6SLiu Bo } 2583aca1bba6SLiu Bo } 2584aca1bba6SLiu Bo 2585aca1bba6SLiu Bo static noinline_for_stack 258694404e82SDavid Sterba void merge_reloc_roots(struct reloc_control *rc) 25873fd0a558SYan, Zheng { 25880b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 25895d4f98a2SYan Zheng struct btrfs_root *root; 25905d4f98a2SYan Zheng struct btrfs_root *reloc_root; 25913fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 25923fd0a558SYan, Zheng int found = 0; 2593aca1bba6SLiu Bo int ret = 0; 25943fd0a558SYan, Zheng again: 25953fd0a558SYan, Zheng root = rc->extent_root; 25967585717fSChris Mason 25977585717fSChris Mason /* 25987585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 25997585717fSChris Mason * we have to make sure nobody is in the middle of 26007585717fSChris Mason * adding their roots to the list while we are 26017585717fSChris Mason * doing this splice 26027585717fSChris Mason */ 26030b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 26043fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 26050b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 26065d4f98a2SYan Zheng 26073fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 26083fd0a558SYan, Zheng found = 1; 26093fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 26103fd0a558SYan, Zheng struct btrfs_root, root_list); 26115d4f98a2SYan Zheng 26125d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 26130b246afaSJeff Mahoney root = read_fs_root(fs_info, 26145d4f98a2SYan Zheng reloc_root->root_key.offset); 26155d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 26165d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 26175d4f98a2SYan Zheng 26183fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 261900246528SJosef Bacik btrfs_put_root(root); 2620b37b39cdSJosef Bacik if (ret) { 262125e293c2SWang Shilong if (list_empty(&reloc_root->root_list)) 262225e293c2SWang Shilong list_add_tail(&reloc_root->root_list, 262325e293c2SWang Shilong &reloc_roots); 2624aca1bba6SLiu Bo goto out; 2625b37b39cdSJosef Bacik } 26263fd0a558SYan, Zheng } else { 26273fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 262830d40577SQu Wenruo /* Don't forget to queue this reloc root for cleanup */ 262930d40577SQu Wenruo list_add_tail(&reloc_root->reloc_dirty_list, 263030d40577SQu Wenruo &rc->dirty_subvol_roots); 26313fd0a558SYan, Zheng } 26325d4f98a2SYan Zheng } 26335d4f98a2SYan Zheng 26343fd0a558SYan, Zheng if (found) { 26353fd0a558SYan, Zheng found = 0; 26363fd0a558SYan, Zheng goto again; 26375d4f98a2SYan Zheng } 2638aca1bba6SLiu Bo out: 2639aca1bba6SLiu Bo if (ret) { 26400b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, NULL); 2641aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 2642aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 2643467bb1d2SWang Shilong 2644467bb1d2SWang Shilong /* new reloc root may be added */ 26450b246afaSJeff Mahoney mutex_lock(&fs_info->reloc_mutex); 2646467bb1d2SWang Shilong list_splice_init(&rc->reloc_roots, &reloc_roots); 26470b246afaSJeff Mahoney mutex_unlock(&fs_info->reloc_mutex); 2648467bb1d2SWang Shilong if (!list_empty(&reloc_roots)) 2649467bb1d2SWang Shilong free_reloc_roots(&reloc_roots); 2650aca1bba6SLiu Bo } 2651aca1bba6SLiu Bo 26527b7b7431SJosef Bacik /* 26537b7b7431SJosef Bacik * We used to have 26547b7b7431SJosef Bacik * 26557b7b7431SJosef Bacik * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 26567b7b7431SJosef Bacik * 26577b7b7431SJosef Bacik * here, but it's wrong. If we fail to start the transaction in 26587b7b7431SJosef Bacik * prepare_to_merge() we will have only 0 ref reloc roots, none of which 26597b7b7431SJosef Bacik * have actually been removed from the reloc_root_tree rb tree. This is 26607b7b7431SJosef Bacik * fine because we're bailing here, and we hold a reference on the root 26617b7b7431SJosef Bacik * for the list that holds it, so these roots will be cleaned up when we 26627b7b7431SJosef Bacik * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root 26637b7b7431SJosef Bacik * will be cleaned up on unmount. 26647b7b7431SJosef Bacik * 26657b7b7431SJosef Bacik * The remaining nodes will be cleaned up by free_reloc_control. 26667b7b7431SJosef Bacik */ 26675d4f98a2SYan Zheng } 26685d4f98a2SYan Zheng 26695d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 26705d4f98a2SYan Zheng { 26715d4f98a2SYan Zheng struct tree_block *block; 26725d4f98a2SYan Zheng struct rb_node *rb_node; 26735d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 26745d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 26755d4f98a2SYan Zheng rb_erase(rb_node, blocks); 26765d4f98a2SYan Zheng kfree(block); 26775d4f98a2SYan Zheng } 26785d4f98a2SYan Zheng } 26795d4f98a2SYan Zheng 26805d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 26815d4f98a2SYan Zheng struct btrfs_root *reloc_root) 26825d4f98a2SYan Zheng { 26830b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = reloc_root->fs_info; 26845d4f98a2SYan Zheng struct btrfs_root *root; 2685442b1ac5SJosef Bacik int ret; 26865d4f98a2SYan Zheng 26875d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 26885d4f98a2SYan Zheng return 0; 26895d4f98a2SYan Zheng 26900b246afaSJeff Mahoney root = read_fs_root(fs_info, reloc_root->root_key.offset); 26915d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 26925d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 2693442b1ac5SJosef Bacik ret = btrfs_record_root_in_trans(trans, root); 269400246528SJosef Bacik btrfs_put_root(root); 26955d4f98a2SYan Zheng 2696442b1ac5SJosef Bacik return ret; 26975d4f98a2SYan Zheng } 26985d4f98a2SYan Zheng 26993fd0a558SYan, Zheng static noinline_for_stack 27003fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 27013fd0a558SYan, Zheng struct reloc_control *rc, 2702a26195a5SQu Wenruo struct btrfs_backref_node *node, 2703a26195a5SQu Wenruo struct btrfs_backref_edge *edges[]) 27045d4f98a2SYan Zheng { 2705a26195a5SQu Wenruo struct btrfs_backref_node *next; 27065d4f98a2SYan Zheng struct btrfs_root *root; 27073fd0a558SYan, Zheng int index = 0; 27083fd0a558SYan, Zheng 27095d4f98a2SYan Zheng next = node; 27105d4f98a2SYan Zheng while (1) { 27115d4f98a2SYan Zheng cond_resched(); 27125d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 27135d4f98a2SYan Zheng root = next->root; 27143fd0a558SYan, Zheng BUG_ON(!root); 271527cdeb70SMiao Xie BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state)); 27165d4f98a2SYan Zheng 27175d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 27185d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 27195d4f98a2SYan Zheng break; 27205d4f98a2SYan Zheng } 27215d4f98a2SYan Zheng 27225d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 27233fd0a558SYan, Zheng root = root->reloc_root; 27243fd0a558SYan, Zheng 27253fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 27263fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 27273fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 27283fd0a558SYan, Zheng next->new_bytenr = root->node->start; 272900246528SJosef Bacik btrfs_put_root(next->root); 273000246528SJosef Bacik next->root = btrfs_grab_root(root); 27310b530bc5SJosef Bacik ASSERT(next->root); 27323fd0a558SYan, Zheng list_add_tail(&next->list, 27333fd0a558SYan, Zheng &rc->backref_cache.changed); 27349569cc20SQu Wenruo mark_block_processed(rc, next); 27355d4f98a2SYan Zheng break; 27365d4f98a2SYan Zheng } 27375d4f98a2SYan Zheng 27383fd0a558SYan, Zheng WARN_ON(1); 27395d4f98a2SYan Zheng root = NULL; 27405d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 27415d4f98a2SYan Zheng if (!next || next->level <= node->level) 27425d4f98a2SYan Zheng break; 27435d4f98a2SYan Zheng } 27443fd0a558SYan, Zheng if (!root) 27453fd0a558SYan, Zheng return NULL; 27465d4f98a2SYan Zheng 27473fd0a558SYan, Zheng next = node; 27483fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 27493fd0a558SYan, Zheng while (1) { 27503fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 27513fd0a558SYan, Zheng if (--index < 0) 27523fd0a558SYan, Zheng break; 27533fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 27543fd0a558SYan, Zheng } 27555d4f98a2SYan Zheng return root; 27565d4f98a2SYan Zheng } 27575d4f98a2SYan Zheng 27583fd0a558SYan, Zheng /* 27593fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 27603fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 27613fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 27623fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 27633fd0a558SYan, Zheng */ 27645d4f98a2SYan Zheng static noinline_for_stack 2765a26195a5SQu Wenruo struct btrfs_root *select_one_root(struct btrfs_backref_node *node) 27665d4f98a2SYan Zheng { 2767a26195a5SQu Wenruo struct btrfs_backref_node *next; 27683fd0a558SYan, Zheng struct btrfs_root *root; 27693fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 2770a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 27713fd0a558SYan, Zheng int index = 0; 27723fd0a558SYan, Zheng 27733fd0a558SYan, Zheng next = node; 27743fd0a558SYan, Zheng while (1) { 27753fd0a558SYan, Zheng cond_resched(); 27763fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 27773fd0a558SYan, Zheng root = next->root; 27783fd0a558SYan, Zheng BUG_ON(!root); 27793fd0a558SYan, Zheng 278025985edcSLucas De Marchi /* no other choice for non-references counted tree */ 278127cdeb70SMiao Xie if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state)) 27823fd0a558SYan, Zheng return root; 27833fd0a558SYan, Zheng 27843fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 27853fd0a558SYan, Zheng fs_root = root; 27863fd0a558SYan, Zheng 27873fd0a558SYan, Zheng if (next != node) 27883fd0a558SYan, Zheng return NULL; 27893fd0a558SYan, Zheng 27903fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 27913fd0a558SYan, Zheng if (!next || next->level <= node->level) 27923fd0a558SYan, Zheng break; 27933fd0a558SYan, Zheng } 27943fd0a558SYan, Zheng 27953fd0a558SYan, Zheng if (!fs_root) 27963fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 27973fd0a558SYan, Zheng return fs_root; 27985d4f98a2SYan Zheng } 27995d4f98a2SYan Zheng 28005d4f98a2SYan Zheng static noinline_for_stack 28013fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 2802a26195a5SQu Wenruo struct btrfs_backref_node *node, int reserve) 28035d4f98a2SYan Zheng { 28040b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2805a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 2806a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2807a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 28083fd0a558SYan, Zheng u64 num_bytes = 0; 28093fd0a558SYan, Zheng int index = 0; 28105d4f98a2SYan Zheng 28113fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 28123fd0a558SYan, Zheng 28133fd0a558SYan, Zheng while (next) { 28143fd0a558SYan, Zheng cond_resched(); 28155d4f98a2SYan Zheng while (1) { 28163fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 28175d4f98a2SYan Zheng break; 28185d4f98a2SYan Zheng 28190b246afaSJeff Mahoney num_bytes += fs_info->nodesize; 28203fd0a558SYan, Zheng 28213fd0a558SYan, Zheng if (list_empty(&next->upper)) 28223fd0a558SYan, Zheng break; 28233fd0a558SYan, Zheng 28243fd0a558SYan, Zheng edge = list_entry(next->upper.next, 2825a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 28263fd0a558SYan, Zheng edges[index++] = edge; 28273fd0a558SYan, Zheng next = edge->node[UPPER]; 28285d4f98a2SYan Zheng } 28293fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 28303fd0a558SYan, Zheng } 28313fd0a558SYan, Zheng return num_bytes; 28323fd0a558SYan, Zheng } 28333fd0a558SYan, Zheng 28343fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 28353fd0a558SYan, Zheng struct reloc_control *rc, 2836a26195a5SQu Wenruo struct btrfs_backref_node *node) 28373fd0a558SYan, Zheng { 28383fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 2839da17066cSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 28403fd0a558SYan, Zheng u64 num_bytes; 28413fd0a558SYan, Zheng int ret; 28420647bf56SWang Shilong u64 tmp; 28433fd0a558SYan, Zheng 28443fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 28453fd0a558SYan, Zheng 28463fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 28470647bf56SWang Shilong rc->reserved_bytes += num_bytes; 28488ca17f0fSJosef Bacik 28498ca17f0fSJosef Bacik /* 28508ca17f0fSJosef Bacik * We are under a transaction here so we can only do limited flushing. 28518ca17f0fSJosef Bacik * If we get an enospc just kick back -EAGAIN so we know to drop the 28528ca17f0fSJosef Bacik * transaction and try to refill when we can flush all the things. 28538ca17f0fSJosef Bacik */ 28540647bf56SWang Shilong ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes, 28558ca17f0fSJosef Bacik BTRFS_RESERVE_FLUSH_LIMIT); 28563fd0a558SYan, Zheng if (ret) { 2857da17066cSJeff Mahoney tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES; 28580647bf56SWang Shilong while (tmp <= rc->reserved_bytes) 28590647bf56SWang Shilong tmp <<= 1; 28600647bf56SWang Shilong /* 28610647bf56SWang Shilong * only one thread can access block_rsv at this point, 28620647bf56SWang Shilong * so we don't need hold lock to protect block_rsv. 28630647bf56SWang Shilong * we expand more reservation size here to allow enough 286452042d8eSAndrea Gelmini * space for relocation and we will return earlier in 28650647bf56SWang Shilong * enospc case. 28660647bf56SWang Shilong */ 2867da17066cSJeff Mahoney rc->block_rsv->size = tmp + fs_info->nodesize * 28680647bf56SWang Shilong RELOCATION_RESERVED_NODES; 28698ca17f0fSJosef Bacik return -EAGAIN; 28703fd0a558SYan, Zheng } 28713fd0a558SYan, Zheng 28723fd0a558SYan, Zheng return 0; 28733fd0a558SYan, Zheng } 28743fd0a558SYan, Zheng 28755d4f98a2SYan Zheng /* 28765d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 28775d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 28785d4f98a2SYan Zheng * 28795d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 28805d4f98a2SYan Zheng * in that case this function just updates pointers. 28815d4f98a2SYan Zheng */ 28825d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 28833fd0a558SYan, Zheng struct reloc_control *rc, 2884a26195a5SQu Wenruo struct btrfs_backref_node *node, 28855d4f98a2SYan Zheng struct btrfs_key *key, 28865d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 28875d4f98a2SYan Zheng { 28882ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 2889a26195a5SQu Wenruo struct btrfs_backref_node *upper; 2890a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 2891a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 28925d4f98a2SYan Zheng struct btrfs_root *root; 28935d4f98a2SYan Zheng struct extent_buffer *eb; 28945d4f98a2SYan Zheng u32 blocksize; 28955d4f98a2SYan Zheng u64 bytenr; 28965d4f98a2SYan Zheng u64 generation; 28975d4f98a2SYan Zheng int slot; 28985d4f98a2SYan Zheng int ret; 28995d4f98a2SYan Zheng int err = 0; 29005d4f98a2SYan Zheng 29015d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 29025d4f98a2SYan Zheng 29035d4f98a2SYan Zheng path->lowest_level = node->level + 1; 29043fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 29055d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 2906581c1760SQu Wenruo struct btrfs_key first_key; 290782fa113fSQu Wenruo struct btrfs_ref ref = { 0 }; 2908581c1760SQu Wenruo 29095d4f98a2SYan Zheng cond_resched(); 29105d4f98a2SYan Zheng 29115d4f98a2SYan Zheng upper = edge->node[UPPER]; 2912dc4103f9SWang Shilong root = select_reloc_root(trans, rc, upper, edges); 29133fd0a558SYan, Zheng BUG_ON(!root); 29145d4f98a2SYan Zheng 29153fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 29163fd0a558SYan, Zheng if (!lowest) { 29173fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 29183fd0a558SYan, Zheng upper->level, &slot); 2919cbca7d59SFilipe Manana if (ret < 0) { 2920cbca7d59SFilipe Manana err = ret; 2921cbca7d59SFilipe Manana goto next; 2922cbca7d59SFilipe Manana } 29233fd0a558SYan, Zheng BUG_ON(ret); 29243fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 29253fd0a558SYan, Zheng if (node->eb->start == bytenr) 29263fd0a558SYan, Zheng goto next; 29273fd0a558SYan, Zheng } 29285d4f98a2SYan Zheng drop_node_buffer(upper); 29293fd0a558SYan, Zheng } 29305d4f98a2SYan Zheng 29315d4f98a2SYan Zheng if (!upper->eb) { 29325d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 29333561b9dbSLiu Bo if (ret) { 29343561b9dbSLiu Bo if (ret < 0) 29355d4f98a2SYan Zheng err = ret; 29363561b9dbSLiu Bo else 29373561b9dbSLiu Bo err = -ENOENT; 29383561b9dbSLiu Bo 29393561b9dbSLiu Bo btrfs_release_path(path); 29405d4f98a2SYan Zheng break; 29415d4f98a2SYan Zheng } 29425d4f98a2SYan Zheng 29433fd0a558SYan, Zheng if (!upper->eb) { 29443fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 29453fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 29463fd0a558SYan, Zheng } else { 29473fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 29483fd0a558SYan, Zheng } 29493fd0a558SYan, Zheng 29503fd0a558SYan, Zheng upper->locked = 1; 29513fd0a558SYan, Zheng path->locks[upper->level] = 0; 29523fd0a558SYan, Zheng 29535d4f98a2SYan Zheng slot = path->slots[upper->level]; 2954b3b4aa74SDavid Sterba btrfs_release_path(path); 29555d4f98a2SYan Zheng } else { 29565d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 29575d4f98a2SYan Zheng &slot); 2958cbca7d59SFilipe Manana if (ret < 0) { 2959cbca7d59SFilipe Manana err = ret; 2960cbca7d59SFilipe Manana goto next; 2961cbca7d59SFilipe Manana } 29625d4f98a2SYan Zheng BUG_ON(ret); 29635d4f98a2SYan Zheng } 29645d4f98a2SYan Zheng 29655d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 29663fd0a558SYan, Zheng if (lowest) { 29674547f4d8SLiu Bo if (bytenr != node->bytenr) { 29684547f4d8SLiu Bo btrfs_err(root->fs_info, 29694547f4d8SLiu Bo "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu", 29704547f4d8SLiu Bo bytenr, node->bytenr, slot, 29714547f4d8SLiu Bo upper->eb->start); 29724547f4d8SLiu Bo err = -EIO; 29734547f4d8SLiu Bo goto next; 29744547f4d8SLiu Bo } 29755d4f98a2SYan Zheng } else { 29763fd0a558SYan, Zheng if (node->eb->start == bytenr) 29773fd0a558SYan, Zheng goto next; 29785d4f98a2SYan Zheng } 29795d4f98a2SYan Zheng 2980da17066cSJeff Mahoney blocksize = root->fs_info->nodesize; 29815d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 2982581c1760SQu Wenruo btrfs_node_key_to_cpu(upper->eb, &first_key, slot); 2983581c1760SQu Wenruo eb = read_tree_block(fs_info, bytenr, generation, 2984581c1760SQu Wenruo upper->level - 1, &first_key); 298564c043deSLiu Bo if (IS_ERR(eb)) { 298664c043deSLiu Bo err = PTR_ERR(eb); 298764c043deSLiu Bo goto next; 298864c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 2989416bc658SJosef Bacik free_extent_buffer(eb); 299097d9a8a4STsutomu Itoh err = -EIO; 299197d9a8a4STsutomu Itoh goto next; 299297d9a8a4STsutomu Itoh } 29935d4f98a2SYan Zheng btrfs_tree_lock(eb); 29948bead258SDavid Sterba btrfs_set_lock_blocking_write(eb); 29955d4f98a2SYan Zheng 29965d4f98a2SYan Zheng if (!node->eb) { 29975d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 29985d4f98a2SYan Zheng slot, &eb); 29993fd0a558SYan, Zheng btrfs_tree_unlock(eb); 30003fd0a558SYan, Zheng free_extent_buffer(eb); 30015d4f98a2SYan Zheng if (ret < 0) { 30025d4f98a2SYan Zheng err = ret; 30033fd0a558SYan, Zheng goto next; 30045d4f98a2SYan Zheng } 30053fd0a558SYan, Zheng BUG_ON(node->eb != eb); 30065d4f98a2SYan Zheng } else { 30075d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 30085d4f98a2SYan Zheng node->eb->start); 30095d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 30105d4f98a2SYan Zheng trans->transid); 30115d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 30125d4f98a2SYan Zheng 301382fa113fSQu Wenruo btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, 30145d4f98a2SYan Zheng node->eb->start, blocksize, 301582fa113fSQu Wenruo upper->eb->start); 301682fa113fSQu Wenruo ref.real_root = root->root_key.objectid; 301782fa113fSQu Wenruo btrfs_init_tree_ref(&ref, node->level, 301882fa113fSQu Wenruo btrfs_header_owner(upper->eb)); 301982fa113fSQu Wenruo ret = btrfs_inc_extent_ref(trans, &ref); 30205d4f98a2SYan Zheng BUG_ON(ret); 30215d4f98a2SYan Zheng 30225d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 30235d4f98a2SYan Zheng BUG_ON(ret); 30245d4f98a2SYan Zheng } 30253fd0a558SYan, Zheng next: 30263fd0a558SYan, Zheng if (!upper->pending) 30273fd0a558SYan, Zheng drop_node_buffer(upper); 30283fd0a558SYan, Zheng else 30293fd0a558SYan, Zheng unlock_node_buffer(upper); 30303fd0a558SYan, Zheng if (err) 30313fd0a558SYan, Zheng break; 30325d4f98a2SYan Zheng } 30333fd0a558SYan, Zheng 30343fd0a558SYan, Zheng if (!err && node->pending) { 30353fd0a558SYan, Zheng drop_node_buffer(node); 30363fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 30373fd0a558SYan, Zheng node->pending = 0; 30385d4f98a2SYan Zheng } 30393fd0a558SYan, Zheng 30405d4f98a2SYan Zheng path->lowest_level = 0; 30413fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 30425d4f98a2SYan Zheng return err; 30435d4f98a2SYan Zheng } 30445d4f98a2SYan Zheng 30455d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 30463fd0a558SYan, Zheng struct reloc_control *rc, 3047a26195a5SQu Wenruo struct btrfs_backref_node *node, 30485d4f98a2SYan Zheng struct btrfs_path *path) 30495d4f98a2SYan Zheng { 30505d4f98a2SYan Zheng struct btrfs_key key; 30515d4f98a2SYan Zheng 30525d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 30533fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 30545d4f98a2SYan Zheng } 30555d4f98a2SYan Zheng 30565d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 30573fd0a558SYan, Zheng struct reloc_control *rc, 30583fd0a558SYan, Zheng struct btrfs_path *path, int err) 30595d4f98a2SYan Zheng { 30603fd0a558SYan, Zheng LIST_HEAD(list); 3061a26195a5SQu Wenruo struct btrfs_backref_cache *cache = &rc->backref_cache; 3062a26195a5SQu Wenruo struct btrfs_backref_node *node; 30635d4f98a2SYan Zheng int level; 30645d4f98a2SYan Zheng int ret; 30655d4f98a2SYan Zheng 30665d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 30675d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 30685d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 3069a26195a5SQu Wenruo struct btrfs_backref_node, list); 30703fd0a558SYan, Zheng list_move_tail(&node->list, &list); 30713fd0a558SYan, Zheng BUG_ON(!node->pending); 30725d4f98a2SYan Zheng 30733fd0a558SYan, Zheng if (!err) { 30743fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 30755d4f98a2SYan Zheng if (ret < 0) 30765d4f98a2SYan Zheng err = ret; 30775d4f98a2SYan Zheng } 30785d4f98a2SYan Zheng } 30793fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 30803fd0a558SYan, Zheng } 30815d4f98a2SYan Zheng return err; 30825d4f98a2SYan Zheng } 30835d4f98a2SYan Zheng 30845d4f98a2SYan Zheng /* 30855d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 30865d4f98a2SYan Zheng * as processed. 30875d4f98a2SYan Zheng */ 30885d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 3089a26195a5SQu Wenruo struct btrfs_backref_node *node) 30905d4f98a2SYan Zheng { 3091a26195a5SQu Wenruo struct btrfs_backref_node *next = node; 3092a26195a5SQu Wenruo struct btrfs_backref_edge *edge; 3093a26195a5SQu Wenruo struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 30945d4f98a2SYan Zheng int index = 0; 30955d4f98a2SYan Zheng 30965d4f98a2SYan Zheng while (next) { 30975d4f98a2SYan Zheng cond_resched(); 30985d4f98a2SYan Zheng while (1) { 30995d4f98a2SYan Zheng if (next->processed) 31005d4f98a2SYan Zheng break; 31015d4f98a2SYan Zheng 31029569cc20SQu Wenruo mark_block_processed(rc, next); 31035d4f98a2SYan Zheng 31045d4f98a2SYan Zheng if (list_empty(&next->upper)) 31055d4f98a2SYan Zheng break; 31065d4f98a2SYan Zheng 31075d4f98a2SYan Zheng edge = list_entry(next->upper.next, 3108a26195a5SQu Wenruo struct btrfs_backref_edge, list[LOWER]); 31095d4f98a2SYan Zheng edges[index++] = edge; 31105d4f98a2SYan Zheng next = edge->node[UPPER]; 31115d4f98a2SYan Zheng } 31125d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 31135d4f98a2SYan Zheng } 31145d4f98a2SYan Zheng } 31155d4f98a2SYan Zheng 31167476dfdaSDavid Sterba static int tree_block_processed(u64 bytenr, struct reloc_control *rc) 31175d4f98a2SYan Zheng { 3118da17066cSJeff Mahoney u32 blocksize = rc->extent_root->fs_info->nodesize; 31197476dfdaSDavid Sterba 31205d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 31219655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 31225d4f98a2SYan Zheng return 1; 31235d4f98a2SYan Zheng return 0; 31245d4f98a2SYan Zheng } 31255d4f98a2SYan Zheng 31262ff7e61eSJeff Mahoney static int get_tree_block_key(struct btrfs_fs_info *fs_info, 31275d4f98a2SYan Zheng struct tree_block *block) 31285d4f98a2SYan Zheng { 31295d4f98a2SYan Zheng struct extent_buffer *eb; 31305d4f98a2SYan Zheng 3131581c1760SQu Wenruo eb = read_tree_block(fs_info, block->bytenr, block->key.offset, 3132581c1760SQu Wenruo block->level, NULL); 313364c043deSLiu Bo if (IS_ERR(eb)) { 313464c043deSLiu Bo return PTR_ERR(eb); 313564c043deSLiu Bo } else if (!extent_buffer_uptodate(eb)) { 3136416bc658SJosef Bacik free_extent_buffer(eb); 3137416bc658SJosef Bacik return -EIO; 3138416bc658SJosef Bacik } 31395d4f98a2SYan Zheng if (block->level == 0) 31405d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 31415d4f98a2SYan Zheng else 31425d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 31435d4f98a2SYan Zheng free_extent_buffer(eb); 31445d4f98a2SYan Zheng block->key_ready = 1; 31455d4f98a2SYan Zheng return 0; 31465d4f98a2SYan Zheng } 31475d4f98a2SYan Zheng 31485d4f98a2SYan Zheng /* 31495d4f98a2SYan Zheng * helper function to relocate a tree block 31505d4f98a2SYan Zheng */ 31515d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 31525d4f98a2SYan Zheng struct reloc_control *rc, 3153a26195a5SQu Wenruo struct btrfs_backref_node *node, 31545d4f98a2SYan Zheng struct btrfs_key *key, 31555d4f98a2SYan Zheng struct btrfs_path *path) 31565d4f98a2SYan Zheng { 31575d4f98a2SYan Zheng struct btrfs_root *root; 31583fd0a558SYan, Zheng int ret = 0; 31595d4f98a2SYan Zheng 31603fd0a558SYan, Zheng if (!node) 31615d4f98a2SYan Zheng return 0; 31623fd0a558SYan, Zheng 31635f6b2e5cSJosef Bacik /* 31645f6b2e5cSJosef Bacik * If we fail here we want to drop our backref_node because we are going 31655f6b2e5cSJosef Bacik * to start over and regenerate the tree for it. 31665f6b2e5cSJosef Bacik */ 31675f6b2e5cSJosef Bacik ret = reserve_metadata_space(trans, rc, node); 31685f6b2e5cSJosef Bacik if (ret) 31695f6b2e5cSJosef Bacik goto out; 31705f6b2e5cSJosef Bacik 31713fd0a558SYan, Zheng BUG_ON(node->processed); 3172147d256eSZhaolei root = select_one_root(node); 31733fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 31743fd0a558SYan, Zheng update_processed_blocks(rc, node); 31753fd0a558SYan, Zheng goto out; 31765d4f98a2SYan Zheng } 31775d4f98a2SYan Zheng 31783fd0a558SYan, Zheng if (root) { 317927cdeb70SMiao Xie if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) { 31803fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 31813fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 31823fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 31833fd0a558SYan, Zheng root = root->reloc_root; 31843fd0a558SYan, Zheng node->new_bytenr = root->node->start; 318500246528SJosef Bacik btrfs_put_root(node->root); 318600246528SJosef Bacik node->root = btrfs_grab_root(root); 31870b530bc5SJosef Bacik ASSERT(node->root); 31883fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 31893fd0a558SYan, Zheng } else { 31905d4f98a2SYan Zheng path->lowest_level = node->level; 31915d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 3192b3b4aa74SDavid Sterba btrfs_release_path(path); 31933fd0a558SYan, Zheng if (ret > 0) 31945d4f98a2SYan Zheng ret = 0; 31953fd0a558SYan, Zheng } 31963fd0a558SYan, Zheng if (!ret) 31973fd0a558SYan, Zheng update_processed_blocks(rc, node); 31983fd0a558SYan, Zheng } else { 31993fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 32003fd0a558SYan, Zheng } 32015d4f98a2SYan Zheng out: 32020647bf56SWang Shilong if (ret || node->level == 0 || node->cowonly) 32033fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 32045d4f98a2SYan Zheng return ret; 32055d4f98a2SYan Zheng } 32065d4f98a2SYan Zheng 32075d4f98a2SYan Zheng /* 32085d4f98a2SYan Zheng * relocate a list of blocks 32095d4f98a2SYan Zheng */ 32105d4f98a2SYan Zheng static noinline_for_stack 32115d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 32125d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 32135d4f98a2SYan Zheng { 32142ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 3215a26195a5SQu Wenruo struct btrfs_backref_node *node; 32165d4f98a2SYan Zheng struct btrfs_path *path; 32175d4f98a2SYan Zheng struct tree_block *block; 321898ff7b94SQu Wenruo struct tree_block *next; 32195d4f98a2SYan Zheng int ret; 32205d4f98a2SYan Zheng int err = 0; 32215d4f98a2SYan Zheng 32225d4f98a2SYan Zheng path = btrfs_alloc_path(); 3223e1a12670SLiu Bo if (!path) { 3224e1a12670SLiu Bo err = -ENOMEM; 322534c2b290SDavid Sterba goto out_free_blocks; 3226e1a12670SLiu Bo } 32275d4f98a2SYan Zheng 322898ff7b94SQu Wenruo /* Kick in readahead for tree blocks with missing keys */ 322998ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 32305d4f98a2SYan Zheng if (!block->key_ready) 32312ff7e61eSJeff Mahoney readahead_tree_block(fs_info, block->bytenr); 32325d4f98a2SYan Zheng } 32335d4f98a2SYan Zheng 323498ff7b94SQu Wenruo /* Get first keys */ 323598ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 323634c2b290SDavid Sterba if (!block->key_ready) { 32372ff7e61eSJeff Mahoney err = get_tree_block_key(fs_info, block); 323834c2b290SDavid Sterba if (err) 323934c2b290SDavid Sterba goto out_free_path; 324034c2b290SDavid Sterba } 32415d4f98a2SYan Zheng } 32425d4f98a2SYan Zheng 324398ff7b94SQu Wenruo /* Do tree relocation */ 324498ff7b94SQu Wenruo rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) { 32453fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 32465d4f98a2SYan Zheng block->level, block->bytenr); 32475d4f98a2SYan Zheng if (IS_ERR(node)) { 32485d4f98a2SYan Zheng err = PTR_ERR(node); 32495d4f98a2SYan Zheng goto out; 32505d4f98a2SYan Zheng } 32515d4f98a2SYan Zheng 32525d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 32535d4f98a2SYan Zheng path); 32545d4f98a2SYan Zheng if (ret < 0) { 32555d4f98a2SYan Zheng err = ret; 325650dbbb71SJosef Bacik break; 32575d4f98a2SYan Zheng } 32585d4f98a2SYan Zheng } 32595d4f98a2SYan Zheng out: 32603fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 32615d4f98a2SYan Zheng 326234c2b290SDavid Sterba out_free_path: 32635d4f98a2SYan Zheng btrfs_free_path(path); 326434c2b290SDavid Sterba out_free_blocks: 3265e1a12670SLiu Bo free_block_list(blocks); 32665d4f98a2SYan Zheng return err; 32675d4f98a2SYan Zheng } 32685d4f98a2SYan Zheng 32695d4f98a2SYan Zheng static noinline_for_stack 3270efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 3271efa56464SYan, Zheng struct file_extent_cluster *cluster) 3272efa56464SYan, Zheng { 3273efa56464SYan, Zheng u64 alloc_hint = 0; 3274efa56464SYan, Zheng u64 start; 3275efa56464SYan, Zheng u64 end; 3276efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 3277efa56464SYan, Zheng u64 num_bytes; 3278efa56464SYan, Zheng int nr = 0; 3279efa56464SYan, Zheng int ret = 0; 3280dcb40c19SWang Xiaoguang u64 prealloc_start = cluster->start - offset; 3281dcb40c19SWang Xiaoguang u64 prealloc_end = cluster->end - offset; 328218513091SWang Xiaoguang u64 cur_offset; 3283364ecf36SQu Wenruo struct extent_changeset *data_reserved = NULL; 3284efa56464SYan, Zheng 3285efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 32865955102cSAl Viro inode_lock(inode); 3287efa56464SYan, Zheng 3288364ecf36SQu Wenruo ret = btrfs_check_data_free_space(inode, &data_reserved, prealloc_start, 3289dcb40c19SWang Xiaoguang prealloc_end + 1 - prealloc_start); 3290efa56464SYan, Zheng if (ret) 3291efa56464SYan, Zheng goto out; 3292efa56464SYan, Zheng 329318513091SWang Xiaoguang cur_offset = prealloc_start; 3294efa56464SYan, Zheng while (nr < cluster->nr) { 3295efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 3296efa56464SYan, Zheng if (nr + 1 < cluster->nr) 3297efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 3298efa56464SYan, Zheng else 3299efa56464SYan, Zheng end = cluster->end - offset; 3300efa56464SYan, Zheng 3301d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 3302efa56464SYan, Zheng num_bytes = end + 1 - start; 330318513091SWang Xiaoguang if (cur_offset < start) 3304bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3305bc42bda2SQu Wenruo cur_offset, start - cur_offset); 3306efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 3307efa56464SYan, Zheng num_bytes, num_bytes, 3308efa56464SYan, Zheng end + 1, &alloc_hint); 330918513091SWang Xiaoguang cur_offset = end + 1; 3310d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 3311efa56464SYan, Zheng if (ret) 3312efa56464SYan, Zheng break; 3313efa56464SYan, Zheng nr++; 3314efa56464SYan, Zheng } 331518513091SWang Xiaoguang if (cur_offset < prealloc_end) 3316bc42bda2SQu Wenruo btrfs_free_reserved_data_space(inode, data_reserved, 3317bc42bda2SQu Wenruo cur_offset, prealloc_end + 1 - cur_offset); 3318efa56464SYan, Zheng out: 33195955102cSAl Viro inode_unlock(inode); 3320364ecf36SQu Wenruo extent_changeset_free(data_reserved); 3321efa56464SYan, Zheng return ret; 3322efa56464SYan, Zheng } 3323efa56464SYan, Zheng 3324efa56464SYan, Zheng static noinline_for_stack 33250257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 33260257bb82SYan, Zheng u64 block_start) 33275d4f98a2SYan Zheng { 33285d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 33295d4f98a2SYan Zheng struct extent_map *em; 33300257bb82SYan, Zheng int ret = 0; 33315d4f98a2SYan Zheng 3332172ddd60SDavid Sterba em = alloc_extent_map(); 33330257bb82SYan, Zheng if (!em) 33340257bb82SYan, Zheng return -ENOMEM; 33350257bb82SYan, Zheng 33365d4f98a2SYan Zheng em->start = start; 33370257bb82SYan, Zheng em->len = end + 1 - start; 33380257bb82SYan, Zheng em->block_len = em->len; 33390257bb82SYan, Zheng em->block_start = block_start; 33405d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 33415d4f98a2SYan Zheng 3342d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 33435d4f98a2SYan Zheng while (1) { 3344890871beSChris Mason write_lock(&em_tree->lock); 334509a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 3346890871beSChris Mason write_unlock(&em_tree->lock); 33475d4f98a2SYan Zheng if (ret != -EEXIST) { 33485d4f98a2SYan Zheng free_extent_map(em); 33495d4f98a2SYan Zheng break; 33505d4f98a2SYan Zheng } 3351dcdbc059SNikolay Borisov btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0); 33525d4f98a2SYan Zheng } 3353d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 33540257bb82SYan, Zheng return ret; 33550257bb82SYan, Zheng } 33565d4f98a2SYan Zheng 3357726a3421SQu Wenruo /* 3358726a3421SQu Wenruo * Allow error injection to test balance cancellation 3359726a3421SQu Wenruo */ 3360726a3421SQu Wenruo int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info) 3361726a3421SQu Wenruo { 3362726a3421SQu Wenruo return atomic_read(&fs_info->balance_cancel_req); 3363726a3421SQu Wenruo } 3364726a3421SQu Wenruo ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE); 3365726a3421SQu Wenruo 33660257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 33670257bb82SYan, Zheng struct file_extent_cluster *cluster) 33680257bb82SYan, Zheng { 33692ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 33700257bb82SYan, Zheng u64 page_start; 33710257bb82SYan, Zheng u64 page_end; 33720257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 33730257bb82SYan, Zheng unsigned long index; 33740257bb82SYan, Zheng unsigned long last_index; 33750257bb82SYan, Zheng struct page *page; 33760257bb82SYan, Zheng struct file_ra_state *ra; 33773b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 33780257bb82SYan, Zheng int nr = 0; 33790257bb82SYan, Zheng int ret = 0; 33800257bb82SYan, Zheng 33810257bb82SYan, Zheng if (!cluster->nr) 33820257bb82SYan, Zheng return 0; 33830257bb82SYan, Zheng 33840257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 33850257bb82SYan, Zheng if (!ra) 33860257bb82SYan, Zheng return -ENOMEM; 33870257bb82SYan, Zheng 3388efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 33890257bb82SYan, Zheng if (ret) 3390efa56464SYan, Zheng goto out; 33910257bb82SYan, Zheng 33920257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 33930257bb82SYan, Zheng 3394efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 3395efa56464SYan, Zheng cluster->end - offset, cluster->start); 3396efa56464SYan, Zheng if (ret) 3397efa56464SYan, Zheng goto out; 3398efa56464SYan, Zheng 339909cbfeafSKirill A. Shutemov index = (cluster->start - offset) >> PAGE_SHIFT; 340009cbfeafSKirill A. Shutemov last_index = (cluster->end - offset) >> PAGE_SHIFT; 34010257bb82SYan, Zheng while (index <= last_index) { 34029f3db423SNikolay Borisov ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), 34039f3db423SNikolay Borisov PAGE_SIZE); 3404efa56464SYan, Zheng if (ret) 3405efa56464SYan, Zheng goto out; 3406efa56464SYan, Zheng 34070257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 34080257bb82SYan, Zheng if (!page) { 34090257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 34100257bb82SYan, Zheng ra, NULL, index, 34110257bb82SYan, Zheng last_index + 1 - index); 3412a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 34133b16a4e3SJosef Bacik mask); 34140257bb82SYan, Zheng if (!page) { 3415691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 341643b18595SQu Wenruo PAGE_SIZE, true); 341744db1216SFilipe Manana btrfs_delalloc_release_extents(BTRFS_I(inode), 34188702ba93SQu Wenruo PAGE_SIZE); 34190257bb82SYan, Zheng ret = -ENOMEM; 3420efa56464SYan, Zheng goto out; 34210257bb82SYan, Zheng } 34220257bb82SYan, Zheng } 34230257bb82SYan, Zheng 34240257bb82SYan, Zheng if (PageReadahead(page)) { 34250257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 34260257bb82SYan, Zheng ra, NULL, page, index, 34270257bb82SYan, Zheng last_index + 1 - index); 34280257bb82SYan, Zheng } 34290257bb82SYan, Zheng 34300257bb82SYan, Zheng if (!PageUptodate(page)) { 34310257bb82SYan, Zheng btrfs_readpage(NULL, page); 34320257bb82SYan, Zheng lock_page(page); 34330257bb82SYan, Zheng if (!PageUptodate(page)) { 34340257bb82SYan, Zheng unlock_page(page); 343509cbfeafSKirill A. Shutemov put_page(page); 3436691fa059SNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 343743b18595SQu Wenruo PAGE_SIZE, true); 34388b62f87bSJosef Bacik btrfs_delalloc_release_extents(BTRFS_I(inode), 34398702ba93SQu Wenruo PAGE_SIZE); 34400257bb82SYan, Zheng ret = -EIO; 3441efa56464SYan, Zheng goto out; 34420257bb82SYan, Zheng } 34430257bb82SYan, Zheng } 34440257bb82SYan, Zheng 34454eee4fa4SMiao Xie page_start = page_offset(page); 344609cbfeafSKirill A. Shutemov page_end = page_start + PAGE_SIZE - 1; 34470257bb82SYan, Zheng 3448d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 34490257bb82SYan, Zheng 34500257bb82SYan, Zheng set_page_extent_mapped(page); 34510257bb82SYan, Zheng 34520257bb82SYan, Zheng if (nr < cluster->nr && 34530257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 34540257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 34550257bb82SYan, Zheng page_start, page_end, 3456ceeb0ae7SDavid Sterba EXTENT_BOUNDARY); 34570257bb82SYan, Zheng nr++; 34580257bb82SYan, Zheng } 34590257bb82SYan, Zheng 3460765f3cebSNikolay Borisov ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0, 3461330a5827SNikolay Borisov NULL); 3462765f3cebSNikolay Borisov if (ret) { 3463765f3cebSNikolay Borisov unlock_page(page); 3464765f3cebSNikolay Borisov put_page(page); 3465765f3cebSNikolay Borisov btrfs_delalloc_release_metadata(BTRFS_I(inode), 346643b18595SQu Wenruo PAGE_SIZE, true); 3467765f3cebSNikolay Borisov btrfs_delalloc_release_extents(BTRFS_I(inode), 34688702ba93SQu Wenruo PAGE_SIZE); 3469765f3cebSNikolay Borisov 3470765f3cebSNikolay Borisov clear_extent_bits(&BTRFS_I(inode)->io_tree, 3471765f3cebSNikolay Borisov page_start, page_end, 3472765f3cebSNikolay Borisov EXTENT_LOCKED | EXTENT_BOUNDARY); 3473765f3cebSNikolay Borisov goto out; 3474765f3cebSNikolay Borisov 3475765f3cebSNikolay Borisov } 34760257bb82SYan, Zheng set_page_dirty(page); 34770257bb82SYan, Zheng 34780257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 3479d0082371SJeff Mahoney page_start, page_end); 34800257bb82SYan, Zheng unlock_page(page); 348109cbfeafSKirill A. Shutemov put_page(page); 34820257bb82SYan, Zheng 34830257bb82SYan, Zheng index++; 34848702ba93SQu Wenruo btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE); 3485efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 34862ff7e61eSJeff Mahoney btrfs_throttle(fs_info); 34877f913c7cSQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 34887f913c7cSQu Wenruo ret = -ECANCELED; 34897f913c7cSQu Wenruo goto out; 34907f913c7cSQu Wenruo } 34910257bb82SYan, Zheng } 34920257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 3493efa56464SYan, Zheng out: 34940257bb82SYan, Zheng kfree(ra); 34950257bb82SYan, Zheng return ret; 34960257bb82SYan, Zheng } 34970257bb82SYan, Zheng 34980257bb82SYan, Zheng static noinline_for_stack 34990257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 35000257bb82SYan, Zheng struct file_extent_cluster *cluster) 35010257bb82SYan, Zheng { 35020257bb82SYan, Zheng int ret; 35030257bb82SYan, Zheng 35040257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 35050257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 35060257bb82SYan, Zheng if (ret) 35070257bb82SYan, Zheng return ret; 35080257bb82SYan, Zheng cluster->nr = 0; 35090257bb82SYan, Zheng } 35100257bb82SYan, Zheng 35110257bb82SYan, Zheng if (!cluster->nr) 35120257bb82SYan, Zheng cluster->start = extent_key->objectid; 35130257bb82SYan, Zheng else 35140257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 35150257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 35160257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 35170257bb82SYan, Zheng cluster->nr++; 35180257bb82SYan, Zheng 35190257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 35200257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 35210257bb82SYan, Zheng if (ret) 35220257bb82SYan, Zheng return ret; 35230257bb82SYan, Zheng cluster->nr = 0; 35240257bb82SYan, Zheng } 35250257bb82SYan, Zheng return 0; 35265d4f98a2SYan Zheng } 35275d4f98a2SYan Zheng 35285d4f98a2SYan Zheng /* 35295d4f98a2SYan Zheng * helper to add a tree block to the list. 35305d4f98a2SYan Zheng * the major work is getting the generation and level of the block 35315d4f98a2SYan Zheng */ 35325d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 35335d4f98a2SYan Zheng struct btrfs_key *extent_key, 35345d4f98a2SYan Zheng struct btrfs_path *path, 35355d4f98a2SYan Zheng struct rb_root *blocks) 35365d4f98a2SYan Zheng { 35375d4f98a2SYan Zheng struct extent_buffer *eb; 35385d4f98a2SYan Zheng struct btrfs_extent_item *ei; 35395d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 35405d4f98a2SYan Zheng struct tree_block *block; 35415d4f98a2SYan Zheng struct rb_node *rb_node; 35425d4f98a2SYan Zheng u32 item_size; 35435d4f98a2SYan Zheng int level = -1; 35447fdf4b60SWang Shilong u64 generation; 35455d4f98a2SYan Zheng 35465d4f98a2SYan Zheng eb = path->nodes[0]; 35475d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 35485d4f98a2SYan Zheng 35493173a18fSJosef Bacik if (extent_key->type == BTRFS_METADATA_ITEM_KEY || 35503173a18fSJosef Bacik item_size >= sizeof(*ei) + sizeof(*bi)) { 35515d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 35525d4f98a2SYan Zheng struct btrfs_extent_item); 35533173a18fSJosef Bacik if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) { 35545d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 35555d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 35565d4f98a2SYan Zheng } else { 35573173a18fSJosef Bacik level = (int)extent_key->offset; 35583173a18fSJosef Bacik } 35593173a18fSJosef Bacik generation = btrfs_extent_generation(eb, ei); 35606d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 3561ba3c2b19SNikolay Borisov btrfs_print_v0_err(eb->fs_info); 3562ba3c2b19SNikolay Borisov btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL); 3563ba3c2b19SNikolay Borisov return -EINVAL; 35643173a18fSJosef Bacik } else { 35655d4f98a2SYan Zheng BUG(); 35665d4f98a2SYan Zheng } 35675d4f98a2SYan Zheng 3568b3b4aa74SDavid Sterba btrfs_release_path(path); 35695d4f98a2SYan Zheng 35705d4f98a2SYan Zheng BUG_ON(level == -1); 35715d4f98a2SYan Zheng 35725d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 35735d4f98a2SYan Zheng if (!block) 35745d4f98a2SYan Zheng return -ENOMEM; 35755d4f98a2SYan Zheng 35765d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 3577da17066cSJeff Mahoney block->key.objectid = rc->extent_root->fs_info->nodesize; 35785d4f98a2SYan Zheng block->key.offset = generation; 35795d4f98a2SYan Zheng block->level = level; 35805d4f98a2SYan Zheng block->key_ready = 0; 35815d4f98a2SYan Zheng 3582e9a28dc5SQu Wenruo rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node); 358343c04fb1SJeff Mahoney if (rb_node) 358443c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, block->bytenr); 35855d4f98a2SYan Zheng 35865d4f98a2SYan Zheng return 0; 35875d4f98a2SYan Zheng } 35885d4f98a2SYan Zheng 35895d4f98a2SYan Zheng /* 35905d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 35915d4f98a2SYan Zheng */ 35925d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 35935d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 35945d4f98a2SYan Zheng struct rb_root *blocks) 35955d4f98a2SYan Zheng { 35960b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 35975d4f98a2SYan Zheng struct btrfs_path *path; 35985d4f98a2SYan Zheng struct btrfs_key key; 35995d4f98a2SYan Zheng int ret; 36000b246afaSJeff Mahoney bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 36015d4f98a2SYan Zheng 36027476dfdaSDavid Sterba if (tree_block_processed(bytenr, rc)) 36035d4f98a2SYan Zheng return 0; 36045d4f98a2SYan Zheng 3605e9a28dc5SQu Wenruo if (rb_simple_search(blocks, bytenr)) 36065d4f98a2SYan Zheng return 0; 36075d4f98a2SYan Zheng 36085d4f98a2SYan Zheng path = btrfs_alloc_path(); 36095d4f98a2SYan Zheng if (!path) 36105d4f98a2SYan Zheng return -ENOMEM; 3611aee68ee5SJosef Bacik again: 36125d4f98a2SYan Zheng key.objectid = bytenr; 3613aee68ee5SJosef Bacik if (skinny) { 3614aee68ee5SJosef Bacik key.type = BTRFS_METADATA_ITEM_KEY; 3615aee68ee5SJosef Bacik key.offset = (u64)-1; 3616aee68ee5SJosef Bacik } else { 36175d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 36185d4f98a2SYan Zheng key.offset = blocksize; 3619aee68ee5SJosef Bacik } 36205d4f98a2SYan Zheng 36215d4f98a2SYan Zheng path->search_commit_root = 1; 36225d4f98a2SYan Zheng path->skip_locking = 1; 36235d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 36245d4f98a2SYan Zheng if (ret < 0) 36255d4f98a2SYan Zheng goto out; 36265d4f98a2SYan Zheng 3627aee68ee5SJosef Bacik if (ret > 0 && skinny) { 3628aee68ee5SJosef Bacik if (path->slots[0]) { 3629aee68ee5SJosef Bacik path->slots[0]--; 3630aee68ee5SJosef Bacik btrfs_item_key_to_cpu(path->nodes[0], &key, 3631aee68ee5SJosef Bacik path->slots[0]); 36323173a18fSJosef Bacik if (key.objectid == bytenr && 3633aee68ee5SJosef Bacik (key.type == BTRFS_METADATA_ITEM_KEY || 3634aee68ee5SJosef Bacik (key.type == BTRFS_EXTENT_ITEM_KEY && 3635aee68ee5SJosef Bacik key.offset == blocksize))) 36363173a18fSJosef Bacik ret = 0; 36373173a18fSJosef Bacik } 3638aee68ee5SJosef Bacik 3639aee68ee5SJosef Bacik if (ret) { 3640aee68ee5SJosef Bacik skinny = false; 3641aee68ee5SJosef Bacik btrfs_release_path(path); 3642aee68ee5SJosef Bacik goto again; 3643aee68ee5SJosef Bacik } 3644aee68ee5SJosef Bacik } 3645cdccee99SLiu Bo if (ret) { 3646cdccee99SLiu Bo ASSERT(ret == 1); 3647cdccee99SLiu Bo btrfs_print_leaf(path->nodes[0]); 3648cdccee99SLiu Bo btrfs_err(fs_info, 3649cdccee99SLiu Bo "tree block extent item (%llu) is not found in extent tree", 3650cdccee99SLiu Bo bytenr); 3651cdccee99SLiu Bo WARN_ON(1); 3652cdccee99SLiu Bo ret = -EINVAL; 3653cdccee99SLiu Bo goto out; 3654cdccee99SLiu Bo } 36553173a18fSJosef Bacik 36565d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 36575d4f98a2SYan Zheng out: 36585d4f98a2SYan Zheng btrfs_free_path(path); 36595d4f98a2SYan Zheng return ret; 36605d4f98a2SYan Zheng } 36615d4f98a2SYan Zheng 36620af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 366332da5386SDavid Sterba struct btrfs_block_group *block_group, 36641bbc621eSChris Mason struct inode *inode, 36651bbc621eSChris Mason u64 ino) 36660af3d00bSJosef Bacik { 36670af3d00bSJosef Bacik struct btrfs_key key; 36680af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 36690af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 36700af3d00bSJosef Bacik int ret = 0; 36710af3d00bSJosef Bacik 36720af3d00bSJosef Bacik if (inode) 36730af3d00bSJosef Bacik goto truncate; 36740af3d00bSJosef Bacik 36750af3d00bSJosef Bacik key.objectid = ino; 36760af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 36770af3d00bSJosef Bacik key.offset = 0; 36780af3d00bSJosef Bacik 36794c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 36802e19f1f9SAl Viro if (IS_ERR(inode)) 36810af3d00bSJosef Bacik return -ENOENT; 36820af3d00bSJosef Bacik 36830af3d00bSJosef Bacik truncate: 36842ff7e61eSJeff Mahoney ret = btrfs_check_trunc_cache_free_space(fs_info, 36857b61cd92SMiao Xie &fs_info->global_block_rsv); 36867b61cd92SMiao Xie if (ret) 36877b61cd92SMiao Xie goto out; 36887b61cd92SMiao Xie 36897a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 36900af3d00bSJosef Bacik if (IS_ERR(trans)) { 36913612b495STsutomu Itoh ret = PTR_ERR(trans); 36920af3d00bSJosef Bacik goto out; 36930af3d00bSJosef Bacik } 36940af3d00bSJosef Bacik 369577ab86bfSJeff Mahoney ret = btrfs_truncate_free_space_cache(trans, block_group, inode); 36960af3d00bSJosef Bacik 36973a45bb20SJeff Mahoney btrfs_end_transaction(trans); 36982ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 36990af3d00bSJosef Bacik out: 37000af3d00bSJosef Bacik iput(inode); 37010af3d00bSJosef Bacik return ret; 37020af3d00bSJosef Bacik } 37030af3d00bSJosef Bacik 37045d4f98a2SYan Zheng /* 370519b546d7SQu Wenruo * Locate the free space cache EXTENT_DATA in root tree leaf and delete the 370619b546d7SQu Wenruo * cache inode, to avoid free space cache data extent blocking data relocation. 37075d4f98a2SYan Zheng */ 370819b546d7SQu Wenruo static int delete_v1_space_cache(struct extent_buffer *leaf, 370919b546d7SQu Wenruo struct btrfs_block_group *block_group, 371019b546d7SQu Wenruo u64 data_bytenr) 37115d4f98a2SYan Zheng { 371219b546d7SQu Wenruo u64 space_cache_ino; 371319b546d7SQu Wenruo struct btrfs_file_extent_item *ei; 37145d4f98a2SYan Zheng struct btrfs_key key; 371519b546d7SQu Wenruo bool found = false; 371619b546d7SQu Wenruo int i; 37175d4f98a2SYan Zheng int ret; 37185d4f98a2SYan Zheng 371919b546d7SQu Wenruo if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID) 372019b546d7SQu Wenruo return 0; 37215d4f98a2SYan Zheng 372219b546d7SQu Wenruo for (i = 0; i < btrfs_header_nritems(leaf); i++) { 372319b546d7SQu Wenruo btrfs_item_key_to_cpu(leaf, &key, i); 372419b546d7SQu Wenruo if (key.type != BTRFS_EXTENT_DATA_KEY) 372519b546d7SQu Wenruo continue; 372619b546d7SQu Wenruo ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 372719b546d7SQu Wenruo if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG && 372819b546d7SQu Wenruo btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) { 372919b546d7SQu Wenruo found = true; 373019b546d7SQu Wenruo space_cache_ino = key.objectid; 373119b546d7SQu Wenruo break; 373219b546d7SQu Wenruo } 373319b546d7SQu Wenruo } 373419b546d7SQu Wenruo if (!found) 373519b546d7SQu Wenruo return -ENOENT; 373619b546d7SQu Wenruo ret = delete_block_group_cache(leaf->fs_info, block_group, NULL, 373719b546d7SQu Wenruo space_cache_ino); 37380af3d00bSJosef Bacik return ret; 37395d4f98a2SYan Zheng } 37405d4f98a2SYan Zheng 37415d4f98a2SYan Zheng /* 37422c016dc2SLiu Bo * helper to find all tree blocks that reference a given data extent 37435d4f98a2SYan Zheng */ 37445d4f98a2SYan Zheng static noinline_for_stack 37455d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 37465d4f98a2SYan Zheng struct btrfs_key *extent_key, 37475d4f98a2SYan Zheng struct btrfs_path *path, 37485d4f98a2SYan Zheng struct rb_root *blocks) 37495d4f98a2SYan Zheng { 375019b546d7SQu Wenruo struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 375119b546d7SQu Wenruo struct ulist *leaves = NULL; 375219b546d7SQu Wenruo struct ulist_iterator leaf_uiter; 375319b546d7SQu Wenruo struct ulist_node *ref_node = NULL; 375419b546d7SQu Wenruo const u32 blocksize = fs_info->nodesize; 3755647f63bdSFilipe David Borba Manana int ret = 0; 37565d4f98a2SYan Zheng 3757b3b4aa74SDavid Sterba btrfs_release_path(path); 375819b546d7SQu Wenruo ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid, 375919b546d7SQu Wenruo 0, &leaves, NULL, true); 376019b546d7SQu Wenruo if (ret < 0) 376119b546d7SQu Wenruo return ret; 376219b546d7SQu Wenruo 376319b546d7SQu Wenruo ULIST_ITER_INIT(&leaf_uiter); 376419b546d7SQu Wenruo while ((ref_node = ulist_next(leaves, &leaf_uiter))) { 376519b546d7SQu Wenruo struct extent_buffer *eb; 376619b546d7SQu Wenruo 376719b546d7SQu Wenruo eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL); 376819b546d7SQu Wenruo if (IS_ERR(eb)) { 376919b546d7SQu Wenruo ret = PTR_ERR(eb); 377019b546d7SQu Wenruo break; 377119b546d7SQu Wenruo } 377219b546d7SQu Wenruo ret = delete_v1_space_cache(eb, rc->block_group, 377319b546d7SQu Wenruo extent_key->objectid); 377419b546d7SQu Wenruo free_extent_buffer(eb); 377519b546d7SQu Wenruo if (ret < 0) 377619b546d7SQu Wenruo break; 377719b546d7SQu Wenruo ret = __add_tree_block(rc, ref_node->val, blocksize, blocks); 377819b546d7SQu Wenruo if (ret < 0) 377919b546d7SQu Wenruo break; 378019b546d7SQu Wenruo } 378119b546d7SQu Wenruo if (ret < 0) 37825d4f98a2SYan Zheng free_block_list(blocks); 378319b546d7SQu Wenruo ulist_free(leaves); 378419b546d7SQu Wenruo return ret; 37855d4f98a2SYan Zheng } 37865d4f98a2SYan Zheng 37875d4f98a2SYan Zheng /* 37882c016dc2SLiu Bo * helper to find next unprocessed extent 37895d4f98a2SYan Zheng */ 37905d4f98a2SYan Zheng static noinline_for_stack 3791147d256eSZhaolei int find_next_extent(struct reloc_control *rc, struct btrfs_path *path, 37923fd0a558SYan, Zheng struct btrfs_key *extent_key) 37935d4f98a2SYan Zheng { 37940b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 37955d4f98a2SYan Zheng struct btrfs_key key; 37965d4f98a2SYan Zheng struct extent_buffer *leaf; 37975d4f98a2SYan Zheng u64 start, end, last; 37985d4f98a2SYan Zheng int ret; 37995d4f98a2SYan Zheng 3800b3470b5dSDavid Sterba last = rc->block_group->start + rc->block_group->length; 38015d4f98a2SYan Zheng while (1) { 38025d4f98a2SYan Zheng cond_resched(); 38035d4f98a2SYan Zheng if (rc->search_start >= last) { 38045d4f98a2SYan Zheng ret = 1; 38055d4f98a2SYan Zheng break; 38065d4f98a2SYan Zheng } 38075d4f98a2SYan Zheng 38085d4f98a2SYan Zheng key.objectid = rc->search_start; 38095d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 38105d4f98a2SYan Zheng key.offset = 0; 38115d4f98a2SYan Zheng 38125d4f98a2SYan Zheng path->search_commit_root = 1; 38135d4f98a2SYan Zheng path->skip_locking = 1; 38145d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 38155d4f98a2SYan Zheng 0, 0); 38165d4f98a2SYan Zheng if (ret < 0) 38175d4f98a2SYan Zheng break; 38185d4f98a2SYan Zheng next: 38195d4f98a2SYan Zheng leaf = path->nodes[0]; 38205d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 38215d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 38225d4f98a2SYan Zheng if (ret != 0) 38235d4f98a2SYan Zheng break; 38245d4f98a2SYan Zheng leaf = path->nodes[0]; 38255d4f98a2SYan Zheng } 38265d4f98a2SYan Zheng 38275d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 38285d4f98a2SYan Zheng if (key.objectid >= last) { 38295d4f98a2SYan Zheng ret = 1; 38305d4f98a2SYan Zheng break; 38315d4f98a2SYan Zheng } 38325d4f98a2SYan Zheng 38333173a18fSJosef Bacik if (key.type != BTRFS_EXTENT_ITEM_KEY && 38343173a18fSJosef Bacik key.type != BTRFS_METADATA_ITEM_KEY) { 38353173a18fSJosef Bacik path->slots[0]++; 38363173a18fSJosef Bacik goto next; 38373173a18fSJosef Bacik } 38383173a18fSJosef Bacik 38393173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY && 38405d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 38415d4f98a2SYan Zheng path->slots[0]++; 38425d4f98a2SYan Zheng goto next; 38435d4f98a2SYan Zheng } 38445d4f98a2SYan Zheng 38453173a18fSJosef Bacik if (key.type == BTRFS_METADATA_ITEM_KEY && 38460b246afaSJeff Mahoney key.objectid + fs_info->nodesize <= 38473173a18fSJosef Bacik rc->search_start) { 38483173a18fSJosef Bacik path->slots[0]++; 38493173a18fSJosef Bacik goto next; 38503173a18fSJosef Bacik } 38513173a18fSJosef Bacik 38525d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 38535d4f98a2SYan Zheng key.objectid, &start, &end, 3854e6138876SJosef Bacik EXTENT_DIRTY, NULL); 38555d4f98a2SYan Zheng 38565d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3857b3b4aa74SDavid Sterba btrfs_release_path(path); 38585d4f98a2SYan Zheng rc->search_start = end + 1; 38595d4f98a2SYan Zheng } else { 38603173a18fSJosef Bacik if (key.type == BTRFS_EXTENT_ITEM_KEY) 38615d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 38623173a18fSJosef Bacik else 38633173a18fSJosef Bacik rc->search_start = key.objectid + 38640b246afaSJeff Mahoney fs_info->nodesize; 38653fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 38665d4f98a2SYan Zheng return 0; 38675d4f98a2SYan Zheng } 38685d4f98a2SYan Zheng } 3869b3b4aa74SDavid Sterba btrfs_release_path(path); 38705d4f98a2SYan Zheng return ret; 38715d4f98a2SYan Zheng } 38725d4f98a2SYan Zheng 38735d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 38745d4f98a2SYan Zheng { 38755d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 38767585717fSChris Mason 38777585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 38785d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 38797585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 38805d4f98a2SYan Zheng } 38815d4f98a2SYan Zheng 38825d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 38835d4f98a2SYan Zheng { 38845d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 38857585717fSChris Mason 38867585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 38875d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 38887585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 38895d4f98a2SYan Zheng } 38905d4f98a2SYan Zheng 38915d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 38925d4f98a2SYan Zheng { 38935d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 38945d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 38955d4f98a2SYan Zheng return 1; 38965d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 38975d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 38985d4f98a2SYan Zheng return 1; 38995d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 39005d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 39015d4f98a2SYan Zheng return 1; 39025d4f98a2SYan Zheng return 0; 39035d4f98a2SYan Zheng } 39045d4f98a2SYan Zheng 39053fd0a558SYan, Zheng static noinline_for_stack 39063fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 39073fd0a558SYan, Zheng { 39083fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 3909ac2fabacSJosef Bacik int ret; 39103fd0a558SYan, Zheng 39112ff7e61eSJeff Mahoney rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info, 391266d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 39133fd0a558SYan, Zheng if (!rc->block_rsv) 39143fd0a558SYan, Zheng return -ENOMEM; 39153fd0a558SYan, Zheng 39163fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 3917b3470b5dSDavid Sterba rc->search_start = rc->block_group->start; 39183fd0a558SYan, Zheng rc->extents_found = 0; 39193fd0a558SYan, Zheng rc->nodes_relocated = 0; 39203fd0a558SYan, Zheng rc->merging_rsv_size = 0; 39210647bf56SWang Shilong rc->reserved_bytes = 0; 3922da17066cSJeff Mahoney rc->block_rsv->size = rc->extent_root->fs_info->nodesize * 39230647bf56SWang Shilong RELOCATION_RESERVED_NODES; 3924ac2fabacSJosef Bacik ret = btrfs_block_rsv_refill(rc->extent_root, 3925ac2fabacSJosef Bacik rc->block_rsv, rc->block_rsv->size, 3926ac2fabacSJosef Bacik BTRFS_RESERVE_FLUSH_ALL); 3927ac2fabacSJosef Bacik if (ret) 3928ac2fabacSJosef Bacik return ret; 39293fd0a558SYan, Zheng 39303fd0a558SYan, Zheng rc->create_reloc_tree = 1; 39313fd0a558SYan, Zheng set_reloc_control(rc); 39323fd0a558SYan, Zheng 39337a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 393428818947SLiu Bo if (IS_ERR(trans)) { 393528818947SLiu Bo unset_reloc_control(rc); 393628818947SLiu Bo /* 393728818947SLiu Bo * extent tree is not a ref_cow tree and has no reloc_root to 393828818947SLiu Bo * cleanup. And callers are responsible to free the above 393928818947SLiu Bo * block rsv. 394028818947SLiu Bo */ 394128818947SLiu Bo return PTR_ERR(trans); 394228818947SLiu Bo } 39433a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 39443fd0a558SYan, Zheng return 0; 39453fd0a558SYan, Zheng } 394676dda93cSYan, Zheng 39475d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 39485d4f98a2SYan Zheng { 39492ff7e61eSJeff Mahoney struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 39505d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 39515d4f98a2SYan Zheng struct btrfs_key key; 39525d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 39535d4f98a2SYan Zheng struct btrfs_path *path; 39545d4f98a2SYan Zheng struct btrfs_extent_item *ei; 39555d4f98a2SYan Zheng u64 flags; 39565d4f98a2SYan Zheng u32 item_size; 39575d4f98a2SYan Zheng int ret; 39585d4f98a2SYan Zheng int err = 0; 3959c87f08caSChris Mason int progress = 0; 39605d4f98a2SYan Zheng 39615d4f98a2SYan Zheng path = btrfs_alloc_path(); 39623fd0a558SYan, Zheng if (!path) 39635d4f98a2SYan Zheng return -ENOMEM; 3964e4058b54SDavid Sterba path->reada = READA_FORWARD; 39653fd0a558SYan, Zheng 39663fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 39673fd0a558SYan, Zheng if (ret) { 39683fd0a558SYan, Zheng err = ret; 39693fd0a558SYan, Zheng goto out_free; 39702423fdfbSJiri Slaby } 39715d4f98a2SYan Zheng 39725d4f98a2SYan Zheng while (1) { 39730647bf56SWang Shilong rc->reserved_bytes = 0; 39740647bf56SWang Shilong ret = btrfs_block_rsv_refill(rc->extent_root, 39750647bf56SWang Shilong rc->block_rsv, rc->block_rsv->size, 39760647bf56SWang Shilong BTRFS_RESERVE_FLUSH_ALL); 39770647bf56SWang Shilong if (ret) { 39780647bf56SWang Shilong err = ret; 39790647bf56SWang Shilong break; 39800647bf56SWang Shilong } 3981c87f08caSChris Mason progress++; 3982a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 39830f788c58SLiu Bo if (IS_ERR(trans)) { 39840f788c58SLiu Bo err = PTR_ERR(trans); 39850f788c58SLiu Bo trans = NULL; 39860f788c58SLiu Bo break; 39870f788c58SLiu Bo } 3988c87f08caSChris Mason restart: 39893fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 39903a45bb20SJeff Mahoney btrfs_end_transaction(trans); 399142a657f5SPan Bian trans = NULL; 39923fd0a558SYan, Zheng continue; 39933fd0a558SYan, Zheng } 39943fd0a558SYan, Zheng 3995147d256eSZhaolei ret = find_next_extent(rc, path, &key); 39965d4f98a2SYan Zheng if (ret < 0) 39975d4f98a2SYan Zheng err = ret; 39985d4f98a2SYan Zheng if (ret != 0) 39995d4f98a2SYan Zheng break; 40005d4f98a2SYan Zheng 40015d4f98a2SYan Zheng rc->extents_found++; 40025d4f98a2SYan Zheng 40035d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 40045d4f98a2SYan Zheng struct btrfs_extent_item); 40053fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 40065d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 40075d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 40085d4f98a2SYan Zheng ret = check_extent_flags(flags); 40095d4f98a2SYan Zheng BUG_ON(ret); 40106d8ff4e4SDavid Sterba } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) { 4011ba3c2b19SNikolay Borisov err = -EINVAL; 4012ba3c2b19SNikolay Borisov btrfs_print_v0_err(trans->fs_info); 4013ba3c2b19SNikolay Borisov btrfs_abort_transaction(trans, err); 4014ba3c2b19SNikolay Borisov break; 40155d4f98a2SYan Zheng } else { 40165d4f98a2SYan Zheng BUG(); 40175d4f98a2SYan Zheng } 40185d4f98a2SYan Zheng 40195d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 40205d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 40215d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 40225d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 40235d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 40245d4f98a2SYan Zheng } else { 4025b3b4aa74SDavid Sterba btrfs_release_path(path); 40265d4f98a2SYan Zheng ret = 0; 40275d4f98a2SYan Zheng } 40285d4f98a2SYan Zheng if (ret < 0) { 40293fd0a558SYan, Zheng err = ret; 40305d4f98a2SYan Zheng break; 40315d4f98a2SYan Zheng } 40325d4f98a2SYan Zheng 40335d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 40345d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 40355d4f98a2SYan Zheng if (ret < 0) { 40363fd0a558SYan, Zheng if (ret != -EAGAIN) { 40375d4f98a2SYan Zheng err = ret; 40385d4f98a2SYan Zheng break; 40395d4f98a2SYan Zheng } 40403fd0a558SYan, Zheng rc->extents_found--; 40413fd0a558SYan, Zheng rc->search_start = key.objectid; 40423fd0a558SYan, Zheng } 40435d4f98a2SYan Zheng } 40445d4f98a2SYan Zheng 40453a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 40462ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 40473fd0a558SYan, Zheng trans = NULL; 40485d4f98a2SYan Zheng 40495d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 40505d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 40515d4f98a2SYan Zheng rc->found_file_extent = 1; 40520257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 40533fd0a558SYan, Zheng &key, &rc->cluster); 40545d4f98a2SYan Zheng if (ret < 0) { 40555d4f98a2SYan Zheng err = ret; 40565d4f98a2SYan Zheng break; 40575d4f98a2SYan Zheng } 40585d4f98a2SYan Zheng } 4059f31ea088SQu Wenruo if (btrfs_should_cancel_balance(fs_info)) { 4060f31ea088SQu Wenruo err = -ECANCELED; 4061f31ea088SQu Wenruo break; 4062f31ea088SQu Wenruo } 40635d4f98a2SYan Zheng } 4064c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 406543a7e99dSNikolay Borisov ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags); 40669689457bSShilong Wang if (ret == 1) { 4067c87f08caSChris Mason err = 0; 4068c87f08caSChris Mason progress = 0; 4069c87f08caSChris Mason goto restart; 4070c87f08caSChris Mason } 4071c87f08caSChris Mason } 40723fd0a558SYan, Zheng 4073b3b4aa74SDavid Sterba btrfs_release_path(path); 407491166212SDavid Sterba clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY); 40755d4f98a2SYan Zheng 40765d4f98a2SYan Zheng if (trans) { 40773a45bb20SJeff Mahoney btrfs_end_transaction_throttle(trans); 40782ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 40795d4f98a2SYan Zheng } 40805d4f98a2SYan Zheng 40810257bb82SYan, Zheng if (!err) { 40823fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 40833fd0a558SYan, Zheng &rc->cluster); 40840257bb82SYan, Zheng if (ret < 0) 40850257bb82SYan, Zheng err = ret; 40860257bb82SYan, Zheng } 40870257bb82SYan, Zheng 40883fd0a558SYan, Zheng rc->create_reloc_tree = 0; 40893fd0a558SYan, Zheng set_reloc_control(rc); 40900257bb82SYan, Zheng 40913fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 409263f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 40935d4f98a2SYan Zheng 40947f913c7cSQu Wenruo /* 40957f913c7cSQu Wenruo * Even in the case when the relocation is cancelled, we should all go 40967f913c7cSQu Wenruo * through prepare_to_merge() and merge_reloc_roots(). 40977f913c7cSQu Wenruo * 40987f913c7cSQu Wenruo * For error (including cancelled balance), prepare_to_merge() will 40997f913c7cSQu Wenruo * mark all reloc trees orphan, then queue them for cleanup in 41007f913c7cSQu Wenruo * merge_reloc_roots() 41017f913c7cSQu Wenruo */ 41023fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 41035d4f98a2SYan Zheng 41045d4f98a2SYan Zheng merge_reloc_roots(rc); 41055d4f98a2SYan Zheng 41063fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 41075d4f98a2SYan Zheng unset_reloc_control(rc); 410863f018beSNikolay Borisov btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL); 41095d4f98a2SYan Zheng 41105d4f98a2SYan Zheng /* get rid of pinned extents */ 41117a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 411262b99540SQu Wenruo if (IS_ERR(trans)) { 41133612b495STsutomu Itoh err = PTR_ERR(trans); 411462b99540SQu Wenruo goto out_free; 411562b99540SQu Wenruo } 41163a45bb20SJeff Mahoney btrfs_commit_transaction(trans); 41176217b0faSJosef Bacik out_free: 4118d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4119d2311e69SQu Wenruo if (ret < 0 && !err) 4120d2311e69SQu Wenruo err = ret; 41212ff7e61eSJeff Mahoney btrfs_free_block_rsv(fs_info, rc->block_rsv); 41223fd0a558SYan, Zheng btrfs_free_path(path); 41235d4f98a2SYan Zheng return err; 41245d4f98a2SYan Zheng } 41255d4f98a2SYan Zheng 41265d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 41270257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 41285d4f98a2SYan Zheng { 41295d4f98a2SYan Zheng struct btrfs_path *path; 41305d4f98a2SYan Zheng struct btrfs_inode_item *item; 41315d4f98a2SYan Zheng struct extent_buffer *leaf; 41325d4f98a2SYan Zheng int ret; 41335d4f98a2SYan Zheng 41345d4f98a2SYan Zheng path = btrfs_alloc_path(); 41355d4f98a2SYan Zheng if (!path) 41365d4f98a2SYan Zheng return -ENOMEM; 41375d4f98a2SYan Zheng 41385d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 41395d4f98a2SYan Zheng if (ret) 41405d4f98a2SYan Zheng goto out; 41415d4f98a2SYan Zheng 41425d4f98a2SYan Zheng leaf = path->nodes[0]; 41435d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 4144b159fa28SDavid Sterba memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item)); 41455d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 41460257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 41475d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 41483fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 41493fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 41505d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 41515d4f98a2SYan Zheng out: 41525d4f98a2SYan Zheng btrfs_free_path(path); 41535d4f98a2SYan Zheng return ret; 41545d4f98a2SYan Zheng } 41555d4f98a2SYan Zheng 41565d4f98a2SYan Zheng /* 41575d4f98a2SYan Zheng * helper to create inode for data relocation. 41585d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 41595d4f98a2SYan Zheng */ 41603fd0a558SYan, Zheng static noinline_for_stack 41613fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 416232da5386SDavid Sterba struct btrfs_block_group *group) 41635d4f98a2SYan Zheng { 41645d4f98a2SYan Zheng struct inode *inode = NULL; 41655d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 41665d4f98a2SYan Zheng struct btrfs_root *root; 41675d4f98a2SYan Zheng struct btrfs_key key; 41684624900dSZhaolei u64 objectid; 41695d4f98a2SYan Zheng int err = 0; 41705d4f98a2SYan Zheng 41715d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 41725d4f98a2SYan Zheng if (IS_ERR(root)) 41735d4f98a2SYan Zheng return ERR_CAST(root); 41745d4f98a2SYan Zheng 4175a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 417676deacf0SJosef Bacik if (IS_ERR(trans)) { 417700246528SJosef Bacik btrfs_put_root(root); 41783fd0a558SYan, Zheng return ERR_CAST(trans); 417976deacf0SJosef Bacik } 41805d4f98a2SYan Zheng 4181581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 41825d4f98a2SYan Zheng if (err) 41835d4f98a2SYan Zheng goto out; 41845d4f98a2SYan Zheng 41850257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 41865d4f98a2SYan Zheng BUG_ON(err); 41875d4f98a2SYan Zheng 41885d4f98a2SYan Zheng key.objectid = objectid; 41895d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 41905d4f98a2SYan Zheng key.offset = 0; 41914c66e0d4SDavid Sterba inode = btrfs_iget(fs_info->sb, &key, root); 41922e19f1f9SAl Viro BUG_ON(IS_ERR(inode)); 4193b3470b5dSDavid Sterba BTRFS_I(inode)->index_cnt = group->start; 41945d4f98a2SYan Zheng 419573f2e545SNikolay Borisov err = btrfs_orphan_add(trans, BTRFS_I(inode)); 41965d4f98a2SYan Zheng out: 419700246528SJosef Bacik btrfs_put_root(root); 41983a45bb20SJeff Mahoney btrfs_end_transaction(trans); 41992ff7e61eSJeff Mahoney btrfs_btree_balance_dirty(fs_info); 42005d4f98a2SYan Zheng if (err) { 42015d4f98a2SYan Zheng if (inode) 42025d4f98a2SYan Zheng iput(inode); 42035d4f98a2SYan Zheng inode = ERR_PTR(err); 42045d4f98a2SYan Zheng } 42055d4f98a2SYan Zheng return inode; 42065d4f98a2SYan Zheng } 42075d4f98a2SYan Zheng 4208c258d6e3SQu Wenruo static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info) 42093fd0a558SYan, Zheng { 42103fd0a558SYan, Zheng struct reloc_control *rc; 42113fd0a558SYan, Zheng 42123fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 42133fd0a558SYan, Zheng if (!rc) 42143fd0a558SYan, Zheng return NULL; 42153fd0a558SYan, Zheng 42163fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 4217d2311e69SQu Wenruo INIT_LIST_HEAD(&rc->dirty_subvol_roots); 4218584fb121SQu Wenruo btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1); 42193fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 422043eb5f29SQu Wenruo extent_io_tree_init(fs_info, &rc->processed_blocks, 422143eb5f29SQu Wenruo IO_TREE_RELOC_BLOCKS, NULL); 42223fd0a558SYan, Zheng return rc; 42233fd0a558SYan, Zheng } 42243fd0a558SYan, Zheng 42251a0afa0eSJosef Bacik static void free_reloc_control(struct reloc_control *rc) 42261a0afa0eSJosef Bacik { 42271a0afa0eSJosef Bacik struct mapping_node *node, *tmp; 42281a0afa0eSJosef Bacik 42291a0afa0eSJosef Bacik free_reloc_roots(&rc->reloc_roots); 42301a0afa0eSJosef Bacik rbtree_postorder_for_each_entry_safe(node, tmp, 42311a0afa0eSJosef Bacik &rc->reloc_root_tree.rb_root, rb_node) 42321a0afa0eSJosef Bacik kfree(node); 42331a0afa0eSJosef Bacik 42341a0afa0eSJosef Bacik kfree(rc); 42351a0afa0eSJosef Bacik } 42361a0afa0eSJosef Bacik 42375d4f98a2SYan Zheng /* 4238ebce0e01SAdam Borowski * Print the block group being relocated 4239ebce0e01SAdam Borowski */ 4240ebce0e01SAdam Borowski static void describe_relocation(struct btrfs_fs_info *fs_info, 424132da5386SDavid Sterba struct btrfs_block_group *block_group) 4242ebce0e01SAdam Borowski { 4243f89e09cfSAnand Jain char buf[128] = {'\0'}; 4244ebce0e01SAdam Borowski 4245f89e09cfSAnand Jain btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf)); 4246ebce0e01SAdam Borowski 4247ebce0e01SAdam Borowski btrfs_info(fs_info, 4248ebce0e01SAdam Borowski "relocating block group %llu flags %s", 4249b3470b5dSDavid Sterba block_group->start, buf); 4250ebce0e01SAdam Borowski } 4251ebce0e01SAdam Borowski 4252430640e3SQu Wenruo static const char *stage_to_string(int stage) 4253430640e3SQu Wenruo { 4254430640e3SQu Wenruo if (stage == MOVE_DATA_EXTENTS) 4255430640e3SQu Wenruo return "move data extents"; 4256430640e3SQu Wenruo if (stage == UPDATE_DATA_PTRS) 4257430640e3SQu Wenruo return "update data pointers"; 4258430640e3SQu Wenruo return "unknown"; 4259430640e3SQu Wenruo } 4260430640e3SQu Wenruo 4261ebce0e01SAdam Borowski /* 42625d4f98a2SYan Zheng * function to relocate all extents in a block group. 42635d4f98a2SYan Zheng */ 42646bccf3abSJeff Mahoney int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start) 42655d4f98a2SYan Zheng { 426632da5386SDavid Sterba struct btrfs_block_group *bg; 42676bccf3abSJeff Mahoney struct btrfs_root *extent_root = fs_info->extent_root; 42685d4f98a2SYan Zheng struct reloc_control *rc; 42690af3d00bSJosef Bacik struct inode *inode; 42700af3d00bSJosef Bacik struct btrfs_path *path; 42715d4f98a2SYan Zheng int ret; 4272f0486c68SYan, Zheng int rw = 0; 42735d4f98a2SYan Zheng int err = 0; 42745d4f98a2SYan Zheng 4275eede2bf3SOmar Sandoval bg = btrfs_lookup_block_group(fs_info, group_start); 4276eede2bf3SOmar Sandoval if (!bg) 4277eede2bf3SOmar Sandoval return -ENOENT; 4278eede2bf3SOmar Sandoval 4279eede2bf3SOmar Sandoval if (btrfs_pinned_by_swapfile(fs_info, bg)) { 4280eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 4281eede2bf3SOmar Sandoval return -ETXTBSY; 4282eede2bf3SOmar Sandoval } 4283eede2bf3SOmar Sandoval 4284c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 4285eede2bf3SOmar Sandoval if (!rc) { 4286eede2bf3SOmar Sandoval btrfs_put_block_group(bg); 42875d4f98a2SYan Zheng return -ENOMEM; 4288eede2bf3SOmar Sandoval } 42895d4f98a2SYan Zheng 4290f0486c68SYan, Zheng rc->extent_root = extent_root; 4291eede2bf3SOmar Sandoval rc->block_group = bg; 42925d4f98a2SYan Zheng 4293b12de528SQu Wenruo ret = btrfs_inc_block_group_ro(rc->block_group, true); 4294f0486c68SYan, Zheng if (ret) { 4295f0486c68SYan, Zheng err = ret; 4296f0486c68SYan, Zheng goto out; 4297f0486c68SYan, Zheng } 4298f0486c68SYan, Zheng rw = 1; 4299f0486c68SYan, Zheng 43000af3d00bSJosef Bacik path = btrfs_alloc_path(); 43010af3d00bSJosef Bacik if (!path) { 43020af3d00bSJosef Bacik err = -ENOMEM; 43030af3d00bSJosef Bacik goto out; 43040af3d00bSJosef Bacik } 43050af3d00bSJosef Bacik 43067949f339SDavid Sterba inode = lookup_free_space_inode(rc->block_group, path); 43070af3d00bSJosef Bacik btrfs_free_path(path); 43080af3d00bSJosef Bacik 43090af3d00bSJosef Bacik if (!IS_ERR(inode)) 43101bbc621eSChris Mason ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0); 43110af3d00bSJosef Bacik else 43120af3d00bSJosef Bacik ret = PTR_ERR(inode); 43130af3d00bSJosef Bacik 43140af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 43150af3d00bSJosef Bacik err = ret; 43160af3d00bSJosef Bacik goto out; 43170af3d00bSJosef Bacik } 43180af3d00bSJosef Bacik 43195d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 43205d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 43215d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 43225d4f98a2SYan Zheng rc->data_inode = NULL; 43235d4f98a2SYan Zheng goto out; 43245d4f98a2SYan Zheng } 43255d4f98a2SYan Zheng 43260b246afaSJeff Mahoney describe_relocation(fs_info, rc->block_group); 43275d4f98a2SYan Zheng 43289cfa3e34SFilipe Manana btrfs_wait_block_group_reservations(rc->block_group); 4329f78c436cSFilipe Manana btrfs_wait_nocow_writers(rc->block_group); 43306374e57aSChris Mason btrfs_wait_ordered_roots(fs_info, U64_MAX, 4331b3470b5dSDavid Sterba rc->block_group->start, 4332b3470b5dSDavid Sterba rc->block_group->length); 43335d4f98a2SYan Zheng 43345d4f98a2SYan Zheng while (1) { 4335430640e3SQu Wenruo int finishes_stage; 4336430640e3SQu Wenruo 433776dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 43385d4f98a2SYan Zheng ret = relocate_block_group(rc); 433976dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 4340ff612ba7SJosef Bacik if (ret < 0) 43415d4f98a2SYan Zheng err = ret; 4342ff612ba7SJosef Bacik 4343430640e3SQu Wenruo finishes_stage = rc->stage; 4344ff612ba7SJosef Bacik /* 4345ff612ba7SJosef Bacik * We may have gotten ENOSPC after we already dirtied some 4346ff612ba7SJosef Bacik * extents. If writeout happens while we're relocating a 4347ff612ba7SJosef Bacik * different block group we could end up hitting the 4348ff612ba7SJosef Bacik * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in 4349ff612ba7SJosef Bacik * btrfs_reloc_cow_block. Make sure we write everything out 4350ff612ba7SJosef Bacik * properly so we don't trip over this problem, and then break 4351ff612ba7SJosef Bacik * out of the loop if we hit an error. 4352ff612ba7SJosef Bacik */ 4353ff612ba7SJosef Bacik if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 4354ff612ba7SJosef Bacik ret = btrfs_wait_ordered_range(rc->data_inode, 0, 4355ff612ba7SJosef Bacik (u64)-1); 4356ff612ba7SJosef Bacik if (ret) 4357ff612ba7SJosef Bacik err = ret; 4358ff612ba7SJosef Bacik invalidate_mapping_pages(rc->data_inode->i_mapping, 4359ff612ba7SJosef Bacik 0, -1); 4360ff612ba7SJosef Bacik rc->stage = UPDATE_DATA_PTRS; 43615d4f98a2SYan Zheng } 43625d4f98a2SYan Zheng 4363ff612ba7SJosef Bacik if (err < 0) 4364ff612ba7SJosef Bacik goto out; 4365ff612ba7SJosef Bacik 43665d4f98a2SYan Zheng if (rc->extents_found == 0) 43675d4f98a2SYan Zheng break; 43685d4f98a2SYan Zheng 4369430640e3SQu Wenruo btrfs_info(fs_info, "found %llu extents, stage: %s", 4370430640e3SQu Wenruo rc->extents_found, stage_to_string(finishes_stage)); 43715d4f98a2SYan Zheng } 43725d4f98a2SYan Zheng 43735d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 43745d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 4375bf38be65SDavid Sterba WARN_ON(rc->block_group->used > 0); 43765d4f98a2SYan Zheng out: 4377f0486c68SYan, Zheng if (err && rw) 43782ff7e61eSJeff Mahoney btrfs_dec_block_group_ro(rc->block_group); 43795d4f98a2SYan Zheng iput(rc->data_inode); 43805d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 43811a0afa0eSJosef Bacik free_reloc_control(rc); 43825d4f98a2SYan Zheng return err; 43835d4f98a2SYan Zheng } 43845d4f98a2SYan Zheng 438576dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 438676dda93cSYan, Zheng { 43870b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 438876dda93cSYan, Zheng struct btrfs_trans_handle *trans; 438979787eaaSJeff Mahoney int ret, err; 439076dda93cSYan, Zheng 43910b246afaSJeff Mahoney trans = btrfs_start_transaction(fs_info->tree_root, 0); 439279787eaaSJeff Mahoney if (IS_ERR(trans)) 439379787eaaSJeff Mahoney return PTR_ERR(trans); 439476dda93cSYan, Zheng 439576dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 439676dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 439776dda93cSYan, Zheng root->root_item.drop_level = 0; 439876dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 43990b246afaSJeff Mahoney ret = btrfs_update_root(trans, fs_info->tree_root, 440076dda93cSYan, Zheng &root->root_key, &root->root_item); 440176dda93cSYan, Zheng 44023a45bb20SJeff Mahoney err = btrfs_end_transaction(trans); 440379787eaaSJeff Mahoney if (err) 440479787eaaSJeff Mahoney return err; 440579787eaaSJeff Mahoney return ret; 440676dda93cSYan, Zheng } 440776dda93cSYan, Zheng 44085d4f98a2SYan Zheng /* 44095d4f98a2SYan Zheng * recover relocation interrupted by system crash. 44105d4f98a2SYan Zheng * 44115d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 44125d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 44135d4f98a2SYan Zheng */ 44145d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 44155d4f98a2SYan Zheng { 44160b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 44175d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 44185d4f98a2SYan Zheng struct btrfs_key key; 44195d4f98a2SYan Zheng struct btrfs_root *fs_root; 44205d4f98a2SYan Zheng struct btrfs_root *reloc_root; 44215d4f98a2SYan Zheng struct btrfs_path *path; 44225d4f98a2SYan Zheng struct extent_buffer *leaf; 44235d4f98a2SYan Zheng struct reloc_control *rc = NULL; 44245d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 44255d4f98a2SYan Zheng int ret; 44265d4f98a2SYan Zheng int err = 0; 44275d4f98a2SYan Zheng 44285d4f98a2SYan Zheng path = btrfs_alloc_path(); 44295d4f98a2SYan Zheng if (!path) 44305d4f98a2SYan Zheng return -ENOMEM; 4431e4058b54SDavid Sterba path->reada = READA_BACK; 44325d4f98a2SYan Zheng 44335d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 44345d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 44355d4f98a2SYan Zheng key.offset = (u64)-1; 44365d4f98a2SYan Zheng 44375d4f98a2SYan Zheng while (1) { 44380b246afaSJeff Mahoney ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, 44395d4f98a2SYan Zheng path, 0, 0); 44405d4f98a2SYan Zheng if (ret < 0) { 44415d4f98a2SYan Zheng err = ret; 44425d4f98a2SYan Zheng goto out; 44435d4f98a2SYan Zheng } 44445d4f98a2SYan Zheng if (ret > 0) { 44455d4f98a2SYan Zheng if (path->slots[0] == 0) 44465d4f98a2SYan Zheng break; 44475d4f98a2SYan Zheng path->slots[0]--; 44485d4f98a2SYan Zheng } 44495d4f98a2SYan Zheng leaf = path->nodes[0]; 44505d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4451b3b4aa74SDavid Sterba btrfs_release_path(path); 44525d4f98a2SYan Zheng 44535d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 44545d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 44555d4f98a2SYan Zheng break; 44565d4f98a2SYan Zheng 44573dbf1738SJosef Bacik reloc_root = btrfs_read_tree_root(root, &key); 44585d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 44595d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 44605d4f98a2SYan Zheng goto out; 44615d4f98a2SYan Zheng } 44625d4f98a2SYan Zheng 44633dbf1738SJosef Bacik set_bit(BTRFS_ROOT_REF_COWS, &reloc_root->state); 44645d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 44655d4f98a2SYan Zheng 44665d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 44670b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, 44685d4f98a2SYan Zheng reloc_root->root_key.offset); 44695d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 447076dda93cSYan, Zheng ret = PTR_ERR(fs_root); 447176dda93cSYan, Zheng if (ret != -ENOENT) { 447276dda93cSYan, Zheng err = ret; 44735d4f98a2SYan Zheng goto out; 44745d4f98a2SYan Zheng } 447579787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 447679787eaaSJeff Mahoney if (ret < 0) { 447779787eaaSJeff Mahoney err = ret; 447879787eaaSJeff Mahoney goto out; 447979787eaaSJeff Mahoney } 4480932fd26dSJosef Bacik } else { 448100246528SJosef Bacik btrfs_put_root(fs_root); 448276dda93cSYan, Zheng } 44835d4f98a2SYan Zheng } 44845d4f98a2SYan Zheng 44855d4f98a2SYan Zheng if (key.offset == 0) 44865d4f98a2SYan Zheng break; 44875d4f98a2SYan Zheng 44885d4f98a2SYan Zheng key.offset--; 44895d4f98a2SYan Zheng } 4490b3b4aa74SDavid Sterba btrfs_release_path(path); 44915d4f98a2SYan Zheng 44925d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 44935d4f98a2SYan Zheng goto out; 44945d4f98a2SYan Zheng 4495c258d6e3SQu Wenruo rc = alloc_reloc_control(fs_info); 44965d4f98a2SYan Zheng if (!rc) { 44975d4f98a2SYan Zheng err = -ENOMEM; 44985d4f98a2SYan Zheng goto out; 44995d4f98a2SYan Zheng } 45005d4f98a2SYan Zheng 45010b246afaSJeff Mahoney rc->extent_root = fs_info->extent_root; 45025d4f98a2SYan Zheng 45035d4f98a2SYan Zheng set_reloc_control(rc); 45045d4f98a2SYan Zheng 45057a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 45063612b495STsutomu Itoh if (IS_ERR(trans)) { 45073612b495STsutomu Itoh err = PTR_ERR(trans); 4508fb2d83eeSJosef Bacik goto out_unset; 45093612b495STsutomu Itoh } 45103fd0a558SYan, Zheng 45113fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 45123fd0a558SYan, Zheng 45135d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 45145d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 45155d4f98a2SYan Zheng struct btrfs_root, root_list); 45165d4f98a2SYan Zheng list_del(&reloc_root->root_list); 45175d4f98a2SYan Zheng 45185d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 45195d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 45205d4f98a2SYan Zheng &rc->reloc_roots); 45215d4f98a2SYan Zheng continue; 45225d4f98a2SYan Zheng } 45235d4f98a2SYan Zheng 45240b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, reloc_root->root_key.offset); 452579787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 452679787eaaSJeff Mahoney err = PTR_ERR(fs_root); 4527ca1aa281SJosef Bacik list_add_tail(&reloc_root->root_list, &reloc_roots); 45281402d17dSXiyu Yang btrfs_end_transaction(trans); 4529fb2d83eeSJosef Bacik goto out_unset; 453079787eaaSJeff Mahoney } 45315d4f98a2SYan Zheng 4532ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 453379787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 4534f44deb74SJosef Bacik fs_root->reloc_root = btrfs_grab_root(reloc_root); 453500246528SJosef Bacik btrfs_put_root(fs_root); 45365d4f98a2SYan Zheng } 45375d4f98a2SYan Zheng 45383a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 453979787eaaSJeff Mahoney if (err) 4540fb2d83eeSJosef Bacik goto out_unset; 45415d4f98a2SYan Zheng 45425d4f98a2SYan Zheng merge_reloc_roots(rc); 45435d4f98a2SYan Zheng 45445d4f98a2SYan Zheng unset_reloc_control(rc); 45455d4f98a2SYan Zheng 45467a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 454762b99540SQu Wenruo if (IS_ERR(trans)) { 45483612b495STsutomu Itoh err = PTR_ERR(trans); 45496217b0faSJosef Bacik goto out_clean; 455062b99540SQu Wenruo } 45513a45bb20SJeff Mahoney err = btrfs_commit_transaction(trans); 45526217b0faSJosef Bacik out_clean: 4553d2311e69SQu Wenruo ret = clean_dirty_subvols(rc); 4554d2311e69SQu Wenruo if (ret < 0 && !err) 4555d2311e69SQu Wenruo err = ret; 4556fb2d83eeSJosef Bacik out_unset: 4557fb2d83eeSJosef Bacik unset_reloc_control(rc); 45581a0afa0eSJosef Bacik free_reloc_control(rc); 45593612b495STsutomu Itoh out: 4560aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 4561aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4562aca1bba6SLiu Bo 45635d4f98a2SYan Zheng btrfs_free_path(path); 45645d4f98a2SYan Zheng 45655d4f98a2SYan Zheng if (err == 0) { 45665d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 45670b246afaSJeff Mahoney fs_root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 4568932fd26dSJosef Bacik if (IS_ERR(fs_root)) { 45695d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4570932fd26dSJosef Bacik } else { 457166b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 457200246528SJosef Bacik btrfs_put_root(fs_root); 4573932fd26dSJosef Bacik } 4574932fd26dSJosef Bacik } 45755d4f98a2SYan Zheng return err; 45765d4f98a2SYan Zheng } 45775d4f98a2SYan Zheng 45785d4f98a2SYan Zheng /* 45795d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 45805d4f98a2SYan Zheng * 45815d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 45825d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 45835d4f98a2SYan Zheng */ 45845d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 45855d4f98a2SYan Zheng { 45860b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 45875d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 45885d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 45895d4f98a2SYan Zheng int ret; 45905d4f98a2SYan Zheng u64 disk_bytenr; 45914577b014SJosef Bacik u64 new_bytenr; 45925d4f98a2SYan Zheng LIST_HEAD(list); 45935d4f98a2SYan Zheng 45945d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 4595bffe633eSOmar Sandoval BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len); 45965d4f98a2SYan Zheng 45975d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 45980b246afaSJeff Mahoney ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr, 4599a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 460079787eaaSJeff Mahoney if (ret) 460179787eaaSJeff Mahoney goto out; 46025d4f98a2SYan Zheng 46035d4f98a2SYan Zheng while (!list_empty(&list)) { 46045d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 46055d4f98a2SYan Zheng list_del_init(&sums->list); 46065d4f98a2SYan Zheng 46074577b014SJosef Bacik /* 46084577b014SJosef Bacik * We need to offset the new_bytenr based on where the csum is. 46094577b014SJosef Bacik * We need to do this because we will read in entire prealloc 46104577b014SJosef Bacik * extents but we may have written to say the middle of the 46114577b014SJosef Bacik * prealloc extent, so we need to make sure the csum goes with 46124577b014SJosef Bacik * the right disk offset. 46134577b014SJosef Bacik * 46144577b014SJosef Bacik * We can do this because the data reloc inode refers strictly 46154577b014SJosef Bacik * to the on disk bytes, so we don't have to worry about 46164577b014SJosef Bacik * disk_len vs real len like with real inodes since it's all 46174577b014SJosef Bacik * disk length. 46184577b014SJosef Bacik */ 4619bffe633eSOmar Sandoval new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr; 46204577b014SJosef Bacik sums->bytenr = new_bytenr; 46215d4f98a2SYan Zheng 4622f9756261SNikolay Borisov btrfs_add_ordered_sum(ordered, sums); 46235d4f98a2SYan Zheng } 462479787eaaSJeff Mahoney out: 46255d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4626411fc6bcSAndi Kleen return ret; 46275d4f98a2SYan Zheng } 46283fd0a558SYan, Zheng 462983d4cfd4SJosef Bacik int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 46303fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 46313fd0a558SYan, Zheng struct extent_buffer *cow) 46323fd0a558SYan, Zheng { 46330b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 46343fd0a558SYan, Zheng struct reloc_control *rc; 4635a26195a5SQu Wenruo struct btrfs_backref_node *node; 46363fd0a558SYan, Zheng int first_cow = 0; 46373fd0a558SYan, Zheng int level; 463883d4cfd4SJosef Bacik int ret = 0; 46393fd0a558SYan, Zheng 46400b246afaSJeff Mahoney rc = fs_info->reloc_ctl; 46413fd0a558SYan, Zheng if (!rc) 464283d4cfd4SJosef Bacik return 0; 46433fd0a558SYan, Zheng 46443fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 46453fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 46463fd0a558SYan, Zheng 46473fd0a558SYan, Zheng level = btrfs_header_level(buf); 46483fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 46493fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 46503fd0a558SYan, Zheng first_cow = 1; 46513fd0a558SYan, Zheng 46523fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 46533fd0a558SYan, Zheng rc->create_reloc_tree) { 46543fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 46553fd0a558SYan, Zheng 46563fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 46573fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 46583fd0a558SYan, Zheng node->new_bytenr != buf->start); 46593fd0a558SYan, Zheng 46603fd0a558SYan, Zheng drop_node_buffer(node); 466167439dadSDavid Sterba atomic_inc(&cow->refs); 46623fd0a558SYan, Zheng node->eb = cow; 46633fd0a558SYan, Zheng node->new_bytenr = cow->start; 46643fd0a558SYan, Zheng 46653fd0a558SYan, Zheng if (!node->pending) { 46663fd0a558SYan, Zheng list_move_tail(&node->list, 46673fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 46683fd0a558SYan, Zheng node->pending = 1; 46693fd0a558SYan, Zheng } 46703fd0a558SYan, Zheng 46713fd0a558SYan, Zheng if (first_cow) 46729569cc20SQu Wenruo mark_block_processed(rc, node); 46733fd0a558SYan, Zheng 46743fd0a558SYan, Zheng if (first_cow && level > 0) 46753fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 46763fd0a558SYan, Zheng } 46773fd0a558SYan, Zheng 467883d4cfd4SJosef Bacik if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) 46793fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 468083d4cfd4SJosef Bacik return ret; 46813fd0a558SYan, Zheng } 46823fd0a558SYan, Zheng 46833fd0a558SYan, Zheng /* 46843fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 468501327610SNicholas D Steeves * required for relocating tree blocks in the snapshot 46863fd0a558SYan, Zheng */ 4687147d256eSZhaolei void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending, 46883fd0a558SYan, Zheng u64 *bytes_to_reserve) 46893fd0a558SYan, Zheng { 469010995c04SQu Wenruo struct btrfs_root *root = pending->root; 469110995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 46923fd0a558SYan, Zheng 46936282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 46943fd0a558SYan, Zheng return; 46953fd0a558SYan, Zheng 46963fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 46973fd0a558SYan, Zheng return; 46983fd0a558SYan, Zheng 46993fd0a558SYan, Zheng root = root->reloc_root; 47003fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 47013fd0a558SYan, Zheng /* 47023fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 47033fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 47043fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 47053fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 47063fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 47073fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 47083fd0a558SYan, Zheng * tree may use more space than it frees. so we need 47093fd0a558SYan, Zheng * reserve extra space. 47103fd0a558SYan, Zheng */ 47113fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 47123fd0a558SYan, Zheng } 47133fd0a558SYan, Zheng 47143fd0a558SYan, Zheng /* 47153fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 47163fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 4717f44deb74SJosef Bacik * 4718f44deb74SJosef Bacik * This is similar to btrfs_init_reloc_root(), we come out of here with two 4719f44deb74SJosef Bacik * references held on the reloc_root, one for root->reloc_root and one for 4720f44deb74SJosef Bacik * rc->reloc_roots. 47213fd0a558SYan, Zheng */ 472249b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 47233fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 47243fd0a558SYan, Zheng { 47253fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 47263fd0a558SYan, Zheng struct btrfs_root *reloc_root; 47273fd0a558SYan, Zheng struct btrfs_root *new_root; 472810995c04SQu Wenruo struct reloc_control *rc = root->fs_info->reloc_ctl; 47293fd0a558SYan, Zheng int ret; 47303fd0a558SYan, Zheng 47316282675eSQu Wenruo if (!rc || !have_reloc_root(root)) 473249b25e05SJeff Mahoney return 0; 47333fd0a558SYan, Zheng 47343fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 47353fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 47363fd0a558SYan, Zheng 47373fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 47383fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 47393fd0a558SYan, Zheng rc->block_rsv, 47403a584174SLu Fengqi rc->nodes_relocated, true); 474149b25e05SJeff Mahoney if (ret) 474249b25e05SJeff Mahoney return ret; 47433fd0a558SYan, Zheng } 47443fd0a558SYan, Zheng 47453fd0a558SYan, Zheng new_root = pending->snap; 47463fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 47473fd0a558SYan, Zheng new_root->root_key.objectid); 474849b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 474949b25e05SJeff Mahoney return PTR_ERR(reloc_root); 47503fd0a558SYan, Zheng 4751ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4752ffd7b339SJeff Mahoney BUG_ON(ret < 0); 4753f44deb74SJosef Bacik new_root->reloc_root = btrfs_grab_root(reloc_root); 47543fd0a558SYan, Zheng 475549b25e05SJeff Mahoney if (rc->create_reloc_tree) 47563fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 475749b25e05SJeff Mahoney return ret; 47583fd0a558SYan, Zheng } 4759