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 3295d4f98a2SYan Zheng /* 3305d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 3315d4f98a2SYan Zheng */ 3325d4f98a2SYan Zheng static struct backref_node *walk_up_backref(struct backref_node *node, 3335d4f98a2SYan Zheng struct backref_edge *edges[], 3345d4f98a2SYan Zheng int *index) 3355d4f98a2SYan Zheng { 3365d4f98a2SYan Zheng struct backref_edge *edge; 3375d4f98a2SYan Zheng int idx = *index; 3385d4f98a2SYan Zheng 3395d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 3405d4f98a2SYan Zheng edge = list_entry(node->upper.next, 3415d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 3425d4f98a2SYan Zheng edges[idx++] = edge; 3435d4f98a2SYan Zheng node = edge->node[UPPER]; 3445d4f98a2SYan Zheng } 3453fd0a558SYan, Zheng BUG_ON(node->detached); 3465d4f98a2SYan Zheng *index = idx; 3475d4f98a2SYan Zheng return node; 3485d4f98a2SYan Zheng } 3495d4f98a2SYan Zheng 3505d4f98a2SYan Zheng /* 3515d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 3525d4f98a2SYan Zheng */ 3535d4f98a2SYan Zheng static struct backref_node *walk_down_backref(struct backref_edge *edges[], 3545d4f98a2SYan Zheng int *index) 3555d4f98a2SYan Zheng { 3565d4f98a2SYan Zheng struct backref_edge *edge; 3575d4f98a2SYan Zheng struct backref_node *lower; 3585d4f98a2SYan Zheng int idx = *index; 3595d4f98a2SYan Zheng 3605d4f98a2SYan Zheng while (idx > 0) { 3615d4f98a2SYan Zheng edge = edges[idx - 1]; 3625d4f98a2SYan Zheng lower = edge->node[LOWER]; 3635d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 3645d4f98a2SYan Zheng idx--; 3655d4f98a2SYan Zheng continue; 3665d4f98a2SYan Zheng } 3675d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 3685d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 3695d4f98a2SYan Zheng edges[idx - 1] = edge; 3705d4f98a2SYan Zheng *index = idx; 3715d4f98a2SYan Zheng return edge->node[UPPER]; 3725d4f98a2SYan Zheng } 3735d4f98a2SYan Zheng *index = 0; 3745d4f98a2SYan Zheng return NULL; 3755d4f98a2SYan Zheng } 3765d4f98a2SYan Zheng 3773fd0a558SYan, Zheng static void unlock_node_buffer(struct backref_node *node) 3785d4f98a2SYan Zheng { 3795d4f98a2SYan Zheng if (node->locked) { 3805d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 3815d4f98a2SYan Zheng node->locked = 0; 3825d4f98a2SYan Zheng } 3833fd0a558SYan, Zheng } 3843fd0a558SYan, Zheng 3853fd0a558SYan, Zheng static void drop_node_buffer(struct backref_node *node) 3863fd0a558SYan, Zheng { 3873fd0a558SYan, Zheng if (node->eb) { 3883fd0a558SYan, Zheng unlock_node_buffer(node); 3895d4f98a2SYan Zheng free_extent_buffer(node->eb); 3905d4f98a2SYan Zheng node->eb = NULL; 3915d4f98a2SYan Zheng } 3925d4f98a2SYan Zheng } 3935d4f98a2SYan Zheng 3945d4f98a2SYan Zheng static void drop_backref_node(struct backref_cache *tree, 3955d4f98a2SYan Zheng struct backref_node *node) 3965d4f98a2SYan Zheng { 3975d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 3985d4f98a2SYan Zheng 3995d4f98a2SYan Zheng drop_node_buffer(node); 4003fd0a558SYan, Zheng list_del(&node->list); 4015d4f98a2SYan Zheng list_del(&node->lower); 4023fd0a558SYan, Zheng if (!RB_EMPTY_NODE(&node->rb_node)) 4035d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 4043fd0a558SYan, Zheng free_backref_node(tree, node); 4055d4f98a2SYan Zheng } 4065d4f98a2SYan Zheng 4075d4f98a2SYan Zheng /* 4085d4f98a2SYan Zheng * remove a backref node from the backref cache 4095d4f98a2SYan Zheng */ 4105d4f98a2SYan Zheng static void remove_backref_node(struct backref_cache *cache, 4115d4f98a2SYan Zheng struct backref_node *node) 4125d4f98a2SYan Zheng { 4135d4f98a2SYan Zheng struct backref_node *upper; 4145d4f98a2SYan Zheng struct backref_edge *edge; 4155d4f98a2SYan Zheng 4165d4f98a2SYan Zheng if (!node) 4175d4f98a2SYan Zheng return; 4185d4f98a2SYan Zheng 4193fd0a558SYan, Zheng BUG_ON(!node->lowest && !node->detached); 4205d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 4215d4f98a2SYan Zheng edge = list_entry(node->upper.next, struct backref_edge, 4225d4f98a2SYan Zheng list[LOWER]); 4235d4f98a2SYan Zheng upper = edge->node[UPPER]; 4245d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 4255d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 4263fd0a558SYan, Zheng free_backref_edge(cache, edge); 4273fd0a558SYan, Zheng 4283fd0a558SYan, Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 4293fd0a558SYan, Zheng BUG_ON(!list_empty(&node->upper)); 4303fd0a558SYan, Zheng drop_backref_node(cache, node); 4313fd0a558SYan, Zheng node = upper; 4323fd0a558SYan, Zheng node->lowest = 1; 4333fd0a558SYan, Zheng continue; 4343fd0a558SYan, Zheng } 4355d4f98a2SYan Zheng /* 4363fd0a558SYan, Zheng * add the node to leaf node list if no other 4375d4f98a2SYan Zheng * child block cached. 4385d4f98a2SYan Zheng */ 4395d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 4403fd0a558SYan, Zheng list_add_tail(&upper->lower, &cache->leaves); 4415d4f98a2SYan Zheng upper->lowest = 1; 4425d4f98a2SYan Zheng } 4435d4f98a2SYan Zheng } 4443fd0a558SYan, Zheng 4455d4f98a2SYan Zheng drop_backref_node(cache, node); 4465d4f98a2SYan Zheng } 4475d4f98a2SYan Zheng 4483fd0a558SYan, Zheng static void update_backref_node(struct backref_cache *cache, 4493fd0a558SYan, Zheng struct backref_node *node, u64 bytenr) 4503fd0a558SYan, Zheng { 4513fd0a558SYan, Zheng struct rb_node *rb_node; 4523fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 4533fd0a558SYan, Zheng node->bytenr = bytenr; 4543fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 4553fd0a558SYan, Zheng BUG_ON(rb_node); 4563fd0a558SYan, Zheng } 4573fd0a558SYan, Zheng 4583fd0a558SYan, Zheng /* 4593fd0a558SYan, Zheng * update backref cache after a transaction commit 4603fd0a558SYan, Zheng */ 4613fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 4623fd0a558SYan, Zheng struct backref_cache *cache) 4633fd0a558SYan, Zheng { 4643fd0a558SYan, Zheng struct backref_node *node; 4653fd0a558SYan, Zheng int level = 0; 4663fd0a558SYan, Zheng 4673fd0a558SYan, Zheng if (cache->last_trans == 0) { 4683fd0a558SYan, Zheng cache->last_trans = trans->transid; 4693fd0a558SYan, Zheng return 0; 4703fd0a558SYan, Zheng } 4713fd0a558SYan, Zheng 4723fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 4733fd0a558SYan, Zheng return 0; 4743fd0a558SYan, Zheng 4753fd0a558SYan, Zheng /* 4763fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 4773fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 4783fd0a558SYan, Zheng * so the detached nodes are no longer useful. 4793fd0a558SYan, Zheng */ 4803fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 4813fd0a558SYan, Zheng node = list_entry(cache->detached.next, 4823fd0a558SYan, Zheng struct backref_node, list); 4833fd0a558SYan, Zheng remove_backref_node(cache, node); 4843fd0a558SYan, Zheng } 4853fd0a558SYan, Zheng 4863fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 4873fd0a558SYan, Zheng node = list_entry(cache->changed.next, 4883fd0a558SYan, Zheng struct backref_node, list); 4893fd0a558SYan, Zheng list_del_init(&node->list); 4903fd0a558SYan, Zheng BUG_ON(node->pending); 4913fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 4923fd0a558SYan, Zheng } 4933fd0a558SYan, Zheng 4943fd0a558SYan, Zheng /* 4953fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 4963fd0a558SYan, Zheng * errors during processing the pending nodes. 4973fd0a558SYan, Zheng */ 4983fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 4993fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 5003fd0a558SYan, Zheng BUG_ON(!node->pending); 5013fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 5023fd0a558SYan, Zheng continue; 5033fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 5043fd0a558SYan, Zheng } 5053fd0a558SYan, Zheng } 5063fd0a558SYan, Zheng 5073fd0a558SYan, Zheng cache->last_trans = 0; 5083fd0a558SYan, Zheng return 1; 5093fd0a558SYan, Zheng } 5103fd0a558SYan, Zheng 511f2a97a9dSDavid Sterba 5123fd0a558SYan, Zheng static int should_ignore_root(struct btrfs_root *root) 5133fd0a558SYan, Zheng { 5143fd0a558SYan, Zheng struct btrfs_root *reloc_root; 5153fd0a558SYan, Zheng 5163fd0a558SYan, Zheng if (!root->ref_cows) 5173fd0a558SYan, Zheng return 0; 5183fd0a558SYan, Zheng 5193fd0a558SYan, Zheng reloc_root = root->reloc_root; 5203fd0a558SYan, Zheng if (!reloc_root) 5213fd0a558SYan, Zheng return 0; 5223fd0a558SYan, Zheng 5233fd0a558SYan, Zheng if (btrfs_root_last_snapshot(&reloc_root->root_item) == 5243fd0a558SYan, Zheng root->fs_info->running_transaction->transid - 1) 5253fd0a558SYan, Zheng return 0; 5263fd0a558SYan, Zheng /* 5273fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 5283fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 5293fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 5303fd0a558SYan, Zheng * relocation. 5313fd0a558SYan, Zheng */ 5323fd0a558SYan, Zheng return 1; 5333fd0a558SYan, Zheng } 5345d4f98a2SYan Zheng /* 5355d4f98a2SYan Zheng * find reloc tree by address of tree root 5365d4f98a2SYan Zheng */ 5375d4f98a2SYan Zheng static struct btrfs_root *find_reloc_root(struct reloc_control *rc, 5385d4f98a2SYan Zheng u64 bytenr) 5395d4f98a2SYan Zheng { 5405d4f98a2SYan Zheng struct rb_node *rb_node; 5415d4f98a2SYan Zheng struct mapping_node *node; 5425d4f98a2SYan Zheng struct btrfs_root *root = NULL; 5435d4f98a2SYan Zheng 5445d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 5455d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr); 5465d4f98a2SYan Zheng if (rb_node) { 5475d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 5485d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 5495d4f98a2SYan Zheng } 5505d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 5515d4f98a2SYan Zheng return root; 5525d4f98a2SYan Zheng } 5535d4f98a2SYan Zheng 5545d4f98a2SYan Zheng static int is_cowonly_root(u64 root_objectid) 5555d4f98a2SYan Zheng { 5565d4f98a2SYan Zheng if (root_objectid == BTRFS_ROOT_TREE_OBJECTID || 5575d4f98a2SYan Zheng root_objectid == BTRFS_EXTENT_TREE_OBJECTID || 5585d4f98a2SYan Zheng root_objectid == BTRFS_CHUNK_TREE_OBJECTID || 5595d4f98a2SYan Zheng root_objectid == BTRFS_DEV_TREE_OBJECTID || 5605d4f98a2SYan Zheng root_objectid == BTRFS_TREE_LOG_OBJECTID || 5615d4f98a2SYan Zheng root_objectid == BTRFS_CSUM_TREE_OBJECTID) 5625d4f98a2SYan Zheng return 1; 5635d4f98a2SYan Zheng return 0; 5645d4f98a2SYan Zheng } 5655d4f98a2SYan Zheng 5665d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 5675d4f98a2SYan Zheng u64 root_objectid) 5685d4f98a2SYan Zheng { 5695d4f98a2SYan Zheng struct btrfs_key key; 5705d4f98a2SYan Zheng 5715d4f98a2SYan Zheng key.objectid = root_objectid; 5725d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 5735d4f98a2SYan Zheng if (is_cowonly_root(root_objectid)) 5745d4f98a2SYan Zheng key.offset = 0; 5755d4f98a2SYan Zheng else 5765d4f98a2SYan Zheng key.offset = (u64)-1; 5775d4f98a2SYan Zheng 5785d4f98a2SYan Zheng return btrfs_read_fs_root_no_name(fs_info, &key); 5795d4f98a2SYan Zheng } 5805d4f98a2SYan Zheng 5815d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 5825d4f98a2SYan Zheng static noinline_for_stack 5835d4f98a2SYan Zheng struct btrfs_root *find_tree_root(struct reloc_control *rc, 5845d4f98a2SYan Zheng struct extent_buffer *leaf, 5855d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0) 5865d4f98a2SYan Zheng { 5875d4f98a2SYan Zheng struct btrfs_root *root; 5885d4f98a2SYan Zheng u64 root_objectid = btrfs_ref_root_v0(leaf, ref0); 5895d4f98a2SYan Zheng u64 generation = btrfs_ref_generation_v0(leaf, ref0); 5905d4f98a2SYan Zheng 5915d4f98a2SYan Zheng BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID); 5925d4f98a2SYan Zheng 5935d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, root_objectid); 5945d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 5955d4f98a2SYan Zheng 5965d4f98a2SYan Zheng if (root->ref_cows && 5975d4f98a2SYan Zheng generation != btrfs_root_generation(&root->root_item)) 5985d4f98a2SYan Zheng return NULL; 5995d4f98a2SYan Zheng 6005d4f98a2SYan Zheng return root; 6015d4f98a2SYan Zheng } 6025d4f98a2SYan Zheng #endif 6035d4f98a2SYan Zheng 6045d4f98a2SYan Zheng static noinline_for_stack 6055d4f98a2SYan Zheng int find_inline_backref(struct extent_buffer *leaf, int slot, 6065d4f98a2SYan Zheng unsigned long *ptr, unsigned long *end) 6075d4f98a2SYan Zheng { 6085d4f98a2SYan Zheng struct btrfs_extent_item *ei; 6095d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 6105d4f98a2SYan Zheng u32 item_size; 6115d4f98a2SYan Zheng 6125d4f98a2SYan Zheng item_size = btrfs_item_size_nr(leaf, slot); 6135d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 6145d4f98a2SYan Zheng if (item_size < sizeof(*ei)) { 6155d4f98a2SYan Zheng WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 6165d4f98a2SYan Zheng return 1; 6175d4f98a2SYan Zheng } 6185d4f98a2SYan Zheng #endif 6195d4f98a2SYan Zheng ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item); 6205d4f98a2SYan Zheng WARN_ON(!(btrfs_extent_flags(leaf, ei) & 6215d4f98a2SYan Zheng BTRFS_EXTENT_FLAG_TREE_BLOCK)); 6225d4f98a2SYan Zheng 6235d4f98a2SYan Zheng if (item_size <= sizeof(*ei) + sizeof(*bi)) { 6245d4f98a2SYan Zheng WARN_ON(item_size < sizeof(*ei) + sizeof(*bi)); 6255d4f98a2SYan Zheng return 1; 6265d4f98a2SYan Zheng } 6275d4f98a2SYan Zheng 6285d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 6295d4f98a2SYan Zheng *ptr = (unsigned long)(bi + 1); 6305d4f98a2SYan Zheng *end = (unsigned long)ei + item_size; 6315d4f98a2SYan Zheng return 0; 6325d4f98a2SYan Zheng } 6335d4f98a2SYan Zheng 6345d4f98a2SYan Zheng /* 6355d4f98a2SYan Zheng * build backref tree for a given tree block. root of the backref tree 6365d4f98a2SYan Zheng * corresponds the tree block, leaves of the backref tree correspond 6375d4f98a2SYan Zheng * roots of b-trees that reference the tree block. 6385d4f98a2SYan Zheng * 6395d4f98a2SYan Zheng * the basic idea of this function is check backrefs of a given block 6405d4f98a2SYan Zheng * to find upper level blocks that refernece the block, and then check 6415d4f98a2SYan Zheng * bakcrefs of these upper level blocks recursively. the recursion stop 6425d4f98a2SYan Zheng * when tree root is reached or backrefs for the block is cached. 6435d4f98a2SYan Zheng * 6445d4f98a2SYan Zheng * NOTE: if we find backrefs for a block are cached, we know backrefs 6455d4f98a2SYan Zheng * for all upper level blocks that directly/indirectly reference the 6465d4f98a2SYan Zheng * block are also cached. 6475d4f98a2SYan Zheng */ 6483fd0a558SYan, Zheng static noinline_for_stack 6493fd0a558SYan, Zheng struct backref_node *build_backref_tree(struct reloc_control *rc, 6505d4f98a2SYan Zheng struct btrfs_key *node_key, 6515d4f98a2SYan Zheng int level, u64 bytenr) 6525d4f98a2SYan Zheng { 6533fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 6545d4f98a2SYan Zheng struct btrfs_path *path1; 6555d4f98a2SYan Zheng struct btrfs_path *path2; 6565d4f98a2SYan Zheng struct extent_buffer *eb; 6575d4f98a2SYan Zheng struct btrfs_root *root; 6585d4f98a2SYan Zheng struct backref_node *cur; 6595d4f98a2SYan Zheng struct backref_node *upper; 6605d4f98a2SYan Zheng struct backref_node *lower; 6615d4f98a2SYan Zheng struct backref_node *node = NULL; 6625d4f98a2SYan Zheng struct backref_node *exist = NULL; 6635d4f98a2SYan Zheng struct backref_edge *edge; 6645d4f98a2SYan Zheng struct rb_node *rb_node; 6655d4f98a2SYan Zheng struct btrfs_key key; 6665d4f98a2SYan Zheng unsigned long end; 6675d4f98a2SYan Zheng unsigned long ptr; 6685d4f98a2SYan Zheng LIST_HEAD(list); 6693fd0a558SYan, Zheng LIST_HEAD(useless); 6703fd0a558SYan, Zheng int cowonly; 6715d4f98a2SYan Zheng int ret; 6725d4f98a2SYan Zheng int err = 0; 6735d4f98a2SYan Zheng 6745d4f98a2SYan Zheng path1 = btrfs_alloc_path(); 6755d4f98a2SYan Zheng path2 = btrfs_alloc_path(); 6765d4f98a2SYan Zheng if (!path1 || !path2) { 6775d4f98a2SYan Zheng err = -ENOMEM; 6785d4f98a2SYan Zheng goto out; 6795d4f98a2SYan Zheng } 680026fd317SJosef Bacik path1->reada = 1; 681026fd317SJosef Bacik path2->reada = 2; 6825d4f98a2SYan Zheng 6833fd0a558SYan, Zheng node = alloc_backref_node(cache); 6845d4f98a2SYan Zheng if (!node) { 6855d4f98a2SYan Zheng err = -ENOMEM; 6865d4f98a2SYan Zheng goto out; 6875d4f98a2SYan Zheng } 6885d4f98a2SYan Zheng 6895d4f98a2SYan Zheng node->bytenr = bytenr; 6905d4f98a2SYan Zheng node->level = level; 6915d4f98a2SYan Zheng node->lowest = 1; 6925d4f98a2SYan Zheng cur = node; 6935d4f98a2SYan Zheng again: 6945d4f98a2SYan Zheng end = 0; 6955d4f98a2SYan Zheng ptr = 0; 6965d4f98a2SYan Zheng key.objectid = cur->bytenr; 6975d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 6985d4f98a2SYan Zheng key.offset = (u64)-1; 6995d4f98a2SYan Zheng 7005d4f98a2SYan Zheng path1->search_commit_root = 1; 7015d4f98a2SYan Zheng path1->skip_locking = 1; 7025d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1, 7035d4f98a2SYan Zheng 0, 0); 7045d4f98a2SYan Zheng if (ret < 0) { 7055d4f98a2SYan Zheng err = ret; 7065d4f98a2SYan Zheng goto out; 7075d4f98a2SYan Zheng } 7085d4f98a2SYan Zheng BUG_ON(!ret || !path1->slots[0]); 7095d4f98a2SYan Zheng 7105d4f98a2SYan Zheng path1->slots[0]--; 7115d4f98a2SYan Zheng 7125d4f98a2SYan Zheng WARN_ON(cur->checked); 7135d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 7145d4f98a2SYan Zheng /* 71570f23fd6SJustin P. Mattock * the backref was added previously when processing 7165d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 7175d4f98a2SYan Zheng */ 7185d4f98a2SYan Zheng BUG_ON(!list_is_singular(&cur->upper)); 7195d4f98a2SYan Zheng edge = list_entry(cur->upper.next, struct backref_edge, 7205d4f98a2SYan Zheng list[LOWER]); 7215d4f98a2SYan Zheng BUG_ON(!list_empty(&edge->list[UPPER])); 7225d4f98a2SYan Zheng exist = edge->node[UPPER]; 7235d4f98a2SYan Zheng /* 7245d4f98a2SYan Zheng * add the upper level block to pending list if we need 7255d4f98a2SYan Zheng * check its backrefs 7265d4f98a2SYan Zheng */ 7275d4f98a2SYan Zheng if (!exist->checked) 7285d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 7295d4f98a2SYan Zheng } else { 7305d4f98a2SYan Zheng exist = NULL; 7315d4f98a2SYan Zheng } 7325d4f98a2SYan Zheng 7335d4f98a2SYan Zheng while (1) { 7345d4f98a2SYan Zheng cond_resched(); 7355d4f98a2SYan Zheng eb = path1->nodes[0]; 7365d4f98a2SYan Zheng 7375d4f98a2SYan Zheng if (ptr >= end) { 7385d4f98a2SYan Zheng if (path1->slots[0] >= btrfs_header_nritems(eb)) { 7395d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path1); 7405d4f98a2SYan Zheng if (ret < 0) { 7415d4f98a2SYan Zheng err = ret; 7425d4f98a2SYan Zheng goto out; 7435d4f98a2SYan Zheng } 7445d4f98a2SYan Zheng if (ret > 0) 7455d4f98a2SYan Zheng break; 7465d4f98a2SYan Zheng eb = path1->nodes[0]; 7475d4f98a2SYan Zheng } 7485d4f98a2SYan Zheng 7495d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path1->slots[0]); 7505d4f98a2SYan Zheng if (key.objectid != cur->bytenr) { 7515d4f98a2SYan Zheng WARN_ON(exist); 7525d4f98a2SYan Zheng break; 7535d4f98a2SYan Zheng } 7545d4f98a2SYan Zheng 7555d4f98a2SYan Zheng if (key.type == BTRFS_EXTENT_ITEM_KEY) { 7565d4f98a2SYan Zheng ret = find_inline_backref(eb, path1->slots[0], 7575d4f98a2SYan Zheng &ptr, &end); 7585d4f98a2SYan Zheng if (ret) 7595d4f98a2SYan Zheng goto next; 7605d4f98a2SYan Zheng } 7615d4f98a2SYan Zheng } 7625d4f98a2SYan Zheng 7635d4f98a2SYan Zheng if (ptr < end) { 7645d4f98a2SYan Zheng /* update key for inline back ref */ 7655d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 7665d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 7675d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 7685d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 7695d4f98a2SYan Zheng WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY && 7705d4f98a2SYan Zheng key.type != BTRFS_SHARED_BLOCK_REF_KEY); 7715d4f98a2SYan Zheng } 7725d4f98a2SYan Zheng 7735d4f98a2SYan Zheng if (exist && 7745d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 7755d4f98a2SYan Zheng exist->owner == key.offset) || 7765d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 7775d4f98a2SYan Zheng exist->bytenr == key.offset))) { 7785d4f98a2SYan Zheng exist = NULL; 7795d4f98a2SYan Zheng goto next; 7805d4f98a2SYan Zheng } 7815d4f98a2SYan Zheng 7825d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 7835d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY || 7845d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 7853fd0a558SYan, Zheng if (key.type == BTRFS_EXTENT_REF_V0_KEY) { 7865d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 7875d4f98a2SYan Zheng ref0 = btrfs_item_ptr(eb, path1->slots[0], 7885d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 7893fd0a558SYan, Zheng if (key.objectid == key.offset) { 790046f264fSYan, Zheng root = find_tree_root(rc, eb, ref0); 7913fd0a558SYan, Zheng if (root && !should_ignore_root(root)) 7925d4f98a2SYan Zheng cur->root = root; 7935d4f98a2SYan Zheng else 7943fd0a558SYan, Zheng list_add(&cur->list, &useless); 7955d4f98a2SYan Zheng break; 7965d4f98a2SYan Zheng } 797046f264fSYan, Zheng if (is_cowonly_root(btrfs_ref_root_v0(eb, 798046f264fSYan, Zheng ref0))) 799046f264fSYan, Zheng cur->cowonly = 1; 8003fd0a558SYan, Zheng } 8015d4f98a2SYan Zheng #else 8025d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 8035d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 8045d4f98a2SYan Zheng #endif 8055d4f98a2SYan Zheng if (key.objectid == key.offset) { 8065d4f98a2SYan Zheng /* 8075d4f98a2SYan Zheng * only root blocks of reloc trees use 8085d4f98a2SYan Zheng * backref of this type. 8095d4f98a2SYan Zheng */ 8105d4f98a2SYan Zheng root = find_reloc_root(rc, cur->bytenr); 8115d4f98a2SYan Zheng BUG_ON(!root); 8125d4f98a2SYan Zheng cur->root = root; 8135d4f98a2SYan Zheng break; 8145d4f98a2SYan Zheng } 8155d4f98a2SYan Zheng 8163fd0a558SYan, Zheng edge = alloc_backref_edge(cache); 8175d4f98a2SYan Zheng if (!edge) { 8185d4f98a2SYan Zheng err = -ENOMEM; 8195d4f98a2SYan Zheng goto out; 8205d4f98a2SYan Zheng } 8215d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, key.offset); 8225d4f98a2SYan Zheng if (!rb_node) { 8233fd0a558SYan, Zheng upper = alloc_backref_node(cache); 8245d4f98a2SYan Zheng if (!upper) { 8253fd0a558SYan, Zheng free_backref_edge(cache, edge); 8265d4f98a2SYan Zheng err = -ENOMEM; 8275d4f98a2SYan Zheng goto out; 8285d4f98a2SYan Zheng } 8295d4f98a2SYan Zheng upper->bytenr = key.offset; 8305d4f98a2SYan Zheng upper->level = cur->level + 1; 8315d4f98a2SYan Zheng /* 8325d4f98a2SYan Zheng * backrefs for the upper level block isn't 8335d4f98a2SYan Zheng * cached, add the block to pending list 8345d4f98a2SYan Zheng */ 8355d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 8365d4f98a2SYan Zheng } else { 8375d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 8385d4f98a2SYan Zheng rb_node); 8393fd0a558SYan, Zheng BUG_ON(!upper->checked); 8405d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 8415d4f98a2SYan Zheng } 8423fd0a558SYan, Zheng list_add_tail(&edge->list[LOWER], &cur->upper); 8435d4f98a2SYan Zheng edge->node[LOWER] = cur; 8443fd0a558SYan, Zheng edge->node[UPPER] = upper; 8455d4f98a2SYan Zheng 8465d4f98a2SYan Zheng goto next; 8475d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 8485d4f98a2SYan Zheng goto next; 8495d4f98a2SYan Zheng } 8505d4f98a2SYan Zheng 8515d4f98a2SYan Zheng /* key.type == BTRFS_TREE_BLOCK_REF_KEY */ 8525d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, key.offset); 8535d4f98a2SYan Zheng if (IS_ERR(root)) { 8545d4f98a2SYan Zheng err = PTR_ERR(root); 8555d4f98a2SYan Zheng goto out; 8565d4f98a2SYan Zheng } 8575d4f98a2SYan Zheng 8583fd0a558SYan, Zheng if (!root->ref_cows) 8593fd0a558SYan, Zheng cur->cowonly = 1; 8603fd0a558SYan, Zheng 8615d4f98a2SYan Zheng if (btrfs_root_level(&root->root_item) == cur->level) { 8625d4f98a2SYan Zheng /* tree root */ 8635d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 8645d4f98a2SYan Zheng cur->bytenr); 8653fd0a558SYan, Zheng if (should_ignore_root(root)) 8663fd0a558SYan, Zheng list_add(&cur->list, &useless); 8673fd0a558SYan, Zheng else 8685d4f98a2SYan Zheng cur->root = root; 8695d4f98a2SYan Zheng break; 8705d4f98a2SYan Zheng } 8715d4f98a2SYan Zheng 8725d4f98a2SYan Zheng level = cur->level + 1; 8735d4f98a2SYan Zheng 8745d4f98a2SYan Zheng /* 8755d4f98a2SYan Zheng * searching the tree to find upper level blocks 8765d4f98a2SYan Zheng * reference the block. 8775d4f98a2SYan Zheng */ 8785d4f98a2SYan Zheng path2->search_commit_root = 1; 8795d4f98a2SYan Zheng path2->skip_locking = 1; 8805d4f98a2SYan Zheng path2->lowest_level = level; 8815d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0); 8825d4f98a2SYan Zheng path2->lowest_level = 0; 8835d4f98a2SYan Zheng if (ret < 0) { 8845d4f98a2SYan Zheng err = ret; 8855d4f98a2SYan Zheng goto out; 8865d4f98a2SYan Zheng } 88733c66f43SYan Zheng if (ret > 0 && path2->slots[level] > 0) 88833c66f43SYan Zheng path2->slots[level]--; 8895d4f98a2SYan Zheng 8905d4f98a2SYan Zheng eb = path2->nodes[level]; 8915d4f98a2SYan Zheng WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) != 8925d4f98a2SYan Zheng cur->bytenr); 8935d4f98a2SYan Zheng 8945d4f98a2SYan Zheng lower = cur; 8955d4f98a2SYan Zheng for (; level < BTRFS_MAX_LEVEL; level++) { 8965d4f98a2SYan Zheng if (!path2->nodes[level]) { 8975d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 8985d4f98a2SYan Zheng lower->bytenr); 8993fd0a558SYan, Zheng if (should_ignore_root(root)) 9003fd0a558SYan, Zheng list_add(&lower->list, &useless); 9013fd0a558SYan, Zheng else 9025d4f98a2SYan Zheng lower->root = root; 9035d4f98a2SYan Zheng break; 9045d4f98a2SYan Zheng } 9055d4f98a2SYan Zheng 9063fd0a558SYan, Zheng edge = alloc_backref_edge(cache); 9075d4f98a2SYan Zheng if (!edge) { 9085d4f98a2SYan Zheng err = -ENOMEM; 9095d4f98a2SYan Zheng goto out; 9105d4f98a2SYan Zheng } 9115d4f98a2SYan Zheng 9125d4f98a2SYan Zheng eb = path2->nodes[level]; 9135d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, eb->start); 9145d4f98a2SYan Zheng if (!rb_node) { 9153fd0a558SYan, Zheng upper = alloc_backref_node(cache); 9165d4f98a2SYan Zheng if (!upper) { 9173fd0a558SYan, Zheng free_backref_edge(cache, edge); 9185d4f98a2SYan Zheng err = -ENOMEM; 9195d4f98a2SYan Zheng goto out; 9205d4f98a2SYan Zheng } 9215d4f98a2SYan Zheng upper->bytenr = eb->start; 9225d4f98a2SYan Zheng upper->owner = btrfs_header_owner(eb); 9235d4f98a2SYan Zheng upper->level = lower->level + 1; 9243fd0a558SYan, Zheng if (!root->ref_cows) 9253fd0a558SYan, Zheng upper->cowonly = 1; 9265d4f98a2SYan Zheng 9275d4f98a2SYan Zheng /* 9285d4f98a2SYan Zheng * if we know the block isn't shared 9295d4f98a2SYan Zheng * we can void checking its backrefs. 9305d4f98a2SYan Zheng */ 9315d4f98a2SYan Zheng if (btrfs_block_can_be_shared(root, eb)) 9325d4f98a2SYan Zheng upper->checked = 0; 9335d4f98a2SYan Zheng else 9345d4f98a2SYan Zheng upper->checked = 1; 9355d4f98a2SYan Zheng 9365d4f98a2SYan Zheng /* 9375d4f98a2SYan Zheng * add the block to pending list if we 9385d4f98a2SYan Zheng * need check its backrefs. only block 9395d4f98a2SYan Zheng * at 'cur->level + 1' is added to the 9405d4f98a2SYan Zheng * tail of pending list. this guarantees 9415d4f98a2SYan Zheng * we check backrefs from lower level 9425d4f98a2SYan Zheng * blocks to upper level blocks. 9435d4f98a2SYan Zheng */ 9445d4f98a2SYan Zheng if (!upper->checked && 9455d4f98a2SYan Zheng level == cur->level + 1) { 9465d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], 9475d4f98a2SYan Zheng &list); 9485d4f98a2SYan Zheng } else 9495d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 9505d4f98a2SYan Zheng } else { 9515d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 9525d4f98a2SYan Zheng rb_node); 9535d4f98a2SYan Zheng BUG_ON(!upper->checked); 9545d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 9553fd0a558SYan, Zheng if (!upper->owner) 9563fd0a558SYan, Zheng upper->owner = btrfs_header_owner(eb); 9575d4f98a2SYan Zheng } 9585d4f98a2SYan Zheng list_add_tail(&edge->list[LOWER], &lower->upper); 9595d4f98a2SYan Zheng edge->node[LOWER] = lower; 9603fd0a558SYan, Zheng edge->node[UPPER] = upper; 9615d4f98a2SYan Zheng 9625d4f98a2SYan Zheng if (rb_node) 9635d4f98a2SYan Zheng break; 9645d4f98a2SYan Zheng lower = upper; 9655d4f98a2SYan Zheng upper = NULL; 9665d4f98a2SYan Zheng } 967b3b4aa74SDavid Sterba btrfs_release_path(path2); 9685d4f98a2SYan Zheng next: 9695d4f98a2SYan Zheng if (ptr < end) { 9705d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 9715d4f98a2SYan Zheng if (ptr >= end) { 9725d4f98a2SYan Zheng WARN_ON(ptr > end); 9735d4f98a2SYan Zheng ptr = 0; 9745d4f98a2SYan Zheng end = 0; 9755d4f98a2SYan Zheng } 9765d4f98a2SYan Zheng } 9775d4f98a2SYan Zheng if (ptr >= end) 9785d4f98a2SYan Zheng path1->slots[0]++; 9795d4f98a2SYan Zheng } 980b3b4aa74SDavid Sterba btrfs_release_path(path1); 9815d4f98a2SYan Zheng 9825d4f98a2SYan Zheng cur->checked = 1; 9835d4f98a2SYan Zheng WARN_ON(exist); 9845d4f98a2SYan Zheng 9855d4f98a2SYan Zheng /* the pending list isn't empty, take the first block to process */ 9865d4f98a2SYan Zheng if (!list_empty(&list)) { 9875d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 9885d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 9895d4f98a2SYan Zheng cur = edge->node[UPPER]; 9905d4f98a2SYan Zheng goto again; 9915d4f98a2SYan Zheng } 9925d4f98a2SYan Zheng 9935d4f98a2SYan Zheng /* 9945d4f98a2SYan Zheng * everything goes well, connect backref nodes and insert backref nodes 9955d4f98a2SYan Zheng * into the cache. 9965d4f98a2SYan Zheng */ 9975d4f98a2SYan Zheng BUG_ON(!node->checked); 9983fd0a558SYan, Zheng cowonly = node->cowonly; 9993fd0a558SYan, Zheng if (!cowonly) { 10003fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, 10013fd0a558SYan, Zheng &node->rb_node); 10025d4f98a2SYan Zheng BUG_ON(rb_node); 10033fd0a558SYan, Zheng list_add_tail(&node->lower, &cache->leaves); 10043fd0a558SYan, Zheng } 10055d4f98a2SYan Zheng 10065d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) 10075d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 10085d4f98a2SYan Zheng 10095d4f98a2SYan Zheng while (!list_empty(&list)) { 10105d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 10115d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 10125d4f98a2SYan Zheng upper = edge->node[UPPER]; 10133fd0a558SYan, Zheng if (upper->detached) { 10143fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 10153fd0a558SYan, Zheng lower = edge->node[LOWER]; 10163fd0a558SYan, Zheng free_backref_edge(cache, edge); 10173fd0a558SYan, Zheng if (list_empty(&lower->upper)) 10183fd0a558SYan, Zheng list_add(&lower->list, &useless); 10193fd0a558SYan, Zheng continue; 10203fd0a558SYan, Zheng } 10215d4f98a2SYan Zheng 10225d4f98a2SYan Zheng if (!RB_EMPTY_NODE(&upper->rb_node)) { 10235d4f98a2SYan Zheng if (upper->lowest) { 10245d4f98a2SYan Zheng list_del_init(&upper->lower); 10255d4f98a2SYan Zheng upper->lowest = 0; 10265d4f98a2SYan Zheng } 10275d4f98a2SYan Zheng 10285d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 10295d4f98a2SYan Zheng continue; 10305d4f98a2SYan Zheng } 10315d4f98a2SYan Zheng 10325d4f98a2SYan Zheng BUG_ON(!upper->checked); 10333fd0a558SYan, Zheng BUG_ON(cowonly != upper->cowonly); 10343fd0a558SYan, Zheng if (!cowonly) { 10355d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, upper->bytenr, 10365d4f98a2SYan Zheng &upper->rb_node); 10375d4f98a2SYan Zheng BUG_ON(rb_node); 10383fd0a558SYan, Zheng } 10395d4f98a2SYan Zheng 10405d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 10415d4f98a2SYan Zheng 10425d4f98a2SYan Zheng list_for_each_entry(edge, &upper->upper, list[LOWER]) 10435d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 10445d4f98a2SYan Zheng } 10453fd0a558SYan, Zheng /* 10463fd0a558SYan, Zheng * process useless backref nodes. backref nodes for tree leaves 10473fd0a558SYan, Zheng * are deleted from the cache. backref nodes for upper level 10483fd0a558SYan, Zheng * tree blocks are left in the cache to avoid unnecessary backref 10493fd0a558SYan, Zheng * lookup. 10503fd0a558SYan, Zheng */ 10513fd0a558SYan, Zheng while (!list_empty(&useless)) { 10523fd0a558SYan, Zheng upper = list_entry(useless.next, struct backref_node, list); 10533fd0a558SYan, Zheng list_del_init(&upper->list); 10543fd0a558SYan, Zheng BUG_ON(!list_empty(&upper->upper)); 10553fd0a558SYan, Zheng if (upper == node) 10563fd0a558SYan, Zheng node = NULL; 10573fd0a558SYan, Zheng if (upper->lowest) { 10583fd0a558SYan, Zheng list_del_init(&upper->lower); 10593fd0a558SYan, Zheng upper->lowest = 0; 10603fd0a558SYan, Zheng } 10613fd0a558SYan, Zheng while (!list_empty(&upper->lower)) { 10623fd0a558SYan, Zheng edge = list_entry(upper->lower.next, 10633fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 10643fd0a558SYan, Zheng list_del(&edge->list[UPPER]); 10653fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 10663fd0a558SYan, Zheng lower = edge->node[LOWER]; 10673fd0a558SYan, Zheng free_backref_edge(cache, edge); 10683fd0a558SYan, Zheng 10693fd0a558SYan, Zheng if (list_empty(&lower->upper)) 10703fd0a558SYan, Zheng list_add(&lower->list, &useless); 10713fd0a558SYan, Zheng } 10723fd0a558SYan, Zheng __mark_block_processed(rc, upper); 10733fd0a558SYan, Zheng if (upper->level > 0) { 10743fd0a558SYan, Zheng list_add(&upper->list, &cache->detached); 10753fd0a558SYan, Zheng upper->detached = 1; 10763fd0a558SYan, Zheng } else { 10773fd0a558SYan, Zheng rb_erase(&upper->rb_node, &cache->rb_root); 10783fd0a558SYan, Zheng free_backref_node(cache, upper); 10793fd0a558SYan, Zheng } 10803fd0a558SYan, Zheng } 10815d4f98a2SYan Zheng out: 10825d4f98a2SYan Zheng btrfs_free_path(path1); 10835d4f98a2SYan Zheng btrfs_free_path(path2); 10845d4f98a2SYan Zheng if (err) { 10853fd0a558SYan, Zheng while (!list_empty(&useless)) { 10863fd0a558SYan, Zheng lower = list_entry(useless.next, 10873fd0a558SYan, Zheng struct backref_node, upper); 10883fd0a558SYan, Zheng list_del_init(&lower->upper); 10893fd0a558SYan, Zheng } 10905d4f98a2SYan Zheng upper = node; 10913fd0a558SYan, Zheng INIT_LIST_HEAD(&list); 10925d4f98a2SYan Zheng while (upper) { 10935d4f98a2SYan Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 10945d4f98a2SYan Zheng list_splice_tail(&upper->upper, &list); 10953fd0a558SYan, Zheng free_backref_node(cache, upper); 10965d4f98a2SYan Zheng } 10975d4f98a2SYan Zheng 10985d4f98a2SYan Zheng if (list_empty(&list)) 10995d4f98a2SYan Zheng break; 11005d4f98a2SYan Zheng 11015d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, 11025d4f98a2SYan Zheng list[LOWER]); 11033fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 11045d4f98a2SYan Zheng upper = edge->node[UPPER]; 11053fd0a558SYan, Zheng free_backref_edge(cache, edge); 11065d4f98a2SYan Zheng } 11075d4f98a2SYan Zheng return ERR_PTR(err); 11085d4f98a2SYan Zheng } 11093fd0a558SYan, Zheng BUG_ON(node && node->detached); 11105d4f98a2SYan Zheng return node; 11115d4f98a2SYan Zheng } 11125d4f98a2SYan Zheng 11135d4f98a2SYan Zheng /* 11143fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 11153fd0a558SYan, Zheng * the backref node is created by cloning backref node that 11163fd0a558SYan, Zheng * corresponds to root of source tree 11173fd0a558SYan, Zheng */ 11183fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 11193fd0a558SYan, Zheng struct reloc_control *rc, 11203fd0a558SYan, Zheng struct btrfs_root *src, 11213fd0a558SYan, Zheng struct btrfs_root *dest) 11223fd0a558SYan, Zheng { 11233fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 11243fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 11253fd0a558SYan, Zheng struct backref_node *node = NULL; 11263fd0a558SYan, Zheng struct backref_node *new_node; 11273fd0a558SYan, Zheng struct backref_edge *edge; 11283fd0a558SYan, Zheng struct backref_edge *new_edge; 11293fd0a558SYan, Zheng struct rb_node *rb_node; 11303fd0a558SYan, Zheng 11313fd0a558SYan, Zheng if (cache->last_trans > 0) 11323fd0a558SYan, Zheng update_backref_cache(trans, cache); 11333fd0a558SYan, Zheng 11343fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, src->commit_root->start); 11353fd0a558SYan, Zheng if (rb_node) { 11363fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, rb_node); 11373fd0a558SYan, Zheng if (node->detached) 11383fd0a558SYan, Zheng node = NULL; 11393fd0a558SYan, Zheng else 11403fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 11413fd0a558SYan, Zheng } 11423fd0a558SYan, Zheng 11433fd0a558SYan, Zheng if (!node) { 11443fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, 11453fd0a558SYan, Zheng reloc_root->commit_root->start); 11463fd0a558SYan, Zheng if (rb_node) { 11473fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, 11483fd0a558SYan, Zheng rb_node); 11493fd0a558SYan, Zheng BUG_ON(node->detached); 11503fd0a558SYan, Zheng } 11513fd0a558SYan, Zheng } 11523fd0a558SYan, Zheng 11533fd0a558SYan, Zheng if (!node) 11543fd0a558SYan, Zheng return 0; 11553fd0a558SYan, Zheng 11563fd0a558SYan, Zheng new_node = alloc_backref_node(cache); 11573fd0a558SYan, Zheng if (!new_node) 11583fd0a558SYan, Zheng return -ENOMEM; 11593fd0a558SYan, Zheng 11603fd0a558SYan, Zheng new_node->bytenr = dest->node->start; 11613fd0a558SYan, Zheng new_node->level = node->level; 11623fd0a558SYan, Zheng new_node->lowest = node->lowest; 11636848ad64SYan, Zheng new_node->checked = 1; 11643fd0a558SYan, Zheng new_node->root = dest; 11653fd0a558SYan, Zheng 11663fd0a558SYan, Zheng if (!node->lowest) { 11673fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 11683fd0a558SYan, Zheng new_edge = alloc_backref_edge(cache); 11693fd0a558SYan, Zheng if (!new_edge) 11703fd0a558SYan, Zheng goto fail; 11713fd0a558SYan, Zheng 11723fd0a558SYan, Zheng new_edge->node[UPPER] = new_node; 11733fd0a558SYan, Zheng new_edge->node[LOWER] = edge->node[LOWER]; 11743fd0a558SYan, Zheng list_add_tail(&new_edge->list[UPPER], 11753fd0a558SYan, Zheng &new_node->lower); 11763fd0a558SYan, Zheng } 11773fd0a558SYan, Zheng } 11783fd0a558SYan, Zheng 11793fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, new_node->bytenr, 11803fd0a558SYan, Zheng &new_node->rb_node); 11813fd0a558SYan, Zheng BUG_ON(rb_node); 11823fd0a558SYan, Zheng 11833fd0a558SYan, Zheng if (!new_node->lowest) { 11843fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 11853fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 11863fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 11873fd0a558SYan, Zheng } 11883fd0a558SYan, Zheng } 11893fd0a558SYan, Zheng return 0; 11903fd0a558SYan, Zheng fail: 11913fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 11923fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 11933fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 11943fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 11953fd0a558SYan, Zheng free_backref_edge(cache, new_edge); 11963fd0a558SYan, Zheng } 11973fd0a558SYan, Zheng free_backref_node(cache, new_node); 11983fd0a558SYan, Zheng return -ENOMEM; 11993fd0a558SYan, Zheng } 12003fd0a558SYan, Zheng 12013fd0a558SYan, Zheng /* 12025d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 12035d4f98a2SYan Zheng */ 12045d4f98a2SYan Zheng static int __add_reloc_root(struct btrfs_root *root) 12055d4f98a2SYan Zheng { 12065d4f98a2SYan Zheng struct rb_node *rb_node; 12075d4f98a2SYan Zheng struct mapping_node *node; 12085d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 12095d4f98a2SYan Zheng 12105d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 12115d4f98a2SYan Zheng BUG_ON(!node); 12125d4f98a2SYan Zheng 12135d4f98a2SYan Zheng node->bytenr = root->node->start; 12145d4f98a2SYan Zheng node->data = root; 12155d4f98a2SYan Zheng 12165d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12175d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 12185d4f98a2SYan Zheng node->bytenr, &node->rb_node); 12195d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 12205d4f98a2SYan Zheng BUG_ON(rb_node); 12215d4f98a2SYan Zheng 12225d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 12235d4f98a2SYan Zheng return 0; 12245d4f98a2SYan Zheng } 12255d4f98a2SYan Zheng 12265d4f98a2SYan Zheng /* 12275d4f98a2SYan Zheng * helper to update/delete the 'address of tree root -> reloc tree' 12285d4f98a2SYan Zheng * mapping 12295d4f98a2SYan Zheng */ 12305d4f98a2SYan Zheng static int __update_reloc_root(struct btrfs_root *root, int del) 12315d4f98a2SYan Zheng { 12325d4f98a2SYan Zheng struct rb_node *rb_node; 12335d4f98a2SYan Zheng struct mapping_node *node = NULL; 12345d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 12355d4f98a2SYan Zheng 12365d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12375d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, 12385d4f98a2SYan Zheng root->commit_root->start); 12395d4f98a2SYan Zheng if (rb_node) { 12405d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 12415d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 12425d4f98a2SYan Zheng } 12435d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 12445d4f98a2SYan Zheng 12455d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 12465d4f98a2SYan Zheng 12475d4f98a2SYan Zheng if (!del) { 12485d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12495d4f98a2SYan Zheng node->bytenr = root->node->start; 12505d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 12515d4f98a2SYan Zheng node->bytenr, &node->rb_node); 12525d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 12535d4f98a2SYan Zheng BUG_ON(rb_node); 12545d4f98a2SYan Zheng } else { 12555d4f98a2SYan Zheng list_del_init(&root->root_list); 12565d4f98a2SYan Zheng kfree(node); 12575d4f98a2SYan Zheng } 12585d4f98a2SYan Zheng return 0; 12595d4f98a2SYan Zheng } 12605d4f98a2SYan Zheng 12613fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 12623fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 12635d4f98a2SYan Zheng { 12645d4f98a2SYan Zheng struct btrfs_root *reloc_root; 12655d4f98a2SYan Zheng struct extent_buffer *eb; 12665d4f98a2SYan Zheng struct btrfs_root_item *root_item; 12675d4f98a2SYan Zheng struct btrfs_key root_key; 12685d4f98a2SYan Zheng int ret; 12695d4f98a2SYan Zheng 12705d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 12715d4f98a2SYan Zheng BUG_ON(!root_item); 12725d4f98a2SYan Zheng 12735d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 12745d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 12753fd0a558SYan, Zheng root_key.offset = objectid; 12765d4f98a2SYan Zheng 12773fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 12783fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 12795d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 12805d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 12815d4f98a2SYan Zheng BUG_ON(ret); 12825d4f98a2SYan Zheng 12833fd0a558SYan, Zheng btrfs_set_root_last_snapshot(&root->root_item, 12843fd0a558SYan, Zheng trans->transid - 1); 12853fd0a558SYan, Zheng } else { 12863fd0a558SYan, Zheng /* 12873fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 12883fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 12893fd0a558SYan, Zheng * modified after it was created have RELOC flag 12903fd0a558SYan, Zheng * set in their headers. so it's OK to not update 12913fd0a558SYan, Zheng * the 'last_snapshot'. 12923fd0a558SYan, Zheng */ 12933fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 12943fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 12953fd0a558SYan, Zheng BUG_ON(ret); 12963fd0a558SYan, Zheng } 12973fd0a558SYan, Zheng 12985d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 12995d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 13005d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 13015d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 13023fd0a558SYan, Zheng 13033fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 13043fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 13053fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 13063fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 13075d4f98a2SYan Zheng root_item->drop_level = 0; 13083fd0a558SYan, Zheng } 13095d4f98a2SYan Zheng 13105d4f98a2SYan Zheng btrfs_tree_unlock(eb); 13115d4f98a2SYan Zheng free_extent_buffer(eb); 13125d4f98a2SYan Zheng 13135d4f98a2SYan Zheng ret = btrfs_insert_root(trans, root->fs_info->tree_root, 13145d4f98a2SYan Zheng &root_key, root_item); 13155d4f98a2SYan Zheng BUG_ON(ret); 13165d4f98a2SYan Zheng kfree(root_item); 13175d4f98a2SYan Zheng 13185d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root, 13195d4f98a2SYan Zheng &root_key); 13205d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 13215d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 13223fd0a558SYan, Zheng return reloc_root; 13233fd0a558SYan, Zheng } 13243fd0a558SYan, Zheng 13253fd0a558SYan, Zheng /* 13263fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 13273fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 13283fd0a558SYan, Zheng */ 13293fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 13303fd0a558SYan, Zheng struct btrfs_root *root) 13313fd0a558SYan, Zheng { 13323fd0a558SYan, Zheng struct btrfs_root *reloc_root; 13333fd0a558SYan, Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 13343fd0a558SYan, Zheng int clear_rsv = 0; 13353fd0a558SYan, Zheng 13363fd0a558SYan, Zheng if (root->reloc_root) { 13373fd0a558SYan, Zheng reloc_root = root->reloc_root; 13383fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 13393fd0a558SYan, Zheng return 0; 13403fd0a558SYan, Zheng } 13413fd0a558SYan, Zheng 13423fd0a558SYan, Zheng if (!rc || !rc->create_reloc_tree || 13433fd0a558SYan, Zheng root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 13443fd0a558SYan, Zheng return 0; 13453fd0a558SYan, Zheng 13463fd0a558SYan, Zheng if (!trans->block_rsv) { 13473fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 13483fd0a558SYan, Zheng clear_rsv = 1; 13493fd0a558SYan, Zheng } 13503fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 13513fd0a558SYan, Zheng if (clear_rsv) 13523fd0a558SYan, Zheng trans->block_rsv = NULL; 13535d4f98a2SYan Zheng 13545d4f98a2SYan Zheng __add_reloc_root(reloc_root); 13555d4f98a2SYan Zheng root->reloc_root = reloc_root; 13565d4f98a2SYan Zheng return 0; 13575d4f98a2SYan Zheng } 13585d4f98a2SYan Zheng 13595d4f98a2SYan Zheng /* 13605d4f98a2SYan Zheng * update root item of reloc tree 13615d4f98a2SYan Zheng */ 13625d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 13635d4f98a2SYan Zheng struct btrfs_root *root) 13645d4f98a2SYan Zheng { 13655d4f98a2SYan Zheng struct btrfs_root *reloc_root; 13665d4f98a2SYan Zheng struct btrfs_root_item *root_item; 13675d4f98a2SYan Zheng int del = 0; 13685d4f98a2SYan Zheng int ret; 13695d4f98a2SYan Zheng 13705d4f98a2SYan Zheng if (!root->reloc_root) 13717585717fSChris Mason goto out; 13725d4f98a2SYan Zheng 13735d4f98a2SYan Zheng reloc_root = root->reloc_root; 13745d4f98a2SYan Zheng root_item = &reloc_root->root_item; 13755d4f98a2SYan Zheng 13763fd0a558SYan, Zheng if (root->fs_info->reloc_ctl->merge_reloc_tree && 13773fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 13785d4f98a2SYan Zheng root->reloc_root = NULL; 13795d4f98a2SYan Zheng del = 1; 13805d4f98a2SYan Zheng } 13815d4f98a2SYan Zheng 13825d4f98a2SYan Zheng __update_reloc_root(reloc_root, del); 13835d4f98a2SYan Zheng 13845d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 13855d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 13865d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 13875d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 13885d4f98a2SYan Zheng } 13895d4f98a2SYan Zheng 13905d4f98a2SYan Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 13915d4f98a2SYan Zheng &reloc_root->root_key, root_item); 13925d4f98a2SYan Zheng BUG_ON(ret); 13937585717fSChris Mason 13947585717fSChris Mason out: 13955d4f98a2SYan Zheng return 0; 13965d4f98a2SYan Zheng } 13975d4f98a2SYan Zheng 13985d4f98a2SYan Zheng /* 13995d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 14005d4f98a2SYan Zheng * in a subvolume 14015d4f98a2SYan Zheng */ 14025d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 14035d4f98a2SYan Zheng { 14045d4f98a2SYan Zheng struct rb_node *node; 14055d4f98a2SYan Zheng struct rb_node *prev; 14065d4f98a2SYan Zheng struct btrfs_inode *entry; 14075d4f98a2SYan Zheng struct inode *inode; 14085d4f98a2SYan Zheng 14095d4f98a2SYan Zheng spin_lock(&root->inode_lock); 14105d4f98a2SYan Zheng again: 14115d4f98a2SYan Zheng node = root->inode_tree.rb_node; 14125d4f98a2SYan Zheng prev = NULL; 14135d4f98a2SYan Zheng while (node) { 14145d4f98a2SYan Zheng prev = node; 14155d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 14165d4f98a2SYan Zheng 141733345d01SLi Zefan if (objectid < btrfs_ino(&entry->vfs_inode)) 14185d4f98a2SYan Zheng node = node->rb_left; 141933345d01SLi Zefan else if (objectid > btrfs_ino(&entry->vfs_inode)) 14205d4f98a2SYan Zheng node = node->rb_right; 14215d4f98a2SYan Zheng else 14225d4f98a2SYan Zheng break; 14235d4f98a2SYan Zheng } 14245d4f98a2SYan Zheng if (!node) { 14255d4f98a2SYan Zheng while (prev) { 14265d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 142733345d01SLi Zefan if (objectid <= btrfs_ino(&entry->vfs_inode)) { 14285d4f98a2SYan Zheng node = prev; 14295d4f98a2SYan Zheng break; 14305d4f98a2SYan Zheng } 14315d4f98a2SYan Zheng prev = rb_next(prev); 14325d4f98a2SYan Zheng } 14335d4f98a2SYan Zheng } 14345d4f98a2SYan Zheng while (node) { 14355d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 14365d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 14375d4f98a2SYan Zheng if (inode) { 14385d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 14395d4f98a2SYan Zheng return inode; 14405d4f98a2SYan Zheng } 14415d4f98a2SYan Zheng 144233345d01SLi Zefan objectid = btrfs_ino(&entry->vfs_inode) + 1; 14435d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 14445d4f98a2SYan Zheng goto again; 14455d4f98a2SYan Zheng 14465d4f98a2SYan Zheng node = rb_next(node); 14475d4f98a2SYan Zheng } 14485d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 14495d4f98a2SYan Zheng return NULL; 14505d4f98a2SYan Zheng } 14515d4f98a2SYan Zheng 14525d4f98a2SYan Zheng static int in_block_group(u64 bytenr, 14535d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group) 14545d4f98a2SYan Zheng { 14555d4f98a2SYan Zheng if (bytenr >= block_group->key.objectid && 14565d4f98a2SYan Zheng bytenr < block_group->key.objectid + block_group->key.offset) 14575d4f98a2SYan Zheng return 1; 14585d4f98a2SYan Zheng return 0; 14595d4f98a2SYan Zheng } 14605d4f98a2SYan Zheng 14615d4f98a2SYan Zheng /* 14625d4f98a2SYan Zheng * get new location of data 14635d4f98a2SYan Zheng */ 14645d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 14655d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 14665d4f98a2SYan Zheng { 14675d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 14685d4f98a2SYan Zheng struct btrfs_path *path; 14695d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 14705d4f98a2SYan Zheng struct extent_buffer *leaf; 14715d4f98a2SYan Zheng int ret; 14725d4f98a2SYan Zheng 14735d4f98a2SYan Zheng path = btrfs_alloc_path(); 14745d4f98a2SYan Zheng if (!path) 14755d4f98a2SYan Zheng return -ENOMEM; 14765d4f98a2SYan Zheng 14775d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 147833345d01SLi Zefan ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode), 14795d4f98a2SYan Zheng bytenr, 0); 14805d4f98a2SYan Zheng if (ret < 0) 14815d4f98a2SYan Zheng goto out; 14825d4f98a2SYan Zheng if (ret > 0) { 14835d4f98a2SYan Zheng ret = -ENOENT; 14845d4f98a2SYan Zheng goto out; 14855d4f98a2SYan Zheng } 14865d4f98a2SYan Zheng 14875d4f98a2SYan Zheng leaf = path->nodes[0]; 14885d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 14895d4f98a2SYan Zheng struct btrfs_file_extent_item); 14905d4f98a2SYan Zheng 14915d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 14925d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 14935d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 14945d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 14955d4f98a2SYan Zheng 14965d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 14975d4f98a2SYan Zheng ret = 1; 14985d4f98a2SYan Zheng goto out; 14995d4f98a2SYan Zheng } 15005d4f98a2SYan Zheng 15015d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 15025d4f98a2SYan Zheng ret = 0; 15035d4f98a2SYan Zheng out: 15045d4f98a2SYan Zheng btrfs_free_path(path); 15055d4f98a2SYan Zheng return ret; 15065d4f98a2SYan Zheng } 15075d4f98a2SYan Zheng 15085d4f98a2SYan Zheng /* 15095d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 15105d4f98a2SYan Zheng * the new locations. 15115d4f98a2SYan Zheng */ 15123fd0a558SYan, Zheng static noinline_for_stack 15133fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 15145d4f98a2SYan Zheng struct reloc_control *rc, 15155d4f98a2SYan Zheng struct btrfs_root *root, 15163fd0a558SYan, Zheng struct extent_buffer *leaf) 15175d4f98a2SYan Zheng { 15185d4f98a2SYan Zheng struct btrfs_key key; 15195d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 15205d4f98a2SYan Zheng struct inode *inode = NULL; 15215d4f98a2SYan Zheng u64 parent; 15225d4f98a2SYan Zheng u64 bytenr; 15233fd0a558SYan, Zheng u64 new_bytenr = 0; 15245d4f98a2SYan Zheng u64 num_bytes; 15255d4f98a2SYan Zheng u64 end; 15265d4f98a2SYan Zheng u32 nritems; 15275d4f98a2SYan Zheng u32 i; 15285d4f98a2SYan Zheng int ret; 15295d4f98a2SYan Zheng int first = 1; 15305d4f98a2SYan Zheng int dirty = 0; 15315d4f98a2SYan Zheng 15325d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 15335d4f98a2SYan Zheng return 0; 15345d4f98a2SYan Zheng 15355d4f98a2SYan Zheng /* reloc trees always use full backref */ 15365d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 15375d4f98a2SYan Zheng parent = leaf->start; 15385d4f98a2SYan Zheng else 15395d4f98a2SYan Zheng parent = 0; 15405d4f98a2SYan Zheng 15415d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 15425d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 15435d4f98a2SYan Zheng cond_resched(); 15445d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 15455d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 15465d4f98a2SYan Zheng continue; 15475d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 15485d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 15495d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 15505d4f98a2SYan Zheng continue; 15515d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 15525d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 15535d4f98a2SYan Zheng if (bytenr == 0) 15545d4f98a2SYan Zheng continue; 15555d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group)) 15565d4f98a2SYan Zheng continue; 15575d4f98a2SYan Zheng 15585d4f98a2SYan Zheng /* 15595d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 15605d4f98a2SYan Zheng * to complete and drop the extent cache 15615d4f98a2SYan Zheng */ 15625d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 15635d4f98a2SYan Zheng if (first) { 15645d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 15655d4f98a2SYan Zheng first = 0; 156633345d01SLi Zefan } else if (inode && btrfs_ino(inode) < key.objectid) { 15673fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 15685d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 15695d4f98a2SYan Zheng } 157033345d01SLi Zefan if (inode && btrfs_ino(inode) == key.objectid) { 15715d4f98a2SYan Zheng end = key.offset + 15725d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 15735d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 15745d4f98a2SYan Zheng root->sectorsize)); 15755d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 15765d4f98a2SYan Zheng end--; 15775d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 15785d4f98a2SYan Zheng key.offset, end, 15795d4f98a2SYan Zheng GFP_NOFS); 15805d4f98a2SYan Zheng if (!ret) 15815d4f98a2SYan Zheng continue; 15825d4f98a2SYan Zheng 15835d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, key.offset, end, 15845d4f98a2SYan Zheng 1); 15855d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 15865d4f98a2SYan Zheng key.offset, end, GFP_NOFS); 15875d4f98a2SYan Zheng } 15885d4f98a2SYan Zheng } 15895d4f98a2SYan Zheng 15905d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 15915d4f98a2SYan Zheng bytenr, num_bytes); 15923fd0a558SYan, Zheng if (ret > 0) { 15933fd0a558SYan, Zheng WARN_ON(1); 15945d4f98a2SYan Zheng continue; 15953fd0a558SYan, Zheng } 15965d4f98a2SYan Zheng BUG_ON(ret < 0); 15975d4f98a2SYan Zheng 15985d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 15995d4f98a2SYan Zheng dirty = 1; 16005d4f98a2SYan Zheng 16015d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 16025d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, new_bytenr, 16035d4f98a2SYan Zheng num_bytes, parent, 16045d4f98a2SYan Zheng btrfs_header_owner(leaf), 16055d4f98a2SYan Zheng key.objectid, key.offset); 16065d4f98a2SYan Zheng BUG_ON(ret); 16075d4f98a2SYan Zheng 16085d4f98a2SYan Zheng ret = btrfs_free_extent(trans, root, bytenr, num_bytes, 16095d4f98a2SYan Zheng parent, btrfs_header_owner(leaf), 16105d4f98a2SYan Zheng key.objectid, key.offset); 16115d4f98a2SYan Zheng BUG_ON(ret); 16125d4f98a2SYan Zheng } 16135d4f98a2SYan Zheng if (dirty) 16145d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 16153fd0a558SYan, Zheng if (inode) 16163fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 16175d4f98a2SYan Zheng return 0; 16185d4f98a2SYan Zheng } 16195d4f98a2SYan Zheng 16205d4f98a2SYan Zheng static noinline_for_stack 16215d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 16225d4f98a2SYan Zheng struct btrfs_path *path, int level) 16235d4f98a2SYan Zheng { 16245d4f98a2SYan Zheng struct btrfs_disk_key key1; 16255d4f98a2SYan Zheng struct btrfs_disk_key key2; 16265d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 16275d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 16285d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 16295d4f98a2SYan Zheng } 16305d4f98a2SYan Zheng 16315d4f98a2SYan Zheng /* 16325d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 16335d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 16345d4f98a2SYan Zheng * reloc tree was create can be replaced. 16355d4f98a2SYan Zheng * 16365d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 16375d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 16385d4f98a2SYan Zheng * errors, a negative error number is returned. 16395d4f98a2SYan Zheng */ 16403fd0a558SYan, Zheng static noinline_for_stack 16413fd0a558SYan, Zheng int replace_path(struct btrfs_trans_handle *trans, 16425d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 16435d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 16445d4f98a2SYan Zheng int lowest_level, int max_level) 16455d4f98a2SYan Zheng { 16465d4f98a2SYan Zheng struct extent_buffer *eb; 16475d4f98a2SYan Zheng struct extent_buffer *parent; 16485d4f98a2SYan Zheng struct btrfs_key key; 16495d4f98a2SYan Zheng u64 old_bytenr; 16505d4f98a2SYan Zheng u64 new_bytenr; 16515d4f98a2SYan Zheng u64 old_ptr_gen; 16525d4f98a2SYan Zheng u64 new_ptr_gen; 16535d4f98a2SYan Zheng u64 last_snapshot; 16545d4f98a2SYan Zheng u32 blocksize; 16553fd0a558SYan, Zheng int cow = 0; 16565d4f98a2SYan Zheng int level; 16575d4f98a2SYan Zheng int ret; 16585d4f98a2SYan Zheng int slot; 16595d4f98a2SYan Zheng 16605d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 16615d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 16625d4f98a2SYan Zheng 16635d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 16643fd0a558SYan, Zheng again: 16655d4f98a2SYan Zheng slot = path->slots[lowest_level]; 16665d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 16675d4f98a2SYan Zheng 16685d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 16695d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 16705d4f98a2SYan Zheng level = btrfs_header_level(eb); 16715d4f98a2SYan Zheng 16725d4f98a2SYan Zheng if (level < lowest_level) { 16735d4f98a2SYan Zheng btrfs_tree_unlock(eb); 16745d4f98a2SYan Zheng free_extent_buffer(eb); 16755d4f98a2SYan Zheng return 0; 16765d4f98a2SYan Zheng } 16775d4f98a2SYan Zheng 16783fd0a558SYan, Zheng if (cow) { 16795d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 16805d4f98a2SYan Zheng BUG_ON(ret); 16813fd0a558SYan, Zheng } 16825d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 16835d4f98a2SYan Zheng 16845d4f98a2SYan Zheng if (next_key) { 16855d4f98a2SYan Zheng next_key->objectid = (u64)-1; 16865d4f98a2SYan Zheng next_key->type = (u8)-1; 16875d4f98a2SYan Zheng next_key->offset = (u64)-1; 16885d4f98a2SYan Zheng } 16895d4f98a2SYan Zheng 16905d4f98a2SYan Zheng parent = eb; 16915d4f98a2SYan Zheng while (1) { 16925d4f98a2SYan Zheng level = btrfs_header_level(parent); 16935d4f98a2SYan Zheng BUG_ON(level < lowest_level); 16945d4f98a2SYan Zheng 16955d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 16965d4f98a2SYan Zheng if (ret && slot > 0) 16975d4f98a2SYan Zheng slot--; 16985d4f98a2SYan Zheng 16995d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 17005d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 17015d4f98a2SYan Zheng 17025d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 17035d4f98a2SYan Zheng blocksize = btrfs_level_size(dest, level - 1); 17045d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 17055d4f98a2SYan Zheng 17065d4f98a2SYan Zheng if (level <= max_level) { 17075d4f98a2SYan Zheng eb = path->nodes[level]; 17085d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 17095d4f98a2SYan Zheng path->slots[level]); 17105d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 17115d4f98a2SYan Zheng path->slots[level]); 17125d4f98a2SYan Zheng } else { 17135d4f98a2SYan Zheng new_bytenr = 0; 17145d4f98a2SYan Zheng new_ptr_gen = 0; 17155d4f98a2SYan Zheng } 17165d4f98a2SYan Zheng 17175d4f98a2SYan Zheng if (new_bytenr > 0 && new_bytenr == old_bytenr) { 17185d4f98a2SYan Zheng WARN_ON(1); 17195d4f98a2SYan Zheng ret = level; 17205d4f98a2SYan Zheng break; 17215d4f98a2SYan Zheng } 17225d4f98a2SYan Zheng 17235d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 17245d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 17253fd0a558SYan, Zheng if (level <= lowest_level) { 17265d4f98a2SYan Zheng ret = 0; 17275d4f98a2SYan Zheng break; 17285d4f98a2SYan Zheng } 17295d4f98a2SYan Zheng 17305d4f98a2SYan Zheng eb = read_tree_block(dest, old_bytenr, blocksize, 17315d4f98a2SYan Zheng old_ptr_gen); 173297d9a8a4STsutomu Itoh BUG_ON(!eb); 17335d4f98a2SYan Zheng btrfs_tree_lock(eb); 17343fd0a558SYan, Zheng if (cow) { 17355d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 17365d4f98a2SYan Zheng slot, &eb); 17375d4f98a2SYan Zheng BUG_ON(ret); 17385d4f98a2SYan Zheng } 17393fd0a558SYan, Zheng btrfs_set_lock_blocking(eb); 17405d4f98a2SYan Zheng 17415d4f98a2SYan Zheng btrfs_tree_unlock(parent); 17425d4f98a2SYan Zheng free_extent_buffer(parent); 17435d4f98a2SYan Zheng 17445d4f98a2SYan Zheng parent = eb; 17455d4f98a2SYan Zheng continue; 17465d4f98a2SYan Zheng } 17475d4f98a2SYan Zheng 17483fd0a558SYan, Zheng if (!cow) { 17493fd0a558SYan, Zheng btrfs_tree_unlock(parent); 17503fd0a558SYan, Zheng free_extent_buffer(parent); 17513fd0a558SYan, Zheng cow = 1; 17523fd0a558SYan, Zheng goto again; 17533fd0a558SYan, Zheng } 17543fd0a558SYan, Zheng 17555d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17565d4f98a2SYan Zheng path->slots[level]); 1757b3b4aa74SDavid Sterba btrfs_release_path(path); 17585d4f98a2SYan Zheng 17595d4f98a2SYan Zheng path->lowest_level = level; 17605d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 17615d4f98a2SYan Zheng path->lowest_level = 0; 17625d4f98a2SYan Zheng BUG_ON(ret); 17635d4f98a2SYan Zheng 17645d4f98a2SYan Zheng /* 17655d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 17665d4f98a2SYan Zheng */ 17675d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 17685d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 17695d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 17705d4f98a2SYan Zheng 17715d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 17725d4f98a2SYan Zheng path->slots[level], old_bytenr); 17735d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 17745d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 17755d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 17765d4f98a2SYan Zheng 17775d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize, 17785d4f98a2SYan Zheng path->nodes[level]->start, 17795d4f98a2SYan Zheng src->root_key.objectid, level - 1, 0); 17805d4f98a2SYan Zheng BUG_ON(ret); 17815d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize, 17825d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 17835d4f98a2SYan Zheng 0); 17845d4f98a2SYan Zheng BUG_ON(ret); 17855d4f98a2SYan Zheng 17865d4f98a2SYan Zheng ret = btrfs_free_extent(trans, src, new_bytenr, blocksize, 17875d4f98a2SYan Zheng path->nodes[level]->start, 17885d4f98a2SYan Zheng src->root_key.objectid, level - 1, 0); 17895d4f98a2SYan Zheng BUG_ON(ret); 17905d4f98a2SYan Zheng 17915d4f98a2SYan Zheng ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize, 17925d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 17935d4f98a2SYan Zheng 0); 17945d4f98a2SYan Zheng BUG_ON(ret); 17955d4f98a2SYan Zheng 17965d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 17975d4f98a2SYan Zheng 17985d4f98a2SYan Zheng ret = level; 17995d4f98a2SYan Zheng break; 18005d4f98a2SYan Zheng } 18015d4f98a2SYan Zheng btrfs_tree_unlock(parent); 18025d4f98a2SYan Zheng free_extent_buffer(parent); 18035d4f98a2SYan Zheng return ret; 18045d4f98a2SYan Zheng } 18055d4f98a2SYan Zheng 18065d4f98a2SYan Zheng /* 18075d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 18085d4f98a2SYan Zheng */ 18095d4f98a2SYan Zheng static noinline_for_stack 18105d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 18115d4f98a2SYan Zheng int *level) 18125d4f98a2SYan Zheng { 18135d4f98a2SYan Zheng struct extent_buffer *eb; 18145d4f98a2SYan Zheng int i; 18155d4f98a2SYan Zheng u64 last_snapshot; 18165d4f98a2SYan Zheng u32 nritems; 18175d4f98a2SYan Zheng 18185d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 18195d4f98a2SYan Zheng 18205d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 18215d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 18225d4f98a2SYan Zheng path->nodes[i] = NULL; 18235d4f98a2SYan Zheng } 18245d4f98a2SYan Zheng 18255d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 18265d4f98a2SYan Zheng eb = path->nodes[i]; 18275d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 18285d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 18295d4f98a2SYan Zheng path->slots[i]++; 18305d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 18315d4f98a2SYan Zheng last_snapshot) 18325d4f98a2SYan Zheng continue; 18335d4f98a2SYan Zheng 18345d4f98a2SYan Zheng *level = i; 18355d4f98a2SYan Zheng return 0; 18365d4f98a2SYan Zheng } 18375d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 18385d4f98a2SYan Zheng path->nodes[i] = NULL; 18395d4f98a2SYan Zheng } 18405d4f98a2SYan Zheng return 1; 18415d4f98a2SYan Zheng } 18425d4f98a2SYan Zheng 18435d4f98a2SYan Zheng /* 18445d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 18455d4f98a2SYan Zheng */ 18465d4f98a2SYan Zheng static noinline_for_stack 18475d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 18485d4f98a2SYan Zheng int *level) 18495d4f98a2SYan Zheng { 18505d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 18515d4f98a2SYan Zheng int i; 18525d4f98a2SYan Zheng u64 bytenr; 18535d4f98a2SYan Zheng u64 ptr_gen = 0; 18545d4f98a2SYan Zheng u64 last_snapshot; 18555d4f98a2SYan Zheng u32 blocksize; 18565d4f98a2SYan Zheng u32 nritems; 18575d4f98a2SYan Zheng 18585d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 18595d4f98a2SYan Zheng 18605d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 18615d4f98a2SYan Zheng eb = path->nodes[i]; 18625d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 18635d4f98a2SYan Zheng while (path->slots[i] < nritems) { 18645d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 18655d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 18665d4f98a2SYan Zheng break; 18675d4f98a2SYan Zheng path->slots[i]++; 18685d4f98a2SYan Zheng } 18695d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 18705d4f98a2SYan Zheng if (i == *level) 18715d4f98a2SYan Zheng break; 18725d4f98a2SYan Zheng *level = i + 1; 18735d4f98a2SYan Zheng return 0; 18745d4f98a2SYan Zheng } 18755d4f98a2SYan Zheng if (i == 1) { 18765d4f98a2SYan Zheng *level = i; 18775d4f98a2SYan Zheng return 0; 18785d4f98a2SYan Zheng } 18795d4f98a2SYan Zheng 18805d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 18815d4f98a2SYan Zheng blocksize = btrfs_level_size(root, i - 1); 18825d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, ptr_gen); 18835d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 18845d4f98a2SYan Zheng path->nodes[i - 1] = eb; 18855d4f98a2SYan Zheng path->slots[i - 1] = 0; 18865d4f98a2SYan Zheng } 18875d4f98a2SYan Zheng return 1; 18885d4f98a2SYan Zheng } 18895d4f98a2SYan Zheng 18905d4f98a2SYan Zheng /* 18915d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 18925d4f98a2SYan Zheng * [min_key, max_key) 18935d4f98a2SYan Zheng */ 18945d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 18955d4f98a2SYan Zheng struct btrfs_key *min_key, 18965d4f98a2SYan Zheng struct btrfs_key *max_key) 18975d4f98a2SYan Zheng { 18985d4f98a2SYan Zheng struct inode *inode = NULL; 18995d4f98a2SYan Zheng u64 objectid; 19005d4f98a2SYan Zheng u64 start, end; 190133345d01SLi Zefan u64 ino; 19025d4f98a2SYan Zheng 19035d4f98a2SYan Zheng objectid = min_key->objectid; 19045d4f98a2SYan Zheng while (1) { 19055d4f98a2SYan Zheng cond_resched(); 19065d4f98a2SYan Zheng iput(inode); 19075d4f98a2SYan Zheng 19085d4f98a2SYan Zheng if (objectid > max_key->objectid) 19095d4f98a2SYan Zheng break; 19105d4f98a2SYan Zheng 19115d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 19125d4f98a2SYan Zheng if (!inode) 19135d4f98a2SYan Zheng break; 191433345d01SLi Zefan ino = btrfs_ino(inode); 19155d4f98a2SYan Zheng 191633345d01SLi Zefan if (ino > max_key->objectid) { 19175d4f98a2SYan Zheng iput(inode); 19185d4f98a2SYan Zheng break; 19195d4f98a2SYan Zheng } 19205d4f98a2SYan Zheng 192133345d01SLi Zefan objectid = ino + 1; 19225d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 19235d4f98a2SYan Zheng continue; 19245d4f98a2SYan Zheng 192533345d01SLi Zefan if (unlikely(min_key->objectid == ino)) { 19265d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 19275d4f98a2SYan Zheng continue; 19285d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 19295d4f98a2SYan Zheng start = 0; 19305d4f98a2SYan Zheng else { 19315d4f98a2SYan Zheng start = min_key->offset; 19325d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(start, root->sectorsize)); 19335d4f98a2SYan Zheng } 19345d4f98a2SYan Zheng } else { 19355d4f98a2SYan Zheng start = 0; 19365d4f98a2SYan Zheng } 19375d4f98a2SYan Zheng 193833345d01SLi Zefan if (unlikely(max_key->objectid == ino)) { 19395d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 19405d4f98a2SYan Zheng continue; 19415d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 19425d4f98a2SYan Zheng end = (u64)-1; 19435d4f98a2SYan Zheng } else { 19445d4f98a2SYan Zheng if (max_key->offset == 0) 19455d4f98a2SYan Zheng continue; 19465d4f98a2SYan Zheng end = max_key->offset; 19475d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 19485d4f98a2SYan Zheng end--; 19495d4f98a2SYan Zheng } 19505d4f98a2SYan Zheng } else { 19515d4f98a2SYan Zheng end = (u64)-1; 19525d4f98a2SYan Zheng } 19535d4f98a2SYan Zheng 19545d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 19555d4f98a2SYan Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 19565d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 1); 19575d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 19585d4f98a2SYan Zheng } 19595d4f98a2SYan Zheng return 0; 19605d4f98a2SYan Zheng } 19615d4f98a2SYan Zheng 19625d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 19635d4f98a2SYan Zheng struct btrfs_key *key) 19645d4f98a2SYan Zheng 19655d4f98a2SYan Zheng { 19665d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 19675d4f98a2SYan Zheng if (!path->nodes[level]) 19685d4f98a2SYan Zheng break; 19695d4f98a2SYan Zheng if (path->slots[level] + 1 < 19705d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 19715d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 19725d4f98a2SYan Zheng path->slots[level] + 1); 19735d4f98a2SYan Zheng return 0; 19745d4f98a2SYan Zheng } 19755d4f98a2SYan Zheng level++; 19765d4f98a2SYan Zheng } 19775d4f98a2SYan Zheng return 1; 19785d4f98a2SYan Zheng } 19795d4f98a2SYan Zheng 19805d4f98a2SYan Zheng /* 19815d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 19825d4f98a2SYan Zheng * fs tree. 19835d4f98a2SYan Zheng */ 19845d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 19855d4f98a2SYan Zheng struct btrfs_root *root) 19865d4f98a2SYan Zheng { 19875d4f98a2SYan Zheng LIST_HEAD(inode_list); 19885d4f98a2SYan Zheng struct btrfs_key key; 19895d4f98a2SYan Zheng struct btrfs_key next_key; 19905d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 19915d4f98a2SYan Zheng struct btrfs_root *reloc_root; 19925d4f98a2SYan Zheng struct btrfs_root_item *root_item; 19935d4f98a2SYan Zheng struct btrfs_path *path; 19943fd0a558SYan, Zheng struct extent_buffer *leaf; 19955d4f98a2SYan Zheng unsigned long nr; 19965d4f98a2SYan Zheng int level; 19975d4f98a2SYan Zheng int max_level; 19985d4f98a2SYan Zheng int replaced = 0; 19995d4f98a2SYan Zheng int ret; 20005d4f98a2SYan Zheng int err = 0; 20013fd0a558SYan, Zheng u32 min_reserved; 20025d4f98a2SYan Zheng 20035d4f98a2SYan Zheng path = btrfs_alloc_path(); 20045d4f98a2SYan Zheng if (!path) 20055d4f98a2SYan Zheng return -ENOMEM; 2006026fd317SJosef Bacik path->reada = 1; 20075d4f98a2SYan Zheng 20085d4f98a2SYan Zheng reloc_root = root->reloc_root; 20095d4f98a2SYan Zheng root_item = &reloc_root->root_item; 20105d4f98a2SYan Zheng 20115d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 20125d4f98a2SYan Zheng level = btrfs_root_level(root_item); 20135d4f98a2SYan Zheng extent_buffer_get(reloc_root->node); 20145d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 20155d4f98a2SYan Zheng path->slots[level] = 0; 20165d4f98a2SYan Zheng } else { 20175d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 20185d4f98a2SYan Zheng 20195d4f98a2SYan Zheng level = root_item->drop_level; 20205d4f98a2SYan Zheng BUG_ON(level == 0); 20215d4f98a2SYan Zheng path->lowest_level = level; 20225d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 202333c66f43SYan Zheng path->lowest_level = 0; 20245d4f98a2SYan Zheng if (ret < 0) { 20255d4f98a2SYan Zheng btrfs_free_path(path); 20265d4f98a2SYan Zheng return ret; 20275d4f98a2SYan Zheng } 20285d4f98a2SYan Zheng 20295d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 20305d4f98a2SYan Zheng path->slots[level]); 20315d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 20325d4f98a2SYan Zheng 20335d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 20345d4f98a2SYan Zheng } 20355d4f98a2SYan Zheng 20363fd0a558SYan, Zheng min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 20375d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 20385d4f98a2SYan Zheng 20395d4f98a2SYan Zheng while (1) { 2040a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 0); 204198d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 20423fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 20433fd0a558SYan, Zheng 204436ba022aSJosef Bacik ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved); 20453fd0a558SYan, Zheng if (ret) { 20463fd0a558SYan, Zheng BUG_ON(ret != -EAGAIN); 20473fd0a558SYan, Zheng ret = btrfs_commit_transaction(trans, root); 20483fd0a558SYan, Zheng BUG_ON(ret); 20493fd0a558SYan, Zheng continue; 20503fd0a558SYan, Zheng } 20513fd0a558SYan, Zheng 20523fd0a558SYan, Zheng replaced = 0; 20535d4f98a2SYan Zheng max_level = level; 20545d4f98a2SYan Zheng 20555d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 20565d4f98a2SYan Zheng if (ret < 0) { 20575d4f98a2SYan Zheng err = ret; 20585d4f98a2SYan Zheng goto out; 20595d4f98a2SYan Zheng } 20605d4f98a2SYan Zheng if (ret > 0) 20615d4f98a2SYan Zheng break; 20625d4f98a2SYan Zheng 20635d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 20645d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 20655d4f98a2SYan Zheng ret = 0; 20665d4f98a2SYan Zheng } else { 20673fd0a558SYan, Zheng ret = replace_path(trans, root, reloc_root, path, 20683fd0a558SYan, Zheng &next_key, level, max_level); 20695d4f98a2SYan Zheng } 20705d4f98a2SYan Zheng if (ret < 0) { 20715d4f98a2SYan Zheng err = ret; 20725d4f98a2SYan Zheng goto out; 20735d4f98a2SYan Zheng } 20745d4f98a2SYan Zheng 20755d4f98a2SYan Zheng if (ret > 0) { 20765d4f98a2SYan Zheng level = ret; 20775d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 20785d4f98a2SYan Zheng path->slots[level]); 20795d4f98a2SYan Zheng replaced = 1; 20805d4f98a2SYan Zheng } 20815d4f98a2SYan Zheng 20825d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 20835d4f98a2SYan Zheng if (ret > 0) 20845d4f98a2SYan Zheng break; 20855d4f98a2SYan Zheng 20865d4f98a2SYan Zheng BUG_ON(level == 0); 20875d4f98a2SYan Zheng /* 20885d4f98a2SYan Zheng * save the merging progress in the drop_progress. 20895d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 20905d4f98a2SYan Zheng */ 20915d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 20925d4f98a2SYan Zheng path->slots[level]); 20935d4f98a2SYan Zheng root_item->drop_level = level; 20945d4f98a2SYan Zheng 20955d4f98a2SYan Zheng nr = trans->blocks_used; 20963fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, root); 20975d4f98a2SYan Zheng 20985d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 20995d4f98a2SYan Zheng 21005d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 21015d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 21025d4f98a2SYan Zheng } 21035d4f98a2SYan Zheng 21045d4f98a2SYan Zheng /* 21055d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 21065d4f98a2SYan Zheng * relocated and the block is tree root. 21075d4f98a2SYan Zheng */ 21085d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 21095d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 21105d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 21115d4f98a2SYan Zheng free_extent_buffer(leaf); 21125d4f98a2SYan Zheng if (ret < 0) 21135d4f98a2SYan Zheng err = ret; 21145d4f98a2SYan Zheng out: 21155d4f98a2SYan Zheng btrfs_free_path(path); 21165d4f98a2SYan Zheng 21175d4f98a2SYan Zheng if (err == 0) { 21185d4f98a2SYan Zheng memset(&root_item->drop_progress, 0, 21195d4f98a2SYan Zheng sizeof(root_item->drop_progress)); 21205d4f98a2SYan Zheng root_item->drop_level = 0; 21215d4f98a2SYan Zheng btrfs_set_root_refs(root_item, 0); 21223fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 21235d4f98a2SYan Zheng } 21245d4f98a2SYan Zheng 21255d4f98a2SYan Zheng nr = trans->blocks_used; 21263fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, root); 21275d4f98a2SYan Zheng 21285d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 21295d4f98a2SYan Zheng 21305d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 21315d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 21325d4f98a2SYan Zheng 21335d4f98a2SYan Zheng return err; 21345d4f98a2SYan Zheng } 21355d4f98a2SYan Zheng 21363fd0a558SYan, Zheng static noinline_for_stack 21373fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 21385d4f98a2SYan Zheng { 21393fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 21403fd0a558SYan, Zheng struct btrfs_root *reloc_root; 21415d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 21423fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 21433fd0a558SYan, Zheng u64 num_bytes = 0; 21443fd0a558SYan, Zheng int ret; 21453fd0a558SYan, Zheng 21467585717fSChris Mason mutex_lock(&root->fs_info->reloc_mutex); 21473fd0a558SYan, Zheng rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 21483fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 21497585717fSChris Mason mutex_unlock(&root->fs_info->reloc_mutex); 21507585717fSChris Mason 21513fd0a558SYan, Zheng again: 21523fd0a558SYan, Zheng if (!err) { 21533fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 21544a92b1b8SJosef Bacik ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes); 21553fd0a558SYan, Zheng if (ret) 21563fd0a558SYan, Zheng err = ret; 21573fd0a558SYan, Zheng } 21583fd0a558SYan, Zheng 21597a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 21603612b495STsutomu Itoh if (IS_ERR(trans)) { 21613612b495STsutomu Itoh if (!err) 21623612b495STsutomu Itoh btrfs_block_rsv_release(rc->extent_root, 21633612b495STsutomu Itoh rc->block_rsv, num_bytes); 21643612b495STsutomu Itoh return PTR_ERR(trans); 21653612b495STsutomu Itoh } 21663fd0a558SYan, Zheng 21673fd0a558SYan, Zheng if (!err) { 21683fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 21693fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 21703fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, 21713fd0a558SYan, Zheng rc->block_rsv, num_bytes); 21723fd0a558SYan, Zheng goto again; 21733fd0a558SYan, Zheng } 21743fd0a558SYan, Zheng } 21753fd0a558SYan, Zheng 21763fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 21773fd0a558SYan, Zheng 21783fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 21793fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 21803fd0a558SYan, Zheng struct btrfs_root, root_list); 21813fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 21823fd0a558SYan, Zheng 21833fd0a558SYan, Zheng root = read_fs_root(reloc_root->fs_info, 21843fd0a558SYan, Zheng reloc_root->root_key.offset); 21853fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 21863fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 21873fd0a558SYan, Zheng 21883fd0a558SYan, Zheng /* 21893fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 21903fd0a558SYan, Zheng * knows it should resumes merging 21913fd0a558SYan, Zheng */ 21923fd0a558SYan, Zheng if (!err) 21933fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 21943fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 21953fd0a558SYan, Zheng 21963fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 21973fd0a558SYan, Zheng } 21983fd0a558SYan, Zheng 21993fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 22003fd0a558SYan, Zheng 22013fd0a558SYan, Zheng if (!err) 22023fd0a558SYan, Zheng btrfs_commit_transaction(trans, rc->extent_root); 22033fd0a558SYan, Zheng else 22043fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 22053fd0a558SYan, Zheng return err; 22063fd0a558SYan, Zheng } 22073fd0a558SYan, Zheng 22083fd0a558SYan, Zheng static noinline_for_stack 22093fd0a558SYan, Zheng int merge_reloc_roots(struct reloc_control *rc) 22103fd0a558SYan, Zheng { 22115d4f98a2SYan Zheng struct btrfs_root *root; 22125d4f98a2SYan Zheng struct btrfs_root *reloc_root; 22133fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 22143fd0a558SYan, Zheng int found = 0; 22153fd0a558SYan, Zheng int ret; 22163fd0a558SYan, Zheng again: 22173fd0a558SYan, Zheng root = rc->extent_root; 22187585717fSChris Mason 22197585717fSChris Mason /* 22207585717fSChris Mason * this serializes us with btrfs_record_root_in_transaction, 22217585717fSChris Mason * we have to make sure nobody is in the middle of 22227585717fSChris Mason * adding their roots to the list while we are 22237585717fSChris Mason * doing this splice 22247585717fSChris Mason */ 22257585717fSChris Mason mutex_lock(&root->fs_info->reloc_mutex); 22263fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 22277585717fSChris Mason mutex_unlock(&root->fs_info->reloc_mutex); 22285d4f98a2SYan Zheng 22293fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 22303fd0a558SYan, Zheng found = 1; 22313fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 22323fd0a558SYan, Zheng struct btrfs_root, root_list); 22335d4f98a2SYan Zheng 22345d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 22355d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, 22365d4f98a2SYan Zheng reloc_root->root_key.offset); 22375d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 22385d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 22395d4f98a2SYan Zheng 22403fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 22413fd0a558SYan, Zheng BUG_ON(ret); 22423fd0a558SYan, Zheng } else { 22433fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 22443fd0a558SYan, Zheng } 22453fd0a558SYan, Zheng btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0); 22465d4f98a2SYan Zheng } 22475d4f98a2SYan Zheng 22483fd0a558SYan, Zheng if (found) { 22493fd0a558SYan, Zheng found = 0; 22503fd0a558SYan, Zheng goto again; 22515d4f98a2SYan Zheng } 22525d4f98a2SYan Zheng BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 22535d4f98a2SYan Zheng return 0; 22545d4f98a2SYan Zheng } 22555d4f98a2SYan Zheng 22565d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 22575d4f98a2SYan Zheng { 22585d4f98a2SYan Zheng struct tree_block *block; 22595d4f98a2SYan Zheng struct rb_node *rb_node; 22605d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 22615d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 22625d4f98a2SYan Zheng rb_erase(rb_node, blocks); 22635d4f98a2SYan Zheng kfree(block); 22645d4f98a2SYan Zheng } 22655d4f98a2SYan Zheng } 22665d4f98a2SYan Zheng 22675d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 22685d4f98a2SYan Zheng struct btrfs_root *reloc_root) 22695d4f98a2SYan Zheng { 22705d4f98a2SYan Zheng struct btrfs_root *root; 22715d4f98a2SYan Zheng 22725d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 22735d4f98a2SYan Zheng return 0; 22745d4f98a2SYan Zheng 22755d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset); 22765d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 22775d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 22785d4f98a2SYan Zheng 22795d4f98a2SYan Zheng return btrfs_record_root_in_trans(trans, root); 22805d4f98a2SYan Zheng } 22815d4f98a2SYan Zheng 22823fd0a558SYan, Zheng static noinline_for_stack 22833fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 22843fd0a558SYan, Zheng struct reloc_control *rc, 22855d4f98a2SYan Zheng struct backref_node *node, 22863fd0a558SYan, Zheng struct backref_edge *edges[], int *nr) 22875d4f98a2SYan Zheng { 22885d4f98a2SYan Zheng struct backref_node *next; 22895d4f98a2SYan Zheng struct btrfs_root *root; 22903fd0a558SYan, Zheng int index = 0; 22913fd0a558SYan, Zheng 22925d4f98a2SYan Zheng next = node; 22935d4f98a2SYan Zheng while (1) { 22945d4f98a2SYan Zheng cond_resched(); 22955d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 22965d4f98a2SYan Zheng root = next->root; 22973fd0a558SYan, Zheng BUG_ON(!root); 22983fd0a558SYan, Zheng BUG_ON(!root->ref_cows); 22995d4f98a2SYan Zheng 23005d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 23015d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 23025d4f98a2SYan Zheng break; 23035d4f98a2SYan Zheng } 23045d4f98a2SYan Zheng 23055d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 23063fd0a558SYan, Zheng root = root->reloc_root; 23073fd0a558SYan, Zheng 23083fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 23093fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 23103fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 23113fd0a558SYan, Zheng next->new_bytenr = root->node->start; 23123fd0a558SYan, Zheng next->root = root; 23133fd0a558SYan, Zheng list_add_tail(&next->list, 23143fd0a558SYan, Zheng &rc->backref_cache.changed); 23153fd0a558SYan, Zheng __mark_block_processed(rc, next); 23165d4f98a2SYan Zheng break; 23175d4f98a2SYan Zheng } 23185d4f98a2SYan Zheng 23193fd0a558SYan, Zheng WARN_ON(1); 23205d4f98a2SYan Zheng root = NULL; 23215d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 23225d4f98a2SYan Zheng if (!next || next->level <= node->level) 23235d4f98a2SYan Zheng break; 23245d4f98a2SYan Zheng } 23253fd0a558SYan, Zheng if (!root) 23263fd0a558SYan, Zheng return NULL; 23275d4f98a2SYan Zheng 23285d4f98a2SYan Zheng *nr = index; 23293fd0a558SYan, Zheng next = node; 23303fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 23313fd0a558SYan, Zheng while (1) { 23323fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 23333fd0a558SYan, Zheng if (--index < 0) 23343fd0a558SYan, Zheng break; 23353fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 23363fd0a558SYan, Zheng } 23375d4f98a2SYan Zheng return root; 23385d4f98a2SYan Zheng } 23395d4f98a2SYan Zheng 23403fd0a558SYan, Zheng /* 23413fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 23423fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 23433fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 23443fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 23453fd0a558SYan, Zheng */ 23465d4f98a2SYan Zheng static noinline_for_stack 23475d4f98a2SYan Zheng struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans, 23485d4f98a2SYan Zheng struct backref_node *node) 23495d4f98a2SYan Zheng { 23503fd0a558SYan, Zheng struct backref_node *next; 23513fd0a558SYan, Zheng struct btrfs_root *root; 23523fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 23535d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 23543fd0a558SYan, Zheng int index = 0; 23553fd0a558SYan, Zheng 23563fd0a558SYan, Zheng next = node; 23573fd0a558SYan, Zheng while (1) { 23583fd0a558SYan, Zheng cond_resched(); 23593fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 23603fd0a558SYan, Zheng root = next->root; 23613fd0a558SYan, Zheng BUG_ON(!root); 23623fd0a558SYan, Zheng 236325985edcSLucas De Marchi /* no other choice for non-references counted tree */ 23643fd0a558SYan, Zheng if (!root->ref_cows) 23653fd0a558SYan, Zheng return root; 23663fd0a558SYan, Zheng 23673fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 23683fd0a558SYan, Zheng fs_root = root; 23693fd0a558SYan, Zheng 23703fd0a558SYan, Zheng if (next != node) 23713fd0a558SYan, Zheng return NULL; 23723fd0a558SYan, Zheng 23733fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 23743fd0a558SYan, Zheng if (!next || next->level <= node->level) 23753fd0a558SYan, Zheng break; 23763fd0a558SYan, Zheng } 23773fd0a558SYan, Zheng 23783fd0a558SYan, Zheng if (!fs_root) 23793fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 23803fd0a558SYan, Zheng return fs_root; 23815d4f98a2SYan Zheng } 23825d4f98a2SYan Zheng 23835d4f98a2SYan Zheng static noinline_for_stack 23843fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 23853fd0a558SYan, Zheng struct backref_node *node, int reserve) 23865d4f98a2SYan Zheng { 23873fd0a558SYan, Zheng struct backref_node *next = node; 23883fd0a558SYan, Zheng struct backref_edge *edge; 23893fd0a558SYan, Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 23903fd0a558SYan, Zheng u64 num_bytes = 0; 23913fd0a558SYan, Zheng int index = 0; 23925d4f98a2SYan Zheng 23933fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 23943fd0a558SYan, Zheng 23953fd0a558SYan, Zheng while (next) { 23963fd0a558SYan, Zheng cond_resched(); 23975d4f98a2SYan Zheng while (1) { 23983fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 23995d4f98a2SYan Zheng break; 24005d4f98a2SYan Zheng 24013fd0a558SYan, Zheng num_bytes += btrfs_level_size(rc->extent_root, 24023fd0a558SYan, Zheng next->level); 24033fd0a558SYan, Zheng 24043fd0a558SYan, Zheng if (list_empty(&next->upper)) 24053fd0a558SYan, Zheng break; 24063fd0a558SYan, Zheng 24073fd0a558SYan, Zheng edge = list_entry(next->upper.next, 24083fd0a558SYan, Zheng struct backref_edge, list[LOWER]); 24093fd0a558SYan, Zheng edges[index++] = edge; 24103fd0a558SYan, Zheng next = edge->node[UPPER]; 24115d4f98a2SYan Zheng } 24123fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 24133fd0a558SYan, Zheng } 24143fd0a558SYan, Zheng return num_bytes; 24153fd0a558SYan, Zheng } 24163fd0a558SYan, Zheng 24173fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 24183fd0a558SYan, Zheng struct reloc_control *rc, 24193fd0a558SYan, Zheng struct backref_node *node) 24203fd0a558SYan, Zheng { 24213fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 24223fd0a558SYan, Zheng u64 num_bytes; 24233fd0a558SYan, Zheng int ret; 24243fd0a558SYan, Zheng 24253fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 24263fd0a558SYan, Zheng 24273fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 24284a92b1b8SJosef Bacik ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes); 24293fd0a558SYan, Zheng if (ret) { 24303fd0a558SYan, Zheng if (ret == -EAGAIN) 24313fd0a558SYan, Zheng rc->commit_transaction = 1; 24323fd0a558SYan, Zheng return ret; 24333fd0a558SYan, Zheng } 24343fd0a558SYan, Zheng 24353fd0a558SYan, Zheng return 0; 24363fd0a558SYan, Zheng } 24373fd0a558SYan, Zheng 24383fd0a558SYan, Zheng static void release_metadata_space(struct reloc_control *rc, 24393fd0a558SYan, Zheng struct backref_node *node) 24403fd0a558SYan, Zheng { 24413fd0a558SYan, Zheng u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2; 24423fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes); 24435d4f98a2SYan Zheng } 24445d4f98a2SYan Zheng 24455d4f98a2SYan Zheng /* 24465d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 24475d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 24485d4f98a2SYan Zheng * 24495d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 24505d4f98a2SYan Zheng * in that case this function just updates pointers. 24515d4f98a2SYan Zheng */ 24525d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 24533fd0a558SYan, Zheng struct reloc_control *rc, 24545d4f98a2SYan Zheng struct backref_node *node, 24555d4f98a2SYan Zheng struct btrfs_key *key, 24565d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 24575d4f98a2SYan Zheng { 24585d4f98a2SYan Zheng struct backref_node *upper; 24595d4f98a2SYan Zheng struct backref_edge *edge; 24605d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24615d4f98a2SYan Zheng struct btrfs_root *root; 24625d4f98a2SYan Zheng struct extent_buffer *eb; 24635d4f98a2SYan Zheng u32 blocksize; 24645d4f98a2SYan Zheng u64 bytenr; 24655d4f98a2SYan Zheng u64 generation; 24665d4f98a2SYan Zheng int nr; 24675d4f98a2SYan Zheng int slot; 24685d4f98a2SYan Zheng int ret; 24695d4f98a2SYan Zheng int err = 0; 24705d4f98a2SYan Zheng 24715d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 24725d4f98a2SYan Zheng 24735d4f98a2SYan Zheng path->lowest_level = node->level + 1; 24743fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 24755d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 24765d4f98a2SYan Zheng cond_resched(); 24775d4f98a2SYan Zheng 24785d4f98a2SYan Zheng upper = edge->node[UPPER]; 24793fd0a558SYan, Zheng root = select_reloc_root(trans, rc, upper, edges, &nr); 24803fd0a558SYan, Zheng BUG_ON(!root); 24815d4f98a2SYan Zheng 24823fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 24833fd0a558SYan, Zheng if (!lowest) { 24843fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 24853fd0a558SYan, Zheng upper->level, &slot); 24863fd0a558SYan, Zheng BUG_ON(ret); 24873fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 24883fd0a558SYan, Zheng if (node->eb->start == bytenr) 24893fd0a558SYan, Zheng goto next; 24903fd0a558SYan, Zheng } 24915d4f98a2SYan Zheng drop_node_buffer(upper); 24923fd0a558SYan, Zheng } 24935d4f98a2SYan Zheng 24945d4f98a2SYan Zheng if (!upper->eb) { 24955d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 24965d4f98a2SYan Zheng if (ret < 0) { 24975d4f98a2SYan Zheng err = ret; 24985d4f98a2SYan Zheng break; 24995d4f98a2SYan Zheng } 25005d4f98a2SYan Zheng BUG_ON(ret > 0); 25015d4f98a2SYan Zheng 25023fd0a558SYan, Zheng if (!upper->eb) { 25033fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 25043fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 25053fd0a558SYan, Zheng } else { 25063fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 25073fd0a558SYan, Zheng } 25083fd0a558SYan, Zheng 25093fd0a558SYan, Zheng upper->locked = 1; 25103fd0a558SYan, Zheng path->locks[upper->level] = 0; 25113fd0a558SYan, Zheng 25125d4f98a2SYan Zheng slot = path->slots[upper->level]; 2513b3b4aa74SDavid Sterba btrfs_release_path(path); 25145d4f98a2SYan Zheng } else { 25155d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 25165d4f98a2SYan Zheng &slot); 25175d4f98a2SYan Zheng BUG_ON(ret); 25185d4f98a2SYan Zheng } 25195d4f98a2SYan Zheng 25205d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 25213fd0a558SYan, Zheng if (lowest) { 25223fd0a558SYan, Zheng BUG_ON(bytenr != node->bytenr); 25235d4f98a2SYan Zheng } else { 25243fd0a558SYan, Zheng if (node->eb->start == bytenr) 25253fd0a558SYan, Zheng goto next; 25265d4f98a2SYan Zheng } 25275d4f98a2SYan Zheng 25285d4f98a2SYan Zheng blocksize = btrfs_level_size(root, node->level); 25295d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 25305d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, generation); 253197d9a8a4STsutomu Itoh if (!eb) { 253297d9a8a4STsutomu Itoh err = -EIO; 253397d9a8a4STsutomu Itoh goto next; 253497d9a8a4STsutomu Itoh } 25355d4f98a2SYan Zheng btrfs_tree_lock(eb); 25365d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 25375d4f98a2SYan Zheng 25385d4f98a2SYan Zheng if (!node->eb) { 25395d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 25405d4f98a2SYan Zheng slot, &eb); 25413fd0a558SYan, Zheng btrfs_tree_unlock(eb); 25423fd0a558SYan, Zheng free_extent_buffer(eb); 25435d4f98a2SYan Zheng if (ret < 0) { 25445d4f98a2SYan Zheng err = ret; 25453fd0a558SYan, Zheng goto next; 25465d4f98a2SYan Zheng } 25473fd0a558SYan, Zheng BUG_ON(node->eb != eb); 25485d4f98a2SYan Zheng } else { 25495d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 25505d4f98a2SYan Zheng node->eb->start); 25515d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 25525d4f98a2SYan Zheng trans->transid); 25535d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 25545d4f98a2SYan Zheng 25555d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, 25565d4f98a2SYan Zheng node->eb->start, blocksize, 25575d4f98a2SYan Zheng upper->eb->start, 25585d4f98a2SYan Zheng btrfs_header_owner(upper->eb), 25595d4f98a2SYan Zheng node->level, 0); 25605d4f98a2SYan Zheng BUG_ON(ret); 25615d4f98a2SYan Zheng 25625d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 25635d4f98a2SYan Zheng BUG_ON(ret); 25645d4f98a2SYan Zheng } 25653fd0a558SYan, Zheng next: 25663fd0a558SYan, Zheng if (!upper->pending) 25673fd0a558SYan, Zheng drop_node_buffer(upper); 25683fd0a558SYan, Zheng else 25693fd0a558SYan, Zheng unlock_node_buffer(upper); 25703fd0a558SYan, Zheng if (err) 25713fd0a558SYan, Zheng break; 25725d4f98a2SYan Zheng } 25733fd0a558SYan, Zheng 25743fd0a558SYan, Zheng if (!err && node->pending) { 25753fd0a558SYan, Zheng drop_node_buffer(node); 25763fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 25773fd0a558SYan, Zheng node->pending = 0; 25785d4f98a2SYan Zheng } 25793fd0a558SYan, Zheng 25805d4f98a2SYan Zheng path->lowest_level = 0; 25813fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 25825d4f98a2SYan Zheng return err; 25835d4f98a2SYan Zheng } 25845d4f98a2SYan Zheng 25855d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 25863fd0a558SYan, Zheng struct reloc_control *rc, 25875d4f98a2SYan Zheng struct backref_node *node, 25885d4f98a2SYan Zheng struct btrfs_path *path) 25895d4f98a2SYan Zheng { 25905d4f98a2SYan Zheng struct btrfs_key key; 25915d4f98a2SYan Zheng 25925d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 25933fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 25945d4f98a2SYan Zheng } 25955d4f98a2SYan Zheng 25965d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 25973fd0a558SYan, Zheng struct reloc_control *rc, 25983fd0a558SYan, Zheng struct btrfs_path *path, int err) 25995d4f98a2SYan Zheng { 26003fd0a558SYan, Zheng LIST_HEAD(list); 26013fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 26025d4f98a2SYan Zheng struct backref_node *node; 26035d4f98a2SYan Zheng int level; 26045d4f98a2SYan Zheng int ret; 26055d4f98a2SYan Zheng 26065d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 26075d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 26085d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 26093fd0a558SYan, Zheng struct backref_node, list); 26103fd0a558SYan, Zheng list_move_tail(&node->list, &list); 26113fd0a558SYan, Zheng BUG_ON(!node->pending); 26125d4f98a2SYan Zheng 26133fd0a558SYan, Zheng if (!err) { 26143fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 26155d4f98a2SYan Zheng if (ret < 0) 26165d4f98a2SYan Zheng err = ret; 26175d4f98a2SYan Zheng } 26185d4f98a2SYan Zheng } 26193fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 26203fd0a558SYan, Zheng } 26215d4f98a2SYan Zheng return err; 26225d4f98a2SYan Zheng } 26235d4f98a2SYan Zheng 26245d4f98a2SYan Zheng static void mark_block_processed(struct reloc_control *rc, 26253fd0a558SYan, Zheng u64 bytenr, u32 blocksize) 26263fd0a558SYan, Zheng { 26273fd0a558SYan, Zheng set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1, 26283fd0a558SYan, Zheng EXTENT_DIRTY, GFP_NOFS); 26293fd0a558SYan, Zheng } 26303fd0a558SYan, Zheng 26313fd0a558SYan, Zheng static void __mark_block_processed(struct reloc_control *rc, 26325d4f98a2SYan Zheng struct backref_node *node) 26335d4f98a2SYan Zheng { 26345d4f98a2SYan Zheng u32 blocksize; 26355d4f98a2SYan Zheng if (node->level == 0 || 26365d4f98a2SYan Zheng in_block_group(node->bytenr, rc->block_group)) { 26375d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, node->level); 26383fd0a558SYan, Zheng mark_block_processed(rc, node->bytenr, blocksize); 26395d4f98a2SYan Zheng } 26405d4f98a2SYan Zheng node->processed = 1; 26415d4f98a2SYan Zheng } 26425d4f98a2SYan Zheng 26435d4f98a2SYan Zheng /* 26445d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 26455d4f98a2SYan Zheng * as processed. 26465d4f98a2SYan Zheng */ 26475d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 26485d4f98a2SYan Zheng struct backref_node *node) 26495d4f98a2SYan Zheng { 26505d4f98a2SYan Zheng struct backref_node *next = node; 26515d4f98a2SYan Zheng struct backref_edge *edge; 26525d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 26535d4f98a2SYan Zheng int index = 0; 26545d4f98a2SYan Zheng 26555d4f98a2SYan Zheng while (next) { 26565d4f98a2SYan Zheng cond_resched(); 26575d4f98a2SYan Zheng while (1) { 26585d4f98a2SYan Zheng if (next->processed) 26595d4f98a2SYan Zheng break; 26605d4f98a2SYan Zheng 26613fd0a558SYan, Zheng __mark_block_processed(rc, next); 26625d4f98a2SYan Zheng 26635d4f98a2SYan Zheng if (list_empty(&next->upper)) 26645d4f98a2SYan Zheng break; 26655d4f98a2SYan Zheng 26665d4f98a2SYan Zheng edge = list_entry(next->upper.next, 26675d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 26685d4f98a2SYan Zheng edges[index++] = edge; 26695d4f98a2SYan Zheng next = edge->node[UPPER]; 26705d4f98a2SYan Zheng } 26715d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 26725d4f98a2SYan Zheng } 26735d4f98a2SYan Zheng } 26745d4f98a2SYan Zheng 26755d4f98a2SYan Zheng static int tree_block_processed(u64 bytenr, u32 blocksize, 26765d4f98a2SYan Zheng struct reloc_control *rc) 26775d4f98a2SYan Zheng { 26785d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 26799655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 26805d4f98a2SYan Zheng return 1; 26815d4f98a2SYan Zheng return 0; 26825d4f98a2SYan Zheng } 26835d4f98a2SYan Zheng 26845d4f98a2SYan Zheng static int get_tree_block_key(struct reloc_control *rc, 26855d4f98a2SYan Zheng struct tree_block *block) 26865d4f98a2SYan Zheng { 26875d4f98a2SYan Zheng struct extent_buffer *eb; 26885d4f98a2SYan Zheng 26895d4f98a2SYan Zheng BUG_ON(block->key_ready); 26905d4f98a2SYan Zheng eb = read_tree_block(rc->extent_root, block->bytenr, 26915d4f98a2SYan Zheng block->key.objectid, block->key.offset); 269297d9a8a4STsutomu Itoh BUG_ON(!eb); 26935d4f98a2SYan Zheng WARN_ON(btrfs_header_level(eb) != block->level); 26945d4f98a2SYan Zheng if (block->level == 0) 26955d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 26965d4f98a2SYan Zheng else 26975d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 26985d4f98a2SYan Zheng free_extent_buffer(eb); 26995d4f98a2SYan Zheng block->key_ready = 1; 27005d4f98a2SYan Zheng return 0; 27015d4f98a2SYan Zheng } 27025d4f98a2SYan Zheng 27035d4f98a2SYan Zheng static int reada_tree_block(struct reloc_control *rc, 27045d4f98a2SYan Zheng struct tree_block *block) 27055d4f98a2SYan Zheng { 27065d4f98a2SYan Zheng BUG_ON(block->key_ready); 27075d4f98a2SYan Zheng readahead_tree_block(rc->extent_root, block->bytenr, 27085d4f98a2SYan Zheng block->key.objectid, block->key.offset); 27095d4f98a2SYan Zheng return 0; 27105d4f98a2SYan Zheng } 27115d4f98a2SYan Zheng 27125d4f98a2SYan Zheng /* 27135d4f98a2SYan Zheng * helper function to relocate a tree block 27145d4f98a2SYan Zheng */ 27155d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 27165d4f98a2SYan Zheng struct reloc_control *rc, 27175d4f98a2SYan Zheng struct backref_node *node, 27185d4f98a2SYan Zheng struct btrfs_key *key, 27195d4f98a2SYan Zheng struct btrfs_path *path) 27205d4f98a2SYan Zheng { 27215d4f98a2SYan Zheng struct btrfs_root *root; 27223fd0a558SYan, Zheng int release = 0; 27233fd0a558SYan, Zheng int ret = 0; 27245d4f98a2SYan Zheng 27253fd0a558SYan, Zheng if (!node) 27265d4f98a2SYan Zheng return 0; 27273fd0a558SYan, Zheng 27283fd0a558SYan, Zheng BUG_ON(node->processed); 27293fd0a558SYan, Zheng root = select_one_root(trans, node); 27303fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 27313fd0a558SYan, Zheng update_processed_blocks(rc, node); 27323fd0a558SYan, Zheng goto out; 27335d4f98a2SYan Zheng } 27345d4f98a2SYan Zheng 27353fd0a558SYan, Zheng if (!root || root->ref_cows) { 27363fd0a558SYan, Zheng ret = reserve_metadata_space(trans, rc, node); 27373fd0a558SYan, Zheng if (ret) 27385d4f98a2SYan Zheng goto out; 27393fd0a558SYan, Zheng release = 1; 27405d4f98a2SYan Zheng } 27413fd0a558SYan, Zheng 27423fd0a558SYan, Zheng if (root) { 27433fd0a558SYan, Zheng if (root->ref_cows) { 27443fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 27453fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 27463fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 27473fd0a558SYan, Zheng root = root->reloc_root; 27483fd0a558SYan, Zheng node->new_bytenr = root->node->start; 27493fd0a558SYan, Zheng node->root = root; 27503fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 27513fd0a558SYan, Zheng } else { 27525d4f98a2SYan Zheng path->lowest_level = node->level; 27535d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2754b3b4aa74SDavid Sterba btrfs_release_path(path); 27553fd0a558SYan, Zheng if (ret > 0) 27565d4f98a2SYan Zheng ret = 0; 27573fd0a558SYan, Zheng } 27583fd0a558SYan, Zheng if (!ret) 27593fd0a558SYan, Zheng update_processed_blocks(rc, node); 27603fd0a558SYan, Zheng } else { 27613fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 27623fd0a558SYan, Zheng } 27635d4f98a2SYan Zheng out: 27643fd0a558SYan, Zheng if (ret || node->level == 0 || node->cowonly) { 27653fd0a558SYan, Zheng if (release) 27663fd0a558SYan, Zheng release_metadata_space(rc, node); 27673fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 27683fd0a558SYan, Zheng } 27695d4f98a2SYan Zheng return ret; 27705d4f98a2SYan Zheng } 27715d4f98a2SYan Zheng 27725d4f98a2SYan Zheng /* 27735d4f98a2SYan Zheng * relocate a list of blocks 27745d4f98a2SYan Zheng */ 27755d4f98a2SYan Zheng static noinline_for_stack 27765d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 27775d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 27785d4f98a2SYan Zheng { 27795d4f98a2SYan Zheng struct backref_node *node; 27805d4f98a2SYan Zheng struct btrfs_path *path; 27815d4f98a2SYan Zheng struct tree_block *block; 27825d4f98a2SYan Zheng struct rb_node *rb_node; 27835d4f98a2SYan Zheng int ret; 27845d4f98a2SYan Zheng int err = 0; 27855d4f98a2SYan Zheng 27865d4f98a2SYan Zheng path = btrfs_alloc_path(); 27875d4f98a2SYan Zheng if (!path) 27885d4f98a2SYan Zheng return -ENOMEM; 27895d4f98a2SYan Zheng 27905d4f98a2SYan Zheng rb_node = rb_first(blocks); 27915d4f98a2SYan Zheng while (rb_node) { 27925d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 27935d4f98a2SYan Zheng if (!block->key_ready) 27945d4f98a2SYan Zheng reada_tree_block(rc, block); 27955d4f98a2SYan Zheng rb_node = rb_next(rb_node); 27965d4f98a2SYan Zheng } 27975d4f98a2SYan Zheng 27985d4f98a2SYan Zheng rb_node = rb_first(blocks); 27995d4f98a2SYan Zheng while (rb_node) { 28005d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 28015d4f98a2SYan Zheng if (!block->key_ready) 28025d4f98a2SYan Zheng get_tree_block_key(rc, block); 28035d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28045d4f98a2SYan Zheng } 28055d4f98a2SYan Zheng 28065d4f98a2SYan Zheng rb_node = rb_first(blocks); 28075d4f98a2SYan Zheng while (rb_node) { 28085d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 28095d4f98a2SYan Zheng 28103fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 28115d4f98a2SYan Zheng block->level, block->bytenr); 28125d4f98a2SYan Zheng if (IS_ERR(node)) { 28135d4f98a2SYan Zheng err = PTR_ERR(node); 28145d4f98a2SYan Zheng goto out; 28155d4f98a2SYan Zheng } 28165d4f98a2SYan Zheng 28175d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 28185d4f98a2SYan Zheng path); 28195d4f98a2SYan Zheng if (ret < 0) { 28203fd0a558SYan, Zheng if (ret != -EAGAIN || rb_node == rb_first(blocks)) 28215d4f98a2SYan Zheng err = ret; 28225d4f98a2SYan Zheng goto out; 28235d4f98a2SYan Zheng } 28245d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28255d4f98a2SYan Zheng } 28265d4f98a2SYan Zheng out: 28275d4f98a2SYan Zheng free_block_list(blocks); 28283fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 28295d4f98a2SYan Zheng 28305d4f98a2SYan Zheng btrfs_free_path(path); 28315d4f98a2SYan Zheng return err; 28325d4f98a2SYan Zheng } 28335d4f98a2SYan Zheng 28345d4f98a2SYan Zheng static noinline_for_stack 2835efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 2836efa56464SYan, Zheng struct file_extent_cluster *cluster) 2837efa56464SYan, Zheng { 2838efa56464SYan, Zheng u64 alloc_hint = 0; 2839efa56464SYan, Zheng u64 start; 2840efa56464SYan, Zheng u64 end; 2841efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 2842efa56464SYan, Zheng u64 num_bytes; 2843efa56464SYan, Zheng int nr = 0; 2844efa56464SYan, Zheng int ret = 0; 2845efa56464SYan, Zheng 2846efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2847efa56464SYan, Zheng mutex_lock(&inode->i_mutex); 2848efa56464SYan, Zheng 2849efa56464SYan, Zheng ret = btrfs_check_data_free_space(inode, cluster->end + 2850efa56464SYan, Zheng 1 - cluster->start); 2851efa56464SYan, Zheng if (ret) 2852efa56464SYan, Zheng goto out; 2853efa56464SYan, Zheng 2854efa56464SYan, Zheng while (nr < cluster->nr) { 2855efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2856efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2857efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2858efa56464SYan, Zheng else 2859efa56464SYan, Zheng end = cluster->end - offset; 2860efa56464SYan, Zheng 2861efa56464SYan, Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 2862efa56464SYan, Zheng num_bytes = end + 1 - start; 2863efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 2864efa56464SYan, Zheng num_bytes, num_bytes, 2865efa56464SYan, Zheng end + 1, &alloc_hint); 2866efa56464SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 2867efa56464SYan, Zheng if (ret) 2868efa56464SYan, Zheng break; 2869efa56464SYan, Zheng nr++; 2870efa56464SYan, Zheng } 2871efa56464SYan, Zheng btrfs_free_reserved_data_space(inode, cluster->end + 2872efa56464SYan, Zheng 1 - cluster->start); 2873efa56464SYan, Zheng out: 2874efa56464SYan, Zheng mutex_unlock(&inode->i_mutex); 2875efa56464SYan, Zheng return ret; 2876efa56464SYan, Zheng } 2877efa56464SYan, Zheng 2878efa56464SYan, Zheng static noinline_for_stack 28790257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 28800257bb82SYan, Zheng u64 block_start) 28815d4f98a2SYan Zheng { 28825d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 28835d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 28845d4f98a2SYan Zheng struct extent_map *em; 28850257bb82SYan, Zheng int ret = 0; 28865d4f98a2SYan Zheng 2887172ddd60SDavid Sterba em = alloc_extent_map(); 28880257bb82SYan, Zheng if (!em) 28890257bb82SYan, Zheng return -ENOMEM; 28900257bb82SYan, Zheng 28915d4f98a2SYan Zheng em->start = start; 28920257bb82SYan, Zheng em->len = end + 1 - start; 28930257bb82SYan, Zheng em->block_len = em->len; 28940257bb82SYan, Zheng em->block_start = block_start; 28955d4f98a2SYan Zheng em->bdev = root->fs_info->fs_devices->latest_bdev; 28965d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 28975d4f98a2SYan Zheng 28985d4f98a2SYan Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 28995d4f98a2SYan Zheng while (1) { 2900890871beSChris Mason write_lock(&em_tree->lock); 29015d4f98a2SYan Zheng ret = add_extent_mapping(em_tree, em); 2902890871beSChris Mason write_unlock(&em_tree->lock); 29035d4f98a2SYan Zheng if (ret != -EEXIST) { 29045d4f98a2SYan Zheng free_extent_map(em); 29055d4f98a2SYan Zheng break; 29065d4f98a2SYan Zheng } 29075d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 0); 29085d4f98a2SYan Zheng } 29095d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 29100257bb82SYan, Zheng return ret; 29110257bb82SYan, Zheng } 29125d4f98a2SYan Zheng 29130257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 29140257bb82SYan, Zheng struct file_extent_cluster *cluster) 29150257bb82SYan, Zheng { 29160257bb82SYan, Zheng u64 page_start; 29170257bb82SYan, Zheng u64 page_end; 29180257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 29190257bb82SYan, Zheng unsigned long index; 29200257bb82SYan, Zheng unsigned long last_index; 29210257bb82SYan, Zheng struct page *page; 29220257bb82SYan, Zheng struct file_ra_state *ra; 29233b16a4e3SJosef Bacik gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping); 29240257bb82SYan, Zheng int nr = 0; 29250257bb82SYan, Zheng int ret = 0; 29260257bb82SYan, Zheng 29270257bb82SYan, Zheng if (!cluster->nr) 29280257bb82SYan, Zheng return 0; 29290257bb82SYan, Zheng 29300257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 29310257bb82SYan, Zheng if (!ra) 29320257bb82SYan, Zheng return -ENOMEM; 29330257bb82SYan, Zheng 2934efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 29350257bb82SYan, Zheng if (ret) 2936efa56464SYan, Zheng goto out; 29370257bb82SYan, Zheng 29380257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 29390257bb82SYan, Zheng 2940efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2941efa56464SYan, Zheng cluster->end - offset, cluster->start); 2942efa56464SYan, Zheng if (ret) 2943efa56464SYan, Zheng goto out; 2944efa56464SYan, Zheng 2945efa56464SYan, Zheng index = (cluster->start - offset) >> PAGE_CACHE_SHIFT; 2946efa56464SYan, Zheng last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT; 29470257bb82SYan, Zheng while (index <= last_index) { 2948efa56464SYan, Zheng ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE); 2949efa56464SYan, Zheng if (ret) 2950efa56464SYan, Zheng goto out; 2951efa56464SYan, Zheng 29520257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 29530257bb82SYan, Zheng if (!page) { 29540257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 29550257bb82SYan, Zheng ra, NULL, index, 29560257bb82SYan, Zheng last_index + 1 - index); 2957a94733d0SJosef Bacik page = find_or_create_page(inode->i_mapping, index, 29583b16a4e3SJosef Bacik mask); 29590257bb82SYan, Zheng if (!page) { 2960efa56464SYan, Zheng btrfs_delalloc_release_metadata(inode, 2961efa56464SYan, Zheng PAGE_CACHE_SIZE); 29620257bb82SYan, Zheng ret = -ENOMEM; 2963efa56464SYan, Zheng goto out; 29640257bb82SYan, Zheng } 29650257bb82SYan, Zheng } 29660257bb82SYan, Zheng 29670257bb82SYan, Zheng if (PageReadahead(page)) { 29680257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 29690257bb82SYan, Zheng ra, NULL, page, index, 29700257bb82SYan, Zheng last_index + 1 - index); 29710257bb82SYan, Zheng } 29720257bb82SYan, Zheng 29730257bb82SYan, Zheng if (!PageUptodate(page)) { 29740257bb82SYan, Zheng btrfs_readpage(NULL, page); 29750257bb82SYan, Zheng lock_page(page); 29760257bb82SYan, Zheng if (!PageUptodate(page)) { 29770257bb82SYan, Zheng unlock_page(page); 29780257bb82SYan, Zheng page_cache_release(page); 2979efa56464SYan, Zheng btrfs_delalloc_release_metadata(inode, 2980efa56464SYan, Zheng PAGE_CACHE_SIZE); 29810257bb82SYan, Zheng ret = -EIO; 2982efa56464SYan, Zheng goto out; 29830257bb82SYan, Zheng } 29840257bb82SYan, Zheng } 29850257bb82SYan, Zheng 29860257bb82SYan, Zheng page_start = (u64)page->index << PAGE_CACHE_SHIFT; 29870257bb82SYan, Zheng page_end = page_start + PAGE_CACHE_SIZE - 1; 29880257bb82SYan, Zheng 29890257bb82SYan, Zheng lock_extent(&BTRFS_I(inode)->io_tree, 29900257bb82SYan, Zheng page_start, page_end, GFP_NOFS); 29910257bb82SYan, Zheng 29920257bb82SYan, Zheng set_page_extent_mapped(page); 29930257bb82SYan, Zheng 29940257bb82SYan, Zheng if (nr < cluster->nr && 29950257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 29960257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 29970257bb82SYan, Zheng page_start, page_end, 29980257bb82SYan, Zheng EXTENT_BOUNDARY, GFP_NOFS); 29990257bb82SYan, Zheng nr++; 30000257bb82SYan, Zheng } 30010257bb82SYan, Zheng 3002efa56464SYan, Zheng btrfs_set_extent_delalloc(inode, page_start, page_end, NULL); 30030257bb82SYan, Zheng set_page_dirty(page); 30040257bb82SYan, Zheng 30050257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 30060257bb82SYan, Zheng page_start, page_end, GFP_NOFS); 30070257bb82SYan, Zheng unlock_page(page); 30080257bb82SYan, Zheng page_cache_release(page); 30090257bb82SYan, Zheng 30100257bb82SYan, Zheng index++; 3011efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 3012efa56464SYan, Zheng btrfs_throttle(BTRFS_I(inode)->root); 30130257bb82SYan, Zheng } 30140257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 3015efa56464SYan, Zheng out: 30160257bb82SYan, Zheng kfree(ra); 30170257bb82SYan, Zheng return ret; 30180257bb82SYan, Zheng } 30190257bb82SYan, Zheng 30200257bb82SYan, Zheng static noinline_for_stack 30210257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 30220257bb82SYan, Zheng struct file_extent_cluster *cluster) 30230257bb82SYan, Zheng { 30240257bb82SYan, Zheng int ret; 30250257bb82SYan, Zheng 30260257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 30270257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30280257bb82SYan, Zheng if (ret) 30290257bb82SYan, Zheng return ret; 30300257bb82SYan, Zheng cluster->nr = 0; 30310257bb82SYan, Zheng } 30320257bb82SYan, Zheng 30330257bb82SYan, Zheng if (!cluster->nr) 30340257bb82SYan, Zheng cluster->start = extent_key->objectid; 30350257bb82SYan, Zheng else 30360257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 30370257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 30380257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 30390257bb82SYan, Zheng cluster->nr++; 30400257bb82SYan, Zheng 30410257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 30420257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30430257bb82SYan, Zheng if (ret) 30440257bb82SYan, Zheng return ret; 30450257bb82SYan, Zheng cluster->nr = 0; 30460257bb82SYan, Zheng } 30470257bb82SYan, Zheng return 0; 30485d4f98a2SYan Zheng } 30495d4f98a2SYan Zheng 30505d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 30515d4f98a2SYan Zheng static int get_ref_objectid_v0(struct reloc_control *rc, 30525d4f98a2SYan Zheng struct btrfs_path *path, 30535d4f98a2SYan Zheng struct btrfs_key *extent_key, 30545d4f98a2SYan Zheng u64 *ref_objectid, int *path_change) 30555d4f98a2SYan Zheng { 30565d4f98a2SYan Zheng struct btrfs_key key; 30575d4f98a2SYan Zheng struct extent_buffer *leaf; 30585d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 30595d4f98a2SYan Zheng int ret; 30605d4f98a2SYan Zheng int slot; 30615d4f98a2SYan Zheng 30625d4f98a2SYan Zheng leaf = path->nodes[0]; 30635d4f98a2SYan Zheng slot = path->slots[0]; 30645d4f98a2SYan Zheng while (1) { 30655d4f98a2SYan Zheng if (slot >= btrfs_header_nritems(leaf)) { 30665d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 30675d4f98a2SYan Zheng if (ret < 0) 30685d4f98a2SYan Zheng return ret; 30695d4f98a2SYan Zheng BUG_ON(ret > 0); 30705d4f98a2SYan Zheng leaf = path->nodes[0]; 30715d4f98a2SYan Zheng slot = path->slots[0]; 30725d4f98a2SYan Zheng if (path_change) 30735d4f98a2SYan Zheng *path_change = 1; 30745d4f98a2SYan Zheng } 30755d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, slot); 30765d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 30775d4f98a2SYan Zheng return -ENOENT; 30785d4f98a2SYan Zheng 30795d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_REF_V0_KEY) { 30805d4f98a2SYan Zheng slot++; 30815d4f98a2SYan Zheng continue; 30825d4f98a2SYan Zheng } 30835d4f98a2SYan Zheng ref0 = btrfs_item_ptr(leaf, slot, 30845d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 30855d4f98a2SYan Zheng *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0); 30865d4f98a2SYan Zheng break; 30875d4f98a2SYan Zheng } 30885d4f98a2SYan Zheng return 0; 30895d4f98a2SYan Zheng } 30905d4f98a2SYan Zheng #endif 30915d4f98a2SYan Zheng 30925d4f98a2SYan Zheng /* 30935d4f98a2SYan Zheng * helper to add a tree block to the list. 30945d4f98a2SYan Zheng * the major work is getting the generation and level of the block 30955d4f98a2SYan Zheng */ 30965d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 30975d4f98a2SYan Zheng struct btrfs_key *extent_key, 30985d4f98a2SYan Zheng struct btrfs_path *path, 30995d4f98a2SYan Zheng struct rb_root *blocks) 31005d4f98a2SYan Zheng { 31015d4f98a2SYan Zheng struct extent_buffer *eb; 31025d4f98a2SYan Zheng struct btrfs_extent_item *ei; 31035d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 31045d4f98a2SYan Zheng struct tree_block *block; 31055d4f98a2SYan Zheng struct rb_node *rb_node; 31065d4f98a2SYan Zheng u32 item_size; 31075d4f98a2SYan Zheng int level = -1; 31085d4f98a2SYan Zheng int generation; 31095d4f98a2SYan Zheng 31105d4f98a2SYan Zheng eb = path->nodes[0]; 31115d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 31125d4f98a2SYan Zheng 31135d4f98a2SYan Zheng if (item_size >= sizeof(*ei) + sizeof(*bi)) { 31145d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 31155d4f98a2SYan Zheng struct btrfs_extent_item); 31165d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 31175d4f98a2SYan Zheng generation = btrfs_extent_generation(eb, ei); 31185d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 31195d4f98a2SYan Zheng } else { 31205d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 31215d4f98a2SYan Zheng u64 ref_owner; 31225d4f98a2SYan Zheng int ret; 31235d4f98a2SYan Zheng 31245d4f98a2SYan Zheng BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 31255d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, extent_key, 31265d4f98a2SYan Zheng &ref_owner, NULL); 3127411fc6bcSAndi Kleen if (ret < 0) 3128411fc6bcSAndi Kleen return ret; 31295d4f98a2SYan Zheng BUG_ON(ref_owner >= BTRFS_MAX_LEVEL); 31305d4f98a2SYan Zheng level = (int)ref_owner; 31315d4f98a2SYan Zheng /* FIXME: get real generation */ 31325d4f98a2SYan Zheng generation = 0; 31335d4f98a2SYan Zheng #else 31345d4f98a2SYan Zheng BUG(); 31355d4f98a2SYan Zheng #endif 31365d4f98a2SYan Zheng } 31375d4f98a2SYan Zheng 3138b3b4aa74SDavid Sterba btrfs_release_path(path); 31395d4f98a2SYan Zheng 31405d4f98a2SYan Zheng BUG_ON(level == -1); 31415d4f98a2SYan Zheng 31425d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 31435d4f98a2SYan Zheng if (!block) 31445d4f98a2SYan Zheng return -ENOMEM; 31455d4f98a2SYan Zheng 31465d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 31475d4f98a2SYan Zheng block->key.objectid = extent_key->offset; 31485d4f98a2SYan Zheng block->key.offset = generation; 31495d4f98a2SYan Zheng block->level = level; 31505d4f98a2SYan Zheng block->key_ready = 0; 31515d4f98a2SYan Zheng 31525d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 31535d4f98a2SYan Zheng BUG_ON(rb_node); 31545d4f98a2SYan Zheng 31555d4f98a2SYan Zheng return 0; 31565d4f98a2SYan Zheng } 31575d4f98a2SYan Zheng 31585d4f98a2SYan Zheng /* 31595d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 31605d4f98a2SYan Zheng */ 31615d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 31625d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 31635d4f98a2SYan Zheng struct rb_root *blocks) 31645d4f98a2SYan Zheng { 31655d4f98a2SYan Zheng struct btrfs_path *path; 31665d4f98a2SYan Zheng struct btrfs_key key; 31675d4f98a2SYan Zheng int ret; 31685d4f98a2SYan Zheng 31695d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 31705d4f98a2SYan Zheng return 0; 31715d4f98a2SYan Zheng 31725d4f98a2SYan Zheng if (tree_search(blocks, bytenr)) 31735d4f98a2SYan Zheng return 0; 31745d4f98a2SYan Zheng 31755d4f98a2SYan Zheng path = btrfs_alloc_path(); 31765d4f98a2SYan Zheng if (!path) 31775d4f98a2SYan Zheng return -ENOMEM; 31785d4f98a2SYan Zheng 31795d4f98a2SYan Zheng key.objectid = bytenr; 31805d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31815d4f98a2SYan Zheng key.offset = blocksize; 31825d4f98a2SYan Zheng 31835d4f98a2SYan Zheng path->search_commit_root = 1; 31845d4f98a2SYan Zheng path->skip_locking = 1; 31855d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 31865d4f98a2SYan Zheng if (ret < 0) 31875d4f98a2SYan Zheng goto out; 31885d4f98a2SYan Zheng BUG_ON(ret); 31895d4f98a2SYan Zheng 31905d4f98a2SYan Zheng btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 31915d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 31925d4f98a2SYan Zheng out: 31935d4f98a2SYan Zheng btrfs_free_path(path); 31945d4f98a2SYan Zheng return ret; 31955d4f98a2SYan Zheng } 31965d4f98a2SYan Zheng 31975d4f98a2SYan Zheng /* 31985d4f98a2SYan Zheng * helper to check if the block use full backrefs for pointers in it 31995d4f98a2SYan Zheng */ 32005d4f98a2SYan Zheng static int block_use_full_backref(struct reloc_control *rc, 32015d4f98a2SYan Zheng struct extent_buffer *eb) 32025d4f98a2SYan Zheng { 32035d4f98a2SYan Zheng u64 flags; 32045d4f98a2SYan Zheng int ret; 32055d4f98a2SYan Zheng 32065d4f98a2SYan Zheng if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) || 32075d4f98a2SYan Zheng btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV) 32085d4f98a2SYan Zheng return 1; 32095d4f98a2SYan Zheng 32103fd0a558SYan, Zheng ret = btrfs_lookup_extent_info(NULL, rc->extent_root, 32113fd0a558SYan, Zheng eb->start, eb->len, NULL, &flags); 32125d4f98a2SYan Zheng BUG_ON(ret); 32135d4f98a2SYan Zheng 32145d4f98a2SYan Zheng if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) 32155d4f98a2SYan Zheng ret = 1; 32165d4f98a2SYan Zheng else 32175d4f98a2SYan Zheng ret = 0; 32185d4f98a2SYan Zheng return ret; 32195d4f98a2SYan Zheng } 32205d4f98a2SYan Zheng 32210af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 32220af3d00bSJosef Bacik struct inode *inode, u64 ino) 32230af3d00bSJosef Bacik { 32240af3d00bSJosef Bacik struct btrfs_key key; 32250af3d00bSJosef Bacik struct btrfs_path *path; 32260af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 32270af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 32280af3d00bSJosef Bacik unsigned long nr; 32290af3d00bSJosef Bacik int ret = 0; 32300af3d00bSJosef Bacik 32310af3d00bSJosef Bacik if (inode) 32320af3d00bSJosef Bacik goto truncate; 32330af3d00bSJosef Bacik 32340af3d00bSJosef Bacik key.objectid = ino; 32350af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 32360af3d00bSJosef Bacik key.offset = 0; 32370af3d00bSJosef Bacik 32380af3d00bSJosef Bacik inode = btrfs_iget(fs_info->sb, &key, root, NULL); 3239c704005dSDavid Sterba if (IS_ERR_OR_NULL(inode) || is_bad_inode(inode)) { 32400af3d00bSJosef Bacik if (inode && !IS_ERR(inode)) 32410af3d00bSJosef Bacik iput(inode); 32420af3d00bSJosef Bacik return -ENOENT; 32430af3d00bSJosef Bacik } 32440af3d00bSJosef Bacik 32450af3d00bSJosef Bacik truncate: 32460af3d00bSJosef Bacik path = btrfs_alloc_path(); 32470af3d00bSJosef Bacik if (!path) { 32480af3d00bSJosef Bacik ret = -ENOMEM; 32490af3d00bSJosef Bacik goto out; 32500af3d00bSJosef Bacik } 32510af3d00bSJosef Bacik 32527a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 32530af3d00bSJosef Bacik if (IS_ERR(trans)) { 32540af3d00bSJosef Bacik btrfs_free_path(path); 32553612b495STsutomu Itoh ret = PTR_ERR(trans); 32560af3d00bSJosef Bacik goto out; 32570af3d00bSJosef Bacik } 32580af3d00bSJosef Bacik 32590af3d00bSJosef Bacik ret = btrfs_truncate_free_space_cache(root, trans, path, inode); 32600af3d00bSJosef Bacik 32610af3d00bSJosef Bacik btrfs_free_path(path); 32620af3d00bSJosef Bacik nr = trans->blocks_used; 32630af3d00bSJosef Bacik btrfs_end_transaction(trans, root); 32640af3d00bSJosef Bacik btrfs_btree_balance_dirty(root, nr); 32650af3d00bSJosef Bacik out: 32660af3d00bSJosef Bacik iput(inode); 32670af3d00bSJosef Bacik return ret; 32680af3d00bSJosef Bacik } 32690af3d00bSJosef Bacik 32705d4f98a2SYan Zheng /* 32715d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY 32725d4f98a2SYan Zheng * this function scans fs tree to find blocks reference the data extent 32735d4f98a2SYan Zheng */ 32745d4f98a2SYan Zheng static int find_data_references(struct reloc_control *rc, 32755d4f98a2SYan Zheng struct btrfs_key *extent_key, 32765d4f98a2SYan Zheng struct extent_buffer *leaf, 32775d4f98a2SYan Zheng struct btrfs_extent_data_ref *ref, 32785d4f98a2SYan Zheng struct rb_root *blocks) 32795d4f98a2SYan Zheng { 32805d4f98a2SYan Zheng struct btrfs_path *path; 32815d4f98a2SYan Zheng struct tree_block *block; 32825d4f98a2SYan Zheng struct btrfs_root *root; 32835d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 32845d4f98a2SYan Zheng struct rb_node *rb_node; 32855d4f98a2SYan Zheng struct btrfs_key key; 32865d4f98a2SYan Zheng u64 ref_root; 32875d4f98a2SYan Zheng u64 ref_objectid; 32885d4f98a2SYan Zheng u64 ref_offset; 32895d4f98a2SYan Zheng u32 ref_count; 32905d4f98a2SYan Zheng u32 nritems; 32915d4f98a2SYan Zheng int err = 0; 32925d4f98a2SYan Zheng int added = 0; 32935d4f98a2SYan Zheng int counted; 32945d4f98a2SYan Zheng int ret; 32955d4f98a2SYan Zheng 32965d4f98a2SYan Zheng ref_root = btrfs_extent_data_ref_root(leaf, ref); 32975d4f98a2SYan Zheng ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref); 32985d4f98a2SYan Zheng ref_offset = btrfs_extent_data_ref_offset(leaf, ref); 32995d4f98a2SYan Zheng ref_count = btrfs_extent_data_ref_count(leaf, ref); 33005d4f98a2SYan Zheng 33010af3d00bSJosef Bacik /* 33020af3d00bSJosef Bacik * This is an extent belonging to the free space cache, lets just delete 33030af3d00bSJosef Bacik * it and redo the search. 33040af3d00bSJosef Bacik */ 33050af3d00bSJosef Bacik if (ref_root == BTRFS_ROOT_TREE_OBJECTID) { 33060af3d00bSJosef Bacik ret = delete_block_group_cache(rc->extent_root->fs_info, 33070af3d00bSJosef Bacik NULL, ref_objectid); 33080af3d00bSJosef Bacik if (ret != -ENOENT) 33090af3d00bSJosef Bacik return ret; 33100af3d00bSJosef Bacik ret = 0; 33110af3d00bSJosef Bacik } 33120af3d00bSJosef Bacik 33130af3d00bSJosef Bacik path = btrfs_alloc_path(); 33140af3d00bSJosef Bacik if (!path) 33150af3d00bSJosef Bacik return -ENOMEM; 3316026fd317SJosef Bacik path->reada = 1; 33170af3d00bSJosef Bacik 33185d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, ref_root); 33195d4f98a2SYan Zheng if (IS_ERR(root)) { 33205d4f98a2SYan Zheng err = PTR_ERR(root); 33215d4f98a2SYan Zheng goto out; 33225d4f98a2SYan Zheng } 33235d4f98a2SYan Zheng 33245d4f98a2SYan Zheng key.objectid = ref_objectid; 33255d4f98a2SYan Zheng key.type = BTRFS_EXTENT_DATA_KEY; 332684850e8dSYan, Zheng if (ref_offset > ((u64)-1 << 32)) 332784850e8dSYan, Zheng key.offset = 0; 332884850e8dSYan, Zheng else 332984850e8dSYan, Zheng key.offset = ref_offset; 33305d4f98a2SYan Zheng 33315d4f98a2SYan Zheng path->search_commit_root = 1; 33325d4f98a2SYan Zheng path->skip_locking = 1; 33335d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 33345d4f98a2SYan Zheng if (ret < 0) { 33355d4f98a2SYan Zheng err = ret; 33365d4f98a2SYan Zheng goto out; 33375d4f98a2SYan Zheng } 33385d4f98a2SYan Zheng 33395d4f98a2SYan Zheng leaf = path->nodes[0]; 33405d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 33415d4f98a2SYan Zheng /* 33425d4f98a2SYan Zheng * the references in tree blocks that use full backrefs 33435d4f98a2SYan Zheng * are not counted in 33445d4f98a2SYan Zheng */ 33455d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 33465d4f98a2SYan Zheng counted = 0; 33475d4f98a2SYan Zheng else 33485d4f98a2SYan Zheng counted = 1; 33495d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 33505d4f98a2SYan Zheng if (rb_node) { 33515d4f98a2SYan Zheng if (counted) 33525d4f98a2SYan Zheng added = 1; 33535d4f98a2SYan Zheng else 33545d4f98a2SYan Zheng path->slots[0] = nritems; 33555d4f98a2SYan Zheng } 33565d4f98a2SYan Zheng 33575d4f98a2SYan Zheng while (ref_count > 0) { 33585d4f98a2SYan Zheng while (path->slots[0] >= nritems) { 33595d4f98a2SYan Zheng ret = btrfs_next_leaf(root, path); 33605d4f98a2SYan Zheng if (ret < 0) { 33615d4f98a2SYan Zheng err = ret; 33625d4f98a2SYan Zheng goto out; 33635d4f98a2SYan Zheng } 33645d4f98a2SYan Zheng if (ret > 0) { 33655d4f98a2SYan Zheng WARN_ON(1); 33665d4f98a2SYan Zheng goto out; 33675d4f98a2SYan Zheng } 33685d4f98a2SYan Zheng 33695d4f98a2SYan Zheng leaf = path->nodes[0]; 33705d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 33715d4f98a2SYan Zheng added = 0; 33725d4f98a2SYan Zheng 33735d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 33745d4f98a2SYan Zheng counted = 0; 33755d4f98a2SYan Zheng else 33765d4f98a2SYan Zheng counted = 1; 33775d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 33785d4f98a2SYan Zheng if (rb_node) { 33795d4f98a2SYan Zheng if (counted) 33805d4f98a2SYan Zheng added = 1; 33815d4f98a2SYan Zheng else 33825d4f98a2SYan Zheng path->slots[0] = nritems; 33835d4f98a2SYan Zheng } 33845d4f98a2SYan Zheng } 33855d4f98a2SYan Zheng 33865d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 33875d4f98a2SYan Zheng if (key.objectid != ref_objectid || 33885d4f98a2SYan Zheng key.type != BTRFS_EXTENT_DATA_KEY) { 33895d4f98a2SYan Zheng WARN_ON(1); 33905d4f98a2SYan Zheng break; 33915d4f98a2SYan Zheng } 33925d4f98a2SYan Zheng 33935d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 33945d4f98a2SYan Zheng struct btrfs_file_extent_item); 33955d4f98a2SYan Zheng 33965d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 33975d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 33985d4f98a2SYan Zheng goto next; 33995d4f98a2SYan Zheng 34005d4f98a2SYan Zheng if (btrfs_file_extent_disk_bytenr(leaf, fi) != 34015d4f98a2SYan Zheng extent_key->objectid) 34025d4f98a2SYan Zheng goto next; 34035d4f98a2SYan Zheng 34045d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 34055d4f98a2SYan Zheng if (key.offset != ref_offset) 34065d4f98a2SYan Zheng goto next; 34075d4f98a2SYan Zheng 34085d4f98a2SYan Zheng if (counted) 34095d4f98a2SYan Zheng ref_count--; 34105d4f98a2SYan Zheng if (added) 34115d4f98a2SYan Zheng goto next; 34125d4f98a2SYan Zheng 34135d4f98a2SYan Zheng if (!tree_block_processed(leaf->start, leaf->len, rc)) { 34145d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 34155d4f98a2SYan Zheng if (!block) { 34165d4f98a2SYan Zheng err = -ENOMEM; 34175d4f98a2SYan Zheng break; 34185d4f98a2SYan Zheng } 34195d4f98a2SYan Zheng block->bytenr = leaf->start; 34205d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &block->key, 0); 34215d4f98a2SYan Zheng block->level = 0; 34225d4f98a2SYan Zheng block->key_ready = 1; 34235d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, 34245d4f98a2SYan Zheng &block->rb_node); 34255d4f98a2SYan Zheng BUG_ON(rb_node); 34265d4f98a2SYan Zheng } 34275d4f98a2SYan Zheng if (counted) 34285d4f98a2SYan Zheng added = 1; 34295d4f98a2SYan Zheng else 34305d4f98a2SYan Zheng path->slots[0] = nritems; 34315d4f98a2SYan Zheng next: 34325d4f98a2SYan Zheng path->slots[0]++; 34335d4f98a2SYan Zheng 34345d4f98a2SYan Zheng } 34355d4f98a2SYan Zheng out: 34365d4f98a2SYan Zheng btrfs_free_path(path); 34375d4f98a2SYan Zheng return err; 34385d4f98a2SYan Zheng } 34395d4f98a2SYan Zheng 34405d4f98a2SYan Zheng /* 34415d4f98a2SYan Zheng * hepler to find all tree blocks that reference a given data extent 34425d4f98a2SYan Zheng */ 34435d4f98a2SYan Zheng static noinline_for_stack 34445d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 34455d4f98a2SYan Zheng struct btrfs_key *extent_key, 34465d4f98a2SYan Zheng struct btrfs_path *path, 34475d4f98a2SYan Zheng struct rb_root *blocks) 34485d4f98a2SYan Zheng { 34495d4f98a2SYan Zheng struct btrfs_key key; 34505d4f98a2SYan Zheng struct extent_buffer *eb; 34515d4f98a2SYan Zheng struct btrfs_extent_data_ref *dref; 34525d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 34535d4f98a2SYan Zheng unsigned long ptr; 34545d4f98a2SYan Zheng unsigned long end; 34553fd0a558SYan, Zheng u32 blocksize = btrfs_level_size(rc->extent_root, 0); 34565d4f98a2SYan Zheng int ret; 34575d4f98a2SYan Zheng int err = 0; 34585d4f98a2SYan Zheng 34595d4f98a2SYan Zheng eb = path->nodes[0]; 34605d4f98a2SYan Zheng ptr = btrfs_item_ptr_offset(eb, path->slots[0]); 34615d4f98a2SYan Zheng end = ptr + btrfs_item_size_nr(eb, path->slots[0]); 34625d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 34635d4f98a2SYan Zheng if (ptr + sizeof(struct btrfs_extent_item_v0) == end) 34645d4f98a2SYan Zheng ptr = end; 34655d4f98a2SYan Zheng else 34665d4f98a2SYan Zheng #endif 34675d4f98a2SYan Zheng ptr += sizeof(struct btrfs_extent_item); 34685d4f98a2SYan Zheng 34695d4f98a2SYan Zheng while (ptr < end) { 34705d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 34715d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 34725d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 34735d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 34745d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 34755d4f98a2SYan Zheng blocks); 34765d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 34775d4f98a2SYan Zheng dref = (struct btrfs_extent_data_ref *)(&iref->offset); 34785d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 34795d4f98a2SYan Zheng eb, dref, blocks); 34805d4f98a2SYan Zheng } else { 34815d4f98a2SYan Zheng BUG(); 34825d4f98a2SYan Zheng } 34835d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 34845d4f98a2SYan Zheng } 34855d4f98a2SYan Zheng WARN_ON(ptr > end); 34865d4f98a2SYan Zheng 34875d4f98a2SYan Zheng while (1) { 34885d4f98a2SYan Zheng cond_resched(); 34895d4f98a2SYan Zheng eb = path->nodes[0]; 34905d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(eb)) { 34915d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 34925d4f98a2SYan Zheng if (ret < 0) { 34935d4f98a2SYan Zheng err = ret; 34945d4f98a2SYan Zheng break; 34955d4f98a2SYan Zheng } 34965d4f98a2SYan Zheng if (ret > 0) 34975d4f98a2SYan Zheng break; 34985d4f98a2SYan Zheng eb = path->nodes[0]; 34995d4f98a2SYan Zheng } 35005d4f98a2SYan Zheng 35015d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path->slots[0]); 35025d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 35035d4f98a2SYan Zheng break; 35045d4f98a2SYan Zheng 35055d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 35065d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY || 35075d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 35085d4f98a2SYan Zheng #else 35095d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 35105d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 35115d4f98a2SYan Zheng #endif 35125d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 35135d4f98a2SYan Zheng blocks); 35145d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 35155d4f98a2SYan Zheng dref = btrfs_item_ptr(eb, path->slots[0], 35165d4f98a2SYan Zheng struct btrfs_extent_data_ref); 35175d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 35185d4f98a2SYan Zheng eb, dref, blocks); 35195d4f98a2SYan Zheng } else { 35205d4f98a2SYan Zheng ret = 0; 35215d4f98a2SYan Zheng } 35225d4f98a2SYan Zheng if (ret) { 35235d4f98a2SYan Zheng err = ret; 35245d4f98a2SYan Zheng break; 35255d4f98a2SYan Zheng } 35265d4f98a2SYan Zheng path->slots[0]++; 35275d4f98a2SYan Zheng } 3528b3b4aa74SDavid Sterba btrfs_release_path(path); 35295d4f98a2SYan Zheng if (err) 35305d4f98a2SYan Zheng free_block_list(blocks); 35315d4f98a2SYan Zheng return err; 35325d4f98a2SYan Zheng } 35335d4f98a2SYan Zheng 35345d4f98a2SYan Zheng /* 35355d4f98a2SYan Zheng * hepler to find next unprocessed extent 35365d4f98a2SYan Zheng */ 35375d4f98a2SYan Zheng static noinline_for_stack 35385d4f98a2SYan Zheng int find_next_extent(struct btrfs_trans_handle *trans, 35393fd0a558SYan, Zheng struct reloc_control *rc, struct btrfs_path *path, 35403fd0a558SYan, Zheng struct btrfs_key *extent_key) 35415d4f98a2SYan Zheng { 35425d4f98a2SYan Zheng struct btrfs_key key; 35435d4f98a2SYan Zheng struct extent_buffer *leaf; 35445d4f98a2SYan Zheng u64 start, end, last; 35455d4f98a2SYan Zheng int ret; 35465d4f98a2SYan Zheng 35475d4f98a2SYan Zheng last = rc->block_group->key.objectid + rc->block_group->key.offset; 35485d4f98a2SYan Zheng while (1) { 35495d4f98a2SYan Zheng cond_resched(); 35505d4f98a2SYan Zheng if (rc->search_start >= last) { 35515d4f98a2SYan Zheng ret = 1; 35525d4f98a2SYan Zheng break; 35535d4f98a2SYan Zheng } 35545d4f98a2SYan Zheng 35555d4f98a2SYan Zheng key.objectid = rc->search_start; 35565d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 35575d4f98a2SYan Zheng key.offset = 0; 35585d4f98a2SYan Zheng 35595d4f98a2SYan Zheng path->search_commit_root = 1; 35605d4f98a2SYan Zheng path->skip_locking = 1; 35615d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 35625d4f98a2SYan Zheng 0, 0); 35635d4f98a2SYan Zheng if (ret < 0) 35645d4f98a2SYan Zheng break; 35655d4f98a2SYan Zheng next: 35665d4f98a2SYan Zheng leaf = path->nodes[0]; 35675d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 35685d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 35695d4f98a2SYan Zheng if (ret != 0) 35705d4f98a2SYan Zheng break; 35715d4f98a2SYan Zheng leaf = path->nodes[0]; 35725d4f98a2SYan Zheng } 35735d4f98a2SYan Zheng 35745d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 35755d4f98a2SYan Zheng if (key.objectid >= last) { 35765d4f98a2SYan Zheng ret = 1; 35775d4f98a2SYan Zheng break; 35785d4f98a2SYan Zheng } 35795d4f98a2SYan Zheng 35805d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_ITEM_KEY || 35815d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 35825d4f98a2SYan Zheng path->slots[0]++; 35835d4f98a2SYan Zheng goto next; 35845d4f98a2SYan Zheng } 35855d4f98a2SYan Zheng 35865d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 35875d4f98a2SYan Zheng key.objectid, &start, &end, 35885d4f98a2SYan Zheng EXTENT_DIRTY); 35895d4f98a2SYan Zheng 35905d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3591b3b4aa74SDavid Sterba btrfs_release_path(path); 35925d4f98a2SYan Zheng rc->search_start = end + 1; 35935d4f98a2SYan Zheng } else { 35945d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 35953fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 35965d4f98a2SYan Zheng return 0; 35975d4f98a2SYan Zheng } 35985d4f98a2SYan Zheng } 3599b3b4aa74SDavid Sterba btrfs_release_path(path); 36005d4f98a2SYan Zheng return ret; 36015d4f98a2SYan Zheng } 36025d4f98a2SYan Zheng 36035d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 36045d4f98a2SYan Zheng { 36055d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36067585717fSChris Mason 36077585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 36085d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 36097585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 36105d4f98a2SYan Zheng } 36115d4f98a2SYan Zheng 36125d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 36135d4f98a2SYan Zheng { 36145d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 36157585717fSChris Mason 36167585717fSChris Mason mutex_lock(&fs_info->reloc_mutex); 36175d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 36187585717fSChris Mason mutex_unlock(&fs_info->reloc_mutex); 36195d4f98a2SYan Zheng } 36205d4f98a2SYan Zheng 36215d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 36225d4f98a2SYan Zheng { 36235d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 36245d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 36255d4f98a2SYan Zheng return 1; 36265d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 36275d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 36285d4f98a2SYan Zheng return 1; 36295d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 36305d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 36315d4f98a2SYan Zheng return 1; 36325d4f98a2SYan Zheng return 0; 36335d4f98a2SYan Zheng } 36345d4f98a2SYan Zheng 36353fd0a558SYan, Zheng static noinline_for_stack 36363fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 36373fd0a558SYan, Zheng { 36383fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 36393fd0a558SYan, Zheng int ret; 36403fd0a558SYan, Zheng 36413fd0a558SYan, Zheng rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root); 36423fd0a558SYan, Zheng if (!rc->block_rsv) 36433fd0a558SYan, Zheng return -ENOMEM; 36443fd0a558SYan, Zheng 36453fd0a558SYan, Zheng /* 36463fd0a558SYan, Zheng * reserve some space for creating reloc trees. 36473fd0a558SYan, Zheng * btrfs_init_reloc_root will use them when there 36483fd0a558SYan, Zheng * is no reservation in transaction handle. 36493fd0a558SYan, Zheng */ 36504a92b1b8SJosef Bacik ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv, 36518bb8ab2eSJosef Bacik rc->extent_root->nodesize * 256); 36523fd0a558SYan, Zheng if (ret) 36533fd0a558SYan, Zheng return ret; 36543fd0a558SYan, Zheng 36553fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 36563fd0a558SYan, Zheng rc->search_start = rc->block_group->key.objectid; 36573fd0a558SYan, Zheng rc->extents_found = 0; 36583fd0a558SYan, Zheng rc->nodes_relocated = 0; 36593fd0a558SYan, Zheng rc->merging_rsv_size = 0; 36603fd0a558SYan, Zheng 36613fd0a558SYan, Zheng rc->create_reloc_tree = 1; 36623fd0a558SYan, Zheng set_reloc_control(rc); 36633fd0a558SYan, Zheng 36647a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 36653612b495STsutomu Itoh BUG_ON(IS_ERR(trans)); 36663fd0a558SYan, Zheng btrfs_commit_transaction(trans, rc->extent_root); 36673fd0a558SYan, Zheng return 0; 36683fd0a558SYan, Zheng } 366976dda93cSYan, Zheng 36705d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 36715d4f98a2SYan Zheng { 36725d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 36735d4f98a2SYan Zheng struct btrfs_key key; 36745d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 36755d4f98a2SYan Zheng struct btrfs_path *path; 36765d4f98a2SYan Zheng struct btrfs_extent_item *ei; 36775d4f98a2SYan Zheng unsigned long nr; 36785d4f98a2SYan Zheng u64 flags; 36795d4f98a2SYan Zheng u32 item_size; 36805d4f98a2SYan Zheng int ret; 36815d4f98a2SYan Zheng int err = 0; 3682c87f08caSChris Mason int progress = 0; 36835d4f98a2SYan Zheng 36845d4f98a2SYan Zheng path = btrfs_alloc_path(); 36853fd0a558SYan, Zheng if (!path) 36865d4f98a2SYan Zheng return -ENOMEM; 3687026fd317SJosef Bacik path->reada = 1; 36883fd0a558SYan, Zheng 36893fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 36903fd0a558SYan, Zheng if (ret) { 36913fd0a558SYan, Zheng err = ret; 36923fd0a558SYan, Zheng goto out_free; 36932423fdfbSJiri Slaby } 36945d4f98a2SYan Zheng 36955d4f98a2SYan Zheng while (1) { 3696c87f08caSChris Mason progress++; 3697a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 369898d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 3699c87f08caSChris Mason restart: 37003fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 37013fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 37023fd0a558SYan, Zheng continue; 37033fd0a558SYan, Zheng } 37043fd0a558SYan, Zheng 37053fd0a558SYan, Zheng ret = find_next_extent(trans, rc, path, &key); 37065d4f98a2SYan Zheng if (ret < 0) 37075d4f98a2SYan Zheng err = ret; 37085d4f98a2SYan Zheng if (ret != 0) 37095d4f98a2SYan Zheng break; 37105d4f98a2SYan Zheng 37115d4f98a2SYan Zheng rc->extents_found++; 37125d4f98a2SYan Zheng 37135d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 37145d4f98a2SYan Zheng struct btrfs_extent_item); 37153fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 37165d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 37175d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 37185d4f98a2SYan Zheng ret = check_extent_flags(flags); 37195d4f98a2SYan Zheng BUG_ON(ret); 37205d4f98a2SYan Zheng 37215d4f98a2SYan Zheng } else { 37225d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 37235d4f98a2SYan Zheng u64 ref_owner; 37245d4f98a2SYan Zheng int path_change = 0; 37255d4f98a2SYan Zheng 37265d4f98a2SYan Zheng BUG_ON(item_size != 37275d4f98a2SYan Zheng sizeof(struct btrfs_extent_item_v0)); 37285d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, &key, &ref_owner, 37295d4f98a2SYan Zheng &path_change); 37305d4f98a2SYan Zheng if (ref_owner < BTRFS_FIRST_FREE_OBJECTID) 37315d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_TREE_BLOCK; 37325d4f98a2SYan Zheng else 37335d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_DATA; 37345d4f98a2SYan Zheng 37355d4f98a2SYan Zheng if (path_change) { 3736b3b4aa74SDavid Sterba btrfs_release_path(path); 37375d4f98a2SYan Zheng 37385d4f98a2SYan Zheng path->search_commit_root = 1; 37395d4f98a2SYan Zheng path->skip_locking = 1; 37405d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, 37415d4f98a2SYan Zheng &key, path, 0, 0); 37425d4f98a2SYan Zheng if (ret < 0) { 37435d4f98a2SYan Zheng err = ret; 37445d4f98a2SYan Zheng break; 37455d4f98a2SYan Zheng } 37465d4f98a2SYan Zheng BUG_ON(ret > 0); 37475d4f98a2SYan Zheng } 37485d4f98a2SYan Zheng #else 37495d4f98a2SYan Zheng BUG(); 37505d4f98a2SYan Zheng #endif 37515d4f98a2SYan Zheng } 37525d4f98a2SYan Zheng 37535d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 37545d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 37555d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 37565d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 37575d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 37585d4f98a2SYan Zheng } else { 3759b3b4aa74SDavid Sterba btrfs_release_path(path); 37605d4f98a2SYan Zheng ret = 0; 37615d4f98a2SYan Zheng } 37625d4f98a2SYan Zheng if (ret < 0) { 37633fd0a558SYan, Zheng err = ret; 37645d4f98a2SYan Zheng break; 37655d4f98a2SYan Zheng } 37665d4f98a2SYan Zheng 37675d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 37685d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 37695d4f98a2SYan Zheng if (ret < 0) { 37703fd0a558SYan, Zheng if (ret != -EAGAIN) { 37715d4f98a2SYan Zheng err = ret; 37725d4f98a2SYan Zheng break; 37735d4f98a2SYan Zheng } 37743fd0a558SYan, Zheng rc->extents_found--; 37753fd0a558SYan, Zheng rc->search_start = key.objectid; 37763fd0a558SYan, Zheng } 37775d4f98a2SYan Zheng } 37785d4f98a2SYan Zheng 377936ba022aSJosef Bacik ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5); 37803fd0a558SYan, Zheng if (ret < 0) { 37813fd0a558SYan, Zheng if (ret != -EAGAIN) { 37823fd0a558SYan, Zheng err = ret; 37833fd0a558SYan, Zheng WARN_ON(1); 37843fd0a558SYan, Zheng break; 37853fd0a558SYan, Zheng } 37863fd0a558SYan, Zheng rc->commit_transaction = 1; 37873fd0a558SYan, Zheng } 37883fd0a558SYan, Zheng 37893fd0a558SYan, Zheng if (rc->commit_transaction) { 37903fd0a558SYan, Zheng rc->commit_transaction = 0; 37913fd0a558SYan, Zheng ret = btrfs_commit_transaction(trans, rc->extent_root); 37923fd0a558SYan, Zheng BUG_ON(ret); 37933fd0a558SYan, Zheng } else { 37945d4f98a2SYan Zheng nr = trans->blocks_used; 37953fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 37965d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 37973fd0a558SYan, Zheng } 37983fd0a558SYan, Zheng trans = NULL; 37995d4f98a2SYan Zheng 38005d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 38015d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 38025d4f98a2SYan Zheng rc->found_file_extent = 1; 38030257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 38043fd0a558SYan, Zheng &key, &rc->cluster); 38055d4f98a2SYan Zheng if (ret < 0) { 38065d4f98a2SYan Zheng err = ret; 38075d4f98a2SYan Zheng break; 38085d4f98a2SYan Zheng } 38095d4f98a2SYan Zheng } 38105d4f98a2SYan Zheng } 3811c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 3812c87f08caSChris Mason ret = btrfs_force_chunk_alloc(trans, rc->extent_root, 3813c87f08caSChris Mason rc->block_group->flags); 3814c87f08caSChris Mason if (ret == 0) { 3815c87f08caSChris Mason err = 0; 3816c87f08caSChris Mason progress = 0; 3817c87f08caSChris Mason goto restart; 3818c87f08caSChris Mason } 3819c87f08caSChris Mason } 38203fd0a558SYan, Zheng 3821b3b4aa74SDavid Sterba btrfs_release_path(path); 38223fd0a558SYan, Zheng clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY, 38233fd0a558SYan, Zheng GFP_NOFS); 38245d4f98a2SYan Zheng 38255d4f98a2SYan Zheng if (trans) { 38265d4f98a2SYan Zheng nr = trans->blocks_used; 38273fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 38285d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 38295d4f98a2SYan Zheng } 38305d4f98a2SYan Zheng 38310257bb82SYan, Zheng if (!err) { 38323fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 38333fd0a558SYan, Zheng &rc->cluster); 38340257bb82SYan, Zheng if (ret < 0) 38350257bb82SYan, Zheng err = ret; 38360257bb82SYan, Zheng } 38370257bb82SYan, Zheng 38383fd0a558SYan, Zheng rc->create_reloc_tree = 0; 38393fd0a558SYan, Zheng set_reloc_control(rc); 38400257bb82SYan, Zheng 38413fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 38423fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1); 38435d4f98a2SYan Zheng 38443fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 38455d4f98a2SYan Zheng 38465d4f98a2SYan Zheng merge_reloc_roots(rc); 38475d4f98a2SYan Zheng 38483fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 38495d4f98a2SYan Zheng unset_reloc_control(rc); 38503fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1); 38515d4f98a2SYan Zheng 38525d4f98a2SYan Zheng /* get rid of pinned extents */ 38537a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 38543612b495STsutomu Itoh if (IS_ERR(trans)) 38553612b495STsutomu Itoh err = PTR_ERR(trans); 38563612b495STsutomu Itoh else 38575d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 38583fd0a558SYan, Zheng out_free: 38593fd0a558SYan, Zheng btrfs_free_block_rsv(rc->extent_root, rc->block_rsv); 38603fd0a558SYan, Zheng btrfs_free_path(path); 38615d4f98a2SYan Zheng return err; 38625d4f98a2SYan Zheng } 38635d4f98a2SYan Zheng 38645d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 38650257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 38665d4f98a2SYan Zheng { 38675d4f98a2SYan Zheng struct btrfs_path *path; 38685d4f98a2SYan Zheng struct btrfs_inode_item *item; 38695d4f98a2SYan Zheng struct extent_buffer *leaf; 38705d4f98a2SYan Zheng int ret; 38715d4f98a2SYan Zheng 38725d4f98a2SYan Zheng path = btrfs_alloc_path(); 38735d4f98a2SYan Zheng if (!path) 38745d4f98a2SYan Zheng return -ENOMEM; 38755d4f98a2SYan Zheng 38765d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 38775d4f98a2SYan Zheng if (ret) 38785d4f98a2SYan Zheng goto out; 38795d4f98a2SYan Zheng 38805d4f98a2SYan Zheng leaf = path->nodes[0]; 38815d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 38825d4f98a2SYan Zheng memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item)); 38835d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 38840257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 38855d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 38863fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 38873fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 38885d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 3889b3b4aa74SDavid Sterba btrfs_release_path(path); 38905d4f98a2SYan Zheng out: 38915d4f98a2SYan Zheng btrfs_free_path(path); 38925d4f98a2SYan Zheng return ret; 38935d4f98a2SYan Zheng } 38945d4f98a2SYan Zheng 38955d4f98a2SYan Zheng /* 38965d4f98a2SYan Zheng * helper to create inode for data relocation. 38975d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 38985d4f98a2SYan Zheng */ 38993fd0a558SYan, Zheng static noinline_for_stack 39003fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 39015d4f98a2SYan Zheng struct btrfs_block_group_cache *group) 39025d4f98a2SYan Zheng { 39035d4f98a2SYan Zheng struct inode *inode = NULL; 39045d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 39055d4f98a2SYan Zheng struct btrfs_root *root; 39065d4f98a2SYan Zheng struct btrfs_key key; 39075d4f98a2SYan Zheng unsigned long nr; 39085d4f98a2SYan Zheng u64 objectid = BTRFS_FIRST_FREE_OBJECTID; 39095d4f98a2SYan Zheng int err = 0; 39105d4f98a2SYan Zheng 39115d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 39125d4f98a2SYan Zheng if (IS_ERR(root)) 39135d4f98a2SYan Zheng return ERR_CAST(root); 39145d4f98a2SYan Zheng 3915a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 39163fd0a558SYan, Zheng if (IS_ERR(trans)) 39173fd0a558SYan, Zheng return ERR_CAST(trans); 39185d4f98a2SYan Zheng 3919581bb050SLi Zefan err = btrfs_find_free_objectid(root, &objectid); 39205d4f98a2SYan Zheng if (err) 39215d4f98a2SYan Zheng goto out; 39225d4f98a2SYan Zheng 39230257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 39245d4f98a2SYan Zheng BUG_ON(err); 39255d4f98a2SYan Zheng 39265d4f98a2SYan Zheng key.objectid = objectid; 39275d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 39285d4f98a2SYan Zheng key.offset = 0; 392973f73415SJosef Bacik inode = btrfs_iget(root->fs_info->sb, &key, root, NULL); 39305d4f98a2SYan Zheng BUG_ON(IS_ERR(inode) || is_bad_inode(inode)); 39315d4f98a2SYan Zheng BTRFS_I(inode)->index_cnt = group->key.objectid; 39325d4f98a2SYan Zheng 39335d4f98a2SYan Zheng err = btrfs_orphan_add(trans, inode); 39345d4f98a2SYan Zheng out: 39355d4f98a2SYan Zheng nr = trans->blocks_used; 39365d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 39375d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 39385d4f98a2SYan Zheng if (err) { 39395d4f98a2SYan Zheng if (inode) 39405d4f98a2SYan Zheng iput(inode); 39415d4f98a2SYan Zheng inode = ERR_PTR(err); 39425d4f98a2SYan Zheng } 39435d4f98a2SYan Zheng return inode; 39445d4f98a2SYan Zheng } 39455d4f98a2SYan Zheng 39463fd0a558SYan, Zheng static struct reloc_control *alloc_reloc_control(void) 39473fd0a558SYan, Zheng { 39483fd0a558SYan, Zheng struct reloc_control *rc; 39493fd0a558SYan, Zheng 39503fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 39513fd0a558SYan, Zheng if (!rc) 39523fd0a558SYan, Zheng return NULL; 39533fd0a558SYan, Zheng 39543fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 39553fd0a558SYan, Zheng backref_cache_init(&rc->backref_cache); 39563fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 3957f993c883SDavid Sterba extent_io_tree_init(&rc->processed_blocks, NULL); 39583fd0a558SYan, Zheng return rc; 39593fd0a558SYan, Zheng } 39603fd0a558SYan, Zheng 39615d4f98a2SYan Zheng /* 39625d4f98a2SYan Zheng * function to relocate all extents in a block group. 39635d4f98a2SYan Zheng */ 39645d4f98a2SYan Zheng int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start) 39655d4f98a2SYan Zheng { 39665d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = extent_root->fs_info; 39675d4f98a2SYan Zheng struct reloc_control *rc; 39680af3d00bSJosef Bacik struct inode *inode; 39690af3d00bSJosef Bacik struct btrfs_path *path; 39705d4f98a2SYan Zheng int ret; 3971f0486c68SYan, Zheng int rw = 0; 39725d4f98a2SYan Zheng int err = 0; 39735d4f98a2SYan Zheng 39743fd0a558SYan, Zheng rc = alloc_reloc_control(); 39755d4f98a2SYan Zheng if (!rc) 39765d4f98a2SYan Zheng return -ENOMEM; 39775d4f98a2SYan Zheng 3978f0486c68SYan, Zheng rc->extent_root = extent_root; 39793fd0a558SYan, Zheng 39805d4f98a2SYan Zheng rc->block_group = btrfs_lookup_block_group(fs_info, group_start); 39815d4f98a2SYan Zheng BUG_ON(!rc->block_group); 39825d4f98a2SYan Zheng 3983f0486c68SYan, Zheng if (!rc->block_group->ro) { 3984f0486c68SYan, Zheng ret = btrfs_set_block_group_ro(extent_root, rc->block_group); 3985f0486c68SYan, Zheng if (ret) { 3986f0486c68SYan, Zheng err = ret; 3987f0486c68SYan, Zheng goto out; 3988f0486c68SYan, Zheng } 3989f0486c68SYan, Zheng rw = 1; 3990f0486c68SYan, Zheng } 3991f0486c68SYan, Zheng 39920af3d00bSJosef Bacik path = btrfs_alloc_path(); 39930af3d00bSJosef Bacik if (!path) { 39940af3d00bSJosef Bacik err = -ENOMEM; 39950af3d00bSJosef Bacik goto out; 39960af3d00bSJosef Bacik } 39970af3d00bSJosef Bacik 39980af3d00bSJosef Bacik inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group, 39990af3d00bSJosef Bacik path); 40000af3d00bSJosef Bacik btrfs_free_path(path); 40010af3d00bSJosef Bacik 40020af3d00bSJosef Bacik if (!IS_ERR(inode)) 40030af3d00bSJosef Bacik ret = delete_block_group_cache(fs_info, inode, 0); 40040af3d00bSJosef Bacik else 40050af3d00bSJosef Bacik ret = PTR_ERR(inode); 40060af3d00bSJosef Bacik 40070af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 40080af3d00bSJosef Bacik err = ret; 40090af3d00bSJosef Bacik goto out; 40100af3d00bSJosef Bacik } 40110af3d00bSJosef Bacik 40125d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 40135d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 40145d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 40155d4f98a2SYan Zheng rc->data_inode = NULL; 40165d4f98a2SYan Zheng goto out; 40175d4f98a2SYan Zheng } 40185d4f98a2SYan Zheng 40195d4f98a2SYan Zheng printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n", 40205d4f98a2SYan Zheng (unsigned long long)rc->block_group->key.objectid, 40215d4f98a2SYan Zheng (unsigned long long)rc->block_group->flags); 40225d4f98a2SYan Zheng 402324bbcf04SYan, Zheng btrfs_start_delalloc_inodes(fs_info->tree_root, 0); 402424bbcf04SYan, Zheng btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0); 40255d4f98a2SYan Zheng 40265d4f98a2SYan Zheng while (1) { 402776dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 402876dda93cSYan, Zheng 402976dda93cSYan, Zheng btrfs_clean_old_snapshots(fs_info->tree_root); 40305d4f98a2SYan Zheng ret = relocate_block_group(rc); 403176dda93cSYan, Zheng 403276dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 40335d4f98a2SYan Zheng if (ret < 0) { 40345d4f98a2SYan Zheng err = ret; 40353fd0a558SYan, Zheng goto out; 40365d4f98a2SYan Zheng } 40375d4f98a2SYan Zheng 40385d4f98a2SYan Zheng if (rc->extents_found == 0) 40395d4f98a2SYan Zheng break; 40405d4f98a2SYan Zheng 40415d4f98a2SYan Zheng printk(KERN_INFO "btrfs: found %llu extents\n", 40425d4f98a2SYan Zheng (unsigned long long)rc->extents_found); 40435d4f98a2SYan Zheng 40445d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 40455d4f98a2SYan Zheng btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1); 40465d4f98a2SYan Zheng invalidate_mapping_pages(rc->data_inode->i_mapping, 40475d4f98a2SYan Zheng 0, -1); 40485d4f98a2SYan Zheng rc->stage = UPDATE_DATA_PTRS; 40495d4f98a2SYan Zheng } 40505d4f98a2SYan Zheng } 40515d4f98a2SYan Zheng 40520257bb82SYan, Zheng filemap_write_and_wait_range(fs_info->btree_inode->i_mapping, 40535d4f98a2SYan Zheng rc->block_group->key.objectid, 40545d4f98a2SYan Zheng rc->block_group->key.objectid + 40555d4f98a2SYan Zheng rc->block_group->key.offset - 1); 40565d4f98a2SYan Zheng 40575d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 40585d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 40595d4f98a2SYan Zheng WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0); 40605d4f98a2SYan Zheng out: 4061f0486c68SYan, Zheng if (err && rw) 4062f0486c68SYan, Zheng btrfs_set_block_group_rw(extent_root, rc->block_group); 40635d4f98a2SYan Zheng iput(rc->data_inode); 40645d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 40655d4f98a2SYan Zheng kfree(rc); 40665d4f98a2SYan Zheng return err; 40675d4f98a2SYan Zheng } 40685d4f98a2SYan Zheng 406976dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 407076dda93cSYan, Zheng { 407176dda93cSYan, Zheng struct btrfs_trans_handle *trans; 407276dda93cSYan, Zheng int ret; 407376dda93cSYan, Zheng 4074a22285a6SYan, Zheng trans = btrfs_start_transaction(root->fs_info->tree_root, 0); 407598d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 407676dda93cSYan, Zheng 407776dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 407876dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 407976dda93cSYan, Zheng root->root_item.drop_level = 0; 408076dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 408176dda93cSYan, Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 408276dda93cSYan, Zheng &root->root_key, &root->root_item); 408376dda93cSYan, Zheng BUG_ON(ret); 408476dda93cSYan, Zheng 408576dda93cSYan, Zheng ret = btrfs_end_transaction(trans, root->fs_info->tree_root); 408676dda93cSYan, Zheng BUG_ON(ret); 408776dda93cSYan, Zheng return 0; 408876dda93cSYan, Zheng } 408976dda93cSYan, Zheng 40905d4f98a2SYan Zheng /* 40915d4f98a2SYan Zheng * recover relocation interrupted by system crash. 40925d4f98a2SYan Zheng * 40935d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 40945d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 40955d4f98a2SYan Zheng */ 40965d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 40975d4f98a2SYan Zheng { 40985d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 40995d4f98a2SYan Zheng struct btrfs_key key; 41005d4f98a2SYan Zheng struct btrfs_root *fs_root; 41015d4f98a2SYan Zheng struct btrfs_root *reloc_root; 41025d4f98a2SYan Zheng struct btrfs_path *path; 41035d4f98a2SYan Zheng struct extent_buffer *leaf; 41045d4f98a2SYan Zheng struct reloc_control *rc = NULL; 41055d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 41065d4f98a2SYan Zheng int ret; 41075d4f98a2SYan Zheng int err = 0; 41085d4f98a2SYan Zheng 41095d4f98a2SYan Zheng path = btrfs_alloc_path(); 41105d4f98a2SYan Zheng if (!path) 41115d4f98a2SYan Zheng return -ENOMEM; 4112026fd317SJosef Bacik path->reada = -1; 41135d4f98a2SYan Zheng 41145d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 41155d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 41165d4f98a2SYan Zheng key.offset = (u64)-1; 41175d4f98a2SYan Zheng 41185d4f98a2SYan Zheng while (1) { 41195d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, 41205d4f98a2SYan Zheng path, 0, 0); 41215d4f98a2SYan Zheng if (ret < 0) { 41225d4f98a2SYan Zheng err = ret; 41235d4f98a2SYan Zheng goto out; 41245d4f98a2SYan Zheng } 41255d4f98a2SYan Zheng if (ret > 0) { 41265d4f98a2SYan Zheng if (path->slots[0] == 0) 41275d4f98a2SYan Zheng break; 41285d4f98a2SYan Zheng path->slots[0]--; 41295d4f98a2SYan Zheng } 41305d4f98a2SYan Zheng leaf = path->nodes[0]; 41315d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4132b3b4aa74SDavid Sterba btrfs_release_path(path); 41335d4f98a2SYan Zheng 41345d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 41355d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 41365d4f98a2SYan Zheng break; 41375d4f98a2SYan Zheng 41385d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root, &key); 41395d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 41405d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 41415d4f98a2SYan Zheng goto out; 41425d4f98a2SYan Zheng } 41435d4f98a2SYan Zheng 41445d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 41455d4f98a2SYan Zheng 41465d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 41475d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 41485d4f98a2SYan Zheng reloc_root->root_key.offset); 41495d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 415076dda93cSYan, Zheng ret = PTR_ERR(fs_root); 415176dda93cSYan, Zheng if (ret != -ENOENT) { 415276dda93cSYan, Zheng err = ret; 41535d4f98a2SYan Zheng goto out; 41545d4f98a2SYan Zheng } 415576dda93cSYan, Zheng mark_garbage_root(reloc_root); 415676dda93cSYan, Zheng } 41575d4f98a2SYan Zheng } 41585d4f98a2SYan Zheng 41595d4f98a2SYan Zheng if (key.offset == 0) 41605d4f98a2SYan Zheng break; 41615d4f98a2SYan Zheng 41625d4f98a2SYan Zheng key.offset--; 41635d4f98a2SYan Zheng } 4164b3b4aa74SDavid Sterba btrfs_release_path(path); 41655d4f98a2SYan Zheng 41665d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 41675d4f98a2SYan Zheng goto out; 41685d4f98a2SYan Zheng 41693fd0a558SYan, Zheng rc = alloc_reloc_control(); 41705d4f98a2SYan Zheng if (!rc) { 41715d4f98a2SYan Zheng err = -ENOMEM; 41725d4f98a2SYan Zheng goto out; 41735d4f98a2SYan Zheng } 41745d4f98a2SYan Zheng 41755d4f98a2SYan Zheng rc->extent_root = root->fs_info->extent_root; 41765d4f98a2SYan Zheng 41775d4f98a2SYan Zheng set_reloc_control(rc); 41785d4f98a2SYan Zheng 41797a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 41803612b495STsutomu Itoh if (IS_ERR(trans)) { 41813612b495STsutomu Itoh unset_reloc_control(rc); 41823612b495STsutomu Itoh err = PTR_ERR(trans); 41833612b495STsutomu Itoh goto out_free; 41843612b495STsutomu Itoh } 41853fd0a558SYan, Zheng 41863fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 41873fd0a558SYan, Zheng 41885d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 41895d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 41905d4f98a2SYan Zheng struct btrfs_root, root_list); 41915d4f98a2SYan Zheng list_del(&reloc_root->root_list); 41925d4f98a2SYan Zheng 41935d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 41945d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 41955d4f98a2SYan Zheng &rc->reloc_roots); 41965d4f98a2SYan Zheng continue; 41975d4f98a2SYan Zheng } 41985d4f98a2SYan Zheng 41995d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 42005d4f98a2SYan Zheng reloc_root->root_key.offset); 42015d4f98a2SYan Zheng BUG_ON(IS_ERR(fs_root)); 42025d4f98a2SYan Zheng 42035d4f98a2SYan Zheng __add_reloc_root(reloc_root); 42045d4f98a2SYan Zheng fs_root->reloc_root = reloc_root; 42055d4f98a2SYan Zheng } 42065d4f98a2SYan Zheng 42075d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 42085d4f98a2SYan Zheng 42095d4f98a2SYan Zheng merge_reloc_roots(rc); 42105d4f98a2SYan Zheng 42115d4f98a2SYan Zheng unset_reloc_control(rc); 42125d4f98a2SYan Zheng 42137a7eaa40SJosef Bacik trans = btrfs_join_transaction(rc->extent_root); 42143612b495STsutomu Itoh if (IS_ERR(trans)) 42153612b495STsutomu Itoh err = PTR_ERR(trans); 42163612b495STsutomu Itoh else 42175d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 42183612b495STsutomu Itoh out_free: 42195d4f98a2SYan Zheng kfree(rc); 42203612b495STsutomu Itoh out: 42215d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 42225d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 42235d4f98a2SYan Zheng struct btrfs_root, root_list); 42245d4f98a2SYan Zheng list_del(&reloc_root->root_list); 42255d4f98a2SYan Zheng free_extent_buffer(reloc_root->node); 42265d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 42275d4f98a2SYan Zheng kfree(reloc_root); 42285d4f98a2SYan Zheng } 42295d4f98a2SYan Zheng btrfs_free_path(path); 42305d4f98a2SYan Zheng 42315d4f98a2SYan Zheng if (err == 0) { 42325d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 42335d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 42345d4f98a2SYan Zheng BTRFS_DATA_RELOC_TREE_OBJECTID); 42355d4f98a2SYan Zheng if (IS_ERR(fs_root)) 42365d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4237d7ce5843SMiao Xie else 423866b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 42395d4f98a2SYan Zheng } 42405d4f98a2SYan Zheng return err; 42415d4f98a2SYan Zheng } 42425d4f98a2SYan Zheng 42435d4f98a2SYan Zheng /* 42445d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 42455d4f98a2SYan Zheng * 42465d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 42475d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 42485d4f98a2SYan Zheng */ 42495d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 42505d4f98a2SYan Zheng { 42515d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 42525d4f98a2SYan Zheng struct btrfs_sector_sum *sector_sum; 42535d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 42545d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 42555d4f98a2SYan Zheng size_t offset; 42565d4f98a2SYan Zheng int ret; 42575d4f98a2SYan Zheng u64 disk_bytenr; 42585d4f98a2SYan Zheng LIST_HEAD(list); 42595d4f98a2SYan Zheng 42605d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 42615d4f98a2SYan Zheng BUG_ON(ordered->file_offset != file_pos || ordered->len != len); 42625d4f98a2SYan Zheng 42635d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 42645d4f98a2SYan Zheng ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr, 4265a2de733cSArne Jansen disk_bytenr + len - 1, &list, 0); 42665d4f98a2SYan Zheng 42675d4f98a2SYan Zheng while (!list_empty(&list)) { 42685d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 42695d4f98a2SYan Zheng list_del_init(&sums->list); 42705d4f98a2SYan Zheng 42715d4f98a2SYan Zheng sector_sum = sums->sums; 42725d4f98a2SYan Zheng sums->bytenr = ordered->start; 42735d4f98a2SYan Zheng 42745d4f98a2SYan Zheng offset = 0; 42755d4f98a2SYan Zheng while (offset < sums->len) { 42765d4f98a2SYan Zheng sector_sum->bytenr += ordered->start - disk_bytenr; 42775d4f98a2SYan Zheng sector_sum++; 42785d4f98a2SYan Zheng offset += root->sectorsize; 42795d4f98a2SYan Zheng } 42805d4f98a2SYan Zheng 42815d4f98a2SYan Zheng btrfs_add_ordered_sum(inode, ordered, sums); 42825d4f98a2SYan Zheng } 42835d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4284411fc6bcSAndi Kleen return ret; 42855d4f98a2SYan Zheng } 42863fd0a558SYan, Zheng 42873fd0a558SYan, Zheng void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 42883fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 42893fd0a558SYan, Zheng struct extent_buffer *cow) 42903fd0a558SYan, Zheng { 42913fd0a558SYan, Zheng struct reloc_control *rc; 42923fd0a558SYan, Zheng struct backref_node *node; 42933fd0a558SYan, Zheng int first_cow = 0; 42943fd0a558SYan, Zheng int level; 42953fd0a558SYan, Zheng int ret; 42963fd0a558SYan, Zheng 42973fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 42983fd0a558SYan, Zheng if (!rc) 42993fd0a558SYan, Zheng return; 43003fd0a558SYan, Zheng 43013fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 43023fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 43033fd0a558SYan, Zheng 43043fd0a558SYan, Zheng level = btrfs_header_level(buf); 43053fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 43063fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 43073fd0a558SYan, Zheng first_cow = 1; 43083fd0a558SYan, Zheng 43093fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 43103fd0a558SYan, Zheng rc->create_reloc_tree) { 43113fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 43123fd0a558SYan, Zheng 43133fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 43143fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 43153fd0a558SYan, Zheng node->new_bytenr != buf->start); 43163fd0a558SYan, Zheng 43173fd0a558SYan, Zheng drop_node_buffer(node); 43183fd0a558SYan, Zheng extent_buffer_get(cow); 43193fd0a558SYan, Zheng node->eb = cow; 43203fd0a558SYan, Zheng node->new_bytenr = cow->start; 43213fd0a558SYan, Zheng 43223fd0a558SYan, Zheng if (!node->pending) { 43233fd0a558SYan, Zheng list_move_tail(&node->list, 43243fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 43253fd0a558SYan, Zheng node->pending = 1; 43263fd0a558SYan, Zheng } 43273fd0a558SYan, Zheng 43283fd0a558SYan, Zheng if (first_cow) 43293fd0a558SYan, Zheng __mark_block_processed(rc, node); 43303fd0a558SYan, Zheng 43313fd0a558SYan, Zheng if (first_cow && level > 0) 43323fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 43333fd0a558SYan, Zheng } 43343fd0a558SYan, Zheng 43353fd0a558SYan, Zheng if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) { 43363fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 43373fd0a558SYan, Zheng BUG_ON(ret); 43383fd0a558SYan, Zheng } 43393fd0a558SYan, Zheng } 43403fd0a558SYan, Zheng 43413fd0a558SYan, Zheng /* 43423fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 43433fd0a558SYan, Zheng * requried for relocating tree blocks in the snapshot 43443fd0a558SYan, Zheng */ 43453fd0a558SYan, Zheng void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans, 43463fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending, 43473fd0a558SYan, Zheng u64 *bytes_to_reserve) 43483fd0a558SYan, Zheng { 43493fd0a558SYan, Zheng struct btrfs_root *root; 43503fd0a558SYan, Zheng struct reloc_control *rc; 43513fd0a558SYan, Zheng 43523fd0a558SYan, Zheng root = pending->root; 43533fd0a558SYan, Zheng if (!root->reloc_root) 43543fd0a558SYan, Zheng return; 43553fd0a558SYan, Zheng 43563fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 43573fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 43583fd0a558SYan, Zheng return; 43593fd0a558SYan, Zheng 43603fd0a558SYan, Zheng root = root->reloc_root; 43613fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 43623fd0a558SYan, Zheng /* 43633fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 43643fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 43653fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 43663fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 43673fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 43683fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 43693fd0a558SYan, Zheng * tree may use more space than it frees. so we need 43703fd0a558SYan, Zheng * reserve extra space. 43713fd0a558SYan, Zheng */ 43723fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 43733fd0a558SYan, Zheng } 43743fd0a558SYan, Zheng 43753fd0a558SYan, Zheng /* 43763fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 43773fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 43783fd0a558SYan, Zheng */ 43793fd0a558SYan, Zheng void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 43803fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 43813fd0a558SYan, Zheng { 43823fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 43833fd0a558SYan, Zheng struct btrfs_root *reloc_root; 43843fd0a558SYan, Zheng struct btrfs_root *new_root; 43853fd0a558SYan, Zheng struct reloc_control *rc; 43863fd0a558SYan, Zheng int ret; 43873fd0a558SYan, Zheng 43883fd0a558SYan, Zheng if (!root->reloc_root) 43893fd0a558SYan, Zheng return; 43903fd0a558SYan, Zheng 43913fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 43923fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 43933fd0a558SYan, Zheng 43943fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 43953fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 43963fd0a558SYan, Zheng rc->block_rsv, 43973fd0a558SYan, Zheng rc->nodes_relocated); 43983fd0a558SYan, Zheng BUG_ON(ret); 43993fd0a558SYan, Zheng } 44003fd0a558SYan, Zheng 44013fd0a558SYan, Zheng new_root = pending->snap; 44023fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 44033fd0a558SYan, Zheng new_root->root_key.objectid); 44043fd0a558SYan, Zheng 44053fd0a558SYan, Zheng __add_reloc_root(reloc_root); 44063fd0a558SYan, Zheng new_root->reloc_root = reloc_root; 44073fd0a558SYan, Zheng 44083fd0a558SYan, Zheng if (rc->create_reloc_tree) { 44093fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 44103fd0a558SYan, Zheng BUG_ON(ret); 44113fd0a558SYan, Zheng } 44123fd0a558SYan, Zheng } 4413