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" 335d4f98a2SYan Zheng 345d4f98a2SYan Zheng /* 355d4f98a2SYan Zheng * backref_node, mapping_node and tree_block start with this 365d4f98a2SYan Zheng */ 375d4f98a2SYan Zheng struct tree_entry { 385d4f98a2SYan Zheng struct rb_node rb_node; 395d4f98a2SYan Zheng u64 bytenr; 405d4f98a2SYan Zheng }; 415d4f98a2SYan Zheng 425d4f98a2SYan Zheng /* 435d4f98a2SYan Zheng * present a tree block in the backref cache 445d4f98a2SYan Zheng */ 455d4f98a2SYan Zheng struct backref_node { 465d4f98a2SYan Zheng struct rb_node rb_node; 475d4f98a2SYan Zheng u64 bytenr; 483fd0a558SYan, Zheng 493fd0a558SYan, Zheng u64 new_bytenr; 503fd0a558SYan, Zheng /* objectid of tree block owner, can be not uptodate */ 515d4f98a2SYan Zheng u64 owner; 523fd0a558SYan, Zheng /* link to pending, changed or detached list */ 533fd0a558SYan, Zheng struct list_head list; 545d4f98a2SYan Zheng /* list of upper level blocks reference this block */ 555d4f98a2SYan Zheng struct list_head upper; 565d4f98a2SYan Zheng /* list of child blocks in the cache */ 575d4f98a2SYan Zheng struct list_head lower; 585d4f98a2SYan Zheng /* NULL if this node is not tree root */ 595d4f98a2SYan Zheng struct btrfs_root *root; 605d4f98a2SYan Zheng /* extent buffer got by COW the block */ 615d4f98a2SYan Zheng struct extent_buffer *eb; 625d4f98a2SYan Zheng /* level of tree block */ 635d4f98a2SYan Zheng unsigned int level:8; 643fd0a558SYan, Zheng /* is the block in non-reference counted tree */ 653fd0a558SYan, Zheng unsigned int cowonly:1; 663fd0a558SYan, Zheng /* 1 if no child node in the cache */ 675d4f98a2SYan Zheng unsigned int lowest:1; 685d4f98a2SYan Zheng /* is the extent buffer locked */ 695d4f98a2SYan Zheng unsigned int locked:1; 705d4f98a2SYan Zheng /* has the block been processed */ 715d4f98a2SYan Zheng unsigned int processed:1; 725d4f98a2SYan Zheng /* have backrefs of this block been checked */ 735d4f98a2SYan Zheng unsigned int checked:1; 743fd0a558SYan, Zheng /* 753fd0a558SYan, Zheng * 1 if corresponding block has been cowed but some upper 763fd0a558SYan, Zheng * level block pointers may not point to the new location 773fd0a558SYan, Zheng */ 783fd0a558SYan, Zheng unsigned int pending:1; 793fd0a558SYan, Zheng /* 803fd0a558SYan, Zheng * 1 if the backref node isn't connected to any other 813fd0a558SYan, Zheng * backref node. 823fd0a558SYan, Zheng */ 833fd0a558SYan, Zheng unsigned int detached:1; 845d4f98a2SYan Zheng }; 855d4f98a2SYan Zheng 865d4f98a2SYan Zheng /* 875d4f98a2SYan Zheng * present a block pointer in the backref cache 885d4f98a2SYan Zheng */ 895d4f98a2SYan Zheng struct backref_edge { 905d4f98a2SYan Zheng struct list_head list[2]; 915d4f98a2SYan Zheng struct backref_node *node[2]; 925d4f98a2SYan Zheng }; 935d4f98a2SYan Zheng 945d4f98a2SYan Zheng #define LOWER 0 955d4f98a2SYan Zheng #define UPPER 1 965d4f98a2SYan Zheng 975d4f98a2SYan Zheng struct backref_cache { 985d4f98a2SYan Zheng /* red black tree of all backref nodes in the cache */ 995d4f98a2SYan Zheng struct rb_root rb_root; 1003fd0a558SYan, Zheng /* for passing backref nodes to btrfs_reloc_cow_block */ 1013fd0a558SYan, Zheng struct backref_node *path[BTRFS_MAX_LEVEL]; 1023fd0a558SYan, Zheng /* 1033fd0a558SYan, Zheng * list of blocks that have been cowed but some block 1043fd0a558SYan, Zheng * pointers in upper level blocks may not reflect the 1053fd0a558SYan, Zheng * new location 1063fd0a558SYan, Zheng */ 1075d4f98a2SYan Zheng struct list_head pending[BTRFS_MAX_LEVEL]; 1083fd0a558SYan, Zheng /* list of backref nodes with no child node */ 1093fd0a558SYan, Zheng struct list_head leaves; 1103fd0a558SYan, Zheng /* list of blocks that have been cowed in current transaction */ 1113fd0a558SYan, Zheng struct list_head changed; 1123fd0a558SYan, Zheng /* list of detached backref node. */ 1133fd0a558SYan, Zheng struct list_head detached; 1143fd0a558SYan, Zheng 1153fd0a558SYan, Zheng u64 last_trans; 1163fd0a558SYan, Zheng 1173fd0a558SYan, Zheng int nr_nodes; 1183fd0a558SYan, Zheng int nr_edges; 1195d4f98a2SYan Zheng }; 1205d4f98a2SYan Zheng 1215d4f98a2SYan Zheng /* 1225d4f98a2SYan Zheng * map address of tree root to tree 1235d4f98a2SYan Zheng */ 1245d4f98a2SYan Zheng struct mapping_node { 1255d4f98a2SYan Zheng struct rb_node rb_node; 1265d4f98a2SYan Zheng u64 bytenr; 1275d4f98a2SYan Zheng void *data; 1285d4f98a2SYan Zheng }; 1295d4f98a2SYan Zheng 1305d4f98a2SYan Zheng struct mapping_tree { 1315d4f98a2SYan Zheng struct rb_root rb_root; 1325d4f98a2SYan Zheng spinlock_t lock; 1335d4f98a2SYan Zheng }; 1345d4f98a2SYan Zheng 1355d4f98a2SYan Zheng /* 1365d4f98a2SYan Zheng * present a tree block to process 1375d4f98a2SYan Zheng */ 1385d4f98a2SYan Zheng struct tree_block { 1395d4f98a2SYan Zheng struct rb_node rb_node; 1405d4f98a2SYan Zheng u64 bytenr; 1415d4f98a2SYan Zheng struct btrfs_key key; 1425d4f98a2SYan Zheng unsigned int level:8; 1435d4f98a2SYan Zheng unsigned int key_ready:1; 1445d4f98a2SYan Zheng }; 1455d4f98a2SYan Zheng 1460257bb82SYan, Zheng #define MAX_EXTENTS 128 1470257bb82SYan, Zheng 1480257bb82SYan, Zheng struct file_extent_cluster { 1490257bb82SYan, Zheng u64 start; 1500257bb82SYan, Zheng u64 end; 1510257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 1520257bb82SYan, Zheng unsigned int nr; 1530257bb82SYan, Zheng }; 1540257bb82SYan, Zheng 1555d4f98a2SYan Zheng struct reloc_control { 1565d4f98a2SYan Zheng /* block group to relocate */ 1575d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group; 1585d4f98a2SYan Zheng /* extent tree */ 1595d4f98a2SYan Zheng struct btrfs_root *extent_root; 1605d4f98a2SYan Zheng /* inode for moving data */ 1615d4f98a2SYan Zheng struct inode *data_inode; 1623fd0a558SYan, Zheng 1633fd0a558SYan, Zheng struct btrfs_block_rsv *block_rsv; 1643fd0a558SYan, Zheng 1653fd0a558SYan, Zheng struct backref_cache backref_cache; 1663fd0a558SYan, Zheng 1673fd0a558SYan, Zheng struct file_extent_cluster cluster; 1685d4f98a2SYan Zheng /* tree blocks have been processed */ 1695d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 1705d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 1715d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 1725d4f98a2SYan Zheng /* list of reloc trees */ 1735d4f98a2SYan Zheng struct list_head reloc_roots; 1743fd0a558SYan, Zheng /* size of metadata reservation for merging reloc trees */ 1753fd0a558SYan, Zheng u64 merging_rsv_size; 1763fd0a558SYan, Zheng /* size of relocated tree nodes */ 1773fd0a558SYan, Zheng u64 nodes_relocated; 1783fd0a558SYan, Zheng 1795d4f98a2SYan Zheng u64 search_start; 1805d4f98a2SYan Zheng u64 extents_found; 1813fd0a558SYan, Zheng 1823fd0a558SYan, Zheng unsigned int stage:8; 1833fd0a558SYan, Zheng unsigned int create_reloc_tree:1; 1843fd0a558SYan, Zheng unsigned int merge_reloc_tree:1; 1855d4f98a2SYan Zheng unsigned int found_file_extent:1; 1863fd0a558SYan, Zheng unsigned int commit_transaction:1; 1875d4f98a2SYan Zheng }; 1885d4f98a2SYan Zheng 1895d4f98a2SYan Zheng /* stages of data relocation */ 1905d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 1915d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 1925d4f98a2SYan Zheng 1933fd0a558SYan, Zheng static void remove_backref_node(struct backref_cache *cache, 1943fd0a558SYan, Zheng struct backref_node *node); 1953fd0a558SYan, Zheng static void __mark_block_processed(struct reloc_control *rc, 1963fd0a558SYan, Zheng struct backref_node *node); 1975d4f98a2SYan Zheng 1985d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 1995d4f98a2SYan Zheng { 2006bef4d31SEric Paris tree->rb_root = RB_ROOT; 2015d4f98a2SYan Zheng spin_lock_init(&tree->lock); 2025d4f98a2SYan Zheng } 2035d4f98a2SYan Zheng 2045d4f98a2SYan Zheng static void backref_cache_init(struct backref_cache *cache) 2055d4f98a2SYan Zheng { 2065d4f98a2SYan Zheng int i; 2076bef4d31SEric Paris cache->rb_root = RB_ROOT; 2085d4f98a2SYan Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 2095d4f98a2SYan Zheng INIT_LIST_HEAD(&cache->pending[i]); 2103fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->changed); 2113fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->detached); 2123fd0a558SYan, Zheng INIT_LIST_HEAD(&cache->leaves); 2135d4f98a2SYan Zheng } 2145d4f98a2SYan Zheng 2153fd0a558SYan, Zheng static void backref_cache_cleanup(struct backref_cache *cache) 2165d4f98a2SYan Zheng { 2173fd0a558SYan, Zheng struct backref_node *node; 2183fd0a558SYan, Zheng int i; 2193fd0a558SYan, Zheng 2203fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 2213fd0a558SYan, Zheng node = list_entry(cache->detached.next, 2223fd0a558SYan, Zheng struct backref_node, list); 2233fd0a558SYan, Zheng remove_backref_node(cache, node); 2243fd0a558SYan, Zheng } 2253fd0a558SYan, Zheng 2263fd0a558SYan, Zheng while (!list_empty(&cache->leaves)) { 2273fd0a558SYan, Zheng node = list_entry(cache->leaves.next, 2283fd0a558SYan, Zheng struct backref_node, lower); 2293fd0a558SYan, Zheng remove_backref_node(cache, node); 2303fd0a558SYan, Zheng } 2313fd0a558SYan, Zheng 2323fd0a558SYan, Zheng cache->last_trans = 0; 2333fd0a558SYan, Zheng 2343fd0a558SYan, Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 2353fd0a558SYan, Zheng BUG_ON(!list_empty(&cache->pending[i])); 2363fd0a558SYan, Zheng BUG_ON(!list_empty(&cache->changed)); 2373fd0a558SYan, Zheng BUG_ON(!list_empty(&cache->detached)); 2383fd0a558SYan, Zheng BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root)); 2393fd0a558SYan, Zheng BUG_ON(cache->nr_nodes); 2403fd0a558SYan, Zheng BUG_ON(cache->nr_edges); 2413fd0a558SYan, Zheng } 2423fd0a558SYan, Zheng 2433fd0a558SYan, Zheng static struct backref_node *alloc_backref_node(struct backref_cache *cache) 2443fd0a558SYan, Zheng { 2453fd0a558SYan, Zheng struct backref_node *node; 2463fd0a558SYan, Zheng 2473fd0a558SYan, Zheng node = kzalloc(sizeof(*node), GFP_NOFS); 2483fd0a558SYan, Zheng if (node) { 2493fd0a558SYan, Zheng INIT_LIST_HEAD(&node->list); 2505d4f98a2SYan Zheng INIT_LIST_HEAD(&node->upper); 2515d4f98a2SYan Zheng INIT_LIST_HEAD(&node->lower); 2525d4f98a2SYan Zheng RB_CLEAR_NODE(&node->rb_node); 2533fd0a558SYan, Zheng cache->nr_nodes++; 2543fd0a558SYan, Zheng } 2553fd0a558SYan, Zheng return node; 2563fd0a558SYan, Zheng } 2573fd0a558SYan, Zheng 2583fd0a558SYan, Zheng static void free_backref_node(struct backref_cache *cache, 2593fd0a558SYan, Zheng struct backref_node *node) 2603fd0a558SYan, Zheng { 2613fd0a558SYan, Zheng if (node) { 2623fd0a558SYan, Zheng cache->nr_nodes--; 2633fd0a558SYan, Zheng kfree(node); 2643fd0a558SYan, Zheng } 2653fd0a558SYan, Zheng } 2663fd0a558SYan, Zheng 2673fd0a558SYan, Zheng static struct backref_edge *alloc_backref_edge(struct backref_cache *cache) 2683fd0a558SYan, Zheng { 2693fd0a558SYan, Zheng struct backref_edge *edge; 2703fd0a558SYan, Zheng 2713fd0a558SYan, Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 2723fd0a558SYan, Zheng if (edge) 2733fd0a558SYan, Zheng cache->nr_edges++; 2743fd0a558SYan, Zheng return edge; 2753fd0a558SYan, Zheng } 2763fd0a558SYan, Zheng 2773fd0a558SYan, Zheng static void free_backref_edge(struct backref_cache *cache, 2783fd0a558SYan, Zheng struct backref_edge *edge) 2793fd0a558SYan, Zheng { 2803fd0a558SYan, Zheng if (edge) { 2813fd0a558SYan, Zheng cache->nr_edges--; 2823fd0a558SYan, Zheng kfree(edge); 2833fd0a558SYan, Zheng } 2845d4f98a2SYan Zheng } 2855d4f98a2SYan Zheng 2865d4f98a2SYan Zheng static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr, 2875d4f98a2SYan Zheng struct rb_node *node) 2885d4f98a2SYan Zheng { 2895d4f98a2SYan Zheng struct rb_node **p = &root->rb_node; 2905d4f98a2SYan Zheng struct rb_node *parent = NULL; 2915d4f98a2SYan Zheng struct tree_entry *entry; 2925d4f98a2SYan Zheng 2935d4f98a2SYan Zheng while (*p) { 2945d4f98a2SYan Zheng parent = *p; 2955d4f98a2SYan Zheng entry = rb_entry(parent, struct tree_entry, rb_node); 2965d4f98a2SYan Zheng 2975d4f98a2SYan Zheng if (bytenr < entry->bytenr) 2985d4f98a2SYan Zheng p = &(*p)->rb_left; 2995d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 3005d4f98a2SYan Zheng p = &(*p)->rb_right; 3015d4f98a2SYan Zheng else 3025d4f98a2SYan Zheng return parent; 3035d4f98a2SYan Zheng } 3045d4f98a2SYan Zheng 3055d4f98a2SYan Zheng rb_link_node(node, parent, p); 3065d4f98a2SYan Zheng rb_insert_color(node, root); 3075d4f98a2SYan Zheng return NULL; 3085d4f98a2SYan Zheng } 3095d4f98a2SYan Zheng 3105d4f98a2SYan Zheng static struct rb_node *tree_search(struct rb_root *root, u64 bytenr) 3115d4f98a2SYan Zheng { 3125d4f98a2SYan Zheng struct rb_node *n = root->rb_node; 3135d4f98a2SYan Zheng struct tree_entry *entry; 3145d4f98a2SYan Zheng 3155d4f98a2SYan Zheng while (n) { 3165d4f98a2SYan Zheng entry = rb_entry(n, struct tree_entry, rb_node); 3175d4f98a2SYan Zheng 3185d4f98a2SYan Zheng if (bytenr < entry->bytenr) 3195d4f98a2SYan Zheng n = n->rb_left; 3205d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 3215d4f98a2SYan Zheng n = n->rb_right; 3225d4f98a2SYan Zheng else 3235d4f98a2SYan Zheng return n; 3245d4f98a2SYan Zheng } 3255d4f98a2SYan Zheng return NULL; 3265d4f98a2SYan Zheng } 3275d4f98a2SYan Zheng 3285d4f98a2SYan Zheng /* 3295d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 3305d4f98a2SYan Zheng */ 3315d4f98a2SYan Zheng static struct backref_node *walk_up_backref(struct backref_node *node, 3325d4f98a2SYan Zheng struct backref_edge *edges[], 3335d4f98a2SYan Zheng int *index) 3345d4f98a2SYan Zheng { 3355d4f98a2SYan Zheng struct backref_edge *edge; 3365d4f98a2SYan Zheng int idx = *index; 3375d4f98a2SYan Zheng 3385d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 3395d4f98a2SYan Zheng edge = list_entry(node->upper.next, 3405d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 3415d4f98a2SYan Zheng edges[idx++] = edge; 3425d4f98a2SYan Zheng node = edge->node[UPPER]; 3435d4f98a2SYan Zheng } 3443fd0a558SYan, Zheng BUG_ON(node->detached); 3455d4f98a2SYan Zheng *index = idx; 3465d4f98a2SYan Zheng return node; 3475d4f98a2SYan Zheng } 3485d4f98a2SYan Zheng 3495d4f98a2SYan Zheng /* 3505d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 3515d4f98a2SYan Zheng */ 3525d4f98a2SYan Zheng static struct backref_node *walk_down_backref(struct backref_edge *edges[], 3535d4f98a2SYan Zheng int *index) 3545d4f98a2SYan Zheng { 3555d4f98a2SYan Zheng struct backref_edge *edge; 3565d4f98a2SYan Zheng struct backref_node *lower; 3575d4f98a2SYan Zheng int idx = *index; 3585d4f98a2SYan Zheng 3595d4f98a2SYan Zheng while (idx > 0) { 3605d4f98a2SYan Zheng edge = edges[idx - 1]; 3615d4f98a2SYan Zheng lower = edge->node[LOWER]; 3625d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 3635d4f98a2SYan Zheng idx--; 3645d4f98a2SYan Zheng continue; 3655d4f98a2SYan Zheng } 3665d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 3675d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 3685d4f98a2SYan Zheng edges[idx - 1] = edge; 3695d4f98a2SYan Zheng *index = idx; 3705d4f98a2SYan Zheng return edge->node[UPPER]; 3715d4f98a2SYan Zheng } 3725d4f98a2SYan Zheng *index = 0; 3735d4f98a2SYan Zheng return NULL; 3745d4f98a2SYan Zheng } 3755d4f98a2SYan Zheng 3763fd0a558SYan, Zheng static void unlock_node_buffer(struct backref_node *node) 3775d4f98a2SYan Zheng { 3785d4f98a2SYan Zheng if (node->locked) { 3795d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 3805d4f98a2SYan Zheng node->locked = 0; 3815d4f98a2SYan Zheng } 3823fd0a558SYan, Zheng } 3833fd0a558SYan, Zheng 3843fd0a558SYan, Zheng static void drop_node_buffer(struct backref_node *node) 3853fd0a558SYan, Zheng { 3863fd0a558SYan, Zheng if (node->eb) { 3873fd0a558SYan, Zheng unlock_node_buffer(node); 3885d4f98a2SYan Zheng free_extent_buffer(node->eb); 3895d4f98a2SYan Zheng node->eb = NULL; 3905d4f98a2SYan Zheng } 3915d4f98a2SYan Zheng } 3925d4f98a2SYan Zheng 3935d4f98a2SYan Zheng static void drop_backref_node(struct backref_cache *tree, 3945d4f98a2SYan Zheng struct backref_node *node) 3955d4f98a2SYan Zheng { 3965d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 3975d4f98a2SYan Zheng 3985d4f98a2SYan Zheng drop_node_buffer(node); 3993fd0a558SYan, Zheng list_del(&node->list); 4005d4f98a2SYan Zheng list_del(&node->lower); 4013fd0a558SYan, Zheng if (!RB_EMPTY_NODE(&node->rb_node)) 4025d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 4033fd0a558SYan, Zheng free_backref_node(tree, node); 4045d4f98a2SYan Zheng } 4055d4f98a2SYan Zheng 4065d4f98a2SYan Zheng /* 4075d4f98a2SYan Zheng * remove a backref node from the backref cache 4085d4f98a2SYan Zheng */ 4095d4f98a2SYan Zheng static void remove_backref_node(struct backref_cache *cache, 4105d4f98a2SYan Zheng struct backref_node *node) 4115d4f98a2SYan Zheng { 4125d4f98a2SYan Zheng struct backref_node *upper; 4135d4f98a2SYan Zheng struct backref_edge *edge; 4145d4f98a2SYan Zheng 4155d4f98a2SYan Zheng if (!node) 4165d4f98a2SYan Zheng return; 4175d4f98a2SYan Zheng 4183fd0a558SYan, Zheng BUG_ON(!node->lowest && !node->detached); 4195d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 4205d4f98a2SYan Zheng edge = list_entry(node->upper.next, struct backref_edge, 4215d4f98a2SYan Zheng list[LOWER]); 4225d4f98a2SYan Zheng upper = edge->node[UPPER]; 4235d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 4245d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 4253fd0a558SYan, Zheng free_backref_edge(cache, edge); 4263fd0a558SYan, Zheng 4273fd0a558SYan, Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 4283fd0a558SYan, Zheng BUG_ON(!list_empty(&node->upper)); 4293fd0a558SYan, Zheng drop_backref_node(cache, node); 4303fd0a558SYan, Zheng node = upper; 4313fd0a558SYan, Zheng node->lowest = 1; 4323fd0a558SYan, Zheng continue; 4333fd0a558SYan, Zheng } 4345d4f98a2SYan Zheng /* 4353fd0a558SYan, Zheng * add the node to leaf node list if no other 4365d4f98a2SYan Zheng * child block cached. 4375d4f98a2SYan Zheng */ 4385d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 4393fd0a558SYan, Zheng list_add_tail(&upper->lower, &cache->leaves); 4405d4f98a2SYan Zheng upper->lowest = 1; 4415d4f98a2SYan Zheng } 4425d4f98a2SYan Zheng } 4433fd0a558SYan, Zheng 4445d4f98a2SYan Zheng drop_backref_node(cache, node); 4455d4f98a2SYan Zheng } 4465d4f98a2SYan Zheng 4473fd0a558SYan, Zheng static void update_backref_node(struct backref_cache *cache, 4483fd0a558SYan, Zheng struct backref_node *node, u64 bytenr) 4493fd0a558SYan, Zheng { 4503fd0a558SYan, Zheng struct rb_node *rb_node; 4513fd0a558SYan, Zheng rb_erase(&node->rb_node, &cache->rb_root); 4523fd0a558SYan, Zheng node->bytenr = bytenr; 4533fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 4543fd0a558SYan, Zheng BUG_ON(rb_node); 4553fd0a558SYan, Zheng } 4563fd0a558SYan, Zheng 4573fd0a558SYan, Zheng /* 4583fd0a558SYan, Zheng * update backref cache after a transaction commit 4593fd0a558SYan, Zheng */ 4603fd0a558SYan, Zheng static int update_backref_cache(struct btrfs_trans_handle *trans, 4613fd0a558SYan, Zheng struct backref_cache *cache) 4623fd0a558SYan, Zheng { 4633fd0a558SYan, Zheng struct backref_node *node; 4643fd0a558SYan, Zheng int level = 0; 4653fd0a558SYan, Zheng 4663fd0a558SYan, Zheng if (cache->last_trans == 0) { 4673fd0a558SYan, Zheng cache->last_trans = trans->transid; 4683fd0a558SYan, Zheng return 0; 4693fd0a558SYan, Zheng } 4703fd0a558SYan, Zheng 4713fd0a558SYan, Zheng if (cache->last_trans == trans->transid) 4723fd0a558SYan, Zheng return 0; 4733fd0a558SYan, Zheng 4743fd0a558SYan, Zheng /* 4753fd0a558SYan, Zheng * detached nodes are used to avoid unnecessary backref 4763fd0a558SYan, Zheng * lookup. transaction commit changes the extent tree. 4773fd0a558SYan, Zheng * so the detached nodes are no longer useful. 4783fd0a558SYan, Zheng */ 4793fd0a558SYan, Zheng while (!list_empty(&cache->detached)) { 4803fd0a558SYan, Zheng node = list_entry(cache->detached.next, 4813fd0a558SYan, Zheng struct backref_node, list); 4823fd0a558SYan, Zheng remove_backref_node(cache, node); 4833fd0a558SYan, Zheng } 4843fd0a558SYan, Zheng 4853fd0a558SYan, Zheng while (!list_empty(&cache->changed)) { 4863fd0a558SYan, Zheng node = list_entry(cache->changed.next, 4873fd0a558SYan, Zheng struct backref_node, list); 4883fd0a558SYan, Zheng list_del_init(&node->list); 4893fd0a558SYan, Zheng BUG_ON(node->pending); 4903fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 4913fd0a558SYan, Zheng } 4923fd0a558SYan, Zheng 4933fd0a558SYan, Zheng /* 4943fd0a558SYan, Zheng * some nodes can be left in the pending list if there were 4953fd0a558SYan, Zheng * errors during processing the pending nodes. 4963fd0a558SYan, Zheng */ 4973fd0a558SYan, Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 4983fd0a558SYan, Zheng list_for_each_entry(node, &cache->pending[level], list) { 4993fd0a558SYan, Zheng BUG_ON(!node->pending); 5003fd0a558SYan, Zheng if (node->bytenr == node->new_bytenr) 5013fd0a558SYan, Zheng continue; 5023fd0a558SYan, Zheng update_backref_node(cache, node, node->new_bytenr); 5033fd0a558SYan, Zheng } 5043fd0a558SYan, Zheng } 5053fd0a558SYan, Zheng 5063fd0a558SYan, Zheng cache->last_trans = 0; 5073fd0a558SYan, Zheng return 1; 5083fd0a558SYan, Zheng } 5093fd0a558SYan, Zheng 5103fd0a558SYan, Zheng static int should_ignore_root(struct btrfs_root *root) 5113fd0a558SYan, Zheng { 5123fd0a558SYan, Zheng struct btrfs_root *reloc_root; 5133fd0a558SYan, Zheng 5143fd0a558SYan, Zheng if (!root->ref_cows) 5153fd0a558SYan, Zheng return 0; 5163fd0a558SYan, Zheng 5173fd0a558SYan, Zheng reloc_root = root->reloc_root; 5183fd0a558SYan, Zheng if (!reloc_root) 5193fd0a558SYan, Zheng return 0; 5203fd0a558SYan, Zheng 5213fd0a558SYan, Zheng if (btrfs_root_last_snapshot(&reloc_root->root_item) == 5223fd0a558SYan, Zheng root->fs_info->running_transaction->transid - 1) 5233fd0a558SYan, Zheng return 0; 5243fd0a558SYan, Zheng /* 5253fd0a558SYan, Zheng * if there is reloc tree and it was created in previous 5263fd0a558SYan, Zheng * transaction backref lookup can find the reloc tree, 5273fd0a558SYan, Zheng * so backref node for the fs tree root is useless for 5283fd0a558SYan, Zheng * relocation. 5293fd0a558SYan, Zheng */ 5303fd0a558SYan, Zheng return 1; 5313fd0a558SYan, Zheng } 5323fd0a558SYan, Zheng 5335d4f98a2SYan Zheng /* 5345d4f98a2SYan Zheng * find reloc tree by address of tree root 5355d4f98a2SYan Zheng */ 5365d4f98a2SYan Zheng static struct btrfs_root *find_reloc_root(struct reloc_control *rc, 5375d4f98a2SYan Zheng u64 bytenr) 5385d4f98a2SYan Zheng { 5395d4f98a2SYan Zheng struct rb_node *rb_node; 5405d4f98a2SYan Zheng struct mapping_node *node; 5415d4f98a2SYan Zheng struct btrfs_root *root = NULL; 5425d4f98a2SYan Zheng 5435d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 5445d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr); 5455d4f98a2SYan Zheng if (rb_node) { 5465d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 5475d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 5485d4f98a2SYan Zheng } 5495d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 5505d4f98a2SYan Zheng return root; 5515d4f98a2SYan Zheng } 5525d4f98a2SYan Zheng 5535d4f98a2SYan Zheng static int is_cowonly_root(u64 root_objectid) 5545d4f98a2SYan Zheng { 5555d4f98a2SYan Zheng if (root_objectid == BTRFS_ROOT_TREE_OBJECTID || 5565d4f98a2SYan Zheng root_objectid == BTRFS_EXTENT_TREE_OBJECTID || 5575d4f98a2SYan Zheng root_objectid == BTRFS_CHUNK_TREE_OBJECTID || 5585d4f98a2SYan Zheng root_objectid == BTRFS_DEV_TREE_OBJECTID || 5595d4f98a2SYan Zheng root_objectid == BTRFS_TREE_LOG_OBJECTID || 5605d4f98a2SYan Zheng root_objectid == BTRFS_CSUM_TREE_OBJECTID) 5615d4f98a2SYan Zheng return 1; 5625d4f98a2SYan Zheng return 0; 5635d4f98a2SYan Zheng } 5645d4f98a2SYan Zheng 5655d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 5665d4f98a2SYan Zheng u64 root_objectid) 5675d4f98a2SYan Zheng { 5685d4f98a2SYan Zheng struct btrfs_key key; 5695d4f98a2SYan Zheng 5705d4f98a2SYan Zheng key.objectid = root_objectid; 5715d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 5725d4f98a2SYan Zheng if (is_cowonly_root(root_objectid)) 5735d4f98a2SYan Zheng key.offset = 0; 5745d4f98a2SYan Zheng else 5755d4f98a2SYan Zheng key.offset = (u64)-1; 5765d4f98a2SYan Zheng 5775d4f98a2SYan Zheng return btrfs_read_fs_root_no_name(fs_info, &key); 5785d4f98a2SYan Zheng } 5795d4f98a2SYan Zheng 5805d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 5815d4f98a2SYan Zheng static noinline_for_stack 5825d4f98a2SYan Zheng struct btrfs_root *find_tree_root(struct reloc_control *rc, 5835d4f98a2SYan Zheng struct extent_buffer *leaf, 5845d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0) 5855d4f98a2SYan Zheng { 5865d4f98a2SYan Zheng struct btrfs_root *root; 5875d4f98a2SYan Zheng u64 root_objectid = btrfs_ref_root_v0(leaf, ref0); 5885d4f98a2SYan Zheng u64 generation = btrfs_ref_generation_v0(leaf, ref0); 5895d4f98a2SYan Zheng 5905d4f98a2SYan Zheng BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID); 5915d4f98a2SYan Zheng 5925d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, root_objectid); 5935d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 5945d4f98a2SYan Zheng 5955d4f98a2SYan Zheng if (root->ref_cows && 5965d4f98a2SYan Zheng generation != btrfs_root_generation(&root->root_item)) 5975d4f98a2SYan Zheng return NULL; 5985d4f98a2SYan Zheng 5995d4f98a2SYan Zheng return root; 6005d4f98a2SYan Zheng } 6015d4f98a2SYan Zheng #endif 6025d4f98a2SYan Zheng 6035d4f98a2SYan Zheng static noinline_for_stack 6045d4f98a2SYan Zheng int find_inline_backref(struct extent_buffer *leaf, int slot, 6055d4f98a2SYan Zheng unsigned long *ptr, unsigned long *end) 6065d4f98a2SYan Zheng { 6075d4f98a2SYan Zheng struct btrfs_extent_item *ei; 6085d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 6095d4f98a2SYan Zheng u32 item_size; 6105d4f98a2SYan Zheng 6115d4f98a2SYan Zheng item_size = btrfs_item_size_nr(leaf, slot); 6125d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 6135d4f98a2SYan Zheng if (item_size < sizeof(*ei)) { 6145d4f98a2SYan Zheng WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 6155d4f98a2SYan Zheng return 1; 6165d4f98a2SYan Zheng } 6175d4f98a2SYan Zheng #endif 6185d4f98a2SYan Zheng ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item); 6195d4f98a2SYan Zheng WARN_ON(!(btrfs_extent_flags(leaf, ei) & 6205d4f98a2SYan Zheng BTRFS_EXTENT_FLAG_TREE_BLOCK)); 6215d4f98a2SYan Zheng 6225d4f98a2SYan Zheng if (item_size <= sizeof(*ei) + sizeof(*bi)) { 6235d4f98a2SYan Zheng WARN_ON(item_size < sizeof(*ei) + sizeof(*bi)); 6245d4f98a2SYan Zheng return 1; 6255d4f98a2SYan Zheng } 6265d4f98a2SYan Zheng 6275d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 6285d4f98a2SYan Zheng *ptr = (unsigned long)(bi + 1); 6295d4f98a2SYan Zheng *end = (unsigned long)ei + item_size; 6305d4f98a2SYan Zheng return 0; 6315d4f98a2SYan Zheng } 6325d4f98a2SYan Zheng 6335d4f98a2SYan Zheng /* 6345d4f98a2SYan Zheng * build backref tree for a given tree block. root of the backref tree 6355d4f98a2SYan Zheng * corresponds the tree block, leaves of the backref tree correspond 6365d4f98a2SYan Zheng * roots of b-trees that reference the tree block. 6375d4f98a2SYan Zheng * 6385d4f98a2SYan Zheng * the basic idea of this function is check backrefs of a given block 6395d4f98a2SYan Zheng * to find upper level blocks that refernece the block, and then check 6405d4f98a2SYan Zheng * bakcrefs of these upper level blocks recursively. the recursion stop 6415d4f98a2SYan Zheng * when tree root is reached or backrefs for the block is cached. 6425d4f98a2SYan Zheng * 6435d4f98a2SYan Zheng * NOTE: if we find backrefs for a block are cached, we know backrefs 6445d4f98a2SYan Zheng * for all upper level blocks that directly/indirectly reference the 6455d4f98a2SYan Zheng * block are also cached. 6465d4f98a2SYan Zheng */ 6473fd0a558SYan, Zheng static noinline_for_stack 6483fd0a558SYan, Zheng struct backref_node *build_backref_tree(struct reloc_control *rc, 6495d4f98a2SYan Zheng struct btrfs_key *node_key, 6505d4f98a2SYan Zheng int level, u64 bytenr) 6515d4f98a2SYan Zheng { 6523fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 6535d4f98a2SYan Zheng struct btrfs_path *path1; 6545d4f98a2SYan Zheng struct btrfs_path *path2; 6555d4f98a2SYan Zheng struct extent_buffer *eb; 6565d4f98a2SYan Zheng struct btrfs_root *root; 6575d4f98a2SYan Zheng struct backref_node *cur; 6585d4f98a2SYan Zheng struct backref_node *upper; 6595d4f98a2SYan Zheng struct backref_node *lower; 6605d4f98a2SYan Zheng struct backref_node *node = NULL; 6615d4f98a2SYan Zheng struct backref_node *exist = NULL; 6625d4f98a2SYan Zheng struct backref_edge *edge; 6635d4f98a2SYan Zheng struct rb_node *rb_node; 6645d4f98a2SYan Zheng struct btrfs_key key; 6655d4f98a2SYan Zheng unsigned long end; 6665d4f98a2SYan Zheng unsigned long ptr; 6675d4f98a2SYan Zheng LIST_HEAD(list); 6683fd0a558SYan, Zheng LIST_HEAD(useless); 6693fd0a558SYan, Zheng int cowonly; 6705d4f98a2SYan Zheng int ret; 6715d4f98a2SYan Zheng int err = 0; 6725d4f98a2SYan Zheng 6735d4f98a2SYan Zheng path1 = btrfs_alloc_path(); 6745d4f98a2SYan Zheng path2 = btrfs_alloc_path(); 6755d4f98a2SYan Zheng if (!path1 || !path2) { 6765d4f98a2SYan Zheng err = -ENOMEM; 6775d4f98a2SYan Zheng goto out; 6785d4f98a2SYan Zheng } 6795d4f98a2SYan Zheng 6803fd0a558SYan, Zheng node = alloc_backref_node(cache); 6815d4f98a2SYan Zheng if (!node) { 6825d4f98a2SYan Zheng err = -ENOMEM; 6835d4f98a2SYan Zheng goto out; 6845d4f98a2SYan Zheng } 6855d4f98a2SYan Zheng 6865d4f98a2SYan Zheng node->bytenr = bytenr; 6875d4f98a2SYan Zheng node->level = level; 6885d4f98a2SYan Zheng node->lowest = 1; 6895d4f98a2SYan Zheng cur = node; 6905d4f98a2SYan Zheng again: 6915d4f98a2SYan Zheng end = 0; 6925d4f98a2SYan Zheng ptr = 0; 6935d4f98a2SYan Zheng key.objectid = cur->bytenr; 6945d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 6955d4f98a2SYan Zheng key.offset = (u64)-1; 6965d4f98a2SYan Zheng 6975d4f98a2SYan Zheng path1->search_commit_root = 1; 6985d4f98a2SYan Zheng path1->skip_locking = 1; 6995d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1, 7005d4f98a2SYan Zheng 0, 0); 7015d4f98a2SYan Zheng if (ret < 0) { 7025d4f98a2SYan Zheng err = ret; 7035d4f98a2SYan Zheng goto out; 7045d4f98a2SYan Zheng } 7055d4f98a2SYan Zheng BUG_ON(!ret || !path1->slots[0]); 7065d4f98a2SYan Zheng 7075d4f98a2SYan Zheng path1->slots[0]--; 7085d4f98a2SYan Zheng 7095d4f98a2SYan Zheng WARN_ON(cur->checked); 7105d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 7115d4f98a2SYan Zheng /* 7125d4f98a2SYan Zheng * the backref was added previously when processsing 7135d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 7145d4f98a2SYan Zheng */ 7155d4f98a2SYan Zheng BUG_ON(!list_is_singular(&cur->upper)); 7165d4f98a2SYan Zheng edge = list_entry(cur->upper.next, struct backref_edge, 7175d4f98a2SYan Zheng list[LOWER]); 7185d4f98a2SYan Zheng BUG_ON(!list_empty(&edge->list[UPPER])); 7195d4f98a2SYan Zheng exist = edge->node[UPPER]; 7205d4f98a2SYan Zheng /* 7215d4f98a2SYan Zheng * add the upper level block to pending list if we need 7225d4f98a2SYan Zheng * check its backrefs 7235d4f98a2SYan Zheng */ 7245d4f98a2SYan Zheng if (!exist->checked) 7255d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 7265d4f98a2SYan Zheng } else { 7275d4f98a2SYan Zheng exist = NULL; 7285d4f98a2SYan Zheng } 7295d4f98a2SYan Zheng 7305d4f98a2SYan Zheng while (1) { 7315d4f98a2SYan Zheng cond_resched(); 7325d4f98a2SYan Zheng eb = path1->nodes[0]; 7335d4f98a2SYan Zheng 7345d4f98a2SYan Zheng if (ptr >= end) { 7355d4f98a2SYan Zheng if (path1->slots[0] >= btrfs_header_nritems(eb)) { 7365d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path1); 7375d4f98a2SYan Zheng if (ret < 0) { 7385d4f98a2SYan Zheng err = ret; 7395d4f98a2SYan Zheng goto out; 7405d4f98a2SYan Zheng } 7415d4f98a2SYan Zheng if (ret > 0) 7425d4f98a2SYan Zheng break; 7435d4f98a2SYan Zheng eb = path1->nodes[0]; 7445d4f98a2SYan Zheng } 7455d4f98a2SYan Zheng 7465d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path1->slots[0]); 7475d4f98a2SYan Zheng if (key.objectid != cur->bytenr) { 7485d4f98a2SYan Zheng WARN_ON(exist); 7495d4f98a2SYan Zheng break; 7505d4f98a2SYan Zheng } 7515d4f98a2SYan Zheng 7525d4f98a2SYan Zheng if (key.type == BTRFS_EXTENT_ITEM_KEY) { 7535d4f98a2SYan Zheng ret = find_inline_backref(eb, path1->slots[0], 7545d4f98a2SYan Zheng &ptr, &end); 7555d4f98a2SYan Zheng if (ret) 7565d4f98a2SYan Zheng goto next; 7575d4f98a2SYan Zheng } 7585d4f98a2SYan Zheng } 7595d4f98a2SYan Zheng 7605d4f98a2SYan Zheng if (ptr < end) { 7615d4f98a2SYan Zheng /* update key for inline back ref */ 7625d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 7635d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 7645d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 7655d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 7665d4f98a2SYan Zheng WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY && 7675d4f98a2SYan Zheng key.type != BTRFS_SHARED_BLOCK_REF_KEY); 7685d4f98a2SYan Zheng } 7695d4f98a2SYan Zheng 7705d4f98a2SYan Zheng if (exist && 7715d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 7725d4f98a2SYan Zheng exist->owner == key.offset) || 7735d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 7745d4f98a2SYan Zheng exist->bytenr == key.offset))) { 7755d4f98a2SYan Zheng exist = NULL; 7765d4f98a2SYan Zheng goto next; 7775d4f98a2SYan Zheng } 7785d4f98a2SYan Zheng 7795d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 7805d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY || 7815d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 7823fd0a558SYan, Zheng if (key.type == BTRFS_EXTENT_REF_V0_KEY) { 7835d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 7845d4f98a2SYan Zheng ref0 = btrfs_item_ptr(eb, path1->slots[0], 7855d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 7863fd0a558SYan, Zheng if (key.objectid == key.offset) { 787046f264fSYan, Zheng root = find_tree_root(rc, eb, ref0); 7883fd0a558SYan, Zheng if (root && !should_ignore_root(root)) 7895d4f98a2SYan Zheng cur->root = root; 7905d4f98a2SYan Zheng else 7913fd0a558SYan, Zheng list_add(&cur->list, &useless); 7925d4f98a2SYan Zheng break; 7935d4f98a2SYan Zheng } 794046f264fSYan, Zheng if (is_cowonly_root(btrfs_ref_root_v0(eb, 795046f264fSYan, Zheng ref0))) 796046f264fSYan, Zheng cur->cowonly = 1; 7973fd0a558SYan, Zheng } 7985d4f98a2SYan Zheng #else 7995d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 8005d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 8015d4f98a2SYan Zheng #endif 8025d4f98a2SYan Zheng if (key.objectid == key.offset) { 8035d4f98a2SYan Zheng /* 8045d4f98a2SYan Zheng * only root blocks of reloc trees use 8055d4f98a2SYan Zheng * backref of this type. 8065d4f98a2SYan Zheng */ 8075d4f98a2SYan Zheng root = find_reloc_root(rc, cur->bytenr); 8085d4f98a2SYan Zheng BUG_ON(!root); 8095d4f98a2SYan Zheng cur->root = root; 8105d4f98a2SYan Zheng break; 8115d4f98a2SYan Zheng } 8125d4f98a2SYan Zheng 8133fd0a558SYan, Zheng edge = alloc_backref_edge(cache); 8145d4f98a2SYan Zheng if (!edge) { 8155d4f98a2SYan Zheng err = -ENOMEM; 8165d4f98a2SYan Zheng goto out; 8175d4f98a2SYan Zheng } 8185d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, key.offset); 8195d4f98a2SYan Zheng if (!rb_node) { 8203fd0a558SYan, Zheng upper = alloc_backref_node(cache); 8215d4f98a2SYan Zheng if (!upper) { 8223fd0a558SYan, Zheng free_backref_edge(cache, edge); 8235d4f98a2SYan Zheng err = -ENOMEM; 8245d4f98a2SYan Zheng goto out; 8255d4f98a2SYan Zheng } 8265d4f98a2SYan Zheng upper->bytenr = key.offset; 8275d4f98a2SYan Zheng upper->level = cur->level + 1; 8285d4f98a2SYan Zheng /* 8295d4f98a2SYan Zheng * backrefs for the upper level block isn't 8305d4f98a2SYan Zheng * cached, add the block to pending list 8315d4f98a2SYan Zheng */ 8325d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 8335d4f98a2SYan Zheng } else { 8345d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 8355d4f98a2SYan Zheng rb_node); 8363fd0a558SYan, Zheng BUG_ON(!upper->checked); 8375d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 8385d4f98a2SYan Zheng } 8393fd0a558SYan, Zheng list_add_tail(&edge->list[LOWER], &cur->upper); 8405d4f98a2SYan Zheng edge->node[LOWER] = cur; 8413fd0a558SYan, Zheng edge->node[UPPER] = upper; 8425d4f98a2SYan Zheng 8435d4f98a2SYan Zheng goto next; 8445d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 8455d4f98a2SYan Zheng goto next; 8465d4f98a2SYan Zheng } 8475d4f98a2SYan Zheng 8485d4f98a2SYan Zheng /* key.type == BTRFS_TREE_BLOCK_REF_KEY */ 8495d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, key.offset); 8505d4f98a2SYan Zheng if (IS_ERR(root)) { 8515d4f98a2SYan Zheng err = PTR_ERR(root); 8525d4f98a2SYan Zheng goto out; 8535d4f98a2SYan Zheng } 8545d4f98a2SYan Zheng 8553fd0a558SYan, Zheng if (!root->ref_cows) 8563fd0a558SYan, Zheng cur->cowonly = 1; 8573fd0a558SYan, Zheng 8585d4f98a2SYan Zheng if (btrfs_root_level(&root->root_item) == cur->level) { 8595d4f98a2SYan Zheng /* tree root */ 8605d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 8615d4f98a2SYan Zheng cur->bytenr); 8623fd0a558SYan, Zheng if (should_ignore_root(root)) 8633fd0a558SYan, Zheng list_add(&cur->list, &useless); 8643fd0a558SYan, Zheng else 8655d4f98a2SYan Zheng cur->root = root; 8665d4f98a2SYan Zheng break; 8675d4f98a2SYan Zheng } 8685d4f98a2SYan Zheng 8695d4f98a2SYan Zheng level = cur->level + 1; 8705d4f98a2SYan Zheng 8715d4f98a2SYan Zheng /* 8725d4f98a2SYan Zheng * searching the tree to find upper level blocks 8735d4f98a2SYan Zheng * reference the block. 8745d4f98a2SYan Zheng */ 8755d4f98a2SYan Zheng path2->search_commit_root = 1; 8765d4f98a2SYan Zheng path2->skip_locking = 1; 8775d4f98a2SYan Zheng path2->lowest_level = level; 8785d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0); 8795d4f98a2SYan Zheng path2->lowest_level = 0; 8805d4f98a2SYan Zheng if (ret < 0) { 8815d4f98a2SYan Zheng err = ret; 8825d4f98a2SYan Zheng goto out; 8835d4f98a2SYan Zheng } 88433c66f43SYan Zheng if (ret > 0 && path2->slots[level] > 0) 88533c66f43SYan Zheng path2->slots[level]--; 8865d4f98a2SYan Zheng 8875d4f98a2SYan Zheng eb = path2->nodes[level]; 8885d4f98a2SYan Zheng WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) != 8895d4f98a2SYan Zheng cur->bytenr); 8905d4f98a2SYan Zheng 8915d4f98a2SYan Zheng lower = cur; 8925d4f98a2SYan Zheng for (; level < BTRFS_MAX_LEVEL; level++) { 8935d4f98a2SYan Zheng if (!path2->nodes[level]) { 8945d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 8955d4f98a2SYan Zheng lower->bytenr); 8963fd0a558SYan, Zheng if (should_ignore_root(root)) 8973fd0a558SYan, Zheng list_add(&lower->list, &useless); 8983fd0a558SYan, Zheng else 8995d4f98a2SYan Zheng lower->root = root; 9005d4f98a2SYan Zheng break; 9015d4f98a2SYan Zheng } 9025d4f98a2SYan Zheng 9033fd0a558SYan, Zheng edge = alloc_backref_edge(cache); 9045d4f98a2SYan Zheng if (!edge) { 9055d4f98a2SYan Zheng err = -ENOMEM; 9065d4f98a2SYan Zheng goto out; 9075d4f98a2SYan Zheng } 9085d4f98a2SYan Zheng 9095d4f98a2SYan Zheng eb = path2->nodes[level]; 9105d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, eb->start); 9115d4f98a2SYan Zheng if (!rb_node) { 9123fd0a558SYan, Zheng upper = alloc_backref_node(cache); 9135d4f98a2SYan Zheng if (!upper) { 9143fd0a558SYan, Zheng free_backref_edge(cache, edge); 9155d4f98a2SYan Zheng err = -ENOMEM; 9165d4f98a2SYan Zheng goto out; 9175d4f98a2SYan Zheng } 9185d4f98a2SYan Zheng upper->bytenr = eb->start; 9195d4f98a2SYan Zheng upper->owner = btrfs_header_owner(eb); 9205d4f98a2SYan Zheng upper->level = lower->level + 1; 9213fd0a558SYan, Zheng if (!root->ref_cows) 9223fd0a558SYan, Zheng upper->cowonly = 1; 9235d4f98a2SYan Zheng 9245d4f98a2SYan Zheng /* 9255d4f98a2SYan Zheng * if we know the block isn't shared 9265d4f98a2SYan Zheng * we can void checking its backrefs. 9275d4f98a2SYan Zheng */ 9285d4f98a2SYan Zheng if (btrfs_block_can_be_shared(root, eb)) 9295d4f98a2SYan Zheng upper->checked = 0; 9305d4f98a2SYan Zheng else 9315d4f98a2SYan Zheng upper->checked = 1; 9325d4f98a2SYan Zheng 9335d4f98a2SYan Zheng /* 9345d4f98a2SYan Zheng * add the block to pending list if we 9355d4f98a2SYan Zheng * need check its backrefs. only block 9365d4f98a2SYan Zheng * at 'cur->level + 1' is added to the 9375d4f98a2SYan Zheng * tail of pending list. this guarantees 9385d4f98a2SYan Zheng * we check backrefs from lower level 9395d4f98a2SYan Zheng * blocks to upper level blocks. 9405d4f98a2SYan Zheng */ 9415d4f98a2SYan Zheng if (!upper->checked && 9425d4f98a2SYan Zheng level == cur->level + 1) { 9435d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], 9445d4f98a2SYan Zheng &list); 9455d4f98a2SYan Zheng } else 9465d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 9475d4f98a2SYan Zheng } else { 9485d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 9495d4f98a2SYan Zheng rb_node); 9505d4f98a2SYan Zheng BUG_ON(!upper->checked); 9515d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 9523fd0a558SYan, Zheng if (!upper->owner) 9533fd0a558SYan, Zheng upper->owner = btrfs_header_owner(eb); 9545d4f98a2SYan Zheng } 9555d4f98a2SYan Zheng list_add_tail(&edge->list[LOWER], &lower->upper); 9565d4f98a2SYan Zheng edge->node[LOWER] = lower; 9573fd0a558SYan, Zheng edge->node[UPPER] = upper; 9585d4f98a2SYan Zheng 9595d4f98a2SYan Zheng if (rb_node) 9605d4f98a2SYan Zheng break; 9615d4f98a2SYan Zheng lower = upper; 9625d4f98a2SYan Zheng upper = NULL; 9635d4f98a2SYan Zheng } 964*b3b4aa74SDavid Sterba btrfs_release_path(path2); 9655d4f98a2SYan Zheng next: 9665d4f98a2SYan Zheng if (ptr < end) { 9675d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 9685d4f98a2SYan Zheng if (ptr >= end) { 9695d4f98a2SYan Zheng WARN_ON(ptr > end); 9705d4f98a2SYan Zheng ptr = 0; 9715d4f98a2SYan Zheng end = 0; 9725d4f98a2SYan Zheng } 9735d4f98a2SYan Zheng } 9745d4f98a2SYan Zheng if (ptr >= end) 9755d4f98a2SYan Zheng path1->slots[0]++; 9765d4f98a2SYan Zheng } 977*b3b4aa74SDavid Sterba btrfs_release_path(path1); 9785d4f98a2SYan Zheng 9795d4f98a2SYan Zheng cur->checked = 1; 9805d4f98a2SYan Zheng WARN_ON(exist); 9815d4f98a2SYan Zheng 9825d4f98a2SYan Zheng /* the pending list isn't empty, take the first block to process */ 9835d4f98a2SYan Zheng if (!list_empty(&list)) { 9845d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 9855d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 9865d4f98a2SYan Zheng cur = edge->node[UPPER]; 9875d4f98a2SYan Zheng goto again; 9885d4f98a2SYan Zheng } 9895d4f98a2SYan Zheng 9905d4f98a2SYan Zheng /* 9915d4f98a2SYan Zheng * everything goes well, connect backref nodes and insert backref nodes 9925d4f98a2SYan Zheng * into the cache. 9935d4f98a2SYan Zheng */ 9945d4f98a2SYan Zheng BUG_ON(!node->checked); 9953fd0a558SYan, Zheng cowonly = node->cowonly; 9963fd0a558SYan, Zheng if (!cowonly) { 9973fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, 9983fd0a558SYan, Zheng &node->rb_node); 9995d4f98a2SYan Zheng BUG_ON(rb_node); 10003fd0a558SYan, Zheng list_add_tail(&node->lower, &cache->leaves); 10013fd0a558SYan, Zheng } 10025d4f98a2SYan Zheng 10035d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) 10045d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 10055d4f98a2SYan Zheng 10065d4f98a2SYan Zheng while (!list_empty(&list)) { 10075d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 10085d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 10095d4f98a2SYan Zheng upper = edge->node[UPPER]; 10103fd0a558SYan, Zheng if (upper->detached) { 10113fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 10123fd0a558SYan, Zheng lower = edge->node[LOWER]; 10133fd0a558SYan, Zheng free_backref_edge(cache, edge); 10143fd0a558SYan, Zheng if (list_empty(&lower->upper)) 10153fd0a558SYan, Zheng list_add(&lower->list, &useless); 10163fd0a558SYan, Zheng continue; 10173fd0a558SYan, Zheng } 10185d4f98a2SYan Zheng 10195d4f98a2SYan Zheng if (!RB_EMPTY_NODE(&upper->rb_node)) { 10205d4f98a2SYan Zheng if (upper->lowest) { 10215d4f98a2SYan Zheng list_del_init(&upper->lower); 10225d4f98a2SYan Zheng upper->lowest = 0; 10235d4f98a2SYan Zheng } 10245d4f98a2SYan Zheng 10255d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 10265d4f98a2SYan Zheng continue; 10275d4f98a2SYan Zheng } 10285d4f98a2SYan Zheng 10295d4f98a2SYan Zheng BUG_ON(!upper->checked); 10303fd0a558SYan, Zheng BUG_ON(cowonly != upper->cowonly); 10313fd0a558SYan, Zheng if (!cowonly) { 10325d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, upper->bytenr, 10335d4f98a2SYan Zheng &upper->rb_node); 10345d4f98a2SYan Zheng BUG_ON(rb_node); 10353fd0a558SYan, Zheng } 10365d4f98a2SYan Zheng 10375d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 10385d4f98a2SYan Zheng 10395d4f98a2SYan Zheng list_for_each_entry(edge, &upper->upper, list[LOWER]) 10405d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 10415d4f98a2SYan Zheng } 10423fd0a558SYan, Zheng /* 10433fd0a558SYan, Zheng * process useless backref nodes. backref nodes for tree leaves 10443fd0a558SYan, Zheng * are deleted from the cache. backref nodes for upper level 10453fd0a558SYan, Zheng * tree blocks are left in the cache to avoid unnecessary backref 10463fd0a558SYan, Zheng * lookup. 10473fd0a558SYan, Zheng */ 10483fd0a558SYan, Zheng while (!list_empty(&useless)) { 10493fd0a558SYan, Zheng upper = list_entry(useless.next, struct backref_node, list); 10503fd0a558SYan, Zheng list_del_init(&upper->list); 10513fd0a558SYan, Zheng BUG_ON(!list_empty(&upper->upper)); 10523fd0a558SYan, Zheng if (upper == node) 10533fd0a558SYan, Zheng node = NULL; 10543fd0a558SYan, Zheng if (upper->lowest) { 10553fd0a558SYan, Zheng list_del_init(&upper->lower); 10563fd0a558SYan, Zheng upper->lowest = 0; 10573fd0a558SYan, Zheng } 10583fd0a558SYan, Zheng while (!list_empty(&upper->lower)) { 10593fd0a558SYan, Zheng edge = list_entry(upper->lower.next, 10603fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 10613fd0a558SYan, Zheng list_del(&edge->list[UPPER]); 10623fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 10633fd0a558SYan, Zheng lower = edge->node[LOWER]; 10643fd0a558SYan, Zheng free_backref_edge(cache, edge); 10653fd0a558SYan, Zheng 10663fd0a558SYan, Zheng if (list_empty(&lower->upper)) 10673fd0a558SYan, Zheng list_add(&lower->list, &useless); 10683fd0a558SYan, Zheng } 10693fd0a558SYan, Zheng __mark_block_processed(rc, upper); 10703fd0a558SYan, Zheng if (upper->level > 0) { 10713fd0a558SYan, Zheng list_add(&upper->list, &cache->detached); 10723fd0a558SYan, Zheng upper->detached = 1; 10733fd0a558SYan, Zheng } else { 10743fd0a558SYan, Zheng rb_erase(&upper->rb_node, &cache->rb_root); 10753fd0a558SYan, Zheng free_backref_node(cache, upper); 10763fd0a558SYan, Zheng } 10773fd0a558SYan, Zheng } 10785d4f98a2SYan Zheng out: 10795d4f98a2SYan Zheng btrfs_free_path(path1); 10805d4f98a2SYan Zheng btrfs_free_path(path2); 10815d4f98a2SYan Zheng if (err) { 10823fd0a558SYan, Zheng while (!list_empty(&useless)) { 10833fd0a558SYan, Zheng lower = list_entry(useless.next, 10843fd0a558SYan, Zheng struct backref_node, upper); 10853fd0a558SYan, Zheng list_del_init(&lower->upper); 10863fd0a558SYan, Zheng } 10875d4f98a2SYan Zheng upper = node; 10883fd0a558SYan, Zheng INIT_LIST_HEAD(&list); 10895d4f98a2SYan Zheng while (upper) { 10905d4f98a2SYan Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 10915d4f98a2SYan Zheng list_splice_tail(&upper->upper, &list); 10923fd0a558SYan, Zheng free_backref_node(cache, upper); 10935d4f98a2SYan Zheng } 10945d4f98a2SYan Zheng 10955d4f98a2SYan Zheng if (list_empty(&list)) 10965d4f98a2SYan Zheng break; 10975d4f98a2SYan Zheng 10985d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, 10995d4f98a2SYan Zheng list[LOWER]); 11003fd0a558SYan, Zheng list_del(&edge->list[LOWER]); 11015d4f98a2SYan Zheng upper = edge->node[UPPER]; 11023fd0a558SYan, Zheng free_backref_edge(cache, edge); 11035d4f98a2SYan Zheng } 11045d4f98a2SYan Zheng return ERR_PTR(err); 11055d4f98a2SYan Zheng } 11063fd0a558SYan, Zheng BUG_ON(node && node->detached); 11075d4f98a2SYan Zheng return node; 11085d4f98a2SYan Zheng } 11095d4f98a2SYan Zheng 11105d4f98a2SYan Zheng /* 11113fd0a558SYan, Zheng * helper to add backref node for the newly created snapshot. 11123fd0a558SYan, Zheng * the backref node is created by cloning backref node that 11133fd0a558SYan, Zheng * corresponds to root of source tree 11143fd0a558SYan, Zheng */ 11153fd0a558SYan, Zheng static int clone_backref_node(struct btrfs_trans_handle *trans, 11163fd0a558SYan, Zheng struct reloc_control *rc, 11173fd0a558SYan, Zheng struct btrfs_root *src, 11183fd0a558SYan, Zheng struct btrfs_root *dest) 11193fd0a558SYan, Zheng { 11203fd0a558SYan, Zheng struct btrfs_root *reloc_root = src->reloc_root; 11213fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 11223fd0a558SYan, Zheng struct backref_node *node = NULL; 11233fd0a558SYan, Zheng struct backref_node *new_node; 11243fd0a558SYan, Zheng struct backref_edge *edge; 11253fd0a558SYan, Zheng struct backref_edge *new_edge; 11263fd0a558SYan, Zheng struct rb_node *rb_node; 11273fd0a558SYan, Zheng 11283fd0a558SYan, Zheng if (cache->last_trans > 0) 11293fd0a558SYan, Zheng update_backref_cache(trans, cache); 11303fd0a558SYan, Zheng 11313fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, src->commit_root->start); 11323fd0a558SYan, Zheng if (rb_node) { 11333fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, rb_node); 11343fd0a558SYan, Zheng if (node->detached) 11353fd0a558SYan, Zheng node = NULL; 11363fd0a558SYan, Zheng else 11373fd0a558SYan, Zheng BUG_ON(node->new_bytenr != reloc_root->node->start); 11383fd0a558SYan, Zheng } 11393fd0a558SYan, Zheng 11403fd0a558SYan, Zheng if (!node) { 11413fd0a558SYan, Zheng rb_node = tree_search(&cache->rb_root, 11423fd0a558SYan, Zheng reloc_root->commit_root->start); 11433fd0a558SYan, Zheng if (rb_node) { 11443fd0a558SYan, Zheng node = rb_entry(rb_node, struct backref_node, 11453fd0a558SYan, Zheng rb_node); 11463fd0a558SYan, Zheng BUG_ON(node->detached); 11473fd0a558SYan, Zheng } 11483fd0a558SYan, Zheng } 11493fd0a558SYan, Zheng 11503fd0a558SYan, Zheng if (!node) 11513fd0a558SYan, Zheng return 0; 11523fd0a558SYan, Zheng 11533fd0a558SYan, Zheng new_node = alloc_backref_node(cache); 11543fd0a558SYan, Zheng if (!new_node) 11553fd0a558SYan, Zheng return -ENOMEM; 11563fd0a558SYan, Zheng 11573fd0a558SYan, Zheng new_node->bytenr = dest->node->start; 11583fd0a558SYan, Zheng new_node->level = node->level; 11593fd0a558SYan, Zheng new_node->lowest = node->lowest; 11606848ad64SYan, Zheng new_node->checked = 1; 11613fd0a558SYan, Zheng new_node->root = dest; 11623fd0a558SYan, Zheng 11633fd0a558SYan, Zheng if (!node->lowest) { 11643fd0a558SYan, Zheng list_for_each_entry(edge, &node->lower, list[UPPER]) { 11653fd0a558SYan, Zheng new_edge = alloc_backref_edge(cache); 11663fd0a558SYan, Zheng if (!new_edge) 11673fd0a558SYan, Zheng goto fail; 11683fd0a558SYan, Zheng 11693fd0a558SYan, Zheng new_edge->node[UPPER] = new_node; 11703fd0a558SYan, Zheng new_edge->node[LOWER] = edge->node[LOWER]; 11713fd0a558SYan, Zheng list_add_tail(&new_edge->list[UPPER], 11723fd0a558SYan, Zheng &new_node->lower); 11733fd0a558SYan, Zheng } 11743fd0a558SYan, Zheng } 11753fd0a558SYan, Zheng 11763fd0a558SYan, Zheng rb_node = tree_insert(&cache->rb_root, new_node->bytenr, 11773fd0a558SYan, Zheng &new_node->rb_node); 11783fd0a558SYan, Zheng BUG_ON(rb_node); 11793fd0a558SYan, Zheng 11803fd0a558SYan, Zheng if (!new_node->lowest) { 11813fd0a558SYan, Zheng list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) { 11823fd0a558SYan, Zheng list_add_tail(&new_edge->list[LOWER], 11833fd0a558SYan, Zheng &new_edge->node[LOWER]->upper); 11843fd0a558SYan, Zheng } 11853fd0a558SYan, Zheng } 11863fd0a558SYan, Zheng return 0; 11873fd0a558SYan, Zheng fail: 11883fd0a558SYan, Zheng while (!list_empty(&new_node->lower)) { 11893fd0a558SYan, Zheng new_edge = list_entry(new_node->lower.next, 11903fd0a558SYan, Zheng struct backref_edge, list[UPPER]); 11913fd0a558SYan, Zheng list_del(&new_edge->list[UPPER]); 11923fd0a558SYan, Zheng free_backref_edge(cache, new_edge); 11933fd0a558SYan, Zheng } 11943fd0a558SYan, Zheng free_backref_node(cache, new_node); 11953fd0a558SYan, Zheng return -ENOMEM; 11963fd0a558SYan, Zheng } 11973fd0a558SYan, Zheng 11983fd0a558SYan, Zheng /* 11995d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 12005d4f98a2SYan Zheng */ 12015d4f98a2SYan Zheng static int __add_reloc_root(struct btrfs_root *root) 12025d4f98a2SYan Zheng { 12035d4f98a2SYan Zheng struct rb_node *rb_node; 12045d4f98a2SYan Zheng struct mapping_node *node; 12055d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 12065d4f98a2SYan Zheng 12075d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 12085d4f98a2SYan Zheng BUG_ON(!node); 12095d4f98a2SYan Zheng 12105d4f98a2SYan Zheng node->bytenr = root->node->start; 12115d4f98a2SYan Zheng node->data = root; 12125d4f98a2SYan Zheng 12135d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12145d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 12155d4f98a2SYan Zheng node->bytenr, &node->rb_node); 12165d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 12175d4f98a2SYan Zheng BUG_ON(rb_node); 12185d4f98a2SYan Zheng 12195d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 12205d4f98a2SYan Zheng return 0; 12215d4f98a2SYan Zheng } 12225d4f98a2SYan Zheng 12235d4f98a2SYan Zheng /* 12245d4f98a2SYan Zheng * helper to update/delete the 'address of tree root -> reloc tree' 12255d4f98a2SYan Zheng * mapping 12265d4f98a2SYan Zheng */ 12275d4f98a2SYan Zheng static int __update_reloc_root(struct btrfs_root *root, int del) 12285d4f98a2SYan Zheng { 12295d4f98a2SYan Zheng struct rb_node *rb_node; 12305d4f98a2SYan Zheng struct mapping_node *node = NULL; 12315d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 12325d4f98a2SYan Zheng 12335d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12345d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, 12355d4f98a2SYan Zheng root->commit_root->start); 12365d4f98a2SYan Zheng if (rb_node) { 12375d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 12385d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 12395d4f98a2SYan Zheng } 12405d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 12415d4f98a2SYan Zheng 12425d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 12435d4f98a2SYan Zheng 12445d4f98a2SYan Zheng if (!del) { 12455d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 12465d4f98a2SYan Zheng node->bytenr = root->node->start; 12475d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 12485d4f98a2SYan Zheng node->bytenr, &node->rb_node); 12495d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 12505d4f98a2SYan Zheng BUG_ON(rb_node); 12515d4f98a2SYan Zheng } else { 12525d4f98a2SYan Zheng list_del_init(&root->root_list); 12535d4f98a2SYan Zheng kfree(node); 12545d4f98a2SYan Zheng } 12555d4f98a2SYan Zheng return 0; 12565d4f98a2SYan Zheng } 12575d4f98a2SYan Zheng 12583fd0a558SYan, Zheng static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans, 12593fd0a558SYan, Zheng struct btrfs_root *root, u64 objectid) 12605d4f98a2SYan Zheng { 12615d4f98a2SYan Zheng struct btrfs_root *reloc_root; 12625d4f98a2SYan Zheng struct extent_buffer *eb; 12635d4f98a2SYan Zheng struct btrfs_root_item *root_item; 12645d4f98a2SYan Zheng struct btrfs_key root_key; 12655d4f98a2SYan Zheng int ret; 12665d4f98a2SYan Zheng 12675d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 12685d4f98a2SYan Zheng BUG_ON(!root_item); 12695d4f98a2SYan Zheng 12705d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 12715d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 12723fd0a558SYan, Zheng root_key.offset = objectid; 12735d4f98a2SYan Zheng 12743fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 12753fd0a558SYan, Zheng /* called by btrfs_init_reloc_root */ 12765d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 12775d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 12785d4f98a2SYan Zheng BUG_ON(ret); 12795d4f98a2SYan Zheng 12803fd0a558SYan, Zheng btrfs_set_root_last_snapshot(&root->root_item, 12813fd0a558SYan, Zheng trans->transid - 1); 12823fd0a558SYan, Zheng } else { 12833fd0a558SYan, Zheng /* 12843fd0a558SYan, Zheng * called by btrfs_reloc_post_snapshot_hook. 12853fd0a558SYan, Zheng * the source tree is a reloc tree, all tree blocks 12863fd0a558SYan, Zheng * modified after it was created have RELOC flag 12873fd0a558SYan, Zheng * set in their headers. so it's OK to not update 12883fd0a558SYan, Zheng * the 'last_snapshot'. 12893fd0a558SYan, Zheng */ 12903fd0a558SYan, Zheng ret = btrfs_copy_root(trans, root, root->node, &eb, 12913fd0a558SYan, Zheng BTRFS_TREE_RELOC_OBJECTID); 12923fd0a558SYan, Zheng BUG_ON(ret); 12933fd0a558SYan, Zheng } 12943fd0a558SYan, Zheng 12955d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 12965d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 12975d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 12985d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 12993fd0a558SYan, Zheng 13003fd0a558SYan, Zheng if (root->root_key.objectid == objectid) { 13013fd0a558SYan, Zheng btrfs_set_root_refs(root_item, 0); 13023fd0a558SYan, Zheng memset(&root_item->drop_progress, 0, 13033fd0a558SYan, Zheng sizeof(struct btrfs_disk_key)); 13045d4f98a2SYan Zheng root_item->drop_level = 0; 13053fd0a558SYan, Zheng } 13065d4f98a2SYan Zheng 13075d4f98a2SYan Zheng btrfs_tree_unlock(eb); 13085d4f98a2SYan Zheng free_extent_buffer(eb); 13095d4f98a2SYan Zheng 13105d4f98a2SYan Zheng ret = btrfs_insert_root(trans, root->fs_info->tree_root, 13115d4f98a2SYan Zheng &root_key, root_item); 13125d4f98a2SYan Zheng BUG_ON(ret); 13135d4f98a2SYan Zheng kfree(root_item); 13145d4f98a2SYan Zheng 13155d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root, 13165d4f98a2SYan Zheng &root_key); 13175d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 13185d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 13193fd0a558SYan, Zheng return reloc_root; 13203fd0a558SYan, Zheng } 13213fd0a558SYan, Zheng 13223fd0a558SYan, Zheng /* 13233fd0a558SYan, Zheng * create reloc tree for a given fs tree. reloc tree is just a 13243fd0a558SYan, Zheng * snapshot of the fs tree with special root objectid. 13253fd0a558SYan, Zheng */ 13263fd0a558SYan, Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 13273fd0a558SYan, Zheng struct btrfs_root *root) 13283fd0a558SYan, Zheng { 13293fd0a558SYan, Zheng struct btrfs_root *reloc_root; 13303fd0a558SYan, Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 13313fd0a558SYan, Zheng int clear_rsv = 0; 13323fd0a558SYan, Zheng 13333fd0a558SYan, Zheng if (root->reloc_root) { 13343fd0a558SYan, Zheng reloc_root = root->reloc_root; 13353fd0a558SYan, Zheng reloc_root->last_trans = trans->transid; 13363fd0a558SYan, Zheng return 0; 13373fd0a558SYan, Zheng } 13383fd0a558SYan, Zheng 13393fd0a558SYan, Zheng if (!rc || !rc->create_reloc_tree || 13403fd0a558SYan, Zheng root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 13413fd0a558SYan, Zheng return 0; 13423fd0a558SYan, Zheng 13433fd0a558SYan, Zheng if (!trans->block_rsv) { 13443fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 13453fd0a558SYan, Zheng clear_rsv = 1; 13463fd0a558SYan, Zheng } 13473fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root, root->root_key.objectid); 13483fd0a558SYan, Zheng if (clear_rsv) 13493fd0a558SYan, Zheng trans->block_rsv = NULL; 13505d4f98a2SYan Zheng 13515d4f98a2SYan Zheng __add_reloc_root(reloc_root); 13525d4f98a2SYan Zheng root->reloc_root = reloc_root; 13535d4f98a2SYan Zheng return 0; 13545d4f98a2SYan Zheng } 13555d4f98a2SYan Zheng 13565d4f98a2SYan Zheng /* 13575d4f98a2SYan Zheng * update root item of reloc tree 13585d4f98a2SYan Zheng */ 13595d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 13605d4f98a2SYan Zheng struct btrfs_root *root) 13615d4f98a2SYan Zheng { 13625d4f98a2SYan Zheng struct btrfs_root *reloc_root; 13635d4f98a2SYan Zheng struct btrfs_root_item *root_item; 13645d4f98a2SYan Zheng int del = 0; 13655d4f98a2SYan Zheng int ret; 13665d4f98a2SYan Zheng 13675d4f98a2SYan Zheng if (!root->reloc_root) 13685d4f98a2SYan Zheng return 0; 13695d4f98a2SYan Zheng 13705d4f98a2SYan Zheng reloc_root = root->reloc_root; 13715d4f98a2SYan Zheng root_item = &reloc_root->root_item; 13725d4f98a2SYan Zheng 13733fd0a558SYan, Zheng if (root->fs_info->reloc_ctl->merge_reloc_tree && 13743fd0a558SYan, Zheng btrfs_root_refs(root_item) == 0) { 13755d4f98a2SYan Zheng root->reloc_root = NULL; 13765d4f98a2SYan Zheng del = 1; 13775d4f98a2SYan Zheng } 13785d4f98a2SYan Zheng 13795d4f98a2SYan Zheng __update_reloc_root(reloc_root, del); 13805d4f98a2SYan Zheng 13815d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 13825d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 13835d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 13845d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 13855d4f98a2SYan Zheng } 13865d4f98a2SYan Zheng 13875d4f98a2SYan Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 13885d4f98a2SYan Zheng &reloc_root->root_key, root_item); 13895d4f98a2SYan Zheng BUG_ON(ret); 13905d4f98a2SYan Zheng return 0; 13915d4f98a2SYan Zheng } 13925d4f98a2SYan Zheng 13935d4f98a2SYan Zheng /* 13945d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 13955d4f98a2SYan Zheng * in a subvolume 13965d4f98a2SYan Zheng */ 13975d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 13985d4f98a2SYan Zheng { 13995d4f98a2SYan Zheng struct rb_node *node; 14005d4f98a2SYan Zheng struct rb_node *prev; 14015d4f98a2SYan Zheng struct btrfs_inode *entry; 14025d4f98a2SYan Zheng struct inode *inode; 14035d4f98a2SYan Zheng 14045d4f98a2SYan Zheng spin_lock(&root->inode_lock); 14055d4f98a2SYan Zheng again: 14065d4f98a2SYan Zheng node = root->inode_tree.rb_node; 14075d4f98a2SYan Zheng prev = NULL; 14085d4f98a2SYan Zheng while (node) { 14095d4f98a2SYan Zheng prev = node; 14105d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 14115d4f98a2SYan Zheng 14125d4f98a2SYan Zheng if (objectid < entry->vfs_inode.i_ino) 14135d4f98a2SYan Zheng node = node->rb_left; 14145d4f98a2SYan Zheng else if (objectid > entry->vfs_inode.i_ino) 14155d4f98a2SYan Zheng node = node->rb_right; 14165d4f98a2SYan Zheng else 14175d4f98a2SYan Zheng break; 14185d4f98a2SYan Zheng } 14195d4f98a2SYan Zheng if (!node) { 14205d4f98a2SYan Zheng while (prev) { 14215d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 14225d4f98a2SYan Zheng if (objectid <= entry->vfs_inode.i_ino) { 14235d4f98a2SYan Zheng node = prev; 14245d4f98a2SYan Zheng break; 14255d4f98a2SYan Zheng } 14265d4f98a2SYan Zheng prev = rb_next(prev); 14275d4f98a2SYan Zheng } 14285d4f98a2SYan Zheng } 14295d4f98a2SYan Zheng while (node) { 14305d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 14315d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 14325d4f98a2SYan Zheng if (inode) { 14335d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 14345d4f98a2SYan Zheng return inode; 14355d4f98a2SYan Zheng } 14365d4f98a2SYan Zheng 14375d4f98a2SYan Zheng objectid = entry->vfs_inode.i_ino + 1; 14385d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 14395d4f98a2SYan Zheng goto again; 14405d4f98a2SYan Zheng 14415d4f98a2SYan Zheng node = rb_next(node); 14425d4f98a2SYan Zheng } 14435d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 14445d4f98a2SYan Zheng return NULL; 14455d4f98a2SYan Zheng } 14465d4f98a2SYan Zheng 14475d4f98a2SYan Zheng static int in_block_group(u64 bytenr, 14485d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group) 14495d4f98a2SYan Zheng { 14505d4f98a2SYan Zheng if (bytenr >= block_group->key.objectid && 14515d4f98a2SYan Zheng bytenr < block_group->key.objectid + block_group->key.offset) 14525d4f98a2SYan Zheng return 1; 14535d4f98a2SYan Zheng return 0; 14545d4f98a2SYan Zheng } 14555d4f98a2SYan Zheng 14565d4f98a2SYan Zheng /* 14575d4f98a2SYan Zheng * get new location of data 14585d4f98a2SYan Zheng */ 14595d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 14605d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 14615d4f98a2SYan Zheng { 14625d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 14635d4f98a2SYan Zheng struct btrfs_path *path; 14645d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 14655d4f98a2SYan Zheng struct extent_buffer *leaf; 14665d4f98a2SYan Zheng int ret; 14675d4f98a2SYan Zheng 14685d4f98a2SYan Zheng path = btrfs_alloc_path(); 14695d4f98a2SYan Zheng if (!path) 14705d4f98a2SYan Zheng return -ENOMEM; 14715d4f98a2SYan Zheng 14725d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 14735d4f98a2SYan Zheng ret = btrfs_lookup_file_extent(NULL, root, path, reloc_inode->i_ino, 14745d4f98a2SYan Zheng bytenr, 0); 14755d4f98a2SYan Zheng if (ret < 0) 14765d4f98a2SYan Zheng goto out; 14775d4f98a2SYan Zheng if (ret > 0) { 14785d4f98a2SYan Zheng ret = -ENOENT; 14795d4f98a2SYan Zheng goto out; 14805d4f98a2SYan Zheng } 14815d4f98a2SYan Zheng 14825d4f98a2SYan Zheng leaf = path->nodes[0]; 14835d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 14845d4f98a2SYan Zheng struct btrfs_file_extent_item); 14855d4f98a2SYan Zheng 14865d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 14875d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 14885d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 14895d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 14905d4f98a2SYan Zheng 14915d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 14925d4f98a2SYan Zheng ret = 1; 14935d4f98a2SYan Zheng goto out; 14945d4f98a2SYan Zheng } 14955d4f98a2SYan Zheng 14965d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 14975d4f98a2SYan Zheng ret = 0; 14985d4f98a2SYan Zheng out: 14995d4f98a2SYan Zheng btrfs_free_path(path); 15005d4f98a2SYan Zheng return ret; 15015d4f98a2SYan Zheng } 15025d4f98a2SYan Zheng 15035d4f98a2SYan Zheng /* 15045d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 15055d4f98a2SYan Zheng * the new locations. 15065d4f98a2SYan Zheng */ 15073fd0a558SYan, Zheng static noinline_for_stack 15083fd0a558SYan, Zheng int replace_file_extents(struct btrfs_trans_handle *trans, 15095d4f98a2SYan Zheng struct reloc_control *rc, 15105d4f98a2SYan Zheng struct btrfs_root *root, 15113fd0a558SYan, Zheng struct extent_buffer *leaf) 15125d4f98a2SYan Zheng { 15135d4f98a2SYan Zheng struct btrfs_key key; 15145d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 15155d4f98a2SYan Zheng struct inode *inode = NULL; 15165d4f98a2SYan Zheng u64 parent; 15175d4f98a2SYan Zheng u64 bytenr; 15183fd0a558SYan, Zheng u64 new_bytenr = 0; 15195d4f98a2SYan Zheng u64 num_bytes; 15205d4f98a2SYan Zheng u64 end; 15215d4f98a2SYan Zheng u32 nritems; 15225d4f98a2SYan Zheng u32 i; 15235d4f98a2SYan Zheng int ret; 15245d4f98a2SYan Zheng int first = 1; 15255d4f98a2SYan Zheng int dirty = 0; 15265d4f98a2SYan Zheng 15275d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 15285d4f98a2SYan Zheng return 0; 15295d4f98a2SYan Zheng 15305d4f98a2SYan Zheng /* reloc trees always use full backref */ 15315d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 15325d4f98a2SYan Zheng parent = leaf->start; 15335d4f98a2SYan Zheng else 15345d4f98a2SYan Zheng parent = 0; 15355d4f98a2SYan Zheng 15365d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 15375d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 15385d4f98a2SYan Zheng cond_resched(); 15395d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 15405d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 15415d4f98a2SYan Zheng continue; 15425d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 15435d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 15445d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 15455d4f98a2SYan Zheng continue; 15465d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 15475d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 15485d4f98a2SYan Zheng if (bytenr == 0) 15495d4f98a2SYan Zheng continue; 15505d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group)) 15515d4f98a2SYan Zheng continue; 15525d4f98a2SYan Zheng 15535d4f98a2SYan Zheng /* 15545d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 15555d4f98a2SYan Zheng * to complete and drop the extent cache 15565d4f98a2SYan Zheng */ 15575d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 15585d4f98a2SYan Zheng if (first) { 15595d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 15605d4f98a2SYan Zheng first = 0; 15615d4f98a2SYan Zheng } else if (inode && inode->i_ino < key.objectid) { 15623fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 15635d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 15645d4f98a2SYan Zheng } 15655d4f98a2SYan Zheng if (inode && inode->i_ino == key.objectid) { 15665d4f98a2SYan Zheng end = key.offset + 15675d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 15685d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 15695d4f98a2SYan Zheng root->sectorsize)); 15705d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 15715d4f98a2SYan Zheng end--; 15725d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 15735d4f98a2SYan Zheng key.offset, end, 15745d4f98a2SYan Zheng GFP_NOFS); 15755d4f98a2SYan Zheng if (!ret) 15765d4f98a2SYan Zheng continue; 15775d4f98a2SYan Zheng 15785d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, key.offset, end, 15795d4f98a2SYan Zheng 1); 15805d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 15815d4f98a2SYan Zheng key.offset, end, GFP_NOFS); 15825d4f98a2SYan Zheng } 15835d4f98a2SYan Zheng } 15845d4f98a2SYan Zheng 15855d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 15865d4f98a2SYan Zheng bytenr, num_bytes); 15873fd0a558SYan, Zheng if (ret > 0) { 15883fd0a558SYan, Zheng WARN_ON(1); 15895d4f98a2SYan Zheng continue; 15903fd0a558SYan, Zheng } 15915d4f98a2SYan Zheng BUG_ON(ret < 0); 15925d4f98a2SYan Zheng 15935d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 15945d4f98a2SYan Zheng dirty = 1; 15955d4f98a2SYan Zheng 15965d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 15975d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, new_bytenr, 15985d4f98a2SYan Zheng num_bytes, parent, 15995d4f98a2SYan Zheng btrfs_header_owner(leaf), 16005d4f98a2SYan Zheng key.objectid, key.offset); 16015d4f98a2SYan Zheng BUG_ON(ret); 16025d4f98a2SYan Zheng 16035d4f98a2SYan Zheng ret = btrfs_free_extent(trans, root, bytenr, num_bytes, 16045d4f98a2SYan Zheng parent, btrfs_header_owner(leaf), 16055d4f98a2SYan Zheng key.objectid, key.offset); 16065d4f98a2SYan Zheng BUG_ON(ret); 16075d4f98a2SYan Zheng } 16085d4f98a2SYan Zheng if (dirty) 16095d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 16103fd0a558SYan, Zheng if (inode) 16113fd0a558SYan, Zheng btrfs_add_delayed_iput(inode); 16125d4f98a2SYan Zheng return 0; 16135d4f98a2SYan Zheng } 16145d4f98a2SYan Zheng 16155d4f98a2SYan Zheng static noinline_for_stack 16165d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 16175d4f98a2SYan Zheng struct btrfs_path *path, int level) 16185d4f98a2SYan Zheng { 16195d4f98a2SYan Zheng struct btrfs_disk_key key1; 16205d4f98a2SYan Zheng struct btrfs_disk_key key2; 16215d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 16225d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 16235d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 16245d4f98a2SYan Zheng } 16255d4f98a2SYan Zheng 16265d4f98a2SYan Zheng /* 16275d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 16285d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 16295d4f98a2SYan Zheng * reloc tree was create can be replaced. 16305d4f98a2SYan Zheng * 16315d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 16325d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 16335d4f98a2SYan Zheng * errors, a negative error number is returned. 16345d4f98a2SYan Zheng */ 16353fd0a558SYan, Zheng static noinline_for_stack 16363fd0a558SYan, Zheng int replace_path(struct btrfs_trans_handle *trans, 16375d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 16385d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 16395d4f98a2SYan Zheng int lowest_level, int max_level) 16405d4f98a2SYan Zheng { 16415d4f98a2SYan Zheng struct extent_buffer *eb; 16425d4f98a2SYan Zheng struct extent_buffer *parent; 16435d4f98a2SYan Zheng struct btrfs_key key; 16445d4f98a2SYan Zheng u64 old_bytenr; 16455d4f98a2SYan Zheng u64 new_bytenr; 16465d4f98a2SYan Zheng u64 old_ptr_gen; 16475d4f98a2SYan Zheng u64 new_ptr_gen; 16485d4f98a2SYan Zheng u64 last_snapshot; 16495d4f98a2SYan Zheng u32 blocksize; 16503fd0a558SYan, Zheng int cow = 0; 16515d4f98a2SYan Zheng int level; 16525d4f98a2SYan Zheng int ret; 16535d4f98a2SYan Zheng int slot; 16545d4f98a2SYan Zheng 16555d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 16565d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 16575d4f98a2SYan Zheng 16585d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 16593fd0a558SYan, Zheng again: 16605d4f98a2SYan Zheng slot = path->slots[lowest_level]; 16615d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 16625d4f98a2SYan Zheng 16635d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 16645d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 16655d4f98a2SYan Zheng level = btrfs_header_level(eb); 16665d4f98a2SYan Zheng 16675d4f98a2SYan Zheng if (level < lowest_level) { 16685d4f98a2SYan Zheng btrfs_tree_unlock(eb); 16695d4f98a2SYan Zheng free_extent_buffer(eb); 16705d4f98a2SYan Zheng return 0; 16715d4f98a2SYan Zheng } 16725d4f98a2SYan Zheng 16733fd0a558SYan, Zheng if (cow) { 16745d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 16755d4f98a2SYan Zheng BUG_ON(ret); 16763fd0a558SYan, Zheng } 16775d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 16785d4f98a2SYan Zheng 16795d4f98a2SYan Zheng if (next_key) { 16805d4f98a2SYan Zheng next_key->objectid = (u64)-1; 16815d4f98a2SYan Zheng next_key->type = (u8)-1; 16825d4f98a2SYan Zheng next_key->offset = (u64)-1; 16835d4f98a2SYan Zheng } 16845d4f98a2SYan Zheng 16855d4f98a2SYan Zheng parent = eb; 16865d4f98a2SYan Zheng while (1) { 16875d4f98a2SYan Zheng level = btrfs_header_level(parent); 16885d4f98a2SYan Zheng BUG_ON(level < lowest_level); 16895d4f98a2SYan Zheng 16905d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 16915d4f98a2SYan Zheng if (ret && slot > 0) 16925d4f98a2SYan Zheng slot--; 16935d4f98a2SYan Zheng 16945d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 16955d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 16965d4f98a2SYan Zheng 16975d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 16985d4f98a2SYan Zheng blocksize = btrfs_level_size(dest, level - 1); 16995d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 17005d4f98a2SYan Zheng 17015d4f98a2SYan Zheng if (level <= max_level) { 17025d4f98a2SYan Zheng eb = path->nodes[level]; 17035d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 17045d4f98a2SYan Zheng path->slots[level]); 17055d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 17065d4f98a2SYan Zheng path->slots[level]); 17075d4f98a2SYan Zheng } else { 17085d4f98a2SYan Zheng new_bytenr = 0; 17095d4f98a2SYan Zheng new_ptr_gen = 0; 17105d4f98a2SYan Zheng } 17115d4f98a2SYan Zheng 17125d4f98a2SYan Zheng if (new_bytenr > 0 && new_bytenr == old_bytenr) { 17135d4f98a2SYan Zheng WARN_ON(1); 17145d4f98a2SYan Zheng ret = level; 17155d4f98a2SYan Zheng break; 17165d4f98a2SYan Zheng } 17175d4f98a2SYan Zheng 17185d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 17195d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 17203fd0a558SYan, Zheng if (level <= lowest_level) { 17215d4f98a2SYan Zheng ret = 0; 17225d4f98a2SYan Zheng break; 17235d4f98a2SYan Zheng } 17245d4f98a2SYan Zheng 17255d4f98a2SYan Zheng eb = read_tree_block(dest, old_bytenr, blocksize, 17265d4f98a2SYan Zheng old_ptr_gen); 172797d9a8a4STsutomu Itoh BUG_ON(!eb); 17285d4f98a2SYan Zheng btrfs_tree_lock(eb); 17293fd0a558SYan, Zheng if (cow) { 17305d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 17315d4f98a2SYan Zheng slot, &eb); 17325d4f98a2SYan Zheng BUG_ON(ret); 17335d4f98a2SYan Zheng } 17343fd0a558SYan, Zheng btrfs_set_lock_blocking(eb); 17355d4f98a2SYan Zheng 17365d4f98a2SYan Zheng btrfs_tree_unlock(parent); 17375d4f98a2SYan Zheng free_extent_buffer(parent); 17385d4f98a2SYan Zheng 17395d4f98a2SYan Zheng parent = eb; 17405d4f98a2SYan Zheng continue; 17415d4f98a2SYan Zheng } 17425d4f98a2SYan Zheng 17433fd0a558SYan, Zheng if (!cow) { 17443fd0a558SYan, Zheng btrfs_tree_unlock(parent); 17453fd0a558SYan, Zheng free_extent_buffer(parent); 17463fd0a558SYan, Zheng cow = 1; 17473fd0a558SYan, Zheng goto again; 17483fd0a558SYan, Zheng } 17493fd0a558SYan, Zheng 17505d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17515d4f98a2SYan Zheng path->slots[level]); 1752*b3b4aa74SDavid Sterba btrfs_release_path(path); 17535d4f98a2SYan Zheng 17545d4f98a2SYan Zheng path->lowest_level = level; 17555d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 17565d4f98a2SYan Zheng path->lowest_level = 0; 17575d4f98a2SYan Zheng BUG_ON(ret); 17585d4f98a2SYan Zheng 17595d4f98a2SYan Zheng /* 17605d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 17615d4f98a2SYan Zheng */ 17625d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 17635d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 17645d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 17655d4f98a2SYan Zheng 17665d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 17675d4f98a2SYan Zheng path->slots[level], old_bytenr); 17685d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 17695d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 17705d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 17715d4f98a2SYan Zheng 17725d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize, 17735d4f98a2SYan Zheng path->nodes[level]->start, 17745d4f98a2SYan Zheng src->root_key.objectid, level - 1, 0); 17755d4f98a2SYan Zheng BUG_ON(ret); 17765d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize, 17775d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 17785d4f98a2SYan Zheng 0); 17795d4f98a2SYan Zheng BUG_ON(ret); 17805d4f98a2SYan Zheng 17815d4f98a2SYan Zheng ret = btrfs_free_extent(trans, src, new_bytenr, blocksize, 17825d4f98a2SYan Zheng path->nodes[level]->start, 17835d4f98a2SYan Zheng src->root_key.objectid, level - 1, 0); 17845d4f98a2SYan Zheng BUG_ON(ret); 17855d4f98a2SYan Zheng 17865d4f98a2SYan Zheng ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize, 17875d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 17885d4f98a2SYan Zheng 0); 17895d4f98a2SYan Zheng BUG_ON(ret); 17905d4f98a2SYan Zheng 17915d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 17925d4f98a2SYan Zheng 17935d4f98a2SYan Zheng ret = level; 17945d4f98a2SYan Zheng break; 17955d4f98a2SYan Zheng } 17965d4f98a2SYan Zheng btrfs_tree_unlock(parent); 17975d4f98a2SYan Zheng free_extent_buffer(parent); 17985d4f98a2SYan Zheng return ret; 17995d4f98a2SYan Zheng } 18005d4f98a2SYan Zheng 18015d4f98a2SYan Zheng /* 18025d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 18035d4f98a2SYan Zheng */ 18045d4f98a2SYan Zheng static noinline_for_stack 18055d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 18065d4f98a2SYan Zheng int *level) 18075d4f98a2SYan Zheng { 18085d4f98a2SYan Zheng struct extent_buffer *eb; 18095d4f98a2SYan Zheng int i; 18105d4f98a2SYan Zheng u64 last_snapshot; 18115d4f98a2SYan Zheng u32 nritems; 18125d4f98a2SYan Zheng 18135d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 18145d4f98a2SYan Zheng 18155d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 18165d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 18175d4f98a2SYan Zheng path->nodes[i] = NULL; 18185d4f98a2SYan Zheng } 18195d4f98a2SYan Zheng 18205d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 18215d4f98a2SYan Zheng eb = path->nodes[i]; 18225d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 18235d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 18245d4f98a2SYan Zheng path->slots[i]++; 18255d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 18265d4f98a2SYan Zheng last_snapshot) 18275d4f98a2SYan Zheng continue; 18285d4f98a2SYan Zheng 18295d4f98a2SYan Zheng *level = i; 18305d4f98a2SYan Zheng return 0; 18315d4f98a2SYan Zheng } 18325d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 18335d4f98a2SYan Zheng path->nodes[i] = NULL; 18345d4f98a2SYan Zheng } 18355d4f98a2SYan Zheng return 1; 18365d4f98a2SYan Zheng } 18375d4f98a2SYan Zheng 18385d4f98a2SYan Zheng /* 18395d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 18405d4f98a2SYan Zheng */ 18415d4f98a2SYan Zheng static noinline_for_stack 18425d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 18435d4f98a2SYan Zheng int *level) 18445d4f98a2SYan Zheng { 18455d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 18465d4f98a2SYan Zheng int i; 18475d4f98a2SYan Zheng u64 bytenr; 18485d4f98a2SYan Zheng u64 ptr_gen = 0; 18495d4f98a2SYan Zheng u64 last_snapshot; 18505d4f98a2SYan Zheng u32 blocksize; 18515d4f98a2SYan Zheng u32 nritems; 18525d4f98a2SYan Zheng 18535d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 18545d4f98a2SYan Zheng 18555d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 18565d4f98a2SYan Zheng eb = path->nodes[i]; 18575d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 18585d4f98a2SYan Zheng while (path->slots[i] < nritems) { 18595d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 18605d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 18615d4f98a2SYan Zheng break; 18625d4f98a2SYan Zheng path->slots[i]++; 18635d4f98a2SYan Zheng } 18645d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 18655d4f98a2SYan Zheng if (i == *level) 18665d4f98a2SYan Zheng break; 18675d4f98a2SYan Zheng *level = i + 1; 18685d4f98a2SYan Zheng return 0; 18695d4f98a2SYan Zheng } 18705d4f98a2SYan Zheng if (i == 1) { 18715d4f98a2SYan Zheng *level = i; 18725d4f98a2SYan Zheng return 0; 18735d4f98a2SYan Zheng } 18745d4f98a2SYan Zheng 18755d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 18765d4f98a2SYan Zheng blocksize = btrfs_level_size(root, i - 1); 18775d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, ptr_gen); 18785d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 18795d4f98a2SYan Zheng path->nodes[i - 1] = eb; 18805d4f98a2SYan Zheng path->slots[i - 1] = 0; 18815d4f98a2SYan Zheng } 18825d4f98a2SYan Zheng return 1; 18835d4f98a2SYan Zheng } 18845d4f98a2SYan Zheng 18855d4f98a2SYan Zheng /* 18865d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 18875d4f98a2SYan Zheng * [min_key, max_key) 18885d4f98a2SYan Zheng */ 18895d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 18905d4f98a2SYan Zheng struct btrfs_key *min_key, 18915d4f98a2SYan Zheng struct btrfs_key *max_key) 18925d4f98a2SYan Zheng { 18935d4f98a2SYan Zheng struct inode *inode = NULL; 18945d4f98a2SYan Zheng u64 objectid; 18955d4f98a2SYan Zheng u64 start, end; 18965d4f98a2SYan Zheng 18975d4f98a2SYan Zheng objectid = min_key->objectid; 18985d4f98a2SYan Zheng while (1) { 18995d4f98a2SYan Zheng cond_resched(); 19005d4f98a2SYan Zheng iput(inode); 19015d4f98a2SYan Zheng 19025d4f98a2SYan Zheng if (objectid > max_key->objectid) 19035d4f98a2SYan Zheng break; 19045d4f98a2SYan Zheng 19055d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 19065d4f98a2SYan Zheng if (!inode) 19075d4f98a2SYan Zheng break; 19085d4f98a2SYan Zheng 19095d4f98a2SYan Zheng if (inode->i_ino > max_key->objectid) { 19105d4f98a2SYan Zheng iput(inode); 19115d4f98a2SYan Zheng break; 19125d4f98a2SYan Zheng } 19135d4f98a2SYan Zheng 19145d4f98a2SYan Zheng objectid = inode->i_ino + 1; 19155d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 19165d4f98a2SYan Zheng continue; 19175d4f98a2SYan Zheng 19185d4f98a2SYan Zheng if (unlikely(min_key->objectid == inode->i_ino)) { 19195d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 19205d4f98a2SYan Zheng continue; 19215d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 19225d4f98a2SYan Zheng start = 0; 19235d4f98a2SYan Zheng else { 19245d4f98a2SYan Zheng start = min_key->offset; 19255d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(start, root->sectorsize)); 19265d4f98a2SYan Zheng } 19275d4f98a2SYan Zheng } else { 19285d4f98a2SYan Zheng start = 0; 19295d4f98a2SYan Zheng } 19305d4f98a2SYan Zheng 19315d4f98a2SYan Zheng if (unlikely(max_key->objectid == inode->i_ino)) { 19325d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 19335d4f98a2SYan Zheng continue; 19345d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 19355d4f98a2SYan Zheng end = (u64)-1; 19365d4f98a2SYan Zheng } else { 19375d4f98a2SYan Zheng if (max_key->offset == 0) 19385d4f98a2SYan Zheng continue; 19395d4f98a2SYan Zheng end = max_key->offset; 19405d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 19415d4f98a2SYan Zheng end--; 19425d4f98a2SYan Zheng } 19435d4f98a2SYan Zheng } else { 19445d4f98a2SYan Zheng end = (u64)-1; 19455d4f98a2SYan Zheng } 19465d4f98a2SYan Zheng 19475d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 19485d4f98a2SYan Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 19495d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 1); 19505d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 19515d4f98a2SYan Zheng } 19525d4f98a2SYan Zheng return 0; 19535d4f98a2SYan Zheng } 19545d4f98a2SYan Zheng 19555d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 19565d4f98a2SYan Zheng struct btrfs_key *key) 19575d4f98a2SYan Zheng 19585d4f98a2SYan Zheng { 19595d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 19605d4f98a2SYan Zheng if (!path->nodes[level]) 19615d4f98a2SYan Zheng break; 19625d4f98a2SYan Zheng if (path->slots[level] + 1 < 19635d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 19645d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 19655d4f98a2SYan Zheng path->slots[level] + 1); 19665d4f98a2SYan Zheng return 0; 19675d4f98a2SYan Zheng } 19685d4f98a2SYan Zheng level++; 19695d4f98a2SYan Zheng } 19705d4f98a2SYan Zheng return 1; 19715d4f98a2SYan Zheng } 19725d4f98a2SYan Zheng 19735d4f98a2SYan Zheng /* 19745d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 19755d4f98a2SYan Zheng * fs tree. 19765d4f98a2SYan Zheng */ 19775d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 19785d4f98a2SYan Zheng struct btrfs_root *root) 19795d4f98a2SYan Zheng { 19805d4f98a2SYan Zheng LIST_HEAD(inode_list); 19815d4f98a2SYan Zheng struct btrfs_key key; 19825d4f98a2SYan Zheng struct btrfs_key next_key; 19835d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 19845d4f98a2SYan Zheng struct btrfs_root *reloc_root; 19855d4f98a2SYan Zheng struct btrfs_root_item *root_item; 19865d4f98a2SYan Zheng struct btrfs_path *path; 19873fd0a558SYan, Zheng struct extent_buffer *leaf; 19885d4f98a2SYan Zheng unsigned long nr; 19895d4f98a2SYan Zheng int level; 19905d4f98a2SYan Zheng int max_level; 19915d4f98a2SYan Zheng int replaced = 0; 19925d4f98a2SYan Zheng int ret; 19935d4f98a2SYan Zheng int err = 0; 19943fd0a558SYan, Zheng u32 min_reserved; 19955d4f98a2SYan Zheng 19965d4f98a2SYan Zheng path = btrfs_alloc_path(); 19975d4f98a2SYan Zheng if (!path) 19985d4f98a2SYan Zheng return -ENOMEM; 19995d4f98a2SYan Zheng 20005d4f98a2SYan Zheng reloc_root = root->reloc_root; 20015d4f98a2SYan Zheng root_item = &reloc_root->root_item; 20025d4f98a2SYan Zheng 20035d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 20045d4f98a2SYan Zheng level = btrfs_root_level(root_item); 20055d4f98a2SYan Zheng extent_buffer_get(reloc_root->node); 20065d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 20075d4f98a2SYan Zheng path->slots[level] = 0; 20085d4f98a2SYan Zheng } else { 20095d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 20105d4f98a2SYan Zheng 20115d4f98a2SYan Zheng level = root_item->drop_level; 20125d4f98a2SYan Zheng BUG_ON(level == 0); 20135d4f98a2SYan Zheng path->lowest_level = level; 20145d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 201533c66f43SYan Zheng path->lowest_level = 0; 20165d4f98a2SYan Zheng if (ret < 0) { 20175d4f98a2SYan Zheng btrfs_free_path(path); 20185d4f98a2SYan Zheng return ret; 20195d4f98a2SYan Zheng } 20205d4f98a2SYan Zheng 20215d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 20225d4f98a2SYan Zheng path->slots[level]); 20235d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 20245d4f98a2SYan Zheng 20255d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 20265d4f98a2SYan Zheng } 20275d4f98a2SYan Zheng 20283fd0a558SYan, Zheng min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 20295d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 20305d4f98a2SYan Zheng 20315d4f98a2SYan Zheng while (1) { 2032a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 0); 203398d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 20343fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 20353fd0a558SYan, Zheng 20363fd0a558SYan, Zheng ret = btrfs_block_rsv_check(trans, root, rc->block_rsv, 20373fd0a558SYan, Zheng min_reserved, 0); 20383fd0a558SYan, Zheng if (ret) { 20393fd0a558SYan, Zheng BUG_ON(ret != -EAGAIN); 20403fd0a558SYan, Zheng ret = btrfs_commit_transaction(trans, root); 20413fd0a558SYan, Zheng BUG_ON(ret); 20423fd0a558SYan, Zheng continue; 20433fd0a558SYan, Zheng } 20443fd0a558SYan, Zheng 20453fd0a558SYan, Zheng replaced = 0; 20465d4f98a2SYan Zheng max_level = level; 20475d4f98a2SYan Zheng 20485d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 20495d4f98a2SYan Zheng if (ret < 0) { 20505d4f98a2SYan Zheng err = ret; 20515d4f98a2SYan Zheng goto out; 20525d4f98a2SYan Zheng } 20535d4f98a2SYan Zheng if (ret > 0) 20545d4f98a2SYan Zheng break; 20555d4f98a2SYan Zheng 20565d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 20575d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 20585d4f98a2SYan Zheng ret = 0; 20595d4f98a2SYan Zheng } else { 20603fd0a558SYan, Zheng ret = replace_path(trans, root, reloc_root, path, 20613fd0a558SYan, Zheng &next_key, level, max_level); 20625d4f98a2SYan Zheng } 20635d4f98a2SYan Zheng if (ret < 0) { 20645d4f98a2SYan Zheng err = ret; 20655d4f98a2SYan Zheng goto out; 20665d4f98a2SYan Zheng } 20675d4f98a2SYan Zheng 20685d4f98a2SYan Zheng if (ret > 0) { 20695d4f98a2SYan Zheng level = ret; 20705d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 20715d4f98a2SYan Zheng path->slots[level]); 20725d4f98a2SYan Zheng replaced = 1; 20735d4f98a2SYan Zheng } 20745d4f98a2SYan Zheng 20755d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 20765d4f98a2SYan Zheng if (ret > 0) 20775d4f98a2SYan Zheng break; 20785d4f98a2SYan Zheng 20795d4f98a2SYan Zheng BUG_ON(level == 0); 20805d4f98a2SYan Zheng /* 20815d4f98a2SYan Zheng * save the merging progress in the drop_progress. 20825d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 20835d4f98a2SYan Zheng */ 20845d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 20855d4f98a2SYan Zheng path->slots[level]); 20865d4f98a2SYan Zheng root_item->drop_level = level; 20875d4f98a2SYan Zheng 20885d4f98a2SYan Zheng nr = trans->blocks_used; 20893fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, root); 20905d4f98a2SYan Zheng 20915d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 20925d4f98a2SYan Zheng 20935d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 20945d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 20955d4f98a2SYan Zheng } 20965d4f98a2SYan Zheng 20975d4f98a2SYan Zheng /* 20985d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 20995d4f98a2SYan Zheng * relocated and the block is tree root. 21005d4f98a2SYan Zheng */ 21015d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 21025d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 21035d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 21045d4f98a2SYan Zheng free_extent_buffer(leaf); 21055d4f98a2SYan Zheng if (ret < 0) 21065d4f98a2SYan Zheng err = ret; 21075d4f98a2SYan Zheng out: 21085d4f98a2SYan Zheng btrfs_free_path(path); 21095d4f98a2SYan Zheng 21105d4f98a2SYan Zheng if (err == 0) { 21115d4f98a2SYan Zheng memset(&root_item->drop_progress, 0, 21125d4f98a2SYan Zheng sizeof(root_item->drop_progress)); 21135d4f98a2SYan Zheng root_item->drop_level = 0; 21145d4f98a2SYan Zheng btrfs_set_root_refs(root_item, 0); 21153fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 21165d4f98a2SYan Zheng } 21175d4f98a2SYan Zheng 21185d4f98a2SYan Zheng nr = trans->blocks_used; 21193fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, root); 21205d4f98a2SYan Zheng 21215d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 21225d4f98a2SYan Zheng 21235d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 21245d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 21255d4f98a2SYan Zheng 21265d4f98a2SYan Zheng return err; 21275d4f98a2SYan Zheng } 21285d4f98a2SYan Zheng 21293fd0a558SYan, Zheng static noinline_for_stack 21303fd0a558SYan, Zheng int prepare_to_merge(struct reloc_control *rc, int err) 21315d4f98a2SYan Zheng { 21323fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 21333fd0a558SYan, Zheng struct btrfs_root *reloc_root; 21345d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 21353fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 21363fd0a558SYan, Zheng u64 num_bytes = 0; 21373fd0a558SYan, Zheng int ret; 21383fd0a558SYan, Zheng 21393fd0a558SYan, Zheng mutex_lock(&root->fs_info->trans_mutex); 21403fd0a558SYan, Zheng rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2; 21413fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated * 2; 21423fd0a558SYan, Zheng mutex_unlock(&root->fs_info->trans_mutex); 21433fd0a558SYan, Zheng again: 21443fd0a558SYan, Zheng if (!err) { 21453fd0a558SYan, Zheng num_bytes = rc->merging_rsv_size; 21463fd0a558SYan, Zheng ret = btrfs_block_rsv_add(NULL, root, rc->block_rsv, 21478bb8ab2eSJosef Bacik num_bytes); 21483fd0a558SYan, Zheng if (ret) 21493fd0a558SYan, Zheng err = ret; 21503fd0a558SYan, Zheng } 21513fd0a558SYan, Zheng 21523fd0a558SYan, Zheng trans = btrfs_join_transaction(rc->extent_root, 1); 21533612b495STsutomu Itoh if (IS_ERR(trans)) { 21543612b495STsutomu Itoh if (!err) 21553612b495STsutomu Itoh btrfs_block_rsv_release(rc->extent_root, 21563612b495STsutomu Itoh rc->block_rsv, num_bytes); 21573612b495STsutomu Itoh return PTR_ERR(trans); 21583612b495STsutomu Itoh } 21593fd0a558SYan, Zheng 21603fd0a558SYan, Zheng if (!err) { 21613fd0a558SYan, Zheng if (num_bytes != rc->merging_rsv_size) { 21623fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 21633fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, 21643fd0a558SYan, Zheng rc->block_rsv, num_bytes); 21653fd0a558SYan, Zheng goto again; 21663fd0a558SYan, Zheng } 21673fd0a558SYan, Zheng } 21683fd0a558SYan, Zheng 21693fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 21703fd0a558SYan, Zheng 21713fd0a558SYan, Zheng while (!list_empty(&rc->reloc_roots)) { 21723fd0a558SYan, Zheng reloc_root = list_entry(rc->reloc_roots.next, 21733fd0a558SYan, Zheng struct btrfs_root, root_list); 21743fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 21753fd0a558SYan, Zheng 21763fd0a558SYan, Zheng root = read_fs_root(reloc_root->fs_info, 21773fd0a558SYan, Zheng reloc_root->root_key.offset); 21783fd0a558SYan, Zheng BUG_ON(IS_ERR(root)); 21793fd0a558SYan, Zheng BUG_ON(root->reloc_root != reloc_root); 21803fd0a558SYan, Zheng 21813fd0a558SYan, Zheng /* 21823fd0a558SYan, Zheng * set reference count to 1, so btrfs_recover_relocation 21833fd0a558SYan, Zheng * knows it should resumes merging 21843fd0a558SYan, Zheng */ 21853fd0a558SYan, Zheng if (!err) 21863fd0a558SYan, Zheng btrfs_set_root_refs(&reloc_root->root_item, 1); 21873fd0a558SYan, Zheng btrfs_update_reloc_root(trans, root); 21883fd0a558SYan, Zheng 21893fd0a558SYan, Zheng list_add(&reloc_root->root_list, &reloc_roots); 21903fd0a558SYan, Zheng } 21913fd0a558SYan, Zheng 21923fd0a558SYan, Zheng list_splice(&reloc_roots, &rc->reloc_roots); 21933fd0a558SYan, Zheng 21943fd0a558SYan, Zheng if (!err) 21953fd0a558SYan, Zheng btrfs_commit_transaction(trans, rc->extent_root); 21963fd0a558SYan, Zheng else 21973fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 21983fd0a558SYan, Zheng return err; 21993fd0a558SYan, Zheng } 22003fd0a558SYan, Zheng 22013fd0a558SYan, Zheng static noinline_for_stack 22023fd0a558SYan, Zheng int merge_reloc_roots(struct reloc_control *rc) 22033fd0a558SYan, Zheng { 22045d4f98a2SYan Zheng struct btrfs_root *root; 22055d4f98a2SYan Zheng struct btrfs_root *reloc_root; 22063fd0a558SYan, Zheng LIST_HEAD(reloc_roots); 22073fd0a558SYan, Zheng int found = 0; 22083fd0a558SYan, Zheng int ret; 22093fd0a558SYan, Zheng again: 22103fd0a558SYan, Zheng root = rc->extent_root; 22113fd0a558SYan, Zheng mutex_lock(&root->fs_info->trans_mutex); 22123fd0a558SYan, Zheng list_splice_init(&rc->reloc_roots, &reloc_roots); 22133fd0a558SYan, Zheng mutex_unlock(&root->fs_info->trans_mutex); 22145d4f98a2SYan Zheng 22153fd0a558SYan, Zheng while (!list_empty(&reloc_roots)) { 22163fd0a558SYan, Zheng found = 1; 22173fd0a558SYan, Zheng reloc_root = list_entry(reloc_roots.next, 22183fd0a558SYan, Zheng struct btrfs_root, root_list); 22195d4f98a2SYan Zheng 22205d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 22215d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, 22225d4f98a2SYan Zheng reloc_root->root_key.offset); 22235d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 22245d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 22255d4f98a2SYan Zheng 22263fd0a558SYan, Zheng ret = merge_reloc_root(rc, root); 22273fd0a558SYan, Zheng BUG_ON(ret); 22283fd0a558SYan, Zheng } else { 22293fd0a558SYan, Zheng list_del_init(&reloc_root->root_list); 22303fd0a558SYan, Zheng } 22313fd0a558SYan, Zheng btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0); 22325d4f98a2SYan Zheng } 22335d4f98a2SYan Zheng 22343fd0a558SYan, Zheng if (found) { 22353fd0a558SYan, Zheng found = 0; 22363fd0a558SYan, Zheng goto again; 22375d4f98a2SYan Zheng } 22385d4f98a2SYan Zheng BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 22395d4f98a2SYan Zheng return 0; 22405d4f98a2SYan Zheng } 22415d4f98a2SYan Zheng 22425d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 22435d4f98a2SYan Zheng { 22445d4f98a2SYan Zheng struct tree_block *block; 22455d4f98a2SYan Zheng struct rb_node *rb_node; 22465d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 22475d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 22485d4f98a2SYan Zheng rb_erase(rb_node, blocks); 22495d4f98a2SYan Zheng kfree(block); 22505d4f98a2SYan Zheng } 22515d4f98a2SYan Zheng } 22525d4f98a2SYan Zheng 22535d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 22545d4f98a2SYan Zheng struct btrfs_root *reloc_root) 22555d4f98a2SYan Zheng { 22565d4f98a2SYan Zheng struct btrfs_root *root; 22575d4f98a2SYan Zheng 22585d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 22595d4f98a2SYan Zheng return 0; 22605d4f98a2SYan Zheng 22615d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset); 22625d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 22635d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 22645d4f98a2SYan Zheng 22655d4f98a2SYan Zheng return btrfs_record_root_in_trans(trans, root); 22665d4f98a2SYan Zheng } 22675d4f98a2SYan Zheng 22683fd0a558SYan, Zheng static noinline_for_stack 22693fd0a558SYan, Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 22703fd0a558SYan, Zheng struct reloc_control *rc, 22715d4f98a2SYan Zheng struct backref_node *node, 22723fd0a558SYan, Zheng struct backref_edge *edges[], int *nr) 22735d4f98a2SYan Zheng { 22745d4f98a2SYan Zheng struct backref_node *next; 22755d4f98a2SYan Zheng struct btrfs_root *root; 22763fd0a558SYan, Zheng int index = 0; 22773fd0a558SYan, Zheng 22785d4f98a2SYan Zheng next = node; 22795d4f98a2SYan Zheng while (1) { 22805d4f98a2SYan Zheng cond_resched(); 22815d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 22825d4f98a2SYan Zheng root = next->root; 22833fd0a558SYan, Zheng BUG_ON(!root); 22843fd0a558SYan, Zheng BUG_ON(!root->ref_cows); 22855d4f98a2SYan Zheng 22865d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 22875d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 22885d4f98a2SYan Zheng break; 22895d4f98a2SYan Zheng } 22905d4f98a2SYan Zheng 22915d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 22923fd0a558SYan, Zheng root = root->reloc_root; 22933fd0a558SYan, Zheng 22943fd0a558SYan, Zheng if (next->new_bytenr != root->node->start) { 22953fd0a558SYan, Zheng BUG_ON(next->new_bytenr); 22963fd0a558SYan, Zheng BUG_ON(!list_empty(&next->list)); 22973fd0a558SYan, Zheng next->new_bytenr = root->node->start; 22983fd0a558SYan, Zheng next->root = root; 22993fd0a558SYan, Zheng list_add_tail(&next->list, 23003fd0a558SYan, Zheng &rc->backref_cache.changed); 23013fd0a558SYan, Zheng __mark_block_processed(rc, next); 23025d4f98a2SYan Zheng break; 23035d4f98a2SYan Zheng } 23045d4f98a2SYan Zheng 23053fd0a558SYan, Zheng WARN_ON(1); 23065d4f98a2SYan Zheng root = NULL; 23075d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 23085d4f98a2SYan Zheng if (!next || next->level <= node->level) 23095d4f98a2SYan Zheng break; 23105d4f98a2SYan Zheng } 23113fd0a558SYan, Zheng if (!root) 23123fd0a558SYan, Zheng return NULL; 23135d4f98a2SYan Zheng 23145d4f98a2SYan Zheng *nr = index; 23153fd0a558SYan, Zheng next = node; 23163fd0a558SYan, Zheng /* setup backref node path for btrfs_reloc_cow_block */ 23173fd0a558SYan, Zheng while (1) { 23183fd0a558SYan, Zheng rc->backref_cache.path[next->level] = next; 23193fd0a558SYan, Zheng if (--index < 0) 23203fd0a558SYan, Zheng break; 23213fd0a558SYan, Zheng next = edges[index]->node[UPPER]; 23223fd0a558SYan, Zheng } 23235d4f98a2SYan Zheng return root; 23245d4f98a2SYan Zheng } 23255d4f98a2SYan Zheng 23263fd0a558SYan, Zheng /* 23273fd0a558SYan, Zheng * select a tree root for relocation. return NULL if the block 23283fd0a558SYan, Zheng * is reference counted. we should use do_relocation() in this 23293fd0a558SYan, Zheng * case. return a tree root pointer if the block isn't reference 23303fd0a558SYan, Zheng * counted. return -ENOENT if the block is root of reloc tree. 23313fd0a558SYan, Zheng */ 23325d4f98a2SYan Zheng static noinline_for_stack 23335d4f98a2SYan Zheng struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans, 23345d4f98a2SYan Zheng struct backref_node *node) 23355d4f98a2SYan Zheng { 23363fd0a558SYan, Zheng struct backref_node *next; 23373fd0a558SYan, Zheng struct btrfs_root *root; 23383fd0a558SYan, Zheng struct btrfs_root *fs_root = NULL; 23395d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 23403fd0a558SYan, Zheng int index = 0; 23413fd0a558SYan, Zheng 23423fd0a558SYan, Zheng next = node; 23433fd0a558SYan, Zheng while (1) { 23443fd0a558SYan, Zheng cond_resched(); 23453fd0a558SYan, Zheng next = walk_up_backref(next, edges, &index); 23463fd0a558SYan, Zheng root = next->root; 23473fd0a558SYan, Zheng BUG_ON(!root); 23483fd0a558SYan, Zheng 234925985edcSLucas De Marchi /* no other choice for non-references counted tree */ 23503fd0a558SYan, Zheng if (!root->ref_cows) 23513fd0a558SYan, Zheng return root; 23523fd0a558SYan, Zheng 23533fd0a558SYan, Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) 23543fd0a558SYan, Zheng fs_root = root; 23553fd0a558SYan, Zheng 23563fd0a558SYan, Zheng if (next != node) 23573fd0a558SYan, Zheng return NULL; 23583fd0a558SYan, Zheng 23593fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 23603fd0a558SYan, Zheng if (!next || next->level <= node->level) 23613fd0a558SYan, Zheng break; 23623fd0a558SYan, Zheng } 23633fd0a558SYan, Zheng 23643fd0a558SYan, Zheng if (!fs_root) 23653fd0a558SYan, Zheng return ERR_PTR(-ENOENT); 23663fd0a558SYan, Zheng return fs_root; 23675d4f98a2SYan Zheng } 23685d4f98a2SYan Zheng 23695d4f98a2SYan Zheng static noinline_for_stack 23703fd0a558SYan, Zheng u64 calcu_metadata_size(struct reloc_control *rc, 23713fd0a558SYan, Zheng struct backref_node *node, int reserve) 23725d4f98a2SYan Zheng { 23733fd0a558SYan, Zheng struct backref_node *next = node; 23743fd0a558SYan, Zheng struct backref_edge *edge; 23753fd0a558SYan, Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 23763fd0a558SYan, Zheng u64 num_bytes = 0; 23773fd0a558SYan, Zheng int index = 0; 23785d4f98a2SYan Zheng 23793fd0a558SYan, Zheng BUG_ON(reserve && node->processed); 23803fd0a558SYan, Zheng 23813fd0a558SYan, Zheng while (next) { 23823fd0a558SYan, Zheng cond_resched(); 23835d4f98a2SYan Zheng while (1) { 23843fd0a558SYan, Zheng if (next->processed && (reserve || next != node)) 23855d4f98a2SYan Zheng break; 23865d4f98a2SYan Zheng 23873fd0a558SYan, Zheng num_bytes += btrfs_level_size(rc->extent_root, 23883fd0a558SYan, Zheng next->level); 23893fd0a558SYan, Zheng 23903fd0a558SYan, Zheng if (list_empty(&next->upper)) 23913fd0a558SYan, Zheng break; 23923fd0a558SYan, Zheng 23933fd0a558SYan, Zheng edge = list_entry(next->upper.next, 23943fd0a558SYan, Zheng struct backref_edge, list[LOWER]); 23953fd0a558SYan, Zheng edges[index++] = edge; 23963fd0a558SYan, Zheng next = edge->node[UPPER]; 23975d4f98a2SYan Zheng } 23983fd0a558SYan, Zheng next = walk_down_backref(edges, &index); 23993fd0a558SYan, Zheng } 24003fd0a558SYan, Zheng return num_bytes; 24013fd0a558SYan, Zheng } 24023fd0a558SYan, Zheng 24033fd0a558SYan, Zheng static int reserve_metadata_space(struct btrfs_trans_handle *trans, 24043fd0a558SYan, Zheng struct reloc_control *rc, 24053fd0a558SYan, Zheng struct backref_node *node) 24063fd0a558SYan, Zheng { 24073fd0a558SYan, Zheng struct btrfs_root *root = rc->extent_root; 24083fd0a558SYan, Zheng u64 num_bytes; 24093fd0a558SYan, Zheng int ret; 24103fd0a558SYan, Zheng 24113fd0a558SYan, Zheng num_bytes = calcu_metadata_size(rc, node, 1) * 2; 24123fd0a558SYan, Zheng 24133fd0a558SYan, Zheng trans->block_rsv = rc->block_rsv; 24148bb8ab2eSJosef Bacik ret = btrfs_block_rsv_add(trans, root, rc->block_rsv, num_bytes); 24153fd0a558SYan, Zheng if (ret) { 24163fd0a558SYan, Zheng if (ret == -EAGAIN) 24173fd0a558SYan, Zheng rc->commit_transaction = 1; 24183fd0a558SYan, Zheng return ret; 24193fd0a558SYan, Zheng } 24203fd0a558SYan, Zheng 24213fd0a558SYan, Zheng return 0; 24223fd0a558SYan, Zheng } 24233fd0a558SYan, Zheng 24243fd0a558SYan, Zheng static void release_metadata_space(struct reloc_control *rc, 24253fd0a558SYan, Zheng struct backref_node *node) 24263fd0a558SYan, Zheng { 24273fd0a558SYan, Zheng u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2; 24283fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes); 24295d4f98a2SYan Zheng } 24305d4f98a2SYan Zheng 24315d4f98a2SYan Zheng /* 24325d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 24335d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 24345d4f98a2SYan Zheng * 24355d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 24365d4f98a2SYan Zheng * in that case this function just updates pointers. 24375d4f98a2SYan Zheng */ 24385d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 24393fd0a558SYan, Zheng struct reloc_control *rc, 24405d4f98a2SYan Zheng struct backref_node *node, 24415d4f98a2SYan Zheng struct btrfs_key *key, 24425d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 24435d4f98a2SYan Zheng { 24445d4f98a2SYan Zheng struct backref_node *upper; 24455d4f98a2SYan Zheng struct backref_edge *edge; 24465d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 24475d4f98a2SYan Zheng struct btrfs_root *root; 24485d4f98a2SYan Zheng struct extent_buffer *eb; 24495d4f98a2SYan Zheng u32 blocksize; 24505d4f98a2SYan Zheng u64 bytenr; 24515d4f98a2SYan Zheng u64 generation; 24525d4f98a2SYan Zheng int nr; 24535d4f98a2SYan Zheng int slot; 24545d4f98a2SYan Zheng int ret; 24555d4f98a2SYan Zheng int err = 0; 24565d4f98a2SYan Zheng 24575d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 24585d4f98a2SYan Zheng 24595d4f98a2SYan Zheng path->lowest_level = node->level + 1; 24603fd0a558SYan, Zheng rc->backref_cache.path[node->level] = node; 24615d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 24625d4f98a2SYan Zheng cond_resched(); 24635d4f98a2SYan Zheng 24645d4f98a2SYan Zheng upper = edge->node[UPPER]; 24653fd0a558SYan, Zheng root = select_reloc_root(trans, rc, upper, edges, &nr); 24663fd0a558SYan, Zheng BUG_ON(!root); 24675d4f98a2SYan Zheng 24683fd0a558SYan, Zheng if (upper->eb && !upper->locked) { 24693fd0a558SYan, Zheng if (!lowest) { 24703fd0a558SYan, Zheng ret = btrfs_bin_search(upper->eb, key, 24713fd0a558SYan, Zheng upper->level, &slot); 24723fd0a558SYan, Zheng BUG_ON(ret); 24733fd0a558SYan, Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 24743fd0a558SYan, Zheng if (node->eb->start == bytenr) 24753fd0a558SYan, Zheng goto next; 24763fd0a558SYan, Zheng } 24775d4f98a2SYan Zheng drop_node_buffer(upper); 24783fd0a558SYan, Zheng } 24795d4f98a2SYan Zheng 24805d4f98a2SYan Zheng if (!upper->eb) { 24815d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 24825d4f98a2SYan Zheng if (ret < 0) { 24835d4f98a2SYan Zheng err = ret; 24845d4f98a2SYan Zheng break; 24855d4f98a2SYan Zheng } 24865d4f98a2SYan Zheng BUG_ON(ret > 0); 24875d4f98a2SYan Zheng 24883fd0a558SYan, Zheng if (!upper->eb) { 24893fd0a558SYan, Zheng upper->eb = path->nodes[upper->level]; 24903fd0a558SYan, Zheng path->nodes[upper->level] = NULL; 24913fd0a558SYan, Zheng } else { 24923fd0a558SYan, Zheng BUG_ON(upper->eb != path->nodes[upper->level]); 24933fd0a558SYan, Zheng } 24943fd0a558SYan, Zheng 24953fd0a558SYan, Zheng upper->locked = 1; 24963fd0a558SYan, Zheng path->locks[upper->level] = 0; 24973fd0a558SYan, Zheng 24985d4f98a2SYan Zheng slot = path->slots[upper->level]; 2499*b3b4aa74SDavid Sterba btrfs_release_path(path); 25005d4f98a2SYan Zheng } else { 25015d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 25025d4f98a2SYan Zheng &slot); 25035d4f98a2SYan Zheng BUG_ON(ret); 25045d4f98a2SYan Zheng } 25055d4f98a2SYan Zheng 25065d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 25073fd0a558SYan, Zheng if (lowest) { 25083fd0a558SYan, Zheng BUG_ON(bytenr != node->bytenr); 25095d4f98a2SYan Zheng } else { 25103fd0a558SYan, Zheng if (node->eb->start == bytenr) 25113fd0a558SYan, Zheng goto next; 25125d4f98a2SYan Zheng } 25135d4f98a2SYan Zheng 25145d4f98a2SYan Zheng blocksize = btrfs_level_size(root, node->level); 25155d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 25165d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, generation); 251797d9a8a4STsutomu Itoh if (!eb) { 251897d9a8a4STsutomu Itoh err = -EIO; 251997d9a8a4STsutomu Itoh goto next; 252097d9a8a4STsutomu Itoh } 25215d4f98a2SYan Zheng btrfs_tree_lock(eb); 25225d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 25235d4f98a2SYan Zheng 25245d4f98a2SYan Zheng if (!node->eb) { 25255d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 25265d4f98a2SYan Zheng slot, &eb); 25273fd0a558SYan, Zheng btrfs_tree_unlock(eb); 25283fd0a558SYan, Zheng free_extent_buffer(eb); 25295d4f98a2SYan Zheng if (ret < 0) { 25305d4f98a2SYan Zheng err = ret; 25313fd0a558SYan, Zheng goto next; 25325d4f98a2SYan Zheng } 25333fd0a558SYan, Zheng BUG_ON(node->eb != eb); 25345d4f98a2SYan Zheng } else { 25355d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 25365d4f98a2SYan Zheng node->eb->start); 25375d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 25385d4f98a2SYan Zheng trans->transid); 25395d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 25405d4f98a2SYan Zheng 25415d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, 25425d4f98a2SYan Zheng node->eb->start, blocksize, 25435d4f98a2SYan Zheng upper->eb->start, 25445d4f98a2SYan Zheng btrfs_header_owner(upper->eb), 25455d4f98a2SYan Zheng node->level, 0); 25465d4f98a2SYan Zheng BUG_ON(ret); 25475d4f98a2SYan Zheng 25485d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 25495d4f98a2SYan Zheng BUG_ON(ret); 25505d4f98a2SYan Zheng } 25513fd0a558SYan, Zheng next: 25523fd0a558SYan, Zheng if (!upper->pending) 25533fd0a558SYan, Zheng drop_node_buffer(upper); 25543fd0a558SYan, Zheng else 25553fd0a558SYan, Zheng unlock_node_buffer(upper); 25563fd0a558SYan, Zheng if (err) 25573fd0a558SYan, Zheng break; 25585d4f98a2SYan Zheng } 25593fd0a558SYan, Zheng 25603fd0a558SYan, Zheng if (!err && node->pending) { 25613fd0a558SYan, Zheng drop_node_buffer(node); 25623fd0a558SYan, Zheng list_move_tail(&node->list, &rc->backref_cache.changed); 25633fd0a558SYan, Zheng node->pending = 0; 25645d4f98a2SYan Zheng } 25653fd0a558SYan, Zheng 25665d4f98a2SYan Zheng path->lowest_level = 0; 25673fd0a558SYan, Zheng BUG_ON(err == -ENOSPC); 25685d4f98a2SYan Zheng return err; 25695d4f98a2SYan Zheng } 25705d4f98a2SYan Zheng 25715d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 25723fd0a558SYan, Zheng struct reloc_control *rc, 25735d4f98a2SYan Zheng struct backref_node *node, 25745d4f98a2SYan Zheng struct btrfs_path *path) 25755d4f98a2SYan Zheng { 25765d4f98a2SYan Zheng struct btrfs_key key; 25775d4f98a2SYan Zheng 25785d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 25793fd0a558SYan, Zheng return do_relocation(trans, rc, node, &key, path, 0); 25805d4f98a2SYan Zheng } 25815d4f98a2SYan Zheng 25825d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 25833fd0a558SYan, Zheng struct reloc_control *rc, 25843fd0a558SYan, Zheng struct btrfs_path *path, int err) 25855d4f98a2SYan Zheng { 25863fd0a558SYan, Zheng LIST_HEAD(list); 25873fd0a558SYan, Zheng struct backref_cache *cache = &rc->backref_cache; 25885d4f98a2SYan Zheng struct backref_node *node; 25895d4f98a2SYan Zheng int level; 25905d4f98a2SYan Zheng int ret; 25915d4f98a2SYan Zheng 25925d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 25935d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 25945d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 25953fd0a558SYan, Zheng struct backref_node, list); 25963fd0a558SYan, Zheng list_move_tail(&node->list, &list); 25973fd0a558SYan, Zheng BUG_ON(!node->pending); 25985d4f98a2SYan Zheng 25993fd0a558SYan, Zheng if (!err) { 26003fd0a558SYan, Zheng ret = link_to_upper(trans, rc, node, path); 26015d4f98a2SYan Zheng if (ret < 0) 26025d4f98a2SYan Zheng err = ret; 26035d4f98a2SYan Zheng } 26045d4f98a2SYan Zheng } 26053fd0a558SYan, Zheng list_splice_init(&list, &cache->pending[level]); 26063fd0a558SYan, Zheng } 26075d4f98a2SYan Zheng return err; 26085d4f98a2SYan Zheng } 26095d4f98a2SYan Zheng 26105d4f98a2SYan Zheng static void mark_block_processed(struct reloc_control *rc, 26113fd0a558SYan, Zheng u64 bytenr, u32 blocksize) 26123fd0a558SYan, Zheng { 26133fd0a558SYan, Zheng set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1, 26143fd0a558SYan, Zheng EXTENT_DIRTY, GFP_NOFS); 26153fd0a558SYan, Zheng } 26163fd0a558SYan, Zheng 26173fd0a558SYan, Zheng static void __mark_block_processed(struct reloc_control *rc, 26185d4f98a2SYan Zheng struct backref_node *node) 26195d4f98a2SYan Zheng { 26205d4f98a2SYan Zheng u32 blocksize; 26215d4f98a2SYan Zheng if (node->level == 0 || 26225d4f98a2SYan Zheng in_block_group(node->bytenr, rc->block_group)) { 26235d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, node->level); 26243fd0a558SYan, Zheng mark_block_processed(rc, node->bytenr, blocksize); 26255d4f98a2SYan Zheng } 26265d4f98a2SYan Zheng node->processed = 1; 26275d4f98a2SYan Zheng } 26285d4f98a2SYan Zheng 26295d4f98a2SYan Zheng /* 26305d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 26315d4f98a2SYan Zheng * as processed. 26325d4f98a2SYan Zheng */ 26335d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 26345d4f98a2SYan Zheng struct backref_node *node) 26355d4f98a2SYan Zheng { 26365d4f98a2SYan Zheng struct backref_node *next = node; 26375d4f98a2SYan Zheng struct backref_edge *edge; 26385d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 26395d4f98a2SYan Zheng int index = 0; 26405d4f98a2SYan Zheng 26415d4f98a2SYan Zheng while (next) { 26425d4f98a2SYan Zheng cond_resched(); 26435d4f98a2SYan Zheng while (1) { 26445d4f98a2SYan Zheng if (next->processed) 26455d4f98a2SYan Zheng break; 26465d4f98a2SYan Zheng 26473fd0a558SYan, Zheng __mark_block_processed(rc, next); 26485d4f98a2SYan Zheng 26495d4f98a2SYan Zheng if (list_empty(&next->upper)) 26505d4f98a2SYan Zheng break; 26515d4f98a2SYan Zheng 26525d4f98a2SYan Zheng edge = list_entry(next->upper.next, 26535d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 26545d4f98a2SYan Zheng edges[index++] = edge; 26555d4f98a2SYan Zheng next = edge->node[UPPER]; 26565d4f98a2SYan Zheng } 26575d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 26585d4f98a2SYan Zheng } 26595d4f98a2SYan Zheng } 26605d4f98a2SYan Zheng 26615d4f98a2SYan Zheng static int tree_block_processed(u64 bytenr, u32 blocksize, 26625d4f98a2SYan Zheng struct reloc_control *rc) 26635d4f98a2SYan Zheng { 26645d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 26659655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 26665d4f98a2SYan Zheng return 1; 26675d4f98a2SYan Zheng return 0; 26685d4f98a2SYan Zheng } 26695d4f98a2SYan Zheng 26705d4f98a2SYan Zheng static int get_tree_block_key(struct reloc_control *rc, 26715d4f98a2SYan Zheng struct tree_block *block) 26725d4f98a2SYan Zheng { 26735d4f98a2SYan Zheng struct extent_buffer *eb; 26745d4f98a2SYan Zheng 26755d4f98a2SYan Zheng BUG_ON(block->key_ready); 26765d4f98a2SYan Zheng eb = read_tree_block(rc->extent_root, block->bytenr, 26775d4f98a2SYan Zheng block->key.objectid, block->key.offset); 267897d9a8a4STsutomu Itoh BUG_ON(!eb); 26795d4f98a2SYan Zheng WARN_ON(btrfs_header_level(eb) != block->level); 26805d4f98a2SYan Zheng if (block->level == 0) 26815d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 26825d4f98a2SYan Zheng else 26835d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 26845d4f98a2SYan Zheng free_extent_buffer(eb); 26855d4f98a2SYan Zheng block->key_ready = 1; 26865d4f98a2SYan Zheng return 0; 26875d4f98a2SYan Zheng } 26885d4f98a2SYan Zheng 26895d4f98a2SYan Zheng static int reada_tree_block(struct reloc_control *rc, 26905d4f98a2SYan Zheng struct tree_block *block) 26915d4f98a2SYan Zheng { 26925d4f98a2SYan Zheng BUG_ON(block->key_ready); 26935d4f98a2SYan Zheng readahead_tree_block(rc->extent_root, block->bytenr, 26945d4f98a2SYan Zheng block->key.objectid, block->key.offset); 26955d4f98a2SYan Zheng return 0; 26965d4f98a2SYan Zheng } 26975d4f98a2SYan Zheng 26985d4f98a2SYan Zheng /* 26995d4f98a2SYan Zheng * helper function to relocate a tree block 27005d4f98a2SYan Zheng */ 27015d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 27025d4f98a2SYan Zheng struct reloc_control *rc, 27035d4f98a2SYan Zheng struct backref_node *node, 27045d4f98a2SYan Zheng struct btrfs_key *key, 27055d4f98a2SYan Zheng struct btrfs_path *path) 27065d4f98a2SYan Zheng { 27075d4f98a2SYan Zheng struct btrfs_root *root; 27083fd0a558SYan, Zheng int release = 0; 27093fd0a558SYan, Zheng int ret = 0; 27105d4f98a2SYan Zheng 27113fd0a558SYan, Zheng if (!node) 27125d4f98a2SYan Zheng return 0; 27133fd0a558SYan, Zheng 27143fd0a558SYan, Zheng BUG_ON(node->processed); 27153fd0a558SYan, Zheng root = select_one_root(trans, node); 27163fd0a558SYan, Zheng if (root == ERR_PTR(-ENOENT)) { 27173fd0a558SYan, Zheng update_processed_blocks(rc, node); 27183fd0a558SYan, Zheng goto out; 27195d4f98a2SYan Zheng } 27205d4f98a2SYan Zheng 27213fd0a558SYan, Zheng if (!root || root->ref_cows) { 27223fd0a558SYan, Zheng ret = reserve_metadata_space(trans, rc, node); 27233fd0a558SYan, Zheng if (ret) 27245d4f98a2SYan Zheng goto out; 27253fd0a558SYan, Zheng release = 1; 27265d4f98a2SYan Zheng } 27273fd0a558SYan, Zheng 27283fd0a558SYan, Zheng if (root) { 27293fd0a558SYan, Zheng if (root->ref_cows) { 27303fd0a558SYan, Zheng BUG_ON(node->new_bytenr); 27313fd0a558SYan, Zheng BUG_ON(!list_empty(&node->list)); 27323fd0a558SYan, Zheng btrfs_record_root_in_trans(trans, root); 27333fd0a558SYan, Zheng root = root->reloc_root; 27343fd0a558SYan, Zheng node->new_bytenr = root->node->start; 27353fd0a558SYan, Zheng node->root = root; 27363fd0a558SYan, Zheng list_add_tail(&node->list, &rc->backref_cache.changed); 27373fd0a558SYan, Zheng } else { 27385d4f98a2SYan Zheng path->lowest_level = node->level; 27395d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2740*b3b4aa74SDavid Sterba btrfs_release_path(path); 27413fd0a558SYan, Zheng if (ret > 0) 27425d4f98a2SYan Zheng ret = 0; 27433fd0a558SYan, Zheng } 27443fd0a558SYan, Zheng if (!ret) 27453fd0a558SYan, Zheng update_processed_blocks(rc, node); 27463fd0a558SYan, Zheng } else { 27473fd0a558SYan, Zheng ret = do_relocation(trans, rc, node, key, path, 1); 27483fd0a558SYan, Zheng } 27495d4f98a2SYan Zheng out: 27503fd0a558SYan, Zheng if (ret || node->level == 0 || node->cowonly) { 27513fd0a558SYan, Zheng if (release) 27523fd0a558SYan, Zheng release_metadata_space(rc, node); 27533fd0a558SYan, Zheng remove_backref_node(&rc->backref_cache, node); 27543fd0a558SYan, Zheng } 27555d4f98a2SYan Zheng return ret; 27565d4f98a2SYan Zheng } 27575d4f98a2SYan Zheng 27585d4f98a2SYan Zheng /* 27595d4f98a2SYan Zheng * relocate a list of blocks 27605d4f98a2SYan Zheng */ 27615d4f98a2SYan Zheng static noinline_for_stack 27625d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 27635d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 27645d4f98a2SYan Zheng { 27655d4f98a2SYan Zheng struct backref_node *node; 27665d4f98a2SYan Zheng struct btrfs_path *path; 27675d4f98a2SYan Zheng struct tree_block *block; 27685d4f98a2SYan Zheng struct rb_node *rb_node; 27695d4f98a2SYan Zheng int ret; 27705d4f98a2SYan Zheng int err = 0; 27715d4f98a2SYan Zheng 27725d4f98a2SYan Zheng path = btrfs_alloc_path(); 27735d4f98a2SYan Zheng if (!path) 27745d4f98a2SYan Zheng return -ENOMEM; 27755d4f98a2SYan Zheng 27765d4f98a2SYan Zheng rb_node = rb_first(blocks); 27775d4f98a2SYan Zheng while (rb_node) { 27785d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 27795d4f98a2SYan Zheng if (!block->key_ready) 27805d4f98a2SYan Zheng reada_tree_block(rc, block); 27815d4f98a2SYan Zheng rb_node = rb_next(rb_node); 27825d4f98a2SYan Zheng } 27835d4f98a2SYan Zheng 27845d4f98a2SYan Zheng rb_node = rb_first(blocks); 27855d4f98a2SYan Zheng while (rb_node) { 27865d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 27875d4f98a2SYan Zheng if (!block->key_ready) 27885d4f98a2SYan Zheng get_tree_block_key(rc, block); 27895d4f98a2SYan Zheng rb_node = rb_next(rb_node); 27905d4f98a2SYan Zheng } 27915d4f98a2SYan Zheng 27925d4f98a2SYan Zheng rb_node = rb_first(blocks); 27935d4f98a2SYan Zheng while (rb_node) { 27945d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 27955d4f98a2SYan Zheng 27963fd0a558SYan, Zheng node = build_backref_tree(rc, &block->key, 27975d4f98a2SYan Zheng block->level, block->bytenr); 27985d4f98a2SYan Zheng if (IS_ERR(node)) { 27995d4f98a2SYan Zheng err = PTR_ERR(node); 28005d4f98a2SYan Zheng goto out; 28015d4f98a2SYan Zheng } 28025d4f98a2SYan Zheng 28035d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 28045d4f98a2SYan Zheng path); 28055d4f98a2SYan Zheng if (ret < 0) { 28063fd0a558SYan, Zheng if (ret != -EAGAIN || rb_node == rb_first(blocks)) 28075d4f98a2SYan Zheng err = ret; 28085d4f98a2SYan Zheng goto out; 28095d4f98a2SYan Zheng } 28105d4f98a2SYan Zheng rb_node = rb_next(rb_node); 28115d4f98a2SYan Zheng } 28125d4f98a2SYan Zheng out: 28135d4f98a2SYan Zheng free_block_list(blocks); 28143fd0a558SYan, Zheng err = finish_pending_nodes(trans, rc, path, err); 28155d4f98a2SYan Zheng 28165d4f98a2SYan Zheng btrfs_free_path(path); 28175d4f98a2SYan Zheng return err; 28185d4f98a2SYan Zheng } 28195d4f98a2SYan Zheng 28205d4f98a2SYan Zheng static noinline_for_stack 2821efa56464SYan, Zheng int prealloc_file_extent_cluster(struct inode *inode, 2822efa56464SYan, Zheng struct file_extent_cluster *cluster) 2823efa56464SYan, Zheng { 2824efa56464SYan, Zheng u64 alloc_hint = 0; 2825efa56464SYan, Zheng u64 start; 2826efa56464SYan, Zheng u64 end; 2827efa56464SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 2828efa56464SYan, Zheng u64 num_bytes; 2829efa56464SYan, Zheng int nr = 0; 2830efa56464SYan, Zheng int ret = 0; 2831efa56464SYan, Zheng 2832efa56464SYan, Zheng BUG_ON(cluster->start != cluster->boundary[0]); 2833efa56464SYan, Zheng mutex_lock(&inode->i_mutex); 2834efa56464SYan, Zheng 2835efa56464SYan, Zheng ret = btrfs_check_data_free_space(inode, cluster->end + 2836efa56464SYan, Zheng 1 - cluster->start); 2837efa56464SYan, Zheng if (ret) 2838efa56464SYan, Zheng goto out; 2839efa56464SYan, Zheng 2840efa56464SYan, Zheng while (nr < cluster->nr) { 2841efa56464SYan, Zheng start = cluster->boundary[nr] - offset; 2842efa56464SYan, Zheng if (nr + 1 < cluster->nr) 2843efa56464SYan, Zheng end = cluster->boundary[nr + 1] - 1 - offset; 2844efa56464SYan, Zheng else 2845efa56464SYan, Zheng end = cluster->end - offset; 2846efa56464SYan, Zheng 2847efa56464SYan, Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 2848efa56464SYan, Zheng num_bytes = end + 1 - start; 2849efa56464SYan, Zheng ret = btrfs_prealloc_file_range(inode, 0, start, 2850efa56464SYan, Zheng num_bytes, num_bytes, 2851efa56464SYan, Zheng end + 1, &alloc_hint); 2852efa56464SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 2853efa56464SYan, Zheng if (ret) 2854efa56464SYan, Zheng break; 2855efa56464SYan, Zheng nr++; 2856efa56464SYan, Zheng } 2857efa56464SYan, Zheng btrfs_free_reserved_data_space(inode, cluster->end + 2858efa56464SYan, Zheng 1 - cluster->start); 2859efa56464SYan, Zheng out: 2860efa56464SYan, Zheng mutex_unlock(&inode->i_mutex); 2861efa56464SYan, Zheng return ret; 2862efa56464SYan, Zheng } 2863efa56464SYan, Zheng 2864efa56464SYan, Zheng static noinline_for_stack 28650257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 28660257bb82SYan, Zheng u64 block_start) 28675d4f98a2SYan Zheng { 28685d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 28695d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 28705d4f98a2SYan Zheng struct extent_map *em; 28710257bb82SYan, Zheng int ret = 0; 28725d4f98a2SYan Zheng 2873172ddd60SDavid Sterba em = alloc_extent_map(); 28740257bb82SYan, Zheng if (!em) 28750257bb82SYan, Zheng return -ENOMEM; 28760257bb82SYan, Zheng 28775d4f98a2SYan Zheng em->start = start; 28780257bb82SYan, Zheng em->len = end + 1 - start; 28790257bb82SYan, Zheng em->block_len = em->len; 28800257bb82SYan, Zheng em->block_start = block_start; 28815d4f98a2SYan Zheng em->bdev = root->fs_info->fs_devices->latest_bdev; 28825d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 28835d4f98a2SYan Zheng 28845d4f98a2SYan Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 28855d4f98a2SYan Zheng while (1) { 2886890871beSChris Mason write_lock(&em_tree->lock); 28875d4f98a2SYan Zheng ret = add_extent_mapping(em_tree, em); 2888890871beSChris Mason write_unlock(&em_tree->lock); 28895d4f98a2SYan Zheng if (ret != -EEXIST) { 28905d4f98a2SYan Zheng free_extent_map(em); 28915d4f98a2SYan Zheng break; 28925d4f98a2SYan Zheng } 28935d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 0); 28945d4f98a2SYan Zheng } 28955d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 28960257bb82SYan, Zheng return ret; 28970257bb82SYan, Zheng } 28985d4f98a2SYan Zheng 28990257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 29000257bb82SYan, Zheng struct file_extent_cluster *cluster) 29010257bb82SYan, Zheng { 29020257bb82SYan, Zheng u64 page_start; 29030257bb82SYan, Zheng u64 page_end; 29040257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 29050257bb82SYan, Zheng unsigned long index; 29060257bb82SYan, Zheng unsigned long last_index; 29070257bb82SYan, Zheng struct page *page; 29080257bb82SYan, Zheng struct file_ra_state *ra; 29090257bb82SYan, Zheng int nr = 0; 29100257bb82SYan, Zheng int ret = 0; 29110257bb82SYan, Zheng 29120257bb82SYan, Zheng if (!cluster->nr) 29130257bb82SYan, Zheng return 0; 29140257bb82SYan, Zheng 29150257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 29160257bb82SYan, Zheng if (!ra) 29170257bb82SYan, Zheng return -ENOMEM; 29180257bb82SYan, Zheng 2919efa56464SYan, Zheng ret = prealloc_file_extent_cluster(inode, cluster); 29200257bb82SYan, Zheng if (ret) 2921efa56464SYan, Zheng goto out; 29220257bb82SYan, Zheng 29230257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 29240257bb82SYan, Zheng 2925efa56464SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 2926efa56464SYan, Zheng cluster->end - offset, cluster->start); 2927efa56464SYan, Zheng if (ret) 2928efa56464SYan, Zheng goto out; 2929efa56464SYan, Zheng 2930efa56464SYan, Zheng index = (cluster->start - offset) >> PAGE_CACHE_SHIFT; 2931efa56464SYan, Zheng last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT; 29320257bb82SYan, Zheng while (index <= last_index) { 2933efa56464SYan, Zheng ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE); 2934efa56464SYan, Zheng if (ret) 2935efa56464SYan, Zheng goto out; 2936efa56464SYan, Zheng 29370257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 29380257bb82SYan, Zheng if (!page) { 29390257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 29400257bb82SYan, Zheng ra, NULL, index, 29410257bb82SYan, Zheng last_index + 1 - index); 29420257bb82SYan, Zheng page = grab_cache_page(inode->i_mapping, index); 29430257bb82SYan, Zheng if (!page) { 2944efa56464SYan, Zheng btrfs_delalloc_release_metadata(inode, 2945efa56464SYan, Zheng PAGE_CACHE_SIZE); 29460257bb82SYan, Zheng ret = -ENOMEM; 2947efa56464SYan, Zheng goto out; 29480257bb82SYan, Zheng } 29490257bb82SYan, Zheng } 29500257bb82SYan, Zheng 29510257bb82SYan, Zheng if (PageReadahead(page)) { 29520257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 29530257bb82SYan, Zheng ra, NULL, page, index, 29540257bb82SYan, Zheng last_index + 1 - index); 29550257bb82SYan, Zheng } 29560257bb82SYan, Zheng 29570257bb82SYan, Zheng if (!PageUptodate(page)) { 29580257bb82SYan, Zheng btrfs_readpage(NULL, page); 29590257bb82SYan, Zheng lock_page(page); 29600257bb82SYan, Zheng if (!PageUptodate(page)) { 29610257bb82SYan, Zheng unlock_page(page); 29620257bb82SYan, Zheng page_cache_release(page); 2963efa56464SYan, Zheng btrfs_delalloc_release_metadata(inode, 2964efa56464SYan, Zheng PAGE_CACHE_SIZE); 29650257bb82SYan, Zheng ret = -EIO; 2966efa56464SYan, Zheng goto out; 29670257bb82SYan, Zheng } 29680257bb82SYan, Zheng } 29690257bb82SYan, Zheng 29700257bb82SYan, Zheng page_start = (u64)page->index << PAGE_CACHE_SHIFT; 29710257bb82SYan, Zheng page_end = page_start + PAGE_CACHE_SIZE - 1; 29720257bb82SYan, Zheng 29730257bb82SYan, Zheng lock_extent(&BTRFS_I(inode)->io_tree, 29740257bb82SYan, Zheng page_start, page_end, GFP_NOFS); 29750257bb82SYan, Zheng 29760257bb82SYan, Zheng set_page_extent_mapped(page); 29770257bb82SYan, Zheng 29780257bb82SYan, Zheng if (nr < cluster->nr && 29790257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 29800257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 29810257bb82SYan, Zheng page_start, page_end, 29820257bb82SYan, Zheng EXTENT_BOUNDARY, GFP_NOFS); 29830257bb82SYan, Zheng nr++; 29840257bb82SYan, Zheng } 29850257bb82SYan, Zheng 2986efa56464SYan, Zheng btrfs_set_extent_delalloc(inode, page_start, page_end, NULL); 29870257bb82SYan, Zheng set_page_dirty(page); 29880257bb82SYan, Zheng 29890257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 29900257bb82SYan, Zheng page_start, page_end, GFP_NOFS); 29910257bb82SYan, Zheng unlock_page(page); 29920257bb82SYan, Zheng page_cache_release(page); 29930257bb82SYan, Zheng 29940257bb82SYan, Zheng index++; 2995efa56464SYan, Zheng balance_dirty_pages_ratelimited(inode->i_mapping); 2996efa56464SYan, Zheng btrfs_throttle(BTRFS_I(inode)->root); 29970257bb82SYan, Zheng } 29980257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 2999efa56464SYan, Zheng out: 30000257bb82SYan, Zheng kfree(ra); 30010257bb82SYan, Zheng return ret; 30020257bb82SYan, Zheng } 30030257bb82SYan, Zheng 30040257bb82SYan, Zheng static noinline_for_stack 30050257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 30060257bb82SYan, Zheng struct file_extent_cluster *cluster) 30070257bb82SYan, Zheng { 30080257bb82SYan, Zheng int ret; 30090257bb82SYan, Zheng 30100257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 30110257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30120257bb82SYan, Zheng if (ret) 30130257bb82SYan, Zheng return ret; 30140257bb82SYan, Zheng cluster->nr = 0; 30150257bb82SYan, Zheng } 30160257bb82SYan, Zheng 30170257bb82SYan, Zheng if (!cluster->nr) 30180257bb82SYan, Zheng cluster->start = extent_key->objectid; 30190257bb82SYan, Zheng else 30200257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 30210257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 30220257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 30230257bb82SYan, Zheng cluster->nr++; 30240257bb82SYan, Zheng 30250257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 30260257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 30270257bb82SYan, Zheng if (ret) 30280257bb82SYan, Zheng return ret; 30290257bb82SYan, Zheng cluster->nr = 0; 30300257bb82SYan, Zheng } 30310257bb82SYan, Zheng return 0; 30325d4f98a2SYan Zheng } 30335d4f98a2SYan Zheng 30345d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 30355d4f98a2SYan Zheng static int get_ref_objectid_v0(struct reloc_control *rc, 30365d4f98a2SYan Zheng struct btrfs_path *path, 30375d4f98a2SYan Zheng struct btrfs_key *extent_key, 30385d4f98a2SYan Zheng u64 *ref_objectid, int *path_change) 30395d4f98a2SYan Zheng { 30405d4f98a2SYan Zheng struct btrfs_key key; 30415d4f98a2SYan Zheng struct extent_buffer *leaf; 30425d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 30435d4f98a2SYan Zheng int ret; 30445d4f98a2SYan Zheng int slot; 30455d4f98a2SYan Zheng 30465d4f98a2SYan Zheng leaf = path->nodes[0]; 30475d4f98a2SYan Zheng slot = path->slots[0]; 30485d4f98a2SYan Zheng while (1) { 30495d4f98a2SYan Zheng if (slot >= btrfs_header_nritems(leaf)) { 30505d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 30515d4f98a2SYan Zheng if (ret < 0) 30525d4f98a2SYan Zheng return ret; 30535d4f98a2SYan Zheng BUG_ON(ret > 0); 30545d4f98a2SYan Zheng leaf = path->nodes[0]; 30555d4f98a2SYan Zheng slot = path->slots[0]; 30565d4f98a2SYan Zheng if (path_change) 30575d4f98a2SYan Zheng *path_change = 1; 30585d4f98a2SYan Zheng } 30595d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, slot); 30605d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 30615d4f98a2SYan Zheng return -ENOENT; 30625d4f98a2SYan Zheng 30635d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_REF_V0_KEY) { 30645d4f98a2SYan Zheng slot++; 30655d4f98a2SYan Zheng continue; 30665d4f98a2SYan Zheng } 30675d4f98a2SYan Zheng ref0 = btrfs_item_ptr(leaf, slot, 30685d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 30695d4f98a2SYan Zheng *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0); 30705d4f98a2SYan Zheng break; 30715d4f98a2SYan Zheng } 30725d4f98a2SYan Zheng return 0; 30735d4f98a2SYan Zheng } 30745d4f98a2SYan Zheng #endif 30755d4f98a2SYan Zheng 30765d4f98a2SYan Zheng /* 30775d4f98a2SYan Zheng * helper to add a tree block to the list. 30785d4f98a2SYan Zheng * the major work is getting the generation and level of the block 30795d4f98a2SYan Zheng */ 30805d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 30815d4f98a2SYan Zheng struct btrfs_key *extent_key, 30825d4f98a2SYan Zheng struct btrfs_path *path, 30835d4f98a2SYan Zheng struct rb_root *blocks) 30845d4f98a2SYan Zheng { 30855d4f98a2SYan Zheng struct extent_buffer *eb; 30865d4f98a2SYan Zheng struct btrfs_extent_item *ei; 30875d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 30885d4f98a2SYan Zheng struct tree_block *block; 30895d4f98a2SYan Zheng struct rb_node *rb_node; 30905d4f98a2SYan Zheng u32 item_size; 30915d4f98a2SYan Zheng int level = -1; 30925d4f98a2SYan Zheng int generation; 30935d4f98a2SYan Zheng 30945d4f98a2SYan Zheng eb = path->nodes[0]; 30955d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 30965d4f98a2SYan Zheng 30975d4f98a2SYan Zheng if (item_size >= sizeof(*ei) + sizeof(*bi)) { 30985d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 30995d4f98a2SYan Zheng struct btrfs_extent_item); 31005d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 31015d4f98a2SYan Zheng generation = btrfs_extent_generation(eb, ei); 31025d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 31035d4f98a2SYan Zheng } else { 31045d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 31055d4f98a2SYan Zheng u64 ref_owner; 31065d4f98a2SYan Zheng int ret; 31075d4f98a2SYan Zheng 31085d4f98a2SYan Zheng BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 31095d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, extent_key, 31105d4f98a2SYan Zheng &ref_owner, NULL); 3111411fc6bcSAndi Kleen if (ret < 0) 3112411fc6bcSAndi Kleen return ret; 31135d4f98a2SYan Zheng BUG_ON(ref_owner >= BTRFS_MAX_LEVEL); 31145d4f98a2SYan Zheng level = (int)ref_owner; 31155d4f98a2SYan Zheng /* FIXME: get real generation */ 31165d4f98a2SYan Zheng generation = 0; 31175d4f98a2SYan Zheng #else 31185d4f98a2SYan Zheng BUG(); 31195d4f98a2SYan Zheng #endif 31205d4f98a2SYan Zheng } 31215d4f98a2SYan Zheng 3122*b3b4aa74SDavid Sterba btrfs_release_path(path); 31235d4f98a2SYan Zheng 31245d4f98a2SYan Zheng BUG_ON(level == -1); 31255d4f98a2SYan Zheng 31265d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 31275d4f98a2SYan Zheng if (!block) 31285d4f98a2SYan Zheng return -ENOMEM; 31295d4f98a2SYan Zheng 31305d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 31315d4f98a2SYan Zheng block->key.objectid = extent_key->offset; 31325d4f98a2SYan Zheng block->key.offset = generation; 31335d4f98a2SYan Zheng block->level = level; 31345d4f98a2SYan Zheng block->key_ready = 0; 31355d4f98a2SYan Zheng 31365d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 31375d4f98a2SYan Zheng BUG_ON(rb_node); 31385d4f98a2SYan Zheng 31395d4f98a2SYan Zheng return 0; 31405d4f98a2SYan Zheng } 31415d4f98a2SYan Zheng 31425d4f98a2SYan Zheng /* 31435d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 31445d4f98a2SYan Zheng */ 31455d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 31465d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 31475d4f98a2SYan Zheng struct rb_root *blocks) 31485d4f98a2SYan Zheng { 31495d4f98a2SYan Zheng struct btrfs_path *path; 31505d4f98a2SYan Zheng struct btrfs_key key; 31515d4f98a2SYan Zheng int ret; 31525d4f98a2SYan Zheng 31535d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 31545d4f98a2SYan Zheng return 0; 31555d4f98a2SYan Zheng 31565d4f98a2SYan Zheng if (tree_search(blocks, bytenr)) 31575d4f98a2SYan Zheng return 0; 31585d4f98a2SYan Zheng 31595d4f98a2SYan Zheng path = btrfs_alloc_path(); 31605d4f98a2SYan Zheng if (!path) 31615d4f98a2SYan Zheng return -ENOMEM; 31625d4f98a2SYan Zheng 31635d4f98a2SYan Zheng key.objectid = bytenr; 31645d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31655d4f98a2SYan Zheng key.offset = blocksize; 31665d4f98a2SYan Zheng 31675d4f98a2SYan Zheng path->search_commit_root = 1; 31685d4f98a2SYan Zheng path->skip_locking = 1; 31695d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 31705d4f98a2SYan Zheng if (ret < 0) 31715d4f98a2SYan Zheng goto out; 31725d4f98a2SYan Zheng BUG_ON(ret); 31735d4f98a2SYan Zheng 31745d4f98a2SYan Zheng btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 31755d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 31765d4f98a2SYan Zheng out: 31775d4f98a2SYan Zheng btrfs_free_path(path); 31785d4f98a2SYan Zheng return ret; 31795d4f98a2SYan Zheng } 31805d4f98a2SYan Zheng 31815d4f98a2SYan Zheng /* 31825d4f98a2SYan Zheng * helper to check if the block use full backrefs for pointers in it 31835d4f98a2SYan Zheng */ 31845d4f98a2SYan Zheng static int block_use_full_backref(struct reloc_control *rc, 31855d4f98a2SYan Zheng struct extent_buffer *eb) 31865d4f98a2SYan Zheng { 31875d4f98a2SYan Zheng u64 flags; 31885d4f98a2SYan Zheng int ret; 31895d4f98a2SYan Zheng 31905d4f98a2SYan Zheng if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) || 31915d4f98a2SYan Zheng btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV) 31925d4f98a2SYan Zheng return 1; 31935d4f98a2SYan Zheng 31943fd0a558SYan, Zheng ret = btrfs_lookup_extent_info(NULL, rc->extent_root, 31953fd0a558SYan, Zheng eb->start, eb->len, NULL, &flags); 31965d4f98a2SYan Zheng BUG_ON(ret); 31975d4f98a2SYan Zheng 31985d4f98a2SYan Zheng if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) 31995d4f98a2SYan Zheng ret = 1; 32005d4f98a2SYan Zheng else 32015d4f98a2SYan Zheng ret = 0; 32025d4f98a2SYan Zheng return ret; 32035d4f98a2SYan Zheng } 32045d4f98a2SYan Zheng 32050af3d00bSJosef Bacik static int delete_block_group_cache(struct btrfs_fs_info *fs_info, 32060af3d00bSJosef Bacik struct inode *inode, u64 ino) 32070af3d00bSJosef Bacik { 32080af3d00bSJosef Bacik struct btrfs_key key; 32090af3d00bSJosef Bacik struct btrfs_path *path; 32100af3d00bSJosef Bacik struct btrfs_root *root = fs_info->tree_root; 32110af3d00bSJosef Bacik struct btrfs_trans_handle *trans; 32120af3d00bSJosef Bacik unsigned long nr; 32130af3d00bSJosef Bacik int ret = 0; 32140af3d00bSJosef Bacik 32150af3d00bSJosef Bacik if (inode) 32160af3d00bSJosef Bacik goto truncate; 32170af3d00bSJosef Bacik 32180af3d00bSJosef Bacik key.objectid = ino; 32190af3d00bSJosef Bacik key.type = BTRFS_INODE_ITEM_KEY; 32200af3d00bSJosef Bacik key.offset = 0; 32210af3d00bSJosef Bacik 32220af3d00bSJosef Bacik inode = btrfs_iget(fs_info->sb, &key, root, NULL); 3223c704005dSDavid Sterba if (IS_ERR_OR_NULL(inode) || is_bad_inode(inode)) { 32240af3d00bSJosef Bacik if (inode && !IS_ERR(inode)) 32250af3d00bSJosef Bacik iput(inode); 32260af3d00bSJosef Bacik return -ENOENT; 32270af3d00bSJosef Bacik } 32280af3d00bSJosef Bacik 32290af3d00bSJosef Bacik truncate: 32300af3d00bSJosef Bacik path = btrfs_alloc_path(); 32310af3d00bSJosef Bacik if (!path) { 32320af3d00bSJosef Bacik ret = -ENOMEM; 32330af3d00bSJosef Bacik goto out; 32340af3d00bSJosef Bacik } 32350af3d00bSJosef Bacik 32360af3d00bSJosef Bacik trans = btrfs_join_transaction(root, 0); 32370af3d00bSJosef Bacik if (IS_ERR(trans)) { 32380af3d00bSJosef Bacik btrfs_free_path(path); 32393612b495STsutomu Itoh ret = PTR_ERR(trans); 32400af3d00bSJosef Bacik goto out; 32410af3d00bSJosef Bacik } 32420af3d00bSJosef Bacik 32430af3d00bSJosef Bacik ret = btrfs_truncate_free_space_cache(root, trans, path, inode); 32440af3d00bSJosef Bacik 32450af3d00bSJosef Bacik btrfs_free_path(path); 32460af3d00bSJosef Bacik nr = trans->blocks_used; 32470af3d00bSJosef Bacik btrfs_end_transaction(trans, root); 32480af3d00bSJosef Bacik btrfs_btree_balance_dirty(root, nr); 32490af3d00bSJosef Bacik out: 32500af3d00bSJosef Bacik iput(inode); 32510af3d00bSJosef Bacik return ret; 32520af3d00bSJosef Bacik } 32530af3d00bSJosef Bacik 32545d4f98a2SYan Zheng /* 32555d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY 32565d4f98a2SYan Zheng * this function scans fs tree to find blocks reference the data extent 32575d4f98a2SYan Zheng */ 32585d4f98a2SYan Zheng static int find_data_references(struct reloc_control *rc, 32595d4f98a2SYan Zheng struct btrfs_key *extent_key, 32605d4f98a2SYan Zheng struct extent_buffer *leaf, 32615d4f98a2SYan Zheng struct btrfs_extent_data_ref *ref, 32625d4f98a2SYan Zheng struct rb_root *blocks) 32635d4f98a2SYan Zheng { 32645d4f98a2SYan Zheng struct btrfs_path *path; 32655d4f98a2SYan Zheng struct tree_block *block; 32665d4f98a2SYan Zheng struct btrfs_root *root; 32675d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 32685d4f98a2SYan Zheng struct rb_node *rb_node; 32695d4f98a2SYan Zheng struct btrfs_key key; 32705d4f98a2SYan Zheng u64 ref_root; 32715d4f98a2SYan Zheng u64 ref_objectid; 32725d4f98a2SYan Zheng u64 ref_offset; 32735d4f98a2SYan Zheng u32 ref_count; 32745d4f98a2SYan Zheng u32 nritems; 32755d4f98a2SYan Zheng int err = 0; 32765d4f98a2SYan Zheng int added = 0; 32775d4f98a2SYan Zheng int counted; 32785d4f98a2SYan Zheng int ret; 32795d4f98a2SYan Zheng 32805d4f98a2SYan Zheng ref_root = btrfs_extent_data_ref_root(leaf, ref); 32815d4f98a2SYan Zheng ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref); 32825d4f98a2SYan Zheng ref_offset = btrfs_extent_data_ref_offset(leaf, ref); 32835d4f98a2SYan Zheng ref_count = btrfs_extent_data_ref_count(leaf, ref); 32845d4f98a2SYan Zheng 32850af3d00bSJosef Bacik /* 32860af3d00bSJosef Bacik * This is an extent belonging to the free space cache, lets just delete 32870af3d00bSJosef Bacik * it and redo the search. 32880af3d00bSJosef Bacik */ 32890af3d00bSJosef Bacik if (ref_root == BTRFS_ROOT_TREE_OBJECTID) { 32900af3d00bSJosef Bacik ret = delete_block_group_cache(rc->extent_root->fs_info, 32910af3d00bSJosef Bacik NULL, ref_objectid); 32920af3d00bSJosef Bacik if (ret != -ENOENT) 32930af3d00bSJosef Bacik return ret; 32940af3d00bSJosef Bacik ret = 0; 32950af3d00bSJosef Bacik } 32960af3d00bSJosef Bacik 32970af3d00bSJosef Bacik path = btrfs_alloc_path(); 32980af3d00bSJosef Bacik if (!path) 32990af3d00bSJosef Bacik return -ENOMEM; 33000af3d00bSJosef Bacik 33015d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, ref_root); 33025d4f98a2SYan Zheng if (IS_ERR(root)) { 33035d4f98a2SYan Zheng err = PTR_ERR(root); 33045d4f98a2SYan Zheng goto out; 33055d4f98a2SYan Zheng } 33065d4f98a2SYan Zheng 33075d4f98a2SYan Zheng key.objectid = ref_objectid; 33085d4f98a2SYan Zheng key.offset = ref_offset; 33095d4f98a2SYan Zheng key.type = BTRFS_EXTENT_DATA_KEY; 33105d4f98a2SYan Zheng 33115d4f98a2SYan Zheng path->search_commit_root = 1; 33125d4f98a2SYan Zheng path->skip_locking = 1; 33135d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 33145d4f98a2SYan Zheng if (ret < 0) { 33155d4f98a2SYan Zheng err = ret; 33165d4f98a2SYan Zheng goto out; 33175d4f98a2SYan Zheng } 33185d4f98a2SYan Zheng 33195d4f98a2SYan Zheng leaf = path->nodes[0]; 33205d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 33215d4f98a2SYan Zheng /* 33225d4f98a2SYan Zheng * the references in tree blocks that use full backrefs 33235d4f98a2SYan Zheng * are not counted in 33245d4f98a2SYan Zheng */ 33255d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 33265d4f98a2SYan Zheng counted = 0; 33275d4f98a2SYan Zheng else 33285d4f98a2SYan Zheng counted = 1; 33295d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 33305d4f98a2SYan Zheng if (rb_node) { 33315d4f98a2SYan Zheng if (counted) 33325d4f98a2SYan Zheng added = 1; 33335d4f98a2SYan Zheng else 33345d4f98a2SYan Zheng path->slots[0] = nritems; 33355d4f98a2SYan Zheng } 33365d4f98a2SYan Zheng 33375d4f98a2SYan Zheng while (ref_count > 0) { 33385d4f98a2SYan Zheng while (path->slots[0] >= nritems) { 33395d4f98a2SYan Zheng ret = btrfs_next_leaf(root, path); 33405d4f98a2SYan Zheng if (ret < 0) { 33415d4f98a2SYan Zheng err = ret; 33425d4f98a2SYan Zheng goto out; 33435d4f98a2SYan Zheng } 33445d4f98a2SYan Zheng if (ret > 0) { 33455d4f98a2SYan Zheng WARN_ON(1); 33465d4f98a2SYan Zheng goto out; 33475d4f98a2SYan Zheng } 33485d4f98a2SYan Zheng 33495d4f98a2SYan Zheng leaf = path->nodes[0]; 33505d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 33515d4f98a2SYan Zheng added = 0; 33525d4f98a2SYan Zheng 33535d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 33545d4f98a2SYan Zheng counted = 0; 33555d4f98a2SYan Zheng else 33565d4f98a2SYan Zheng counted = 1; 33575d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 33585d4f98a2SYan Zheng if (rb_node) { 33595d4f98a2SYan Zheng if (counted) 33605d4f98a2SYan Zheng added = 1; 33615d4f98a2SYan Zheng else 33625d4f98a2SYan Zheng path->slots[0] = nritems; 33635d4f98a2SYan Zheng } 33645d4f98a2SYan Zheng } 33655d4f98a2SYan Zheng 33665d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 33675d4f98a2SYan Zheng if (key.objectid != ref_objectid || 33685d4f98a2SYan Zheng key.type != BTRFS_EXTENT_DATA_KEY) { 33695d4f98a2SYan Zheng WARN_ON(1); 33705d4f98a2SYan Zheng break; 33715d4f98a2SYan Zheng } 33725d4f98a2SYan Zheng 33735d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 33745d4f98a2SYan Zheng struct btrfs_file_extent_item); 33755d4f98a2SYan Zheng 33765d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 33775d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 33785d4f98a2SYan Zheng goto next; 33795d4f98a2SYan Zheng 33805d4f98a2SYan Zheng if (btrfs_file_extent_disk_bytenr(leaf, fi) != 33815d4f98a2SYan Zheng extent_key->objectid) 33825d4f98a2SYan Zheng goto next; 33835d4f98a2SYan Zheng 33845d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 33855d4f98a2SYan Zheng if (key.offset != ref_offset) 33865d4f98a2SYan Zheng goto next; 33875d4f98a2SYan Zheng 33885d4f98a2SYan Zheng if (counted) 33895d4f98a2SYan Zheng ref_count--; 33905d4f98a2SYan Zheng if (added) 33915d4f98a2SYan Zheng goto next; 33925d4f98a2SYan Zheng 33935d4f98a2SYan Zheng if (!tree_block_processed(leaf->start, leaf->len, rc)) { 33945d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 33955d4f98a2SYan Zheng if (!block) { 33965d4f98a2SYan Zheng err = -ENOMEM; 33975d4f98a2SYan Zheng break; 33985d4f98a2SYan Zheng } 33995d4f98a2SYan Zheng block->bytenr = leaf->start; 34005d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &block->key, 0); 34015d4f98a2SYan Zheng block->level = 0; 34025d4f98a2SYan Zheng block->key_ready = 1; 34035d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, 34045d4f98a2SYan Zheng &block->rb_node); 34055d4f98a2SYan Zheng BUG_ON(rb_node); 34065d4f98a2SYan Zheng } 34075d4f98a2SYan Zheng if (counted) 34085d4f98a2SYan Zheng added = 1; 34095d4f98a2SYan Zheng else 34105d4f98a2SYan Zheng path->slots[0] = nritems; 34115d4f98a2SYan Zheng next: 34125d4f98a2SYan Zheng path->slots[0]++; 34135d4f98a2SYan Zheng 34145d4f98a2SYan Zheng } 34155d4f98a2SYan Zheng out: 34165d4f98a2SYan Zheng btrfs_free_path(path); 34175d4f98a2SYan Zheng return err; 34185d4f98a2SYan Zheng } 34195d4f98a2SYan Zheng 34205d4f98a2SYan Zheng /* 34215d4f98a2SYan Zheng * hepler to find all tree blocks that reference a given data extent 34225d4f98a2SYan Zheng */ 34235d4f98a2SYan Zheng static noinline_for_stack 34245d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 34255d4f98a2SYan Zheng struct btrfs_key *extent_key, 34265d4f98a2SYan Zheng struct btrfs_path *path, 34275d4f98a2SYan Zheng struct rb_root *blocks) 34285d4f98a2SYan Zheng { 34295d4f98a2SYan Zheng struct btrfs_key key; 34305d4f98a2SYan Zheng struct extent_buffer *eb; 34315d4f98a2SYan Zheng struct btrfs_extent_data_ref *dref; 34325d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 34335d4f98a2SYan Zheng unsigned long ptr; 34345d4f98a2SYan Zheng unsigned long end; 34353fd0a558SYan, Zheng u32 blocksize = btrfs_level_size(rc->extent_root, 0); 34365d4f98a2SYan Zheng int ret; 34375d4f98a2SYan Zheng int err = 0; 34385d4f98a2SYan Zheng 34395d4f98a2SYan Zheng eb = path->nodes[0]; 34405d4f98a2SYan Zheng ptr = btrfs_item_ptr_offset(eb, path->slots[0]); 34415d4f98a2SYan Zheng end = ptr + btrfs_item_size_nr(eb, path->slots[0]); 34425d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 34435d4f98a2SYan Zheng if (ptr + sizeof(struct btrfs_extent_item_v0) == end) 34445d4f98a2SYan Zheng ptr = end; 34455d4f98a2SYan Zheng else 34465d4f98a2SYan Zheng #endif 34475d4f98a2SYan Zheng ptr += sizeof(struct btrfs_extent_item); 34485d4f98a2SYan Zheng 34495d4f98a2SYan Zheng while (ptr < end) { 34505d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 34515d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 34525d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 34535d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 34545d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 34555d4f98a2SYan Zheng blocks); 34565d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 34575d4f98a2SYan Zheng dref = (struct btrfs_extent_data_ref *)(&iref->offset); 34585d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 34595d4f98a2SYan Zheng eb, dref, blocks); 34605d4f98a2SYan Zheng } else { 34615d4f98a2SYan Zheng BUG(); 34625d4f98a2SYan Zheng } 34635d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 34645d4f98a2SYan Zheng } 34655d4f98a2SYan Zheng WARN_ON(ptr > end); 34665d4f98a2SYan Zheng 34675d4f98a2SYan Zheng while (1) { 34685d4f98a2SYan Zheng cond_resched(); 34695d4f98a2SYan Zheng eb = path->nodes[0]; 34705d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(eb)) { 34715d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 34725d4f98a2SYan Zheng if (ret < 0) { 34735d4f98a2SYan Zheng err = ret; 34745d4f98a2SYan Zheng break; 34755d4f98a2SYan Zheng } 34765d4f98a2SYan Zheng if (ret > 0) 34775d4f98a2SYan Zheng break; 34785d4f98a2SYan Zheng eb = path->nodes[0]; 34795d4f98a2SYan Zheng } 34805d4f98a2SYan Zheng 34815d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path->slots[0]); 34825d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 34835d4f98a2SYan Zheng break; 34845d4f98a2SYan Zheng 34855d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 34865d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY || 34875d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 34885d4f98a2SYan Zheng #else 34895d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 34905d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 34915d4f98a2SYan Zheng #endif 34925d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 34935d4f98a2SYan Zheng blocks); 34945d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 34955d4f98a2SYan Zheng dref = btrfs_item_ptr(eb, path->slots[0], 34965d4f98a2SYan Zheng struct btrfs_extent_data_ref); 34975d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 34985d4f98a2SYan Zheng eb, dref, blocks); 34995d4f98a2SYan Zheng } else { 35005d4f98a2SYan Zheng ret = 0; 35015d4f98a2SYan Zheng } 35025d4f98a2SYan Zheng if (ret) { 35035d4f98a2SYan Zheng err = ret; 35045d4f98a2SYan Zheng break; 35055d4f98a2SYan Zheng } 35065d4f98a2SYan Zheng path->slots[0]++; 35075d4f98a2SYan Zheng } 3508*b3b4aa74SDavid Sterba btrfs_release_path(path); 35095d4f98a2SYan Zheng if (err) 35105d4f98a2SYan Zheng free_block_list(blocks); 35115d4f98a2SYan Zheng return err; 35125d4f98a2SYan Zheng } 35135d4f98a2SYan Zheng 35145d4f98a2SYan Zheng /* 35155d4f98a2SYan Zheng * hepler to find next unprocessed extent 35165d4f98a2SYan Zheng */ 35175d4f98a2SYan Zheng static noinline_for_stack 35185d4f98a2SYan Zheng int find_next_extent(struct btrfs_trans_handle *trans, 35193fd0a558SYan, Zheng struct reloc_control *rc, struct btrfs_path *path, 35203fd0a558SYan, Zheng struct btrfs_key *extent_key) 35215d4f98a2SYan Zheng { 35225d4f98a2SYan Zheng struct btrfs_key key; 35235d4f98a2SYan Zheng struct extent_buffer *leaf; 35245d4f98a2SYan Zheng u64 start, end, last; 35255d4f98a2SYan Zheng int ret; 35265d4f98a2SYan Zheng 35275d4f98a2SYan Zheng last = rc->block_group->key.objectid + rc->block_group->key.offset; 35285d4f98a2SYan Zheng while (1) { 35295d4f98a2SYan Zheng cond_resched(); 35305d4f98a2SYan Zheng if (rc->search_start >= last) { 35315d4f98a2SYan Zheng ret = 1; 35325d4f98a2SYan Zheng break; 35335d4f98a2SYan Zheng } 35345d4f98a2SYan Zheng 35355d4f98a2SYan Zheng key.objectid = rc->search_start; 35365d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 35375d4f98a2SYan Zheng key.offset = 0; 35385d4f98a2SYan Zheng 35395d4f98a2SYan Zheng path->search_commit_root = 1; 35405d4f98a2SYan Zheng path->skip_locking = 1; 35415d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 35425d4f98a2SYan Zheng 0, 0); 35435d4f98a2SYan Zheng if (ret < 0) 35445d4f98a2SYan Zheng break; 35455d4f98a2SYan Zheng next: 35465d4f98a2SYan Zheng leaf = path->nodes[0]; 35475d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 35485d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 35495d4f98a2SYan Zheng if (ret != 0) 35505d4f98a2SYan Zheng break; 35515d4f98a2SYan Zheng leaf = path->nodes[0]; 35525d4f98a2SYan Zheng } 35535d4f98a2SYan Zheng 35545d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 35555d4f98a2SYan Zheng if (key.objectid >= last) { 35565d4f98a2SYan Zheng ret = 1; 35575d4f98a2SYan Zheng break; 35585d4f98a2SYan Zheng } 35595d4f98a2SYan Zheng 35605d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_ITEM_KEY || 35615d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 35625d4f98a2SYan Zheng path->slots[0]++; 35635d4f98a2SYan Zheng goto next; 35645d4f98a2SYan Zheng } 35655d4f98a2SYan Zheng 35665d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 35675d4f98a2SYan Zheng key.objectid, &start, &end, 35685d4f98a2SYan Zheng EXTENT_DIRTY); 35695d4f98a2SYan Zheng 35705d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 3571*b3b4aa74SDavid Sterba btrfs_release_path(path); 35725d4f98a2SYan Zheng rc->search_start = end + 1; 35735d4f98a2SYan Zheng } else { 35745d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 35753fd0a558SYan, Zheng memcpy(extent_key, &key, sizeof(key)); 35765d4f98a2SYan Zheng return 0; 35775d4f98a2SYan Zheng } 35785d4f98a2SYan Zheng } 3579*b3b4aa74SDavid Sterba btrfs_release_path(path); 35805d4f98a2SYan Zheng return ret; 35815d4f98a2SYan Zheng } 35825d4f98a2SYan Zheng 35835d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 35845d4f98a2SYan Zheng { 35855d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 35865d4f98a2SYan Zheng mutex_lock(&fs_info->trans_mutex); 35875d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 35885d4f98a2SYan Zheng mutex_unlock(&fs_info->trans_mutex); 35895d4f98a2SYan Zheng } 35905d4f98a2SYan Zheng 35915d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 35925d4f98a2SYan Zheng { 35935d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 35945d4f98a2SYan Zheng mutex_lock(&fs_info->trans_mutex); 35955d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 35965d4f98a2SYan Zheng mutex_unlock(&fs_info->trans_mutex); 35975d4f98a2SYan Zheng } 35985d4f98a2SYan Zheng 35995d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 36005d4f98a2SYan Zheng { 36015d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 36025d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 36035d4f98a2SYan Zheng return 1; 36045d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 36055d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 36065d4f98a2SYan Zheng return 1; 36075d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 36085d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 36095d4f98a2SYan Zheng return 1; 36105d4f98a2SYan Zheng return 0; 36115d4f98a2SYan Zheng } 36125d4f98a2SYan Zheng 36133fd0a558SYan, Zheng static noinline_for_stack 36143fd0a558SYan, Zheng int prepare_to_relocate(struct reloc_control *rc) 36153fd0a558SYan, Zheng { 36163fd0a558SYan, Zheng struct btrfs_trans_handle *trans; 36173fd0a558SYan, Zheng int ret; 36183fd0a558SYan, Zheng 36193fd0a558SYan, Zheng rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root); 36203fd0a558SYan, Zheng if (!rc->block_rsv) 36213fd0a558SYan, Zheng return -ENOMEM; 36223fd0a558SYan, Zheng 36233fd0a558SYan, Zheng /* 36243fd0a558SYan, Zheng * reserve some space for creating reloc trees. 36253fd0a558SYan, Zheng * btrfs_init_reloc_root will use them when there 36263fd0a558SYan, Zheng * is no reservation in transaction handle. 36273fd0a558SYan, Zheng */ 36283fd0a558SYan, Zheng ret = btrfs_block_rsv_add(NULL, rc->extent_root, rc->block_rsv, 36298bb8ab2eSJosef Bacik rc->extent_root->nodesize * 256); 36303fd0a558SYan, Zheng if (ret) 36313fd0a558SYan, Zheng return ret; 36323fd0a558SYan, Zheng 36333fd0a558SYan, Zheng rc->block_rsv->refill_used = 1; 36343fd0a558SYan, Zheng btrfs_add_durable_block_rsv(rc->extent_root->fs_info, rc->block_rsv); 36353fd0a558SYan, Zheng 36363fd0a558SYan, Zheng memset(&rc->cluster, 0, sizeof(rc->cluster)); 36373fd0a558SYan, Zheng rc->search_start = rc->block_group->key.objectid; 36383fd0a558SYan, Zheng rc->extents_found = 0; 36393fd0a558SYan, Zheng rc->nodes_relocated = 0; 36403fd0a558SYan, Zheng rc->merging_rsv_size = 0; 36413fd0a558SYan, Zheng 36423fd0a558SYan, Zheng rc->create_reloc_tree = 1; 36433fd0a558SYan, Zheng set_reloc_control(rc); 36443fd0a558SYan, Zheng 36453fd0a558SYan, Zheng trans = btrfs_join_transaction(rc->extent_root, 1); 36463612b495STsutomu Itoh BUG_ON(IS_ERR(trans)); 36473fd0a558SYan, Zheng btrfs_commit_transaction(trans, rc->extent_root); 36483fd0a558SYan, Zheng return 0; 36493fd0a558SYan, Zheng } 365076dda93cSYan, Zheng 36515d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 36525d4f98a2SYan Zheng { 36535d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 36545d4f98a2SYan Zheng struct btrfs_key key; 36555d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 36565d4f98a2SYan Zheng struct btrfs_path *path; 36575d4f98a2SYan Zheng struct btrfs_extent_item *ei; 36585d4f98a2SYan Zheng unsigned long nr; 36595d4f98a2SYan Zheng u64 flags; 36605d4f98a2SYan Zheng u32 item_size; 36615d4f98a2SYan Zheng int ret; 36625d4f98a2SYan Zheng int err = 0; 3663c87f08caSChris Mason int progress = 0; 36645d4f98a2SYan Zheng 36655d4f98a2SYan Zheng path = btrfs_alloc_path(); 36663fd0a558SYan, Zheng if (!path) 36675d4f98a2SYan Zheng return -ENOMEM; 36683fd0a558SYan, Zheng 36693fd0a558SYan, Zheng ret = prepare_to_relocate(rc); 36703fd0a558SYan, Zheng if (ret) { 36713fd0a558SYan, Zheng err = ret; 36723fd0a558SYan, Zheng goto out_free; 36732423fdfbSJiri Slaby } 36745d4f98a2SYan Zheng 36755d4f98a2SYan Zheng while (1) { 3676c87f08caSChris Mason progress++; 3677a22285a6SYan, Zheng trans = btrfs_start_transaction(rc->extent_root, 0); 367898d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 3679c87f08caSChris Mason restart: 36803fd0a558SYan, Zheng if (update_backref_cache(trans, &rc->backref_cache)) { 36813fd0a558SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 36823fd0a558SYan, Zheng continue; 36833fd0a558SYan, Zheng } 36843fd0a558SYan, Zheng 36853fd0a558SYan, Zheng ret = find_next_extent(trans, rc, path, &key); 36865d4f98a2SYan Zheng if (ret < 0) 36875d4f98a2SYan Zheng err = ret; 36885d4f98a2SYan Zheng if (ret != 0) 36895d4f98a2SYan Zheng break; 36905d4f98a2SYan Zheng 36915d4f98a2SYan Zheng rc->extents_found++; 36925d4f98a2SYan Zheng 36935d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 36945d4f98a2SYan Zheng struct btrfs_extent_item); 36953fd0a558SYan, Zheng item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]); 36965d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 36975d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 36985d4f98a2SYan Zheng ret = check_extent_flags(flags); 36995d4f98a2SYan Zheng BUG_ON(ret); 37005d4f98a2SYan Zheng 37015d4f98a2SYan Zheng } else { 37025d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 37035d4f98a2SYan Zheng u64 ref_owner; 37045d4f98a2SYan Zheng int path_change = 0; 37055d4f98a2SYan Zheng 37065d4f98a2SYan Zheng BUG_ON(item_size != 37075d4f98a2SYan Zheng sizeof(struct btrfs_extent_item_v0)); 37085d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, &key, &ref_owner, 37095d4f98a2SYan Zheng &path_change); 37105d4f98a2SYan Zheng if (ref_owner < BTRFS_FIRST_FREE_OBJECTID) 37115d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_TREE_BLOCK; 37125d4f98a2SYan Zheng else 37135d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_DATA; 37145d4f98a2SYan Zheng 37155d4f98a2SYan Zheng if (path_change) { 3716*b3b4aa74SDavid Sterba btrfs_release_path(path); 37175d4f98a2SYan Zheng 37185d4f98a2SYan Zheng path->search_commit_root = 1; 37195d4f98a2SYan Zheng path->skip_locking = 1; 37205d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, 37215d4f98a2SYan Zheng &key, path, 0, 0); 37225d4f98a2SYan Zheng if (ret < 0) { 37235d4f98a2SYan Zheng err = ret; 37245d4f98a2SYan Zheng break; 37255d4f98a2SYan Zheng } 37265d4f98a2SYan Zheng BUG_ON(ret > 0); 37275d4f98a2SYan Zheng } 37285d4f98a2SYan Zheng #else 37295d4f98a2SYan Zheng BUG(); 37305d4f98a2SYan Zheng #endif 37315d4f98a2SYan Zheng } 37325d4f98a2SYan Zheng 37335d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 37345d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 37355d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 37365d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 37375d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 37385d4f98a2SYan Zheng } else { 3739*b3b4aa74SDavid Sterba btrfs_release_path(path); 37405d4f98a2SYan Zheng ret = 0; 37415d4f98a2SYan Zheng } 37425d4f98a2SYan Zheng if (ret < 0) { 37433fd0a558SYan, Zheng err = ret; 37445d4f98a2SYan Zheng break; 37455d4f98a2SYan Zheng } 37465d4f98a2SYan Zheng 37475d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 37485d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 37495d4f98a2SYan Zheng if (ret < 0) { 37503fd0a558SYan, Zheng if (ret != -EAGAIN) { 37515d4f98a2SYan Zheng err = ret; 37525d4f98a2SYan Zheng break; 37535d4f98a2SYan Zheng } 37543fd0a558SYan, Zheng rc->extents_found--; 37553fd0a558SYan, Zheng rc->search_start = key.objectid; 37563fd0a558SYan, Zheng } 37575d4f98a2SYan Zheng } 37585d4f98a2SYan Zheng 37593fd0a558SYan, Zheng ret = btrfs_block_rsv_check(trans, rc->extent_root, 37603fd0a558SYan, Zheng rc->block_rsv, 0, 5); 37613fd0a558SYan, Zheng if (ret < 0) { 37623fd0a558SYan, Zheng if (ret != -EAGAIN) { 37633fd0a558SYan, Zheng err = ret; 37643fd0a558SYan, Zheng WARN_ON(1); 37653fd0a558SYan, Zheng break; 37663fd0a558SYan, Zheng } 37673fd0a558SYan, Zheng rc->commit_transaction = 1; 37683fd0a558SYan, Zheng } 37693fd0a558SYan, Zheng 37703fd0a558SYan, Zheng if (rc->commit_transaction) { 37713fd0a558SYan, Zheng rc->commit_transaction = 0; 37723fd0a558SYan, Zheng ret = btrfs_commit_transaction(trans, rc->extent_root); 37733fd0a558SYan, Zheng BUG_ON(ret); 37743fd0a558SYan, Zheng } else { 37755d4f98a2SYan Zheng nr = trans->blocks_used; 37763fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 37775d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 37783fd0a558SYan, Zheng } 37793fd0a558SYan, Zheng trans = NULL; 37805d4f98a2SYan Zheng 37815d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 37825d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 37835d4f98a2SYan Zheng rc->found_file_extent = 1; 37840257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 37853fd0a558SYan, Zheng &key, &rc->cluster); 37865d4f98a2SYan Zheng if (ret < 0) { 37875d4f98a2SYan Zheng err = ret; 37885d4f98a2SYan Zheng break; 37895d4f98a2SYan Zheng } 37905d4f98a2SYan Zheng } 37915d4f98a2SYan Zheng } 3792c87f08caSChris Mason if (trans && progress && err == -ENOSPC) { 3793c87f08caSChris Mason ret = btrfs_force_chunk_alloc(trans, rc->extent_root, 3794c87f08caSChris Mason rc->block_group->flags); 3795c87f08caSChris Mason if (ret == 0) { 3796c87f08caSChris Mason err = 0; 3797c87f08caSChris Mason progress = 0; 3798c87f08caSChris Mason goto restart; 3799c87f08caSChris Mason } 3800c87f08caSChris Mason } 38013fd0a558SYan, Zheng 3802*b3b4aa74SDavid Sterba btrfs_release_path(path); 38033fd0a558SYan, Zheng clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY, 38043fd0a558SYan, Zheng GFP_NOFS); 38055d4f98a2SYan Zheng 38065d4f98a2SYan Zheng if (trans) { 38075d4f98a2SYan Zheng nr = trans->blocks_used; 38083fd0a558SYan, Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 38095d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 38105d4f98a2SYan Zheng } 38115d4f98a2SYan Zheng 38120257bb82SYan, Zheng if (!err) { 38133fd0a558SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, 38143fd0a558SYan, Zheng &rc->cluster); 38150257bb82SYan, Zheng if (ret < 0) 38160257bb82SYan, Zheng err = ret; 38170257bb82SYan, Zheng } 38180257bb82SYan, Zheng 38193fd0a558SYan, Zheng rc->create_reloc_tree = 0; 38203fd0a558SYan, Zheng set_reloc_control(rc); 38210257bb82SYan, Zheng 38223fd0a558SYan, Zheng backref_cache_cleanup(&rc->backref_cache); 38233fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1); 38245d4f98a2SYan Zheng 38253fd0a558SYan, Zheng err = prepare_to_merge(rc, err); 38265d4f98a2SYan Zheng 38275d4f98a2SYan Zheng merge_reloc_roots(rc); 38285d4f98a2SYan Zheng 38293fd0a558SYan, Zheng rc->merge_reloc_tree = 0; 38305d4f98a2SYan Zheng unset_reloc_control(rc); 38313fd0a558SYan, Zheng btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1); 38325d4f98a2SYan Zheng 38335d4f98a2SYan Zheng /* get rid of pinned extents */ 3834a22285a6SYan, Zheng trans = btrfs_join_transaction(rc->extent_root, 1); 38353612b495STsutomu Itoh if (IS_ERR(trans)) 38363612b495STsutomu Itoh err = PTR_ERR(trans); 38373612b495STsutomu Itoh else 38385d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 38393fd0a558SYan, Zheng out_free: 38403fd0a558SYan, Zheng btrfs_free_block_rsv(rc->extent_root, rc->block_rsv); 38413fd0a558SYan, Zheng btrfs_free_path(path); 38425d4f98a2SYan Zheng return err; 38435d4f98a2SYan Zheng } 38445d4f98a2SYan Zheng 38455d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 38460257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 38475d4f98a2SYan Zheng { 38485d4f98a2SYan Zheng struct btrfs_path *path; 38495d4f98a2SYan Zheng struct btrfs_inode_item *item; 38505d4f98a2SYan Zheng struct extent_buffer *leaf; 38515d4f98a2SYan Zheng int ret; 38525d4f98a2SYan Zheng 38535d4f98a2SYan Zheng path = btrfs_alloc_path(); 38545d4f98a2SYan Zheng if (!path) 38555d4f98a2SYan Zheng return -ENOMEM; 38565d4f98a2SYan Zheng 38575d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 38585d4f98a2SYan Zheng if (ret) 38595d4f98a2SYan Zheng goto out; 38605d4f98a2SYan Zheng 38615d4f98a2SYan Zheng leaf = path->nodes[0]; 38625d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 38635d4f98a2SYan Zheng memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item)); 38645d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 38650257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 38665d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 38673fd0a558SYan, Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS | 38683fd0a558SYan, Zheng BTRFS_INODE_PREALLOC); 38695d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 3870*b3b4aa74SDavid Sterba btrfs_release_path(path); 38715d4f98a2SYan Zheng out: 38725d4f98a2SYan Zheng btrfs_free_path(path); 38735d4f98a2SYan Zheng return ret; 38745d4f98a2SYan Zheng } 38755d4f98a2SYan Zheng 38765d4f98a2SYan Zheng /* 38775d4f98a2SYan Zheng * helper to create inode for data relocation. 38785d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 38795d4f98a2SYan Zheng */ 38803fd0a558SYan, Zheng static noinline_for_stack 38813fd0a558SYan, Zheng struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 38825d4f98a2SYan Zheng struct btrfs_block_group_cache *group) 38835d4f98a2SYan Zheng { 38845d4f98a2SYan Zheng struct inode *inode = NULL; 38855d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 38865d4f98a2SYan Zheng struct btrfs_root *root; 38875d4f98a2SYan Zheng struct btrfs_key key; 38885d4f98a2SYan Zheng unsigned long nr; 38895d4f98a2SYan Zheng u64 objectid = BTRFS_FIRST_FREE_OBJECTID; 38905d4f98a2SYan Zheng int err = 0; 38915d4f98a2SYan Zheng 38925d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 38935d4f98a2SYan Zheng if (IS_ERR(root)) 38945d4f98a2SYan Zheng return ERR_CAST(root); 38955d4f98a2SYan Zheng 3896a22285a6SYan, Zheng trans = btrfs_start_transaction(root, 6); 38973fd0a558SYan, Zheng if (IS_ERR(trans)) 38983fd0a558SYan, Zheng return ERR_CAST(trans); 38995d4f98a2SYan Zheng 39005d4f98a2SYan Zheng err = btrfs_find_free_objectid(trans, root, objectid, &objectid); 39015d4f98a2SYan Zheng if (err) 39025d4f98a2SYan Zheng goto out; 39035d4f98a2SYan Zheng 39040257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 39055d4f98a2SYan Zheng BUG_ON(err); 39065d4f98a2SYan Zheng 39075d4f98a2SYan Zheng key.objectid = objectid; 39085d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 39095d4f98a2SYan Zheng key.offset = 0; 391073f73415SJosef Bacik inode = btrfs_iget(root->fs_info->sb, &key, root, NULL); 39115d4f98a2SYan Zheng BUG_ON(IS_ERR(inode) || is_bad_inode(inode)); 39125d4f98a2SYan Zheng BTRFS_I(inode)->index_cnt = group->key.objectid; 39135d4f98a2SYan Zheng 39145d4f98a2SYan Zheng err = btrfs_orphan_add(trans, inode); 39155d4f98a2SYan Zheng out: 39165d4f98a2SYan Zheng nr = trans->blocks_used; 39175d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 39185d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 39195d4f98a2SYan Zheng if (err) { 39205d4f98a2SYan Zheng if (inode) 39215d4f98a2SYan Zheng iput(inode); 39225d4f98a2SYan Zheng inode = ERR_PTR(err); 39235d4f98a2SYan Zheng } 39245d4f98a2SYan Zheng return inode; 39255d4f98a2SYan Zheng } 39265d4f98a2SYan Zheng 39273fd0a558SYan, Zheng static struct reloc_control *alloc_reloc_control(void) 39283fd0a558SYan, Zheng { 39293fd0a558SYan, Zheng struct reloc_control *rc; 39303fd0a558SYan, Zheng 39313fd0a558SYan, Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 39323fd0a558SYan, Zheng if (!rc) 39333fd0a558SYan, Zheng return NULL; 39343fd0a558SYan, Zheng 39353fd0a558SYan, Zheng INIT_LIST_HEAD(&rc->reloc_roots); 39363fd0a558SYan, Zheng backref_cache_init(&rc->backref_cache); 39373fd0a558SYan, Zheng mapping_tree_init(&rc->reloc_root_tree); 3938f993c883SDavid Sterba extent_io_tree_init(&rc->processed_blocks, NULL); 39393fd0a558SYan, Zheng return rc; 39403fd0a558SYan, Zheng } 39413fd0a558SYan, Zheng 39425d4f98a2SYan Zheng /* 39435d4f98a2SYan Zheng * function to relocate all extents in a block group. 39445d4f98a2SYan Zheng */ 39455d4f98a2SYan Zheng int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start) 39465d4f98a2SYan Zheng { 39475d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = extent_root->fs_info; 39485d4f98a2SYan Zheng struct reloc_control *rc; 39490af3d00bSJosef Bacik struct inode *inode; 39500af3d00bSJosef Bacik struct btrfs_path *path; 39515d4f98a2SYan Zheng int ret; 3952f0486c68SYan, Zheng int rw = 0; 39535d4f98a2SYan Zheng int err = 0; 39545d4f98a2SYan Zheng 39553fd0a558SYan, Zheng rc = alloc_reloc_control(); 39565d4f98a2SYan Zheng if (!rc) 39575d4f98a2SYan Zheng return -ENOMEM; 39585d4f98a2SYan Zheng 3959f0486c68SYan, Zheng rc->extent_root = extent_root; 39603fd0a558SYan, Zheng 39615d4f98a2SYan Zheng rc->block_group = btrfs_lookup_block_group(fs_info, group_start); 39625d4f98a2SYan Zheng BUG_ON(!rc->block_group); 39635d4f98a2SYan Zheng 3964f0486c68SYan, Zheng if (!rc->block_group->ro) { 3965f0486c68SYan, Zheng ret = btrfs_set_block_group_ro(extent_root, rc->block_group); 3966f0486c68SYan, Zheng if (ret) { 3967f0486c68SYan, Zheng err = ret; 3968f0486c68SYan, Zheng goto out; 3969f0486c68SYan, Zheng } 3970f0486c68SYan, Zheng rw = 1; 3971f0486c68SYan, Zheng } 3972f0486c68SYan, Zheng 39730af3d00bSJosef Bacik path = btrfs_alloc_path(); 39740af3d00bSJosef Bacik if (!path) { 39750af3d00bSJosef Bacik err = -ENOMEM; 39760af3d00bSJosef Bacik goto out; 39770af3d00bSJosef Bacik } 39780af3d00bSJosef Bacik 39790af3d00bSJosef Bacik inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group, 39800af3d00bSJosef Bacik path); 39810af3d00bSJosef Bacik btrfs_free_path(path); 39820af3d00bSJosef Bacik 39830af3d00bSJosef Bacik if (!IS_ERR(inode)) 39840af3d00bSJosef Bacik ret = delete_block_group_cache(fs_info, inode, 0); 39850af3d00bSJosef Bacik else 39860af3d00bSJosef Bacik ret = PTR_ERR(inode); 39870af3d00bSJosef Bacik 39880af3d00bSJosef Bacik if (ret && ret != -ENOENT) { 39890af3d00bSJosef Bacik err = ret; 39900af3d00bSJosef Bacik goto out; 39910af3d00bSJosef Bacik } 39920af3d00bSJosef Bacik 39935d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 39945d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 39955d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 39965d4f98a2SYan Zheng rc->data_inode = NULL; 39975d4f98a2SYan Zheng goto out; 39985d4f98a2SYan Zheng } 39995d4f98a2SYan Zheng 40005d4f98a2SYan Zheng printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n", 40015d4f98a2SYan Zheng (unsigned long long)rc->block_group->key.objectid, 40025d4f98a2SYan Zheng (unsigned long long)rc->block_group->flags); 40035d4f98a2SYan Zheng 400424bbcf04SYan, Zheng btrfs_start_delalloc_inodes(fs_info->tree_root, 0); 400524bbcf04SYan, Zheng btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0); 40065d4f98a2SYan Zheng 40075d4f98a2SYan Zheng while (1) { 400876dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 400976dda93cSYan, Zheng 401076dda93cSYan, Zheng btrfs_clean_old_snapshots(fs_info->tree_root); 40115d4f98a2SYan Zheng ret = relocate_block_group(rc); 401276dda93cSYan, Zheng 401376dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 40145d4f98a2SYan Zheng if (ret < 0) { 40155d4f98a2SYan Zheng err = ret; 40163fd0a558SYan, Zheng goto out; 40175d4f98a2SYan Zheng } 40185d4f98a2SYan Zheng 40195d4f98a2SYan Zheng if (rc->extents_found == 0) 40205d4f98a2SYan Zheng break; 40215d4f98a2SYan Zheng 40225d4f98a2SYan Zheng printk(KERN_INFO "btrfs: found %llu extents\n", 40235d4f98a2SYan Zheng (unsigned long long)rc->extents_found); 40245d4f98a2SYan Zheng 40255d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 40265d4f98a2SYan Zheng btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1); 40275d4f98a2SYan Zheng invalidate_mapping_pages(rc->data_inode->i_mapping, 40285d4f98a2SYan Zheng 0, -1); 40295d4f98a2SYan Zheng rc->stage = UPDATE_DATA_PTRS; 40305d4f98a2SYan Zheng } 40315d4f98a2SYan Zheng } 40325d4f98a2SYan Zheng 40330257bb82SYan, Zheng filemap_write_and_wait_range(fs_info->btree_inode->i_mapping, 40345d4f98a2SYan Zheng rc->block_group->key.objectid, 40355d4f98a2SYan Zheng rc->block_group->key.objectid + 40365d4f98a2SYan Zheng rc->block_group->key.offset - 1); 40375d4f98a2SYan Zheng 40385d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 40395d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 40405d4f98a2SYan Zheng WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0); 40415d4f98a2SYan Zheng out: 4042f0486c68SYan, Zheng if (err && rw) 4043f0486c68SYan, Zheng btrfs_set_block_group_rw(extent_root, rc->block_group); 40445d4f98a2SYan Zheng iput(rc->data_inode); 40455d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 40465d4f98a2SYan Zheng kfree(rc); 40475d4f98a2SYan Zheng return err; 40485d4f98a2SYan Zheng } 40495d4f98a2SYan Zheng 405076dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 405176dda93cSYan, Zheng { 405276dda93cSYan, Zheng struct btrfs_trans_handle *trans; 405376dda93cSYan, Zheng int ret; 405476dda93cSYan, Zheng 4055a22285a6SYan, Zheng trans = btrfs_start_transaction(root->fs_info->tree_root, 0); 405698d5dc13STsutomu Itoh BUG_ON(IS_ERR(trans)); 405776dda93cSYan, Zheng 405876dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 405976dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 406076dda93cSYan, Zheng root->root_item.drop_level = 0; 406176dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 406276dda93cSYan, Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 406376dda93cSYan, Zheng &root->root_key, &root->root_item); 406476dda93cSYan, Zheng BUG_ON(ret); 406576dda93cSYan, Zheng 406676dda93cSYan, Zheng ret = btrfs_end_transaction(trans, root->fs_info->tree_root); 406776dda93cSYan, Zheng BUG_ON(ret); 406876dda93cSYan, Zheng return 0; 406976dda93cSYan, Zheng } 407076dda93cSYan, Zheng 40715d4f98a2SYan Zheng /* 40725d4f98a2SYan Zheng * recover relocation interrupted by system crash. 40735d4f98a2SYan Zheng * 40745d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 40755d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 40765d4f98a2SYan Zheng */ 40775d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 40785d4f98a2SYan Zheng { 40795d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 40805d4f98a2SYan Zheng struct btrfs_key key; 40815d4f98a2SYan Zheng struct btrfs_root *fs_root; 40825d4f98a2SYan Zheng struct btrfs_root *reloc_root; 40835d4f98a2SYan Zheng struct btrfs_path *path; 40845d4f98a2SYan Zheng struct extent_buffer *leaf; 40855d4f98a2SYan Zheng struct reloc_control *rc = NULL; 40865d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 40875d4f98a2SYan Zheng int ret; 40885d4f98a2SYan Zheng int err = 0; 40895d4f98a2SYan Zheng 40905d4f98a2SYan Zheng path = btrfs_alloc_path(); 40915d4f98a2SYan Zheng if (!path) 40925d4f98a2SYan Zheng return -ENOMEM; 40935d4f98a2SYan Zheng 40945d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 40955d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 40965d4f98a2SYan Zheng key.offset = (u64)-1; 40975d4f98a2SYan Zheng 40985d4f98a2SYan Zheng while (1) { 40995d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, 41005d4f98a2SYan Zheng path, 0, 0); 41015d4f98a2SYan Zheng if (ret < 0) { 41025d4f98a2SYan Zheng err = ret; 41035d4f98a2SYan Zheng goto out; 41045d4f98a2SYan Zheng } 41055d4f98a2SYan Zheng if (ret > 0) { 41065d4f98a2SYan Zheng if (path->slots[0] == 0) 41075d4f98a2SYan Zheng break; 41085d4f98a2SYan Zheng path->slots[0]--; 41095d4f98a2SYan Zheng } 41105d4f98a2SYan Zheng leaf = path->nodes[0]; 41115d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 4112*b3b4aa74SDavid Sterba btrfs_release_path(path); 41135d4f98a2SYan Zheng 41145d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 41155d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 41165d4f98a2SYan Zheng break; 41175d4f98a2SYan Zheng 41185d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root, &key); 41195d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 41205d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 41215d4f98a2SYan Zheng goto out; 41225d4f98a2SYan Zheng } 41235d4f98a2SYan Zheng 41245d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 41255d4f98a2SYan Zheng 41265d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 41275d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 41285d4f98a2SYan Zheng reloc_root->root_key.offset); 41295d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 413076dda93cSYan, Zheng ret = PTR_ERR(fs_root); 413176dda93cSYan, Zheng if (ret != -ENOENT) { 413276dda93cSYan, Zheng err = ret; 41335d4f98a2SYan Zheng goto out; 41345d4f98a2SYan Zheng } 413576dda93cSYan, Zheng mark_garbage_root(reloc_root); 413676dda93cSYan, Zheng } 41375d4f98a2SYan Zheng } 41385d4f98a2SYan Zheng 41395d4f98a2SYan Zheng if (key.offset == 0) 41405d4f98a2SYan Zheng break; 41415d4f98a2SYan Zheng 41425d4f98a2SYan Zheng key.offset--; 41435d4f98a2SYan Zheng } 4144*b3b4aa74SDavid Sterba btrfs_release_path(path); 41455d4f98a2SYan Zheng 41465d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 41475d4f98a2SYan Zheng goto out; 41485d4f98a2SYan Zheng 41493fd0a558SYan, Zheng rc = alloc_reloc_control(); 41505d4f98a2SYan Zheng if (!rc) { 41515d4f98a2SYan Zheng err = -ENOMEM; 41525d4f98a2SYan Zheng goto out; 41535d4f98a2SYan Zheng } 41545d4f98a2SYan Zheng 41555d4f98a2SYan Zheng rc->extent_root = root->fs_info->extent_root; 41565d4f98a2SYan Zheng 41575d4f98a2SYan Zheng set_reloc_control(rc); 41585d4f98a2SYan Zheng 41593fd0a558SYan, Zheng trans = btrfs_join_transaction(rc->extent_root, 1); 41603612b495STsutomu Itoh if (IS_ERR(trans)) { 41613612b495STsutomu Itoh unset_reloc_control(rc); 41623612b495STsutomu Itoh err = PTR_ERR(trans); 41633612b495STsutomu Itoh goto out_free; 41643612b495STsutomu Itoh } 41653fd0a558SYan, Zheng 41663fd0a558SYan, Zheng rc->merge_reloc_tree = 1; 41673fd0a558SYan, Zheng 41685d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 41695d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 41705d4f98a2SYan Zheng struct btrfs_root, root_list); 41715d4f98a2SYan Zheng list_del(&reloc_root->root_list); 41725d4f98a2SYan Zheng 41735d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 41745d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 41755d4f98a2SYan Zheng &rc->reloc_roots); 41765d4f98a2SYan Zheng continue; 41775d4f98a2SYan Zheng } 41785d4f98a2SYan Zheng 41795d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 41805d4f98a2SYan Zheng reloc_root->root_key.offset); 41815d4f98a2SYan Zheng BUG_ON(IS_ERR(fs_root)); 41825d4f98a2SYan Zheng 41835d4f98a2SYan Zheng __add_reloc_root(reloc_root); 41845d4f98a2SYan Zheng fs_root->reloc_root = reloc_root; 41855d4f98a2SYan Zheng } 41865d4f98a2SYan Zheng 41875d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 41885d4f98a2SYan Zheng 41895d4f98a2SYan Zheng merge_reloc_roots(rc); 41905d4f98a2SYan Zheng 41915d4f98a2SYan Zheng unset_reloc_control(rc); 41925d4f98a2SYan Zheng 41933fd0a558SYan, Zheng trans = btrfs_join_transaction(rc->extent_root, 1); 41943612b495STsutomu Itoh if (IS_ERR(trans)) 41953612b495STsutomu Itoh err = PTR_ERR(trans); 41963612b495STsutomu Itoh else 41975d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 41983612b495STsutomu Itoh out_free: 41995d4f98a2SYan Zheng kfree(rc); 42003612b495STsutomu Itoh out: 42015d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 42025d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 42035d4f98a2SYan Zheng struct btrfs_root, root_list); 42045d4f98a2SYan Zheng list_del(&reloc_root->root_list); 42055d4f98a2SYan Zheng free_extent_buffer(reloc_root->node); 42065d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 42075d4f98a2SYan Zheng kfree(reloc_root); 42085d4f98a2SYan Zheng } 42095d4f98a2SYan Zheng btrfs_free_path(path); 42105d4f98a2SYan Zheng 42115d4f98a2SYan Zheng if (err == 0) { 42125d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 42135d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 42145d4f98a2SYan Zheng BTRFS_DATA_RELOC_TREE_OBJECTID); 42155d4f98a2SYan Zheng if (IS_ERR(fs_root)) 42165d4f98a2SYan Zheng err = PTR_ERR(fs_root); 4217d7ce5843SMiao Xie else 421866b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(fs_root); 42195d4f98a2SYan Zheng } 42205d4f98a2SYan Zheng return err; 42215d4f98a2SYan Zheng } 42225d4f98a2SYan Zheng 42235d4f98a2SYan Zheng /* 42245d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 42255d4f98a2SYan Zheng * 42265d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 42275d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 42285d4f98a2SYan Zheng */ 42295d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 42305d4f98a2SYan Zheng { 42315d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 42325d4f98a2SYan Zheng struct btrfs_sector_sum *sector_sum; 42335d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 42345d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 42355d4f98a2SYan Zheng size_t offset; 42365d4f98a2SYan Zheng int ret; 42375d4f98a2SYan Zheng u64 disk_bytenr; 42385d4f98a2SYan Zheng LIST_HEAD(list); 42395d4f98a2SYan Zheng 42405d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 42415d4f98a2SYan Zheng BUG_ON(ordered->file_offset != file_pos || ordered->len != len); 42425d4f98a2SYan Zheng 42435d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 42445d4f98a2SYan Zheng ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr, 42455d4f98a2SYan Zheng disk_bytenr + len - 1, &list); 42465d4f98a2SYan Zheng 42475d4f98a2SYan Zheng while (!list_empty(&list)) { 42485d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 42495d4f98a2SYan Zheng list_del_init(&sums->list); 42505d4f98a2SYan Zheng 42515d4f98a2SYan Zheng sector_sum = sums->sums; 42525d4f98a2SYan Zheng sums->bytenr = ordered->start; 42535d4f98a2SYan Zheng 42545d4f98a2SYan Zheng offset = 0; 42555d4f98a2SYan Zheng while (offset < sums->len) { 42565d4f98a2SYan Zheng sector_sum->bytenr += ordered->start - disk_bytenr; 42575d4f98a2SYan Zheng sector_sum++; 42585d4f98a2SYan Zheng offset += root->sectorsize; 42595d4f98a2SYan Zheng } 42605d4f98a2SYan Zheng 42615d4f98a2SYan Zheng btrfs_add_ordered_sum(inode, ordered, sums); 42625d4f98a2SYan Zheng } 42635d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 4264411fc6bcSAndi Kleen return ret; 42655d4f98a2SYan Zheng } 42663fd0a558SYan, Zheng 42673fd0a558SYan, Zheng void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans, 42683fd0a558SYan, Zheng struct btrfs_root *root, struct extent_buffer *buf, 42693fd0a558SYan, Zheng struct extent_buffer *cow) 42703fd0a558SYan, Zheng { 42713fd0a558SYan, Zheng struct reloc_control *rc; 42723fd0a558SYan, Zheng struct backref_node *node; 42733fd0a558SYan, Zheng int first_cow = 0; 42743fd0a558SYan, Zheng int level; 42753fd0a558SYan, Zheng int ret; 42763fd0a558SYan, Zheng 42773fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 42783fd0a558SYan, Zheng if (!rc) 42793fd0a558SYan, Zheng return; 42803fd0a558SYan, Zheng 42813fd0a558SYan, Zheng BUG_ON(rc->stage == UPDATE_DATA_PTRS && 42823fd0a558SYan, Zheng root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID); 42833fd0a558SYan, Zheng 42843fd0a558SYan, Zheng level = btrfs_header_level(buf); 42853fd0a558SYan, Zheng if (btrfs_header_generation(buf) <= 42863fd0a558SYan, Zheng btrfs_root_last_snapshot(&root->root_item)) 42873fd0a558SYan, Zheng first_cow = 1; 42883fd0a558SYan, Zheng 42893fd0a558SYan, Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID && 42903fd0a558SYan, Zheng rc->create_reloc_tree) { 42913fd0a558SYan, Zheng WARN_ON(!first_cow && level == 0); 42923fd0a558SYan, Zheng 42933fd0a558SYan, Zheng node = rc->backref_cache.path[level]; 42943fd0a558SYan, Zheng BUG_ON(node->bytenr != buf->start && 42953fd0a558SYan, Zheng node->new_bytenr != buf->start); 42963fd0a558SYan, Zheng 42973fd0a558SYan, Zheng drop_node_buffer(node); 42983fd0a558SYan, Zheng extent_buffer_get(cow); 42993fd0a558SYan, Zheng node->eb = cow; 43003fd0a558SYan, Zheng node->new_bytenr = cow->start; 43013fd0a558SYan, Zheng 43023fd0a558SYan, Zheng if (!node->pending) { 43033fd0a558SYan, Zheng list_move_tail(&node->list, 43043fd0a558SYan, Zheng &rc->backref_cache.pending[level]); 43053fd0a558SYan, Zheng node->pending = 1; 43063fd0a558SYan, Zheng } 43073fd0a558SYan, Zheng 43083fd0a558SYan, Zheng if (first_cow) 43093fd0a558SYan, Zheng __mark_block_processed(rc, node); 43103fd0a558SYan, Zheng 43113fd0a558SYan, Zheng if (first_cow && level > 0) 43123fd0a558SYan, Zheng rc->nodes_relocated += buf->len; 43133fd0a558SYan, Zheng } 43143fd0a558SYan, Zheng 43153fd0a558SYan, Zheng if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) { 43163fd0a558SYan, Zheng ret = replace_file_extents(trans, rc, root, cow); 43173fd0a558SYan, Zheng BUG_ON(ret); 43183fd0a558SYan, Zheng } 43193fd0a558SYan, Zheng } 43203fd0a558SYan, Zheng 43213fd0a558SYan, Zheng /* 43223fd0a558SYan, Zheng * called before creating snapshot. it calculates metadata reservation 43233fd0a558SYan, Zheng * requried for relocating tree blocks in the snapshot 43243fd0a558SYan, Zheng */ 43253fd0a558SYan, Zheng void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans, 43263fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending, 43273fd0a558SYan, Zheng u64 *bytes_to_reserve) 43283fd0a558SYan, Zheng { 43293fd0a558SYan, Zheng struct btrfs_root *root; 43303fd0a558SYan, Zheng struct reloc_control *rc; 43313fd0a558SYan, Zheng 43323fd0a558SYan, Zheng root = pending->root; 43333fd0a558SYan, Zheng if (!root->reloc_root) 43343fd0a558SYan, Zheng return; 43353fd0a558SYan, Zheng 43363fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 43373fd0a558SYan, Zheng if (!rc->merge_reloc_tree) 43383fd0a558SYan, Zheng return; 43393fd0a558SYan, Zheng 43403fd0a558SYan, Zheng root = root->reloc_root; 43413fd0a558SYan, Zheng BUG_ON(btrfs_root_refs(&root->root_item) == 0); 43423fd0a558SYan, Zheng /* 43433fd0a558SYan, Zheng * relocation is in the stage of merging trees. the space 43443fd0a558SYan, Zheng * used by merging a reloc tree is twice the size of 43453fd0a558SYan, Zheng * relocated tree nodes in the worst case. half for cowing 43463fd0a558SYan, Zheng * the reloc tree, half for cowing the fs tree. the space 43473fd0a558SYan, Zheng * used by cowing the reloc tree will be freed after the 43483fd0a558SYan, Zheng * tree is dropped. if we create snapshot, cowing the fs 43493fd0a558SYan, Zheng * tree may use more space than it frees. so we need 43503fd0a558SYan, Zheng * reserve extra space. 43513fd0a558SYan, Zheng */ 43523fd0a558SYan, Zheng *bytes_to_reserve += rc->nodes_relocated; 43533fd0a558SYan, Zheng } 43543fd0a558SYan, Zheng 43553fd0a558SYan, Zheng /* 43563fd0a558SYan, Zheng * called after snapshot is created. migrate block reservation 43573fd0a558SYan, Zheng * and create reloc root for the newly created snapshot 43583fd0a558SYan, Zheng */ 43593fd0a558SYan, Zheng void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans, 43603fd0a558SYan, Zheng struct btrfs_pending_snapshot *pending) 43613fd0a558SYan, Zheng { 43623fd0a558SYan, Zheng struct btrfs_root *root = pending->root; 43633fd0a558SYan, Zheng struct btrfs_root *reloc_root; 43643fd0a558SYan, Zheng struct btrfs_root *new_root; 43653fd0a558SYan, Zheng struct reloc_control *rc; 43663fd0a558SYan, Zheng int ret; 43673fd0a558SYan, Zheng 43683fd0a558SYan, Zheng if (!root->reloc_root) 43693fd0a558SYan, Zheng return; 43703fd0a558SYan, Zheng 43713fd0a558SYan, Zheng rc = root->fs_info->reloc_ctl; 43723fd0a558SYan, Zheng rc->merging_rsv_size += rc->nodes_relocated; 43733fd0a558SYan, Zheng 43743fd0a558SYan, Zheng if (rc->merge_reloc_tree) { 43753fd0a558SYan, Zheng ret = btrfs_block_rsv_migrate(&pending->block_rsv, 43763fd0a558SYan, Zheng rc->block_rsv, 43773fd0a558SYan, Zheng rc->nodes_relocated); 43783fd0a558SYan, Zheng BUG_ON(ret); 43793fd0a558SYan, Zheng } 43803fd0a558SYan, Zheng 43813fd0a558SYan, Zheng new_root = pending->snap; 43823fd0a558SYan, Zheng reloc_root = create_reloc_root(trans, root->reloc_root, 43833fd0a558SYan, Zheng new_root->root_key.objectid); 43843fd0a558SYan, Zheng 43853fd0a558SYan, Zheng __add_reloc_root(reloc_root); 43863fd0a558SYan, Zheng new_root->reloc_root = reloc_root; 43873fd0a558SYan, Zheng 43883fd0a558SYan, Zheng if (rc->create_reloc_tree) { 43893fd0a558SYan, Zheng ret = clone_backref_node(trans, rc, root, reloc_root); 43903fd0a558SYan, Zheng BUG_ON(ret); 43913fd0a558SYan, Zheng } 43923fd0a558SYan, Zheng } 4393