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 12725d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 12735d4f98a2SYan Zheng 12745d4f98a2SYan Zheng if (!del) { 12755d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12765d4f98a2SYan Zheng node->bytenr = root->node->start; 12775d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 12785d4f98a2SYan Zheng node->bytenr, &node->rb_node); 12795d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 128043c04fb1SJeff Mahoney if (rb_node) 128143c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, node->bytenr); 12825d4f98a2SYan Zheng } else { 12831daf3540SDaniel J Blueman spin_lock(&root->fs_info->trans_lock); 12845d4f98a2SYan Zheng list_del_init(&root->root_list); 12851daf3540SDaniel J Blueman spin_unlock(&root->fs_info->trans_lock); 12865d4f98a2SYan Zheng kfree(node); 12875d4f98a2SYan Zheng } 12885d4f98a2SYan Zheng return 0; 12895d4f98a2SYan Zheng } 12905d4f98a2SYan Zheng 12913fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 12923fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 12935d4f98a2SYan Zheng { 12945d4f98a2SYan Zheng struct btrfs_root *reloc_root; 12955d4f98a2SYan Zheng struct extent_buffer *eb; 12965d4f98a2SYan Zheng struct btrfs_root_item *root_item; 12975d4f98a2SYan Zheng struct btrfs_key root_key; 12985d4f98a2SYan Zheng int ret; 12995d4f98a2SYan Zheng 13005d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 13015d4f98a2SYan Zheng BUG_ON(!root_item); 13025d4f98a2SYan Zheng 13035d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 13045d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 13053fd0a558SYan, Zheng root_key.offset = objectid; 13065d4f98a2SYan Zheng 13073fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 13083fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 13095d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 13105d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 13115d4f98a2SYan Zheng BUG_ON(ret); 13125d4f98a2SYan Zheng 13133fd0a558SYan, Zheng btrfs_set_root_last_snapshot(&root->root_item, 13143fd0a558SYan, Zheng trans->transid - 1); 13153fd0a558SYan, Zheng } else { 13163fd0a558SYan, Zheng /* 13173fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 13183fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 13193fd0a558SYan, Zheng * modified after it was created have RELOC flag 13203fd0a558SYan, Zheng * set in their headers. so it's OK to not update 13213fd0a558SYan, Zheng * the 'last_snapshot'. 13223fd0a558SYan, Zheng */ 13233fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 13243fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 13253fd0a558SYan, Zheng BUG_ON(ret); 13263fd0a558SYan, Zheng } 13273fd0a558SYan, Zheng 13285d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 13295d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 13305d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 13315d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 13323fd0a558SYan, Zheng 13333fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 13343fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 13353fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 13363fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 13375d4f98a2SYan Zheng root_item->drop_level = 0; 13383fd0a558SYan, Zheng } 13395d4f98a2SYan Zheng 13405d4f98a2SYan Zheng btrfs_tree_unlock(eb); 13415d4f98a2SYan Zheng free_extent_buffer(eb); 13425d4f98a2SYan Zheng 13435d4f98a2SYan Zheng ret = btrfs_insert_root(trans, root->fs_info->tree_root, 13445d4f98a2SYan Zheng &root_key, root_item); 13455d4f98a2SYan Zheng BUG_ON(ret); 13465d4f98a2SYan Zheng kfree(root_item); 13475d4f98a2SYan Zheng 13485d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root, 13495d4f98a2SYan Zheng &root_key); 13505d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 13515d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 13523fd0a558SYan, Zheng return reloc_root; 13533fd0a558SYan, Zheng } 13543fd0a558SYan, Zheng 13553fd0a558SYan, Zheng /* 13563fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 13573fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 13583fd0a558SYan, Zheng */ 13593fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 13603fd0a558SYan, Zheng struct btrfs_root *root) 13613fd0a558SYan, Zheng { 13623fd0a558SYan, Zheng struct btrfs_root *reloc_root; 13633fd0a558SYan, Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 13643fd0a558SYan, Zheng int clear_rsv = 0; 1365ffd7b339SJeff Mahoney int ret; 13663fd0a558SYan, Zheng 13673fd0a558SYan, Zheng if (root->reloc_root) { 13683fd0a558SYan, Zheng reloc_root = root->reloc_root; 13693fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 13703fd0a558SYan, Zheng return 0; 13713fd0a558SYan, Zheng } 13723fd0a558SYan, Zheng 13733fd0a558SYan, Zheng if (!rc || !rc->create_reloc_tree || 13743fd0a558SYan, Zheng root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 13753fd0a558SYan, Zheng return 0; 13763fd0a558SYan, Zheng 13773fd0a558SYan, Zheng if (!trans->block_rsv) { 13783fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 13793fd0a558SYan, Zheng clear_rsv = 1; 13803fd0a558SYan, Zheng } 13813fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 13823fd0a558SYan, Zheng if (clear_rsv) 13833fd0a558SYan, Zheng trans->block_rsv = NULL; 13845d4f98a2SYan Zheng 1385ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 1386ffd7b339SJeff Mahoney BUG_ON(ret < 0); 13875d4f98a2SYan Zheng root->reloc_root = reloc_root; 13885d4f98a2SYan Zheng return 0; 13895d4f98a2SYan Zheng } 13905d4f98a2SYan Zheng 13915d4f98a2SYan Zheng /* 13925d4f98a2SYan Zheng * update root item of reloc tree 13935d4f98a2SYan Zheng */ 13945d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 13955d4f98a2SYan Zheng struct btrfs_root *root) 13965d4f98a2SYan Zheng { 13975d4f98a2SYan Zheng struct btrfs_root *reloc_root; 13985d4f98a2SYan Zheng struct btrfs_root_item *root_item; 13995d4f98a2SYan Zheng int del = 0; 14005d4f98a2SYan Zheng int ret; 14015d4f98a2SYan Zheng 14025d4f98a2SYan Zheng if (!root->reloc_root) 14037585717fSChris Mason goto out; 14045d4f98a2SYan Zheng 14055d4f98a2SYan Zheng reloc_root = root->reloc_root; 14065d4f98a2SYan Zheng root_item = &reloc_root->root_item; 14075d4f98a2SYan Zheng 14083fd0a558SYan, Zheng if (root->fs_info->reloc_ctl->merge_reloc_tree && 14093fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 14105d4f98a2SYan Zheng root->reloc_root = NULL; 14115d4f98a2SYan Zheng del = 1; 14125d4f98a2SYan Zheng } 14135d4f98a2SYan Zheng 14145d4f98a2SYan Zheng __update_reloc_root(reloc_root, del); 14155d4f98a2SYan Zheng 14165d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 14175d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 14185d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 14195d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 14205d4f98a2SYan Zheng } 14215d4f98a2SYan Zheng 14225d4f98a2SYan Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 14235d4f98a2SYan Zheng &reloc_root->root_key, root_item); 14245d4f98a2SYan Zheng BUG_ON(ret); 14257585717fSChris Mason 14267585717fSChris Mason out: 14275d4f98a2SYan Zheng return 0; 14285d4f98a2SYan Zheng } 14295d4f98a2SYan Zheng 14305d4f98a2SYan Zheng /* 14315d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 14325d4f98a2SYan Zheng * in a subvolume 14335d4f98a2SYan Zheng */ 14345d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 14355d4f98a2SYan Zheng { 14365d4f98a2SYan Zheng struct rb_node *node; 14375d4f98a2SYan Zheng struct rb_node *prev; 14385d4f98a2SYan Zheng struct btrfs_inode *entry; 14395d4f98a2SYan Zheng struct inode *inode; 14405d4f98a2SYan Zheng 14415d4f98a2SYan Zheng spin_lock(&root->inode_lock); 14425d4f98a2SYan Zheng again: 14435d4f98a2SYan Zheng node = root->inode_tree.rb_node; 14445d4f98a2SYan Zheng prev = NULL; 14455d4f98a2SYan Zheng while (node) { 14465d4f98a2SYan Zheng prev = node; 14475d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 14485d4f98a2SYan Zheng 144933345d01SLi Zefan if (objectid < btrfs_ino(&entry->vfs_inode)) 14505d4f98a2SYan Zheng node = node->rb_left; 145133345d01SLi Zefan else if (objectid > btrfs_ino(&entry->vfs_inode)) 14525d4f98a2SYan Zheng node = node->rb_right; 14535d4f98a2SYan Zheng else 14545d4f98a2SYan Zheng break; 14555d4f98a2SYan Zheng } 14565d4f98a2SYan Zheng if (!node) { 14575d4f98a2SYan Zheng while (prev) { 14585d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 145933345d01SLi Zefan if (objectid <= btrfs_ino(&entry->vfs_inode)) { 14605d4f98a2SYan Zheng node = prev; 14615d4f98a2SYan Zheng break; 14625d4f98a2SYan Zheng } 14635d4f98a2SYan Zheng prev = rb_next(prev); 14645d4f98a2SYan Zheng } 14655d4f98a2SYan Zheng } 14665d4f98a2SYan Zheng while (node) { 14675d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 14685d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 14695d4f98a2SYan Zheng if (inode) { 14705d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 14715d4f98a2SYan Zheng return inode; 14725d4f98a2SYan Zheng } 14735d4f98a2SYan Zheng 147433345d01SLi Zefan objectid = btrfs_ino(&entry->vfs_inode) + 1; 14755d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 14765d4f98a2SYan Zheng goto again; 14775d4f98a2SYan Zheng 14785d4f98a2SYan Zheng node = rb_next(node); 14795d4f98a2SYan Zheng } 14805d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 14815d4f98a2SYan Zheng return NULL; 14825d4f98a2SYan Zheng } 14835d4f98a2SYan Zheng 14845d4f98a2SYan Zheng static int in_block_group(u64 bytenr, 14855d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group) 14865d4f98a2SYan Zheng { 14875d4f98a2SYan Zheng if (bytenr >= block_group->key.objectid && 14885d4f98a2SYan Zheng bytenr < block_group->key.objectid + block_group->key.offset) 14895d4f98a2SYan Zheng return 1; 14905d4f98a2SYan Zheng return 0; 14915d4f98a2SYan Zheng } 14925d4f98a2SYan Zheng 14935d4f98a2SYan Zheng /* 14945d4f98a2SYan Zheng * get new location of data 14955d4f98a2SYan Zheng */ 14965d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 14975d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 14985d4f98a2SYan Zheng { 14995d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 15005d4f98a2SYan Zheng struct btrfs_path *path; 15015d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 15025d4f98a2SYan Zheng struct extent_buffer *leaf; 15035d4f98a2SYan Zheng int ret; 15045d4f98a2SYan Zheng 15055d4f98a2SYan Zheng path = btrfs_alloc_path(); 15065d4f98a2SYan Zheng if (!path) 15075d4f98a2SYan Zheng return -ENOMEM; 15085d4f98a2SYan Zheng 15095d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 151033345d01SLi Zefan ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode), 15115d4f98a2SYan Zheng bytenr, 0); 15125d4f98a2SYan Zheng if (ret < 0) 15135d4f98a2SYan Zheng goto out; 15145d4f98a2SYan Zheng if (ret > 0) { 15155d4f98a2SYan Zheng ret = -ENOENT; 15165d4f98a2SYan Zheng goto out; 15175d4f98a2SYan Zheng } 15185d4f98a2SYan Zheng 15195d4f98a2SYan Zheng leaf = path->nodes[0]; 15205d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 15215d4f98a2SYan Zheng struct btrfs_file_extent_item); 15225d4f98a2SYan Zheng 15235d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 15245d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 15255d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 15265d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 15275d4f98a2SYan Zheng 15285d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 15295d4f98a2SYan Zheng ret = 1; 15305d4f98a2SYan Zheng goto out; 15315d4f98a2SYan Zheng } 15325d4f98a2SYan Zheng 15335d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 15345d4f98a2SYan Zheng ret = 0; 15355d4f98a2SYan Zheng out: 15365d4f98a2SYan Zheng btrfs_free_path(path); 15375d4f98a2SYan Zheng return ret; 15385d4f98a2SYan Zheng } 15395d4f98a2SYan Zheng 15405d4f98a2SYan Zheng /* 15415d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 15425d4f98a2SYan Zheng * the new locations. 15435d4f98a2SYan Zheng */ 15443fd0a558SYan, Zheng static noinline_for_stack 15453fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 15465d4f98a2SYan Zheng struct reloc_control *rc, 15475d4f98a2SYan Zheng struct btrfs_root *root, 15483fd0a558SYan, Zheng struct extent_buffer *leaf) 15495d4f98a2SYan Zheng { 15505d4f98a2SYan Zheng struct btrfs_key key; 15515d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 15525d4f98a2SYan Zheng struct inode *inode = NULL; 15535d4f98a2SYan Zheng u64 parent; 15545d4f98a2SYan Zheng u64 bytenr; 15553fd0a558SYan, Zheng u64 new_bytenr = 0; 15565d4f98a2SYan Zheng u64 num_bytes; 15575d4f98a2SYan Zheng u64 end; 15585d4f98a2SYan Zheng u32 nritems; 15595d4f98a2SYan Zheng u32 i; 15605d4f98a2SYan Zheng int ret; 15615d4f98a2SYan Zheng int first = 1; 15625d4f98a2SYan Zheng int dirty = 0; 15635d4f98a2SYan Zheng 15645d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 15655d4f98a2SYan Zheng return 0; 15665d4f98a2SYan Zheng 15675d4f98a2SYan Zheng /* reloc trees always use full backref */ 15685d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 15695d4f98a2SYan Zheng parent = leaf->start; 15705d4f98a2SYan Zheng else 15715d4f98a2SYan Zheng parent = 0; 15725d4f98a2SYan Zheng 15735d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 15745d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 15755d4f98a2SYan Zheng cond_resched(); 15765d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 15775d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 15785d4f98a2SYan Zheng continue; 15795d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 15805d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 15815d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 15825d4f98a2SYan Zheng continue; 15835d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 15845d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 15855d4f98a2SYan Zheng if (bytenr == 0) 15865d4f98a2SYan Zheng continue; 15875d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group)) 15885d4f98a2SYan Zheng continue; 15895d4f98a2SYan Zheng 15905d4f98a2SYan Zheng /* 15915d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 15925d4f98a2SYan Zheng * to complete and drop the extent cache 15935d4f98a2SYan Zheng */ 15945d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 15955d4f98a2SYan Zheng if (first) { 15965d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 15975d4f98a2SYan Zheng first = 0; 159833345d01SLi Zefan } else if (inode && btrfs_ino(inode) < key.objectid) { 15993fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 16005d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 16015d4f98a2SYan Zheng } 160233345d01SLi Zefan if (inode && btrfs_ino(inode) == key.objectid) { 16035d4f98a2SYan Zheng end = key.offset + 16045d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 16055d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 16065d4f98a2SYan Zheng root->sectorsize)); 16075d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 16085d4f98a2SYan Zheng end--; 16095d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 1610d0082371SJeff Mahoney key.offset, end); 16115d4f98a2SYan Zheng if (!ret) 16125d4f98a2SYan Zheng continue; 16135d4f98a2SYan Zheng 16145d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, key.offset, end, 16155d4f98a2SYan Zheng 1); 16165d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 1617d0082371SJeff Mahoney key.offset, end); 16185d4f98a2SYan Zheng } 16195d4f98a2SYan Zheng } 16205d4f98a2SYan Zheng 16215d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 16225d4f98a2SYan Zheng bytenr, num_bytes); 16233fd0a558SYan, Zheng if (ret > 0) { 16243fd0a558SYan, Zheng WARN_ON(1); 16255d4f98a2SYan Zheng continue; 16263fd0a558SYan, Zheng } 16275d4f98a2SYan Zheng BUG_ON(ret < 0); 16285d4f98a2SYan Zheng 16295d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 16305d4f98a2SYan Zheng dirty = 1; 16315d4f98a2SYan Zheng 16325d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 16335d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, new_bytenr, 16345d4f98a2SYan Zheng num_bytes, parent, 16355d4f98a2SYan Zheng btrfs_header_owner(leaf), 163666d7e7f0SArne Jansen key.objectid, key.offset, 1); 16375d4f98a2SYan Zheng BUG_ON(ret); 16385d4f98a2SYan Zheng 16395d4f98a2SYan Zheng ret = btrfs_free_extent(trans, root, bytenr, num_bytes, 16405d4f98a2SYan Zheng parent, btrfs_header_owner(leaf), 164166d7e7f0SArne Jansen key.objectid, key.offset, 1); 16425d4f98a2SYan Zheng BUG_ON(ret); 16435d4f98a2SYan Zheng } 16445d4f98a2SYan Zheng if (dirty) 16455d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 16463fd0a558SYan, Zheng if (inode) 16473fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 16485d4f98a2SYan Zheng return 0; 16495d4f98a2SYan Zheng } 16505d4f98a2SYan Zheng 16515d4f98a2SYan Zheng static noinline_for_stack 16525d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 16535d4f98a2SYan Zheng struct btrfs_path *path, int level) 16545d4f98a2SYan Zheng { 16555d4f98a2SYan Zheng struct btrfs_disk_key key1; 16565d4f98a2SYan Zheng struct btrfs_disk_key key2; 16575d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 16585d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 16595d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 16605d4f98a2SYan Zheng } 16615d4f98a2SYan Zheng 16625d4f98a2SYan Zheng /* 16635d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 16645d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 16655d4f98a2SYan Zheng * reloc tree was create can be replaced. 16665d4f98a2SYan Zheng * 16675d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 16685d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 16695d4f98a2SYan Zheng * errors, a negative error number is returned. 16705d4f98a2SYan Zheng */ 16713fd0a558SYan, Zheng static noinline_for_stack 16723fd0a558SYan, Zheng int replace_path(struct btrfs_trans_handle *trans, 16735d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 16745d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 16755d4f98a2SYan Zheng int lowest_level, int max_level) 16765d4f98a2SYan Zheng { 16775d4f98a2SYan Zheng struct extent_buffer *eb; 16785d4f98a2SYan Zheng struct extent_buffer *parent; 16795d4f98a2SYan Zheng struct btrfs_key key; 16805d4f98a2SYan Zheng u64 old_bytenr; 16815d4f98a2SYan Zheng u64 new_bytenr; 16825d4f98a2SYan Zheng u64 old_ptr_gen; 16835d4f98a2SYan Zheng u64 new_ptr_gen; 16845d4f98a2SYan Zheng u64 last_snapshot; 16855d4f98a2SYan Zheng u32 blocksize; 16863fd0a558SYan, Zheng int cow = 0; 16875d4f98a2SYan Zheng int level; 16885d4f98a2SYan Zheng int ret; 16895d4f98a2SYan Zheng int slot; 16905d4f98a2SYan Zheng 16915d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 16925d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 16935d4f98a2SYan Zheng 16945d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 16953fd0a558SYan, Zheng again: 16965d4f98a2SYan Zheng slot = path->slots[lowest_level]; 16975d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 16985d4f98a2SYan Zheng 16995d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 17005d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 17015d4f98a2SYan Zheng level = btrfs_header_level(eb); 17025d4f98a2SYan Zheng 17035d4f98a2SYan Zheng if (level < lowest_level) { 17045d4f98a2SYan Zheng btrfs_tree_unlock(eb); 17055d4f98a2SYan Zheng free_extent_buffer(eb); 17065d4f98a2SYan Zheng return 0; 17075d4f98a2SYan Zheng } 17085d4f98a2SYan Zheng 17093fd0a558SYan, Zheng if (cow) { 17105d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 17115d4f98a2SYan Zheng BUG_ON(ret); 17123fd0a558SYan, Zheng } 17135d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 17145d4f98a2SYan Zheng 17155d4f98a2SYan Zheng if (next_key) { 17165d4f98a2SYan Zheng next_key->objectid = (u64)-1; 17175d4f98a2SYan Zheng next_key->type = (u8)-1; 17185d4f98a2SYan Zheng next_key->offset = (u64)-1; 17195d4f98a2SYan Zheng } 17205d4f98a2SYan Zheng 17215d4f98a2SYan Zheng parent = eb; 17225d4f98a2SYan Zheng while (1) { 17235d4f98a2SYan Zheng level = btrfs_header_level(parent); 17245d4f98a2SYan Zheng BUG_ON(level < lowest_level); 17255d4f98a2SYan Zheng 17265d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 17275d4f98a2SYan Zheng if (ret && slot > 0) 17285d4f98a2SYan Zheng slot--; 17295d4f98a2SYan Zheng 17305d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 17315d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 17325d4f98a2SYan Zheng 17335d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 17345d4f98a2SYan Zheng blocksize = btrfs_level_size(dest, level - 1); 17355d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 17365d4f98a2SYan Zheng 17375d4f98a2SYan Zheng if (level <= max_level) { 17385d4f98a2SYan Zheng eb = path->nodes[level]; 17395d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 17405d4f98a2SYan Zheng path->slots[level]); 17415d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 17425d4f98a2SYan Zheng path->slots[level]); 17435d4f98a2SYan Zheng } else { 17445d4f98a2SYan Zheng new_bytenr = 0; 17455d4f98a2SYan Zheng new_ptr_gen = 0; 17465d4f98a2SYan Zheng } 17475d4f98a2SYan Zheng 17485d4f98a2SYan Zheng if (new_bytenr > 0 && new_bytenr == old_bytenr) { 17495d4f98a2SYan Zheng WARN_ON(1); 17505d4f98a2SYan Zheng ret = level; 17515d4f98a2SYan Zheng break; 17525d4f98a2SYan Zheng } 17535d4f98a2SYan Zheng 17545d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 17555d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 17563fd0a558SYan, Zheng if (level <= lowest_level) { 17575d4f98a2SYan Zheng ret = 0; 17585d4f98a2SYan Zheng break; 17595d4f98a2SYan Zheng } 17605d4f98a2SYan Zheng 17615d4f98a2SYan Zheng eb = read_tree_block(dest, old_bytenr, blocksize, 17625d4f98a2SYan Zheng old_ptr_gen); 176397d9a8a4STsutomu Itoh BUG_ON(!eb); 17645d4f98a2SYan Zheng btrfs_tree_lock(eb); 17653fd0a558SYan, Zheng if (cow) { 17665d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 17675d4f98a2SYan Zheng slot, &eb); 17685d4f98a2SYan Zheng BUG_ON(ret); 17695d4f98a2SYan Zheng } 17703fd0a558SYan, Zheng btrfs_set_lock_blocking(eb); 17715d4f98a2SYan Zheng 17725d4f98a2SYan Zheng btrfs_tree_unlock(parent); 17735d4f98a2SYan Zheng free_extent_buffer(parent); 17745d4f98a2SYan Zheng 17755d4f98a2SYan Zheng parent = eb; 17765d4f98a2SYan Zheng continue; 17775d4f98a2SYan Zheng } 17785d4f98a2SYan Zheng 17793fd0a558SYan, Zheng if (!cow) { 17803fd0a558SYan, Zheng btrfs_tree_unlock(parent); 17813fd0a558SYan, Zheng free_extent_buffer(parent); 17823fd0a558SYan, Zheng cow = 1; 17833fd0a558SYan, Zheng goto again; 17843fd0a558SYan, Zheng } 17853fd0a558SYan, Zheng 17865d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17875d4f98a2SYan Zheng path->slots[level]); 1788b3b4aa74SDavid Sterba btrfs_release_path(path); 17895d4f98a2SYan Zheng 17905d4f98a2SYan Zheng path->lowest_level = level; 17915d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 17925d4f98a2SYan Zheng path->lowest_level = 0; 17935d4f98a2SYan Zheng BUG_ON(ret); 17945d4f98a2SYan Zheng 17955d4f98a2SYan Zheng /* 17965d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 17975d4f98a2SYan Zheng */ 17985d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 17995d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 18005d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 18015d4f98a2SYan Zheng 18025d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 18035d4f98a2SYan Zheng path->slots[level], old_bytenr); 18045d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 18055d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 18065d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 18075d4f98a2SYan Zheng 18085d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize, 18095d4f98a2SYan Zheng path->nodes[level]->start, 181066d7e7f0SArne Jansen src->root_key.objectid, level - 1, 0, 181166d7e7f0SArne Jansen 1); 18125d4f98a2SYan Zheng BUG_ON(ret); 18135d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize, 18145d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 181566d7e7f0SArne Jansen 0, 1); 18165d4f98a2SYan Zheng BUG_ON(ret); 18175d4f98a2SYan Zheng 18185d4f98a2SYan Zheng ret = btrfs_free_extent(trans, src, new_bytenr, blocksize, 18195d4f98a2SYan Zheng path->nodes[level]->start, 182066d7e7f0SArne Jansen src->root_key.objectid, level - 1, 0, 182166d7e7f0SArne Jansen 1); 18225d4f98a2SYan Zheng BUG_ON(ret); 18235d4f98a2SYan Zheng 18245d4f98a2SYan Zheng ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize, 18255d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 182666d7e7f0SArne Jansen 0, 1); 18275d4f98a2SYan Zheng BUG_ON(ret); 18285d4f98a2SYan Zheng 18295d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 18305d4f98a2SYan Zheng 18315d4f98a2SYan Zheng ret = level; 18325d4f98a2SYan Zheng break; 18335d4f98a2SYan Zheng } 18345d4f98a2SYan Zheng btrfs_tree_unlock(parent); 18355d4f98a2SYan Zheng free_extent_buffer(parent); 18365d4f98a2SYan Zheng return ret; 18375d4f98a2SYan Zheng } 18385d4f98a2SYan Zheng 18395d4f98a2SYan Zheng /* 18405d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 18415d4f98a2SYan Zheng */ 18425d4f98a2SYan Zheng static noinline_for_stack 18435d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 18445d4f98a2SYan Zheng int *level) 18455d4f98a2SYan Zheng { 18465d4f98a2SYan Zheng struct extent_buffer *eb; 18475d4f98a2SYan Zheng int i; 18485d4f98a2SYan Zheng u64 last_snapshot; 18495d4f98a2SYan Zheng u32 nritems; 18505d4f98a2SYan Zheng 18515d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 18525d4f98a2SYan Zheng 18535d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 18545d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 18555d4f98a2SYan Zheng path->nodes[i] = NULL; 18565d4f98a2SYan Zheng } 18575d4f98a2SYan Zheng 18585d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 18595d4f98a2SYan Zheng eb = path->nodes[i]; 18605d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 18615d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 18625d4f98a2SYan Zheng path->slots[i]++; 18635d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 18645d4f98a2SYan Zheng last_snapshot) 18655d4f98a2SYan Zheng continue; 18665d4f98a2SYan Zheng 18675d4f98a2SYan Zheng *level = i; 18685d4f98a2SYan Zheng return 0; 18695d4f98a2SYan Zheng } 18705d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 18715d4f98a2SYan Zheng path->nodes[i] = NULL; 18725d4f98a2SYan Zheng } 18735d4f98a2SYan Zheng return 1; 18745d4f98a2SYan Zheng } 18755d4f98a2SYan Zheng 18765d4f98a2SYan Zheng /* 18775d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 18785d4f98a2SYan Zheng */ 18795d4f98a2SYan Zheng static noinline_for_stack 18805d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 18815d4f98a2SYan Zheng int *level) 18825d4f98a2SYan Zheng { 18835d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 18845d4f98a2SYan Zheng int i; 18855d4f98a2SYan Zheng u64 bytenr; 18865d4f98a2SYan Zheng u64 ptr_gen = 0; 18875d4f98a2SYan Zheng u64 last_snapshot; 18885d4f98a2SYan Zheng u32 blocksize; 18895d4f98a2SYan Zheng u32 nritems; 18905d4f98a2SYan Zheng 18915d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 18925d4f98a2SYan Zheng 18935d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 18945d4f98a2SYan Zheng eb = path->nodes[i]; 18955d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 18965d4f98a2SYan Zheng while (path->slots[i] < nritems) { 18975d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 18985d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 18995d4f98a2SYan Zheng break; 19005d4f98a2SYan Zheng path->slots[i]++; 19015d4f98a2SYan Zheng } 19025d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 19035d4f98a2SYan Zheng if (i == *level) 19045d4f98a2SYan Zheng break; 19055d4f98a2SYan Zheng *level = i + 1; 19065d4f98a2SYan Zheng return 0; 19075d4f98a2SYan Zheng } 19085d4f98a2SYan Zheng if (i == 1) { 19095d4f98a2SYan Zheng *level = i; 19105d4f98a2SYan Zheng return 0; 19115d4f98a2SYan Zheng } 19125d4f98a2SYan Zheng 19135d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 19145d4f98a2SYan Zheng blocksize = btrfs_level_size(root, i - 1); 19155d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, ptr_gen); 19165d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 19175d4f98a2SYan Zheng path->nodes[i - 1] = eb; 19185d4f98a2SYan Zheng path->slots[i - 1] = 0; 19195d4f98a2SYan Zheng } 19205d4f98a2SYan Zheng return 1; 19215d4f98a2SYan Zheng } 19225d4f98a2SYan Zheng 19235d4f98a2SYan Zheng /* 19245d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 19255d4f98a2SYan Zheng * [min_key, max_key) 19265d4f98a2SYan Zheng */ 19275d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 19285d4f98a2SYan Zheng struct btrfs_key *min_key, 19295d4f98a2SYan Zheng struct btrfs_key *max_key) 19305d4f98a2SYan Zheng { 19315d4f98a2SYan Zheng struct inode *inode = NULL; 19325d4f98a2SYan Zheng u64 objectid; 19335d4f98a2SYan Zheng u64 start, end; 193433345d01SLi Zefan u64 ino; 19355d4f98a2SYan Zheng 19365d4f98a2SYan Zheng objectid = min_key->objectid; 19375d4f98a2SYan Zheng while (1) { 19385d4f98a2SYan Zheng cond_resched(); 19395d4f98a2SYan Zheng iput(inode); 19405d4f98a2SYan Zheng 19415d4f98a2SYan Zheng if (objectid > max_key->objectid) 19425d4f98a2SYan Zheng break; 19435d4f98a2SYan Zheng 19445d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 19455d4f98a2SYan Zheng if (!inode) 19465d4f98a2SYan Zheng break; 194733345d01SLi Zefan ino = btrfs_ino(inode); 19485d4f98a2SYan Zheng 194933345d01SLi Zefan if (ino > max_key->objectid) { 19505d4f98a2SYan Zheng iput(inode); 19515d4f98a2SYan Zheng break; 19525d4f98a2SYan Zheng } 19535d4f98a2SYan Zheng 195433345d01SLi Zefan objectid = ino + 1; 19555d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 19565d4f98a2SYan Zheng continue; 19575d4f98a2SYan Zheng 195833345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 19595d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 19605d4f98a2SYan Zheng continue; 19615d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 19625d4f98a2SYan Zheng start = 0; 19635d4f98a2SYan Zheng else { 19645d4f98a2SYan Zheng start = min_key->offset; 19655d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(start, root->sectorsize)); 19665d4f98a2SYan Zheng } 19675d4f98a2SYan Zheng } else { 19685d4f98a2SYan Zheng start = 0; 19695d4f98a2SYan Zheng } 19705d4f98a2SYan Zheng 197133345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 19725d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 19735d4f98a2SYan Zheng continue; 19745d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 19755d4f98a2SYan Zheng end = (u64)-1; 19765d4f98a2SYan Zheng } else { 19775d4f98a2SYan Zheng if (max_key->offset == 0) 19785d4f98a2SYan Zheng continue; 19795d4f98a2SYan Zheng end = max_key->offset; 19805d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 19815d4f98a2SYan Zheng end--; 19825d4f98a2SYan Zheng } 19835d4f98a2SYan Zheng } else { 19845d4f98a2SYan Zheng end = (u64)-1; 19855d4f98a2SYan Zheng } 19865d4f98a2SYan Zheng 19875d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 1988d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 19895d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 1); 1990d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 19915d4f98a2SYan Zheng } 19925d4f98a2SYan Zheng return 0; 19935d4f98a2SYan Zheng } 19945d4f98a2SYan Zheng 19955d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 19965d4f98a2SYan Zheng struct btrfs_key *key) 19975d4f98a2SYan Zheng 19985d4f98a2SYan Zheng { 19995d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 20005d4f98a2SYan Zheng if (!path->nodes[level]) 20015d4f98a2SYan Zheng break; 20025d4f98a2SYan Zheng if (path->slots[level] + 1 < 20035d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 20045d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 20055d4f98a2SYan Zheng path->slots[level] + 1); 20065d4f98a2SYan Zheng return 0; 20075d4f98a2SYan Zheng } 20085d4f98a2SYan Zheng level++; 20095d4f98a2SYan Zheng } 20105d4f98a2SYan Zheng return 1; 20115d4f98a2SYan Zheng } 20125d4f98a2SYan Zheng 20135d4f98a2SYan Zheng /* 20145d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 20155d4f98a2SYan Zheng * fs tree. 20165d4f98a2SYan Zheng */ 20175d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 20185d4f98a2SYan Zheng struct btrfs_root *root) 20195d4f98a2SYan Zheng { 20205d4f98a2SYan Zheng LIST_HEAD(inode_list); 20215d4f98a2SYan Zheng struct btrfs_key key; 20225d4f98a2SYan Zheng struct btrfs_key next_key; 20235d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 20245d4f98a2SYan Zheng struct btrfs_root *reloc_root; 20255d4f98a2SYan Zheng struct btrfs_root_item *root_item; 20265d4f98a2SYan Zheng struct btrfs_path *path; 20273fd0a558SYan, Zheng struct extent_buffer *leaf; 20285d4f98a2SYan Zheng unsigned long nr; 20295d4f98a2SYan Zheng int level; 20305d4f98a2SYan Zheng int max_level; 20315d4f98a2SYan Zheng int replaced = 0; 20325d4f98a2SYan Zheng int ret; 20335d4f98a2SYan Zheng int err = 0; 20343fd0a558SYan, Zheng u32 min_reserved; 20355d4f98a2SYan Zheng 20365d4f98a2SYan Zheng path = btrfs_alloc_path(); 20375d4f98a2SYan Zheng if (!path) 20385d4f98a2SYan Zheng return -ENOMEM; 2039026fd317SJosef Bacik path->reada = 1; 20405d4f98a2SYan Zheng 20415d4f98a2SYan Zheng reloc_root = root->reloc_root; 20425d4f98a2SYan Zheng root_item = &reloc_root->root_item; 20435d4f98a2SYan Zheng 20445d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 20455d4f98a2SYan Zheng level = btrfs_root_level(root_item); 20465d4f98a2SYan Zheng extent_buffer_get(reloc_root->node); 20475d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 20485d4f98a2SYan Zheng path->slots[level] = 0; 20495d4f98a2SYan Zheng } else { 20505d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 20515d4f98a2SYan Zheng 20525d4f98a2SYan Zheng level = root_item->drop_level; 20535d4f98a2SYan Zheng BUG_ON(level == 0); 20545d4f98a2SYan Zheng path->lowest_level = level; 20555d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 205633c66f43SYan Zheng path->lowest_level = 0; 20575d4f98a2SYan Zheng if (ret < 0) { 20585d4f98a2SYan Zheng btrfs_free_path(path); 20595d4f98a2SYan Zheng return ret; 20605d4f98a2SYan Zheng } 20615d4f98a2SYan Zheng 20625d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 20635d4f98a2SYan Zheng path->slots[level]); 20645d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 20655d4f98a2SYan Zheng 20665d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 20675d4f98a2SYan Zheng } 20685d4f98a2SYan Zheng 20693fd0a558SYan, Zheng min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 20705d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 20715d4f98a2SYan Zheng 20725d4f98a2SYan Zheng while (1) { 2073a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 0); 207498d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 20753fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 20763fd0a558SYan, Zheng 207736ba022aSJosef Bacik ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved); 20783fd0a558SYan, Zheng if (ret) { 20793fd0a558SYan, Zheng BUG_ON(ret != -EAGAIN); 20803fd0a558SYan, Zheng ret = btrfs_commit_transaction(trans, root); 20813fd0a558SYan, Zheng BUG_ON(ret); 20823fd0a558SYan, Zheng continue; 20833fd0a558SYan, Zheng } 20843fd0a558SYan, Zheng 20853fd0a558SYan, Zheng replaced = 0; 20865d4f98a2SYan Zheng max_level = level; 20875d4f98a2SYan Zheng 20885d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 20895d4f98a2SYan Zheng if (ret < 0) { 20905d4f98a2SYan Zheng err = ret; 20915d4f98a2SYan Zheng goto out; 20925d4f98a2SYan Zheng } 20935d4f98a2SYan Zheng if (ret > 0) 20945d4f98a2SYan Zheng break; 20955d4f98a2SYan Zheng 20965d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 20975d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 20985d4f98a2SYan Zheng ret = 0; 20995d4f98a2SYan Zheng } else { 21003fd0a558SYan, Zheng ret = replace_path(trans, root, reloc_root, path, 21013fd0a558SYan, Zheng &next_key, level, max_level); 21025d4f98a2SYan Zheng } 21035d4f98a2SYan Zheng if (ret < 0) { 21045d4f98a2SYan Zheng err = ret; 21055d4f98a2SYan Zheng goto out; 21065d4f98a2SYan Zheng } 21075d4f98a2SYan Zheng 21085d4f98a2SYan Zheng if (ret > 0) { 21095d4f98a2SYan Zheng level = ret; 21105d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 21115d4f98a2SYan Zheng path->slots[level]); 21125d4f98a2SYan Zheng replaced = 1; 21135d4f98a2SYan Zheng } 21145d4f98a2SYan Zheng 21155d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 21165d4f98a2SYan Zheng if (ret > 0) 21175d4f98a2SYan Zheng break; 21185d4f98a2SYan Zheng 21195d4f98a2SYan Zheng BUG_ON(level == 0); 21205d4f98a2SYan Zheng /* 21215d4f98a2SYan Zheng * save the merging progress in the drop_progress. 21225d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 21235d4f98a2SYan Zheng */ 21245d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 21255d4f98a2SYan Zheng path->slots[level]); 21265d4f98a2SYan Zheng root_item->drop_level = level; 21275d4f98a2SYan Zheng 21285d4f98a2SYan Zheng nr = trans->blocks_used; 21293fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, root); 21305d4f98a2SYan Zheng 21315d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 21325d4f98a2SYan Zheng 21335d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 21345d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 21355d4f98a2SYan Zheng } 21365d4f98a2SYan Zheng 21375d4f98a2SYan Zheng /* 21385d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 21395d4f98a2SYan Zheng * relocated and the block is tree root. 21405d4f98a2SYan Zheng */ 21415d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 21425d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 21435d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 21445d4f98a2SYan Zheng free_extent_buffer(leaf); 21455d4f98a2SYan Zheng if (ret < 0) 21465d4f98a2SYan Zheng err = ret; 21475d4f98a2SYan Zheng out: 21485d4f98a2SYan Zheng btrfs_free_path(path); 21495d4f98a2SYan Zheng 21505d4f98a2SYan Zheng if (err == 0) { 21515d4f98a2SYan Zheng memset(&root_item->drop_progress, 0, 21525d4f98a2SYan Zheng sizeof(root_item->drop_progress)); 21535d4f98a2SYan Zheng root_item->drop_level = 0; 21545d4f98a2SYan Zheng btrfs_set_root_refs(root_item, 0); 21553fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 21565d4f98a2SYan Zheng } 21575d4f98a2SYan Zheng 21585d4f98a2SYan Zheng nr = trans->blocks_used; 21593fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, root); 21605d4f98a2SYan Zheng 21615d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 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; 21874a92b1b8SJosef Bacik ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes); 21883fd0a558SYan, Zheng if (ret) 21893fd0a558SYan, Zheng err = ret; 21903fd0a558SYan, Zheng } 21913fd0a558SYan, Zheng 21927a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 21933612b495STsutomu Itoh if (IS_ERR(trans)) { 21943612b495STsutomu Itoh if (!err) 21953612b495STsutomu Itoh btrfs_block_rsv_release(rc->extent_root, 21963612b495STsutomu Itoh rc->block_rsv, num_bytes); 21973612b495STsutomu Itoh return PTR_ERR(trans); 21983612b495STsutomu Itoh } 21993fd0a558SYan, Zheng 22003fd0a558SYan, Zheng if (!err) { 22013fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 22023fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 22033fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, 22043fd0a558SYan, Zheng rc->block_rsv, num_bytes); 22053fd0a558SYan, Zheng goto again; 22063fd0a558SYan, Zheng } 22073fd0a558SYan, Zheng } 22083fd0a558SYan, Zheng 22093fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 22103fd0a558SYan, Zheng 22113fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 22123fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 22133fd0a558SYan, Zheng struct btrfs_root, root_list); 22143fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 22153fd0a558SYan, Zheng 22163fd0a558SYan, Zheng root = read_fs_root(reloc_root->fs_info, 22173fd0a558SYan, Zheng reloc_root->root_key.offset); 22183fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 22193fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 22203fd0a558SYan, Zheng 22213fd0a558SYan, Zheng /* 22223fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 22233fd0a558SYan, Zheng * knows it should resumes merging 22243fd0a558SYan, Zheng */ 22253fd0a558SYan, Zheng if (!err) 22263fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 22273fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 22283fd0a558SYan, Zheng 22293fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 22303fd0a558SYan, Zheng } 22313fd0a558SYan, Zheng 22323fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 22333fd0a558SYan, Zheng 22343fd0a558SYan, Zheng if (!err) 22353fd0a558SYan, Zheng btrfs_commit_transaction(trans, rc->extent_root); 22363fd0a558SYan, Zheng else 22373fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 22383fd0a558SYan, Zheng return err; 22393fd0a558SYan, Zheng } 22403fd0a558SYan, Zheng 22413fd0a558SYan, Zheng static noinline_for_stack 22423fd0a558SYan, Zheng int merge_reloc_roots(struct reloc_control *rc) 22433fd0a558SYan, Zheng { 22445d4f98a2SYan Zheng struct btrfs_root *root; 22455d4f98a2SYan Zheng struct btrfs_root *reloc_root; 22463fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 22473fd0a558SYan, Zheng int found = 0; 22483fd0a558SYan, Zheng int ret; 22493fd0a558SYan, Zheng again: 22503fd0a558SYan, Zheng root = rc->extent_root; 22517585717fSChris Mason 22527585717fSChris Mason /* 22537585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 22547585717fSChris Mason * we have to make sure nobody is in the middle of 22557585717fSChris Mason * adding their roots to the list while we are 22567585717fSChris Mason * doing this splice 22577585717fSChris Mason */ 22587585717fSChris Mason mutex_lock(&root->fs_info->reloc_mutex); 22593fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 22607585717fSChris Mason mutex_unlock(&root->fs_info->reloc_mutex); 22615d4f98a2SYan Zheng 22623fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 22633fd0a558SYan, Zheng found = 1; 22643fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 22653fd0a558SYan, Zheng struct btrfs_root, root_list); 22665d4f98a2SYan Zheng 22675d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 22685d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, 22695d4f98a2SYan Zheng reloc_root->root_key.offset); 22705d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 22715d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 22725d4f98a2SYan Zheng 22733fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 22743fd0a558SYan, Zheng BUG_ON(ret); 22753fd0a558SYan, Zheng } else { 22763fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 22773fd0a558SYan, Zheng } 22782c536799SJeff Mahoney ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1); 22792c536799SJeff Mahoney BUG_ON(ret < 0); 22805d4f98a2SYan Zheng } 22815d4f98a2SYan Zheng 22823fd0a558SYan, Zheng if (found) { 22833fd0a558SYan, Zheng found = 0; 22843fd0a558SYan, Zheng goto again; 22855d4f98a2SYan Zheng } 22865d4f98a2SYan Zheng BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 22875d4f98a2SYan Zheng return 0; 22885d4f98a2SYan Zheng } 22895d4f98a2SYan Zheng 22905d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 22915d4f98a2SYan Zheng { 22925d4f98a2SYan Zheng struct tree_block *block; 22935d4f98a2SYan Zheng struct rb_node *rb_node; 22945d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 22955d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 22965d4f98a2SYan Zheng rb_erase(rb_node, blocks); 22975d4f98a2SYan Zheng kfree(block); 22985d4f98a2SYan Zheng } 22995d4f98a2SYan Zheng } 23005d4f98a2SYan Zheng 23015d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 23025d4f98a2SYan Zheng struct btrfs_root *reloc_root) 23035d4f98a2SYan Zheng { 23045d4f98a2SYan Zheng struct btrfs_root *root; 23055d4f98a2SYan Zheng 23065d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 23075d4f98a2SYan Zheng return 0; 23085d4f98a2SYan Zheng 23095d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset); 23105d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 23115d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 23125d4f98a2SYan Zheng 23135d4f98a2SYan Zheng return btrfs_record_root_in_trans(trans, root); 23145d4f98a2SYan Zheng } 23155d4f98a2SYan Zheng 23163fd0a558SYan, Zheng static noinline_for_stack 23173fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 23183fd0a558SYan, Zheng struct reloc_control *rc, 23195d4f98a2SYan Zheng struct backref_node *node, 23203fd0a558SYan, Zheng struct backref_edge *edges[], int *nr) 23215d4f98a2SYan Zheng { 23225d4f98a2SYan Zheng struct backref_node *next; 23235d4f98a2SYan Zheng struct btrfs_root *root; 23243fd0a558SYan, Zheng int index = 0; 23253fd0a558SYan, Zheng 23265d4f98a2SYan Zheng next = node; 23275d4f98a2SYan Zheng while (1) { 23285d4f98a2SYan Zheng cond_resched(); 23295d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 23305d4f98a2SYan Zheng root = next->root; 23313fd0a558SYan, Zheng BUG_ON(!root); 23323fd0a558SYan, Zheng BUG_ON(!root->ref_cows); 23335d4f98a2SYan Zheng 23345d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 23355d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 23365d4f98a2SYan Zheng break; 23375d4f98a2SYan Zheng } 23385d4f98a2SYan Zheng 23395d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 23403fd0a558SYan, Zheng root = root->reloc_root; 23413fd0a558SYan, Zheng 23423fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 23433fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 23443fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 23453fd0a558SYan, Zheng next->new_bytenr = root->node->start; 23463fd0a558SYan, Zheng next->root = root; 23473fd0a558SYan, Zheng list_add_tail(&next->list, 23483fd0a558SYan, Zheng &rc->backref_cache.changed); 23493fd0a558SYan, Zheng __mark_block_processed(rc, next); 23505d4f98a2SYan Zheng break; 23515d4f98a2SYan Zheng } 23525d4f98a2SYan Zheng 23533fd0a558SYan, Zheng WARN_ON(1); 23545d4f98a2SYan Zheng root = NULL; 23555d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 23565d4f98a2SYan Zheng if (!next || next->level <= node->level) 23575d4f98a2SYan Zheng break; 23585d4f98a2SYan Zheng } 23593fd0a558SYan, Zheng if (!root) 23603fd0a558SYan, Zheng return NULL; 23615d4f98a2SYan Zheng 23625d4f98a2SYan Zheng *nr = index; 23633fd0a558SYan, Zheng next = node; 23643fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 23653fd0a558SYan, Zheng while (1) { 23663fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 23673fd0a558SYan, Zheng if (--index < 0) 23683fd0a558SYan, Zheng break; 23693fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 23703fd0a558SYan, Zheng } 23715d4f98a2SYan Zheng return root; 23725d4f98a2SYan Zheng } 23735d4f98a2SYan Zheng 23743fd0a558SYan, Zheng /* 23753fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 23763fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 23773fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 23783fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 23793fd0a558SYan, Zheng */ 23805d4f98a2SYan Zheng static noinline_for_stack 23815d4f98a2SYan Zheng struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans, 23825d4f98a2SYan Zheng struct backref_node *node) 23835d4f98a2SYan Zheng { 23843fd0a558SYan, Zheng struct backref_node *next; 23853fd0a558SYan, Zheng struct btrfs_root *root; 23863fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 23875d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 23883fd0a558SYan, Zheng int index = 0; 23893fd0a558SYan, Zheng 23903fd0a558SYan, Zheng next = node; 23913fd0a558SYan, Zheng while (1) { 23923fd0a558SYan, Zheng cond_resched(); 23933fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 23943fd0a558SYan, Zheng root = next->root; 23953fd0a558SYan, Zheng BUG_ON(!root); 23963fd0a558SYan, Zheng 239725985edcSLucas De Marchi /* no other choice for non-references counted tree */ 23983fd0a558SYan, Zheng if (!root->ref_cows) 23993fd0a558SYan, Zheng return root; 24003fd0a558SYan, Zheng 24013fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 24023fd0a558SYan, Zheng fs_root = root; 24033fd0a558SYan, Zheng 24043fd0a558SYan, Zheng if (next != node) 24053fd0a558SYan, Zheng return NULL; 24063fd0a558SYan, Zheng 24073fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 24083fd0a558SYan, Zheng if (!next || next->level <= node->level) 24093fd0a558SYan, Zheng break; 24103fd0a558SYan, Zheng } 24113fd0a558SYan, Zheng 24123fd0a558SYan, Zheng if (!fs_root) 24133fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 24143fd0a558SYan, Zheng return fs_root; 24155d4f98a2SYan Zheng } 24165d4f98a2SYan Zheng 24175d4f98a2SYan Zheng static noinline_for_stack 24183fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 24193fd0a558SYan, Zheng struct backref_node *node, int reserve) 24205d4f98a2SYan Zheng { 24213fd0a558SYan, Zheng struct backref_node *next = node; 24223fd0a558SYan, Zheng struct backref_edge *edge; 24233fd0a558SYan, Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24243fd0a558SYan, Zheng u64 num_bytes = 0; 24253fd0a558SYan, Zheng int index = 0; 24265d4f98a2SYan Zheng 24273fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 24283fd0a558SYan, Zheng 24293fd0a558SYan, Zheng while (next) { 24303fd0a558SYan, Zheng cond_resched(); 24315d4f98a2SYan Zheng while (1) { 24323fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 24335d4f98a2SYan Zheng break; 24345d4f98a2SYan Zheng 24353fd0a558SYan, Zheng num_bytes += btrfs_level_size(rc->extent_root, 24363fd0a558SYan, Zheng next->level); 24373fd0a558SYan, Zheng 24383fd0a558SYan, Zheng if (list_empty(&next->upper)) 24393fd0a558SYan, Zheng break; 24403fd0a558SYan, Zheng 24413fd0a558SYan, Zheng edge = list_entry(next->upper.next, 24423fd0a558SYan, Zheng struct backref_edge, list[LOWER]); 24433fd0a558SYan, Zheng edges[index++] = edge; 24443fd0a558SYan, Zheng next = edge->node[UPPER]; 24455d4f98a2SYan Zheng } 24463fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 24473fd0a558SYan, Zheng } 24483fd0a558SYan, Zheng return num_bytes; 24493fd0a558SYan, Zheng } 24503fd0a558SYan, Zheng 24513fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 24523fd0a558SYan, Zheng struct reloc_control *rc, 24533fd0a558SYan, Zheng struct backref_node *node) 24543fd0a558SYan, Zheng { 24553fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 24563fd0a558SYan, Zheng u64 num_bytes; 24573fd0a558SYan, Zheng int ret; 24583fd0a558SYan, Zheng 24593fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 24603fd0a558SYan, Zheng 24613fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 24624a92b1b8SJosef Bacik ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes); 24633fd0a558SYan, Zheng if (ret) { 24643fd0a558SYan, Zheng if (ret == -EAGAIN) 24653fd0a558SYan, Zheng rc->commit_transaction = 1; 24663fd0a558SYan, Zheng return ret; 24673fd0a558SYan, Zheng } 24683fd0a558SYan, Zheng 24693fd0a558SYan, Zheng return 0; 24703fd0a558SYan, Zheng } 24713fd0a558SYan, Zheng 24723fd0a558SYan, Zheng static void release_metadata_space(struct reloc_control *rc, 24733fd0a558SYan, Zheng struct backref_node *node) 24743fd0a558SYan, Zheng { 24753fd0a558SYan, Zheng u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2; 24763fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes); 24775d4f98a2SYan Zheng } 24785d4f98a2SYan Zheng 24795d4f98a2SYan Zheng /* 24805d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 24815d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 24825d4f98a2SYan Zheng * 24835d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 24845d4f98a2SYan Zheng * in that case this function just updates pointers. 24855d4f98a2SYan Zheng */ 24865d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 24873fd0a558SYan, Zheng struct reloc_control *rc, 24885d4f98a2SYan Zheng struct backref_node *node, 24895d4f98a2SYan Zheng struct btrfs_key *key, 24905d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 24915d4f98a2SYan Zheng { 24925d4f98a2SYan Zheng struct backref_node *upper; 24935d4f98a2SYan Zheng struct backref_edge *edge; 24945d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24955d4f98a2SYan Zheng struct btrfs_root *root; 24965d4f98a2SYan Zheng struct extent_buffer *eb; 24975d4f98a2SYan Zheng u32 blocksize; 24985d4f98a2SYan Zheng u64 bytenr; 24995d4f98a2SYan Zheng u64 generation; 25005d4f98a2SYan Zheng int nr; 25015d4f98a2SYan Zheng int slot; 25025d4f98a2SYan Zheng int ret; 25035d4f98a2SYan Zheng int err = 0; 25045d4f98a2SYan Zheng 25055d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 25065d4f98a2SYan Zheng 25075d4f98a2SYan Zheng path->lowest_level = node->level + 1; 25083fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 25095d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 25105d4f98a2SYan Zheng cond_resched(); 25115d4f98a2SYan Zheng 25125d4f98a2SYan Zheng upper = edge->node[UPPER]; 25133fd0a558SYan, Zheng root = select_reloc_root(trans, rc, upper, edges, &nr); 25143fd0a558SYan, Zheng BUG_ON(!root); 25155d4f98a2SYan Zheng 25163fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 25173fd0a558SYan, Zheng if (!lowest) { 25183fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 25193fd0a558SYan, Zheng upper->level, &slot); 25203fd0a558SYan, Zheng BUG_ON(ret); 25213fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 25223fd0a558SYan, Zheng if (node->eb->start == bytenr) 25233fd0a558SYan, Zheng goto next; 25243fd0a558SYan, Zheng } 25255d4f98a2SYan Zheng drop_node_buffer(upper); 25263fd0a558SYan, Zheng } 25275d4f98a2SYan Zheng 25285d4f98a2SYan Zheng if (!upper->eb) { 25295d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 25305d4f98a2SYan Zheng if (ret < 0) { 25315d4f98a2SYan Zheng err = ret; 25325d4f98a2SYan Zheng break; 25335d4f98a2SYan Zheng } 25345d4f98a2SYan Zheng BUG_ON(ret > 0); 25355d4f98a2SYan Zheng 25363fd0a558SYan, Zheng if (!upper->eb) { 25373fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 25383fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 25393fd0a558SYan, Zheng } else { 25403fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 25413fd0a558SYan, Zheng } 25423fd0a558SYan, Zheng 25433fd0a558SYan, Zheng upper->locked = 1; 25443fd0a558SYan, Zheng path->locks[upper->level] = 0; 25453fd0a558SYan, Zheng 25465d4f98a2SYan Zheng slot = path->slots[upper->level]; 2547b3b4aa74SDavid Sterba btrfs_release_path(path); 25485d4f98a2SYan Zheng } else { 25495d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 25505d4f98a2SYan Zheng &slot); 25515d4f98a2SYan Zheng BUG_ON(ret); 25525d4f98a2SYan Zheng } 25535d4f98a2SYan Zheng 25545d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 25553fd0a558SYan, Zheng if (lowest) { 25563fd0a558SYan, Zheng BUG_ON(bytenr != node->bytenr); 25575d4f98a2SYan Zheng } else { 25583fd0a558SYan, Zheng if (node->eb->start == bytenr) 25593fd0a558SYan, Zheng goto next; 25605d4f98a2SYan Zheng } 25615d4f98a2SYan Zheng 25625d4f98a2SYan Zheng blocksize = btrfs_level_size(root, node->level); 25635d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 25645d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, generation); 256597d9a8a4STsutomu Itoh if (!eb) { 256697d9a8a4STsutomu Itoh err = -EIO; 256797d9a8a4STsutomu Itoh goto next; 256897d9a8a4STsutomu Itoh } 25695d4f98a2SYan Zheng btrfs_tree_lock(eb); 25705d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 25715d4f98a2SYan Zheng 25725d4f98a2SYan Zheng if (!node->eb) { 25735d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 25745d4f98a2SYan Zheng slot, &eb); 25753fd0a558SYan, Zheng btrfs_tree_unlock(eb); 25763fd0a558SYan, Zheng free_extent_buffer(eb); 25775d4f98a2SYan Zheng if (ret < 0) { 25785d4f98a2SYan Zheng err = ret; 25793fd0a558SYan, Zheng goto next; 25805d4f98a2SYan Zheng } 25813fd0a558SYan, Zheng BUG_ON(node->eb != eb); 25825d4f98a2SYan Zheng } else { 25835d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 25845d4f98a2SYan Zheng node->eb->start); 25855d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 25865d4f98a2SYan Zheng trans->transid); 25875d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 25885d4f98a2SYan Zheng 25895d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, 25905d4f98a2SYan Zheng node->eb->start, blocksize, 25915d4f98a2SYan Zheng upper->eb->start, 25925d4f98a2SYan Zheng btrfs_header_owner(upper->eb), 259366d7e7f0SArne Jansen node->level, 0, 1); 25945d4f98a2SYan Zheng BUG_ON(ret); 25955d4f98a2SYan Zheng 25965d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 25975d4f98a2SYan Zheng BUG_ON(ret); 25985d4f98a2SYan Zheng } 25993fd0a558SYan, Zheng next: 26003fd0a558SYan, Zheng if (!upper->pending) 26013fd0a558SYan, Zheng drop_node_buffer(upper); 26023fd0a558SYan, Zheng else 26033fd0a558SYan, Zheng unlock_node_buffer(upper); 26043fd0a558SYan, Zheng if (err) 26053fd0a558SYan, Zheng break; 26065d4f98a2SYan Zheng } 26073fd0a558SYan, Zheng 26083fd0a558SYan, Zheng if (!err && node->pending) { 26093fd0a558SYan, Zheng drop_node_buffer(node); 26103fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 26113fd0a558SYan, Zheng node->pending = 0; 26125d4f98a2SYan Zheng } 26133fd0a558SYan, Zheng 26145d4f98a2SYan Zheng path->lowest_level = 0; 26153fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 26165d4f98a2SYan Zheng return err; 26175d4f98a2SYan Zheng } 26185d4f98a2SYan Zheng 26195d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 26203fd0a558SYan, Zheng struct reloc_control *rc, 26215d4f98a2SYan Zheng struct backref_node *node, 26225d4f98a2SYan Zheng struct btrfs_path *path) 26235d4f98a2SYan Zheng { 26245d4f98a2SYan Zheng struct btrfs_key key; 26255d4f98a2SYan Zheng 26265d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 26273fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 26285d4f98a2SYan Zheng } 26295d4f98a2SYan Zheng 26305d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 26313fd0a558SYan, Zheng struct reloc_control *rc, 26323fd0a558SYan, Zheng struct btrfs_path *path, int err) 26335d4f98a2SYan Zheng { 26343fd0a558SYan, Zheng LIST_HEAD(list); 26353fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 26365d4f98a2SYan Zheng struct backref_node *node; 26375d4f98a2SYan Zheng int level; 26385d4f98a2SYan Zheng int ret; 26395d4f98a2SYan Zheng 26405d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 26415d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 26425d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 26433fd0a558SYan, Zheng struct backref_node, list); 26443fd0a558SYan, Zheng list_move_tail(&node->list, &list); 26453fd0a558SYan, Zheng BUG_ON(!node->pending); 26465d4f98a2SYan Zheng 26473fd0a558SYan, Zheng if (!err) { 26483fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 26495d4f98a2SYan Zheng if (ret < 0) 26505d4f98a2SYan Zheng err = ret; 26515d4f98a2SYan Zheng } 26525d4f98a2SYan Zheng } 26533fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 26543fd0a558SYan, Zheng } 26555d4f98a2SYan Zheng return err; 26565d4f98a2SYan Zheng } 26575d4f98a2SYan Zheng 26585d4f98a2SYan Zheng static void mark_block_processed(struct reloc_control *rc, 26593fd0a558SYan, Zheng u64 bytenr, u32 blocksize) 26603fd0a558SYan, Zheng { 26613fd0a558SYan, Zheng set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1, 26623fd0a558SYan, Zheng EXTENT_DIRTY, GFP_NOFS); 26633fd0a558SYan, Zheng } 26643fd0a558SYan, Zheng 26653fd0a558SYan, Zheng static void __mark_block_processed(struct reloc_control *rc, 26665d4f98a2SYan Zheng struct backref_node *node) 26675d4f98a2SYan Zheng { 26685d4f98a2SYan Zheng u32 blocksize; 26695d4f98a2SYan Zheng if (node->level == 0 || 26705d4f98a2SYan Zheng in_block_group(node->bytenr, rc->block_group)) { 26715d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, node->level); 26723fd0a558SYan, Zheng mark_block_processed(rc, node->bytenr, blocksize); 26735d4f98a2SYan Zheng } 26745d4f98a2SYan Zheng node->processed = 1; 26755d4f98a2SYan Zheng } 26765d4f98a2SYan Zheng 26775d4f98a2SYan Zheng /* 26785d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 26795d4f98a2SYan Zheng * as processed. 26805d4f98a2SYan Zheng */ 26815d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 26825d4f98a2SYan Zheng struct backref_node *node) 26835d4f98a2SYan Zheng { 26845d4f98a2SYan Zheng struct backref_node *next = node; 26855d4f98a2SYan Zheng struct backref_edge *edge; 26865d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 26875d4f98a2SYan Zheng int index = 0; 26885d4f98a2SYan Zheng 26895d4f98a2SYan Zheng while (next) { 26905d4f98a2SYan Zheng cond_resched(); 26915d4f98a2SYan Zheng while (1) { 26925d4f98a2SYan Zheng if (next->processed) 26935d4f98a2SYan Zheng break; 26945d4f98a2SYan Zheng 26953fd0a558SYan, Zheng __mark_block_processed(rc, next); 26965d4f98a2SYan Zheng 26975d4f98a2SYan Zheng if (list_empty(&next->upper)) 26985d4f98a2SYan Zheng break; 26995d4f98a2SYan Zheng 27005d4f98a2SYan Zheng edge = list_entry(next->upper.next, 27015d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 27025d4f98a2SYan Zheng edges[index++] = edge; 27035d4f98a2SYan Zheng next = edge->node[UPPER]; 27045d4f98a2SYan Zheng } 27055d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 27065d4f98a2SYan Zheng } 27075d4f98a2SYan Zheng } 27085d4f98a2SYan Zheng 27095d4f98a2SYan Zheng static int tree_block_processed(u64 bytenr, u32 blocksize, 27105d4f98a2SYan Zheng struct reloc_control *rc) 27115d4f98a2SYan Zheng { 27125d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 27139655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 27145d4f98a2SYan Zheng return 1; 27155d4f98a2SYan Zheng return 0; 27165d4f98a2SYan Zheng } 27175d4f98a2SYan Zheng 27185d4f98a2SYan Zheng static int get_tree_block_key(struct reloc_control *rc, 27195d4f98a2SYan Zheng struct tree_block *block) 27205d4f98a2SYan Zheng { 27215d4f98a2SYan Zheng struct extent_buffer *eb; 27225d4f98a2SYan Zheng 27235d4f98a2SYan Zheng BUG_ON(block->key_ready); 27245d4f98a2SYan Zheng eb = read_tree_block(rc->extent_root, block->bytenr, 27255d4f98a2SYan Zheng block->key.objectid, block->key.offset); 272697d9a8a4STsutomu Itoh BUG_ON(!eb); 27275d4f98a2SYan Zheng WARN_ON(btrfs_header_level(eb) != block->level); 27285d4f98a2SYan Zheng if (block->level == 0) 27295d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 27305d4f98a2SYan Zheng else 27315d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 27325d4f98a2SYan Zheng free_extent_buffer(eb); 27335d4f98a2SYan Zheng block->key_ready = 1; 27345d4f98a2SYan Zheng return 0; 27355d4f98a2SYan Zheng } 27365d4f98a2SYan Zheng 27375d4f98a2SYan Zheng static int reada_tree_block(struct reloc_control *rc, 27385d4f98a2SYan Zheng struct tree_block *block) 27395d4f98a2SYan Zheng { 27405d4f98a2SYan Zheng BUG_ON(block->key_ready); 27415d4f98a2SYan Zheng readahead_tree_block(rc->extent_root, block->bytenr, 27425d4f98a2SYan Zheng block->key.objectid, block->key.offset); 27435d4f98a2SYan Zheng return 0; 27445d4f98a2SYan Zheng } 27455d4f98a2SYan Zheng 27465d4f98a2SYan Zheng /* 27475d4f98a2SYan Zheng * helper function to relocate a tree block 27485d4f98a2SYan Zheng */ 27495d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 27505d4f98a2SYan Zheng struct reloc_control *rc, 27515d4f98a2SYan Zheng struct backref_node *node, 27525d4f98a2SYan Zheng struct btrfs_key *key, 27535d4f98a2SYan Zheng struct btrfs_path *path) 27545d4f98a2SYan Zheng { 27555d4f98a2SYan Zheng struct btrfs_root *root; 27563fd0a558SYan, Zheng int release = 0; 27573fd0a558SYan, Zheng int ret = 0; 27585d4f98a2SYan Zheng 27593fd0a558SYan, Zheng if (!node) 27605d4f98a2SYan Zheng return 0; 27613fd0a558SYan, Zheng 27623fd0a558SYan, Zheng BUG_ON(node->processed); 27633fd0a558SYan, Zheng root = select_one_root(trans, node); 27643fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 27653fd0a558SYan, Zheng update_processed_blocks(rc, node); 27663fd0a558SYan, Zheng goto out; 27675d4f98a2SYan Zheng } 27685d4f98a2SYan Zheng 27693fd0a558SYan, Zheng if (!root || root->ref_cows) { 27703fd0a558SYan, Zheng ret = reserve_metadata_space(trans, rc, node); 27713fd0a558SYan, Zheng if (ret) 27725d4f98a2SYan Zheng goto out; 27733fd0a558SYan, Zheng release = 1; 27745d4f98a2SYan Zheng } 27753fd0a558SYan, Zheng 27763fd0a558SYan, Zheng if (root) { 27773fd0a558SYan, Zheng if (root->ref_cows) { 27783fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 27793fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 27803fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 27813fd0a558SYan, Zheng root = root->reloc_root; 27823fd0a558SYan, Zheng node->new_bytenr = root->node->start; 27833fd0a558SYan, Zheng node->root = root; 27843fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 27853fd0a558SYan, Zheng } else { 27865d4f98a2SYan Zheng path->lowest_level = node->level; 27875d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2788b3b4aa74SDavid Sterba btrfs_release_path(path); 27893fd0a558SYan, Zheng if (ret > 0) 27905d4f98a2SYan Zheng ret = 0; 27913fd0a558SYan, Zheng } 27923fd0a558SYan, Zheng if (!ret) 27933fd0a558SYan, Zheng update_processed_blocks(rc, node); 27943fd0a558SYan, Zheng } else { 27953fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 27963fd0a558SYan, Zheng } 27975d4f98a2SYan Zheng out: 27983fd0a558SYan, Zheng if (ret || node->level == 0 || node->cowonly) { 27993fd0a558SYan, Zheng if (release) 28003fd0a558SYan, Zheng release_metadata_space(rc, node); 28013fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 28023fd0a558SYan, Zheng } 28035d4f98a2SYan Zheng return ret; 28045d4f98a2SYan Zheng } 28055d4f98a2SYan Zheng 28065d4f98a2SYan Zheng /* 28075d4f98a2SYan Zheng * relocate a list of blocks 28085d4f98a2SYan Zheng */ 28095d4f98a2SYan Zheng static noinline_for_stack 28105d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 28115d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 28125d4f98a2SYan Zheng { 28135d4f98a2SYan Zheng struct backref_node *node; 28145d4f98a2SYan Zheng struct btrfs_path *path; 28155d4f98a2SYan Zheng struct tree_block *block; 28165d4f98a2SYan Zheng struct rb_node *rb_node; 28175d4f98a2SYan Zheng int ret; 28185d4f98a2SYan Zheng int err = 0; 28195d4f98a2SYan Zheng 28205d4f98a2SYan Zheng path = btrfs_alloc_path(); 28215d4f98a2SYan Zheng if (!path) 28225d4f98a2SYan Zheng return -ENOMEM; 28235d4f98a2SYan Zheng 28245d4f98a2SYan Zheng rb_node = rb_first(blocks); 28255d4f98a2SYan Zheng while (rb_node) { 28265d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 28275d4f98a2SYan Zheng if (!block->key_ready) 28285d4f98a2SYan Zheng reada_tree_block(rc, block); 28295d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28305d4f98a2SYan Zheng } 28315d4f98a2SYan Zheng 28325d4f98a2SYan Zheng rb_node = rb_first(blocks); 28335d4f98a2SYan Zheng while (rb_node) { 28345d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 28355d4f98a2SYan Zheng if (!block->key_ready) 28365d4f98a2SYan Zheng get_tree_block_key(rc, block); 28375d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28385d4f98a2SYan Zheng } 28395d4f98a2SYan Zheng 28405d4f98a2SYan Zheng rb_node = rb_first(blocks); 28415d4f98a2SYan Zheng while (rb_node) { 28425d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 28435d4f98a2SYan Zheng 28443fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 28455d4f98a2SYan Zheng block->level, block->bytenr); 28465d4f98a2SYan Zheng if (IS_ERR(node)) { 28475d4f98a2SYan Zheng err = PTR_ERR(node); 28485d4f98a2SYan Zheng goto out; 28495d4f98a2SYan Zheng } 28505d4f98a2SYan Zheng 28515d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 28525d4f98a2SYan Zheng path); 28535d4f98a2SYan Zheng if (ret < 0) { 28543fd0a558SYan, Zheng if (ret != -EAGAIN || rb_node == rb_first(blocks)) 28555d4f98a2SYan Zheng err = ret; 28565d4f98a2SYan Zheng goto out; 28575d4f98a2SYan Zheng } 28585d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28595d4f98a2SYan Zheng } 28605d4f98a2SYan Zheng out: 28615d4f98a2SYan Zheng free_block_list(blocks); 28623fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 28635d4f98a2SYan Zheng 28645d4f98a2SYan Zheng btrfs_free_path(path); 28655d4f98a2SYan Zheng return err; 28665d4f98a2SYan Zheng } 28675d4f98a2SYan Zheng 28685d4f98a2SYan Zheng static noinline_for_stack 2869efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 2870efa56464SYan, Zheng struct file_extent_cluster *cluster) 2871efa56464SYan, Zheng { 2872efa56464SYan, Zheng u64 alloc_hint = 0; 2873efa56464SYan, Zheng u64 start; 2874efa56464SYan, Zheng u64 end; 2875efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 2876efa56464SYan, Zheng u64 num_bytes; 2877efa56464SYan, Zheng int nr = 0; 2878efa56464SYan, Zheng int ret = 0; 2879efa56464SYan, Zheng 2880efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2881efa56464SYan, Zheng mutex_lock(&inode->i_mutex); 2882efa56464SYan, Zheng 2883efa56464SYan, Zheng ret = btrfs_check_data_free_space(inode, cluster->end + 2884efa56464SYan, Zheng 1 - cluster->start); 2885efa56464SYan, Zheng if (ret) 2886efa56464SYan, Zheng goto out; 2887efa56464SYan, Zheng 2888efa56464SYan, Zheng while (nr < cluster->nr) { 2889efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2890efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2891efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2892efa56464SYan, Zheng else 2893efa56464SYan, Zheng end = cluster->end - offset; 2894efa56464SYan, Zheng 2895d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 2896efa56464SYan, Zheng num_bytes = end + 1 - start; 2897efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 2898efa56464SYan, Zheng num_bytes, num_bytes, 2899efa56464SYan, Zheng end + 1, &alloc_hint); 2900d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 2901efa56464SYan, Zheng if (ret) 2902efa56464SYan, Zheng break; 2903efa56464SYan, Zheng nr++; 2904efa56464SYan, Zheng } 2905efa56464SYan, Zheng btrfs_free_reserved_data_space(inode, cluster->end + 2906efa56464SYan, Zheng 1 - cluster->start); 2907efa56464SYan, Zheng out: 2908efa56464SYan, Zheng mutex_unlock(&inode->i_mutex); 2909efa56464SYan, Zheng return ret; 2910efa56464SYan, Zheng } 2911efa56464SYan, Zheng 2912efa56464SYan, Zheng static noinline_for_stack 29130257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 29140257bb82SYan, Zheng u64 block_start) 29155d4f98a2SYan Zheng { 29165d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 29175d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 29185d4f98a2SYan Zheng struct extent_map *em; 29190257bb82SYan, Zheng int ret = 0; 29205d4f98a2SYan Zheng 2921172ddd60SDavid Sterba em = alloc_extent_map(); 29220257bb82SYan, Zheng if (!em) 29230257bb82SYan, Zheng return -ENOMEM; 29240257bb82SYan, Zheng 29255d4f98a2SYan Zheng em->start = start; 29260257bb82SYan, Zheng em->len = end + 1 - start; 29270257bb82SYan, Zheng em->block_len = em->len; 29280257bb82SYan, Zheng em->block_start = block_start; 29295d4f98a2SYan Zheng em->bdev = root->fs_info->fs_devices->latest_bdev; 29305d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 29315d4f98a2SYan Zheng 2932d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, start, end); 29335d4f98a2SYan Zheng while (1) { 2934890871beSChris Mason write_lock(&em_tree->lock); 29355d4f98a2SYan Zheng ret = add_extent_mapping(em_tree, em); 2936890871beSChris Mason write_unlock(&em_tree->lock); 29375d4f98a2SYan Zheng if (ret != -EEXIST) { 29385d4f98a2SYan Zheng free_extent_map(em); 29395d4f98a2SYan Zheng break; 29405d4f98a2SYan Zheng } 29415d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 0); 29425d4f98a2SYan Zheng } 2943d0082371SJeff Mahoney unlock_extent(&BTRFS_I(inode)->io_tree, start, end); 29440257bb82SYan, Zheng return ret; 29450257bb82SYan, Zheng } 29465d4f98a2SYan Zheng 29470257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 29480257bb82SYan, Zheng struct file_extent_cluster *cluster) 29490257bb82SYan, Zheng { 29500257bb82SYan, Zheng u64 page_start; 29510257bb82SYan, Zheng u64 page_end; 29520257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 29530257bb82SYan, Zheng unsigned long index; 29540257bb82SYan, Zheng unsigned long last_index; 29550257bb82SYan, Zheng struct page *page; 29560257bb82SYan, Zheng struct file_ra_state *ra; 29573b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 29580257bb82SYan, Zheng int nr = 0; 29590257bb82SYan, Zheng int ret = 0; 29600257bb82SYan, Zheng 29610257bb82SYan, Zheng if (!cluster->nr) 29620257bb82SYan, Zheng return 0; 29630257bb82SYan, Zheng 29640257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 29650257bb82SYan, Zheng if (!ra) 29660257bb82SYan, Zheng return -ENOMEM; 29670257bb82SYan, Zheng 2968efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 29690257bb82SYan, Zheng if (ret) 2970efa56464SYan, Zheng goto out; 29710257bb82SYan, Zheng 29720257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 29730257bb82SYan, Zheng 2974efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2975efa56464SYan, Zheng cluster->end - offset, cluster->start); 2976efa56464SYan, Zheng if (ret) 2977efa56464SYan, Zheng goto out; 2978efa56464SYan, Zheng 2979efa56464SYan, Zheng index = (cluster->start - offset) >> PAGE_CACHE_SHIFT; 2980efa56464SYan, Zheng last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT; 29810257bb82SYan, Zheng while (index <= last_index) { 2982efa56464SYan, Zheng ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE); 2983efa56464SYan, Zheng if (ret) 2984efa56464SYan, Zheng goto out; 2985efa56464SYan, Zheng 29860257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 29870257bb82SYan, Zheng if (!page) { 29880257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 29890257bb82SYan, Zheng ra, NULL, index, 29900257bb82SYan, Zheng last_index + 1 - index); 2991a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 29923b16a4e3SJosef Bacik mask); 29930257bb82SYan, Zheng if (!page) { 2994efa56464SYan, Zheng btrfs_delalloc_release_metadata(inode, 2995efa56464SYan, Zheng PAGE_CACHE_SIZE); 29960257bb82SYan, Zheng ret = -ENOMEM; 2997efa56464SYan, Zheng goto out; 29980257bb82SYan, Zheng } 29990257bb82SYan, Zheng } 30000257bb82SYan, Zheng 30010257bb82SYan, Zheng if (PageReadahead(page)) { 30020257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 30030257bb82SYan, Zheng ra, NULL, page, index, 30040257bb82SYan, Zheng last_index + 1 - index); 30050257bb82SYan, Zheng } 30060257bb82SYan, Zheng 30070257bb82SYan, Zheng if (!PageUptodate(page)) { 30080257bb82SYan, Zheng btrfs_readpage(NULL, page); 30090257bb82SYan, Zheng lock_page(page); 30100257bb82SYan, Zheng if (!PageUptodate(page)) { 30110257bb82SYan, Zheng unlock_page(page); 30120257bb82SYan, Zheng page_cache_release(page); 3013efa56464SYan, Zheng btrfs_delalloc_release_metadata(inode, 3014efa56464SYan, Zheng PAGE_CACHE_SIZE); 30150257bb82SYan, Zheng ret = -EIO; 3016efa56464SYan, Zheng goto out; 30170257bb82SYan, Zheng } 30180257bb82SYan, Zheng } 30190257bb82SYan, Zheng 30200257bb82SYan, Zheng page_start = (u64)page->index << PAGE_CACHE_SHIFT; 30210257bb82SYan, Zheng page_end = page_start + PAGE_CACHE_SIZE - 1; 30220257bb82SYan, Zheng 3023d0082371SJeff Mahoney lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end); 30240257bb82SYan, Zheng 30250257bb82SYan, Zheng set_page_extent_mapped(page); 30260257bb82SYan, Zheng 30270257bb82SYan, Zheng if (nr < cluster->nr && 30280257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 30290257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 30300257bb82SYan, Zheng page_start, page_end, 30310257bb82SYan, Zheng EXTENT_BOUNDARY, GFP_NOFS); 30320257bb82SYan, Zheng nr++; 30330257bb82SYan, Zheng } 30340257bb82SYan, Zheng 3035efa56464SYan, Zheng btrfs_set_extent_delalloc(inode, page_start, page_end, NULL); 30360257bb82SYan, Zheng set_page_dirty(page); 30370257bb82SYan, Zheng 30380257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 3039d0082371SJeff Mahoney page_start, page_end); 30400257bb82SYan, Zheng unlock_page(page); 30410257bb82SYan, Zheng page_cache_release(page); 30420257bb82SYan, Zheng 30430257bb82SYan, Zheng index++; 3044efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 3045efa56464SYan, Zheng btrfs_throttle(BTRFS_I(inode)->root); 30460257bb82SYan, Zheng } 30470257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 3048efa56464SYan, Zheng out: 30490257bb82SYan, Zheng kfree(ra); 30500257bb82SYan, Zheng return ret; 30510257bb82SYan, Zheng } 30520257bb82SYan, Zheng 30530257bb82SYan, Zheng static noinline_for_stack 30540257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 30550257bb82SYan, Zheng struct file_extent_cluster *cluster) 30560257bb82SYan, Zheng { 30570257bb82SYan, Zheng int ret; 30580257bb82SYan, Zheng 30590257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 30600257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30610257bb82SYan, Zheng if (ret) 30620257bb82SYan, Zheng return ret; 30630257bb82SYan, Zheng cluster->nr = 0; 30640257bb82SYan, Zheng } 30650257bb82SYan, Zheng 30660257bb82SYan, Zheng if (!cluster->nr) 30670257bb82SYan, Zheng cluster->start = extent_key->objectid; 30680257bb82SYan, Zheng else 30690257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 30700257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 30710257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 30720257bb82SYan, Zheng cluster->nr++; 30730257bb82SYan, Zheng 30740257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 30750257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30760257bb82SYan, Zheng if (ret) 30770257bb82SYan, Zheng return ret; 30780257bb82SYan, Zheng cluster->nr = 0; 30790257bb82SYan, Zheng } 30800257bb82SYan, Zheng return 0; 30815d4f98a2SYan Zheng } 30825d4f98a2SYan Zheng 30835d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 30845d4f98a2SYan Zheng static int get_ref_objectid_v0(struct reloc_control *rc, 30855d4f98a2SYan Zheng struct btrfs_path *path, 30865d4f98a2SYan Zheng struct btrfs_key *extent_key, 30875d4f98a2SYan Zheng u64 *ref_objectid, int *path_change) 30885d4f98a2SYan Zheng { 30895d4f98a2SYan Zheng struct btrfs_key key; 30905d4f98a2SYan Zheng struct extent_buffer *leaf; 30915d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 30925d4f98a2SYan Zheng int ret; 30935d4f98a2SYan Zheng int slot; 30945d4f98a2SYan Zheng 30955d4f98a2SYan Zheng leaf = path->nodes[0]; 30965d4f98a2SYan Zheng slot = path->slots[0]; 30975d4f98a2SYan Zheng while (1) { 30985d4f98a2SYan Zheng if (slot >= btrfs_header_nritems(leaf)) { 30995d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 31005d4f98a2SYan Zheng if (ret < 0) 31015d4f98a2SYan Zheng return ret; 31025d4f98a2SYan Zheng BUG_ON(ret > 0); 31035d4f98a2SYan Zheng leaf = path->nodes[0]; 31045d4f98a2SYan Zheng slot = path->slots[0]; 31055d4f98a2SYan Zheng if (path_change) 31065d4f98a2SYan Zheng *path_change = 1; 31075d4f98a2SYan Zheng } 31085d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, slot); 31095d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 31105d4f98a2SYan Zheng return -ENOENT; 31115d4f98a2SYan Zheng 31125d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_REF_V0_KEY) { 31135d4f98a2SYan Zheng slot++; 31145d4f98a2SYan Zheng continue; 31155d4f98a2SYan Zheng } 31165d4f98a2SYan Zheng ref0 = btrfs_item_ptr(leaf, slot, 31175d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 31185d4f98a2SYan Zheng *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0); 31195d4f98a2SYan Zheng break; 31205d4f98a2SYan Zheng } 31215d4f98a2SYan Zheng return 0; 31225d4f98a2SYan Zheng } 31235d4f98a2SYan Zheng #endif 31245d4f98a2SYan Zheng 31255d4f98a2SYan Zheng /* 31265d4f98a2SYan Zheng * helper to add a tree block to the list. 31275d4f98a2SYan Zheng * the major work is getting the generation and level of the block 31285d4f98a2SYan Zheng */ 31295d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 31305d4f98a2SYan Zheng struct btrfs_key *extent_key, 31315d4f98a2SYan Zheng struct btrfs_path *path, 31325d4f98a2SYan Zheng struct rb_root *blocks) 31335d4f98a2SYan Zheng { 31345d4f98a2SYan Zheng struct extent_buffer *eb; 31355d4f98a2SYan Zheng struct btrfs_extent_item *ei; 31365d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 31375d4f98a2SYan Zheng struct tree_block *block; 31385d4f98a2SYan Zheng struct rb_node *rb_node; 31395d4f98a2SYan Zheng u32 item_size; 31405d4f98a2SYan Zheng int level = -1; 31415d4f98a2SYan Zheng int generation; 31425d4f98a2SYan Zheng 31435d4f98a2SYan Zheng eb = path->nodes[0]; 31445d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 31455d4f98a2SYan Zheng 31465d4f98a2SYan Zheng if (item_size >= sizeof(*ei) + sizeof(*bi)) { 31475d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 31485d4f98a2SYan Zheng struct btrfs_extent_item); 31495d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 31505d4f98a2SYan Zheng generation = btrfs_extent_generation(eb, ei); 31515d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 31525d4f98a2SYan Zheng } else { 31535d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 31545d4f98a2SYan Zheng u64 ref_owner; 31555d4f98a2SYan Zheng int ret; 31565d4f98a2SYan Zheng 31575d4f98a2SYan Zheng BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 31585d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, extent_key, 31595d4f98a2SYan Zheng &ref_owner, NULL); 3160411fc6bcSAndi Kleen if (ret < 0) 3161411fc6bcSAndi Kleen return ret; 31625d4f98a2SYan Zheng BUG_ON(ref_owner >= BTRFS_MAX_LEVEL); 31635d4f98a2SYan Zheng level = (int)ref_owner; 31645d4f98a2SYan Zheng /* FIXME: get real generation */ 31655d4f98a2SYan Zheng generation = 0; 31665d4f98a2SYan Zheng #else 31675d4f98a2SYan Zheng BUG(); 31685d4f98a2SYan Zheng #endif 31695d4f98a2SYan Zheng } 31705d4f98a2SYan Zheng 3171b3b4aa74SDavid Sterba btrfs_release_path(path); 31725d4f98a2SYan Zheng 31735d4f98a2SYan Zheng BUG_ON(level == -1); 31745d4f98a2SYan Zheng 31755d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 31765d4f98a2SYan Zheng if (!block) 31775d4f98a2SYan Zheng return -ENOMEM; 31785d4f98a2SYan Zheng 31795d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 31805d4f98a2SYan Zheng block->key.objectid = extent_key->offset; 31815d4f98a2SYan Zheng block->key.offset = generation; 31825d4f98a2SYan Zheng block->level = level; 31835d4f98a2SYan Zheng block->key_ready = 0; 31845d4f98a2SYan Zheng 31855d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 318643c04fb1SJeff Mahoney if (rb_node) 318743c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, block->bytenr); 31885d4f98a2SYan Zheng 31895d4f98a2SYan Zheng return 0; 31905d4f98a2SYan Zheng } 31915d4f98a2SYan Zheng 31925d4f98a2SYan Zheng /* 31935d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 31945d4f98a2SYan Zheng */ 31955d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 31965d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 31975d4f98a2SYan Zheng struct rb_root *blocks) 31985d4f98a2SYan Zheng { 31995d4f98a2SYan Zheng struct btrfs_path *path; 32005d4f98a2SYan Zheng struct btrfs_key key; 32015d4f98a2SYan Zheng int ret; 32025d4f98a2SYan Zheng 32035d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 32045d4f98a2SYan Zheng return 0; 32055d4f98a2SYan Zheng 32065d4f98a2SYan Zheng if (tree_search(blocks, bytenr)) 32075d4f98a2SYan Zheng return 0; 32085d4f98a2SYan Zheng 32095d4f98a2SYan Zheng path = btrfs_alloc_path(); 32105d4f98a2SYan Zheng if (!path) 32115d4f98a2SYan Zheng return -ENOMEM; 32125d4f98a2SYan Zheng 32135d4f98a2SYan Zheng key.objectid = bytenr; 32145d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 32155d4f98a2SYan Zheng key.offset = blocksize; 32165d4f98a2SYan Zheng 32175d4f98a2SYan Zheng path->search_commit_root = 1; 32185d4f98a2SYan Zheng path->skip_locking = 1; 32195d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 32205d4f98a2SYan Zheng if (ret < 0) 32215d4f98a2SYan Zheng goto out; 32225d4f98a2SYan Zheng BUG_ON(ret); 32235d4f98a2SYan Zheng 32245d4f98a2SYan Zheng btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 32255d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 32265d4f98a2SYan Zheng out: 32275d4f98a2SYan Zheng btrfs_free_path(path); 32285d4f98a2SYan Zheng return ret; 32295d4f98a2SYan Zheng } 32305d4f98a2SYan Zheng 32315d4f98a2SYan Zheng /* 32325d4f98a2SYan Zheng * helper to check if the block use full backrefs for pointers in it 32335d4f98a2SYan Zheng */ 32345d4f98a2SYan Zheng static int block_use_full_backref(struct reloc_control *rc, 32355d4f98a2SYan Zheng struct extent_buffer *eb) 32365d4f98a2SYan Zheng { 32375d4f98a2SYan Zheng u64 flags; 32385d4f98a2SYan Zheng int ret; 32395d4f98a2SYan Zheng 32405d4f98a2SYan Zheng if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) || 32415d4f98a2SYan Zheng btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV) 32425d4f98a2SYan Zheng return 1; 32435d4f98a2SYan Zheng 32443fd0a558SYan, Zheng ret = btrfs_lookup_extent_info(NULL, rc->extent_root, 32453fd0a558SYan, Zheng eb->start, eb->len, NULL, &flags); 32465d4f98a2SYan Zheng BUG_ON(ret); 32475d4f98a2SYan Zheng 32485d4f98a2SYan Zheng if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) 32495d4f98a2SYan Zheng ret = 1; 32505d4f98a2SYan Zheng else 32515d4f98a2SYan Zheng ret = 0; 32525d4f98a2SYan Zheng return ret; 32535d4f98a2SYan Zheng } 32545d4f98a2SYan Zheng 32550af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 32560af3d00bSJosef Bacik struct inode *inode, u64 ino) 32570af3d00bSJosef Bacik { 32580af3d00bSJosef Bacik struct btrfs_key key; 32590af3d00bSJosef Bacik struct btrfs_path *path; 32600af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 32610af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 32620af3d00bSJosef Bacik unsigned long nr; 32630af3d00bSJosef Bacik int ret = 0; 32640af3d00bSJosef Bacik 32650af3d00bSJosef Bacik if (inode) 32660af3d00bSJosef Bacik goto truncate; 32670af3d00bSJosef Bacik 32680af3d00bSJosef Bacik key.objectid = ino; 32690af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 32700af3d00bSJosef Bacik key.offset = 0; 32710af3d00bSJosef Bacik 32720af3d00bSJosef Bacik inode = btrfs_iget(fs_info->sb, &key, root, NULL); 3273f54fb859STsutomu Itoh if (IS_ERR(inode) || is_bad_inode(inode)) { 3274f54fb859STsutomu Itoh if (!IS_ERR(inode)) 32750af3d00bSJosef Bacik iput(inode); 32760af3d00bSJosef Bacik return -ENOENT; 32770af3d00bSJosef Bacik } 32780af3d00bSJosef Bacik 32790af3d00bSJosef Bacik truncate: 32800af3d00bSJosef Bacik path = btrfs_alloc_path(); 32810af3d00bSJosef Bacik if (!path) { 32820af3d00bSJosef Bacik ret = -ENOMEM; 32830af3d00bSJosef Bacik goto out; 32840af3d00bSJosef Bacik } 32850af3d00bSJosef Bacik 32867a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 32870af3d00bSJosef Bacik if (IS_ERR(trans)) { 32880af3d00bSJosef Bacik btrfs_free_path(path); 32893612b495STsutomu Itoh ret = PTR_ERR(trans); 32900af3d00bSJosef Bacik goto out; 32910af3d00bSJosef Bacik } 32920af3d00bSJosef Bacik 32930af3d00bSJosef Bacik ret = btrfs_truncate_free_space_cache(root, trans, path, inode); 32940af3d00bSJosef Bacik 32950af3d00bSJosef Bacik btrfs_free_path(path); 32960af3d00bSJosef Bacik nr = trans->blocks_used; 32970af3d00bSJosef Bacik btrfs_end_transaction(trans, root); 32980af3d00bSJosef Bacik btrfs_btree_balance_dirty(root, nr); 32990af3d00bSJosef Bacik out: 33000af3d00bSJosef Bacik iput(inode); 33010af3d00bSJosef Bacik return ret; 33020af3d00bSJosef Bacik } 33030af3d00bSJosef Bacik 33045d4f98a2SYan Zheng /* 33055d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY 33065d4f98a2SYan Zheng * this function scans fs tree to find blocks reference the data extent 33075d4f98a2SYan Zheng */ 33085d4f98a2SYan Zheng static int find_data_references(struct reloc_control *rc, 33095d4f98a2SYan Zheng struct btrfs_key *extent_key, 33105d4f98a2SYan Zheng struct extent_buffer *leaf, 33115d4f98a2SYan Zheng struct btrfs_extent_data_ref *ref, 33125d4f98a2SYan Zheng struct rb_root *blocks) 33135d4f98a2SYan Zheng { 33145d4f98a2SYan Zheng struct btrfs_path *path; 33155d4f98a2SYan Zheng struct tree_block *block; 33165d4f98a2SYan Zheng struct btrfs_root *root; 33175d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 33185d4f98a2SYan Zheng struct rb_node *rb_node; 33195d4f98a2SYan Zheng struct btrfs_key key; 33205d4f98a2SYan Zheng u64 ref_root; 33215d4f98a2SYan Zheng u64 ref_objectid; 33225d4f98a2SYan Zheng u64 ref_offset; 33235d4f98a2SYan Zheng u32 ref_count; 33245d4f98a2SYan Zheng u32 nritems; 33255d4f98a2SYan Zheng int err = 0; 33265d4f98a2SYan Zheng int added = 0; 33275d4f98a2SYan Zheng int counted; 33285d4f98a2SYan Zheng int ret; 33295d4f98a2SYan Zheng 33305d4f98a2SYan Zheng ref_root = btrfs_extent_data_ref_root(leaf, ref); 33315d4f98a2SYan Zheng ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref); 33325d4f98a2SYan Zheng ref_offset = btrfs_extent_data_ref_offset(leaf, ref); 33335d4f98a2SYan Zheng ref_count = btrfs_extent_data_ref_count(leaf, ref); 33345d4f98a2SYan Zheng 33350af3d00bSJosef Bacik /* 33360af3d00bSJosef Bacik * This is an extent belonging to the free space cache, lets just delete 33370af3d00bSJosef Bacik * it and redo the search. 33380af3d00bSJosef Bacik */ 33390af3d00bSJosef Bacik if (ref_root == BTRFS_ROOT_TREE_OBJECTID) { 33400af3d00bSJosef Bacik ret = delete_block_group_cache(rc->extent_root->fs_info, 33410af3d00bSJosef Bacik NULL, ref_objectid); 33420af3d00bSJosef Bacik if (ret != -ENOENT) 33430af3d00bSJosef Bacik return ret; 33440af3d00bSJosef Bacik ret = 0; 33450af3d00bSJosef Bacik } 33460af3d00bSJosef Bacik 33470af3d00bSJosef Bacik path = btrfs_alloc_path(); 33480af3d00bSJosef Bacik if (!path) 33490af3d00bSJosef Bacik return -ENOMEM; 3350026fd317SJosef Bacik path->reada = 1; 33510af3d00bSJosef Bacik 33525d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, ref_root); 33535d4f98a2SYan Zheng if (IS_ERR(root)) { 33545d4f98a2SYan Zheng err = PTR_ERR(root); 33555d4f98a2SYan Zheng goto out; 33565d4f98a2SYan Zheng } 33575d4f98a2SYan Zheng 33585d4f98a2SYan Zheng key.objectid = ref_objectid; 33595d4f98a2SYan Zheng key.type = BTRFS_EXTENT_DATA_KEY; 336084850e8dSYan, Zheng if (ref_offset > ((u64)-1 << 32)) 336184850e8dSYan, Zheng key.offset = 0; 336284850e8dSYan, Zheng else 336384850e8dSYan, Zheng key.offset = ref_offset; 33645d4f98a2SYan Zheng 33655d4f98a2SYan Zheng path->search_commit_root = 1; 33665d4f98a2SYan Zheng path->skip_locking = 1; 33675d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 33685d4f98a2SYan Zheng if (ret < 0) { 33695d4f98a2SYan Zheng err = ret; 33705d4f98a2SYan Zheng goto out; 33715d4f98a2SYan Zheng } 33725d4f98a2SYan Zheng 33735d4f98a2SYan Zheng leaf = path->nodes[0]; 33745d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 33755d4f98a2SYan Zheng /* 33765d4f98a2SYan Zheng * the references in tree blocks that use full backrefs 33775d4f98a2SYan Zheng * are not counted in 33785d4f98a2SYan Zheng */ 33795d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 33805d4f98a2SYan Zheng counted = 0; 33815d4f98a2SYan Zheng else 33825d4f98a2SYan Zheng counted = 1; 33835d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 33845d4f98a2SYan Zheng if (rb_node) { 33855d4f98a2SYan Zheng if (counted) 33865d4f98a2SYan Zheng added = 1; 33875d4f98a2SYan Zheng else 33885d4f98a2SYan Zheng path->slots[0] = nritems; 33895d4f98a2SYan Zheng } 33905d4f98a2SYan Zheng 33915d4f98a2SYan Zheng while (ref_count > 0) { 33925d4f98a2SYan Zheng while (path->slots[0] >= nritems) { 33935d4f98a2SYan Zheng ret = btrfs_next_leaf(root, path); 33945d4f98a2SYan Zheng if (ret < 0) { 33955d4f98a2SYan Zheng err = ret; 33965d4f98a2SYan Zheng goto out; 33975d4f98a2SYan Zheng } 33985d4f98a2SYan Zheng if (ret > 0) { 33995d4f98a2SYan Zheng WARN_ON(1); 34005d4f98a2SYan Zheng goto out; 34015d4f98a2SYan Zheng } 34025d4f98a2SYan Zheng 34035d4f98a2SYan Zheng leaf = path->nodes[0]; 34045d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 34055d4f98a2SYan Zheng added = 0; 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 34205d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 34215d4f98a2SYan Zheng if (key.objectid != ref_objectid || 34225d4f98a2SYan Zheng key.type != BTRFS_EXTENT_DATA_KEY) { 34235d4f98a2SYan Zheng WARN_ON(1); 34245d4f98a2SYan Zheng break; 34255d4f98a2SYan Zheng } 34265d4f98a2SYan Zheng 34275d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 34285d4f98a2SYan Zheng struct btrfs_file_extent_item); 34295d4f98a2SYan Zheng 34305d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 34315d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 34325d4f98a2SYan Zheng goto next; 34335d4f98a2SYan Zheng 34345d4f98a2SYan Zheng if (btrfs_file_extent_disk_bytenr(leaf, fi) != 34355d4f98a2SYan Zheng extent_key->objectid) 34365d4f98a2SYan Zheng goto next; 34375d4f98a2SYan Zheng 34385d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 34395d4f98a2SYan Zheng if (key.offset != ref_offset) 34405d4f98a2SYan Zheng goto next; 34415d4f98a2SYan Zheng 34425d4f98a2SYan Zheng if (counted) 34435d4f98a2SYan Zheng ref_count--; 34445d4f98a2SYan Zheng if (added) 34455d4f98a2SYan Zheng goto next; 34465d4f98a2SYan Zheng 34475d4f98a2SYan Zheng if (!tree_block_processed(leaf->start, leaf->len, rc)) { 34485d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 34495d4f98a2SYan Zheng if (!block) { 34505d4f98a2SYan Zheng err = -ENOMEM; 34515d4f98a2SYan Zheng break; 34525d4f98a2SYan Zheng } 34535d4f98a2SYan Zheng block->bytenr = leaf->start; 34545d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &block->key, 0); 34555d4f98a2SYan Zheng block->level = 0; 34565d4f98a2SYan Zheng block->key_ready = 1; 34575d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, 34585d4f98a2SYan Zheng &block->rb_node); 345943c04fb1SJeff Mahoney if (rb_node) 346043c04fb1SJeff Mahoney backref_tree_panic(rb_node, -EEXIST, 346143c04fb1SJeff Mahoney block->bytenr); 34625d4f98a2SYan Zheng } 34635d4f98a2SYan Zheng if (counted) 34645d4f98a2SYan Zheng added = 1; 34655d4f98a2SYan Zheng else 34665d4f98a2SYan Zheng path->slots[0] = nritems; 34675d4f98a2SYan Zheng next: 34685d4f98a2SYan Zheng path->slots[0]++; 34695d4f98a2SYan Zheng 34705d4f98a2SYan Zheng } 34715d4f98a2SYan Zheng out: 34725d4f98a2SYan Zheng btrfs_free_path(path); 34735d4f98a2SYan Zheng return err; 34745d4f98a2SYan Zheng } 34755d4f98a2SYan Zheng 34765d4f98a2SYan Zheng /* 34775d4f98a2SYan Zheng * hepler to find all tree blocks that reference a given data extent 34785d4f98a2SYan Zheng */ 34795d4f98a2SYan Zheng static noinline_for_stack 34805d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 34815d4f98a2SYan Zheng struct btrfs_key *extent_key, 34825d4f98a2SYan Zheng struct btrfs_path *path, 34835d4f98a2SYan Zheng struct rb_root *blocks) 34845d4f98a2SYan Zheng { 34855d4f98a2SYan Zheng struct btrfs_key key; 34865d4f98a2SYan Zheng struct extent_buffer *eb; 34875d4f98a2SYan Zheng struct btrfs_extent_data_ref *dref; 34885d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 34895d4f98a2SYan Zheng unsigned long ptr; 34905d4f98a2SYan Zheng unsigned long end; 34913fd0a558SYan, Zheng u32 blocksize = btrfs_level_size(rc->extent_root, 0); 34925d4f98a2SYan Zheng int ret; 34935d4f98a2SYan Zheng int err = 0; 34945d4f98a2SYan Zheng 34955d4f98a2SYan Zheng eb = path->nodes[0]; 34965d4f98a2SYan Zheng ptr = btrfs_item_ptr_offset(eb, path->slots[0]); 34975d4f98a2SYan Zheng end = ptr + btrfs_item_size_nr(eb, path->slots[0]); 34985d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 34995d4f98a2SYan Zheng if (ptr + sizeof(struct btrfs_extent_item_v0) == end) 35005d4f98a2SYan Zheng ptr = end; 35015d4f98a2SYan Zheng else 35025d4f98a2SYan Zheng #endif 35035d4f98a2SYan Zheng ptr += sizeof(struct btrfs_extent_item); 35045d4f98a2SYan Zheng 35055d4f98a2SYan Zheng while (ptr < end) { 35065d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 35075d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 35085d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 35095d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 35105d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 35115d4f98a2SYan Zheng blocks); 35125d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 35135d4f98a2SYan Zheng dref = (struct btrfs_extent_data_ref *)(&iref->offset); 35145d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 35155d4f98a2SYan Zheng eb, dref, blocks); 35165d4f98a2SYan Zheng } else { 35175d4f98a2SYan Zheng BUG(); 35185d4f98a2SYan Zheng } 35195d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 35205d4f98a2SYan Zheng } 35215d4f98a2SYan Zheng WARN_ON(ptr > end); 35225d4f98a2SYan Zheng 35235d4f98a2SYan Zheng while (1) { 35245d4f98a2SYan Zheng cond_resched(); 35255d4f98a2SYan Zheng eb = path->nodes[0]; 35265d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(eb)) { 35275d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 35285d4f98a2SYan Zheng if (ret < 0) { 35295d4f98a2SYan Zheng err = ret; 35305d4f98a2SYan Zheng break; 35315d4f98a2SYan Zheng } 35325d4f98a2SYan Zheng if (ret > 0) 35335d4f98a2SYan Zheng break; 35345d4f98a2SYan Zheng eb = path->nodes[0]; 35355d4f98a2SYan Zheng } 35365d4f98a2SYan Zheng 35375d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path->slots[0]); 35385d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 35395d4f98a2SYan Zheng break; 35405d4f98a2SYan Zheng 35415d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 35425d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY || 35435d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 35445d4f98a2SYan Zheng #else 35455d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 35465d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 35475d4f98a2SYan Zheng #endif 35485d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 35495d4f98a2SYan Zheng blocks); 35505d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 35515d4f98a2SYan Zheng dref = btrfs_item_ptr(eb, path->slots[0], 35525d4f98a2SYan Zheng struct btrfs_extent_data_ref); 35535d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 35545d4f98a2SYan Zheng eb, dref, blocks); 35555d4f98a2SYan Zheng } else { 35565d4f98a2SYan Zheng ret = 0; 35575d4f98a2SYan Zheng } 35585d4f98a2SYan Zheng if (ret) { 35595d4f98a2SYan Zheng err = ret; 35605d4f98a2SYan Zheng break; 35615d4f98a2SYan Zheng } 35625d4f98a2SYan Zheng path->slots[0]++; 35635d4f98a2SYan Zheng } 3564b3b4aa74SDavid Sterba btrfs_release_path(path); 35655d4f98a2SYan Zheng if (err) 35665d4f98a2SYan Zheng free_block_list(blocks); 35675d4f98a2SYan Zheng return err; 35685d4f98a2SYan Zheng } 35695d4f98a2SYan Zheng 35705d4f98a2SYan Zheng /* 35715d4f98a2SYan Zheng * hepler to find next unprocessed extent 35725d4f98a2SYan Zheng */ 35735d4f98a2SYan Zheng static noinline_for_stack 35745d4f98a2SYan Zheng int find_next_extent(struct btrfs_trans_handle *trans, 35753fd0a558SYan, Zheng struct reloc_control *rc, struct btrfs_path *path, 35763fd0a558SYan, Zheng struct btrfs_key *extent_key) 35775d4f98a2SYan Zheng { 35785d4f98a2SYan Zheng struct btrfs_key key; 35795d4f98a2SYan Zheng struct extent_buffer *leaf; 35805d4f98a2SYan Zheng u64 start, end, last; 35815d4f98a2SYan Zheng int ret; 35825d4f98a2SYan Zheng 35835d4f98a2SYan Zheng last = rc->block_group->key.objectid + rc->block_group->key.offset; 35845d4f98a2SYan Zheng while (1) { 35855d4f98a2SYan Zheng cond_resched(); 35865d4f98a2SYan Zheng if (rc->search_start >= last) { 35875d4f98a2SYan Zheng ret = 1; 35885d4f98a2SYan Zheng break; 35895d4f98a2SYan Zheng } 35905d4f98a2SYan Zheng 35915d4f98a2SYan Zheng key.objectid = rc->search_start; 35925d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 35935d4f98a2SYan Zheng key.offset = 0; 35945d4f98a2SYan Zheng 35955d4f98a2SYan Zheng path->search_commit_root = 1; 35965d4f98a2SYan Zheng path->skip_locking = 1; 35975d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 35985d4f98a2SYan Zheng 0, 0); 35995d4f98a2SYan Zheng if (ret < 0) 36005d4f98a2SYan Zheng break; 36015d4f98a2SYan Zheng next: 36025d4f98a2SYan Zheng leaf = path->nodes[0]; 36035d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 36045d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 36055d4f98a2SYan Zheng if (ret != 0) 36065d4f98a2SYan Zheng break; 36075d4f98a2SYan Zheng leaf = path->nodes[0]; 36085d4f98a2SYan Zheng } 36095d4f98a2SYan Zheng 36105d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 36115d4f98a2SYan Zheng if (key.objectid >= last) { 36125d4f98a2SYan Zheng ret = 1; 36135d4f98a2SYan Zheng break; 36145d4f98a2SYan Zheng } 36155d4f98a2SYan Zheng 36165d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_ITEM_KEY || 36175d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 36185d4f98a2SYan Zheng path->slots[0]++; 36195d4f98a2SYan Zheng goto next; 36205d4f98a2SYan Zheng } 36215d4f98a2SYan Zheng 36225d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 36235d4f98a2SYan Zheng key.objectid, &start, &end, 3624*e6138876SJosef Bacik EXTENT_DIRTY, NULL); 36255d4f98a2SYan Zheng 36265d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3627b3b4aa74SDavid Sterba btrfs_release_path(path); 36285d4f98a2SYan Zheng rc->search_start = end + 1; 36295d4f98a2SYan Zheng } else { 36305d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 36313fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 36325d4f98a2SYan Zheng return 0; 36335d4f98a2SYan Zheng } 36345d4f98a2SYan Zheng } 3635b3b4aa74SDavid Sterba btrfs_release_path(path); 36365d4f98a2SYan Zheng return ret; 36375d4f98a2SYan Zheng } 36385d4f98a2SYan Zheng 36395d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 36405d4f98a2SYan Zheng { 36415d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36427585717fSChris Mason 36437585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 36445d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 36457585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 36465d4f98a2SYan Zheng } 36475d4f98a2SYan Zheng 36485d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 36495d4f98a2SYan Zheng { 36505d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36517585717fSChris Mason 36527585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 36535d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 36547585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 36555d4f98a2SYan Zheng } 36565d4f98a2SYan Zheng 36575d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 36585d4f98a2SYan Zheng { 36595d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 36605d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 36615d4f98a2SYan Zheng return 1; 36625d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 36635d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 36645d4f98a2SYan Zheng return 1; 36655d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 36665d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 36675d4f98a2SYan Zheng return 1; 36685d4f98a2SYan Zheng return 0; 36695d4f98a2SYan Zheng } 36705d4f98a2SYan Zheng 36713fd0a558SYan, Zheng static noinline_for_stack 36723fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 36733fd0a558SYan, Zheng { 36743fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 36753fd0a558SYan, Zheng int ret; 36763fd0a558SYan, Zheng 367766d8f3ddSMiao Xie rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root, 367866d8f3ddSMiao Xie BTRFS_BLOCK_RSV_TEMP); 36793fd0a558SYan, Zheng if (!rc->block_rsv) 36803fd0a558SYan, Zheng return -ENOMEM; 36813fd0a558SYan, Zheng 36823fd0a558SYan, Zheng /* 36833fd0a558SYan, Zheng * reserve some space for creating reloc trees. 36843fd0a558SYan, Zheng * btrfs_init_reloc_root will use them when there 36853fd0a558SYan, Zheng * is no reservation in transaction handle. 36863fd0a558SYan, Zheng */ 36874a92b1b8SJosef Bacik ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv, 36888bb8ab2eSJosef Bacik rc->extent_root->nodesize * 256); 36893fd0a558SYan, Zheng if (ret) 36903fd0a558SYan, Zheng return ret; 36913fd0a558SYan, Zheng 36923fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 36933fd0a558SYan, Zheng rc->search_start = rc->block_group->key.objectid; 36943fd0a558SYan, Zheng rc->extents_found = 0; 36953fd0a558SYan, Zheng rc->nodes_relocated = 0; 36963fd0a558SYan, Zheng rc->merging_rsv_size = 0; 36973fd0a558SYan, Zheng 36983fd0a558SYan, Zheng rc->create_reloc_tree = 1; 36993fd0a558SYan, Zheng set_reloc_control(rc); 37003fd0a558SYan, Zheng 37017a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 37023612b495STsutomu Itoh BUG_ON(IS_ERR(trans)); 37033fd0a558SYan, Zheng btrfs_commit_transaction(trans, rc->extent_root); 37043fd0a558SYan, Zheng return 0; 37053fd0a558SYan, Zheng } 370676dda93cSYan, Zheng 37075d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 37085d4f98a2SYan Zheng { 37095d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 37105d4f98a2SYan Zheng struct btrfs_key key; 37115d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 37125d4f98a2SYan Zheng struct btrfs_path *path; 37135d4f98a2SYan Zheng struct btrfs_extent_item *ei; 37145d4f98a2SYan Zheng unsigned long nr; 37155d4f98a2SYan Zheng u64 flags; 37165d4f98a2SYan Zheng u32 item_size; 37175d4f98a2SYan Zheng int ret; 37185d4f98a2SYan Zheng int err = 0; 3719c87f08caSChris Mason int progress = 0; 37205d4f98a2SYan Zheng 37215d4f98a2SYan Zheng path = btrfs_alloc_path(); 37223fd0a558SYan, Zheng if (!path) 37235d4f98a2SYan Zheng return -ENOMEM; 3724026fd317SJosef Bacik path->reada = 1; 37253fd0a558SYan, Zheng 37263fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 37273fd0a558SYan, Zheng if (ret) { 37283fd0a558SYan, Zheng err = ret; 37293fd0a558SYan, Zheng goto out_free; 37302423fdfbSJiri Slaby } 37315d4f98a2SYan Zheng 37325d4f98a2SYan Zheng while (1) { 3733c87f08caSChris Mason progress++; 3734a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 373598d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 3736c87f08caSChris Mason restart: 37373fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 37383fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 37393fd0a558SYan, Zheng continue; 37403fd0a558SYan, Zheng } 37413fd0a558SYan, Zheng 37423fd0a558SYan, Zheng ret = find_next_extent(trans, rc, path, &key); 37435d4f98a2SYan Zheng if (ret < 0) 37445d4f98a2SYan Zheng err = ret; 37455d4f98a2SYan Zheng if (ret != 0) 37465d4f98a2SYan Zheng break; 37475d4f98a2SYan Zheng 37485d4f98a2SYan Zheng rc->extents_found++; 37495d4f98a2SYan Zheng 37505d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 37515d4f98a2SYan Zheng struct btrfs_extent_item); 37523fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 37535d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 37545d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 37555d4f98a2SYan Zheng ret = check_extent_flags(flags); 37565d4f98a2SYan Zheng BUG_ON(ret); 37575d4f98a2SYan Zheng 37585d4f98a2SYan Zheng } else { 37595d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 37605d4f98a2SYan Zheng u64 ref_owner; 37615d4f98a2SYan Zheng int path_change = 0; 37625d4f98a2SYan Zheng 37635d4f98a2SYan Zheng BUG_ON(item_size != 37645d4f98a2SYan Zheng sizeof(struct btrfs_extent_item_v0)); 37655d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, &key, &ref_owner, 37665d4f98a2SYan Zheng &path_change); 37675d4f98a2SYan Zheng if (ref_owner < BTRFS_FIRST_FREE_OBJECTID) 37685d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_TREE_BLOCK; 37695d4f98a2SYan Zheng else 37705d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_DATA; 37715d4f98a2SYan Zheng 37725d4f98a2SYan Zheng if (path_change) { 3773b3b4aa74SDavid Sterba btrfs_release_path(path); 37745d4f98a2SYan Zheng 37755d4f98a2SYan Zheng path->search_commit_root = 1; 37765d4f98a2SYan Zheng path->skip_locking = 1; 37775d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, 37785d4f98a2SYan Zheng &key, path, 0, 0); 37795d4f98a2SYan Zheng if (ret < 0) { 37805d4f98a2SYan Zheng err = ret; 37815d4f98a2SYan Zheng break; 37825d4f98a2SYan Zheng } 37835d4f98a2SYan Zheng BUG_ON(ret > 0); 37845d4f98a2SYan Zheng } 37855d4f98a2SYan Zheng #else 37865d4f98a2SYan Zheng BUG(); 37875d4f98a2SYan Zheng #endif 37885d4f98a2SYan Zheng } 37895d4f98a2SYan Zheng 37905d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 37915d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 37925d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 37935d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 37945d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 37955d4f98a2SYan Zheng } else { 3796b3b4aa74SDavid Sterba btrfs_release_path(path); 37975d4f98a2SYan Zheng ret = 0; 37985d4f98a2SYan Zheng } 37995d4f98a2SYan Zheng if (ret < 0) { 38003fd0a558SYan, Zheng err = ret; 38015d4f98a2SYan Zheng break; 38025d4f98a2SYan Zheng } 38035d4f98a2SYan Zheng 38045d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 38055d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 38065d4f98a2SYan Zheng if (ret < 0) { 38073fd0a558SYan, Zheng if (ret != -EAGAIN) { 38085d4f98a2SYan Zheng err = ret; 38095d4f98a2SYan Zheng break; 38105d4f98a2SYan Zheng } 38113fd0a558SYan, Zheng rc->extents_found--; 38123fd0a558SYan, Zheng rc->search_start = key.objectid; 38133fd0a558SYan, Zheng } 38145d4f98a2SYan Zheng } 38155d4f98a2SYan Zheng 381636ba022aSJosef Bacik ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5); 38173fd0a558SYan, Zheng if (ret < 0) { 38187654b724SDaniel J Blueman if (ret != -ENOSPC) { 38193fd0a558SYan, Zheng err = ret; 38203fd0a558SYan, Zheng WARN_ON(1); 38213fd0a558SYan, Zheng break; 38223fd0a558SYan, Zheng } 38233fd0a558SYan, Zheng rc->commit_transaction = 1; 38243fd0a558SYan, Zheng } 38253fd0a558SYan, Zheng 38263fd0a558SYan, Zheng if (rc->commit_transaction) { 38273fd0a558SYan, Zheng rc->commit_transaction = 0; 38283fd0a558SYan, Zheng ret = btrfs_commit_transaction(trans, rc->extent_root); 38293fd0a558SYan, Zheng BUG_ON(ret); 38303fd0a558SYan, Zheng } else { 38315d4f98a2SYan Zheng nr = trans->blocks_used; 38323fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 38335d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 38343fd0a558SYan, Zheng } 38353fd0a558SYan, Zheng trans = NULL; 38365d4f98a2SYan Zheng 38375d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 38385d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 38395d4f98a2SYan Zheng rc->found_file_extent = 1; 38400257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 38413fd0a558SYan, Zheng &key, &rc->cluster); 38425d4f98a2SYan Zheng if (ret < 0) { 38435d4f98a2SYan Zheng err = ret; 38445d4f98a2SYan Zheng break; 38455d4f98a2SYan Zheng } 38465d4f98a2SYan Zheng } 38475d4f98a2SYan Zheng } 3848c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 3849c87f08caSChris Mason ret = btrfs_force_chunk_alloc(trans, rc->extent_root, 3850c87f08caSChris Mason rc->block_group->flags); 3851c87f08caSChris Mason if (ret == 0) { 3852c87f08caSChris Mason err = 0; 3853c87f08caSChris Mason progress = 0; 3854c87f08caSChris Mason goto restart; 3855c87f08caSChris Mason } 3856c87f08caSChris Mason } 38573fd0a558SYan, Zheng 3858b3b4aa74SDavid Sterba btrfs_release_path(path); 38593fd0a558SYan, Zheng clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY, 38603fd0a558SYan, Zheng GFP_NOFS); 38615d4f98a2SYan Zheng 38625d4f98a2SYan Zheng if (trans) { 38635d4f98a2SYan Zheng nr = trans->blocks_used; 38643fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 38655d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 38665d4f98a2SYan Zheng } 38675d4f98a2SYan Zheng 38680257bb82SYan, Zheng if (!err) { 38693fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 38703fd0a558SYan, Zheng &rc->cluster); 38710257bb82SYan, Zheng if (ret < 0) 38720257bb82SYan, Zheng err = ret; 38730257bb82SYan, Zheng } 38740257bb82SYan, Zheng 38753fd0a558SYan, Zheng rc->create_reloc_tree = 0; 38763fd0a558SYan, Zheng set_reloc_control(rc); 38770257bb82SYan, Zheng 38783fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 38793fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1); 38805d4f98a2SYan Zheng 38813fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 38825d4f98a2SYan Zheng 38835d4f98a2SYan Zheng merge_reloc_roots(rc); 38845d4f98a2SYan Zheng 38853fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 38865d4f98a2SYan Zheng unset_reloc_control(rc); 38873fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1); 38885d4f98a2SYan Zheng 38895d4f98a2SYan Zheng /* get rid of pinned extents */ 38907a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 38913612b495STsutomu Itoh if (IS_ERR(trans)) 38923612b495STsutomu Itoh err = PTR_ERR(trans); 38933612b495STsutomu Itoh else 38945d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 38953fd0a558SYan, Zheng out_free: 38963fd0a558SYan, Zheng btrfs_free_block_rsv(rc->extent_root, rc->block_rsv); 38973fd0a558SYan, Zheng btrfs_free_path(path); 38985d4f98a2SYan Zheng return err; 38995d4f98a2SYan Zheng } 39005d4f98a2SYan Zheng 39015d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 39020257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 39035d4f98a2SYan Zheng { 39045d4f98a2SYan Zheng struct btrfs_path *path; 39055d4f98a2SYan Zheng struct btrfs_inode_item *item; 39065d4f98a2SYan Zheng struct extent_buffer *leaf; 39075d4f98a2SYan Zheng int ret; 39085d4f98a2SYan Zheng 39095d4f98a2SYan Zheng path = btrfs_alloc_path(); 39105d4f98a2SYan Zheng if (!path) 39115d4f98a2SYan Zheng return -ENOMEM; 39125d4f98a2SYan Zheng 39135d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 39145d4f98a2SYan Zheng if (ret) 39155d4f98a2SYan Zheng goto out; 39165d4f98a2SYan Zheng 39175d4f98a2SYan Zheng leaf = path->nodes[0]; 39185d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 39195d4f98a2SYan Zheng memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item)); 39205d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 39210257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 39225d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 39233fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 39243fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 39255d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 3926b3b4aa74SDavid Sterba btrfs_release_path(path); 39275d4f98a2SYan Zheng out: 39285d4f98a2SYan Zheng btrfs_free_path(path); 39295d4f98a2SYan Zheng return ret; 39305d4f98a2SYan Zheng } 39315d4f98a2SYan Zheng 39325d4f98a2SYan Zheng /* 39335d4f98a2SYan Zheng * helper to create inode for data relocation. 39345d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 39355d4f98a2SYan Zheng */ 39363fd0a558SYan, Zheng static noinline_for_stack 39373fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 39385d4f98a2SYan Zheng struct btrfs_block_group_cache *group) 39395d4f98a2SYan Zheng { 39405d4f98a2SYan Zheng struct inode *inode = NULL; 39415d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 39425d4f98a2SYan Zheng struct btrfs_root *root; 39435d4f98a2SYan Zheng struct btrfs_key key; 39445d4f98a2SYan Zheng unsigned long nr; 39455d4f98a2SYan Zheng u64 objectid = BTRFS_FIRST_FREE_OBJECTID; 39465d4f98a2SYan Zheng int err = 0; 39475d4f98a2SYan Zheng 39485d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 39495d4f98a2SYan Zheng if (IS_ERR(root)) 39505d4f98a2SYan Zheng return ERR_CAST(root); 39515d4f98a2SYan Zheng 3952a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 39533fd0a558SYan, Zheng if (IS_ERR(trans)) 39543fd0a558SYan, Zheng return ERR_CAST(trans); 39555d4f98a2SYan Zheng 3956581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 39575d4f98a2SYan Zheng if (err) 39585d4f98a2SYan Zheng goto out; 39595d4f98a2SYan Zheng 39600257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 39615d4f98a2SYan Zheng BUG_ON(err); 39625d4f98a2SYan Zheng 39635d4f98a2SYan Zheng key.objectid = objectid; 39645d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 39655d4f98a2SYan Zheng key.offset = 0; 396673f73415SJosef Bacik inode = btrfs_iget(root->fs_info->sb, &key, root, NULL); 39675d4f98a2SYan Zheng BUG_ON(IS_ERR(inode) || is_bad_inode(inode)); 39685d4f98a2SYan Zheng BTRFS_I(inode)->index_cnt = group->key.objectid; 39695d4f98a2SYan Zheng 39705d4f98a2SYan Zheng err = btrfs_orphan_add(trans, inode); 39715d4f98a2SYan Zheng out: 39725d4f98a2SYan Zheng nr = trans->blocks_used; 39735d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 39745d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 39755d4f98a2SYan Zheng if (err) { 39765d4f98a2SYan Zheng if (inode) 39775d4f98a2SYan Zheng iput(inode); 39785d4f98a2SYan Zheng inode = ERR_PTR(err); 39795d4f98a2SYan Zheng } 39805d4f98a2SYan Zheng return inode; 39815d4f98a2SYan Zheng } 39825d4f98a2SYan Zheng 39833fd0a558SYan, Zheng static struct reloc_control *alloc_reloc_control(void) 39843fd0a558SYan, Zheng { 39853fd0a558SYan, Zheng struct reloc_control *rc; 39863fd0a558SYan, Zheng 39873fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 39883fd0a558SYan, Zheng if (!rc) 39893fd0a558SYan, Zheng return NULL; 39903fd0a558SYan, Zheng 39913fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 39923fd0a558SYan, Zheng backref_cache_init(&rc->backref_cache); 39933fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 3994f993c883SDavid Sterba extent_io_tree_init(&rc->processed_blocks, NULL); 39953fd0a558SYan, Zheng return rc; 39963fd0a558SYan, Zheng } 39973fd0a558SYan, Zheng 39985d4f98a2SYan Zheng /* 39995d4f98a2SYan Zheng * function to relocate all extents in a block group. 40005d4f98a2SYan Zheng */ 40015d4f98a2SYan Zheng int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start) 40025d4f98a2SYan Zheng { 40035d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = extent_root->fs_info; 40045d4f98a2SYan Zheng struct reloc_control *rc; 40050af3d00bSJosef Bacik struct inode *inode; 40060af3d00bSJosef Bacik struct btrfs_path *path; 40075d4f98a2SYan Zheng int ret; 4008f0486c68SYan, Zheng int rw = 0; 40095d4f98a2SYan Zheng int err = 0; 40105d4f98a2SYan Zheng 40113fd0a558SYan, Zheng rc = alloc_reloc_control(); 40125d4f98a2SYan Zheng if (!rc) 40135d4f98a2SYan Zheng return -ENOMEM; 40145d4f98a2SYan Zheng 4015f0486c68SYan, Zheng rc->extent_root = extent_root; 40163fd0a558SYan, Zheng 40175d4f98a2SYan Zheng rc->block_group = btrfs_lookup_block_group(fs_info, group_start); 40185d4f98a2SYan Zheng BUG_ON(!rc->block_group); 40195d4f98a2SYan Zheng 4020f0486c68SYan, Zheng if (!rc->block_group->ro) { 4021f0486c68SYan, Zheng ret = btrfs_set_block_group_ro(extent_root, rc->block_group); 4022f0486c68SYan, Zheng if (ret) { 4023f0486c68SYan, Zheng err = ret; 4024f0486c68SYan, Zheng goto out; 4025f0486c68SYan, Zheng } 4026f0486c68SYan, Zheng rw = 1; 4027f0486c68SYan, Zheng } 4028f0486c68SYan, Zheng 40290af3d00bSJosef Bacik path = btrfs_alloc_path(); 40300af3d00bSJosef Bacik if (!path) { 40310af3d00bSJosef Bacik err = -ENOMEM; 40320af3d00bSJosef Bacik goto out; 40330af3d00bSJosef Bacik } 40340af3d00bSJosef Bacik 40350af3d00bSJosef Bacik inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group, 40360af3d00bSJosef Bacik path); 40370af3d00bSJosef Bacik btrfs_free_path(path); 40380af3d00bSJosef Bacik 40390af3d00bSJosef Bacik if (!IS_ERR(inode)) 40400af3d00bSJosef Bacik ret = delete_block_group_cache(fs_info, inode, 0); 40410af3d00bSJosef Bacik else 40420af3d00bSJosef Bacik ret = PTR_ERR(inode); 40430af3d00bSJosef Bacik 40440af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 40450af3d00bSJosef Bacik err = ret; 40460af3d00bSJosef Bacik goto out; 40470af3d00bSJosef Bacik } 40480af3d00bSJosef Bacik 40495d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 40505d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 40515d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 40525d4f98a2SYan Zheng rc->data_inode = NULL; 40535d4f98a2SYan Zheng goto out; 40545d4f98a2SYan Zheng } 40555d4f98a2SYan Zheng 40565d4f98a2SYan Zheng printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n", 40575d4f98a2SYan Zheng (unsigned long long)rc->block_group->key.objectid, 40585d4f98a2SYan Zheng (unsigned long long)rc->block_group->flags); 40595d4f98a2SYan Zheng 406024bbcf04SYan, Zheng btrfs_start_delalloc_inodes(fs_info->tree_root, 0); 40616bbe3a9cSLiu Bo btrfs_wait_ordered_extents(fs_info->tree_root, 0); 40625d4f98a2SYan Zheng 40635d4f98a2SYan Zheng while (1) { 406476dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 406576dda93cSYan, Zheng 406676dda93cSYan, Zheng btrfs_clean_old_snapshots(fs_info->tree_root); 40675d4f98a2SYan Zheng ret = relocate_block_group(rc); 406876dda93cSYan, Zheng 406976dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 40705d4f98a2SYan Zheng if (ret < 0) { 40715d4f98a2SYan Zheng err = ret; 40723fd0a558SYan, Zheng goto out; 40735d4f98a2SYan Zheng } 40745d4f98a2SYan Zheng 40755d4f98a2SYan Zheng if (rc->extents_found == 0) 40765d4f98a2SYan Zheng break; 40775d4f98a2SYan Zheng 40785d4f98a2SYan Zheng printk(KERN_INFO "btrfs: found %llu extents\n", 40795d4f98a2SYan Zheng (unsigned long long)rc->extents_found); 40805d4f98a2SYan Zheng 40815d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 40825d4f98a2SYan Zheng btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1); 40835d4f98a2SYan Zheng invalidate_mapping_pages(rc->data_inode->i_mapping, 40845d4f98a2SYan Zheng 0, -1); 40855d4f98a2SYan Zheng rc->stage = UPDATE_DATA_PTRS; 40865d4f98a2SYan Zheng } 40875d4f98a2SYan Zheng } 40885d4f98a2SYan Zheng 40890257bb82SYan, Zheng filemap_write_and_wait_range(fs_info->btree_inode->i_mapping, 40905d4f98a2SYan Zheng rc->block_group->key.objectid, 40915d4f98a2SYan Zheng rc->block_group->key.objectid + 40925d4f98a2SYan Zheng rc->block_group->key.offset - 1); 40935d4f98a2SYan Zheng 40945d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 40955d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 40965d4f98a2SYan Zheng WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0); 40975d4f98a2SYan Zheng out: 4098f0486c68SYan, Zheng if (err && rw) 4099f0486c68SYan, Zheng btrfs_set_block_group_rw(extent_root, rc->block_group); 41005d4f98a2SYan Zheng iput(rc->data_inode); 41015d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 41025d4f98a2SYan Zheng kfree(rc); 41035d4f98a2SYan Zheng return err; 41045d4f98a2SYan Zheng } 41055d4f98a2SYan Zheng 410676dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 410776dda93cSYan, Zheng { 410876dda93cSYan, Zheng struct btrfs_trans_handle *trans; 410979787eaaSJeff Mahoney int ret, err; 411076dda93cSYan, Zheng 4111a22285a6SYan, Zheng trans = btrfs_start_transaction(root->fs_info->tree_root, 0); 411279787eaaSJeff Mahoney if (IS_ERR(trans)) 411379787eaaSJeff Mahoney return PTR_ERR(trans); 411476dda93cSYan, Zheng 411576dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 411676dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 411776dda93cSYan, Zheng root->root_item.drop_level = 0; 411876dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 411976dda93cSYan, Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 412076dda93cSYan, Zheng &root->root_key, &root->root_item); 412176dda93cSYan, Zheng 412279787eaaSJeff Mahoney err = btrfs_end_transaction(trans, root->fs_info->tree_root); 412379787eaaSJeff Mahoney if (err) 412479787eaaSJeff Mahoney return err; 412579787eaaSJeff Mahoney return ret; 412676dda93cSYan, Zheng } 412776dda93cSYan, Zheng 41285d4f98a2SYan Zheng /* 41295d4f98a2SYan Zheng * recover relocation interrupted by system crash. 41305d4f98a2SYan Zheng * 41315d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 41325d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 41335d4f98a2SYan Zheng */ 41345d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 41355d4f98a2SYan Zheng { 41365d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 41375d4f98a2SYan Zheng struct btrfs_key key; 41385d4f98a2SYan Zheng struct btrfs_root *fs_root; 41395d4f98a2SYan Zheng struct btrfs_root *reloc_root; 41405d4f98a2SYan Zheng struct btrfs_path *path; 41415d4f98a2SYan Zheng struct extent_buffer *leaf; 41425d4f98a2SYan Zheng struct reloc_control *rc = NULL; 41435d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 41445d4f98a2SYan Zheng int ret; 41455d4f98a2SYan Zheng int err = 0; 41465d4f98a2SYan Zheng 41475d4f98a2SYan Zheng path = btrfs_alloc_path(); 41485d4f98a2SYan Zheng if (!path) 41495d4f98a2SYan Zheng return -ENOMEM; 4150026fd317SJosef Bacik path->reada = -1; 41515d4f98a2SYan Zheng 41525d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 41535d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 41545d4f98a2SYan Zheng key.offset = (u64)-1; 41555d4f98a2SYan Zheng 41565d4f98a2SYan Zheng while (1) { 41575d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, 41585d4f98a2SYan Zheng path, 0, 0); 41595d4f98a2SYan Zheng if (ret < 0) { 41605d4f98a2SYan Zheng err = ret; 41615d4f98a2SYan Zheng goto out; 41625d4f98a2SYan Zheng } 41635d4f98a2SYan Zheng if (ret > 0) { 41645d4f98a2SYan Zheng if (path->slots[0] == 0) 41655d4f98a2SYan Zheng break; 41665d4f98a2SYan Zheng path->slots[0]--; 41675d4f98a2SYan Zheng } 41685d4f98a2SYan Zheng leaf = path->nodes[0]; 41695d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4170b3b4aa74SDavid Sterba btrfs_release_path(path); 41715d4f98a2SYan Zheng 41725d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 41735d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 41745d4f98a2SYan Zheng break; 41755d4f98a2SYan Zheng 41765d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root, &key); 41775d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 41785d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 41795d4f98a2SYan Zheng goto out; 41805d4f98a2SYan Zheng } 41815d4f98a2SYan Zheng 41825d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 41835d4f98a2SYan Zheng 41845d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 41855d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 41865d4f98a2SYan Zheng reloc_root->root_key.offset); 41875d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 418876dda93cSYan, Zheng ret = PTR_ERR(fs_root); 418976dda93cSYan, Zheng if (ret != -ENOENT) { 419076dda93cSYan, Zheng err = ret; 41915d4f98a2SYan Zheng goto out; 41925d4f98a2SYan Zheng } 419379787eaaSJeff Mahoney ret = mark_garbage_root(reloc_root); 419479787eaaSJeff Mahoney if (ret < 0) { 419579787eaaSJeff Mahoney err = ret; 419679787eaaSJeff Mahoney goto out; 419779787eaaSJeff Mahoney } 419876dda93cSYan, Zheng } 41995d4f98a2SYan Zheng } 42005d4f98a2SYan Zheng 42015d4f98a2SYan Zheng if (key.offset == 0) 42025d4f98a2SYan Zheng break; 42035d4f98a2SYan Zheng 42045d4f98a2SYan Zheng key.offset--; 42055d4f98a2SYan Zheng } 4206b3b4aa74SDavid Sterba btrfs_release_path(path); 42075d4f98a2SYan Zheng 42085d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 42095d4f98a2SYan Zheng goto out; 42105d4f98a2SYan Zheng 42113fd0a558SYan, Zheng rc = alloc_reloc_control(); 42125d4f98a2SYan Zheng if (!rc) { 42135d4f98a2SYan Zheng err = -ENOMEM; 42145d4f98a2SYan Zheng goto out; 42155d4f98a2SYan Zheng } 42165d4f98a2SYan Zheng 42175d4f98a2SYan Zheng rc->extent_root = root->fs_info->extent_root; 42185d4f98a2SYan Zheng 42195d4f98a2SYan Zheng set_reloc_control(rc); 42205d4f98a2SYan Zheng 42217a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 42223612b495STsutomu Itoh if (IS_ERR(trans)) { 42233612b495STsutomu Itoh unset_reloc_control(rc); 42243612b495STsutomu Itoh err = PTR_ERR(trans); 42253612b495STsutomu Itoh goto out_free; 42263612b495STsutomu Itoh } 42273fd0a558SYan, Zheng 42283fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 42293fd0a558SYan, Zheng 42305d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 42315d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 42325d4f98a2SYan Zheng struct btrfs_root, root_list); 42335d4f98a2SYan Zheng list_del(&reloc_root->root_list); 42345d4f98a2SYan Zheng 42355d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 42365d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 42375d4f98a2SYan Zheng &rc->reloc_roots); 42385d4f98a2SYan Zheng continue; 42395d4f98a2SYan Zheng } 42405d4f98a2SYan Zheng 42415d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 42425d4f98a2SYan Zheng reloc_root->root_key.offset); 424379787eaaSJeff Mahoney if (IS_ERR(fs_root)) { 424479787eaaSJeff Mahoney err = PTR_ERR(fs_root); 424579787eaaSJeff Mahoney goto out_free; 424679787eaaSJeff Mahoney } 42475d4f98a2SYan Zheng 4248ffd7b339SJeff Mahoney err = __add_reloc_root(reloc_root); 424979787eaaSJeff Mahoney BUG_ON(err < 0); /* -ENOMEM or logic error */ 42505d4f98a2SYan Zheng fs_root->reloc_root = reloc_root; 42515d4f98a2SYan Zheng } 42525d4f98a2SYan Zheng 425379787eaaSJeff Mahoney err = btrfs_commit_transaction(trans, rc->extent_root); 425479787eaaSJeff Mahoney if (err) 425579787eaaSJeff Mahoney goto out_free; 42565d4f98a2SYan Zheng 42575d4f98a2SYan Zheng merge_reloc_roots(rc); 42585d4f98a2SYan Zheng 42595d4f98a2SYan Zheng unset_reloc_control(rc); 42605d4f98a2SYan Zheng 42617a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 42623612b495STsutomu Itoh if (IS_ERR(trans)) 42633612b495STsutomu Itoh err = PTR_ERR(trans); 42643612b495STsutomu Itoh else 426579787eaaSJeff Mahoney err = btrfs_commit_transaction(trans, rc->extent_root); 42663612b495STsutomu Itoh out_free: 42675d4f98a2SYan Zheng kfree(rc); 42683612b495STsutomu Itoh out: 42695d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 42705d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 42715d4f98a2SYan Zheng struct btrfs_root, root_list); 42725d4f98a2SYan Zheng list_del(&reloc_root->root_list); 42735d4f98a2SYan Zheng free_extent_buffer(reloc_root->node); 42745d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 42755d4f98a2SYan Zheng kfree(reloc_root); 42765d4f98a2SYan Zheng } 42775d4f98a2SYan Zheng btrfs_free_path(path); 42785d4f98a2SYan Zheng 42795d4f98a2SYan Zheng if (err == 0) { 42805d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 42815d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 42825d4f98a2SYan Zheng BTRFS_DATA_RELOC_TREE_OBJECTID); 42835d4f98a2SYan Zheng if (IS_ERR(fs_root)) 42845d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4285d7ce5843SMiao Xie else 428666b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 42875d4f98a2SYan Zheng } 42885d4f98a2SYan Zheng return err; 42895d4f98a2SYan Zheng } 42905d4f98a2SYan Zheng 42915d4f98a2SYan Zheng /* 42925d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 42935d4f98a2SYan Zheng * 42945d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 42955d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 42965d4f98a2SYan Zheng */ 42975d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 42985d4f98a2SYan Zheng { 42995d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 43005d4f98a2SYan Zheng struct btrfs_sector_sum *sector_sum; 43015d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 43025d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 43035d4f98a2SYan Zheng size_t offset; 43045d4f98a2SYan Zheng int ret; 43055d4f98a2SYan Zheng u64 disk_bytenr; 43065d4f98a2SYan Zheng LIST_HEAD(list); 43075d4f98a2SYan Zheng 43085d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 43095d4f98a2SYan Zheng BUG_ON(ordered->file_offset != file_pos || ordered->len != len); 43105d4f98a2SYan Zheng 43115d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 43125d4f98a2SYan Zheng ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr, 4313a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 431479787eaaSJeff Mahoney if (ret) 431579787eaaSJeff Mahoney goto out; 43165d4f98a2SYan Zheng 43175d4f98a2SYan Zheng while (!list_empty(&list)) { 43185d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 43195d4f98a2SYan Zheng list_del_init(&sums->list); 43205d4f98a2SYan Zheng 43215d4f98a2SYan Zheng sector_sum = sums->sums; 43225d4f98a2SYan Zheng sums->bytenr = ordered->start; 43235d4f98a2SYan Zheng 43245d4f98a2SYan Zheng offset = 0; 43255d4f98a2SYan Zheng while (offset < sums->len) { 43265d4f98a2SYan Zheng sector_sum->bytenr += ordered->start - disk_bytenr; 43275d4f98a2SYan Zheng sector_sum++; 43285d4f98a2SYan Zheng offset += root->sectorsize; 43295d4f98a2SYan Zheng } 43305d4f98a2SYan Zheng 43315d4f98a2SYan Zheng btrfs_add_ordered_sum(inode, ordered, sums); 43325d4f98a2SYan Zheng } 433379787eaaSJeff Mahoney out: 43345d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4335411fc6bcSAndi Kleen return ret; 43365d4f98a2SYan Zheng } 43373fd0a558SYan, Zheng 43383fd0a558SYan, Zheng void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 43393fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 43403fd0a558SYan, Zheng struct extent_buffer *cow) 43413fd0a558SYan, Zheng { 43423fd0a558SYan, Zheng struct reloc_control *rc; 43433fd0a558SYan, Zheng struct backref_node *node; 43443fd0a558SYan, Zheng int first_cow = 0; 43453fd0a558SYan, Zheng int level; 43463fd0a558SYan, Zheng int ret; 43473fd0a558SYan, Zheng 43483fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 43493fd0a558SYan, Zheng if (!rc) 43503fd0a558SYan, Zheng return; 43513fd0a558SYan, Zheng 43523fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 43533fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 43543fd0a558SYan, Zheng 43553fd0a558SYan, Zheng level = btrfs_header_level(buf); 43563fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 43573fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 43583fd0a558SYan, Zheng first_cow = 1; 43593fd0a558SYan, Zheng 43603fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 43613fd0a558SYan, Zheng rc->create_reloc_tree) { 43623fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 43633fd0a558SYan, Zheng 43643fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 43653fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 43663fd0a558SYan, Zheng node->new_bytenr != buf->start); 43673fd0a558SYan, Zheng 43683fd0a558SYan, Zheng drop_node_buffer(node); 43693fd0a558SYan, Zheng extent_buffer_get(cow); 43703fd0a558SYan, Zheng node->eb = cow; 43713fd0a558SYan, Zheng node->new_bytenr = cow->start; 43723fd0a558SYan, Zheng 43733fd0a558SYan, Zheng if (!node->pending) { 43743fd0a558SYan, Zheng list_move_tail(&node->list, 43753fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 43763fd0a558SYan, Zheng node->pending = 1; 43773fd0a558SYan, Zheng } 43783fd0a558SYan, Zheng 43793fd0a558SYan, Zheng if (first_cow) 43803fd0a558SYan, Zheng __mark_block_processed(rc, node); 43813fd0a558SYan, Zheng 43823fd0a558SYan, Zheng if (first_cow && level > 0) 43833fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 43843fd0a558SYan, Zheng } 43853fd0a558SYan, Zheng 43863fd0a558SYan, Zheng if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) { 43873fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 43883fd0a558SYan, Zheng BUG_ON(ret); 43893fd0a558SYan, Zheng } 43903fd0a558SYan, Zheng } 43913fd0a558SYan, Zheng 43923fd0a558SYan, Zheng /* 43933fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 43943fd0a558SYan, Zheng * requried for relocating tree blocks in the snapshot 43953fd0a558SYan, Zheng */ 43963fd0a558SYan, Zheng void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans, 43973fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending, 43983fd0a558SYan, Zheng u64 *bytes_to_reserve) 43993fd0a558SYan, Zheng { 44003fd0a558SYan, Zheng struct btrfs_root *root; 44013fd0a558SYan, Zheng struct reloc_control *rc; 44023fd0a558SYan, Zheng 44033fd0a558SYan, Zheng root = pending->root; 44043fd0a558SYan, Zheng if (!root->reloc_root) 44053fd0a558SYan, Zheng return; 44063fd0a558SYan, Zheng 44073fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 44083fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 44093fd0a558SYan, Zheng return; 44103fd0a558SYan, Zheng 44113fd0a558SYan, Zheng root = root->reloc_root; 44123fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 44133fd0a558SYan, Zheng /* 44143fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 44153fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 44163fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 44173fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 44183fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 44193fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 44203fd0a558SYan, Zheng * tree may use more space than it frees. so we need 44213fd0a558SYan, Zheng * reserve extra space. 44223fd0a558SYan, Zheng */ 44233fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 44243fd0a558SYan, Zheng } 44253fd0a558SYan, Zheng 44263fd0a558SYan, Zheng /* 44273fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 44283fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 44293fd0a558SYan, Zheng */ 443049b25e05SJeff Mahoney int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 44313fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 44323fd0a558SYan, Zheng { 44333fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 44343fd0a558SYan, Zheng struct btrfs_root *reloc_root; 44353fd0a558SYan, Zheng struct btrfs_root *new_root; 44363fd0a558SYan, Zheng struct reloc_control *rc; 44373fd0a558SYan, Zheng int ret; 44383fd0a558SYan, Zheng 44393fd0a558SYan, Zheng if (!root->reloc_root) 444049b25e05SJeff Mahoney return 0; 44413fd0a558SYan, Zheng 44423fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 44433fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 44443fd0a558SYan, Zheng 44453fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 44463fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 44473fd0a558SYan, Zheng rc->block_rsv, 44483fd0a558SYan, Zheng rc->nodes_relocated); 444949b25e05SJeff Mahoney if (ret) 445049b25e05SJeff Mahoney return ret; 44513fd0a558SYan, Zheng } 44523fd0a558SYan, Zheng 44533fd0a558SYan, Zheng new_root = pending->snap; 44543fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 44553fd0a558SYan, Zheng new_root->root_key.objectid); 445649b25e05SJeff Mahoney if (IS_ERR(reloc_root)) 445749b25e05SJeff Mahoney return PTR_ERR(reloc_root); 44583fd0a558SYan, Zheng 4459ffd7b339SJeff Mahoney ret = __add_reloc_root(reloc_root); 4460ffd7b339SJeff Mahoney BUG_ON(ret < 0); 44613fd0a558SYan, Zheng new_root->reloc_root = reloc_root; 44623fd0a558SYan, Zheng 446349b25e05SJeff Mahoney if (rc->create_reloc_tree) 44643fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 446549b25e05SJeff Mahoney return ret; 44663fd0a558SYan, Zheng } 4467