15d4f98a2SYan Zheng /* 25d4f98a2SYan Zheng * Copyright (C) 2009 Oracle. All rights reserved. 35d4f98a2SYan Zheng * 45d4f98a2SYan Zheng * This program is free software; you can redistribute it and/or 55d4f98a2SYan Zheng * modify it under the terms of the GNU General Public 65d4f98a2SYan Zheng * License v2 as published by the Free Software Foundation. 75d4f98a2SYan Zheng * 85d4f98a2SYan Zheng * This program is distributed in the hope that it will be useful, 95d4f98a2SYan Zheng * but WITHOUT ANY WARRANTY; without even the implied warranty of 105d4f98a2SYan Zheng * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 115d4f98a2SYan Zheng * General Public License for more details. 125d4f98a2SYan Zheng * 135d4f98a2SYan Zheng * You should have received a copy of the GNU General Public 145d4f98a2SYan Zheng * License along with this program; if not, write to the 155d4f98a2SYan Zheng * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 165d4f98a2SYan Zheng * Boston, MA 021110-1307, USA. 175d4f98a2SYan Zheng */ 185d4f98a2SYan Zheng 195d4f98a2SYan Zheng #include <linux/sched.h> 205d4f98a2SYan Zheng #include <linux/pagemap.h> 215d4f98a2SYan Zheng #include <linux/writeback.h> 225d4f98a2SYan Zheng #include <linux/blkdev.h> 235d4f98a2SYan Zheng #include <linux/rbtree.h> 245a0e3ad6STejun Heo #include <linux/slab.h> 255d4f98a2SYan Zheng #include "ctree.h" 265d4f98a2SYan Zheng #include "disk-io.h" 275d4f98a2SYan Zheng #include "transaction.h" 285d4f98a2SYan Zheng #include "volumes.h" 295d4f98a2SYan Zheng #include "locking.h" 305d4f98a2SYan Zheng #include "btrfs_inode.h" 315d4f98a2SYan Zheng #include "async-thread.h" 320af3d00bSJosef Bacik #include "free-space-cache.h" 33581bb050SLi Zefan #include "inode-map.h" 345d4f98a2SYan Zheng 355d4f98a2SYan Zheng /* 365d4f98a2SYan Zheng * backref_node, mapping_node and tree_block start with this 375d4f98a2SYan Zheng */ 385d4f98a2SYan Zheng struct tree_entry { 395d4f98a2SYan Zheng struct rb_node rb_node; 405d4f98a2SYan Zheng u64 bytenr; 415d4f98a2SYan Zheng }; 425d4f98a2SYan Zheng 435d4f98a2SYan Zheng /* 445d4f98a2SYan Zheng * present a tree block in the backref cache 455d4f98a2SYan Zheng */ 465d4f98a2SYan Zheng struct backref_node { 475d4f98a2SYan Zheng struct rb_node rb_node; 485d4f98a2SYan Zheng u64 bytenr; 493fd0a558SYan, Zheng 503fd0a558SYan, Zheng u64 new_bytenr; 513fd0a558SYan, Zheng /* objectid of tree block owner, can be not uptodate */ 525d4f98a2SYan Zheng u64 owner; 533fd0a558SYan, Zheng /* link to pending, changed or detached list */ 543fd0a558SYan, Zheng struct list_head list; 555d4f98a2SYan Zheng /* list of upper level blocks reference this block */ 565d4f98a2SYan Zheng struct list_head upper; 575d4f98a2SYan Zheng /* list of child blocks in the cache */ 585d4f98a2SYan Zheng struct list_head lower; 595d4f98a2SYan Zheng /* NULL if this node is not tree root */ 605d4f98a2SYan Zheng struct btrfs_root *root; 615d4f98a2SYan Zheng /* extent buffer got by COW the block */ 625d4f98a2SYan Zheng struct extent_buffer *eb; 635d4f98a2SYan Zheng /* level of tree block */ 645d4f98a2SYan Zheng unsigned int level:8; 653fd0a558SYan, Zheng /* is the block in non-reference counted tree */ 663fd0a558SYan, Zheng unsigned int cowonly:1; 673fd0a558SYan, Zheng /* 1 if no child node in the cache */ 685d4f98a2SYan Zheng unsigned int lowest:1; 695d4f98a2SYan Zheng /* is the extent buffer locked */ 705d4f98a2SYan Zheng unsigned int locked:1; 715d4f98a2SYan Zheng /* has the block been processed */ 725d4f98a2SYan Zheng unsigned int processed:1; 735d4f98a2SYan Zheng /* have backrefs of this block been checked */ 745d4f98a2SYan Zheng unsigned int checked:1; 753fd0a558SYan, Zheng /* 763fd0a558SYan, Zheng * 1 if corresponding block has been cowed but some upper 773fd0a558SYan, Zheng * level block pointers may not point to the new location 783fd0a558SYan, Zheng */ 793fd0a558SYan, Zheng unsigned int pending:1; 803fd0a558SYan, Zheng /* 813fd0a558SYan, Zheng * 1 if the backref node isn't connected to any other 823fd0a558SYan, Zheng * backref node. 833fd0a558SYan, Zheng */ 843fd0a558SYan, Zheng unsigned int detached:1; 855d4f98a2SYan Zheng }; 865d4f98a2SYan Zheng 875d4f98a2SYan Zheng /* 885d4f98a2SYan Zheng * present a block pointer in the backref cache 895d4f98a2SYan Zheng */ 905d4f98a2SYan Zheng struct backref_edge { 915d4f98a2SYan Zheng struct list_head list[2]; 925d4f98a2SYan Zheng struct backref_node *node[2]; 935d4f98a2SYan Zheng }; 945d4f98a2SYan Zheng 955d4f98a2SYan Zheng #define LOWER 0 965d4f98a2SYan Zheng #define UPPER 1 975d4f98a2SYan Zheng 985d4f98a2SYan Zheng struct backref_cache { 995d4f98a2SYan Zheng /* red black tree of all backref nodes in the cache */ 1005d4f98a2SYan Zheng struct rb_root rb_root; 1013fd0a558SYan, Zheng /* for passing backref nodes to btrfs_reloc_cow_block */ 1023fd0a558SYan, Zheng struct backref_node *path[BTRFS_MAX_LEVEL]; 1033fd0a558SYan, Zheng /* 1043fd0a558SYan, Zheng * list of blocks that have been cowed but some block 1053fd0a558SYan, Zheng * pointers in upper level blocks may not reflect the 1063fd0a558SYan, Zheng * new location 1073fd0a558SYan, Zheng */ 1085d4f98a2SYan Zheng struct list_head pending[BTRFS_MAX_LEVEL]; 1093fd0a558SYan, Zheng /* list of backref nodes with no child node */ 1103fd0a558SYan, Zheng struct list_head leaves; 1113fd0a558SYan, Zheng /* list of blocks that have been cowed in current transaction */ 1123fd0a558SYan, Zheng struct list_head changed; 1133fd0a558SYan, Zheng /* list of detached backref node. */ 1143fd0a558SYan, Zheng struct list_head detached; 1153fd0a558SYan, Zheng 1163fd0a558SYan, Zheng u64 last_trans; 1173fd0a558SYan, Zheng 1183fd0a558SYan, Zheng int nr_nodes; 1193fd0a558SYan, Zheng int nr_edges; 1205d4f98a2SYan Zheng }; 1215d4f98a2SYan Zheng 1225d4f98a2SYan Zheng /* 1235d4f98a2SYan Zheng * map address of tree root to tree 1245d4f98a2SYan Zheng */ 1255d4f98a2SYan Zheng struct mapping_node { 1265d4f98a2SYan Zheng struct rb_node rb_node; 1275d4f98a2SYan Zheng u64 bytenr; 1285d4f98a2SYan Zheng void *data; 1295d4f98a2SYan Zheng }; 1305d4f98a2SYan Zheng 1315d4f98a2SYan Zheng struct mapping_tree { 1325d4f98a2SYan Zheng struct rb_root rb_root; 1335d4f98a2SYan Zheng spinlock_t lock; 1345d4f98a2SYan Zheng }; 1355d4f98a2SYan Zheng 1365d4f98a2SYan Zheng /* 1375d4f98a2SYan Zheng * present a tree block to process 1385d4f98a2SYan Zheng */ 1395d4f98a2SYan Zheng struct tree_block { 1405d4f98a2SYan Zheng struct rb_node rb_node; 1415d4f98a2SYan Zheng u64 bytenr; 1425d4f98a2SYan Zheng struct btrfs_key key; 1435d4f98a2SYan Zheng unsigned int level:8; 1445d4f98a2SYan Zheng unsigned int key_ready:1; 1455d4f98a2SYan Zheng }; 1465d4f98a2SYan Zheng 1470257bb82SYan, Zheng #define MAX_EXTENTS 128 1480257bb82SYan, Zheng 1490257bb82SYan, Zheng struct file_extent_cluster { 1500257bb82SYan, Zheng u64 start; 1510257bb82SYan, Zheng u64 end; 1520257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 1530257bb82SYan, Zheng unsigned int nr; 1540257bb82SYan, Zheng }; 1550257bb82SYan, Zheng 1565d4f98a2SYan Zheng struct reloc_control { 1575d4f98a2SYan Zheng /* block group to relocate */ 1585d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group; 1595d4f98a2SYan Zheng /* extent tree */ 1605d4f98a2SYan Zheng struct btrfs_root *extent_root; 1615d4f98a2SYan Zheng /* inode for moving data */ 1625d4f98a2SYan Zheng struct inode *data_inode; 1633fd0a558SYan, Zheng 1643fd0a558SYan, Zheng struct btrfs_block_rsv *block_rsv; 1653fd0a558SYan, Zheng 1663fd0a558SYan, Zheng struct backref_cache backref_cache; 1673fd0a558SYan, Zheng 1683fd0a558SYan, Zheng struct file_extent_cluster cluster; 1695d4f98a2SYan Zheng /* tree blocks have been processed */ 1705d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 1715d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 1725d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 1735d4f98a2SYan Zheng /* list of reloc trees */ 1745d4f98a2SYan Zheng struct list_head reloc_roots; 1753fd0a558SYan, Zheng /* size of metadata reservation for merging reloc trees */ 1763fd0a558SYan, Zheng u64 merging_rsv_size; 1773fd0a558SYan, Zheng /* size of relocated tree nodes */ 1783fd0a558SYan, Zheng u64 nodes_relocated; 1793fd0a558SYan, Zheng 1805d4f98a2SYan Zheng u64 search_start; 1815d4f98a2SYan Zheng u64 extents_found; 1823fd0a558SYan, Zheng 1833fd0a558SYan, Zheng unsigned int stage:8; 1843fd0a558SYan, Zheng unsigned int create_reloc_tree:1; 1853fd0a558SYan, Zheng unsigned int merge_reloc_tree:1; 1865d4f98a2SYan Zheng unsigned int found_file_extent:1; 1873fd0a558SYan, Zheng unsigned int commit_transaction:1; 1885d4f98a2SYan Zheng }; 1895d4f98a2SYan Zheng 1905d4f98a2SYan Zheng /* stages of data relocation */ 1915d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 1925d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 1935d4f98a2SYan Zheng 1943fd0a558SYan, Zheng static void remove_backref_node(struct backref_cache *cache, 1953fd0a558SYan, Zheng struct backref_node *node); 1963fd0a558SYan, Zheng static void __mark_block_processed(struct reloc_control *rc, 1973fd0a558SYan, Zheng struct backref_node *node); 1985d4f98a2SYan Zheng 1995d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 2005d4f98a2SYan Zheng { 2016bef4d31SEric Paris tree->rb_root = RB_ROOT; 2025d4f98a2SYan Zheng spin_lock_init(&tree->lock); 2035d4f98a2SYan Zheng } 2045d4f98a2SYan Zheng 2055d4f98a2SYan Zheng static void backref_cache_init(struct backref_cache *cache) 2065d4f98a2SYan Zheng { 2075d4f98a2SYan Zheng int i; 2086bef4d31SEric Paris cache->rb_root = RB_ROOT; 2095d4f98a2SYan Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 2105d4f98a2SYan Zheng INIT_LIST_HEAD(&cache->pending[i]); 2113fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->changed); 2123fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->detached); 2133fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->leaves); 2145d4f98a2SYan Zheng } 2155d4f98a2SYan Zheng 2163fd0a558SYan, Zheng static void backref_cache_cleanup(struct backref_cache *cache) 2175d4f98a2SYan Zheng { 2183fd0a558SYan, Zheng struct backref_node *node; 2193fd0a558SYan, Zheng int i; 2203fd0a558SYan, Zheng 2213fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 2223fd0a558SYan, Zheng node = list_entry(cache->detached.next, 2233fd0a558SYan, Zheng struct backref_node, list); 2243fd0a558SYan, Zheng remove_backref_node(cache, node); 2253fd0a558SYan, Zheng } 2263fd0a558SYan, Zheng 2273fd0a558SYan, Zheng while (!list_empty(&cache->leaves)) { 2283fd0a558SYan, Zheng node = list_entry(cache->leaves.next, 2293fd0a558SYan, Zheng struct backref_node, lower); 2303fd0a558SYan, Zheng remove_backref_node(cache, node); 2313fd0a558SYan, Zheng } 2323fd0a558SYan, Zheng 2333fd0a558SYan, Zheng cache->last_trans = 0; 2343fd0a558SYan, Zheng 2353fd0a558SYan, Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 2363fd0a558SYan, Zheng BUG_ON(!list_empty(&cache->pending[i])); 2373fd0a558SYan, Zheng BUG_ON(!list_empty(&cache->changed)); 2383fd0a558SYan, Zheng BUG_ON(!list_empty(&cache->detached)); 2393fd0a558SYan, Zheng BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root)); 2403fd0a558SYan, Zheng BUG_ON(cache->nr_nodes); 2413fd0a558SYan, Zheng BUG_ON(cache->nr_edges); 2423fd0a558SYan, Zheng } 2433fd0a558SYan, Zheng 2443fd0a558SYan, Zheng static struct backref_node *alloc_backref_node(struct backref_cache *cache) 2453fd0a558SYan, Zheng { 2463fd0a558SYan, Zheng struct backref_node *node; 2473fd0a558SYan, Zheng 2483fd0a558SYan, Zheng node = kzalloc(sizeof(*node), GFP_NOFS); 2493fd0a558SYan, Zheng if (node) { 2503fd0a558SYan, Zheng INIT_LIST_HEAD(&node->list); 2515d4f98a2SYan Zheng INIT_LIST_HEAD(&node->upper); 2525d4f98a2SYan Zheng INIT_LIST_HEAD(&node->lower); 2535d4f98a2SYan Zheng RB_CLEAR_NODE(&node->rb_node); 2543fd0a558SYan, Zheng cache->nr_nodes++; 2553fd0a558SYan, Zheng } 2563fd0a558SYan, Zheng return node; 2573fd0a558SYan, Zheng } 2583fd0a558SYan, Zheng 2593fd0a558SYan, Zheng static void free_backref_node(struct backref_cache *cache, 2603fd0a558SYan, Zheng struct backref_node *node) 2613fd0a558SYan, Zheng { 2623fd0a558SYan, Zheng if (node) { 2633fd0a558SYan, Zheng cache->nr_nodes--; 2643fd0a558SYan, Zheng kfree(node); 2653fd0a558SYan, Zheng } 2663fd0a558SYan, Zheng } 2673fd0a558SYan, Zheng 2683fd0a558SYan, Zheng static struct backref_edge *alloc_backref_edge(struct backref_cache *cache) 2693fd0a558SYan, Zheng { 2703fd0a558SYan, Zheng struct backref_edge *edge; 2713fd0a558SYan, Zheng 2723fd0a558SYan, Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 2733fd0a558SYan, Zheng if (edge) 2743fd0a558SYan, Zheng cache->nr_edges++; 2753fd0a558SYan, Zheng return edge; 2763fd0a558SYan, Zheng } 2773fd0a558SYan, Zheng 2783fd0a558SYan, Zheng static void free_backref_edge(struct backref_cache *cache, 2793fd0a558SYan, Zheng struct backref_edge *edge) 2803fd0a558SYan, Zheng { 2813fd0a558SYan, Zheng if (edge) { 2823fd0a558SYan, Zheng cache->nr_edges--; 2833fd0a558SYan, Zheng kfree(edge); 2843fd0a558SYan, Zheng } 2855d4f98a2SYan Zheng } 2865d4f98a2SYan Zheng 2875d4f98a2SYan Zheng static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr, 2885d4f98a2SYan Zheng struct rb_node *node) 2895d4f98a2SYan Zheng { 2905d4f98a2SYan Zheng struct rb_node **p = &root->rb_node; 2915d4f98a2SYan Zheng struct rb_node *parent = NULL; 2925d4f98a2SYan Zheng struct tree_entry *entry; 2935d4f98a2SYan Zheng 2945d4f98a2SYan Zheng while (*p) { 2955d4f98a2SYan Zheng parent = *p; 2965d4f98a2SYan Zheng entry = rb_entry(parent, struct tree_entry, rb_node); 2975d4f98a2SYan Zheng 2985d4f98a2SYan Zheng if (bytenr < entry->bytenr) 2995d4f98a2SYan Zheng p = &(*p)->rb_left; 3005d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 3015d4f98a2SYan Zheng p = &(*p)->rb_right; 3025d4f98a2SYan Zheng else 3035d4f98a2SYan Zheng return parent; 3045d4f98a2SYan Zheng } 3055d4f98a2SYan Zheng 3065d4f98a2SYan Zheng rb_link_node(node, parent, p); 3075d4f98a2SYan Zheng rb_insert_color(node, root); 3085d4f98a2SYan Zheng return NULL; 3095d4f98a2SYan Zheng } 3105d4f98a2SYan Zheng 3115d4f98a2SYan Zheng static struct rb_node *tree_search(struct rb_root *root, u64 bytenr) 3125d4f98a2SYan Zheng { 3135d4f98a2SYan Zheng struct rb_node *n = root->rb_node; 3145d4f98a2SYan Zheng struct tree_entry *entry; 3155d4f98a2SYan Zheng 3165d4f98a2SYan Zheng while (n) { 3175d4f98a2SYan Zheng entry = rb_entry(n, struct tree_entry, rb_node); 3185d4f98a2SYan Zheng 3195d4f98a2SYan Zheng if (bytenr < entry->bytenr) 3205d4f98a2SYan Zheng n = n->rb_left; 3215d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 3225d4f98a2SYan Zheng n = n->rb_right; 3235d4f98a2SYan Zheng else 3245d4f98a2SYan Zheng return n; 3255d4f98a2SYan Zheng } 3265d4f98a2SYan Zheng return NULL; 3275d4f98a2SYan Zheng } 3285d4f98a2SYan Zheng 32943c04fb1SJeff Mahoney void backref_tree_panic(struct rb_node *rb_node, int errno, 33043c04fb1SJeff Mahoney u64 bytenr) 33143c04fb1SJeff Mahoney { 33243c04fb1SJeff Mahoney 33343c04fb1SJeff Mahoney struct btrfs_fs_info *fs_info = NULL; 33443c04fb1SJeff Mahoney struct backref_node *bnode = rb_entry(rb_node, struct backref_node, 33543c04fb1SJeff Mahoney rb_node); 33643c04fb1SJeff Mahoney if (bnode->root) 33743c04fb1SJeff Mahoney fs_info = bnode->root->fs_info; 33843c04fb1SJeff Mahoney btrfs_panic(fs_info, errno, "Inconsistency in backref cache " 33943c04fb1SJeff Mahoney "found at offset %llu\n", (unsigned long long)bytenr); 34043c04fb1SJeff Mahoney } 34143c04fb1SJeff Mahoney 3425d4f98a2SYan Zheng /* 3435d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 3445d4f98a2SYan Zheng */ 3455d4f98a2SYan Zheng static struct backref_node *walk_up_backref(struct backref_node *node, 3465d4f98a2SYan Zheng struct backref_edge *edges[], 3475d4f98a2SYan Zheng int *index) 3485d4f98a2SYan Zheng { 3495d4f98a2SYan Zheng struct backref_edge *edge; 3505d4f98a2SYan Zheng int idx = *index; 3515d4f98a2SYan Zheng 3525d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 3535d4f98a2SYan Zheng edge = list_entry(node->upper.next, 3545d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 3555d4f98a2SYan Zheng edges[idx++] = edge; 3565d4f98a2SYan Zheng node = edge->node[UPPER]; 3575d4f98a2SYan Zheng } 3583fd0a558SYan, Zheng BUG_ON(node->detached); 3595d4f98a2SYan Zheng *index = idx; 3605d4f98a2SYan Zheng return node; 3615d4f98a2SYan Zheng } 3625d4f98a2SYan Zheng 3635d4f98a2SYan Zheng /* 3645d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 3655d4f98a2SYan Zheng */ 3665d4f98a2SYan Zheng static struct backref_node *walk_down_backref(struct backref_edge *edges[], 3675d4f98a2SYan Zheng int *index) 3685d4f98a2SYan Zheng { 3695d4f98a2SYan Zheng struct backref_edge *edge; 3705d4f98a2SYan Zheng struct backref_node *lower; 3715d4f98a2SYan Zheng int idx = *index; 3725d4f98a2SYan Zheng 3735d4f98a2SYan Zheng while (idx > 0) { 3745d4f98a2SYan Zheng edge = edges[idx - 1]; 3755d4f98a2SYan Zheng lower = edge->node[LOWER]; 3765d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 3775d4f98a2SYan Zheng idx--; 3785d4f98a2SYan Zheng continue; 3795d4f98a2SYan Zheng } 3805d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 3815d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 3825d4f98a2SYan Zheng edges[idx - 1] = edge; 3835d4f98a2SYan Zheng *index = idx; 3845d4f98a2SYan Zheng return edge->node[UPPER]; 3855d4f98a2SYan Zheng } 3865d4f98a2SYan Zheng *index = 0; 3875d4f98a2SYan Zheng return NULL; 3885d4f98a2SYan Zheng } 3895d4f98a2SYan Zheng 3903fd0a558SYan, Zheng static void unlock_node_buffer(struct backref_node *node) 3915d4f98a2SYan Zheng { 3925d4f98a2SYan Zheng if (node->locked) { 3935d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 3945d4f98a2SYan Zheng node->locked = 0; 3955d4f98a2SYan Zheng } 3963fd0a558SYan, Zheng } 3973fd0a558SYan, Zheng 3983fd0a558SYan, Zheng static void drop_node_buffer(struct backref_node *node) 3993fd0a558SYan, Zheng { 4003fd0a558SYan, Zheng if (node->eb) { 4013fd0a558SYan, Zheng unlock_node_buffer(node); 4025d4f98a2SYan Zheng free_extent_buffer(node->eb); 4035d4f98a2SYan Zheng node->eb = NULL; 4045d4f98a2SYan Zheng } 4055d4f98a2SYan Zheng } 4065d4f98a2SYan Zheng 4075d4f98a2SYan Zheng static void drop_backref_node(struct backref_cache *tree, 4085d4f98a2SYan Zheng struct backref_node *node) 4095d4f98a2SYan Zheng { 4105d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 4115d4f98a2SYan Zheng 4125d4f98a2SYan Zheng drop_node_buffer(node); 4133fd0a558SYan, Zheng list_del(&node->list); 4145d4f98a2SYan Zheng list_del(&node->lower); 4153fd0a558SYan, Zheng if (!RB_EMPTY_NODE(&node->rb_node)) 4165d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 4173fd0a558SYan, Zheng free_backref_node(tree, node); 4185d4f98a2SYan Zheng } 4195d4f98a2SYan Zheng 4205d4f98a2SYan Zheng /* 4215d4f98a2SYan Zheng * remove a backref node from the backref cache 4225d4f98a2SYan Zheng */ 4235d4f98a2SYan Zheng static void remove_backref_node(struct backref_cache *cache, 4245d4f98a2SYan Zheng struct backref_node *node) 4255d4f98a2SYan Zheng { 4265d4f98a2SYan Zheng struct backref_node *upper; 4275d4f98a2SYan Zheng struct backref_edge *edge; 4285d4f98a2SYan Zheng 4295d4f98a2SYan Zheng if (!node) 4305d4f98a2SYan Zheng return; 4315d4f98a2SYan Zheng 4323fd0a558SYan, Zheng BUG_ON(!node->lowest && !node->detached); 4335d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 4345d4f98a2SYan Zheng edge = list_entry(node->upper.next, struct backref_edge, 4355d4f98a2SYan Zheng list[LOWER]); 4365d4f98a2SYan Zheng upper = edge->node[UPPER]; 4375d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 4385d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 4393fd0a558SYan, Zheng free_backref_edge(cache, edge); 4403fd0a558SYan, Zheng 4413fd0a558SYan, Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 4423fd0a558SYan, Zheng BUG_ON(!list_empty(&node->upper)); 4433fd0a558SYan, Zheng drop_backref_node(cache, node); 4443fd0a558SYan, Zheng node = upper; 4453fd0a558SYan, Zheng node->lowest = 1; 4463fd0a558SYan, Zheng continue; 4473fd0a558SYan, Zheng } 4485d4f98a2SYan Zheng /* 4493fd0a558SYan, Zheng * add the node to leaf node list if no other 4505d4f98a2SYan Zheng * child block cached. 4515d4f98a2SYan Zheng */ 4525d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 4533fd0a558SYan, Zheng list_add_tail(&upper->lower, &cache->leaves); 4545d4f98a2SYan Zheng upper->lowest = 1; 4555d4f98a2SYan Zheng } 4565d4f98a2SYan Zheng } 4573fd0a558SYan, Zheng 4585d4f98a2SYan Zheng drop_backref_node(cache, node); 4595d4f98a2SYan Zheng } 4605d4f98a2SYan Zheng 4613fd0a558SYan, Zheng static void update_backref_node(struct backref_cache *cache, 4623fd0a558SYan, Zheng struct backref_node *node, u64 bytenr) 4633fd0a558SYan, Zheng { 4643fd0a558SYan, Zheng struct rb_node *rb_node; 4653fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 4663fd0a558SYan, Zheng node->bytenr = bytenr; 4673fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 46843c04fb1SJeff Mahoney if (rb_node) 46943c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, bytenr); 4703fd0a558SYan, Zheng } 4713fd0a558SYan, Zheng 4723fd0a558SYan, Zheng /* 4733fd0a558SYan, Zheng * update backref cache after a transaction commit 4743fd0a558SYan, Zheng */ 4753fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 4763fd0a558SYan, Zheng struct backref_cache *cache) 4773fd0a558SYan, Zheng { 4783fd0a558SYan, Zheng struct backref_node *node; 4793fd0a558SYan, Zheng int level = 0; 4803fd0a558SYan, Zheng 4813fd0a558SYan, Zheng if (cache->last_trans == 0) { 4823fd0a558SYan, Zheng cache->last_trans = trans->transid; 4833fd0a558SYan, Zheng return 0; 4843fd0a558SYan, Zheng } 4853fd0a558SYan, Zheng 4863fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 4873fd0a558SYan, Zheng return 0; 4883fd0a558SYan, Zheng 4893fd0a558SYan, Zheng /* 4903fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 4913fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 4923fd0a558SYan, Zheng * so the detached nodes are no longer useful. 4933fd0a558SYan, Zheng */ 4943fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 4953fd0a558SYan, Zheng node = list_entry(cache->detached.next, 4963fd0a558SYan, Zheng struct backref_node, list); 4973fd0a558SYan, Zheng remove_backref_node(cache, node); 4983fd0a558SYan, Zheng } 4993fd0a558SYan, Zheng 5003fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 5013fd0a558SYan, Zheng node = list_entry(cache->changed.next, 5023fd0a558SYan, Zheng struct backref_node, list); 5033fd0a558SYan, Zheng list_del_init(&node->list); 5043fd0a558SYan, Zheng BUG_ON(node->pending); 5053fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 5063fd0a558SYan, Zheng } 5073fd0a558SYan, Zheng 5083fd0a558SYan, Zheng /* 5093fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 5103fd0a558SYan, Zheng * errors during processing the pending nodes. 5113fd0a558SYan, Zheng */ 5123fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 5133fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 5143fd0a558SYan, Zheng BUG_ON(!node->pending); 5153fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 5163fd0a558SYan, Zheng continue; 5173fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 5183fd0a558SYan, Zheng } 5193fd0a558SYan, Zheng } 5203fd0a558SYan, Zheng 5213fd0a558SYan, Zheng cache->last_trans = 0; 5223fd0a558SYan, Zheng return 1; 5233fd0a558SYan, Zheng } 5243fd0a558SYan, Zheng 525f2a97a9dSDavid Sterba 5263fd0a558SYan, Zheng static int should_ignore_root(struct btrfs_root *root) 5273fd0a558SYan, Zheng { 5283fd0a558SYan, Zheng struct btrfs_root *reloc_root; 5293fd0a558SYan, Zheng 5303fd0a558SYan, Zheng if (!root->ref_cows) 5313fd0a558SYan, Zheng return 0; 5323fd0a558SYan, Zheng 5333fd0a558SYan, Zheng reloc_root = root->reloc_root; 5343fd0a558SYan, Zheng if (!reloc_root) 5353fd0a558SYan, Zheng return 0; 5363fd0a558SYan, Zheng 5373fd0a558SYan, Zheng if (btrfs_root_last_snapshot(&reloc_root->root_item) == 5383fd0a558SYan, Zheng root->fs_info->running_transaction->transid - 1) 5393fd0a558SYan, Zheng return 0; 5403fd0a558SYan, Zheng /* 5413fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 5423fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 5433fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 5443fd0a558SYan, Zheng * relocation. 5453fd0a558SYan, Zheng */ 5463fd0a558SYan, Zheng return 1; 5473fd0a558SYan, Zheng } 5485d4f98a2SYan Zheng /* 5495d4f98a2SYan Zheng * find reloc tree by address of tree root 5505d4f98a2SYan Zheng */ 5515d4f98a2SYan Zheng static struct btrfs_root *find_reloc_root(struct reloc_control *rc, 5525d4f98a2SYan Zheng u64 bytenr) 5535d4f98a2SYan Zheng { 5545d4f98a2SYan Zheng struct rb_node *rb_node; 5555d4f98a2SYan Zheng struct mapping_node *node; 5565d4f98a2SYan Zheng struct btrfs_root *root = NULL; 5575d4f98a2SYan Zheng 5585d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 5595d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr); 5605d4f98a2SYan Zheng if (rb_node) { 5615d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 5625d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 5635d4f98a2SYan Zheng } 5645d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 5655d4f98a2SYan Zheng return root; 5665d4f98a2SYan Zheng } 5675d4f98a2SYan Zheng 5685d4f98a2SYan Zheng static int is_cowonly_root(u64 root_objectid) 5695d4f98a2SYan Zheng { 5705d4f98a2SYan Zheng if (root_objectid == BTRFS_ROOT_TREE_OBJECTID || 5715d4f98a2SYan Zheng root_objectid == BTRFS_EXTENT_TREE_OBJECTID || 5725d4f98a2SYan Zheng root_objectid == BTRFS_CHUNK_TREE_OBJECTID || 5735d4f98a2SYan Zheng root_objectid == BTRFS_DEV_TREE_OBJECTID || 5745d4f98a2SYan Zheng root_objectid == BTRFS_TREE_LOG_OBJECTID || 5755d4f98a2SYan Zheng root_objectid == BTRFS_CSUM_TREE_OBJECTID) 5765d4f98a2SYan Zheng return 1; 5775d4f98a2SYan Zheng return 0; 5785d4f98a2SYan Zheng } 5795d4f98a2SYan Zheng 5805d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 5815d4f98a2SYan Zheng u64 root_objectid) 5825d4f98a2SYan Zheng { 5835d4f98a2SYan Zheng struct btrfs_key key; 5845d4f98a2SYan Zheng 5855d4f98a2SYan Zheng key.objectid = root_objectid; 5865d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 5875d4f98a2SYan Zheng if (is_cowonly_root(root_objectid)) 5885d4f98a2SYan Zheng key.offset = 0; 5895d4f98a2SYan Zheng else 5905d4f98a2SYan Zheng key.offset = (u64)-1; 5915d4f98a2SYan Zheng 5925d4f98a2SYan Zheng return btrfs_read_fs_root_no_name(fs_info, &key); 5935d4f98a2SYan Zheng } 5945d4f98a2SYan Zheng 5955d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 5965d4f98a2SYan Zheng static noinline_for_stack 5975d4f98a2SYan Zheng struct btrfs_root *find_tree_root(struct reloc_control *rc, 5985d4f98a2SYan Zheng struct extent_buffer *leaf, 5995d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0) 6005d4f98a2SYan Zheng { 6015d4f98a2SYan Zheng struct btrfs_root *root; 6025d4f98a2SYan Zheng u64 root_objectid = btrfs_ref_root_v0(leaf, ref0); 6035d4f98a2SYan Zheng u64 generation = btrfs_ref_generation_v0(leaf, ref0); 6045d4f98a2SYan Zheng 6055d4f98a2SYan Zheng BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID); 6065d4f98a2SYan Zheng 6075d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, root_objectid); 6085d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 6095d4f98a2SYan Zheng 6105d4f98a2SYan Zheng if (root->ref_cows && 6115d4f98a2SYan Zheng generation != btrfs_root_generation(&root->root_item)) 6125d4f98a2SYan Zheng return NULL; 6135d4f98a2SYan Zheng 6145d4f98a2SYan Zheng return root; 6155d4f98a2SYan Zheng } 6165d4f98a2SYan Zheng #endif 6175d4f98a2SYan Zheng 6185d4f98a2SYan Zheng static noinline_for_stack 6195d4f98a2SYan Zheng int find_inline_backref(struct extent_buffer *leaf, int slot, 6205d4f98a2SYan Zheng unsigned long *ptr, unsigned long *end) 6215d4f98a2SYan Zheng { 6225d4f98a2SYan Zheng struct btrfs_extent_item *ei; 6235d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 6245d4f98a2SYan Zheng u32 item_size; 6255d4f98a2SYan Zheng 6265d4f98a2SYan Zheng item_size = btrfs_item_size_nr(leaf, slot); 6275d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 6285d4f98a2SYan Zheng if (item_size < sizeof(*ei)) { 6295d4f98a2SYan Zheng WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 6305d4f98a2SYan Zheng return 1; 6315d4f98a2SYan Zheng } 6325d4f98a2SYan Zheng #endif 6335d4f98a2SYan Zheng ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item); 6345d4f98a2SYan Zheng WARN_ON(!(btrfs_extent_flags(leaf, ei) & 6355d4f98a2SYan Zheng BTRFS_EXTENT_FLAG_TREE_BLOCK)); 6365d4f98a2SYan Zheng 6375d4f98a2SYan Zheng if (item_size <= sizeof(*ei) + sizeof(*bi)) { 6385d4f98a2SYan Zheng WARN_ON(item_size < sizeof(*ei) + sizeof(*bi)); 6395d4f98a2SYan Zheng return 1; 6405d4f98a2SYan Zheng } 6415d4f98a2SYan Zheng 6425d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 6435d4f98a2SYan Zheng *ptr = (unsigned long)(bi + 1); 6445d4f98a2SYan Zheng *end = (unsigned long)ei + item_size; 6455d4f98a2SYan Zheng return 0; 6465d4f98a2SYan Zheng } 6475d4f98a2SYan Zheng 6485d4f98a2SYan Zheng /* 6495d4f98a2SYan Zheng * build backref tree for a given tree block. root of the backref tree 6505d4f98a2SYan Zheng * corresponds the tree block, leaves of the backref tree correspond 6515d4f98a2SYan Zheng * roots of b-trees that reference the tree block. 6525d4f98a2SYan Zheng * 6535d4f98a2SYan Zheng * the basic idea of this function is check backrefs of a given block 6545d4f98a2SYan Zheng * to find upper level blocks that refernece the block, and then check 6555d4f98a2SYan Zheng * bakcrefs of these upper level blocks recursively. the recursion stop 6565d4f98a2SYan Zheng * when tree root is reached or backrefs for the block is cached. 6575d4f98a2SYan Zheng * 6585d4f98a2SYan Zheng * NOTE: if we find backrefs for a block are cached, we know backrefs 6595d4f98a2SYan Zheng * for all upper level blocks that directly/indirectly reference the 6605d4f98a2SYan Zheng * block are also cached. 6615d4f98a2SYan Zheng */ 6623fd0a558SYan, Zheng static noinline_for_stack 6633fd0a558SYan, Zheng struct backref_node *build_backref_tree(struct reloc_control *rc, 6645d4f98a2SYan Zheng struct btrfs_key *node_key, 6655d4f98a2SYan Zheng int level, u64 bytenr) 6665d4f98a2SYan Zheng { 6673fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 6685d4f98a2SYan Zheng struct btrfs_path *path1; 6695d4f98a2SYan Zheng struct btrfs_path *path2; 6705d4f98a2SYan Zheng struct extent_buffer *eb; 6715d4f98a2SYan Zheng struct btrfs_root *root; 6725d4f98a2SYan Zheng struct backref_node *cur; 6735d4f98a2SYan Zheng struct backref_node *upper; 6745d4f98a2SYan Zheng struct backref_node *lower; 6755d4f98a2SYan Zheng struct backref_node *node = NULL; 6765d4f98a2SYan Zheng struct backref_node *exist = NULL; 6775d4f98a2SYan Zheng struct backref_edge *edge; 6785d4f98a2SYan Zheng struct rb_node *rb_node; 6795d4f98a2SYan Zheng struct btrfs_key key; 6805d4f98a2SYan Zheng unsigned long end; 6815d4f98a2SYan Zheng unsigned long ptr; 6825d4f98a2SYan Zheng LIST_HEAD(list); 6833fd0a558SYan, Zheng LIST_HEAD(useless); 6843fd0a558SYan, Zheng int cowonly; 6855d4f98a2SYan Zheng int ret; 6865d4f98a2SYan Zheng int err = 0; 6875d4f98a2SYan Zheng 6885d4f98a2SYan Zheng path1 = btrfs_alloc_path(); 6895d4f98a2SYan Zheng path2 = btrfs_alloc_path(); 6905d4f98a2SYan Zheng if (!path1 || !path2) { 6915d4f98a2SYan Zheng err = -ENOMEM; 6925d4f98a2SYan Zheng goto out; 6935d4f98a2SYan Zheng } 694026fd317SJosef Bacik path1->reada = 1; 695026fd317SJosef Bacik path2->reada = 2; 6965d4f98a2SYan Zheng 6973fd0a558SYan, Zheng node = alloc_backref_node(cache); 6985d4f98a2SYan Zheng if (!node) { 6995d4f98a2SYan Zheng err = -ENOMEM; 7005d4f98a2SYan Zheng goto out; 7015d4f98a2SYan Zheng } 7025d4f98a2SYan Zheng 7035d4f98a2SYan Zheng node->bytenr = bytenr; 7045d4f98a2SYan Zheng node->level = level; 7055d4f98a2SYan Zheng node->lowest = 1; 7065d4f98a2SYan Zheng cur = node; 7075d4f98a2SYan Zheng again: 7085d4f98a2SYan Zheng end = 0; 7095d4f98a2SYan Zheng ptr = 0; 7105d4f98a2SYan Zheng key.objectid = cur->bytenr; 7115d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 7125d4f98a2SYan Zheng key.offset = (u64)-1; 7135d4f98a2SYan Zheng 7145d4f98a2SYan Zheng path1->search_commit_root = 1; 7155d4f98a2SYan Zheng path1->skip_locking = 1; 7165d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1, 7175d4f98a2SYan Zheng 0, 0); 7185d4f98a2SYan Zheng if (ret < 0) { 7195d4f98a2SYan Zheng err = ret; 7205d4f98a2SYan Zheng goto out; 7215d4f98a2SYan Zheng } 7225d4f98a2SYan Zheng BUG_ON(!ret || !path1->slots[0]); 7235d4f98a2SYan Zheng 7245d4f98a2SYan Zheng path1->slots[0]--; 7255d4f98a2SYan Zheng 7265d4f98a2SYan Zheng WARN_ON(cur->checked); 7275d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 7285d4f98a2SYan Zheng /* 72970f23fd6SJustin P. Mattock * the backref was added previously when processing 7305d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 7315d4f98a2SYan Zheng */ 7325d4f98a2SYan Zheng BUG_ON(!list_is_singular(&cur->upper)); 7335d4f98a2SYan Zheng edge = list_entry(cur->upper.next, struct backref_edge, 7345d4f98a2SYan Zheng list[LOWER]); 7355d4f98a2SYan Zheng BUG_ON(!list_empty(&edge->list[UPPER])); 7365d4f98a2SYan Zheng exist = edge->node[UPPER]; 7375d4f98a2SYan Zheng /* 7385d4f98a2SYan Zheng * add the upper level block to pending list if we need 7395d4f98a2SYan Zheng * check its backrefs 7405d4f98a2SYan Zheng */ 7415d4f98a2SYan Zheng if (!exist->checked) 7425d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 7435d4f98a2SYan Zheng } else { 7445d4f98a2SYan Zheng exist = NULL; 7455d4f98a2SYan Zheng } 7465d4f98a2SYan Zheng 7475d4f98a2SYan Zheng while (1) { 7485d4f98a2SYan Zheng cond_resched(); 7495d4f98a2SYan Zheng eb = path1->nodes[0]; 7505d4f98a2SYan Zheng 7515d4f98a2SYan Zheng if (ptr >= end) { 7525d4f98a2SYan Zheng if (path1->slots[0] >= btrfs_header_nritems(eb)) { 7535d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path1); 7545d4f98a2SYan Zheng if (ret < 0) { 7555d4f98a2SYan Zheng err = ret; 7565d4f98a2SYan Zheng goto out; 7575d4f98a2SYan Zheng } 7585d4f98a2SYan Zheng if (ret > 0) 7595d4f98a2SYan Zheng break; 7605d4f98a2SYan Zheng eb = path1->nodes[0]; 7615d4f98a2SYan Zheng } 7625d4f98a2SYan Zheng 7635d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path1->slots[0]); 7645d4f98a2SYan Zheng if (key.objectid != cur->bytenr) { 7655d4f98a2SYan Zheng WARN_ON(exist); 7665d4f98a2SYan Zheng break; 7675d4f98a2SYan Zheng } 7685d4f98a2SYan Zheng 7695d4f98a2SYan Zheng if (key.type == BTRFS_EXTENT_ITEM_KEY) { 7705d4f98a2SYan Zheng ret = find_inline_backref(eb, path1->slots[0], 7715d4f98a2SYan Zheng &ptr, &end); 7725d4f98a2SYan Zheng if (ret) 7735d4f98a2SYan Zheng goto next; 7745d4f98a2SYan Zheng } 7755d4f98a2SYan Zheng } 7765d4f98a2SYan Zheng 7775d4f98a2SYan Zheng if (ptr < end) { 7785d4f98a2SYan Zheng /* update key for inline back ref */ 7795d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 7805d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 7815d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 7825d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 7835d4f98a2SYan Zheng WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY && 7845d4f98a2SYan Zheng key.type != BTRFS_SHARED_BLOCK_REF_KEY); 7855d4f98a2SYan Zheng } 7865d4f98a2SYan Zheng 7875d4f98a2SYan Zheng if (exist && 7885d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 7895d4f98a2SYan Zheng exist->owner == key.offset) || 7905d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 7915d4f98a2SYan Zheng exist->bytenr == key.offset))) { 7925d4f98a2SYan Zheng exist = NULL; 7935d4f98a2SYan Zheng goto next; 7945d4f98a2SYan Zheng } 7955d4f98a2SYan Zheng 7965d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 7975d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY || 7985d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 7993fd0a558SYan, Zheng if (key.type == BTRFS_EXTENT_REF_V0_KEY) { 8005d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 8015d4f98a2SYan Zheng ref0 = btrfs_item_ptr(eb, path1->slots[0], 8025d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 8033fd0a558SYan, Zheng if (key.objectid == key.offset) { 804046f264fSYan, Zheng root = find_tree_root(rc, eb, ref0); 8053fd0a558SYan, Zheng if (root && !should_ignore_root(root)) 8065d4f98a2SYan Zheng cur->root = root; 8075d4f98a2SYan Zheng else 8083fd0a558SYan, Zheng list_add(&cur->list, &useless); 8095d4f98a2SYan Zheng break; 8105d4f98a2SYan Zheng } 811046f264fSYan, Zheng if (is_cowonly_root(btrfs_ref_root_v0(eb, 812046f264fSYan, Zheng ref0))) 813046f264fSYan, Zheng cur->cowonly = 1; 8143fd0a558SYan, Zheng } 8155d4f98a2SYan Zheng #else 8165d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 8175d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 8185d4f98a2SYan Zheng #endif 8195d4f98a2SYan Zheng if (key.objectid == key.offset) { 8205d4f98a2SYan Zheng /* 8215d4f98a2SYan Zheng * only root blocks of reloc trees use 8225d4f98a2SYan Zheng * backref of this type. 8235d4f98a2SYan Zheng */ 8245d4f98a2SYan Zheng root = find_reloc_root(rc, cur->bytenr); 8255d4f98a2SYan Zheng BUG_ON(!root); 8265d4f98a2SYan Zheng cur->root = root; 8275d4f98a2SYan Zheng break; 8285d4f98a2SYan Zheng } 8295d4f98a2SYan Zheng 8303fd0a558SYan, Zheng edge = alloc_backref_edge(cache); 8315d4f98a2SYan Zheng if (!edge) { 8325d4f98a2SYan Zheng err = -ENOMEM; 8335d4f98a2SYan Zheng goto out; 8345d4f98a2SYan Zheng } 8355d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, key.offset); 8365d4f98a2SYan Zheng if (!rb_node) { 8373fd0a558SYan, Zheng upper = alloc_backref_node(cache); 8385d4f98a2SYan Zheng if (!upper) { 8393fd0a558SYan, Zheng free_backref_edge(cache, edge); 8405d4f98a2SYan Zheng err = -ENOMEM; 8415d4f98a2SYan Zheng goto out; 8425d4f98a2SYan Zheng } 8435d4f98a2SYan Zheng upper->bytenr = key.offset; 8445d4f98a2SYan Zheng upper->level = cur->level + 1; 8455d4f98a2SYan Zheng /* 8465d4f98a2SYan Zheng * backrefs for the upper level block isn't 8475d4f98a2SYan Zheng * cached, add the block to pending list 8485d4f98a2SYan Zheng */ 8495d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 8505d4f98a2SYan Zheng } else { 8515d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 8525d4f98a2SYan Zheng rb_node); 8533fd0a558SYan, Zheng BUG_ON(!upper->checked); 8545d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 8555d4f98a2SYan Zheng } 8563fd0a558SYan, Zheng list_add_tail(&edge->list[LOWER], &cur->upper); 8575d4f98a2SYan Zheng edge->node[LOWER] = cur; 8583fd0a558SYan, Zheng edge->node[UPPER] = upper; 8595d4f98a2SYan Zheng 8605d4f98a2SYan Zheng goto next; 8615d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 8625d4f98a2SYan Zheng goto next; 8635d4f98a2SYan Zheng } 8645d4f98a2SYan Zheng 8655d4f98a2SYan Zheng /* key.type == BTRFS_TREE_BLOCK_REF_KEY */ 8665d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, key.offset); 8675d4f98a2SYan Zheng if (IS_ERR(root)) { 8685d4f98a2SYan Zheng err = PTR_ERR(root); 8695d4f98a2SYan Zheng goto out; 8705d4f98a2SYan Zheng } 8715d4f98a2SYan Zheng 8723fd0a558SYan, Zheng if (!root->ref_cows) 8733fd0a558SYan, Zheng cur->cowonly = 1; 8743fd0a558SYan, Zheng 8755d4f98a2SYan Zheng if (btrfs_root_level(&root->root_item) == cur->level) { 8765d4f98a2SYan Zheng /* tree root */ 8775d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 8785d4f98a2SYan Zheng cur->bytenr); 8793fd0a558SYan, Zheng if (should_ignore_root(root)) 8803fd0a558SYan, Zheng list_add(&cur->list, &useless); 8813fd0a558SYan, Zheng else 8825d4f98a2SYan Zheng cur->root = root; 8835d4f98a2SYan Zheng break; 8845d4f98a2SYan Zheng } 8855d4f98a2SYan Zheng 8865d4f98a2SYan Zheng level = cur->level + 1; 8875d4f98a2SYan Zheng 8885d4f98a2SYan Zheng /* 8895d4f98a2SYan Zheng * searching the tree to find upper level blocks 8905d4f98a2SYan Zheng * reference the block. 8915d4f98a2SYan Zheng */ 8925d4f98a2SYan Zheng path2->search_commit_root = 1; 8935d4f98a2SYan Zheng path2->skip_locking = 1; 8945d4f98a2SYan Zheng path2->lowest_level = level; 8955d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0); 8965d4f98a2SYan Zheng path2->lowest_level = 0; 8975d4f98a2SYan Zheng if (ret < 0) { 8985d4f98a2SYan Zheng err = ret; 8995d4f98a2SYan Zheng goto out; 9005d4f98a2SYan Zheng } 90133c66f43SYan Zheng if (ret > 0 && path2->slots[level] > 0) 90233c66f43SYan Zheng path2->slots[level]--; 9035d4f98a2SYan Zheng 9045d4f98a2SYan Zheng eb = path2->nodes[level]; 9055d4f98a2SYan Zheng WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) != 9065d4f98a2SYan Zheng cur->bytenr); 9075d4f98a2SYan Zheng 9085d4f98a2SYan Zheng lower = cur; 9095d4f98a2SYan Zheng for (; level < BTRFS_MAX_LEVEL; level++) { 9105d4f98a2SYan Zheng if (!path2->nodes[level]) { 9115d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 9125d4f98a2SYan Zheng lower->bytenr); 9133fd0a558SYan, Zheng if (should_ignore_root(root)) 9143fd0a558SYan, Zheng list_add(&lower->list, &useless); 9153fd0a558SYan, Zheng else 9165d4f98a2SYan Zheng lower->root = root; 9175d4f98a2SYan Zheng break; 9185d4f98a2SYan Zheng } 9195d4f98a2SYan Zheng 9203fd0a558SYan, Zheng edge = alloc_backref_edge(cache); 9215d4f98a2SYan Zheng if (!edge) { 9225d4f98a2SYan Zheng err = -ENOMEM; 9235d4f98a2SYan Zheng goto out; 9245d4f98a2SYan Zheng } 9255d4f98a2SYan Zheng 9265d4f98a2SYan Zheng eb = path2->nodes[level]; 9275d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, eb->start); 9285d4f98a2SYan Zheng if (!rb_node) { 9293fd0a558SYan, Zheng upper = alloc_backref_node(cache); 9305d4f98a2SYan Zheng if (!upper) { 9313fd0a558SYan, Zheng free_backref_edge(cache, edge); 9325d4f98a2SYan Zheng err = -ENOMEM; 9335d4f98a2SYan Zheng goto out; 9345d4f98a2SYan Zheng } 9355d4f98a2SYan Zheng upper->bytenr = eb->start; 9365d4f98a2SYan Zheng upper->owner = btrfs_header_owner(eb); 9375d4f98a2SYan Zheng upper->level = lower->level + 1; 9383fd0a558SYan, Zheng if (!root->ref_cows) 9393fd0a558SYan, Zheng upper->cowonly = 1; 9405d4f98a2SYan Zheng 9415d4f98a2SYan Zheng /* 9425d4f98a2SYan Zheng * if we know the block isn't shared 9435d4f98a2SYan Zheng * we can void checking its backrefs. 9445d4f98a2SYan Zheng */ 9455d4f98a2SYan Zheng if (btrfs_block_can_be_shared(root, eb)) 9465d4f98a2SYan Zheng upper->checked = 0; 9475d4f98a2SYan Zheng else 9485d4f98a2SYan Zheng upper->checked = 1; 9495d4f98a2SYan Zheng 9505d4f98a2SYan Zheng /* 9515d4f98a2SYan Zheng * add the block to pending list if we 9525d4f98a2SYan Zheng * need check its backrefs. only block 9535d4f98a2SYan Zheng * at 'cur->level + 1' is added to the 9545d4f98a2SYan Zheng * tail of pending list. this guarantees 9555d4f98a2SYan Zheng * we check backrefs from lower level 9565d4f98a2SYan Zheng * blocks to upper level blocks. 9575d4f98a2SYan Zheng */ 9585d4f98a2SYan Zheng if (!upper->checked && 9595d4f98a2SYan Zheng level == cur->level + 1) { 9605d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], 9615d4f98a2SYan Zheng &list); 9625d4f98a2SYan Zheng } else 9635d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 9645d4f98a2SYan Zheng } else { 9655d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 9665d4f98a2SYan Zheng rb_node); 9675d4f98a2SYan Zheng BUG_ON(!upper->checked); 9685d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 9693fd0a558SYan, Zheng if (!upper->owner) 9703fd0a558SYan, Zheng upper->owner = btrfs_header_owner(eb); 9715d4f98a2SYan Zheng } 9725d4f98a2SYan Zheng list_add_tail(&edge->list[LOWER], &lower->upper); 9735d4f98a2SYan Zheng edge->node[LOWER] = lower; 9743fd0a558SYan, Zheng edge->node[UPPER] = upper; 9755d4f98a2SYan Zheng 9765d4f98a2SYan Zheng if (rb_node) 9775d4f98a2SYan Zheng break; 9785d4f98a2SYan Zheng lower = upper; 9795d4f98a2SYan Zheng upper = NULL; 9805d4f98a2SYan Zheng } 981b3b4aa74SDavid Sterba btrfs_release_path(path2); 9825d4f98a2SYan Zheng next: 9835d4f98a2SYan Zheng if (ptr < end) { 9845d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 9855d4f98a2SYan Zheng if (ptr >= end) { 9865d4f98a2SYan Zheng WARN_ON(ptr > end); 9875d4f98a2SYan Zheng ptr = 0; 9885d4f98a2SYan Zheng end = 0; 9895d4f98a2SYan Zheng } 9905d4f98a2SYan Zheng } 9915d4f98a2SYan Zheng if (ptr >= end) 9925d4f98a2SYan Zheng path1->slots[0]++; 9935d4f98a2SYan Zheng } 994b3b4aa74SDavid Sterba btrfs_release_path(path1); 9955d4f98a2SYan Zheng 9965d4f98a2SYan Zheng cur->checked = 1; 9975d4f98a2SYan Zheng WARN_ON(exist); 9985d4f98a2SYan Zheng 9995d4f98a2SYan Zheng /* the pending list isn't empty, take the first block to process */ 10005d4f98a2SYan Zheng if (!list_empty(&list)) { 10015d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 10025d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 10035d4f98a2SYan Zheng cur = edge->node[UPPER]; 10045d4f98a2SYan Zheng goto again; 10055d4f98a2SYan Zheng } 10065d4f98a2SYan Zheng 10075d4f98a2SYan Zheng /* 10085d4f98a2SYan Zheng * everything goes well, connect backref nodes and insert backref nodes 10095d4f98a2SYan Zheng * into the cache. 10105d4f98a2SYan Zheng */ 10115d4f98a2SYan Zheng BUG_ON(!node->checked); 10123fd0a558SYan, Zheng cowonly = node->cowonly; 10133fd0a558SYan, Zheng if (!cowonly) { 10143fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, 10153fd0a558SYan, Zheng &node->rb_node); 101643c04fb1SJeff Mahoney if (rb_node) 101743c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 10183fd0a558SYan, Zheng list_add_tail(&node->lower, &cache->leaves); 10193fd0a558SYan, Zheng } 10205d4f98a2SYan Zheng 10215d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) 10225d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 10235d4f98a2SYan Zheng 10245d4f98a2SYan Zheng while (!list_empty(&list)) { 10255d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 10265d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 10275d4f98a2SYan Zheng upper = edge->node[UPPER]; 10283fd0a558SYan, Zheng if (upper->detached) { 10293fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 10303fd0a558SYan, Zheng lower = edge->node[LOWER]; 10313fd0a558SYan, Zheng free_backref_edge(cache, edge); 10323fd0a558SYan, Zheng if (list_empty(&lower->upper)) 10333fd0a558SYan, Zheng list_add(&lower->list, &useless); 10343fd0a558SYan, Zheng continue; 10353fd0a558SYan, Zheng } 10365d4f98a2SYan Zheng 10375d4f98a2SYan Zheng if (!RB_EMPTY_NODE(&upper->rb_node)) { 10385d4f98a2SYan Zheng if (upper->lowest) { 10395d4f98a2SYan Zheng list_del_init(&upper->lower); 10405d4f98a2SYan Zheng upper->lowest = 0; 10415d4f98a2SYan Zheng } 10425d4f98a2SYan Zheng 10435d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 10445d4f98a2SYan Zheng continue; 10455d4f98a2SYan Zheng } 10465d4f98a2SYan Zheng 10475d4f98a2SYan Zheng BUG_ON(!upper->checked); 10483fd0a558SYan, Zheng BUG_ON(cowonly != upper->cowonly); 10493fd0a558SYan, Zheng if (!cowonly) { 10505d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, upper->bytenr, 10515d4f98a2SYan Zheng &upper->rb_node); 105243c04fb1SJeff Mahoney if (rb_node) 105343c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, 105443c04fb1SJeff Mahoney upper->bytenr); 10553fd0a558SYan, Zheng } 10565d4f98a2SYan Zheng 10575d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 10585d4f98a2SYan Zheng 10595d4f98a2SYan Zheng list_for_each_entry(edge, &upper->upper, list[LOWER]) 10605d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 10615d4f98a2SYan Zheng } 10623fd0a558SYan, Zheng /* 10633fd0a558SYan, Zheng * process useless backref nodes. backref nodes for tree leaves 10643fd0a558SYan, Zheng * are deleted from the cache. backref nodes for upper level 10653fd0a558SYan, Zheng * tree blocks are left in the cache to avoid unnecessary backref 10663fd0a558SYan, Zheng * lookup. 10673fd0a558SYan, Zheng */ 10683fd0a558SYan, Zheng while (!list_empty(&useless)) { 10693fd0a558SYan, Zheng upper = list_entry(useless.next, struct backref_node, list); 10703fd0a558SYan, Zheng list_del_init(&upper->list); 10713fd0a558SYan, Zheng BUG_ON(!list_empty(&upper->upper)); 10723fd0a558SYan, Zheng if (upper == node) 10733fd0a558SYan, Zheng node = NULL; 10743fd0a558SYan, Zheng if (upper->lowest) { 10753fd0a558SYan, Zheng list_del_init(&upper->lower); 10763fd0a558SYan, Zheng upper->lowest = 0; 10773fd0a558SYan, Zheng } 10783fd0a558SYan, Zheng while (!list_empty(&upper->lower)) { 10793fd0a558SYan, Zheng edge = list_entry(upper->lower.next, 10803fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 10813fd0a558SYan, Zheng list_del(&edge->list[UPPER]); 10823fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 10833fd0a558SYan, Zheng lower = edge->node[LOWER]; 10843fd0a558SYan, Zheng free_backref_edge(cache, edge); 10853fd0a558SYan, Zheng 10863fd0a558SYan, Zheng if (list_empty(&lower->upper)) 10873fd0a558SYan, Zheng list_add(&lower->list, &useless); 10883fd0a558SYan, Zheng } 10893fd0a558SYan, Zheng __mark_block_processed(rc, upper); 10903fd0a558SYan, Zheng if (upper->level > 0) { 10913fd0a558SYan, Zheng list_add(&upper->list, &cache->detached); 10923fd0a558SYan, Zheng upper->detached = 1; 10933fd0a558SYan, Zheng } else { 10943fd0a558SYan, Zheng rb_erase(&upper->rb_node, &cache->rb_root); 10953fd0a558SYan, Zheng free_backref_node(cache, upper); 10963fd0a558SYan, Zheng } 10973fd0a558SYan, Zheng } 10985d4f98a2SYan Zheng out: 10995d4f98a2SYan Zheng btrfs_free_path(path1); 11005d4f98a2SYan Zheng btrfs_free_path(path2); 11015d4f98a2SYan Zheng if (err) { 11023fd0a558SYan, Zheng while (!list_empty(&useless)) { 11033fd0a558SYan, Zheng lower = list_entry(useless.next, 11043fd0a558SYan, Zheng struct backref_node, upper); 11053fd0a558SYan, Zheng list_del_init(&lower->upper); 11063fd0a558SYan, Zheng } 11075d4f98a2SYan Zheng upper = node; 11083fd0a558SYan, Zheng INIT_LIST_HEAD(&list); 11095d4f98a2SYan Zheng while (upper) { 11105d4f98a2SYan Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 11115d4f98a2SYan Zheng list_splice_tail(&upper->upper, &list); 11123fd0a558SYan, Zheng free_backref_node(cache, upper); 11135d4f98a2SYan Zheng } 11145d4f98a2SYan Zheng 11155d4f98a2SYan Zheng if (list_empty(&list)) 11165d4f98a2SYan Zheng break; 11175d4f98a2SYan Zheng 11185d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, 11195d4f98a2SYan Zheng list[LOWER]); 11203fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 11215d4f98a2SYan Zheng upper = edge->node[UPPER]; 11223fd0a558SYan, Zheng free_backref_edge(cache, edge); 11235d4f98a2SYan Zheng } 11245d4f98a2SYan Zheng return ERR_PTR(err); 11255d4f98a2SYan Zheng } 11263fd0a558SYan, Zheng BUG_ON(node && node->detached); 11275d4f98a2SYan Zheng return node; 11285d4f98a2SYan Zheng } 11295d4f98a2SYan Zheng 11305d4f98a2SYan Zheng /* 11313fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 11323fd0a558SYan, Zheng * the backref node is created by cloning backref node that 11333fd0a558SYan, Zheng * corresponds to root of source tree 11343fd0a558SYan, Zheng */ 11353fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 11363fd0a558SYan, Zheng struct reloc_control *rc, 11373fd0a558SYan, Zheng struct btrfs_root *src, 11383fd0a558SYan, Zheng struct btrfs_root *dest) 11393fd0a558SYan, Zheng { 11403fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 11413fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 11423fd0a558SYan, Zheng struct backref_node *node = NULL; 11433fd0a558SYan, Zheng struct backref_node *new_node; 11443fd0a558SYan, Zheng struct backref_edge *edge; 11453fd0a558SYan, Zheng struct backref_edge *new_edge; 11463fd0a558SYan, Zheng struct rb_node *rb_node; 11473fd0a558SYan, Zheng 11483fd0a558SYan, Zheng if (cache->last_trans > 0) 11493fd0a558SYan, Zheng update_backref_cache(trans, cache); 11503fd0a558SYan, Zheng 11513fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, src->commit_root->start); 11523fd0a558SYan, Zheng if (rb_node) { 11533fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, rb_node); 11543fd0a558SYan, Zheng if (node->detached) 11553fd0a558SYan, Zheng node = NULL; 11563fd0a558SYan, Zheng else 11573fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 11583fd0a558SYan, Zheng } 11593fd0a558SYan, Zheng 11603fd0a558SYan, Zheng if (!node) { 11613fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, 11623fd0a558SYan, Zheng reloc_root->commit_root->start); 11633fd0a558SYan, Zheng if (rb_node) { 11643fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, 11653fd0a558SYan, Zheng rb_node); 11663fd0a558SYan, Zheng BUG_ON(node->detached); 11673fd0a558SYan, Zheng } 11683fd0a558SYan, Zheng } 11693fd0a558SYan, Zheng 11703fd0a558SYan, Zheng if (!node) 11713fd0a558SYan, Zheng return 0; 11723fd0a558SYan, Zheng 11733fd0a558SYan, Zheng new_node = alloc_backref_node(cache); 11743fd0a558SYan, Zheng if (!new_node) 11753fd0a558SYan, Zheng return -ENOMEM; 11763fd0a558SYan, Zheng 11773fd0a558SYan, Zheng new_node->bytenr = dest->node->start; 11783fd0a558SYan, Zheng new_node->level = node->level; 11793fd0a558SYan, Zheng new_node->lowest = node->lowest; 11806848ad64SYan, Zheng new_node->checked = 1; 11813fd0a558SYan, Zheng new_node->root = dest; 11823fd0a558SYan, Zheng 11833fd0a558SYan, Zheng if (!node->lowest) { 11843fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 11853fd0a558SYan, Zheng new_edge = alloc_backref_edge(cache); 11863fd0a558SYan, Zheng if (!new_edge) 11873fd0a558SYan, Zheng goto fail; 11883fd0a558SYan, Zheng 11893fd0a558SYan, Zheng new_edge->node[UPPER] = new_node; 11903fd0a558SYan, Zheng new_edge->node[LOWER] = edge->node[LOWER]; 11913fd0a558SYan, Zheng list_add_tail(&new_edge->list[UPPER], 11923fd0a558SYan, Zheng &new_node->lower); 11933fd0a558SYan, Zheng } 119476b9e23dSMiao Xie } else { 119576b9e23dSMiao Xie list_add_tail(&new_node->lower, &cache->leaves); 11963fd0a558SYan, Zheng } 11973fd0a558SYan, Zheng 11983fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, new_node->bytenr, 11993fd0a558SYan, Zheng &new_node->rb_node); 120043c04fb1SJeff Mahoney if (rb_node) 120143c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, new_node->bytenr); 12023fd0a558SYan, Zheng 12033fd0a558SYan, Zheng if (!new_node->lowest) { 12043fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 12053fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 12063fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 12073fd0a558SYan, Zheng } 12083fd0a558SYan, Zheng } 12093fd0a558SYan, Zheng return 0; 12103fd0a558SYan, Zheng fail: 12113fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 12123fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 12133fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 12143fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 12153fd0a558SYan, Zheng free_backref_edge(cache, new_edge); 12163fd0a558SYan, Zheng } 12173fd0a558SYan, Zheng free_backref_node(cache, new_node); 12183fd0a558SYan, Zheng return -ENOMEM; 12193fd0a558SYan, Zheng } 12203fd0a558SYan, Zheng 12213fd0a558SYan, Zheng /* 12225d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 12235d4f98a2SYan Zheng */ 1224ffd7b339SJeff Mahoney static int __must_check __add_reloc_root(struct btrfs_root *root) 12255d4f98a2SYan Zheng { 12265d4f98a2SYan Zheng struct rb_node *rb_node; 12275d4f98a2SYan Zheng struct mapping_node *node; 12285d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 12295d4f98a2SYan Zheng 12305d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 1231ffd7b339SJeff Mahoney if (!node) 1232ffd7b339SJeff Mahoney return -ENOMEM; 12335d4f98a2SYan Zheng 12345d4f98a2SYan Zheng node->bytenr = root->node->start; 12355d4f98a2SYan Zheng node->data = root; 12365d4f98a2SYan Zheng 12375d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12385d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 12395d4f98a2SYan Zheng node->bytenr, &node->rb_node); 12405d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 1241ffd7b339SJeff Mahoney if (rb_node) { 1242ffd7b339SJeff Mahoney btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found " 1243ffd7b339SJeff Mahoney "for start=%llu while inserting into relocation " 1244533574c6SJoe Perches "tree\n", node->bytenr); 124523291a04SDan Carpenter kfree(node); 124623291a04SDan Carpenter return -EEXIST; 1247ffd7b339SJeff Mahoney } 12485d4f98a2SYan Zheng 12495d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 12505d4f98a2SYan Zheng return 0; 12515d4f98a2SYan Zheng } 12525d4f98a2SYan Zheng 12535d4f98a2SYan Zheng /* 12545d4f98a2SYan Zheng * helper to update/delete the 'address of tree root -> reloc tree' 12555d4f98a2SYan Zheng * mapping 12565d4f98a2SYan Zheng */ 12575d4f98a2SYan Zheng static int __update_reloc_root(struct btrfs_root *root, int del) 12585d4f98a2SYan Zheng { 12595d4f98a2SYan Zheng struct rb_node *rb_node; 12605d4f98a2SYan Zheng struct mapping_node *node = NULL; 12615d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 12625d4f98a2SYan Zheng 12635d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12645d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, 12655d4f98a2SYan Zheng root->commit_root->start); 12665d4f98a2SYan Zheng if (rb_node) { 12675d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 12685d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 12695d4f98a2SYan Zheng } 12705d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 12715d4f98a2SYan Zheng 12728f71f3e0SLiu Bo if (!node) 12738f71f3e0SLiu Bo return 0; 12745d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 12755d4f98a2SYan Zheng 12765d4f98a2SYan Zheng if (!del) { 12775d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12785d4f98a2SYan Zheng node->bytenr = root->node->start; 12795d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 12805d4f98a2SYan Zheng node->bytenr, &node->rb_node); 12815d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 128243c04fb1SJeff Mahoney if (rb_node) 128343c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 12845d4f98a2SYan Zheng } else { 12851daf3540SDaniel J Blueman spin_lock(&root->fs_info->trans_lock); 12865d4f98a2SYan Zheng list_del_init(&root->root_list); 12871daf3540SDaniel J Blueman spin_unlock(&root->fs_info->trans_lock); 12885d4f98a2SYan Zheng kfree(node); 12895d4f98a2SYan Zheng } 12905d4f98a2SYan Zheng return 0; 12915d4f98a2SYan Zheng } 12925d4f98a2SYan Zheng 12933fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 12943fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 12955d4f98a2SYan Zheng { 12965d4f98a2SYan Zheng struct btrfs_root *reloc_root; 12975d4f98a2SYan Zheng struct extent_buffer *eb; 12985d4f98a2SYan Zheng struct btrfs_root_item *root_item; 12995d4f98a2SYan Zheng struct btrfs_key root_key; 13005d4f98a2SYan Zheng int ret; 13015d4f98a2SYan Zheng 13025d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 13035d4f98a2SYan Zheng BUG_ON(!root_item); 13045d4f98a2SYan Zheng 13055d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 13065d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 13073fd0a558SYan, Zheng root_key.offset = objectid; 13085d4f98a2SYan Zheng 13093fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 13103fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 13115d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 13125d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 13135d4f98a2SYan Zheng BUG_ON(ret); 13145d4f98a2SYan Zheng 13153fd0a558SYan, Zheng btrfs_set_root_last_snapshot(&root->root_item, 13163fd0a558SYan, Zheng trans->transid - 1); 13173fd0a558SYan, Zheng } else { 13183fd0a558SYan, Zheng /* 13193fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 13203fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 13213fd0a558SYan, Zheng * modified after it was created have RELOC flag 13223fd0a558SYan, Zheng * set in their headers. so it's OK to not update 13233fd0a558SYan, Zheng * the 'last_snapshot'. 13243fd0a558SYan, Zheng */ 13253fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 13263fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 13273fd0a558SYan, Zheng BUG_ON(ret); 13283fd0a558SYan, Zheng } 13293fd0a558SYan, Zheng 13305d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 13315d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 13325d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 13335d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 13343fd0a558SYan, Zheng 13353fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 13363fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 13373fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 13383fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 13395d4f98a2SYan Zheng root_item->drop_level = 0; 13403fd0a558SYan, Zheng } 13415d4f98a2SYan Zheng 13425d4f98a2SYan Zheng btrfs_tree_unlock(eb); 13435d4f98a2SYan Zheng free_extent_buffer(eb); 13445d4f98a2SYan Zheng 13455d4f98a2SYan Zheng ret = btrfs_insert_root(trans, root->fs_info->tree_root, 13465d4f98a2SYan Zheng &root_key, root_item); 13475d4f98a2SYan Zheng BUG_ON(ret); 13485d4f98a2SYan Zheng kfree(root_item); 13495d4f98a2SYan Zheng 13505d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root, 13515d4f98a2SYan Zheng &root_key); 13525d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 13535d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 13543fd0a558SYan, Zheng return reloc_root; 13553fd0a558SYan, Zheng } 13563fd0a558SYan, Zheng 13573fd0a558SYan, Zheng /* 13583fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 13593fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 13603fd0a558SYan, Zheng */ 13613fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 13623fd0a558SYan, Zheng struct btrfs_root *root) 13633fd0a558SYan, Zheng { 13643fd0a558SYan, Zheng struct btrfs_root *reloc_root; 13653fd0a558SYan, Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 13663fd0a558SYan, Zheng int clear_rsv = 0; 1367ffd7b339SJeff Mahoney int ret; 13683fd0a558SYan, Zheng 13693fd0a558SYan, Zheng if (root->reloc_root) { 13703fd0a558SYan, Zheng reloc_root = root->reloc_root; 13713fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 13723fd0a558SYan, Zheng return 0; 13733fd0a558SYan, Zheng } 13743fd0a558SYan, Zheng 13753fd0a558SYan, Zheng if (!rc || !rc->create_reloc_tree || 13763fd0a558SYan, Zheng root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 13773fd0a558SYan, Zheng return 0; 13783fd0a558SYan, Zheng 13793fd0a558SYan, Zheng if (!trans->block_rsv) { 13803fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 13813fd0a558SYan, Zheng clear_rsv = 1; 13823fd0a558SYan, Zheng } 13833fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 13843fd0a558SYan, Zheng if (clear_rsv) 13853fd0a558SYan, Zheng trans->block_rsv = NULL; 13865d4f98a2SYan Zheng 1387ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 1388ffd7b339SJeff Mahoney BUG_ON(ret < 0); 13895d4f98a2SYan Zheng root->reloc_root = reloc_root; 13905d4f98a2SYan Zheng return 0; 13915d4f98a2SYan Zheng } 13925d4f98a2SYan Zheng 13935d4f98a2SYan Zheng /* 13945d4f98a2SYan Zheng * update root item of reloc tree 13955d4f98a2SYan Zheng */ 13965d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 13975d4f98a2SYan Zheng struct btrfs_root *root) 13985d4f98a2SYan Zheng { 13995d4f98a2SYan Zheng struct btrfs_root *reloc_root; 14005d4f98a2SYan Zheng struct btrfs_root_item *root_item; 14015d4f98a2SYan Zheng int del = 0; 14025d4f98a2SYan Zheng int ret; 14035d4f98a2SYan Zheng 14045d4f98a2SYan Zheng if (!root->reloc_root) 14057585717fSChris Mason goto out; 14065d4f98a2SYan Zheng 14075d4f98a2SYan Zheng reloc_root = root->reloc_root; 14085d4f98a2SYan Zheng root_item = &reloc_root->root_item; 14095d4f98a2SYan Zheng 14103fd0a558SYan, Zheng if (root->fs_info->reloc_ctl->merge_reloc_tree && 14113fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 14125d4f98a2SYan Zheng root->reloc_root = NULL; 14135d4f98a2SYan Zheng del = 1; 14145d4f98a2SYan Zheng } 14155d4f98a2SYan Zheng 14165d4f98a2SYan Zheng __update_reloc_root(reloc_root, del); 14175d4f98a2SYan Zheng 14185d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 14195d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 14205d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 14215d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 14225d4f98a2SYan Zheng } 14235d4f98a2SYan Zheng 14245d4f98a2SYan Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 14255d4f98a2SYan Zheng &reloc_root->root_key, root_item); 14265d4f98a2SYan Zheng BUG_ON(ret); 14277585717fSChris Mason 14287585717fSChris Mason out: 14295d4f98a2SYan Zheng return 0; 14305d4f98a2SYan Zheng } 14315d4f98a2SYan Zheng 14325d4f98a2SYan Zheng /* 14335d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 14345d4f98a2SYan Zheng * in a subvolume 14355d4f98a2SYan Zheng */ 14365d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 14375d4f98a2SYan Zheng { 14385d4f98a2SYan Zheng struct rb_node *node; 14395d4f98a2SYan Zheng struct rb_node *prev; 14405d4f98a2SYan Zheng struct btrfs_inode *entry; 14415d4f98a2SYan Zheng struct inode *inode; 14425d4f98a2SYan Zheng 14435d4f98a2SYan Zheng spin_lock(&root->inode_lock); 14445d4f98a2SYan Zheng again: 14455d4f98a2SYan Zheng node = root->inode_tree.rb_node; 14465d4f98a2SYan Zheng prev = NULL; 14475d4f98a2SYan Zheng while (node) { 14485d4f98a2SYan Zheng prev = node; 14495d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 14505d4f98a2SYan Zheng 145133345d01SLi Zefan if (objectid < btrfs_ino(&entry->vfs_inode)) 14525d4f98a2SYan Zheng node = node->rb_left; 145333345d01SLi Zefan else if (objectid > btrfs_ino(&entry->vfs_inode)) 14545d4f98a2SYan Zheng node = node->rb_right; 14555d4f98a2SYan Zheng else 14565d4f98a2SYan Zheng break; 14575d4f98a2SYan Zheng } 14585d4f98a2SYan Zheng if (!node) { 14595d4f98a2SYan Zheng while (prev) { 14605d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 146133345d01SLi Zefan if (objectid <= btrfs_ino(&entry->vfs_inode)) { 14625d4f98a2SYan Zheng node = prev; 14635d4f98a2SYan Zheng break; 14645d4f98a2SYan Zheng } 14655d4f98a2SYan Zheng prev = rb_next(prev); 14665d4f98a2SYan Zheng } 14675d4f98a2SYan Zheng } 14685d4f98a2SYan Zheng while (node) { 14695d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 14705d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 14715d4f98a2SYan Zheng if (inode) { 14725d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 14735d4f98a2SYan Zheng return inode; 14745d4f98a2SYan Zheng } 14755d4f98a2SYan Zheng 147633345d01SLi Zefan objectid = btrfs_ino(&entry->vfs_inode) + 1; 14775d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 14785d4f98a2SYan Zheng goto again; 14795d4f98a2SYan Zheng 14805d4f98a2SYan Zheng node = rb_next(node); 14815d4f98a2SYan Zheng } 14825d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 14835d4f98a2SYan Zheng return NULL; 14845d4f98a2SYan Zheng } 14855d4f98a2SYan Zheng 14865d4f98a2SYan Zheng static int in_block_group(u64 bytenr, 14875d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group) 14885d4f98a2SYan Zheng { 14895d4f98a2SYan Zheng if (bytenr >= block_group->key.objectid && 14905d4f98a2SYan Zheng bytenr < block_group->key.objectid + block_group->key.offset) 14915d4f98a2SYan Zheng return 1; 14925d4f98a2SYan Zheng return 0; 14935d4f98a2SYan Zheng } 14945d4f98a2SYan Zheng 14955d4f98a2SYan Zheng /* 14965d4f98a2SYan Zheng * get new location of data 14975d4f98a2SYan Zheng */ 14985d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 14995d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 15005d4f98a2SYan Zheng { 15015d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 15025d4f98a2SYan Zheng struct btrfs_path *path; 15035d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 15045d4f98a2SYan Zheng struct extent_buffer *leaf; 15055d4f98a2SYan Zheng int ret; 15065d4f98a2SYan Zheng 15075d4f98a2SYan Zheng path = btrfs_alloc_path(); 15085d4f98a2SYan Zheng if (!path) 15095d4f98a2SYan Zheng return -ENOMEM; 15105d4f98a2SYan Zheng 15115d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 151233345d01SLi Zefan ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode), 15135d4f98a2SYan Zheng bytenr, 0); 15145d4f98a2SYan Zheng if (ret < 0) 15155d4f98a2SYan Zheng goto out; 15165d4f98a2SYan Zheng if (ret > 0) { 15175d4f98a2SYan Zheng ret = -ENOENT; 15185d4f98a2SYan Zheng goto out; 15195d4f98a2SYan Zheng } 15205d4f98a2SYan Zheng 15215d4f98a2SYan Zheng leaf = path->nodes[0]; 15225d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 15235d4f98a2SYan Zheng struct btrfs_file_extent_item); 15245d4f98a2SYan Zheng 15255d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 15265d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 15275d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 15285d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 15295d4f98a2SYan Zheng 15305d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 15315d4f98a2SYan Zheng ret = 1; 15325d4f98a2SYan Zheng goto out; 15335d4f98a2SYan Zheng } 15345d4f98a2SYan Zheng 15355d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 15365d4f98a2SYan Zheng ret = 0; 15375d4f98a2SYan Zheng out: 15385d4f98a2SYan Zheng btrfs_free_path(path); 15395d4f98a2SYan Zheng return ret; 15405d4f98a2SYan Zheng } 15415d4f98a2SYan Zheng 15425d4f98a2SYan Zheng /* 15435d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 15445d4f98a2SYan Zheng * the new locations. 15455d4f98a2SYan Zheng */ 15463fd0a558SYan, Zheng static noinline_for_stack 15473fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 15485d4f98a2SYan Zheng struct reloc_control *rc, 15495d4f98a2SYan Zheng struct btrfs_root *root, 15503fd0a558SYan, Zheng struct extent_buffer *leaf) 15515d4f98a2SYan Zheng { 15525d4f98a2SYan Zheng struct btrfs_key key; 15535d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 15545d4f98a2SYan Zheng struct inode *inode = NULL; 15555d4f98a2SYan Zheng u64 parent; 15565d4f98a2SYan Zheng u64 bytenr; 15573fd0a558SYan, Zheng u64 new_bytenr = 0; 15585d4f98a2SYan Zheng u64 num_bytes; 15595d4f98a2SYan Zheng u64 end; 15605d4f98a2SYan Zheng u32 nritems; 15615d4f98a2SYan Zheng u32 i; 15625d4f98a2SYan Zheng int ret; 15635d4f98a2SYan Zheng int first = 1; 15645d4f98a2SYan Zheng int dirty = 0; 15655d4f98a2SYan Zheng 15665d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 15675d4f98a2SYan Zheng return 0; 15685d4f98a2SYan Zheng 15695d4f98a2SYan Zheng /* reloc trees always use full backref */ 15705d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 15715d4f98a2SYan Zheng parent = leaf->start; 15725d4f98a2SYan Zheng else 15735d4f98a2SYan Zheng parent = 0; 15745d4f98a2SYan Zheng 15755d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 15765d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 15775d4f98a2SYan Zheng cond_resched(); 15785d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 15795d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 15805d4f98a2SYan Zheng continue; 15815d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 15825d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 15835d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 15845d4f98a2SYan Zheng continue; 15855d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 15865d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 15875d4f98a2SYan Zheng if (bytenr == 0) 15885d4f98a2SYan Zheng continue; 15895d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group)) 15905d4f98a2SYan Zheng continue; 15915d4f98a2SYan Zheng 15925d4f98a2SYan Zheng /* 15935d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 15945d4f98a2SYan Zheng * to complete and drop the extent cache 15955d4f98a2SYan Zheng */ 15965d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 15975d4f98a2SYan Zheng if (first) { 15985d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 15995d4f98a2SYan Zheng first = 0; 160033345d01SLi Zefan } else if (inode && btrfs_ino(inode) < key.objectid) { 16013fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 16025d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 16035d4f98a2SYan Zheng } 160433345d01SLi Zefan if (inode && btrfs_ino(inode) == key.objectid) { 16055d4f98a2SYan Zheng end = key.offset + 16065d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 16075d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 16085d4f98a2SYan Zheng root->sectorsize)); 16095d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 16105d4f98a2SYan Zheng end--; 16115d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1612d0082371SJeff Mahoney key.offset, end); 16135d4f98a2SYan Zheng if (!ret) 16145d4f98a2SYan Zheng continue; 16155d4f98a2SYan Zheng 16165d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, key.offset, end, 16175d4f98a2SYan Zheng 1); 16185d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1619d0082371SJeff Mahoney key.offset, end); 16205d4f98a2SYan Zheng } 16215d4f98a2SYan Zheng } 16225d4f98a2SYan Zheng 16235d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 16245d4f98a2SYan Zheng bytenr, num_bytes); 16253fd0a558SYan, Zheng if (ret > 0) { 16263fd0a558SYan, Zheng WARN_ON(1); 16275d4f98a2SYan Zheng continue; 16283fd0a558SYan, Zheng } 16295d4f98a2SYan Zheng BUG_ON(ret < 0); 16305d4f98a2SYan Zheng 16315d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 16325d4f98a2SYan Zheng dirty = 1; 16335d4f98a2SYan Zheng 16345d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 16355d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, new_bytenr, 16365d4f98a2SYan Zheng num_bytes, parent, 16375d4f98a2SYan Zheng btrfs_header_owner(leaf), 163866d7e7f0SArne Jansen key.objectid, key.offset, 1); 16395d4f98a2SYan Zheng BUG_ON(ret); 16405d4f98a2SYan Zheng 16415d4f98a2SYan Zheng ret = btrfs_free_extent(trans, root, bytenr, num_bytes, 16425d4f98a2SYan Zheng parent, btrfs_header_owner(leaf), 164366d7e7f0SArne Jansen key.objectid, key.offset, 1); 16445d4f98a2SYan Zheng BUG_ON(ret); 16455d4f98a2SYan Zheng } 16465d4f98a2SYan Zheng if (dirty) 16475d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 16483fd0a558SYan, Zheng if (inode) 16493fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 16505d4f98a2SYan Zheng return 0; 16515d4f98a2SYan Zheng } 16525d4f98a2SYan Zheng 16535d4f98a2SYan Zheng static noinline_for_stack 16545d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 16555d4f98a2SYan Zheng struct btrfs_path *path, int level) 16565d4f98a2SYan Zheng { 16575d4f98a2SYan Zheng struct btrfs_disk_key key1; 16585d4f98a2SYan Zheng struct btrfs_disk_key key2; 16595d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 16605d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 16615d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 16625d4f98a2SYan Zheng } 16635d4f98a2SYan Zheng 16645d4f98a2SYan Zheng /* 16655d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 16665d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 16675d4f98a2SYan Zheng * reloc tree was create can be replaced. 16685d4f98a2SYan Zheng * 16695d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 16705d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 16715d4f98a2SYan Zheng * errors, a negative error number is returned. 16725d4f98a2SYan Zheng */ 16733fd0a558SYan, Zheng static noinline_for_stack 16743fd0a558SYan, Zheng int replace_path(struct btrfs_trans_handle *trans, 16755d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 16765d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 16775d4f98a2SYan Zheng int lowest_level, int max_level) 16785d4f98a2SYan Zheng { 16795d4f98a2SYan Zheng struct extent_buffer *eb; 16805d4f98a2SYan Zheng struct extent_buffer *parent; 16815d4f98a2SYan Zheng struct btrfs_key key; 16825d4f98a2SYan Zheng u64 old_bytenr; 16835d4f98a2SYan Zheng u64 new_bytenr; 16845d4f98a2SYan Zheng u64 old_ptr_gen; 16855d4f98a2SYan Zheng u64 new_ptr_gen; 16865d4f98a2SYan Zheng u64 last_snapshot; 16875d4f98a2SYan Zheng u32 blocksize; 16883fd0a558SYan, Zheng int cow = 0; 16895d4f98a2SYan Zheng int level; 16905d4f98a2SYan Zheng int ret; 16915d4f98a2SYan Zheng int slot; 16925d4f98a2SYan Zheng 16935d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 16945d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 16955d4f98a2SYan Zheng 16965d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 16973fd0a558SYan, Zheng again: 16985d4f98a2SYan Zheng slot = path->slots[lowest_level]; 16995d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 17005d4f98a2SYan Zheng 17015d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 17025d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 17035d4f98a2SYan Zheng level = btrfs_header_level(eb); 17045d4f98a2SYan Zheng 17055d4f98a2SYan Zheng if (level < lowest_level) { 17065d4f98a2SYan Zheng btrfs_tree_unlock(eb); 17075d4f98a2SYan Zheng free_extent_buffer(eb); 17085d4f98a2SYan Zheng return 0; 17095d4f98a2SYan Zheng } 17105d4f98a2SYan Zheng 17113fd0a558SYan, Zheng if (cow) { 17125d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 17135d4f98a2SYan Zheng BUG_ON(ret); 17143fd0a558SYan, Zheng } 17155d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 17165d4f98a2SYan Zheng 17175d4f98a2SYan Zheng if (next_key) { 17185d4f98a2SYan Zheng next_key->objectid = (u64)-1; 17195d4f98a2SYan Zheng next_key->type = (u8)-1; 17205d4f98a2SYan Zheng next_key->offset = (u64)-1; 17215d4f98a2SYan Zheng } 17225d4f98a2SYan Zheng 17235d4f98a2SYan Zheng parent = eb; 17245d4f98a2SYan Zheng while (1) { 17255d4f98a2SYan Zheng level = btrfs_header_level(parent); 17265d4f98a2SYan Zheng BUG_ON(level < lowest_level); 17275d4f98a2SYan Zheng 17285d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 17295d4f98a2SYan Zheng if (ret && slot > 0) 17305d4f98a2SYan Zheng slot--; 17315d4f98a2SYan Zheng 17325d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 17335d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 17345d4f98a2SYan Zheng 17355d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 17365d4f98a2SYan Zheng blocksize = btrfs_level_size(dest, level - 1); 17375d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 17385d4f98a2SYan Zheng 17395d4f98a2SYan Zheng if (level <= max_level) { 17405d4f98a2SYan Zheng eb = path->nodes[level]; 17415d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 17425d4f98a2SYan Zheng path->slots[level]); 17435d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 17445d4f98a2SYan Zheng path->slots[level]); 17455d4f98a2SYan Zheng } else { 17465d4f98a2SYan Zheng new_bytenr = 0; 17475d4f98a2SYan Zheng new_ptr_gen = 0; 17485d4f98a2SYan Zheng } 17495d4f98a2SYan Zheng 17505d4f98a2SYan Zheng if (new_bytenr > 0 && new_bytenr == old_bytenr) { 17515d4f98a2SYan Zheng WARN_ON(1); 17525d4f98a2SYan Zheng ret = level; 17535d4f98a2SYan Zheng break; 17545d4f98a2SYan Zheng } 17555d4f98a2SYan Zheng 17565d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 17575d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 17583fd0a558SYan, Zheng if (level <= lowest_level) { 17595d4f98a2SYan Zheng ret = 0; 17605d4f98a2SYan Zheng break; 17615d4f98a2SYan Zheng } 17625d4f98a2SYan Zheng 17635d4f98a2SYan Zheng eb = read_tree_block(dest, old_bytenr, blocksize, 17645d4f98a2SYan Zheng old_ptr_gen); 176597d9a8a4STsutomu Itoh BUG_ON(!eb); 17665d4f98a2SYan Zheng btrfs_tree_lock(eb); 17673fd0a558SYan, Zheng if (cow) { 17685d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 17695d4f98a2SYan Zheng slot, &eb); 17705d4f98a2SYan Zheng BUG_ON(ret); 17715d4f98a2SYan Zheng } 17723fd0a558SYan, Zheng btrfs_set_lock_blocking(eb); 17735d4f98a2SYan Zheng 17745d4f98a2SYan Zheng btrfs_tree_unlock(parent); 17755d4f98a2SYan Zheng free_extent_buffer(parent); 17765d4f98a2SYan Zheng 17775d4f98a2SYan Zheng parent = eb; 17785d4f98a2SYan Zheng continue; 17795d4f98a2SYan Zheng } 17805d4f98a2SYan Zheng 17813fd0a558SYan, Zheng if (!cow) { 17823fd0a558SYan, Zheng btrfs_tree_unlock(parent); 17833fd0a558SYan, Zheng free_extent_buffer(parent); 17843fd0a558SYan, Zheng cow = 1; 17853fd0a558SYan, Zheng goto again; 17863fd0a558SYan, Zheng } 17873fd0a558SYan, Zheng 17885d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17895d4f98a2SYan Zheng path->slots[level]); 1790b3b4aa74SDavid Sterba btrfs_release_path(path); 17915d4f98a2SYan Zheng 17925d4f98a2SYan Zheng path->lowest_level = level; 17935d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 17945d4f98a2SYan Zheng path->lowest_level = 0; 17955d4f98a2SYan Zheng BUG_ON(ret); 17965d4f98a2SYan Zheng 17975d4f98a2SYan Zheng /* 17985d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 17995d4f98a2SYan Zheng */ 18005d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 18015d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 18025d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 18035d4f98a2SYan Zheng 18045d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 18055d4f98a2SYan Zheng path->slots[level], old_bytenr); 18065d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 18075d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 18085d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 18095d4f98a2SYan Zheng 18105d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize, 18115d4f98a2SYan Zheng path->nodes[level]->start, 181266d7e7f0SArne Jansen src->root_key.objectid, level - 1, 0, 181366d7e7f0SArne Jansen 1); 18145d4f98a2SYan Zheng BUG_ON(ret); 18155d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize, 18165d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 181766d7e7f0SArne Jansen 0, 1); 18185d4f98a2SYan Zheng BUG_ON(ret); 18195d4f98a2SYan Zheng 18205d4f98a2SYan Zheng ret = btrfs_free_extent(trans, src, new_bytenr, blocksize, 18215d4f98a2SYan Zheng path->nodes[level]->start, 182266d7e7f0SArne Jansen src->root_key.objectid, level - 1, 0, 182366d7e7f0SArne Jansen 1); 18245d4f98a2SYan Zheng BUG_ON(ret); 18255d4f98a2SYan Zheng 18265d4f98a2SYan Zheng ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize, 18275d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 182866d7e7f0SArne Jansen 0, 1); 18295d4f98a2SYan Zheng BUG_ON(ret); 18305d4f98a2SYan Zheng 18315d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 18325d4f98a2SYan Zheng 18335d4f98a2SYan Zheng ret = level; 18345d4f98a2SYan Zheng break; 18355d4f98a2SYan Zheng } 18365d4f98a2SYan Zheng btrfs_tree_unlock(parent); 18375d4f98a2SYan Zheng free_extent_buffer(parent); 18385d4f98a2SYan Zheng return ret; 18395d4f98a2SYan Zheng } 18405d4f98a2SYan Zheng 18415d4f98a2SYan Zheng /* 18425d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 18435d4f98a2SYan Zheng */ 18445d4f98a2SYan Zheng static noinline_for_stack 18455d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 18465d4f98a2SYan Zheng int *level) 18475d4f98a2SYan Zheng { 18485d4f98a2SYan Zheng struct extent_buffer *eb; 18495d4f98a2SYan Zheng int i; 18505d4f98a2SYan Zheng u64 last_snapshot; 18515d4f98a2SYan Zheng u32 nritems; 18525d4f98a2SYan Zheng 18535d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 18545d4f98a2SYan Zheng 18555d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 18565d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 18575d4f98a2SYan Zheng path->nodes[i] = NULL; 18585d4f98a2SYan Zheng } 18595d4f98a2SYan Zheng 18605d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 18615d4f98a2SYan Zheng eb = path->nodes[i]; 18625d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 18635d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 18645d4f98a2SYan Zheng path->slots[i]++; 18655d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 18665d4f98a2SYan Zheng last_snapshot) 18675d4f98a2SYan Zheng continue; 18685d4f98a2SYan Zheng 18695d4f98a2SYan Zheng *level = i; 18705d4f98a2SYan Zheng return 0; 18715d4f98a2SYan Zheng } 18725d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 18735d4f98a2SYan Zheng path->nodes[i] = NULL; 18745d4f98a2SYan Zheng } 18755d4f98a2SYan Zheng return 1; 18765d4f98a2SYan Zheng } 18775d4f98a2SYan Zheng 18785d4f98a2SYan Zheng /* 18795d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 18805d4f98a2SYan Zheng */ 18815d4f98a2SYan Zheng static noinline_for_stack 18825d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 18835d4f98a2SYan Zheng int *level) 18845d4f98a2SYan Zheng { 18855d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 18865d4f98a2SYan Zheng int i; 18875d4f98a2SYan Zheng u64 bytenr; 18885d4f98a2SYan Zheng u64 ptr_gen = 0; 18895d4f98a2SYan Zheng u64 last_snapshot; 18905d4f98a2SYan Zheng u32 blocksize; 18915d4f98a2SYan Zheng u32 nritems; 18925d4f98a2SYan Zheng 18935d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 18945d4f98a2SYan Zheng 18955d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 18965d4f98a2SYan Zheng eb = path->nodes[i]; 18975d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 18985d4f98a2SYan Zheng while (path->slots[i] < nritems) { 18995d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 19005d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 19015d4f98a2SYan Zheng break; 19025d4f98a2SYan Zheng path->slots[i]++; 19035d4f98a2SYan Zheng } 19045d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 19055d4f98a2SYan Zheng if (i == *level) 19065d4f98a2SYan Zheng break; 19075d4f98a2SYan Zheng *level = i + 1; 19085d4f98a2SYan Zheng return 0; 19095d4f98a2SYan Zheng } 19105d4f98a2SYan Zheng if (i == 1) { 19115d4f98a2SYan Zheng *level = i; 19125d4f98a2SYan Zheng return 0; 19135d4f98a2SYan Zheng } 19145d4f98a2SYan Zheng 19155d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 19165d4f98a2SYan Zheng blocksize = btrfs_level_size(root, i - 1); 19175d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, ptr_gen); 19185d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 19195d4f98a2SYan Zheng path->nodes[i - 1] = eb; 19205d4f98a2SYan Zheng path->slots[i - 1] = 0; 19215d4f98a2SYan Zheng } 19225d4f98a2SYan Zheng return 1; 19235d4f98a2SYan Zheng } 19245d4f98a2SYan Zheng 19255d4f98a2SYan Zheng /* 19265d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 19275d4f98a2SYan Zheng * [min_key, max_key) 19285d4f98a2SYan Zheng */ 19295d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 19305d4f98a2SYan Zheng struct btrfs_key *min_key, 19315d4f98a2SYan Zheng struct btrfs_key *max_key) 19325d4f98a2SYan Zheng { 19335d4f98a2SYan Zheng struct inode *inode = NULL; 19345d4f98a2SYan Zheng u64 objectid; 19355d4f98a2SYan Zheng u64 start, end; 193633345d01SLi Zefan u64 ino; 19375d4f98a2SYan Zheng 19385d4f98a2SYan Zheng objectid = min_key->objectid; 19395d4f98a2SYan Zheng while (1) { 19405d4f98a2SYan Zheng cond_resched(); 19415d4f98a2SYan Zheng iput(inode); 19425d4f98a2SYan Zheng 19435d4f98a2SYan Zheng if (objectid > max_key->objectid) 19445d4f98a2SYan Zheng break; 19455d4f98a2SYan Zheng 19465d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 19475d4f98a2SYan Zheng if (!inode) 19485d4f98a2SYan Zheng break; 194933345d01SLi Zefan ino = btrfs_ino(inode); 19505d4f98a2SYan Zheng 195133345d01SLi Zefan if (ino > max_key->objectid) { 19525d4f98a2SYan Zheng iput(inode); 19535d4f98a2SYan Zheng break; 19545d4f98a2SYan Zheng } 19555d4f98a2SYan Zheng 195633345d01SLi Zefan objectid = ino + 1; 19575d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 19585d4f98a2SYan Zheng continue; 19595d4f98a2SYan Zheng 196033345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 19615d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 19625d4f98a2SYan Zheng continue; 19635d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 19645d4f98a2SYan Zheng start = 0; 19655d4f98a2SYan Zheng else { 19665d4f98a2SYan Zheng start = min_key->offset; 19675d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(start, root->sectorsize)); 19685d4f98a2SYan Zheng } 19695d4f98a2SYan Zheng } else { 19705d4f98a2SYan Zheng start = 0; 19715d4f98a2SYan Zheng } 19725d4f98a2SYan Zheng 197333345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 19745d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 19755d4f98a2SYan Zheng continue; 19765d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 19775d4f98a2SYan Zheng end = (u64)-1; 19785d4f98a2SYan Zheng } else { 19795d4f98a2SYan Zheng if (max_key->offset == 0) 19805d4f98a2SYan Zheng continue; 19815d4f98a2SYan Zheng end = max_key->offset; 19825d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 19835d4f98a2SYan Zheng end--; 19845d4f98a2SYan Zheng } 19855d4f98a2SYan Zheng } else { 19865d4f98a2SYan Zheng end = (u64)-1; 19875d4f98a2SYan Zheng } 19885d4f98a2SYan Zheng 19895d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 1990d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 19915d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 1); 1992d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 19935d4f98a2SYan Zheng } 19945d4f98a2SYan Zheng return 0; 19955d4f98a2SYan Zheng } 19965d4f98a2SYan Zheng 19975d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 19985d4f98a2SYan Zheng struct btrfs_key *key) 19995d4f98a2SYan Zheng 20005d4f98a2SYan Zheng { 20015d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 20025d4f98a2SYan Zheng if (!path->nodes[level]) 20035d4f98a2SYan Zheng break; 20045d4f98a2SYan Zheng if (path->slots[level] + 1 < 20055d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 20065d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 20075d4f98a2SYan Zheng path->slots[level] + 1); 20085d4f98a2SYan Zheng return 0; 20095d4f98a2SYan Zheng } 20105d4f98a2SYan Zheng level++; 20115d4f98a2SYan Zheng } 20125d4f98a2SYan Zheng return 1; 20135d4f98a2SYan Zheng } 20145d4f98a2SYan Zheng 20155d4f98a2SYan Zheng /* 20165d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 20175d4f98a2SYan Zheng * fs tree. 20185d4f98a2SYan Zheng */ 20195d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 20205d4f98a2SYan Zheng struct btrfs_root *root) 20215d4f98a2SYan Zheng { 20225d4f98a2SYan Zheng LIST_HEAD(inode_list); 20235d4f98a2SYan Zheng struct btrfs_key key; 20245d4f98a2SYan Zheng struct btrfs_key next_key; 20255d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 20265d4f98a2SYan Zheng struct btrfs_root *reloc_root; 20275d4f98a2SYan Zheng struct btrfs_root_item *root_item; 20285d4f98a2SYan Zheng struct btrfs_path *path; 20293fd0a558SYan, Zheng struct extent_buffer *leaf; 20305d4f98a2SYan Zheng int level; 20315d4f98a2SYan Zheng int max_level; 20325d4f98a2SYan Zheng int replaced = 0; 20335d4f98a2SYan Zheng int ret; 20345d4f98a2SYan Zheng int err = 0; 20353fd0a558SYan, Zheng u32 min_reserved; 20365d4f98a2SYan Zheng 20375d4f98a2SYan Zheng path = btrfs_alloc_path(); 20385d4f98a2SYan Zheng if (!path) 20395d4f98a2SYan Zheng return -ENOMEM; 2040026fd317SJosef Bacik path->reada = 1; 20415d4f98a2SYan Zheng 20425d4f98a2SYan Zheng reloc_root = root->reloc_root; 20435d4f98a2SYan Zheng root_item = &reloc_root->root_item; 20445d4f98a2SYan Zheng 20455d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 20465d4f98a2SYan Zheng level = btrfs_root_level(root_item); 20475d4f98a2SYan Zheng extent_buffer_get(reloc_root->node); 20485d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 20495d4f98a2SYan Zheng path->slots[level] = 0; 20505d4f98a2SYan Zheng } else { 20515d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 20525d4f98a2SYan Zheng 20535d4f98a2SYan Zheng level = root_item->drop_level; 20545d4f98a2SYan Zheng BUG_ON(level == 0); 20555d4f98a2SYan Zheng path->lowest_level = level; 20565d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 205733c66f43SYan Zheng path->lowest_level = 0; 20585d4f98a2SYan Zheng if (ret < 0) { 20595d4f98a2SYan Zheng btrfs_free_path(path); 20605d4f98a2SYan Zheng return ret; 20615d4f98a2SYan Zheng } 20625d4f98a2SYan Zheng 20635d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 20645d4f98a2SYan Zheng path->slots[level]); 20655d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 20665d4f98a2SYan Zheng 20675d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 20685d4f98a2SYan Zheng } 20695d4f98a2SYan Zheng 20703fd0a558SYan, Zheng min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 20715d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 20725d4f98a2SYan Zheng 20735d4f98a2SYan Zheng while (1) { 2074a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 0); 207598d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 20763fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 20773fd0a558SYan, Zheng 207808e007d2SMiao Xie ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved, 207908e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 20803fd0a558SYan, Zheng if (ret) { 20813fd0a558SYan, Zheng BUG_ON(ret != -EAGAIN); 20823fd0a558SYan, Zheng ret = btrfs_commit_transaction(trans, root); 20833fd0a558SYan, Zheng BUG_ON(ret); 20843fd0a558SYan, Zheng continue; 20853fd0a558SYan, Zheng } 20863fd0a558SYan, Zheng 20873fd0a558SYan, Zheng replaced = 0; 20885d4f98a2SYan Zheng max_level = level; 20895d4f98a2SYan Zheng 20905d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 20915d4f98a2SYan Zheng if (ret < 0) { 20925d4f98a2SYan Zheng err = ret; 20935d4f98a2SYan Zheng goto out; 20945d4f98a2SYan Zheng } 20955d4f98a2SYan Zheng if (ret > 0) 20965d4f98a2SYan Zheng break; 20975d4f98a2SYan Zheng 20985d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 20995d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 21005d4f98a2SYan Zheng ret = 0; 21015d4f98a2SYan Zheng } else { 21023fd0a558SYan, Zheng ret = replace_path(trans, root, reloc_root, path, 21033fd0a558SYan, Zheng &next_key, level, max_level); 21045d4f98a2SYan Zheng } 21055d4f98a2SYan Zheng if (ret < 0) { 21065d4f98a2SYan Zheng err = ret; 21075d4f98a2SYan Zheng goto out; 21085d4f98a2SYan Zheng } 21095d4f98a2SYan Zheng 21105d4f98a2SYan Zheng if (ret > 0) { 21115d4f98a2SYan Zheng level = ret; 21125d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 21135d4f98a2SYan Zheng path->slots[level]); 21145d4f98a2SYan Zheng replaced = 1; 21155d4f98a2SYan Zheng } 21165d4f98a2SYan Zheng 21175d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 21185d4f98a2SYan Zheng if (ret > 0) 21195d4f98a2SYan Zheng break; 21205d4f98a2SYan Zheng 21215d4f98a2SYan Zheng BUG_ON(level == 0); 21225d4f98a2SYan Zheng /* 21235d4f98a2SYan Zheng * save the merging progress in the drop_progress. 21245d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 21255d4f98a2SYan Zheng */ 21265d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 21275d4f98a2SYan Zheng path->slots[level]); 21285d4f98a2SYan Zheng root_item->drop_level = level; 21295d4f98a2SYan Zheng 21303fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, root); 21315d4f98a2SYan Zheng 2132b53d3f5dSLiu Bo btrfs_btree_balance_dirty(root); 21335d4f98a2SYan Zheng 21345d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 21355d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 21365d4f98a2SYan Zheng } 21375d4f98a2SYan Zheng 21385d4f98a2SYan Zheng /* 21395d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 21405d4f98a2SYan Zheng * relocated and the block is tree root. 21415d4f98a2SYan Zheng */ 21425d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 21435d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 21445d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 21455d4f98a2SYan Zheng free_extent_buffer(leaf); 21465d4f98a2SYan Zheng if (ret < 0) 21475d4f98a2SYan Zheng err = ret; 21485d4f98a2SYan Zheng out: 21495d4f98a2SYan Zheng btrfs_free_path(path); 21505d4f98a2SYan Zheng 21515d4f98a2SYan Zheng if (err == 0) { 21525d4f98a2SYan Zheng memset(&root_item->drop_progress, 0, 21535d4f98a2SYan Zheng sizeof(root_item->drop_progress)); 21545d4f98a2SYan Zheng root_item->drop_level = 0; 21555d4f98a2SYan Zheng btrfs_set_root_refs(root_item, 0); 21563fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 21575d4f98a2SYan Zheng } 21585d4f98a2SYan Zheng 21593fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, root); 21605d4f98a2SYan Zheng 2161b53d3f5dSLiu Bo btrfs_btree_balance_dirty(root); 21625d4f98a2SYan Zheng 21635d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 21645d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 21655d4f98a2SYan Zheng 21665d4f98a2SYan Zheng return err; 21675d4f98a2SYan Zheng } 21685d4f98a2SYan Zheng 21693fd0a558SYan, Zheng static noinline_for_stack 21703fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 21715d4f98a2SYan Zheng { 21723fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 21733fd0a558SYan, Zheng struct btrfs_root *reloc_root; 21745d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 21753fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 21763fd0a558SYan, Zheng u64 num_bytes = 0; 21773fd0a558SYan, Zheng int ret; 21783fd0a558SYan, Zheng 21797585717fSChris Mason mutex_lock(&root->fs_info->reloc_mutex); 21803fd0a558SYan, Zheng rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 21813fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 21827585717fSChris Mason mutex_unlock(&root->fs_info->reloc_mutex); 21837585717fSChris Mason 21843fd0a558SYan, Zheng again: 21853fd0a558SYan, Zheng if (!err) { 21863fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 218708e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 218808e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 21893fd0a558SYan, Zheng if (ret) 21903fd0a558SYan, Zheng err = ret; 21913fd0a558SYan, Zheng } 21923fd0a558SYan, Zheng 21937a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 21943612b495STsutomu Itoh if (IS_ERR(trans)) { 21953612b495STsutomu Itoh if (!err) 21963612b495STsutomu Itoh btrfs_block_rsv_release(rc->extent_root, 21973612b495STsutomu Itoh rc->block_rsv, num_bytes); 21983612b495STsutomu Itoh return PTR_ERR(trans); 21993612b495STsutomu Itoh } 22003fd0a558SYan, Zheng 22013fd0a558SYan, Zheng if (!err) { 22023fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 22033fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 22043fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, 22053fd0a558SYan, Zheng rc->block_rsv, num_bytes); 22063fd0a558SYan, Zheng goto again; 22073fd0a558SYan, Zheng } 22083fd0a558SYan, Zheng } 22093fd0a558SYan, Zheng 22103fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 22113fd0a558SYan, Zheng 22123fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 22133fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 22143fd0a558SYan, Zheng struct btrfs_root, root_list); 22153fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 22163fd0a558SYan, Zheng 22173fd0a558SYan, Zheng root = read_fs_root(reloc_root->fs_info, 22183fd0a558SYan, Zheng reloc_root->root_key.offset); 22193fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 22203fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 22213fd0a558SYan, Zheng 22223fd0a558SYan, Zheng /* 22233fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 22243fd0a558SYan, Zheng * knows it should resumes merging 22253fd0a558SYan, Zheng */ 22263fd0a558SYan, Zheng if (!err) 22273fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 22283fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 22293fd0a558SYan, Zheng 22303fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 22313fd0a558SYan, Zheng } 22323fd0a558SYan, Zheng 22333fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 22343fd0a558SYan, Zheng 22353fd0a558SYan, Zheng if (!err) 22363fd0a558SYan, Zheng btrfs_commit_transaction(trans, rc->extent_root); 22373fd0a558SYan, Zheng else 22383fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 22393fd0a558SYan, Zheng return err; 22403fd0a558SYan, Zheng } 22413fd0a558SYan, Zheng 22423fd0a558SYan, Zheng static noinline_for_stack 2243aca1bba6SLiu Bo void free_reloc_roots(struct list_head *list) 2244aca1bba6SLiu Bo { 2245aca1bba6SLiu Bo struct btrfs_root *reloc_root; 2246aca1bba6SLiu Bo 2247aca1bba6SLiu Bo while (!list_empty(list)) { 2248aca1bba6SLiu Bo reloc_root = list_entry(list->next, struct btrfs_root, 2249aca1bba6SLiu Bo root_list); 2250aca1bba6SLiu Bo __update_reloc_root(reloc_root, 1); 2251aca1bba6SLiu Bo free_extent_buffer(reloc_root->node); 2252aca1bba6SLiu Bo free_extent_buffer(reloc_root->commit_root); 2253aca1bba6SLiu Bo kfree(reloc_root); 2254aca1bba6SLiu Bo } 2255aca1bba6SLiu Bo } 2256aca1bba6SLiu Bo 2257aca1bba6SLiu Bo static noinline_for_stack 22583fd0a558SYan, Zheng int merge_reloc_roots(struct reloc_control *rc) 22593fd0a558SYan, Zheng { 22605d4f98a2SYan Zheng struct btrfs_root *root; 22615d4f98a2SYan Zheng struct btrfs_root *reloc_root; 22623fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 22633fd0a558SYan, Zheng int found = 0; 2264aca1bba6SLiu Bo int ret = 0; 22653fd0a558SYan, Zheng again: 22663fd0a558SYan, Zheng root = rc->extent_root; 22677585717fSChris Mason 22687585717fSChris Mason /* 22697585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 22707585717fSChris Mason * we have to make sure nobody is in the middle of 22717585717fSChris Mason * adding their roots to the list while we are 22727585717fSChris Mason * doing this splice 22737585717fSChris Mason */ 22747585717fSChris Mason mutex_lock(&root->fs_info->reloc_mutex); 22753fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 22767585717fSChris Mason mutex_unlock(&root->fs_info->reloc_mutex); 22775d4f98a2SYan Zheng 22783fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 22793fd0a558SYan, Zheng found = 1; 22803fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 22813fd0a558SYan, Zheng struct btrfs_root, root_list); 22825d4f98a2SYan Zheng 22835d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 22845d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, 22855d4f98a2SYan Zheng reloc_root->root_key.offset); 22865d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 22875d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 22885d4f98a2SYan Zheng 22893fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 2290aca1bba6SLiu Bo if (ret) 2291aca1bba6SLiu Bo goto out; 22923fd0a558SYan, Zheng } else { 22933fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 22943fd0a558SYan, Zheng } 22952c536799SJeff Mahoney ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1); 2296aca1bba6SLiu Bo if (ret < 0) { 2297aca1bba6SLiu Bo if (list_empty(&reloc_root->root_list)) 2298aca1bba6SLiu Bo list_add_tail(&reloc_root->root_list, 2299aca1bba6SLiu Bo &reloc_roots); 2300aca1bba6SLiu Bo goto out; 2301aca1bba6SLiu Bo } 23025d4f98a2SYan Zheng } 23035d4f98a2SYan Zheng 23043fd0a558SYan, Zheng if (found) { 23053fd0a558SYan, Zheng found = 0; 23063fd0a558SYan, Zheng goto again; 23075d4f98a2SYan Zheng } 2308aca1bba6SLiu Bo out: 2309aca1bba6SLiu Bo if (ret) { 2310aca1bba6SLiu Bo btrfs_std_error(root->fs_info, ret); 2311aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 2312aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 2313aca1bba6SLiu Bo } 2314aca1bba6SLiu Bo 23155d4f98a2SYan Zheng BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 2316aca1bba6SLiu Bo return ret; 23175d4f98a2SYan Zheng } 23185d4f98a2SYan Zheng 23195d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 23205d4f98a2SYan Zheng { 23215d4f98a2SYan Zheng struct tree_block *block; 23225d4f98a2SYan Zheng struct rb_node *rb_node; 23235d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 23245d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 23255d4f98a2SYan Zheng rb_erase(rb_node, blocks); 23265d4f98a2SYan Zheng kfree(block); 23275d4f98a2SYan Zheng } 23285d4f98a2SYan Zheng } 23295d4f98a2SYan Zheng 23305d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 23315d4f98a2SYan Zheng struct btrfs_root *reloc_root) 23325d4f98a2SYan Zheng { 23335d4f98a2SYan Zheng struct btrfs_root *root; 23345d4f98a2SYan Zheng 23355d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 23365d4f98a2SYan Zheng return 0; 23375d4f98a2SYan Zheng 23385d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset); 23395d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 23405d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 23415d4f98a2SYan Zheng 23425d4f98a2SYan Zheng return btrfs_record_root_in_trans(trans, root); 23435d4f98a2SYan Zheng } 23445d4f98a2SYan Zheng 23453fd0a558SYan, Zheng static noinline_for_stack 23463fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 23473fd0a558SYan, Zheng struct reloc_control *rc, 23485d4f98a2SYan Zheng struct backref_node *node, 23493fd0a558SYan, Zheng struct backref_edge *edges[], int *nr) 23505d4f98a2SYan Zheng { 23515d4f98a2SYan Zheng struct backref_node *next; 23525d4f98a2SYan Zheng struct btrfs_root *root; 23533fd0a558SYan, Zheng int index = 0; 23543fd0a558SYan, Zheng 23555d4f98a2SYan Zheng next = node; 23565d4f98a2SYan Zheng while (1) { 23575d4f98a2SYan Zheng cond_resched(); 23585d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 23595d4f98a2SYan Zheng root = next->root; 23603fd0a558SYan, Zheng BUG_ON(!root); 23613fd0a558SYan, Zheng BUG_ON(!root->ref_cows); 23625d4f98a2SYan Zheng 23635d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 23645d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 23655d4f98a2SYan Zheng break; 23665d4f98a2SYan Zheng } 23675d4f98a2SYan Zheng 23685d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 23693fd0a558SYan, Zheng root = root->reloc_root; 23703fd0a558SYan, Zheng 23713fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 23723fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 23733fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 23743fd0a558SYan, Zheng next->new_bytenr = root->node->start; 23753fd0a558SYan, Zheng next->root = root; 23763fd0a558SYan, Zheng list_add_tail(&next->list, 23773fd0a558SYan, Zheng &rc->backref_cache.changed); 23783fd0a558SYan, Zheng __mark_block_processed(rc, next); 23795d4f98a2SYan Zheng break; 23805d4f98a2SYan Zheng } 23815d4f98a2SYan Zheng 23823fd0a558SYan, Zheng WARN_ON(1); 23835d4f98a2SYan Zheng root = NULL; 23845d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 23855d4f98a2SYan Zheng if (!next || next->level <= node->level) 23865d4f98a2SYan Zheng break; 23875d4f98a2SYan Zheng } 23883fd0a558SYan, Zheng if (!root) 23893fd0a558SYan, Zheng return NULL; 23905d4f98a2SYan Zheng 23915d4f98a2SYan Zheng *nr = index; 23923fd0a558SYan, Zheng next = node; 23933fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 23943fd0a558SYan, Zheng while (1) { 23953fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 23963fd0a558SYan, Zheng if (--index < 0) 23973fd0a558SYan, Zheng break; 23983fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 23993fd0a558SYan, Zheng } 24005d4f98a2SYan Zheng return root; 24015d4f98a2SYan Zheng } 24025d4f98a2SYan Zheng 24033fd0a558SYan, Zheng /* 24043fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 24053fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 24063fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 24073fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 24083fd0a558SYan, Zheng */ 24095d4f98a2SYan Zheng static noinline_for_stack 24105d4f98a2SYan Zheng struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans, 24115d4f98a2SYan Zheng struct backref_node *node) 24125d4f98a2SYan Zheng { 24133fd0a558SYan, Zheng struct backref_node *next; 24143fd0a558SYan, Zheng struct btrfs_root *root; 24153fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 24165d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24173fd0a558SYan, Zheng int index = 0; 24183fd0a558SYan, Zheng 24193fd0a558SYan, Zheng next = node; 24203fd0a558SYan, Zheng while (1) { 24213fd0a558SYan, Zheng cond_resched(); 24223fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 24233fd0a558SYan, Zheng root = next->root; 24243fd0a558SYan, Zheng BUG_ON(!root); 24253fd0a558SYan, Zheng 242625985edcSLucas De Marchi /* no other choice for non-references counted tree */ 24273fd0a558SYan, Zheng if (!root->ref_cows) 24283fd0a558SYan, Zheng return root; 24293fd0a558SYan, Zheng 24303fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 24313fd0a558SYan, Zheng fs_root = root; 24323fd0a558SYan, Zheng 24333fd0a558SYan, Zheng if (next != node) 24343fd0a558SYan, Zheng return NULL; 24353fd0a558SYan, Zheng 24363fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 24373fd0a558SYan, Zheng if (!next || next->level <= node->level) 24383fd0a558SYan, Zheng break; 24393fd0a558SYan, Zheng } 24403fd0a558SYan, Zheng 24413fd0a558SYan, Zheng if (!fs_root) 24423fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 24433fd0a558SYan, Zheng return fs_root; 24445d4f98a2SYan Zheng } 24455d4f98a2SYan Zheng 24465d4f98a2SYan Zheng static noinline_for_stack 24473fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 24483fd0a558SYan, Zheng struct backref_node *node, int reserve) 24495d4f98a2SYan Zheng { 24503fd0a558SYan, Zheng struct backref_node *next = node; 24513fd0a558SYan, Zheng struct backref_edge *edge; 24523fd0a558SYan, Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24533fd0a558SYan, Zheng u64 num_bytes = 0; 24543fd0a558SYan, Zheng int index = 0; 24555d4f98a2SYan Zheng 24563fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 24573fd0a558SYan, Zheng 24583fd0a558SYan, Zheng while (next) { 24593fd0a558SYan, Zheng cond_resched(); 24605d4f98a2SYan Zheng while (1) { 24613fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 24625d4f98a2SYan Zheng break; 24635d4f98a2SYan Zheng 24643fd0a558SYan, Zheng num_bytes += btrfs_level_size(rc->extent_root, 24653fd0a558SYan, Zheng next->level); 24663fd0a558SYan, Zheng 24673fd0a558SYan, Zheng if (list_empty(&next->upper)) 24683fd0a558SYan, Zheng break; 24693fd0a558SYan, Zheng 24703fd0a558SYan, Zheng edge = list_entry(next->upper.next, 24713fd0a558SYan, Zheng struct backref_edge, list[LOWER]); 24723fd0a558SYan, Zheng edges[index++] = edge; 24733fd0a558SYan, Zheng next = edge->node[UPPER]; 24745d4f98a2SYan Zheng } 24753fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 24763fd0a558SYan, Zheng } 24773fd0a558SYan, Zheng return num_bytes; 24783fd0a558SYan, Zheng } 24793fd0a558SYan, Zheng 24803fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 24813fd0a558SYan, Zheng struct reloc_control *rc, 24823fd0a558SYan, Zheng struct backref_node *node) 24833fd0a558SYan, Zheng { 24843fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 24853fd0a558SYan, Zheng u64 num_bytes; 24863fd0a558SYan, Zheng int ret; 24873fd0a558SYan, Zheng 24883fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 24893fd0a558SYan, Zheng 24903fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 249108e007d2SMiao Xie ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes, 249208e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 24933fd0a558SYan, Zheng if (ret) { 24943fd0a558SYan, Zheng if (ret == -EAGAIN) 24953fd0a558SYan, Zheng rc->commit_transaction = 1; 24963fd0a558SYan, Zheng return ret; 24973fd0a558SYan, Zheng } 24983fd0a558SYan, Zheng 24993fd0a558SYan, Zheng return 0; 25003fd0a558SYan, Zheng } 25013fd0a558SYan, Zheng 25023fd0a558SYan, Zheng static void release_metadata_space(struct reloc_control *rc, 25033fd0a558SYan, Zheng struct backref_node *node) 25043fd0a558SYan, Zheng { 25053fd0a558SYan, Zheng u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2; 25063fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes); 25075d4f98a2SYan Zheng } 25085d4f98a2SYan Zheng 25095d4f98a2SYan Zheng /* 25105d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 25115d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 25125d4f98a2SYan Zheng * 25135d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 25145d4f98a2SYan Zheng * in that case this function just updates pointers. 25155d4f98a2SYan Zheng */ 25165d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 25173fd0a558SYan, Zheng struct reloc_control *rc, 25185d4f98a2SYan Zheng struct backref_node *node, 25195d4f98a2SYan Zheng struct btrfs_key *key, 25205d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 25215d4f98a2SYan Zheng { 25225d4f98a2SYan Zheng struct backref_node *upper; 25235d4f98a2SYan Zheng struct backref_edge *edge; 25245d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 25255d4f98a2SYan Zheng struct btrfs_root *root; 25265d4f98a2SYan Zheng struct extent_buffer *eb; 25275d4f98a2SYan Zheng u32 blocksize; 25285d4f98a2SYan Zheng u64 bytenr; 25295d4f98a2SYan Zheng u64 generation; 25305d4f98a2SYan Zheng int nr; 25315d4f98a2SYan Zheng int slot; 25325d4f98a2SYan Zheng int ret; 25335d4f98a2SYan Zheng int err = 0; 25345d4f98a2SYan Zheng 25355d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 25365d4f98a2SYan Zheng 25375d4f98a2SYan Zheng path->lowest_level = node->level + 1; 25383fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 25395d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 25405d4f98a2SYan Zheng cond_resched(); 25415d4f98a2SYan Zheng 25425d4f98a2SYan Zheng upper = edge->node[UPPER]; 25433fd0a558SYan, Zheng root = select_reloc_root(trans, rc, upper, edges, &nr); 25443fd0a558SYan, Zheng BUG_ON(!root); 25455d4f98a2SYan Zheng 25463fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 25473fd0a558SYan, Zheng if (!lowest) { 25483fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 25493fd0a558SYan, Zheng upper->level, &slot); 25503fd0a558SYan, Zheng BUG_ON(ret); 25513fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 25523fd0a558SYan, Zheng if (node->eb->start == bytenr) 25533fd0a558SYan, Zheng goto next; 25543fd0a558SYan, Zheng } 25555d4f98a2SYan Zheng drop_node_buffer(upper); 25563fd0a558SYan, Zheng } 25575d4f98a2SYan Zheng 25585d4f98a2SYan Zheng if (!upper->eb) { 25595d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 25605d4f98a2SYan Zheng if (ret < 0) { 25615d4f98a2SYan Zheng err = ret; 25625d4f98a2SYan Zheng break; 25635d4f98a2SYan Zheng } 25645d4f98a2SYan Zheng BUG_ON(ret > 0); 25655d4f98a2SYan Zheng 25663fd0a558SYan, Zheng if (!upper->eb) { 25673fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 25683fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 25693fd0a558SYan, Zheng } else { 25703fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 25713fd0a558SYan, Zheng } 25723fd0a558SYan, Zheng 25733fd0a558SYan, Zheng upper->locked = 1; 25743fd0a558SYan, Zheng path->locks[upper->level] = 0; 25753fd0a558SYan, Zheng 25765d4f98a2SYan Zheng slot = path->slots[upper->level]; 2577b3b4aa74SDavid Sterba btrfs_release_path(path); 25785d4f98a2SYan Zheng } else { 25795d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 25805d4f98a2SYan Zheng &slot); 25815d4f98a2SYan Zheng BUG_ON(ret); 25825d4f98a2SYan Zheng } 25835d4f98a2SYan Zheng 25845d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 25853fd0a558SYan, Zheng if (lowest) { 25863fd0a558SYan, Zheng BUG_ON(bytenr != node->bytenr); 25875d4f98a2SYan Zheng } else { 25883fd0a558SYan, Zheng if (node->eb->start == bytenr) 25893fd0a558SYan, Zheng goto next; 25905d4f98a2SYan Zheng } 25915d4f98a2SYan Zheng 25925d4f98a2SYan Zheng blocksize = btrfs_level_size(root, node->level); 25935d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 25945d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, generation); 259597d9a8a4STsutomu Itoh if (!eb) { 259697d9a8a4STsutomu Itoh err = -EIO; 259797d9a8a4STsutomu Itoh goto next; 259897d9a8a4STsutomu Itoh } 25995d4f98a2SYan Zheng btrfs_tree_lock(eb); 26005d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 26015d4f98a2SYan Zheng 26025d4f98a2SYan Zheng if (!node->eb) { 26035d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 26045d4f98a2SYan Zheng slot, &eb); 26053fd0a558SYan, Zheng btrfs_tree_unlock(eb); 26063fd0a558SYan, Zheng free_extent_buffer(eb); 26075d4f98a2SYan Zheng if (ret < 0) { 26085d4f98a2SYan Zheng err = ret; 26093fd0a558SYan, Zheng goto next; 26105d4f98a2SYan Zheng } 26113fd0a558SYan, Zheng BUG_ON(node->eb != eb); 26125d4f98a2SYan Zheng } else { 26135d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 26145d4f98a2SYan Zheng node->eb->start); 26155d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 26165d4f98a2SYan Zheng trans->transid); 26175d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 26185d4f98a2SYan Zheng 26195d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, 26205d4f98a2SYan Zheng node->eb->start, blocksize, 26215d4f98a2SYan Zheng upper->eb->start, 26225d4f98a2SYan Zheng btrfs_header_owner(upper->eb), 262366d7e7f0SArne Jansen node->level, 0, 1); 26245d4f98a2SYan Zheng BUG_ON(ret); 26255d4f98a2SYan Zheng 26265d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 26275d4f98a2SYan Zheng BUG_ON(ret); 26285d4f98a2SYan Zheng } 26293fd0a558SYan, Zheng next: 26303fd0a558SYan, Zheng if (!upper->pending) 26313fd0a558SYan, Zheng drop_node_buffer(upper); 26323fd0a558SYan, Zheng else 26333fd0a558SYan, Zheng unlock_node_buffer(upper); 26343fd0a558SYan, Zheng if (err) 26353fd0a558SYan, Zheng break; 26365d4f98a2SYan Zheng } 26373fd0a558SYan, Zheng 26383fd0a558SYan, Zheng if (!err && node->pending) { 26393fd0a558SYan, Zheng drop_node_buffer(node); 26403fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 26413fd0a558SYan, Zheng node->pending = 0; 26425d4f98a2SYan Zheng } 26433fd0a558SYan, Zheng 26445d4f98a2SYan Zheng path->lowest_level = 0; 26453fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 26465d4f98a2SYan Zheng return err; 26475d4f98a2SYan Zheng } 26485d4f98a2SYan Zheng 26495d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 26503fd0a558SYan, Zheng struct reloc_control *rc, 26515d4f98a2SYan Zheng struct backref_node *node, 26525d4f98a2SYan Zheng struct btrfs_path *path) 26535d4f98a2SYan Zheng { 26545d4f98a2SYan Zheng struct btrfs_key key; 26555d4f98a2SYan Zheng 26565d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 26573fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 26585d4f98a2SYan Zheng } 26595d4f98a2SYan Zheng 26605d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 26613fd0a558SYan, Zheng struct reloc_control *rc, 26623fd0a558SYan, Zheng struct btrfs_path *path, int err) 26635d4f98a2SYan Zheng { 26643fd0a558SYan, Zheng LIST_HEAD(list); 26653fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 26665d4f98a2SYan Zheng struct backref_node *node; 26675d4f98a2SYan Zheng int level; 26685d4f98a2SYan Zheng int ret; 26695d4f98a2SYan Zheng 26705d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 26715d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 26725d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 26733fd0a558SYan, Zheng struct backref_node, list); 26743fd0a558SYan, Zheng list_move_tail(&node->list, &list); 26753fd0a558SYan, Zheng BUG_ON(!node->pending); 26765d4f98a2SYan Zheng 26773fd0a558SYan, Zheng if (!err) { 26783fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 26795d4f98a2SYan Zheng if (ret < 0) 26805d4f98a2SYan Zheng err = ret; 26815d4f98a2SYan Zheng } 26825d4f98a2SYan Zheng } 26833fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 26843fd0a558SYan, Zheng } 26855d4f98a2SYan Zheng return err; 26865d4f98a2SYan Zheng } 26875d4f98a2SYan Zheng 26885d4f98a2SYan Zheng static void mark_block_processed(struct reloc_control *rc, 26893fd0a558SYan, Zheng u64 bytenr, u32 blocksize) 26903fd0a558SYan, Zheng { 26913fd0a558SYan, Zheng set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1, 26923fd0a558SYan, Zheng EXTENT_DIRTY, GFP_NOFS); 26933fd0a558SYan, Zheng } 26943fd0a558SYan, Zheng 26953fd0a558SYan, Zheng static void __mark_block_processed(struct reloc_control *rc, 26965d4f98a2SYan Zheng struct backref_node *node) 26975d4f98a2SYan Zheng { 26985d4f98a2SYan Zheng u32 blocksize; 26995d4f98a2SYan Zheng if (node->level == 0 || 27005d4f98a2SYan Zheng in_block_group(node->bytenr, rc->block_group)) { 27015d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, node->level); 27023fd0a558SYan, Zheng mark_block_processed(rc, node->bytenr, blocksize); 27035d4f98a2SYan Zheng } 27045d4f98a2SYan Zheng node->processed = 1; 27055d4f98a2SYan Zheng } 27065d4f98a2SYan Zheng 27075d4f98a2SYan Zheng /* 27085d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 27095d4f98a2SYan Zheng * as processed. 27105d4f98a2SYan Zheng */ 27115d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 27125d4f98a2SYan Zheng struct backref_node *node) 27135d4f98a2SYan Zheng { 27145d4f98a2SYan Zheng struct backref_node *next = node; 27155d4f98a2SYan Zheng struct backref_edge *edge; 27165d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 27175d4f98a2SYan Zheng int index = 0; 27185d4f98a2SYan Zheng 27195d4f98a2SYan Zheng while (next) { 27205d4f98a2SYan Zheng cond_resched(); 27215d4f98a2SYan Zheng while (1) { 27225d4f98a2SYan Zheng if (next->processed) 27235d4f98a2SYan Zheng break; 27245d4f98a2SYan Zheng 27253fd0a558SYan, Zheng __mark_block_processed(rc, next); 27265d4f98a2SYan Zheng 27275d4f98a2SYan Zheng if (list_empty(&next->upper)) 27285d4f98a2SYan Zheng break; 27295d4f98a2SYan Zheng 27305d4f98a2SYan Zheng edge = list_entry(next->upper.next, 27315d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 27325d4f98a2SYan Zheng edges[index++] = edge; 27335d4f98a2SYan Zheng next = edge->node[UPPER]; 27345d4f98a2SYan Zheng } 27355d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 27365d4f98a2SYan Zheng } 27375d4f98a2SYan Zheng } 27385d4f98a2SYan Zheng 27395d4f98a2SYan Zheng static int tree_block_processed(u64 bytenr, u32 blocksize, 27405d4f98a2SYan Zheng struct reloc_control *rc) 27415d4f98a2SYan Zheng { 27425d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 27439655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 27445d4f98a2SYan Zheng return 1; 27455d4f98a2SYan Zheng return 0; 27465d4f98a2SYan Zheng } 27475d4f98a2SYan Zheng 27485d4f98a2SYan Zheng static int get_tree_block_key(struct reloc_control *rc, 27495d4f98a2SYan Zheng struct tree_block *block) 27505d4f98a2SYan Zheng { 27515d4f98a2SYan Zheng struct extent_buffer *eb; 27525d4f98a2SYan Zheng 27535d4f98a2SYan Zheng BUG_ON(block->key_ready); 27545d4f98a2SYan Zheng eb = read_tree_block(rc->extent_root, block->bytenr, 27555d4f98a2SYan Zheng block->key.objectid, block->key.offset); 275697d9a8a4STsutomu Itoh BUG_ON(!eb); 27575d4f98a2SYan Zheng WARN_ON(btrfs_header_level(eb) != block->level); 27585d4f98a2SYan Zheng if (block->level == 0) 27595d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 27605d4f98a2SYan Zheng else 27615d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 27625d4f98a2SYan Zheng free_extent_buffer(eb); 27635d4f98a2SYan Zheng block->key_ready = 1; 27645d4f98a2SYan Zheng return 0; 27655d4f98a2SYan Zheng } 27665d4f98a2SYan Zheng 27675d4f98a2SYan Zheng static int reada_tree_block(struct reloc_control *rc, 27685d4f98a2SYan Zheng struct tree_block *block) 27695d4f98a2SYan Zheng { 27705d4f98a2SYan Zheng BUG_ON(block->key_ready); 27715d4f98a2SYan Zheng readahead_tree_block(rc->extent_root, block->bytenr, 27725d4f98a2SYan Zheng block->key.objectid, block->key.offset); 27735d4f98a2SYan Zheng return 0; 27745d4f98a2SYan Zheng } 27755d4f98a2SYan Zheng 27765d4f98a2SYan Zheng /* 27775d4f98a2SYan Zheng * helper function to relocate a tree block 27785d4f98a2SYan Zheng */ 27795d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 27805d4f98a2SYan Zheng struct reloc_control *rc, 27815d4f98a2SYan Zheng struct backref_node *node, 27825d4f98a2SYan Zheng struct btrfs_key *key, 27835d4f98a2SYan Zheng struct btrfs_path *path) 27845d4f98a2SYan Zheng { 27855d4f98a2SYan Zheng struct btrfs_root *root; 27863fd0a558SYan, Zheng int release = 0; 27873fd0a558SYan, Zheng int ret = 0; 27885d4f98a2SYan Zheng 27893fd0a558SYan, Zheng if (!node) 27905d4f98a2SYan Zheng return 0; 27913fd0a558SYan, Zheng 27923fd0a558SYan, Zheng BUG_ON(node->processed); 27933fd0a558SYan, Zheng root = select_one_root(trans, node); 27943fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 27953fd0a558SYan, Zheng update_processed_blocks(rc, node); 27963fd0a558SYan, Zheng goto out; 27975d4f98a2SYan Zheng } 27985d4f98a2SYan Zheng 27993fd0a558SYan, Zheng if (!root || root->ref_cows) { 28003fd0a558SYan, Zheng ret = reserve_metadata_space(trans, rc, node); 28013fd0a558SYan, Zheng if (ret) 28025d4f98a2SYan Zheng goto out; 28033fd0a558SYan, Zheng release = 1; 28045d4f98a2SYan Zheng } 28053fd0a558SYan, Zheng 28063fd0a558SYan, Zheng if (root) { 28073fd0a558SYan, Zheng if (root->ref_cows) { 28083fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 28093fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 28103fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 28113fd0a558SYan, Zheng root = root->reloc_root; 28123fd0a558SYan, Zheng node->new_bytenr = root->node->start; 28133fd0a558SYan, Zheng node->root = root; 28143fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 28153fd0a558SYan, Zheng } else { 28165d4f98a2SYan Zheng path->lowest_level = node->level; 28175d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2818b3b4aa74SDavid Sterba btrfs_release_path(path); 28193fd0a558SYan, Zheng if (ret > 0) 28205d4f98a2SYan Zheng ret = 0; 28213fd0a558SYan, Zheng } 28223fd0a558SYan, Zheng if (!ret) 28233fd0a558SYan, Zheng update_processed_blocks(rc, node); 28243fd0a558SYan, Zheng } else { 28253fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 28263fd0a558SYan, Zheng } 28275d4f98a2SYan Zheng out: 28283fd0a558SYan, Zheng if (ret || node->level == 0 || node->cowonly) { 28293fd0a558SYan, Zheng if (release) 28303fd0a558SYan, Zheng release_metadata_space(rc, node); 28313fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 28323fd0a558SYan, Zheng } 28335d4f98a2SYan Zheng return ret; 28345d4f98a2SYan Zheng } 28355d4f98a2SYan Zheng 28365d4f98a2SYan Zheng /* 28375d4f98a2SYan Zheng * relocate a list of blocks 28385d4f98a2SYan Zheng */ 28395d4f98a2SYan Zheng static noinline_for_stack 28405d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 28415d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 28425d4f98a2SYan Zheng { 28435d4f98a2SYan Zheng struct backref_node *node; 28445d4f98a2SYan Zheng struct btrfs_path *path; 28455d4f98a2SYan Zheng struct tree_block *block; 28465d4f98a2SYan Zheng struct rb_node *rb_node; 28475d4f98a2SYan Zheng int ret; 28485d4f98a2SYan Zheng int err = 0; 28495d4f98a2SYan Zheng 28505d4f98a2SYan Zheng path = btrfs_alloc_path(); 28515d4f98a2SYan Zheng if (!path) 28525d4f98a2SYan Zheng return -ENOMEM; 28535d4f98a2SYan Zheng 28545d4f98a2SYan Zheng rb_node = rb_first(blocks); 28555d4f98a2SYan Zheng while (rb_node) { 28565d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 28575d4f98a2SYan Zheng if (!block->key_ready) 28585d4f98a2SYan Zheng reada_tree_block(rc, block); 28595d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28605d4f98a2SYan Zheng } 28615d4f98a2SYan Zheng 28625d4f98a2SYan Zheng rb_node = rb_first(blocks); 28635d4f98a2SYan Zheng while (rb_node) { 28645d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 28655d4f98a2SYan Zheng if (!block->key_ready) 28665d4f98a2SYan Zheng get_tree_block_key(rc, block); 28675d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28685d4f98a2SYan Zheng } 28695d4f98a2SYan Zheng 28705d4f98a2SYan Zheng rb_node = rb_first(blocks); 28715d4f98a2SYan Zheng while (rb_node) { 28725d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 28735d4f98a2SYan Zheng 28743fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 28755d4f98a2SYan Zheng block->level, block->bytenr); 28765d4f98a2SYan Zheng if (IS_ERR(node)) { 28775d4f98a2SYan Zheng err = PTR_ERR(node); 28785d4f98a2SYan Zheng goto out; 28795d4f98a2SYan Zheng } 28805d4f98a2SYan Zheng 28815d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 28825d4f98a2SYan Zheng path); 28835d4f98a2SYan Zheng if (ret < 0) { 28843fd0a558SYan, Zheng if (ret != -EAGAIN || rb_node == rb_first(blocks)) 28855d4f98a2SYan Zheng err = ret; 28865d4f98a2SYan Zheng goto out; 28875d4f98a2SYan Zheng } 28885d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28895d4f98a2SYan Zheng } 28905d4f98a2SYan Zheng out: 28915d4f98a2SYan Zheng free_block_list(blocks); 28923fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 28935d4f98a2SYan Zheng 28945d4f98a2SYan Zheng btrfs_free_path(path); 28955d4f98a2SYan Zheng return err; 28965d4f98a2SYan Zheng } 28975d4f98a2SYan Zheng 28985d4f98a2SYan Zheng static noinline_for_stack 2899efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 2900efa56464SYan, Zheng struct file_extent_cluster *cluster) 2901efa56464SYan, Zheng { 2902efa56464SYan, Zheng u64 alloc_hint = 0; 2903efa56464SYan, Zheng u64 start; 2904efa56464SYan, Zheng u64 end; 2905efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 2906efa56464SYan, Zheng u64 num_bytes; 2907efa56464SYan, Zheng int nr = 0; 2908efa56464SYan, Zheng int ret = 0; 2909efa56464SYan, Zheng 2910efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2911efa56464SYan, Zheng mutex_lock(&inode->i_mutex); 2912efa56464SYan, Zheng 2913efa56464SYan, Zheng ret = btrfs_check_data_free_space(inode, cluster->end + 2914efa56464SYan, Zheng 1 - cluster->start); 2915efa56464SYan, Zheng if (ret) 2916efa56464SYan, Zheng goto out; 2917efa56464SYan, Zheng 2918efa56464SYan, Zheng while (nr < cluster->nr) { 2919efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2920efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2921efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2922efa56464SYan, Zheng else 2923efa56464SYan, Zheng end = cluster->end - offset; 2924efa56464SYan, Zheng 2925d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 2926efa56464SYan, Zheng num_bytes = end + 1 - start; 2927efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 2928efa56464SYan, Zheng num_bytes, num_bytes, 2929efa56464SYan, Zheng end + 1, &alloc_hint); 2930d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 2931efa56464SYan, Zheng if (ret) 2932efa56464SYan, Zheng break; 2933efa56464SYan, Zheng nr++; 2934efa56464SYan, Zheng } 2935efa56464SYan, Zheng btrfs_free_reserved_data_space(inode, cluster->end + 2936efa56464SYan, Zheng 1 - cluster->start); 2937efa56464SYan, Zheng out: 2938efa56464SYan, Zheng mutex_unlock(&inode->i_mutex); 2939efa56464SYan, Zheng return ret; 2940efa56464SYan, Zheng } 2941efa56464SYan, Zheng 2942efa56464SYan, Zheng static noinline_for_stack 29430257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 29440257bb82SYan, Zheng u64 block_start) 29455d4f98a2SYan Zheng { 29465d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 29475d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 29485d4f98a2SYan Zheng struct extent_map *em; 29490257bb82SYan, Zheng int ret = 0; 29505d4f98a2SYan Zheng 2951172ddd60SDavid Sterba em = alloc_extent_map(); 29520257bb82SYan, Zheng if (!em) 29530257bb82SYan, Zheng return -ENOMEM; 29540257bb82SYan, Zheng 29555d4f98a2SYan Zheng em->start = start; 29560257bb82SYan, Zheng em->len = end + 1 - start; 29570257bb82SYan, Zheng em->block_len = em->len; 29580257bb82SYan, Zheng em->block_start = block_start; 29595d4f98a2SYan Zheng em->bdev = root->fs_info->fs_devices->latest_bdev; 29605d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 29615d4f98a2SYan Zheng 2962d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 29635d4f98a2SYan Zheng while (1) { 2964890871beSChris Mason write_lock(&em_tree->lock); 29655d4f98a2SYan Zheng ret = add_extent_mapping(em_tree, em); 2966890871beSChris Mason write_unlock(&em_tree->lock); 29675d4f98a2SYan Zheng if (ret != -EEXIST) { 29685d4f98a2SYan Zheng free_extent_map(em); 29695d4f98a2SYan Zheng break; 29705d4f98a2SYan Zheng } 29715d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 0); 29725d4f98a2SYan Zheng } 2973d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 29740257bb82SYan, Zheng return ret; 29750257bb82SYan, Zheng } 29765d4f98a2SYan Zheng 29770257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 29780257bb82SYan, Zheng struct file_extent_cluster *cluster) 29790257bb82SYan, Zheng { 29800257bb82SYan, Zheng u64 page_start; 29810257bb82SYan, Zheng u64 page_end; 29820257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 29830257bb82SYan, Zheng unsigned long index; 29840257bb82SYan, Zheng unsigned long last_index; 29850257bb82SYan, Zheng struct page *page; 29860257bb82SYan, Zheng struct file_ra_state *ra; 29873b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 29880257bb82SYan, Zheng int nr = 0; 29890257bb82SYan, Zheng int ret = 0; 29900257bb82SYan, Zheng 29910257bb82SYan, Zheng if (!cluster->nr) 29920257bb82SYan, Zheng return 0; 29930257bb82SYan, Zheng 29940257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 29950257bb82SYan, Zheng if (!ra) 29960257bb82SYan, Zheng return -ENOMEM; 29970257bb82SYan, Zheng 2998efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 29990257bb82SYan, Zheng if (ret) 3000efa56464SYan, Zheng goto out; 30010257bb82SYan, Zheng 30020257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 30030257bb82SYan, Zheng 3004efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 3005efa56464SYan, Zheng cluster->end - offset, cluster->start); 3006efa56464SYan, Zheng if (ret) 3007efa56464SYan, Zheng goto out; 3008efa56464SYan, Zheng 3009efa56464SYan, Zheng index = (cluster->start - offset) >> PAGE_CACHE_SHIFT; 3010efa56464SYan, Zheng last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT; 30110257bb82SYan, Zheng while (index <= last_index) { 3012efa56464SYan, Zheng ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE); 3013efa56464SYan, Zheng if (ret) 3014efa56464SYan, Zheng goto out; 3015efa56464SYan, Zheng 30160257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 30170257bb82SYan, Zheng if (!page) { 30180257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 30190257bb82SYan, Zheng ra, NULL, index, 30200257bb82SYan, Zheng last_index + 1 - index); 3021a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 30223b16a4e3SJosef Bacik mask); 30230257bb82SYan, Zheng if (!page) { 3024efa56464SYan, Zheng btrfs_delalloc_release_metadata(inode, 3025efa56464SYan, Zheng PAGE_CACHE_SIZE); 30260257bb82SYan, Zheng ret = -ENOMEM; 3027efa56464SYan, Zheng goto out; 30280257bb82SYan, Zheng } 30290257bb82SYan, Zheng } 30300257bb82SYan, Zheng 30310257bb82SYan, Zheng if (PageReadahead(page)) { 30320257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 30330257bb82SYan, Zheng ra, NULL, page, index, 30340257bb82SYan, Zheng last_index + 1 - index); 30350257bb82SYan, Zheng } 30360257bb82SYan, Zheng 30370257bb82SYan, Zheng if (!PageUptodate(page)) { 30380257bb82SYan, Zheng btrfs_readpage(NULL, page); 30390257bb82SYan, Zheng lock_page(page); 30400257bb82SYan, Zheng if (!PageUptodate(page)) { 30410257bb82SYan, Zheng unlock_page(page); 30420257bb82SYan, Zheng page_cache_release(page); 3043efa56464SYan, Zheng btrfs_delalloc_release_metadata(inode, 3044efa56464SYan, Zheng PAGE_CACHE_SIZE); 30450257bb82SYan, Zheng ret = -EIO; 3046efa56464SYan, Zheng goto out; 30470257bb82SYan, Zheng } 30480257bb82SYan, Zheng } 30490257bb82SYan, Zheng 30504eee4fa4SMiao Xie page_start = page_offset(page); 30510257bb82SYan, Zheng page_end = page_start + PAGE_CACHE_SIZE - 1; 30520257bb82SYan, Zheng 3053d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 30540257bb82SYan, Zheng 30550257bb82SYan, Zheng set_page_extent_mapped(page); 30560257bb82SYan, Zheng 30570257bb82SYan, Zheng if (nr < cluster->nr && 30580257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 30590257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 30600257bb82SYan, Zheng page_start, page_end, 30610257bb82SYan, Zheng EXTENT_BOUNDARY, GFP_NOFS); 30620257bb82SYan, Zheng nr++; 30630257bb82SYan, Zheng } 30640257bb82SYan, Zheng 3065efa56464SYan, Zheng btrfs_set_extent_delalloc(inode, page_start, page_end, NULL); 30660257bb82SYan, Zheng set_page_dirty(page); 30670257bb82SYan, Zheng 30680257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 3069d0082371SJeff Mahoney page_start, page_end); 30700257bb82SYan, Zheng unlock_page(page); 30710257bb82SYan, Zheng page_cache_release(page); 30720257bb82SYan, Zheng 30730257bb82SYan, Zheng index++; 3074efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 3075efa56464SYan, Zheng btrfs_throttle(BTRFS_I(inode)->root); 30760257bb82SYan, Zheng } 30770257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 3078efa56464SYan, Zheng out: 30790257bb82SYan, Zheng kfree(ra); 30800257bb82SYan, Zheng return ret; 30810257bb82SYan, Zheng } 30820257bb82SYan, Zheng 30830257bb82SYan, Zheng static noinline_for_stack 30840257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 30850257bb82SYan, Zheng struct file_extent_cluster *cluster) 30860257bb82SYan, Zheng { 30870257bb82SYan, Zheng int ret; 30880257bb82SYan, Zheng 30890257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 30900257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30910257bb82SYan, Zheng if (ret) 30920257bb82SYan, Zheng return ret; 30930257bb82SYan, Zheng cluster->nr = 0; 30940257bb82SYan, Zheng } 30950257bb82SYan, Zheng 30960257bb82SYan, Zheng if (!cluster->nr) 30970257bb82SYan, Zheng cluster->start = extent_key->objectid; 30980257bb82SYan, Zheng else 30990257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 31000257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 31010257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 31020257bb82SYan, Zheng cluster->nr++; 31030257bb82SYan, Zheng 31040257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 31050257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 31060257bb82SYan, Zheng if (ret) 31070257bb82SYan, Zheng return ret; 31080257bb82SYan, Zheng cluster->nr = 0; 31090257bb82SYan, Zheng } 31100257bb82SYan, Zheng return 0; 31115d4f98a2SYan Zheng } 31125d4f98a2SYan Zheng 31135d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 31145d4f98a2SYan Zheng static int get_ref_objectid_v0(struct reloc_control *rc, 31155d4f98a2SYan Zheng struct btrfs_path *path, 31165d4f98a2SYan Zheng struct btrfs_key *extent_key, 31175d4f98a2SYan Zheng u64 *ref_objectid, int *path_change) 31185d4f98a2SYan Zheng { 31195d4f98a2SYan Zheng struct btrfs_key key; 31205d4f98a2SYan Zheng struct extent_buffer *leaf; 31215d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 31225d4f98a2SYan Zheng int ret; 31235d4f98a2SYan Zheng int slot; 31245d4f98a2SYan Zheng 31255d4f98a2SYan Zheng leaf = path->nodes[0]; 31265d4f98a2SYan Zheng slot = path->slots[0]; 31275d4f98a2SYan Zheng while (1) { 31285d4f98a2SYan Zheng if (slot >= btrfs_header_nritems(leaf)) { 31295d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 31305d4f98a2SYan Zheng if (ret < 0) 31315d4f98a2SYan Zheng return ret; 31325d4f98a2SYan Zheng BUG_ON(ret > 0); 31335d4f98a2SYan Zheng leaf = path->nodes[0]; 31345d4f98a2SYan Zheng slot = path->slots[0]; 31355d4f98a2SYan Zheng if (path_change) 31365d4f98a2SYan Zheng *path_change = 1; 31375d4f98a2SYan Zheng } 31385d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, slot); 31395d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 31405d4f98a2SYan Zheng return -ENOENT; 31415d4f98a2SYan Zheng 31425d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_REF_V0_KEY) { 31435d4f98a2SYan Zheng slot++; 31445d4f98a2SYan Zheng continue; 31455d4f98a2SYan Zheng } 31465d4f98a2SYan Zheng ref0 = btrfs_item_ptr(leaf, slot, 31475d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 31485d4f98a2SYan Zheng *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0); 31495d4f98a2SYan Zheng break; 31505d4f98a2SYan Zheng } 31515d4f98a2SYan Zheng return 0; 31525d4f98a2SYan Zheng } 31535d4f98a2SYan Zheng #endif 31545d4f98a2SYan Zheng 31555d4f98a2SYan Zheng /* 31565d4f98a2SYan Zheng * helper to add a tree block to the list. 31575d4f98a2SYan Zheng * the major work is getting the generation and level of the block 31585d4f98a2SYan Zheng */ 31595d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 31605d4f98a2SYan Zheng struct btrfs_key *extent_key, 31615d4f98a2SYan Zheng struct btrfs_path *path, 31625d4f98a2SYan Zheng struct rb_root *blocks) 31635d4f98a2SYan Zheng { 31645d4f98a2SYan Zheng struct extent_buffer *eb; 31655d4f98a2SYan Zheng struct btrfs_extent_item *ei; 31665d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 31675d4f98a2SYan Zheng struct tree_block *block; 31685d4f98a2SYan Zheng struct rb_node *rb_node; 31695d4f98a2SYan Zheng u32 item_size; 31705d4f98a2SYan Zheng int level = -1; 31715d4f98a2SYan Zheng int generation; 31725d4f98a2SYan Zheng 31735d4f98a2SYan Zheng eb = path->nodes[0]; 31745d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 31755d4f98a2SYan Zheng 31765d4f98a2SYan Zheng if (item_size >= sizeof(*ei) + sizeof(*bi)) { 31775d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 31785d4f98a2SYan Zheng struct btrfs_extent_item); 31795d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 31805d4f98a2SYan Zheng generation = btrfs_extent_generation(eb, ei); 31815d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 31825d4f98a2SYan Zheng } else { 31835d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 31845d4f98a2SYan Zheng u64 ref_owner; 31855d4f98a2SYan Zheng int ret; 31865d4f98a2SYan Zheng 31875d4f98a2SYan Zheng BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 31885d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, extent_key, 31895d4f98a2SYan Zheng &ref_owner, NULL); 3190411fc6bcSAndi Kleen if (ret < 0) 3191411fc6bcSAndi Kleen return ret; 31925d4f98a2SYan Zheng BUG_ON(ref_owner >= BTRFS_MAX_LEVEL); 31935d4f98a2SYan Zheng level = (int)ref_owner; 31945d4f98a2SYan Zheng /* FIXME: get real generation */ 31955d4f98a2SYan Zheng generation = 0; 31965d4f98a2SYan Zheng #else 31975d4f98a2SYan Zheng BUG(); 31985d4f98a2SYan Zheng #endif 31995d4f98a2SYan Zheng } 32005d4f98a2SYan Zheng 3201b3b4aa74SDavid Sterba btrfs_release_path(path); 32025d4f98a2SYan Zheng 32035d4f98a2SYan Zheng BUG_ON(level == -1); 32045d4f98a2SYan Zheng 32055d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 32065d4f98a2SYan Zheng if (!block) 32075d4f98a2SYan Zheng return -ENOMEM; 32085d4f98a2SYan Zheng 32095d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 32105d4f98a2SYan Zheng block->key.objectid = extent_key->offset; 32115d4f98a2SYan Zheng block->key.offset = generation; 32125d4f98a2SYan Zheng block->level = level; 32135d4f98a2SYan Zheng block->key_ready = 0; 32145d4f98a2SYan Zheng 32155d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 321643c04fb1SJeff Mahoney if (rb_node) 321743c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, block->bytenr); 32185d4f98a2SYan Zheng 32195d4f98a2SYan Zheng return 0; 32205d4f98a2SYan Zheng } 32215d4f98a2SYan Zheng 32225d4f98a2SYan Zheng /* 32235d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 32245d4f98a2SYan Zheng */ 32255d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 32265d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 32275d4f98a2SYan Zheng struct rb_root *blocks) 32285d4f98a2SYan Zheng { 32295d4f98a2SYan Zheng struct btrfs_path *path; 32305d4f98a2SYan Zheng struct btrfs_key key; 32315d4f98a2SYan Zheng int ret; 32325d4f98a2SYan Zheng 32335d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 32345d4f98a2SYan Zheng return 0; 32355d4f98a2SYan Zheng 32365d4f98a2SYan Zheng if (tree_search(blocks, bytenr)) 32375d4f98a2SYan Zheng return 0; 32385d4f98a2SYan Zheng 32395d4f98a2SYan Zheng path = btrfs_alloc_path(); 32405d4f98a2SYan Zheng if (!path) 32415d4f98a2SYan Zheng return -ENOMEM; 32425d4f98a2SYan Zheng 32435d4f98a2SYan Zheng key.objectid = bytenr; 32445d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 32455d4f98a2SYan Zheng key.offset = blocksize; 32465d4f98a2SYan Zheng 32475d4f98a2SYan Zheng path->search_commit_root = 1; 32485d4f98a2SYan Zheng path->skip_locking = 1; 32495d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 32505d4f98a2SYan Zheng if (ret < 0) 32515d4f98a2SYan Zheng goto out; 32525d4f98a2SYan Zheng BUG_ON(ret); 32535d4f98a2SYan Zheng 32545d4f98a2SYan Zheng btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 32555d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 32565d4f98a2SYan Zheng out: 32575d4f98a2SYan Zheng btrfs_free_path(path); 32585d4f98a2SYan Zheng return ret; 32595d4f98a2SYan Zheng } 32605d4f98a2SYan Zheng 32615d4f98a2SYan Zheng /* 32625d4f98a2SYan Zheng * helper to check if the block use full backrefs for pointers in it 32635d4f98a2SYan Zheng */ 32645d4f98a2SYan Zheng static int block_use_full_backref(struct reloc_control *rc, 32655d4f98a2SYan Zheng struct extent_buffer *eb) 32665d4f98a2SYan Zheng { 32675d4f98a2SYan Zheng u64 flags; 32685d4f98a2SYan Zheng int ret; 32695d4f98a2SYan Zheng 32705d4f98a2SYan Zheng if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) || 32715d4f98a2SYan Zheng btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV) 32725d4f98a2SYan Zheng return 1; 32735d4f98a2SYan Zheng 32743fd0a558SYan, Zheng ret = btrfs_lookup_extent_info(NULL, rc->extent_root, 32753fd0a558SYan, Zheng eb->start, eb->len, NULL, &flags); 32765d4f98a2SYan Zheng BUG_ON(ret); 32775d4f98a2SYan Zheng 32785d4f98a2SYan Zheng if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) 32795d4f98a2SYan Zheng ret = 1; 32805d4f98a2SYan Zheng else 32815d4f98a2SYan Zheng ret = 0; 32825d4f98a2SYan Zheng return ret; 32835d4f98a2SYan Zheng } 32845d4f98a2SYan Zheng 32850af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 32860af3d00bSJosef Bacik struct inode *inode, u64 ino) 32870af3d00bSJosef Bacik { 32880af3d00bSJosef Bacik struct btrfs_key key; 32890af3d00bSJosef Bacik struct btrfs_path *path; 32900af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 32910af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 32920af3d00bSJosef Bacik int ret = 0; 32930af3d00bSJosef Bacik 32940af3d00bSJosef Bacik if (inode) 32950af3d00bSJosef Bacik goto truncate; 32960af3d00bSJosef Bacik 32970af3d00bSJosef Bacik key.objectid = ino; 32980af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 32990af3d00bSJosef Bacik key.offset = 0; 33000af3d00bSJosef Bacik 33010af3d00bSJosef Bacik inode = btrfs_iget(fs_info->sb, &key, root, NULL); 3302f54fb859STsutomu Itoh if (IS_ERR(inode) || is_bad_inode(inode)) { 3303f54fb859STsutomu Itoh if (!IS_ERR(inode)) 33040af3d00bSJosef Bacik iput(inode); 33050af3d00bSJosef Bacik return -ENOENT; 33060af3d00bSJosef Bacik } 33070af3d00bSJosef Bacik 33080af3d00bSJosef Bacik truncate: 33090af3d00bSJosef Bacik path = btrfs_alloc_path(); 33100af3d00bSJosef Bacik if (!path) { 33110af3d00bSJosef Bacik ret = -ENOMEM; 33120af3d00bSJosef Bacik goto out; 33130af3d00bSJosef Bacik } 33140af3d00bSJosef Bacik 33157a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 33160af3d00bSJosef Bacik if (IS_ERR(trans)) { 33170af3d00bSJosef Bacik btrfs_free_path(path); 33183612b495STsutomu Itoh ret = PTR_ERR(trans); 33190af3d00bSJosef Bacik goto out; 33200af3d00bSJosef Bacik } 33210af3d00bSJosef Bacik 33220af3d00bSJosef Bacik ret = btrfs_truncate_free_space_cache(root, trans, path, inode); 33230af3d00bSJosef Bacik 33240af3d00bSJosef Bacik btrfs_free_path(path); 33250af3d00bSJosef Bacik btrfs_end_transaction(trans, root); 3326b53d3f5dSLiu Bo btrfs_btree_balance_dirty(root); 33270af3d00bSJosef Bacik out: 33280af3d00bSJosef Bacik iput(inode); 33290af3d00bSJosef Bacik return ret; 33300af3d00bSJosef Bacik } 33310af3d00bSJosef Bacik 33325d4f98a2SYan Zheng /* 33335d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY 33345d4f98a2SYan Zheng * this function scans fs tree to find blocks reference the data extent 33355d4f98a2SYan Zheng */ 33365d4f98a2SYan Zheng static int find_data_references(struct reloc_control *rc, 33375d4f98a2SYan Zheng struct btrfs_key *extent_key, 33385d4f98a2SYan Zheng struct extent_buffer *leaf, 33395d4f98a2SYan Zheng struct btrfs_extent_data_ref *ref, 33405d4f98a2SYan Zheng struct rb_root *blocks) 33415d4f98a2SYan Zheng { 33425d4f98a2SYan Zheng struct btrfs_path *path; 33435d4f98a2SYan Zheng struct tree_block *block; 33445d4f98a2SYan Zheng struct btrfs_root *root; 33455d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 33465d4f98a2SYan Zheng struct rb_node *rb_node; 33475d4f98a2SYan Zheng struct btrfs_key key; 33485d4f98a2SYan Zheng u64 ref_root; 33495d4f98a2SYan Zheng u64 ref_objectid; 33505d4f98a2SYan Zheng u64 ref_offset; 33515d4f98a2SYan Zheng u32 ref_count; 33525d4f98a2SYan Zheng u32 nritems; 33535d4f98a2SYan Zheng int err = 0; 33545d4f98a2SYan Zheng int added = 0; 33555d4f98a2SYan Zheng int counted; 33565d4f98a2SYan Zheng int ret; 33575d4f98a2SYan Zheng 33585d4f98a2SYan Zheng ref_root = btrfs_extent_data_ref_root(leaf, ref); 33595d4f98a2SYan Zheng ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref); 33605d4f98a2SYan Zheng ref_offset = btrfs_extent_data_ref_offset(leaf, ref); 33615d4f98a2SYan Zheng ref_count = btrfs_extent_data_ref_count(leaf, ref); 33625d4f98a2SYan Zheng 33630af3d00bSJosef Bacik /* 33640af3d00bSJosef Bacik * This is an extent belonging to the free space cache, lets just delete 33650af3d00bSJosef Bacik * it and redo the search. 33660af3d00bSJosef Bacik */ 33670af3d00bSJosef Bacik if (ref_root == BTRFS_ROOT_TREE_OBJECTID) { 33680af3d00bSJosef Bacik ret = delete_block_group_cache(rc->extent_root->fs_info, 33690af3d00bSJosef Bacik NULL, ref_objectid); 33700af3d00bSJosef Bacik if (ret != -ENOENT) 33710af3d00bSJosef Bacik return ret; 33720af3d00bSJosef Bacik ret = 0; 33730af3d00bSJosef Bacik } 33740af3d00bSJosef Bacik 33750af3d00bSJosef Bacik path = btrfs_alloc_path(); 33760af3d00bSJosef Bacik if (!path) 33770af3d00bSJosef Bacik return -ENOMEM; 3378026fd317SJosef Bacik path->reada = 1; 33790af3d00bSJosef Bacik 33805d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, ref_root); 33815d4f98a2SYan Zheng if (IS_ERR(root)) { 33825d4f98a2SYan Zheng err = PTR_ERR(root); 33835d4f98a2SYan Zheng goto out; 33845d4f98a2SYan Zheng } 33855d4f98a2SYan Zheng 33865d4f98a2SYan Zheng key.objectid = ref_objectid; 33875d4f98a2SYan Zheng key.type = BTRFS_EXTENT_DATA_KEY; 338884850e8dSYan, Zheng if (ref_offset > ((u64)-1 << 32)) 338984850e8dSYan, Zheng key.offset = 0; 339084850e8dSYan, Zheng else 339184850e8dSYan, Zheng key.offset = ref_offset; 33925d4f98a2SYan Zheng 33935d4f98a2SYan Zheng path->search_commit_root = 1; 33945d4f98a2SYan Zheng path->skip_locking = 1; 33955d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 33965d4f98a2SYan Zheng if (ret < 0) { 33975d4f98a2SYan Zheng err = ret; 33985d4f98a2SYan Zheng goto out; 33995d4f98a2SYan Zheng } 34005d4f98a2SYan Zheng 34015d4f98a2SYan Zheng leaf = path->nodes[0]; 34025d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 34035d4f98a2SYan Zheng /* 34045d4f98a2SYan Zheng * the references in tree blocks that use full backrefs 34055d4f98a2SYan Zheng * are not counted in 34065d4f98a2SYan Zheng */ 34075d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 34085d4f98a2SYan Zheng counted = 0; 34095d4f98a2SYan Zheng else 34105d4f98a2SYan Zheng counted = 1; 34115d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 34125d4f98a2SYan Zheng if (rb_node) { 34135d4f98a2SYan Zheng if (counted) 34145d4f98a2SYan Zheng added = 1; 34155d4f98a2SYan Zheng else 34165d4f98a2SYan Zheng path->slots[0] = nritems; 34175d4f98a2SYan Zheng } 34185d4f98a2SYan Zheng 34195d4f98a2SYan Zheng while (ref_count > 0) { 34205d4f98a2SYan Zheng while (path->slots[0] >= nritems) { 34215d4f98a2SYan Zheng ret = btrfs_next_leaf(root, path); 34225d4f98a2SYan Zheng if (ret < 0) { 34235d4f98a2SYan Zheng err = ret; 34245d4f98a2SYan Zheng goto out; 34255d4f98a2SYan Zheng } 34265d4f98a2SYan Zheng if (ret > 0) { 34275d4f98a2SYan Zheng WARN_ON(1); 34285d4f98a2SYan Zheng goto out; 34295d4f98a2SYan Zheng } 34305d4f98a2SYan Zheng 34315d4f98a2SYan Zheng leaf = path->nodes[0]; 34325d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 34335d4f98a2SYan Zheng added = 0; 34345d4f98a2SYan Zheng 34355d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 34365d4f98a2SYan Zheng counted = 0; 34375d4f98a2SYan Zheng else 34385d4f98a2SYan Zheng counted = 1; 34395d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 34405d4f98a2SYan Zheng if (rb_node) { 34415d4f98a2SYan Zheng if (counted) 34425d4f98a2SYan Zheng added = 1; 34435d4f98a2SYan Zheng else 34445d4f98a2SYan Zheng path->slots[0] = nritems; 34455d4f98a2SYan Zheng } 34465d4f98a2SYan Zheng } 34475d4f98a2SYan Zheng 34485d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 34495d4f98a2SYan Zheng if (key.objectid != ref_objectid || 34505d4f98a2SYan Zheng key.type != BTRFS_EXTENT_DATA_KEY) { 34515d4f98a2SYan Zheng WARN_ON(1); 34525d4f98a2SYan Zheng break; 34535d4f98a2SYan Zheng } 34545d4f98a2SYan Zheng 34555d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 34565d4f98a2SYan Zheng struct btrfs_file_extent_item); 34575d4f98a2SYan Zheng 34585d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 34595d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 34605d4f98a2SYan Zheng goto next; 34615d4f98a2SYan Zheng 34625d4f98a2SYan Zheng if (btrfs_file_extent_disk_bytenr(leaf, fi) != 34635d4f98a2SYan Zheng extent_key->objectid) 34645d4f98a2SYan Zheng goto next; 34655d4f98a2SYan Zheng 34665d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 34675d4f98a2SYan Zheng if (key.offset != ref_offset) 34685d4f98a2SYan Zheng goto next; 34695d4f98a2SYan Zheng 34705d4f98a2SYan Zheng if (counted) 34715d4f98a2SYan Zheng ref_count--; 34725d4f98a2SYan Zheng if (added) 34735d4f98a2SYan Zheng goto next; 34745d4f98a2SYan Zheng 34755d4f98a2SYan Zheng if (!tree_block_processed(leaf->start, leaf->len, rc)) { 34765d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 34775d4f98a2SYan Zheng if (!block) { 34785d4f98a2SYan Zheng err = -ENOMEM; 34795d4f98a2SYan Zheng break; 34805d4f98a2SYan Zheng } 34815d4f98a2SYan Zheng block->bytenr = leaf->start; 34825d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &block->key, 0); 34835d4f98a2SYan Zheng block->level = 0; 34845d4f98a2SYan Zheng block->key_ready = 1; 34855d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, 34865d4f98a2SYan Zheng &block->rb_node); 348743c04fb1SJeff Mahoney if (rb_node) 348843c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, 348943c04fb1SJeff Mahoney block->bytenr); 34905d4f98a2SYan Zheng } 34915d4f98a2SYan Zheng if (counted) 34925d4f98a2SYan Zheng added = 1; 34935d4f98a2SYan Zheng else 34945d4f98a2SYan Zheng path->slots[0] = nritems; 34955d4f98a2SYan Zheng next: 34965d4f98a2SYan Zheng path->slots[0]++; 34975d4f98a2SYan Zheng 34985d4f98a2SYan Zheng } 34995d4f98a2SYan Zheng out: 35005d4f98a2SYan Zheng btrfs_free_path(path); 35015d4f98a2SYan Zheng return err; 35025d4f98a2SYan Zheng } 35035d4f98a2SYan Zheng 35045d4f98a2SYan Zheng /* 35055d4f98a2SYan Zheng * hepler to find all tree blocks that reference a given data extent 35065d4f98a2SYan Zheng */ 35075d4f98a2SYan Zheng static noinline_for_stack 35085d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 35095d4f98a2SYan Zheng struct btrfs_key *extent_key, 35105d4f98a2SYan Zheng struct btrfs_path *path, 35115d4f98a2SYan Zheng struct rb_root *blocks) 35125d4f98a2SYan Zheng { 35135d4f98a2SYan Zheng struct btrfs_key key; 35145d4f98a2SYan Zheng struct extent_buffer *eb; 35155d4f98a2SYan Zheng struct btrfs_extent_data_ref *dref; 35165d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 35175d4f98a2SYan Zheng unsigned long ptr; 35185d4f98a2SYan Zheng unsigned long end; 35193fd0a558SYan, Zheng u32 blocksize = btrfs_level_size(rc->extent_root, 0); 35205d4f98a2SYan Zheng int ret; 35215d4f98a2SYan Zheng int err = 0; 35225d4f98a2SYan Zheng 35235d4f98a2SYan Zheng eb = path->nodes[0]; 35245d4f98a2SYan Zheng ptr = btrfs_item_ptr_offset(eb, path->slots[0]); 35255d4f98a2SYan Zheng end = ptr + btrfs_item_size_nr(eb, path->slots[0]); 35265d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 35275d4f98a2SYan Zheng if (ptr + sizeof(struct btrfs_extent_item_v0) == end) 35285d4f98a2SYan Zheng ptr = end; 35295d4f98a2SYan Zheng else 35305d4f98a2SYan Zheng #endif 35315d4f98a2SYan Zheng ptr += sizeof(struct btrfs_extent_item); 35325d4f98a2SYan Zheng 35335d4f98a2SYan Zheng while (ptr < end) { 35345d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 35355d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 35365d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 35375d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 35385d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 35395d4f98a2SYan Zheng blocks); 35405d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 35415d4f98a2SYan Zheng dref = (struct btrfs_extent_data_ref *)(&iref->offset); 35425d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 35435d4f98a2SYan Zheng eb, dref, blocks); 35445d4f98a2SYan Zheng } else { 35455d4f98a2SYan Zheng BUG(); 35465d4f98a2SYan Zheng } 35475d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 35485d4f98a2SYan Zheng } 35495d4f98a2SYan Zheng WARN_ON(ptr > end); 35505d4f98a2SYan Zheng 35515d4f98a2SYan Zheng while (1) { 35525d4f98a2SYan Zheng cond_resched(); 35535d4f98a2SYan Zheng eb = path->nodes[0]; 35545d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(eb)) { 35555d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 35565d4f98a2SYan Zheng if (ret < 0) { 35575d4f98a2SYan Zheng err = ret; 35585d4f98a2SYan Zheng break; 35595d4f98a2SYan Zheng } 35605d4f98a2SYan Zheng if (ret > 0) 35615d4f98a2SYan Zheng break; 35625d4f98a2SYan Zheng eb = path->nodes[0]; 35635d4f98a2SYan Zheng } 35645d4f98a2SYan Zheng 35655d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path->slots[0]); 35665d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 35675d4f98a2SYan Zheng break; 35685d4f98a2SYan Zheng 35695d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 35705d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY || 35715d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 35725d4f98a2SYan Zheng #else 35735d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 35745d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 35755d4f98a2SYan Zheng #endif 35765d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 35775d4f98a2SYan Zheng blocks); 35785d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 35795d4f98a2SYan Zheng dref = btrfs_item_ptr(eb, path->slots[0], 35805d4f98a2SYan Zheng struct btrfs_extent_data_ref); 35815d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 35825d4f98a2SYan Zheng eb, dref, blocks); 35835d4f98a2SYan Zheng } else { 35845d4f98a2SYan Zheng ret = 0; 35855d4f98a2SYan Zheng } 35865d4f98a2SYan Zheng if (ret) { 35875d4f98a2SYan Zheng err = ret; 35885d4f98a2SYan Zheng break; 35895d4f98a2SYan Zheng } 35905d4f98a2SYan Zheng path->slots[0]++; 35915d4f98a2SYan Zheng } 3592b3b4aa74SDavid Sterba btrfs_release_path(path); 35935d4f98a2SYan Zheng if (err) 35945d4f98a2SYan Zheng free_block_list(blocks); 35955d4f98a2SYan Zheng return err; 35965d4f98a2SYan Zheng } 35975d4f98a2SYan Zheng 35985d4f98a2SYan Zheng /* 35995d4f98a2SYan Zheng * hepler to find next unprocessed extent 36005d4f98a2SYan Zheng */ 36015d4f98a2SYan Zheng static noinline_for_stack 36025d4f98a2SYan Zheng int find_next_extent(struct btrfs_trans_handle *trans, 36033fd0a558SYan, Zheng struct reloc_control *rc, struct btrfs_path *path, 36043fd0a558SYan, Zheng struct btrfs_key *extent_key) 36055d4f98a2SYan Zheng { 36065d4f98a2SYan Zheng struct btrfs_key key; 36075d4f98a2SYan Zheng struct extent_buffer *leaf; 36085d4f98a2SYan Zheng u64 start, end, last; 36095d4f98a2SYan Zheng int ret; 36105d4f98a2SYan Zheng 36115d4f98a2SYan Zheng last = rc->block_group->key.objectid + rc->block_group->key.offset; 36125d4f98a2SYan Zheng while (1) { 36135d4f98a2SYan Zheng cond_resched(); 36145d4f98a2SYan Zheng if (rc->search_start >= last) { 36155d4f98a2SYan Zheng ret = 1; 36165d4f98a2SYan Zheng break; 36175d4f98a2SYan Zheng } 36185d4f98a2SYan Zheng 36195d4f98a2SYan Zheng key.objectid = rc->search_start; 36205d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 36215d4f98a2SYan Zheng key.offset = 0; 36225d4f98a2SYan Zheng 36235d4f98a2SYan Zheng path->search_commit_root = 1; 36245d4f98a2SYan Zheng path->skip_locking = 1; 36255d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 36265d4f98a2SYan Zheng 0, 0); 36275d4f98a2SYan Zheng if (ret < 0) 36285d4f98a2SYan Zheng break; 36295d4f98a2SYan Zheng next: 36305d4f98a2SYan Zheng leaf = path->nodes[0]; 36315d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 36325d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 36335d4f98a2SYan Zheng if (ret != 0) 36345d4f98a2SYan Zheng break; 36355d4f98a2SYan Zheng leaf = path->nodes[0]; 36365d4f98a2SYan Zheng } 36375d4f98a2SYan Zheng 36385d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 36395d4f98a2SYan Zheng if (key.objectid >= last) { 36405d4f98a2SYan Zheng ret = 1; 36415d4f98a2SYan Zheng break; 36425d4f98a2SYan Zheng } 36435d4f98a2SYan Zheng 36445d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_ITEM_KEY || 36455d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 36465d4f98a2SYan Zheng path->slots[0]++; 36475d4f98a2SYan Zheng goto next; 36485d4f98a2SYan Zheng } 36495d4f98a2SYan Zheng 36505d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 36515d4f98a2SYan Zheng key.objectid, &start, &end, 3652e6138876SJosef Bacik EXTENT_DIRTY, NULL); 36535d4f98a2SYan Zheng 36545d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3655b3b4aa74SDavid Sterba btrfs_release_path(path); 36565d4f98a2SYan Zheng rc->search_start = end + 1; 36575d4f98a2SYan Zheng } else { 36585d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 36593fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 36605d4f98a2SYan Zheng return 0; 36615d4f98a2SYan Zheng } 36625d4f98a2SYan Zheng } 3663b3b4aa74SDavid Sterba btrfs_release_path(path); 36645d4f98a2SYan Zheng return ret; 36655d4f98a2SYan Zheng } 36665d4f98a2SYan Zheng 36675d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 36685d4f98a2SYan Zheng { 36695d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36707585717fSChris Mason 36717585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 36725d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 36737585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 36745d4f98a2SYan Zheng } 36755d4f98a2SYan Zheng 36765d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 36775d4f98a2SYan Zheng { 36785d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36797585717fSChris Mason 36807585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 36815d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 36827585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 36835d4f98a2SYan Zheng } 36845d4f98a2SYan Zheng 36855d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 36865d4f98a2SYan Zheng { 36875d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 36885d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 36895d4f98a2SYan Zheng return 1; 36905d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 36915d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 36925d4f98a2SYan Zheng return 1; 36935d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 36945d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 36955d4f98a2SYan Zheng return 1; 36965d4f98a2SYan Zheng return 0; 36975d4f98a2SYan Zheng } 36985d4f98a2SYan Zheng 36993fd0a558SYan, Zheng static noinline_for_stack 37003fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 37013fd0a558SYan, Zheng { 37023fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 37033fd0a558SYan, Zheng int ret; 37043fd0a558SYan, Zheng 370566d8f3ddSMiao Xie rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root, 370666d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 37073fd0a558SYan, Zheng if (!rc->block_rsv) 37083fd0a558SYan, Zheng return -ENOMEM; 37093fd0a558SYan, Zheng 37103fd0a558SYan, Zheng /* 37113fd0a558SYan, Zheng * reserve some space for creating reloc trees. 37123fd0a558SYan, Zheng * btrfs_init_reloc_root will use them when there 37133fd0a558SYan, Zheng * is no reservation in transaction handle. 37143fd0a558SYan, Zheng */ 37154a92b1b8SJosef Bacik ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv, 371608e007d2SMiao Xie rc->extent_root->nodesize * 256, 371708e007d2SMiao Xie BTRFS_RESERVE_FLUSH_ALL); 37183fd0a558SYan, Zheng if (ret) 37193fd0a558SYan, Zheng return ret; 37203fd0a558SYan, Zheng 37213fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 37223fd0a558SYan, Zheng rc->search_start = rc->block_group->key.objectid; 37233fd0a558SYan, Zheng rc->extents_found = 0; 37243fd0a558SYan, Zheng rc->nodes_relocated = 0; 37253fd0a558SYan, Zheng rc->merging_rsv_size = 0; 37263fd0a558SYan, Zheng 37273fd0a558SYan, Zheng rc->create_reloc_tree = 1; 37283fd0a558SYan, Zheng set_reloc_control(rc); 37293fd0a558SYan, Zheng 37307a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 37313612b495STsutomu Itoh BUG_ON(IS_ERR(trans)); 37323fd0a558SYan, Zheng btrfs_commit_transaction(trans, rc->extent_root); 37333fd0a558SYan, Zheng return 0; 37343fd0a558SYan, Zheng } 373576dda93cSYan, Zheng 37365d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 37375d4f98a2SYan Zheng { 37385d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 37395d4f98a2SYan Zheng struct btrfs_key key; 37405d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 37415d4f98a2SYan Zheng struct btrfs_path *path; 37425d4f98a2SYan Zheng struct btrfs_extent_item *ei; 37435d4f98a2SYan Zheng u64 flags; 37445d4f98a2SYan Zheng u32 item_size; 37455d4f98a2SYan Zheng int ret; 37465d4f98a2SYan Zheng int err = 0; 3747c87f08caSChris Mason int progress = 0; 37485d4f98a2SYan Zheng 37495d4f98a2SYan Zheng path = btrfs_alloc_path(); 37503fd0a558SYan, Zheng if (!path) 37515d4f98a2SYan Zheng return -ENOMEM; 3752026fd317SJosef Bacik path->reada = 1; 37533fd0a558SYan, Zheng 37543fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 37553fd0a558SYan, Zheng if (ret) { 37563fd0a558SYan, Zheng err = ret; 37573fd0a558SYan, Zheng goto out_free; 37582423fdfbSJiri Slaby } 37595d4f98a2SYan Zheng 37605d4f98a2SYan Zheng while (1) { 3761c87f08caSChris Mason progress++; 3762a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 376398d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 3764c87f08caSChris Mason restart: 37653fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 37663fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 37673fd0a558SYan, Zheng continue; 37683fd0a558SYan, Zheng } 37693fd0a558SYan, Zheng 37703fd0a558SYan, Zheng ret = find_next_extent(trans, rc, path, &key); 37715d4f98a2SYan Zheng if (ret < 0) 37725d4f98a2SYan Zheng err = ret; 37735d4f98a2SYan Zheng if (ret != 0) 37745d4f98a2SYan Zheng break; 37755d4f98a2SYan Zheng 37765d4f98a2SYan Zheng rc->extents_found++; 37775d4f98a2SYan Zheng 37785d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 37795d4f98a2SYan Zheng struct btrfs_extent_item); 37803fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 37815d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 37825d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 37835d4f98a2SYan Zheng ret = check_extent_flags(flags); 37845d4f98a2SYan Zheng BUG_ON(ret); 37855d4f98a2SYan Zheng 37865d4f98a2SYan Zheng } else { 37875d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 37885d4f98a2SYan Zheng u64 ref_owner; 37895d4f98a2SYan Zheng int path_change = 0; 37905d4f98a2SYan Zheng 37915d4f98a2SYan Zheng BUG_ON(item_size != 37925d4f98a2SYan Zheng sizeof(struct btrfs_extent_item_v0)); 37935d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, &key, &ref_owner, 37945d4f98a2SYan Zheng &path_change); 37955d4f98a2SYan Zheng if (ref_owner < BTRFS_FIRST_FREE_OBJECTID) 37965d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_TREE_BLOCK; 37975d4f98a2SYan Zheng else 37985d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_DATA; 37995d4f98a2SYan Zheng 38005d4f98a2SYan Zheng if (path_change) { 3801b3b4aa74SDavid Sterba btrfs_release_path(path); 38025d4f98a2SYan Zheng 38035d4f98a2SYan Zheng path->search_commit_root = 1; 38045d4f98a2SYan Zheng path->skip_locking = 1; 38055d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, 38065d4f98a2SYan Zheng &key, path, 0, 0); 38075d4f98a2SYan Zheng if (ret < 0) { 38085d4f98a2SYan Zheng err = ret; 38095d4f98a2SYan Zheng break; 38105d4f98a2SYan Zheng } 38115d4f98a2SYan Zheng BUG_ON(ret > 0); 38125d4f98a2SYan Zheng } 38135d4f98a2SYan Zheng #else 38145d4f98a2SYan Zheng BUG(); 38155d4f98a2SYan Zheng #endif 38165d4f98a2SYan Zheng } 38175d4f98a2SYan Zheng 38185d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 38195d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 38205d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 38215d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 38225d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 38235d4f98a2SYan Zheng } else { 3824b3b4aa74SDavid Sterba btrfs_release_path(path); 38255d4f98a2SYan Zheng ret = 0; 38265d4f98a2SYan Zheng } 38275d4f98a2SYan Zheng if (ret < 0) { 38283fd0a558SYan, Zheng err = ret; 38295d4f98a2SYan Zheng break; 38305d4f98a2SYan Zheng } 38315d4f98a2SYan Zheng 38325d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 38335d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 38345d4f98a2SYan Zheng if (ret < 0) { 38353fd0a558SYan, Zheng if (ret != -EAGAIN) { 38365d4f98a2SYan Zheng err = ret; 38375d4f98a2SYan Zheng break; 38385d4f98a2SYan Zheng } 38393fd0a558SYan, Zheng rc->extents_found--; 38403fd0a558SYan, Zheng rc->search_start = key.objectid; 38413fd0a558SYan, Zheng } 38425d4f98a2SYan Zheng } 38435d4f98a2SYan Zheng 384436ba022aSJosef Bacik ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5); 38453fd0a558SYan, Zheng if (ret < 0) { 38467654b724SDaniel J Blueman if (ret != -ENOSPC) { 38473fd0a558SYan, Zheng err = ret; 38483fd0a558SYan, Zheng WARN_ON(1); 38493fd0a558SYan, Zheng break; 38503fd0a558SYan, Zheng } 38513fd0a558SYan, Zheng rc->commit_transaction = 1; 38523fd0a558SYan, Zheng } 38533fd0a558SYan, Zheng 38543fd0a558SYan, Zheng if (rc->commit_transaction) { 38553fd0a558SYan, Zheng rc->commit_transaction = 0; 38563fd0a558SYan, Zheng ret = btrfs_commit_transaction(trans, rc->extent_root); 38573fd0a558SYan, Zheng BUG_ON(ret); 38583fd0a558SYan, Zheng } else { 38593fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 3860b53d3f5dSLiu Bo btrfs_btree_balance_dirty(rc->extent_root); 38613fd0a558SYan, Zheng } 38623fd0a558SYan, Zheng trans = NULL; 38635d4f98a2SYan Zheng 38645d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 38655d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 38665d4f98a2SYan Zheng rc->found_file_extent = 1; 38670257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 38683fd0a558SYan, Zheng &key, &rc->cluster); 38695d4f98a2SYan Zheng if (ret < 0) { 38705d4f98a2SYan Zheng err = ret; 38715d4f98a2SYan Zheng break; 38725d4f98a2SYan Zheng } 38735d4f98a2SYan Zheng } 38745d4f98a2SYan Zheng } 3875c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 3876c87f08caSChris Mason ret = btrfs_force_chunk_alloc(trans, rc->extent_root, 3877c87f08caSChris Mason rc->block_group->flags); 3878c87f08caSChris Mason if (ret == 0) { 3879c87f08caSChris Mason err = 0; 3880c87f08caSChris Mason progress = 0; 3881c87f08caSChris Mason goto restart; 3882c87f08caSChris Mason } 3883c87f08caSChris Mason } 38843fd0a558SYan, Zheng 3885b3b4aa74SDavid Sterba btrfs_release_path(path); 38863fd0a558SYan, Zheng clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY, 38873fd0a558SYan, Zheng GFP_NOFS); 38885d4f98a2SYan Zheng 38895d4f98a2SYan Zheng if (trans) { 38903fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 3891b53d3f5dSLiu Bo btrfs_btree_balance_dirty(rc->extent_root); 38925d4f98a2SYan Zheng } 38935d4f98a2SYan Zheng 38940257bb82SYan, Zheng if (!err) { 38953fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 38963fd0a558SYan, Zheng &rc->cluster); 38970257bb82SYan, Zheng if (ret < 0) 38980257bb82SYan, Zheng err = ret; 38990257bb82SYan, Zheng } 39000257bb82SYan, Zheng 39013fd0a558SYan, Zheng rc->create_reloc_tree = 0; 39023fd0a558SYan, Zheng set_reloc_control(rc); 39030257bb82SYan, Zheng 39043fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 39053fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1); 39065d4f98a2SYan Zheng 39073fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 39085d4f98a2SYan Zheng 39095d4f98a2SYan Zheng merge_reloc_roots(rc); 39105d4f98a2SYan Zheng 39113fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 39125d4f98a2SYan Zheng unset_reloc_control(rc); 39133fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1); 39145d4f98a2SYan Zheng 39155d4f98a2SYan Zheng /* get rid of pinned extents */ 39167a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 39173612b495STsutomu Itoh if (IS_ERR(trans)) 39183612b495STsutomu Itoh err = PTR_ERR(trans); 39193612b495STsutomu Itoh else 39205d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 39213fd0a558SYan, Zheng out_free: 39223fd0a558SYan, Zheng btrfs_free_block_rsv(rc->extent_root, rc->block_rsv); 39233fd0a558SYan, Zheng btrfs_free_path(path); 39245d4f98a2SYan Zheng return err; 39255d4f98a2SYan Zheng } 39265d4f98a2SYan Zheng 39275d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 39280257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 39295d4f98a2SYan Zheng { 39305d4f98a2SYan Zheng struct btrfs_path *path; 39315d4f98a2SYan Zheng struct btrfs_inode_item *item; 39325d4f98a2SYan Zheng struct extent_buffer *leaf; 39335d4f98a2SYan Zheng int ret; 39345d4f98a2SYan Zheng 39355d4f98a2SYan Zheng path = btrfs_alloc_path(); 39365d4f98a2SYan Zheng if (!path) 39375d4f98a2SYan Zheng return -ENOMEM; 39385d4f98a2SYan Zheng 39395d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 39405d4f98a2SYan Zheng if (ret) 39415d4f98a2SYan Zheng goto out; 39425d4f98a2SYan Zheng 39435d4f98a2SYan Zheng leaf = path->nodes[0]; 39445d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 39455d4f98a2SYan Zheng memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item)); 39465d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 39470257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 39485d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 39493fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 39503fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 39515d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 3952b3b4aa74SDavid Sterba btrfs_release_path(path); 39535d4f98a2SYan Zheng out: 39545d4f98a2SYan Zheng btrfs_free_path(path); 39555d4f98a2SYan Zheng return ret; 39565d4f98a2SYan Zheng } 39575d4f98a2SYan Zheng 39585d4f98a2SYan Zheng /* 39595d4f98a2SYan Zheng * helper to create inode for data relocation. 39605d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 39615d4f98a2SYan Zheng */ 39623fd0a558SYan, Zheng static noinline_for_stack 39633fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 39645d4f98a2SYan Zheng struct btrfs_block_group_cache *group) 39655d4f98a2SYan Zheng { 39665d4f98a2SYan Zheng struct inode *inode = NULL; 39675d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 39685d4f98a2SYan Zheng struct btrfs_root *root; 39695d4f98a2SYan Zheng struct btrfs_key key; 39705d4f98a2SYan Zheng u64 objectid = BTRFS_FIRST_FREE_OBJECTID; 39715d4f98a2SYan Zheng int err = 0; 39725d4f98a2SYan Zheng 39735d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 39745d4f98a2SYan Zheng if (IS_ERR(root)) 39755d4f98a2SYan Zheng return ERR_CAST(root); 39765d4f98a2SYan Zheng 3977a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 39783fd0a558SYan, Zheng if (IS_ERR(trans)) 39793fd0a558SYan, Zheng return ERR_CAST(trans); 39805d4f98a2SYan Zheng 3981581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 39825d4f98a2SYan Zheng if (err) 39835d4f98a2SYan Zheng goto out; 39845d4f98a2SYan Zheng 39850257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 39865d4f98a2SYan Zheng BUG_ON(err); 39875d4f98a2SYan Zheng 39885d4f98a2SYan Zheng key.objectid = objectid; 39895d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 39905d4f98a2SYan Zheng key.offset = 0; 399173f73415SJosef Bacik inode = btrfs_iget(root->fs_info->sb, &key, root, NULL); 39925d4f98a2SYan Zheng BUG_ON(IS_ERR(inode) || is_bad_inode(inode)); 39935d4f98a2SYan Zheng BTRFS_I(inode)->index_cnt = group->key.objectid; 39945d4f98a2SYan Zheng 39955d4f98a2SYan Zheng err = btrfs_orphan_add(trans, inode); 39965d4f98a2SYan Zheng out: 39975d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 3998b53d3f5dSLiu Bo btrfs_btree_balance_dirty(root); 39995d4f98a2SYan Zheng if (err) { 40005d4f98a2SYan Zheng if (inode) 40015d4f98a2SYan Zheng iput(inode); 40025d4f98a2SYan Zheng inode = ERR_PTR(err); 40035d4f98a2SYan Zheng } 40045d4f98a2SYan Zheng return inode; 40055d4f98a2SYan Zheng } 40065d4f98a2SYan Zheng 40073fd0a558SYan, Zheng static struct reloc_control *alloc_reloc_control(void) 40083fd0a558SYan, Zheng { 40093fd0a558SYan, Zheng struct reloc_control *rc; 40103fd0a558SYan, Zheng 40113fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 40123fd0a558SYan, Zheng if (!rc) 40133fd0a558SYan, Zheng return NULL; 40143fd0a558SYan, Zheng 40153fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 40163fd0a558SYan, Zheng backref_cache_init(&rc->backref_cache); 40173fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 4018f993c883SDavid Sterba extent_io_tree_init(&rc->processed_blocks, NULL); 40193fd0a558SYan, Zheng return rc; 40203fd0a558SYan, Zheng } 40213fd0a558SYan, Zheng 40225d4f98a2SYan Zheng /* 40235d4f98a2SYan Zheng * function to relocate all extents in a block group. 40245d4f98a2SYan Zheng */ 40255d4f98a2SYan Zheng int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start) 40265d4f98a2SYan Zheng { 40275d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = extent_root->fs_info; 40285d4f98a2SYan Zheng struct reloc_control *rc; 40290af3d00bSJosef Bacik struct inode *inode; 40300af3d00bSJosef Bacik struct btrfs_path *path; 40315d4f98a2SYan Zheng int ret; 4032f0486c68SYan, Zheng int rw = 0; 40335d4f98a2SYan Zheng int err = 0; 40345d4f98a2SYan Zheng 40353fd0a558SYan, Zheng rc = alloc_reloc_control(); 40365d4f98a2SYan Zheng if (!rc) 40375d4f98a2SYan Zheng return -ENOMEM; 40385d4f98a2SYan Zheng 4039f0486c68SYan, Zheng rc->extent_root = extent_root; 40403fd0a558SYan, Zheng 40415d4f98a2SYan Zheng rc->block_group = btrfs_lookup_block_group(fs_info, group_start); 40425d4f98a2SYan Zheng BUG_ON(!rc->block_group); 40435d4f98a2SYan Zheng 4044f0486c68SYan, Zheng if (!rc->block_group->ro) { 4045f0486c68SYan, Zheng ret = btrfs_set_block_group_ro(extent_root, rc->block_group); 4046f0486c68SYan, Zheng if (ret) { 4047f0486c68SYan, Zheng err = ret; 4048f0486c68SYan, Zheng goto out; 4049f0486c68SYan, Zheng } 4050f0486c68SYan, Zheng rw = 1; 4051f0486c68SYan, Zheng } 4052f0486c68SYan, Zheng 40530af3d00bSJosef Bacik path = btrfs_alloc_path(); 40540af3d00bSJosef Bacik if (!path) { 40550af3d00bSJosef Bacik err = -ENOMEM; 40560af3d00bSJosef Bacik goto out; 40570af3d00bSJosef Bacik } 40580af3d00bSJosef Bacik 40590af3d00bSJosef Bacik inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group, 40600af3d00bSJosef Bacik path); 40610af3d00bSJosef Bacik btrfs_free_path(path); 40620af3d00bSJosef Bacik 40630af3d00bSJosef Bacik if (!IS_ERR(inode)) 40640af3d00bSJosef Bacik ret = delete_block_group_cache(fs_info, inode, 0); 40650af3d00bSJosef Bacik else 40660af3d00bSJosef Bacik ret = PTR_ERR(inode); 40670af3d00bSJosef Bacik 40680af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 40690af3d00bSJosef Bacik err = ret; 40700af3d00bSJosef Bacik goto out; 40710af3d00bSJosef Bacik } 40720af3d00bSJosef Bacik 40735d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 40745d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 40755d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 40765d4f98a2SYan Zheng rc->data_inode = NULL; 40775d4f98a2SYan Zheng goto out; 40785d4f98a2SYan Zheng } 40795d4f98a2SYan Zheng 40805d4f98a2SYan Zheng printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n", 40815d4f98a2SYan Zheng (unsigned long long)rc->block_group->key.objectid, 40825d4f98a2SYan Zheng (unsigned long long)rc->block_group->flags); 40835d4f98a2SYan Zheng 40848ccf6f19SMiao Xie ret = btrfs_start_delalloc_inodes(fs_info->tree_root, 0); 40858ccf6f19SMiao Xie if (ret < 0) { 40868ccf6f19SMiao Xie err = ret; 40878ccf6f19SMiao Xie goto out; 40888ccf6f19SMiao Xie } 40896bbe3a9cSLiu Bo btrfs_wait_ordered_extents(fs_info->tree_root, 0); 40905d4f98a2SYan Zheng 40915d4f98a2SYan Zheng while (1) { 409276dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 409376dda93cSYan, Zheng 409476dda93cSYan, Zheng btrfs_clean_old_snapshots(fs_info->tree_root); 40955d4f98a2SYan Zheng ret = relocate_block_group(rc); 409676dda93cSYan, Zheng 409776dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 40985d4f98a2SYan Zheng if (ret < 0) { 40995d4f98a2SYan Zheng err = ret; 41003fd0a558SYan, Zheng goto out; 41015d4f98a2SYan Zheng } 41025d4f98a2SYan Zheng 41035d4f98a2SYan Zheng if (rc->extents_found == 0) 41045d4f98a2SYan Zheng break; 41055d4f98a2SYan Zheng 41065d4f98a2SYan Zheng printk(KERN_INFO "btrfs: found %llu extents\n", 41075d4f98a2SYan Zheng (unsigned long long)rc->extents_found); 41085d4f98a2SYan Zheng 41095d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 41105d4f98a2SYan Zheng btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1); 41115d4f98a2SYan Zheng invalidate_mapping_pages(rc->data_inode->i_mapping, 41125d4f98a2SYan Zheng 0, -1); 41135d4f98a2SYan Zheng rc->stage = UPDATE_DATA_PTRS; 41145d4f98a2SYan Zheng } 41155d4f98a2SYan Zheng } 41165d4f98a2SYan Zheng 41170257bb82SYan, Zheng filemap_write_and_wait_range(fs_info->btree_inode->i_mapping, 41185d4f98a2SYan Zheng rc->block_group->key.objectid, 41195d4f98a2SYan Zheng rc->block_group->key.objectid + 41205d4f98a2SYan Zheng rc->block_group->key.offset - 1); 41215d4f98a2SYan Zheng 41225d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 41235d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 41245d4f98a2SYan Zheng WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0); 41255d4f98a2SYan Zheng out: 4126f0486c68SYan, Zheng if (err && rw) 4127f0486c68SYan, Zheng btrfs_set_block_group_rw(extent_root, rc->block_group); 41285d4f98a2SYan Zheng iput(rc->data_inode); 41295d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 41305d4f98a2SYan Zheng kfree(rc); 41315d4f98a2SYan Zheng return err; 41325d4f98a2SYan Zheng } 41335d4f98a2SYan Zheng 413476dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 413576dda93cSYan, Zheng { 413676dda93cSYan, Zheng struct btrfs_trans_handle *trans; 413779787eaaSJeff Mahoney int ret, err; 413876dda93cSYan, Zheng 4139a22285a6SYan, Zheng trans = btrfs_start_transaction(root->fs_info->tree_root, 0); 414079787eaaSJeff Mahoney if (IS_ERR(trans)) 414179787eaaSJeff Mahoney return PTR_ERR(trans); 414276dda93cSYan, Zheng 414376dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 414476dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 414576dda93cSYan, Zheng root->root_item.drop_level = 0; 414676dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 414776dda93cSYan, Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 414876dda93cSYan, Zheng &root->root_key, &root->root_item); 414976dda93cSYan, Zheng 415079787eaaSJeff Mahoney err = btrfs_end_transaction(trans, root->fs_info->tree_root); 415179787eaaSJeff Mahoney if (err) 415279787eaaSJeff Mahoney return err; 415379787eaaSJeff Mahoney return ret; 415476dda93cSYan, Zheng } 415576dda93cSYan, Zheng 41565d4f98a2SYan Zheng /* 41575d4f98a2SYan Zheng * recover relocation interrupted by system crash. 41585d4f98a2SYan Zheng * 41595d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 41605d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 41615d4f98a2SYan Zheng */ 41625d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 41635d4f98a2SYan Zheng { 41645d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 41655d4f98a2SYan Zheng struct btrfs_key key; 41665d4f98a2SYan Zheng struct btrfs_root *fs_root; 41675d4f98a2SYan Zheng struct btrfs_root *reloc_root; 41685d4f98a2SYan Zheng struct btrfs_path *path; 41695d4f98a2SYan Zheng struct extent_buffer *leaf; 41705d4f98a2SYan Zheng struct reloc_control *rc = NULL; 41715d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 41725d4f98a2SYan Zheng int ret; 41735d4f98a2SYan Zheng int err = 0; 41745d4f98a2SYan Zheng 41755d4f98a2SYan Zheng path = btrfs_alloc_path(); 41765d4f98a2SYan Zheng if (!path) 41775d4f98a2SYan Zheng return -ENOMEM; 4178026fd317SJosef Bacik path->reada = -1; 41795d4f98a2SYan Zheng 41805d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 41815d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 41825d4f98a2SYan Zheng key.offset = (u64)-1; 41835d4f98a2SYan Zheng 41845d4f98a2SYan Zheng while (1) { 41855d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, 41865d4f98a2SYan Zheng path, 0, 0); 41875d4f98a2SYan Zheng if (ret < 0) { 41885d4f98a2SYan Zheng err = ret; 41895d4f98a2SYan Zheng goto out; 41905d4f98a2SYan Zheng } 41915d4f98a2SYan Zheng if (ret > 0) { 41925d4f98a2SYan Zheng if (path->slots[0] == 0) 41935d4f98a2SYan Zheng break; 41945d4f98a2SYan Zheng path->slots[0]--; 41955d4f98a2SYan Zheng } 41965d4f98a2SYan Zheng leaf = path->nodes[0]; 41975d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4198b3b4aa74SDavid Sterba btrfs_release_path(path); 41995d4f98a2SYan Zheng 42005d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 42015d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 42025d4f98a2SYan Zheng break; 42035d4f98a2SYan Zheng 42045d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root, &key); 42055d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 42065d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 42075d4f98a2SYan Zheng goto out; 42085d4f98a2SYan Zheng } 42095d4f98a2SYan Zheng 42105d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 42115d4f98a2SYan Zheng 42125d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 42135d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 42145d4f98a2SYan Zheng reloc_root->root_key.offset); 42155d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 421676dda93cSYan, Zheng ret = PTR_ERR(fs_root); 421776dda93cSYan, Zheng if (ret != -ENOENT) { 421876dda93cSYan, Zheng err = ret; 42195d4f98a2SYan Zheng goto out; 42205d4f98a2SYan Zheng } 422179787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 422279787eaaSJeff Mahoney if (ret < 0) { 422379787eaaSJeff Mahoney err = ret; 422479787eaaSJeff Mahoney goto out; 422579787eaaSJeff Mahoney } 422676dda93cSYan, Zheng } 42275d4f98a2SYan Zheng } 42285d4f98a2SYan Zheng 42295d4f98a2SYan Zheng if (key.offset == 0) 42305d4f98a2SYan Zheng break; 42315d4f98a2SYan Zheng 42325d4f98a2SYan Zheng key.offset--; 42335d4f98a2SYan Zheng } 4234b3b4aa74SDavid Sterba btrfs_release_path(path); 42355d4f98a2SYan Zheng 42365d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 42375d4f98a2SYan Zheng goto out; 42385d4f98a2SYan Zheng 42393fd0a558SYan, Zheng rc = alloc_reloc_control(); 42405d4f98a2SYan Zheng if (!rc) { 42415d4f98a2SYan Zheng err = -ENOMEM; 42425d4f98a2SYan Zheng goto out; 42435d4f98a2SYan Zheng } 42445d4f98a2SYan Zheng 42455d4f98a2SYan Zheng rc->extent_root = root->fs_info->extent_root; 42465d4f98a2SYan Zheng 42475d4f98a2SYan Zheng set_reloc_control(rc); 42485d4f98a2SYan Zheng 42497a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 42503612b495STsutomu Itoh if (IS_ERR(trans)) { 42513612b495STsutomu Itoh unset_reloc_control(rc); 42523612b495STsutomu Itoh err = PTR_ERR(trans); 42533612b495STsutomu Itoh goto out_free; 42543612b495STsutomu Itoh } 42553fd0a558SYan, Zheng 42563fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 42573fd0a558SYan, Zheng 42585d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 42595d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 42605d4f98a2SYan Zheng struct btrfs_root, root_list); 42615d4f98a2SYan Zheng list_del(&reloc_root->root_list); 42625d4f98a2SYan Zheng 42635d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 42645d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 42655d4f98a2SYan Zheng &rc->reloc_roots); 42665d4f98a2SYan Zheng continue; 42675d4f98a2SYan Zheng } 42685d4f98a2SYan Zheng 42695d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 42705d4f98a2SYan Zheng reloc_root->root_key.offset); 427179787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 427279787eaaSJeff Mahoney err = PTR_ERR(fs_root); 427379787eaaSJeff Mahoney goto out_free; 427479787eaaSJeff Mahoney } 42755d4f98a2SYan Zheng 4276ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 427779787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 42785d4f98a2SYan Zheng fs_root->reloc_root = reloc_root; 42795d4f98a2SYan Zheng } 42805d4f98a2SYan Zheng 428179787eaaSJeff Mahoney err = btrfs_commit_transaction(trans, rc->extent_root); 428279787eaaSJeff Mahoney if (err) 428379787eaaSJeff Mahoney goto out_free; 42845d4f98a2SYan Zheng 42855d4f98a2SYan Zheng merge_reloc_roots(rc); 42865d4f98a2SYan Zheng 42875d4f98a2SYan Zheng unset_reloc_control(rc); 42885d4f98a2SYan Zheng 42897a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 42903612b495STsutomu Itoh if (IS_ERR(trans)) 42913612b495STsutomu Itoh err = PTR_ERR(trans); 42923612b495STsutomu Itoh else 429379787eaaSJeff Mahoney err = btrfs_commit_transaction(trans, rc->extent_root); 42943612b495STsutomu Itoh out_free: 42955d4f98a2SYan Zheng kfree(rc); 42963612b495STsutomu Itoh out: 4297aca1bba6SLiu Bo if (!list_empty(&reloc_roots)) 4298aca1bba6SLiu Bo free_reloc_roots(&reloc_roots); 4299aca1bba6SLiu Bo 43005d4f98a2SYan Zheng btrfs_free_path(path); 43015d4f98a2SYan Zheng 43025d4f98a2SYan Zheng if (err == 0) { 43035d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 43045d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 43055d4f98a2SYan Zheng BTRFS_DATA_RELOC_TREE_OBJECTID); 43065d4f98a2SYan Zheng if (IS_ERR(fs_root)) 43075d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4308d7ce5843SMiao Xie else 430966b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 43105d4f98a2SYan Zheng } 43115d4f98a2SYan Zheng return err; 43125d4f98a2SYan Zheng } 43135d4f98a2SYan Zheng 43145d4f98a2SYan Zheng /* 43155d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 43165d4f98a2SYan Zheng * 43175d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 43185d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 43195d4f98a2SYan Zheng */ 43205d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 43215d4f98a2SYan Zheng { 43225d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 43235d4f98a2SYan Zheng struct btrfs_sector_sum *sector_sum; 43245d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 43255d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 43265d4f98a2SYan Zheng size_t offset; 43275d4f98a2SYan Zheng int ret; 43285d4f98a2SYan Zheng u64 disk_bytenr; 43295d4f98a2SYan Zheng LIST_HEAD(list); 43305d4f98a2SYan Zheng 43315d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 43325d4f98a2SYan Zheng BUG_ON(ordered->file_offset != file_pos || ordered->len != len); 43335d4f98a2SYan Zheng 43345d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 43355d4f98a2SYan Zheng ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr, 4336a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 433779787eaaSJeff Mahoney if (ret) 433879787eaaSJeff Mahoney goto out; 43395d4f98a2SYan Zheng 43405d4f98a2SYan Zheng while (!list_empty(&list)) { 43415d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 43425d4f98a2SYan Zheng list_del_init(&sums->list); 43435d4f98a2SYan Zheng 43445d4f98a2SYan Zheng sector_sum = sums->sums; 43455d4f98a2SYan Zheng sums->bytenr = ordered->start; 43465d4f98a2SYan Zheng 43475d4f98a2SYan Zheng offset = 0; 43485d4f98a2SYan Zheng while (offset < sums->len) { 43495d4f98a2SYan Zheng sector_sum->bytenr += ordered->start - disk_bytenr; 43505d4f98a2SYan Zheng sector_sum++; 43515d4f98a2SYan Zheng offset += root->sectorsize; 43525d4f98a2SYan Zheng } 43535d4f98a2SYan Zheng 43545d4f98a2SYan Zheng btrfs_add_ordered_sum(inode, ordered, sums); 43555d4f98a2SYan Zheng } 435679787eaaSJeff Mahoney out: 43575d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4358411fc6bcSAndi Kleen return ret; 43595d4f98a2SYan Zheng } 43603fd0a558SYan, Zheng 43613fd0a558SYan, Zheng void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 43623fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 43633fd0a558SYan, Zheng struct extent_buffer *cow) 43643fd0a558SYan, Zheng { 43653fd0a558SYan, Zheng struct reloc_control *rc; 43663fd0a558SYan, Zheng struct backref_node *node; 43673fd0a558SYan, Zheng int first_cow = 0; 43683fd0a558SYan, Zheng int level; 43693fd0a558SYan, Zheng int ret; 43703fd0a558SYan, Zheng 43713fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 43723fd0a558SYan, Zheng if (!rc) 43733fd0a558SYan, Zheng return; 43743fd0a558SYan, Zheng 43753fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 43763fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 43773fd0a558SYan, Zheng 43783fd0a558SYan, Zheng level = btrfs_header_level(buf); 43793fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 43803fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 43813fd0a558SYan, Zheng first_cow = 1; 43823fd0a558SYan, Zheng 43833fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 43843fd0a558SYan, Zheng rc->create_reloc_tree) { 43853fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 43863fd0a558SYan, Zheng 43873fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 43883fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 43893fd0a558SYan, Zheng node->new_bytenr != buf->start); 43903fd0a558SYan, Zheng 43913fd0a558SYan, Zheng drop_node_buffer(node); 43923fd0a558SYan, Zheng extent_buffer_get(cow); 43933fd0a558SYan, Zheng node->eb = cow; 43943fd0a558SYan, Zheng node->new_bytenr = cow->start; 43953fd0a558SYan, Zheng 43963fd0a558SYan, Zheng if (!node->pending) { 43973fd0a558SYan, Zheng list_move_tail(&node->list, 43983fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 43993fd0a558SYan, Zheng node->pending = 1; 44003fd0a558SYan, Zheng } 44013fd0a558SYan, Zheng 44023fd0a558SYan, Zheng if (first_cow) 44033fd0a558SYan, Zheng __mark_block_processed(rc, node); 44043fd0a558SYan, Zheng 44053fd0a558SYan, Zheng if (first_cow && level > 0) 44063fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 44073fd0a558SYan, Zheng } 44083fd0a558SYan, Zheng 44093fd0a558SYan, Zheng if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) { 44103fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 44113fd0a558SYan, Zheng BUG_ON(ret); 44123fd0a558SYan, Zheng } 44133fd0a558SYan, Zheng } 44143fd0a558SYan, Zheng 44153fd0a558SYan, Zheng /* 44163fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 44173fd0a558SYan, Zheng * requried for relocating tree blocks in the snapshot 44183fd0a558SYan, Zheng */ 44193fd0a558SYan, Zheng void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans, 44203fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending, 44213fd0a558SYan, Zheng u64 *bytes_to_reserve) 44223fd0a558SYan, Zheng { 44233fd0a558SYan, Zheng struct btrfs_root *root; 44243fd0a558SYan, Zheng struct reloc_control *rc; 44253fd0a558SYan, Zheng 44263fd0a558SYan, Zheng root = pending->root; 44273fd0a558SYan, Zheng if (!root->reloc_root) 44283fd0a558SYan, Zheng return; 44293fd0a558SYan, Zheng 44303fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 44313fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 44323fd0a558SYan, Zheng return; 44333fd0a558SYan, Zheng 44343fd0a558SYan, Zheng root = root->reloc_root; 44353fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 44363fd0a558SYan, Zheng /* 44373fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 44383fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 44393fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 44403fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 44413fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 44423fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 44433fd0a558SYan, Zheng * tree may use more space than it frees. so we need 44443fd0a558SYan, Zheng * reserve extra space. 44453fd0a558SYan, Zheng */ 44463fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 44473fd0a558SYan, Zheng } 44483fd0a558SYan, Zheng 44493fd0a558SYan, Zheng /* 44503fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 44513fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 44523fd0a558SYan, Zheng */ 445349b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 44543fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 44553fd0a558SYan, Zheng { 44563fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 44573fd0a558SYan, Zheng struct btrfs_root *reloc_root; 44583fd0a558SYan, Zheng struct btrfs_root *new_root; 44593fd0a558SYan, Zheng struct reloc_control *rc; 44603fd0a558SYan, Zheng int ret; 44613fd0a558SYan, Zheng 44623fd0a558SYan, Zheng if (!root->reloc_root) 446349b25e05SJeff Mahoney return 0; 44643fd0a558SYan, Zheng 44653fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 44663fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 44673fd0a558SYan, Zheng 44683fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 44693fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 44703fd0a558SYan, Zheng rc->block_rsv, 44713fd0a558SYan, Zheng rc->nodes_relocated); 447249b25e05SJeff Mahoney if (ret) 447349b25e05SJeff Mahoney return ret; 44743fd0a558SYan, Zheng } 44753fd0a558SYan, Zheng 44763fd0a558SYan, Zheng new_root = pending->snap; 44773fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 44783fd0a558SYan, Zheng new_root->root_key.objectid); 447949b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 448049b25e05SJeff Mahoney return PTR_ERR(reloc_root); 44813fd0a558SYan, Zheng 4482ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4483ffd7b339SJeff Mahoney BUG_ON(ret < 0); 44843fd0a558SYan, Zheng new_root->reloc_root = reloc_root; 44853fd0a558SYan, Zheng 448649b25e05SJeff Mahoney if (rc->create_reloc_tree) 44873fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 448849b25e05SJeff Mahoney return ret; 44893fd0a558SYan, Zheng } 4490