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> 245d4f98a2SYan Zheng #include "ctree.h" 255d4f98a2SYan Zheng #include "disk-io.h" 265d4f98a2SYan Zheng #include "transaction.h" 275d4f98a2SYan Zheng #include "volumes.h" 285d4f98a2SYan Zheng #include "locking.h" 295d4f98a2SYan Zheng #include "btrfs_inode.h" 305d4f98a2SYan Zheng #include "async-thread.h" 315d4f98a2SYan Zheng 325d4f98a2SYan Zheng /* 335d4f98a2SYan Zheng * backref_node, mapping_node and tree_block start with this 345d4f98a2SYan Zheng */ 355d4f98a2SYan Zheng struct tree_entry { 365d4f98a2SYan Zheng struct rb_node rb_node; 375d4f98a2SYan Zheng u64 bytenr; 385d4f98a2SYan Zheng }; 395d4f98a2SYan Zheng 405d4f98a2SYan Zheng /* 415d4f98a2SYan Zheng * present a tree block in the backref cache 425d4f98a2SYan Zheng */ 435d4f98a2SYan Zheng struct backref_node { 445d4f98a2SYan Zheng struct rb_node rb_node; 455d4f98a2SYan Zheng u64 bytenr; 465d4f98a2SYan Zheng /* objectid tree block owner */ 475d4f98a2SYan Zheng u64 owner; 485d4f98a2SYan Zheng /* list of upper level blocks reference this block */ 495d4f98a2SYan Zheng struct list_head upper; 505d4f98a2SYan Zheng /* list of child blocks in the cache */ 515d4f98a2SYan Zheng struct list_head lower; 525d4f98a2SYan Zheng /* NULL if this node is not tree root */ 535d4f98a2SYan Zheng struct btrfs_root *root; 545d4f98a2SYan Zheng /* extent buffer got by COW the block */ 555d4f98a2SYan Zheng struct extent_buffer *eb; 565d4f98a2SYan Zheng /* level of tree block */ 575d4f98a2SYan Zheng unsigned int level:8; 585d4f98a2SYan Zheng /* 1 if the block is root of old snapshot */ 595d4f98a2SYan Zheng unsigned int old_root:1; 605d4f98a2SYan Zheng /* 1 if no child blocks in the cache */ 615d4f98a2SYan Zheng unsigned int lowest:1; 625d4f98a2SYan Zheng /* is the extent buffer locked */ 635d4f98a2SYan Zheng unsigned int locked:1; 645d4f98a2SYan Zheng /* has the block been processed */ 655d4f98a2SYan Zheng unsigned int processed:1; 665d4f98a2SYan Zheng /* have backrefs of this block been checked */ 675d4f98a2SYan Zheng unsigned int checked:1; 685d4f98a2SYan Zheng }; 695d4f98a2SYan Zheng 705d4f98a2SYan Zheng /* 715d4f98a2SYan Zheng * present a block pointer in the backref cache 725d4f98a2SYan Zheng */ 735d4f98a2SYan Zheng struct backref_edge { 745d4f98a2SYan Zheng struct list_head list[2]; 755d4f98a2SYan Zheng struct backref_node *node[2]; 765d4f98a2SYan Zheng u64 blockptr; 775d4f98a2SYan Zheng }; 785d4f98a2SYan Zheng 795d4f98a2SYan Zheng #define LOWER 0 805d4f98a2SYan Zheng #define UPPER 1 815d4f98a2SYan Zheng 825d4f98a2SYan Zheng struct backref_cache { 835d4f98a2SYan Zheng /* red black tree of all backref nodes in the cache */ 845d4f98a2SYan Zheng struct rb_root rb_root; 855d4f98a2SYan Zheng /* list of backref nodes with no child block in the cache */ 865d4f98a2SYan Zheng struct list_head pending[BTRFS_MAX_LEVEL]; 875d4f98a2SYan Zheng spinlock_t lock; 885d4f98a2SYan Zheng }; 895d4f98a2SYan Zheng 905d4f98a2SYan Zheng /* 915d4f98a2SYan Zheng * map address of tree root to tree 925d4f98a2SYan Zheng */ 935d4f98a2SYan Zheng struct mapping_node { 945d4f98a2SYan Zheng struct rb_node rb_node; 955d4f98a2SYan Zheng u64 bytenr; 965d4f98a2SYan Zheng void *data; 975d4f98a2SYan Zheng }; 985d4f98a2SYan Zheng 995d4f98a2SYan Zheng struct mapping_tree { 1005d4f98a2SYan Zheng struct rb_root rb_root; 1015d4f98a2SYan Zheng spinlock_t lock; 1025d4f98a2SYan Zheng }; 1035d4f98a2SYan Zheng 1045d4f98a2SYan Zheng /* 1055d4f98a2SYan Zheng * present a tree block to process 1065d4f98a2SYan Zheng */ 1075d4f98a2SYan Zheng struct tree_block { 1085d4f98a2SYan Zheng struct rb_node rb_node; 1095d4f98a2SYan Zheng u64 bytenr; 1105d4f98a2SYan Zheng struct btrfs_key key; 1115d4f98a2SYan Zheng unsigned int level:8; 1125d4f98a2SYan Zheng unsigned int key_ready:1; 1135d4f98a2SYan Zheng }; 1145d4f98a2SYan Zheng 1155d4f98a2SYan Zheng /* inode vector */ 1165d4f98a2SYan Zheng #define INODEVEC_SIZE 16 1175d4f98a2SYan Zheng 1185d4f98a2SYan Zheng struct inodevec { 1195d4f98a2SYan Zheng struct list_head list; 1205d4f98a2SYan Zheng struct inode *inode[INODEVEC_SIZE]; 1215d4f98a2SYan Zheng int nr; 1225d4f98a2SYan Zheng }; 1235d4f98a2SYan Zheng 1245d4f98a2SYan Zheng struct reloc_control { 1255d4f98a2SYan Zheng /* block group to relocate */ 1265d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group; 1275d4f98a2SYan Zheng /* extent tree */ 1285d4f98a2SYan Zheng struct btrfs_root *extent_root; 1295d4f98a2SYan Zheng /* inode for moving data */ 1305d4f98a2SYan Zheng struct inode *data_inode; 1315d4f98a2SYan Zheng struct btrfs_workers workers; 1325d4f98a2SYan Zheng /* tree blocks have been processed */ 1335d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 1345d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 1355d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 1365d4f98a2SYan Zheng /* list of reloc trees */ 1375d4f98a2SYan Zheng struct list_head reloc_roots; 1385d4f98a2SYan Zheng u64 search_start; 1395d4f98a2SYan Zheng u64 extents_found; 1405d4f98a2SYan Zheng u64 extents_skipped; 1415d4f98a2SYan Zheng int stage; 1425d4f98a2SYan Zheng int create_reloc_root; 1435d4f98a2SYan Zheng unsigned int found_file_extent:1; 1445d4f98a2SYan Zheng unsigned int found_old_snapshot:1; 1455d4f98a2SYan Zheng }; 1465d4f98a2SYan Zheng 1475d4f98a2SYan Zheng /* stages of data relocation */ 1485d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 1495d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 1505d4f98a2SYan Zheng 1515d4f98a2SYan Zheng /* 1525d4f98a2SYan Zheng * merge reloc tree to corresponding fs tree in worker threads 1535d4f98a2SYan Zheng */ 1545d4f98a2SYan Zheng struct async_merge { 1555d4f98a2SYan Zheng struct btrfs_work work; 1565d4f98a2SYan Zheng struct reloc_control *rc; 1575d4f98a2SYan Zheng struct btrfs_root *root; 1585d4f98a2SYan Zheng struct completion *done; 1595d4f98a2SYan Zheng atomic_t *num_pending; 1605d4f98a2SYan Zheng }; 1615d4f98a2SYan Zheng 1625d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 1635d4f98a2SYan Zheng { 1645d4f98a2SYan Zheng tree->rb_root.rb_node = NULL; 1655d4f98a2SYan Zheng spin_lock_init(&tree->lock); 1665d4f98a2SYan Zheng } 1675d4f98a2SYan Zheng 1685d4f98a2SYan Zheng static void backref_cache_init(struct backref_cache *cache) 1695d4f98a2SYan Zheng { 1705d4f98a2SYan Zheng int i; 1715d4f98a2SYan Zheng cache->rb_root.rb_node = NULL; 1725d4f98a2SYan Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 1735d4f98a2SYan Zheng INIT_LIST_HEAD(&cache->pending[i]); 1745d4f98a2SYan Zheng spin_lock_init(&cache->lock); 1755d4f98a2SYan Zheng } 1765d4f98a2SYan Zheng 1775d4f98a2SYan Zheng static void backref_node_init(struct backref_node *node) 1785d4f98a2SYan Zheng { 1795d4f98a2SYan Zheng memset(node, 0, sizeof(*node)); 1805d4f98a2SYan Zheng INIT_LIST_HEAD(&node->upper); 1815d4f98a2SYan Zheng INIT_LIST_HEAD(&node->lower); 1825d4f98a2SYan Zheng RB_CLEAR_NODE(&node->rb_node); 1835d4f98a2SYan Zheng } 1845d4f98a2SYan Zheng 1855d4f98a2SYan Zheng static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr, 1865d4f98a2SYan Zheng struct rb_node *node) 1875d4f98a2SYan Zheng { 1885d4f98a2SYan Zheng struct rb_node **p = &root->rb_node; 1895d4f98a2SYan Zheng struct rb_node *parent = NULL; 1905d4f98a2SYan Zheng struct tree_entry *entry; 1915d4f98a2SYan Zheng 1925d4f98a2SYan Zheng while (*p) { 1935d4f98a2SYan Zheng parent = *p; 1945d4f98a2SYan Zheng entry = rb_entry(parent, struct tree_entry, rb_node); 1955d4f98a2SYan Zheng 1965d4f98a2SYan Zheng if (bytenr < entry->bytenr) 1975d4f98a2SYan Zheng p = &(*p)->rb_left; 1985d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 1995d4f98a2SYan Zheng p = &(*p)->rb_right; 2005d4f98a2SYan Zheng else 2015d4f98a2SYan Zheng return parent; 2025d4f98a2SYan Zheng } 2035d4f98a2SYan Zheng 2045d4f98a2SYan Zheng rb_link_node(node, parent, p); 2055d4f98a2SYan Zheng rb_insert_color(node, root); 2065d4f98a2SYan Zheng return NULL; 2075d4f98a2SYan Zheng } 2085d4f98a2SYan Zheng 2095d4f98a2SYan Zheng static struct rb_node *tree_search(struct rb_root *root, u64 bytenr) 2105d4f98a2SYan Zheng { 2115d4f98a2SYan Zheng struct rb_node *n = root->rb_node; 2125d4f98a2SYan Zheng struct tree_entry *entry; 2135d4f98a2SYan Zheng 2145d4f98a2SYan Zheng while (n) { 2155d4f98a2SYan Zheng entry = rb_entry(n, struct tree_entry, rb_node); 2165d4f98a2SYan Zheng 2175d4f98a2SYan Zheng if (bytenr < entry->bytenr) 2185d4f98a2SYan Zheng n = n->rb_left; 2195d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 2205d4f98a2SYan Zheng n = n->rb_right; 2215d4f98a2SYan Zheng else 2225d4f98a2SYan Zheng return n; 2235d4f98a2SYan Zheng } 2245d4f98a2SYan Zheng return NULL; 2255d4f98a2SYan Zheng } 2265d4f98a2SYan Zheng 2275d4f98a2SYan Zheng /* 2285d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 2295d4f98a2SYan Zheng */ 2305d4f98a2SYan Zheng static struct backref_node *walk_up_backref(struct backref_node *node, 2315d4f98a2SYan Zheng struct backref_edge *edges[], 2325d4f98a2SYan Zheng int *index) 2335d4f98a2SYan Zheng { 2345d4f98a2SYan Zheng struct backref_edge *edge; 2355d4f98a2SYan Zheng int idx = *index; 2365d4f98a2SYan Zheng 2375d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 2385d4f98a2SYan Zheng edge = list_entry(node->upper.next, 2395d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 2405d4f98a2SYan Zheng edges[idx++] = edge; 2415d4f98a2SYan Zheng node = edge->node[UPPER]; 2425d4f98a2SYan Zheng } 2435d4f98a2SYan Zheng *index = idx; 2445d4f98a2SYan Zheng return node; 2455d4f98a2SYan Zheng } 2465d4f98a2SYan Zheng 2475d4f98a2SYan Zheng /* 2485d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 2495d4f98a2SYan Zheng */ 2505d4f98a2SYan Zheng static struct backref_node *walk_down_backref(struct backref_edge *edges[], 2515d4f98a2SYan Zheng int *index) 2525d4f98a2SYan Zheng { 2535d4f98a2SYan Zheng struct backref_edge *edge; 2545d4f98a2SYan Zheng struct backref_node *lower; 2555d4f98a2SYan Zheng int idx = *index; 2565d4f98a2SYan Zheng 2575d4f98a2SYan Zheng while (idx > 0) { 2585d4f98a2SYan Zheng edge = edges[idx - 1]; 2595d4f98a2SYan Zheng lower = edge->node[LOWER]; 2605d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 2615d4f98a2SYan Zheng idx--; 2625d4f98a2SYan Zheng continue; 2635d4f98a2SYan Zheng } 2645d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 2655d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 2665d4f98a2SYan Zheng edges[idx - 1] = edge; 2675d4f98a2SYan Zheng *index = idx; 2685d4f98a2SYan Zheng return edge->node[UPPER]; 2695d4f98a2SYan Zheng } 2705d4f98a2SYan Zheng *index = 0; 2715d4f98a2SYan Zheng return NULL; 2725d4f98a2SYan Zheng } 2735d4f98a2SYan Zheng 2745d4f98a2SYan Zheng static void drop_node_buffer(struct backref_node *node) 2755d4f98a2SYan Zheng { 2765d4f98a2SYan Zheng if (node->eb) { 2775d4f98a2SYan Zheng if (node->locked) { 2785d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 2795d4f98a2SYan Zheng node->locked = 0; 2805d4f98a2SYan Zheng } 2815d4f98a2SYan Zheng free_extent_buffer(node->eb); 2825d4f98a2SYan Zheng node->eb = NULL; 2835d4f98a2SYan Zheng } 2845d4f98a2SYan Zheng } 2855d4f98a2SYan Zheng 2865d4f98a2SYan Zheng static void drop_backref_node(struct backref_cache *tree, 2875d4f98a2SYan Zheng struct backref_node *node) 2885d4f98a2SYan Zheng { 2895d4f98a2SYan Zheng BUG_ON(!node->lowest); 2905d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 2915d4f98a2SYan Zheng 2925d4f98a2SYan Zheng drop_node_buffer(node); 2935d4f98a2SYan Zheng list_del(&node->lower); 2945d4f98a2SYan Zheng 2955d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 2965d4f98a2SYan Zheng kfree(node); 2975d4f98a2SYan Zheng } 2985d4f98a2SYan Zheng 2995d4f98a2SYan Zheng /* 3005d4f98a2SYan Zheng * remove a backref node from the backref cache 3015d4f98a2SYan Zheng */ 3025d4f98a2SYan Zheng static void remove_backref_node(struct backref_cache *cache, 3035d4f98a2SYan Zheng struct backref_node *node) 3045d4f98a2SYan Zheng { 3055d4f98a2SYan Zheng struct backref_node *upper; 3065d4f98a2SYan Zheng struct backref_edge *edge; 3075d4f98a2SYan Zheng 3085d4f98a2SYan Zheng if (!node) 3095d4f98a2SYan Zheng return; 3105d4f98a2SYan Zheng 3115d4f98a2SYan Zheng BUG_ON(!node->lowest); 3125d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 3135d4f98a2SYan Zheng edge = list_entry(node->upper.next, struct backref_edge, 3145d4f98a2SYan Zheng list[LOWER]); 3155d4f98a2SYan Zheng upper = edge->node[UPPER]; 3165d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 3175d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 3185d4f98a2SYan Zheng kfree(edge); 3195d4f98a2SYan Zheng /* 3205d4f98a2SYan Zheng * add the node to pending list if no other 3215d4f98a2SYan Zheng * child block cached. 3225d4f98a2SYan Zheng */ 3235d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 3245d4f98a2SYan Zheng list_add_tail(&upper->lower, 3255d4f98a2SYan Zheng &cache->pending[upper->level]); 3265d4f98a2SYan Zheng upper->lowest = 1; 3275d4f98a2SYan Zheng } 3285d4f98a2SYan Zheng } 3295d4f98a2SYan Zheng drop_backref_node(cache, node); 3305d4f98a2SYan Zheng } 3315d4f98a2SYan Zheng 3325d4f98a2SYan Zheng /* 3335d4f98a2SYan Zheng * find reloc tree by address of tree root 3345d4f98a2SYan Zheng */ 3355d4f98a2SYan Zheng static struct btrfs_root *find_reloc_root(struct reloc_control *rc, 3365d4f98a2SYan Zheng u64 bytenr) 3375d4f98a2SYan Zheng { 3385d4f98a2SYan Zheng struct rb_node *rb_node; 3395d4f98a2SYan Zheng struct mapping_node *node; 3405d4f98a2SYan Zheng struct btrfs_root *root = NULL; 3415d4f98a2SYan Zheng 3425d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 3435d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr); 3445d4f98a2SYan Zheng if (rb_node) { 3455d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 3465d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 3475d4f98a2SYan Zheng } 3485d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 3495d4f98a2SYan Zheng return root; 3505d4f98a2SYan Zheng } 3515d4f98a2SYan Zheng 3525d4f98a2SYan Zheng static int is_cowonly_root(u64 root_objectid) 3535d4f98a2SYan Zheng { 3545d4f98a2SYan Zheng if (root_objectid == BTRFS_ROOT_TREE_OBJECTID || 3555d4f98a2SYan Zheng root_objectid == BTRFS_EXTENT_TREE_OBJECTID || 3565d4f98a2SYan Zheng root_objectid == BTRFS_CHUNK_TREE_OBJECTID || 3575d4f98a2SYan Zheng root_objectid == BTRFS_DEV_TREE_OBJECTID || 3585d4f98a2SYan Zheng root_objectid == BTRFS_TREE_LOG_OBJECTID || 3595d4f98a2SYan Zheng root_objectid == BTRFS_CSUM_TREE_OBJECTID) 3605d4f98a2SYan Zheng return 1; 3615d4f98a2SYan Zheng return 0; 3625d4f98a2SYan Zheng } 3635d4f98a2SYan Zheng 3645d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 3655d4f98a2SYan Zheng u64 root_objectid) 3665d4f98a2SYan Zheng { 3675d4f98a2SYan Zheng struct btrfs_key key; 3685d4f98a2SYan Zheng 3695d4f98a2SYan Zheng key.objectid = root_objectid; 3705d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 3715d4f98a2SYan Zheng if (is_cowonly_root(root_objectid)) 3725d4f98a2SYan Zheng key.offset = 0; 3735d4f98a2SYan Zheng else 3745d4f98a2SYan Zheng key.offset = (u64)-1; 3755d4f98a2SYan Zheng 3765d4f98a2SYan Zheng return btrfs_read_fs_root_no_name(fs_info, &key); 3775d4f98a2SYan Zheng } 3785d4f98a2SYan Zheng 3795d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 3805d4f98a2SYan Zheng static noinline_for_stack 3815d4f98a2SYan Zheng struct btrfs_root *find_tree_root(struct reloc_control *rc, 3825d4f98a2SYan Zheng struct extent_buffer *leaf, 3835d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0) 3845d4f98a2SYan Zheng { 3855d4f98a2SYan Zheng struct btrfs_root *root; 3865d4f98a2SYan Zheng u64 root_objectid = btrfs_ref_root_v0(leaf, ref0); 3875d4f98a2SYan Zheng u64 generation = btrfs_ref_generation_v0(leaf, ref0); 3885d4f98a2SYan Zheng 3895d4f98a2SYan Zheng BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID); 3905d4f98a2SYan Zheng 3915d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, root_objectid); 3925d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 3935d4f98a2SYan Zheng 3945d4f98a2SYan Zheng if (root->ref_cows && 3955d4f98a2SYan Zheng generation != btrfs_root_generation(&root->root_item)) 3965d4f98a2SYan Zheng return NULL; 3975d4f98a2SYan Zheng 3985d4f98a2SYan Zheng return root; 3995d4f98a2SYan Zheng } 4005d4f98a2SYan Zheng #endif 4015d4f98a2SYan Zheng 4025d4f98a2SYan Zheng static noinline_for_stack 4035d4f98a2SYan Zheng int find_inline_backref(struct extent_buffer *leaf, int slot, 4045d4f98a2SYan Zheng unsigned long *ptr, unsigned long *end) 4055d4f98a2SYan Zheng { 4065d4f98a2SYan Zheng struct btrfs_extent_item *ei; 4075d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 4085d4f98a2SYan Zheng u32 item_size; 4095d4f98a2SYan Zheng 4105d4f98a2SYan Zheng item_size = btrfs_item_size_nr(leaf, slot); 4115d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 4125d4f98a2SYan Zheng if (item_size < sizeof(*ei)) { 4135d4f98a2SYan Zheng WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 4145d4f98a2SYan Zheng return 1; 4155d4f98a2SYan Zheng } 4165d4f98a2SYan Zheng #endif 4175d4f98a2SYan Zheng ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item); 4185d4f98a2SYan Zheng WARN_ON(!(btrfs_extent_flags(leaf, ei) & 4195d4f98a2SYan Zheng BTRFS_EXTENT_FLAG_TREE_BLOCK)); 4205d4f98a2SYan Zheng 4215d4f98a2SYan Zheng if (item_size <= sizeof(*ei) + sizeof(*bi)) { 4225d4f98a2SYan Zheng WARN_ON(item_size < sizeof(*ei) + sizeof(*bi)); 4235d4f98a2SYan Zheng return 1; 4245d4f98a2SYan Zheng } 4255d4f98a2SYan Zheng 4265d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 4275d4f98a2SYan Zheng *ptr = (unsigned long)(bi + 1); 4285d4f98a2SYan Zheng *end = (unsigned long)ei + item_size; 4295d4f98a2SYan Zheng return 0; 4305d4f98a2SYan Zheng } 4315d4f98a2SYan Zheng 4325d4f98a2SYan Zheng /* 4335d4f98a2SYan Zheng * build backref tree for a given tree block. root of the backref tree 4345d4f98a2SYan Zheng * corresponds the tree block, leaves of the backref tree correspond 4355d4f98a2SYan Zheng * roots of b-trees that reference the tree block. 4365d4f98a2SYan Zheng * 4375d4f98a2SYan Zheng * the basic idea of this function is check backrefs of a given block 4385d4f98a2SYan Zheng * to find upper level blocks that refernece the block, and then check 4395d4f98a2SYan Zheng * bakcrefs of these upper level blocks recursively. the recursion stop 4405d4f98a2SYan Zheng * when tree root is reached or backrefs for the block is cached. 4415d4f98a2SYan Zheng * 4425d4f98a2SYan Zheng * NOTE: if we find backrefs for a block are cached, we know backrefs 4435d4f98a2SYan Zheng * for all upper level blocks that directly/indirectly reference the 4445d4f98a2SYan Zheng * block are also cached. 4455d4f98a2SYan Zheng */ 4465d4f98a2SYan Zheng static struct backref_node *build_backref_tree(struct reloc_control *rc, 4475d4f98a2SYan Zheng struct backref_cache *cache, 4485d4f98a2SYan Zheng struct btrfs_key *node_key, 4495d4f98a2SYan Zheng int level, u64 bytenr) 4505d4f98a2SYan Zheng { 4515d4f98a2SYan Zheng struct btrfs_path *path1; 4525d4f98a2SYan Zheng struct btrfs_path *path2; 4535d4f98a2SYan Zheng struct extent_buffer *eb; 4545d4f98a2SYan Zheng struct btrfs_root *root; 4555d4f98a2SYan Zheng struct backref_node *cur; 4565d4f98a2SYan Zheng struct backref_node *upper; 4575d4f98a2SYan Zheng struct backref_node *lower; 4585d4f98a2SYan Zheng struct backref_node *node = NULL; 4595d4f98a2SYan Zheng struct backref_node *exist = NULL; 4605d4f98a2SYan Zheng struct backref_edge *edge; 4615d4f98a2SYan Zheng struct rb_node *rb_node; 4625d4f98a2SYan Zheng struct btrfs_key key; 4635d4f98a2SYan Zheng unsigned long end; 4645d4f98a2SYan Zheng unsigned long ptr; 4655d4f98a2SYan Zheng LIST_HEAD(list); 4665d4f98a2SYan Zheng int ret; 4675d4f98a2SYan Zheng int err = 0; 4685d4f98a2SYan Zheng 4695d4f98a2SYan Zheng path1 = btrfs_alloc_path(); 4705d4f98a2SYan Zheng path2 = btrfs_alloc_path(); 4715d4f98a2SYan Zheng if (!path1 || !path2) { 4725d4f98a2SYan Zheng err = -ENOMEM; 4735d4f98a2SYan Zheng goto out; 4745d4f98a2SYan Zheng } 4755d4f98a2SYan Zheng 4765d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 4775d4f98a2SYan Zheng if (!node) { 4785d4f98a2SYan Zheng err = -ENOMEM; 4795d4f98a2SYan Zheng goto out; 4805d4f98a2SYan Zheng } 4815d4f98a2SYan Zheng 4825d4f98a2SYan Zheng backref_node_init(node); 4835d4f98a2SYan Zheng node->bytenr = bytenr; 4845d4f98a2SYan Zheng node->owner = 0; 4855d4f98a2SYan Zheng node->level = level; 4865d4f98a2SYan Zheng node->lowest = 1; 4875d4f98a2SYan Zheng cur = node; 4885d4f98a2SYan Zheng again: 4895d4f98a2SYan Zheng end = 0; 4905d4f98a2SYan Zheng ptr = 0; 4915d4f98a2SYan Zheng key.objectid = cur->bytenr; 4925d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 4935d4f98a2SYan Zheng key.offset = (u64)-1; 4945d4f98a2SYan Zheng 4955d4f98a2SYan Zheng path1->search_commit_root = 1; 4965d4f98a2SYan Zheng path1->skip_locking = 1; 4975d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1, 4985d4f98a2SYan Zheng 0, 0); 4995d4f98a2SYan Zheng if (ret < 0) { 5005d4f98a2SYan Zheng err = ret; 5015d4f98a2SYan Zheng goto out; 5025d4f98a2SYan Zheng } 5035d4f98a2SYan Zheng BUG_ON(!ret || !path1->slots[0]); 5045d4f98a2SYan Zheng 5055d4f98a2SYan Zheng path1->slots[0]--; 5065d4f98a2SYan Zheng 5075d4f98a2SYan Zheng WARN_ON(cur->checked); 5085d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 5095d4f98a2SYan Zheng /* 5105d4f98a2SYan Zheng * the backref was added previously when processsing 5115d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 5125d4f98a2SYan Zheng */ 5135d4f98a2SYan Zheng BUG_ON(!list_is_singular(&cur->upper)); 5145d4f98a2SYan Zheng edge = list_entry(cur->upper.next, struct backref_edge, 5155d4f98a2SYan Zheng list[LOWER]); 5165d4f98a2SYan Zheng BUG_ON(!list_empty(&edge->list[UPPER])); 5175d4f98a2SYan Zheng exist = edge->node[UPPER]; 5185d4f98a2SYan Zheng /* 5195d4f98a2SYan Zheng * add the upper level block to pending list if we need 5205d4f98a2SYan Zheng * check its backrefs 5215d4f98a2SYan Zheng */ 5225d4f98a2SYan Zheng if (!exist->checked) 5235d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 5245d4f98a2SYan Zheng } else { 5255d4f98a2SYan Zheng exist = NULL; 5265d4f98a2SYan Zheng } 5275d4f98a2SYan Zheng 5285d4f98a2SYan Zheng while (1) { 5295d4f98a2SYan Zheng cond_resched(); 5305d4f98a2SYan Zheng eb = path1->nodes[0]; 5315d4f98a2SYan Zheng 5325d4f98a2SYan Zheng if (ptr >= end) { 5335d4f98a2SYan Zheng if (path1->slots[0] >= btrfs_header_nritems(eb)) { 5345d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path1); 5355d4f98a2SYan Zheng if (ret < 0) { 5365d4f98a2SYan Zheng err = ret; 5375d4f98a2SYan Zheng goto out; 5385d4f98a2SYan Zheng } 5395d4f98a2SYan Zheng if (ret > 0) 5405d4f98a2SYan Zheng break; 5415d4f98a2SYan Zheng eb = path1->nodes[0]; 5425d4f98a2SYan Zheng } 5435d4f98a2SYan Zheng 5445d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path1->slots[0]); 5455d4f98a2SYan Zheng if (key.objectid != cur->bytenr) { 5465d4f98a2SYan Zheng WARN_ON(exist); 5475d4f98a2SYan Zheng break; 5485d4f98a2SYan Zheng } 5495d4f98a2SYan Zheng 5505d4f98a2SYan Zheng if (key.type == BTRFS_EXTENT_ITEM_KEY) { 5515d4f98a2SYan Zheng ret = find_inline_backref(eb, path1->slots[0], 5525d4f98a2SYan Zheng &ptr, &end); 5535d4f98a2SYan Zheng if (ret) 5545d4f98a2SYan Zheng goto next; 5555d4f98a2SYan Zheng } 5565d4f98a2SYan Zheng } 5575d4f98a2SYan Zheng 5585d4f98a2SYan Zheng if (ptr < end) { 5595d4f98a2SYan Zheng /* update key for inline back ref */ 5605d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 5615d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 5625d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 5635d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 5645d4f98a2SYan Zheng WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY && 5655d4f98a2SYan Zheng key.type != BTRFS_SHARED_BLOCK_REF_KEY); 5665d4f98a2SYan Zheng } 5675d4f98a2SYan Zheng 5685d4f98a2SYan Zheng if (exist && 5695d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 5705d4f98a2SYan Zheng exist->owner == key.offset) || 5715d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 5725d4f98a2SYan Zheng exist->bytenr == key.offset))) { 5735d4f98a2SYan Zheng exist = NULL; 5745d4f98a2SYan Zheng goto next; 5755d4f98a2SYan Zheng } 5765d4f98a2SYan Zheng 5775d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 5785d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY || 5795d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 5805d4f98a2SYan Zheng if (key.objectid == key.offset && 5815d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 5825d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 5835d4f98a2SYan Zheng ref0 = btrfs_item_ptr(eb, path1->slots[0], 5845d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 5855d4f98a2SYan Zheng root = find_tree_root(rc, eb, ref0); 5865d4f98a2SYan Zheng if (root) 5875d4f98a2SYan Zheng cur->root = root; 5885d4f98a2SYan Zheng else 5895d4f98a2SYan Zheng cur->old_root = 1; 5905d4f98a2SYan Zheng break; 5915d4f98a2SYan Zheng } 5925d4f98a2SYan Zheng #else 5935d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 5945d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 5955d4f98a2SYan Zheng #endif 5965d4f98a2SYan Zheng if (key.objectid == key.offset) { 5975d4f98a2SYan Zheng /* 5985d4f98a2SYan Zheng * only root blocks of reloc trees use 5995d4f98a2SYan Zheng * backref of this type. 6005d4f98a2SYan Zheng */ 6015d4f98a2SYan Zheng root = find_reloc_root(rc, cur->bytenr); 6025d4f98a2SYan Zheng BUG_ON(!root); 6035d4f98a2SYan Zheng cur->root = root; 6045d4f98a2SYan Zheng break; 6055d4f98a2SYan Zheng } 6065d4f98a2SYan Zheng 6075d4f98a2SYan Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 6085d4f98a2SYan Zheng if (!edge) { 6095d4f98a2SYan Zheng err = -ENOMEM; 6105d4f98a2SYan Zheng goto out; 6115d4f98a2SYan Zheng } 6125d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, key.offset); 6135d4f98a2SYan Zheng if (!rb_node) { 6145d4f98a2SYan Zheng upper = kmalloc(sizeof(*upper), GFP_NOFS); 6155d4f98a2SYan Zheng if (!upper) { 6165d4f98a2SYan Zheng kfree(edge); 6175d4f98a2SYan Zheng err = -ENOMEM; 6185d4f98a2SYan Zheng goto out; 6195d4f98a2SYan Zheng } 6205d4f98a2SYan Zheng backref_node_init(upper); 6215d4f98a2SYan Zheng upper->bytenr = key.offset; 6225d4f98a2SYan Zheng upper->owner = 0; 6235d4f98a2SYan Zheng upper->level = cur->level + 1; 6245d4f98a2SYan Zheng /* 6255d4f98a2SYan Zheng * backrefs for the upper level block isn't 6265d4f98a2SYan Zheng * cached, add the block to pending list 6275d4f98a2SYan Zheng */ 6285d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 6295d4f98a2SYan Zheng } else { 6305d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 6315d4f98a2SYan Zheng rb_node); 6325d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 6335d4f98a2SYan Zheng } 6345d4f98a2SYan Zheng list_add(&edge->list[LOWER], &cur->upper); 6355d4f98a2SYan Zheng edge->node[UPPER] = upper; 6365d4f98a2SYan Zheng edge->node[LOWER] = cur; 6375d4f98a2SYan Zheng 6385d4f98a2SYan Zheng goto next; 6395d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 6405d4f98a2SYan Zheng goto next; 6415d4f98a2SYan Zheng } 6425d4f98a2SYan Zheng 6435d4f98a2SYan Zheng /* key.type == BTRFS_TREE_BLOCK_REF_KEY */ 6445d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, key.offset); 6455d4f98a2SYan Zheng if (IS_ERR(root)) { 6465d4f98a2SYan Zheng err = PTR_ERR(root); 6475d4f98a2SYan Zheng goto out; 6485d4f98a2SYan Zheng } 6495d4f98a2SYan Zheng 6505d4f98a2SYan Zheng if (btrfs_root_level(&root->root_item) == cur->level) { 6515d4f98a2SYan Zheng /* tree root */ 6525d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 6535d4f98a2SYan Zheng cur->bytenr); 6545d4f98a2SYan Zheng cur->root = root; 6555d4f98a2SYan Zheng break; 6565d4f98a2SYan Zheng } 6575d4f98a2SYan Zheng 6585d4f98a2SYan Zheng level = cur->level + 1; 6595d4f98a2SYan Zheng 6605d4f98a2SYan Zheng /* 6615d4f98a2SYan Zheng * searching the tree to find upper level blocks 6625d4f98a2SYan Zheng * reference the block. 6635d4f98a2SYan Zheng */ 6645d4f98a2SYan Zheng path2->search_commit_root = 1; 6655d4f98a2SYan Zheng path2->skip_locking = 1; 6665d4f98a2SYan Zheng path2->lowest_level = level; 6675d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0); 6685d4f98a2SYan Zheng path2->lowest_level = 0; 6695d4f98a2SYan Zheng if (ret < 0) { 6705d4f98a2SYan Zheng err = ret; 6715d4f98a2SYan Zheng goto out; 6725d4f98a2SYan Zheng } 6735d4f98a2SYan Zheng 6745d4f98a2SYan Zheng eb = path2->nodes[level]; 6755d4f98a2SYan Zheng WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) != 6765d4f98a2SYan Zheng cur->bytenr); 6775d4f98a2SYan Zheng 6785d4f98a2SYan Zheng lower = cur; 6795d4f98a2SYan Zheng for (; level < BTRFS_MAX_LEVEL; level++) { 6805d4f98a2SYan Zheng if (!path2->nodes[level]) { 6815d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 6825d4f98a2SYan Zheng lower->bytenr); 6835d4f98a2SYan Zheng lower->root = root; 6845d4f98a2SYan Zheng break; 6855d4f98a2SYan Zheng } 6865d4f98a2SYan Zheng 6875d4f98a2SYan Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 6885d4f98a2SYan Zheng if (!edge) { 6895d4f98a2SYan Zheng err = -ENOMEM; 6905d4f98a2SYan Zheng goto out; 6915d4f98a2SYan Zheng } 6925d4f98a2SYan Zheng 6935d4f98a2SYan Zheng eb = path2->nodes[level]; 6945d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, eb->start); 6955d4f98a2SYan Zheng if (!rb_node) { 6965d4f98a2SYan Zheng upper = kmalloc(sizeof(*upper), GFP_NOFS); 6975d4f98a2SYan Zheng if (!upper) { 6985d4f98a2SYan Zheng kfree(edge); 6995d4f98a2SYan Zheng err = -ENOMEM; 7005d4f98a2SYan Zheng goto out; 7015d4f98a2SYan Zheng } 7025d4f98a2SYan Zheng backref_node_init(upper); 7035d4f98a2SYan Zheng upper->bytenr = eb->start; 7045d4f98a2SYan Zheng upper->owner = btrfs_header_owner(eb); 7055d4f98a2SYan Zheng upper->level = lower->level + 1; 7065d4f98a2SYan Zheng 7075d4f98a2SYan Zheng /* 7085d4f98a2SYan Zheng * if we know the block isn't shared 7095d4f98a2SYan Zheng * we can void checking its backrefs. 7105d4f98a2SYan Zheng */ 7115d4f98a2SYan Zheng if (btrfs_block_can_be_shared(root, eb)) 7125d4f98a2SYan Zheng upper->checked = 0; 7135d4f98a2SYan Zheng else 7145d4f98a2SYan Zheng upper->checked = 1; 7155d4f98a2SYan Zheng 7165d4f98a2SYan Zheng /* 7175d4f98a2SYan Zheng * add the block to pending list if we 7185d4f98a2SYan Zheng * need check its backrefs. only block 7195d4f98a2SYan Zheng * at 'cur->level + 1' is added to the 7205d4f98a2SYan Zheng * tail of pending list. this guarantees 7215d4f98a2SYan Zheng * we check backrefs from lower level 7225d4f98a2SYan Zheng * blocks to upper level blocks. 7235d4f98a2SYan Zheng */ 7245d4f98a2SYan Zheng if (!upper->checked && 7255d4f98a2SYan Zheng level == cur->level + 1) { 7265d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], 7275d4f98a2SYan Zheng &list); 7285d4f98a2SYan Zheng } else 7295d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 7305d4f98a2SYan Zheng } else { 7315d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 7325d4f98a2SYan Zheng rb_node); 7335d4f98a2SYan Zheng BUG_ON(!upper->checked); 7345d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 7355d4f98a2SYan Zheng } 7365d4f98a2SYan Zheng list_add_tail(&edge->list[LOWER], &lower->upper); 7375d4f98a2SYan Zheng edge->node[UPPER] = upper; 7385d4f98a2SYan Zheng edge->node[LOWER] = lower; 7395d4f98a2SYan Zheng 7405d4f98a2SYan Zheng if (rb_node) 7415d4f98a2SYan Zheng break; 7425d4f98a2SYan Zheng lower = upper; 7435d4f98a2SYan Zheng upper = NULL; 7445d4f98a2SYan Zheng } 7455d4f98a2SYan Zheng btrfs_release_path(root, path2); 7465d4f98a2SYan Zheng next: 7475d4f98a2SYan Zheng if (ptr < end) { 7485d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 7495d4f98a2SYan Zheng if (ptr >= end) { 7505d4f98a2SYan Zheng WARN_ON(ptr > end); 7515d4f98a2SYan Zheng ptr = 0; 7525d4f98a2SYan Zheng end = 0; 7535d4f98a2SYan Zheng } 7545d4f98a2SYan Zheng } 7555d4f98a2SYan Zheng if (ptr >= end) 7565d4f98a2SYan Zheng path1->slots[0]++; 7575d4f98a2SYan Zheng } 7585d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path1); 7595d4f98a2SYan Zheng 7605d4f98a2SYan Zheng cur->checked = 1; 7615d4f98a2SYan Zheng WARN_ON(exist); 7625d4f98a2SYan Zheng 7635d4f98a2SYan Zheng /* the pending list isn't empty, take the first block to process */ 7645d4f98a2SYan Zheng if (!list_empty(&list)) { 7655d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 7665d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 7675d4f98a2SYan Zheng cur = edge->node[UPPER]; 7685d4f98a2SYan Zheng goto again; 7695d4f98a2SYan Zheng } 7705d4f98a2SYan Zheng 7715d4f98a2SYan Zheng /* 7725d4f98a2SYan Zheng * everything goes well, connect backref nodes and insert backref nodes 7735d4f98a2SYan Zheng * into the cache. 7745d4f98a2SYan Zheng */ 7755d4f98a2SYan Zheng BUG_ON(!node->checked); 7765d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 7775d4f98a2SYan Zheng BUG_ON(rb_node); 7785d4f98a2SYan Zheng 7795d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) 7805d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 7815d4f98a2SYan Zheng 7825d4f98a2SYan Zheng while (!list_empty(&list)) { 7835d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 7845d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 7855d4f98a2SYan Zheng upper = edge->node[UPPER]; 7865d4f98a2SYan Zheng 7875d4f98a2SYan Zheng if (!RB_EMPTY_NODE(&upper->rb_node)) { 7885d4f98a2SYan Zheng if (upper->lowest) { 7895d4f98a2SYan Zheng list_del_init(&upper->lower); 7905d4f98a2SYan Zheng upper->lowest = 0; 7915d4f98a2SYan Zheng } 7925d4f98a2SYan Zheng 7935d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 7945d4f98a2SYan Zheng continue; 7955d4f98a2SYan Zheng } 7965d4f98a2SYan Zheng 7975d4f98a2SYan Zheng BUG_ON(!upper->checked); 7985d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, upper->bytenr, 7995d4f98a2SYan Zheng &upper->rb_node); 8005d4f98a2SYan Zheng BUG_ON(rb_node); 8015d4f98a2SYan Zheng 8025d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 8035d4f98a2SYan Zheng 8045d4f98a2SYan Zheng list_for_each_entry(edge, &upper->upper, list[LOWER]) 8055d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 8065d4f98a2SYan Zheng } 8075d4f98a2SYan Zheng out: 8085d4f98a2SYan Zheng btrfs_free_path(path1); 8095d4f98a2SYan Zheng btrfs_free_path(path2); 8105d4f98a2SYan Zheng if (err) { 8115d4f98a2SYan Zheng INIT_LIST_HEAD(&list); 8125d4f98a2SYan Zheng upper = node; 8135d4f98a2SYan Zheng while (upper) { 8145d4f98a2SYan Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 8155d4f98a2SYan Zheng list_splice_tail(&upper->upper, &list); 8165d4f98a2SYan Zheng kfree(upper); 8175d4f98a2SYan Zheng } 8185d4f98a2SYan Zheng 8195d4f98a2SYan Zheng if (list_empty(&list)) 8205d4f98a2SYan Zheng break; 8215d4f98a2SYan Zheng 8225d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, 8235d4f98a2SYan Zheng list[LOWER]); 8245d4f98a2SYan Zheng upper = edge->node[UPPER]; 8255d4f98a2SYan Zheng kfree(edge); 8265d4f98a2SYan Zheng } 8275d4f98a2SYan Zheng return ERR_PTR(err); 8285d4f98a2SYan Zheng } 8295d4f98a2SYan Zheng return node; 8305d4f98a2SYan Zheng } 8315d4f98a2SYan Zheng 8325d4f98a2SYan Zheng /* 8335d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 8345d4f98a2SYan Zheng */ 8355d4f98a2SYan Zheng static int __add_reloc_root(struct btrfs_root *root) 8365d4f98a2SYan Zheng { 8375d4f98a2SYan Zheng struct rb_node *rb_node; 8385d4f98a2SYan Zheng struct mapping_node *node; 8395d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 8405d4f98a2SYan Zheng 8415d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 8425d4f98a2SYan Zheng BUG_ON(!node); 8435d4f98a2SYan Zheng 8445d4f98a2SYan Zheng node->bytenr = root->node->start; 8455d4f98a2SYan Zheng node->data = root; 8465d4f98a2SYan Zheng 8475d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 8485d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 8495d4f98a2SYan Zheng node->bytenr, &node->rb_node); 8505d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 8515d4f98a2SYan Zheng BUG_ON(rb_node); 8525d4f98a2SYan Zheng 8535d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 8545d4f98a2SYan Zheng return 0; 8555d4f98a2SYan Zheng } 8565d4f98a2SYan Zheng 8575d4f98a2SYan Zheng /* 8585d4f98a2SYan Zheng * helper to update/delete the 'address of tree root -> reloc tree' 8595d4f98a2SYan Zheng * mapping 8605d4f98a2SYan Zheng */ 8615d4f98a2SYan Zheng static int __update_reloc_root(struct btrfs_root *root, int del) 8625d4f98a2SYan Zheng { 8635d4f98a2SYan Zheng struct rb_node *rb_node; 8645d4f98a2SYan Zheng struct mapping_node *node = NULL; 8655d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 8665d4f98a2SYan Zheng 8675d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 8685d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, 8695d4f98a2SYan Zheng root->commit_root->start); 8705d4f98a2SYan Zheng if (rb_node) { 8715d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 8725d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 8735d4f98a2SYan Zheng } 8745d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 8755d4f98a2SYan Zheng 8765d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 8775d4f98a2SYan Zheng 8785d4f98a2SYan Zheng if (!del) { 8795d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 8805d4f98a2SYan Zheng node->bytenr = root->node->start; 8815d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 8825d4f98a2SYan Zheng node->bytenr, &node->rb_node); 8835d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 8845d4f98a2SYan Zheng BUG_ON(rb_node); 8855d4f98a2SYan Zheng } else { 8865d4f98a2SYan Zheng list_del_init(&root->root_list); 8875d4f98a2SYan Zheng kfree(node); 8885d4f98a2SYan Zheng } 8895d4f98a2SYan Zheng return 0; 8905d4f98a2SYan Zheng } 8915d4f98a2SYan Zheng 8925d4f98a2SYan Zheng /* 8935d4f98a2SYan Zheng * create reloc tree for a given fs tree. reloc tree is just a 8945d4f98a2SYan Zheng * snapshot of the fs tree with special root objectid. 8955d4f98a2SYan Zheng */ 8965d4f98a2SYan Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 8975d4f98a2SYan Zheng struct btrfs_root *root) 8985d4f98a2SYan Zheng { 8995d4f98a2SYan Zheng struct btrfs_root *reloc_root; 9005d4f98a2SYan Zheng struct extent_buffer *eb; 9015d4f98a2SYan Zheng struct btrfs_root_item *root_item; 9025d4f98a2SYan Zheng struct btrfs_key root_key; 9035d4f98a2SYan Zheng int ret; 9045d4f98a2SYan Zheng 9055d4f98a2SYan Zheng if (root->reloc_root) { 9065d4f98a2SYan Zheng reloc_root = root->reloc_root; 9075d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 9085d4f98a2SYan Zheng return 0; 9095d4f98a2SYan Zheng } 9105d4f98a2SYan Zheng 9115d4f98a2SYan Zheng if (!root->fs_info->reloc_ctl || 9125d4f98a2SYan Zheng !root->fs_info->reloc_ctl->create_reloc_root || 9135d4f98a2SYan Zheng root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 9145d4f98a2SYan Zheng return 0; 9155d4f98a2SYan Zheng 9165d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 9175d4f98a2SYan Zheng BUG_ON(!root_item); 9185d4f98a2SYan Zheng 9195d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 9205d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 9215d4f98a2SYan Zheng root_key.offset = root->root_key.objectid; 9225d4f98a2SYan Zheng 9235d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 9245d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 9255d4f98a2SYan Zheng BUG_ON(ret); 9265d4f98a2SYan Zheng 9275d4f98a2SYan Zheng btrfs_set_root_last_snapshot(&root->root_item, trans->transid - 1); 9285d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 9295d4f98a2SYan Zheng btrfs_set_root_refs(root_item, 1); 9305d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 9315d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 9325d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 9335d4f98a2SYan Zheng memset(&root_item->drop_progress, 0, sizeof(struct btrfs_disk_key)); 9345d4f98a2SYan Zheng root_item->drop_level = 0; 9355d4f98a2SYan Zheng 9365d4f98a2SYan Zheng btrfs_tree_unlock(eb); 9375d4f98a2SYan Zheng free_extent_buffer(eb); 9385d4f98a2SYan Zheng 9395d4f98a2SYan Zheng ret = btrfs_insert_root(trans, root->fs_info->tree_root, 9405d4f98a2SYan Zheng &root_key, root_item); 9415d4f98a2SYan Zheng BUG_ON(ret); 9425d4f98a2SYan Zheng kfree(root_item); 9435d4f98a2SYan Zheng 9445d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root, 9455d4f98a2SYan Zheng &root_key); 9465d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 9475d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 9485d4f98a2SYan Zheng 9495d4f98a2SYan Zheng __add_reloc_root(reloc_root); 9505d4f98a2SYan Zheng root->reloc_root = reloc_root; 9515d4f98a2SYan Zheng return 0; 9525d4f98a2SYan Zheng } 9535d4f98a2SYan Zheng 9545d4f98a2SYan Zheng /* 9555d4f98a2SYan Zheng * update root item of reloc tree 9565d4f98a2SYan Zheng */ 9575d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 9585d4f98a2SYan Zheng struct btrfs_root *root) 9595d4f98a2SYan Zheng { 9605d4f98a2SYan Zheng struct btrfs_root *reloc_root; 9615d4f98a2SYan Zheng struct btrfs_root_item *root_item; 9625d4f98a2SYan Zheng int del = 0; 9635d4f98a2SYan Zheng int ret; 9645d4f98a2SYan Zheng 9655d4f98a2SYan Zheng if (!root->reloc_root) 9665d4f98a2SYan Zheng return 0; 9675d4f98a2SYan Zheng 9685d4f98a2SYan Zheng reloc_root = root->reloc_root; 9695d4f98a2SYan Zheng root_item = &reloc_root->root_item; 9705d4f98a2SYan Zheng 9715d4f98a2SYan Zheng if (btrfs_root_refs(root_item) == 0) { 9725d4f98a2SYan Zheng root->reloc_root = NULL; 9735d4f98a2SYan Zheng del = 1; 9745d4f98a2SYan Zheng } 9755d4f98a2SYan Zheng 9765d4f98a2SYan Zheng __update_reloc_root(reloc_root, del); 9775d4f98a2SYan Zheng 9785d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 9795d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 9805d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 9815d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 9825d4f98a2SYan Zheng } 9835d4f98a2SYan Zheng 9845d4f98a2SYan Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 9855d4f98a2SYan Zheng &reloc_root->root_key, root_item); 9865d4f98a2SYan Zheng BUG_ON(ret); 9875d4f98a2SYan Zheng return 0; 9885d4f98a2SYan Zheng } 9895d4f98a2SYan Zheng 9905d4f98a2SYan Zheng /* 9915d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 9925d4f98a2SYan Zheng * in a subvolume 9935d4f98a2SYan Zheng */ 9945d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 9955d4f98a2SYan Zheng { 9965d4f98a2SYan Zheng struct rb_node *node; 9975d4f98a2SYan Zheng struct rb_node *prev; 9985d4f98a2SYan Zheng struct btrfs_inode *entry; 9995d4f98a2SYan Zheng struct inode *inode; 10005d4f98a2SYan Zheng 10015d4f98a2SYan Zheng spin_lock(&root->inode_lock); 10025d4f98a2SYan Zheng again: 10035d4f98a2SYan Zheng node = root->inode_tree.rb_node; 10045d4f98a2SYan Zheng prev = NULL; 10055d4f98a2SYan Zheng while (node) { 10065d4f98a2SYan Zheng prev = node; 10075d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 10085d4f98a2SYan Zheng 10095d4f98a2SYan Zheng if (objectid < entry->vfs_inode.i_ino) 10105d4f98a2SYan Zheng node = node->rb_left; 10115d4f98a2SYan Zheng else if (objectid > entry->vfs_inode.i_ino) 10125d4f98a2SYan Zheng node = node->rb_right; 10135d4f98a2SYan Zheng else 10145d4f98a2SYan Zheng break; 10155d4f98a2SYan Zheng } 10165d4f98a2SYan Zheng if (!node) { 10175d4f98a2SYan Zheng while (prev) { 10185d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 10195d4f98a2SYan Zheng if (objectid <= entry->vfs_inode.i_ino) { 10205d4f98a2SYan Zheng node = prev; 10215d4f98a2SYan Zheng break; 10225d4f98a2SYan Zheng } 10235d4f98a2SYan Zheng prev = rb_next(prev); 10245d4f98a2SYan Zheng } 10255d4f98a2SYan Zheng } 10265d4f98a2SYan Zheng while (node) { 10275d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 10285d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 10295d4f98a2SYan Zheng if (inode) { 10305d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 10315d4f98a2SYan Zheng return inode; 10325d4f98a2SYan Zheng } 10335d4f98a2SYan Zheng 10345d4f98a2SYan Zheng objectid = entry->vfs_inode.i_ino + 1; 10355d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 10365d4f98a2SYan Zheng goto again; 10375d4f98a2SYan Zheng 10385d4f98a2SYan Zheng node = rb_next(node); 10395d4f98a2SYan Zheng } 10405d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 10415d4f98a2SYan Zheng return NULL; 10425d4f98a2SYan Zheng } 10435d4f98a2SYan Zheng 10445d4f98a2SYan Zheng static int in_block_group(u64 bytenr, 10455d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group) 10465d4f98a2SYan Zheng { 10475d4f98a2SYan Zheng if (bytenr >= block_group->key.objectid && 10485d4f98a2SYan Zheng bytenr < block_group->key.objectid + block_group->key.offset) 10495d4f98a2SYan Zheng return 1; 10505d4f98a2SYan Zheng return 0; 10515d4f98a2SYan Zheng } 10525d4f98a2SYan Zheng 10535d4f98a2SYan Zheng /* 10545d4f98a2SYan Zheng * get new location of data 10555d4f98a2SYan Zheng */ 10565d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 10575d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 10585d4f98a2SYan Zheng { 10595d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 10605d4f98a2SYan Zheng struct btrfs_path *path; 10615d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10625d4f98a2SYan Zheng struct extent_buffer *leaf; 10635d4f98a2SYan Zheng int ret; 10645d4f98a2SYan Zheng 10655d4f98a2SYan Zheng path = btrfs_alloc_path(); 10665d4f98a2SYan Zheng if (!path) 10675d4f98a2SYan Zheng return -ENOMEM; 10685d4f98a2SYan Zheng 10695d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 10705d4f98a2SYan Zheng ret = btrfs_lookup_file_extent(NULL, root, path, reloc_inode->i_ino, 10715d4f98a2SYan Zheng bytenr, 0); 10725d4f98a2SYan Zheng if (ret < 0) 10735d4f98a2SYan Zheng goto out; 10745d4f98a2SYan Zheng if (ret > 0) { 10755d4f98a2SYan Zheng ret = -ENOENT; 10765d4f98a2SYan Zheng goto out; 10775d4f98a2SYan Zheng } 10785d4f98a2SYan Zheng 10795d4f98a2SYan Zheng leaf = path->nodes[0]; 10805d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 10815d4f98a2SYan Zheng struct btrfs_file_extent_item); 10825d4f98a2SYan Zheng 10835d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 10845d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 10855d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 10865d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 10875d4f98a2SYan Zheng 10885d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 10895d4f98a2SYan Zheng ret = 1; 10905d4f98a2SYan Zheng goto out; 10915d4f98a2SYan Zheng } 10925d4f98a2SYan Zheng 10935d4f98a2SYan Zheng if (new_bytenr) 10945d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 10955d4f98a2SYan Zheng ret = 0; 10965d4f98a2SYan Zheng out: 10975d4f98a2SYan Zheng btrfs_free_path(path); 10985d4f98a2SYan Zheng return ret; 10995d4f98a2SYan Zheng } 11005d4f98a2SYan Zheng 11015d4f98a2SYan Zheng /* 11025d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 11035d4f98a2SYan Zheng * the new locations. 11045d4f98a2SYan Zheng */ 11055d4f98a2SYan Zheng static int replace_file_extents(struct btrfs_trans_handle *trans, 11065d4f98a2SYan Zheng struct reloc_control *rc, 11075d4f98a2SYan Zheng struct btrfs_root *root, 11085d4f98a2SYan Zheng struct extent_buffer *leaf, 11095d4f98a2SYan Zheng struct list_head *inode_list) 11105d4f98a2SYan Zheng { 11115d4f98a2SYan Zheng struct btrfs_key key; 11125d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 11135d4f98a2SYan Zheng struct inode *inode = NULL; 11145d4f98a2SYan Zheng struct inodevec *ivec = NULL; 11155d4f98a2SYan Zheng u64 parent; 11165d4f98a2SYan Zheng u64 bytenr; 11175d4f98a2SYan Zheng u64 new_bytenr; 11185d4f98a2SYan Zheng u64 num_bytes; 11195d4f98a2SYan Zheng u64 end; 11205d4f98a2SYan Zheng u32 nritems; 11215d4f98a2SYan Zheng u32 i; 11225d4f98a2SYan Zheng int ret; 11235d4f98a2SYan Zheng int first = 1; 11245d4f98a2SYan Zheng int dirty = 0; 11255d4f98a2SYan Zheng 11265d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 11275d4f98a2SYan Zheng return 0; 11285d4f98a2SYan Zheng 11295d4f98a2SYan Zheng /* reloc trees always use full backref */ 11305d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 11315d4f98a2SYan Zheng parent = leaf->start; 11325d4f98a2SYan Zheng else 11335d4f98a2SYan Zheng parent = 0; 11345d4f98a2SYan Zheng 11355d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 11365d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 11375d4f98a2SYan Zheng cond_resched(); 11385d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 11395d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 11405d4f98a2SYan Zheng continue; 11415d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 11425d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 11435d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 11445d4f98a2SYan Zheng continue; 11455d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 11465d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 11475d4f98a2SYan Zheng if (bytenr == 0) 11485d4f98a2SYan Zheng continue; 11495d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group)) 11505d4f98a2SYan Zheng continue; 11515d4f98a2SYan Zheng 11525d4f98a2SYan Zheng /* 11535d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 11545d4f98a2SYan Zheng * to complete and drop the extent cache 11555d4f98a2SYan Zheng */ 11565d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 11575d4f98a2SYan Zheng if (!ivec || ivec->nr == INODEVEC_SIZE) { 11585d4f98a2SYan Zheng ivec = kmalloc(sizeof(*ivec), GFP_NOFS); 11595d4f98a2SYan Zheng BUG_ON(!ivec); 11605d4f98a2SYan Zheng ivec->nr = 0; 11615d4f98a2SYan Zheng list_add_tail(&ivec->list, inode_list); 11625d4f98a2SYan Zheng } 11635d4f98a2SYan Zheng if (first) { 11645d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11655d4f98a2SYan Zheng if (inode) 11665d4f98a2SYan Zheng ivec->inode[ivec->nr++] = inode; 11675d4f98a2SYan Zheng first = 0; 11685d4f98a2SYan Zheng } else if (inode && inode->i_ino < key.objectid) { 11695d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11705d4f98a2SYan Zheng if (inode) 11715d4f98a2SYan Zheng ivec->inode[ivec->nr++] = inode; 11725d4f98a2SYan Zheng } 11735d4f98a2SYan Zheng if (inode && inode->i_ino == key.objectid) { 11745d4f98a2SYan Zheng end = key.offset + 11755d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 11765d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 11775d4f98a2SYan Zheng root->sectorsize)); 11785d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 11795d4f98a2SYan Zheng end--; 11805d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 11815d4f98a2SYan Zheng key.offset, end, 11825d4f98a2SYan Zheng GFP_NOFS); 11835d4f98a2SYan Zheng if (!ret) 11845d4f98a2SYan Zheng continue; 11855d4f98a2SYan Zheng 11865d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, key.offset, end, 11875d4f98a2SYan Zheng 1); 11885d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 11895d4f98a2SYan Zheng key.offset, end, GFP_NOFS); 11905d4f98a2SYan Zheng } 11915d4f98a2SYan Zheng } 11925d4f98a2SYan Zheng 11935d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 11945d4f98a2SYan Zheng bytenr, num_bytes); 11955d4f98a2SYan Zheng if (ret > 0) 11965d4f98a2SYan Zheng continue; 11975d4f98a2SYan Zheng BUG_ON(ret < 0); 11985d4f98a2SYan Zheng 11995d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 12005d4f98a2SYan Zheng dirty = 1; 12015d4f98a2SYan Zheng 12025d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 12035d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, new_bytenr, 12045d4f98a2SYan Zheng num_bytes, parent, 12055d4f98a2SYan Zheng btrfs_header_owner(leaf), 12065d4f98a2SYan Zheng key.objectid, key.offset); 12075d4f98a2SYan Zheng BUG_ON(ret); 12085d4f98a2SYan Zheng 12095d4f98a2SYan Zheng ret = btrfs_free_extent(trans, root, bytenr, num_bytes, 12105d4f98a2SYan Zheng parent, btrfs_header_owner(leaf), 12115d4f98a2SYan Zheng key.objectid, key.offset); 12125d4f98a2SYan Zheng BUG_ON(ret); 12135d4f98a2SYan Zheng } 12145d4f98a2SYan Zheng if (dirty) 12155d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 12165d4f98a2SYan Zheng return 0; 12175d4f98a2SYan Zheng } 12185d4f98a2SYan Zheng 12195d4f98a2SYan Zheng static noinline_for_stack 12205d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 12215d4f98a2SYan Zheng struct btrfs_path *path, int level) 12225d4f98a2SYan Zheng { 12235d4f98a2SYan Zheng struct btrfs_disk_key key1; 12245d4f98a2SYan Zheng struct btrfs_disk_key key2; 12255d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 12265d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 12275d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 12285d4f98a2SYan Zheng } 12295d4f98a2SYan Zheng 12305d4f98a2SYan Zheng /* 12315d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 12325d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 12335d4f98a2SYan Zheng * reloc tree was create can be replaced. 12345d4f98a2SYan Zheng * 12355d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 12365d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 12375d4f98a2SYan Zheng * errors, a negative error number is returned. 12385d4f98a2SYan Zheng */ 12395d4f98a2SYan Zheng static int replace_path(struct btrfs_trans_handle *trans, 12405d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 12415d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 12425d4f98a2SYan Zheng struct extent_buffer **leaf, 12435d4f98a2SYan Zheng int lowest_level, int max_level) 12445d4f98a2SYan Zheng { 12455d4f98a2SYan Zheng struct extent_buffer *eb; 12465d4f98a2SYan Zheng struct extent_buffer *parent; 12475d4f98a2SYan Zheng struct btrfs_key key; 12485d4f98a2SYan Zheng u64 old_bytenr; 12495d4f98a2SYan Zheng u64 new_bytenr; 12505d4f98a2SYan Zheng u64 old_ptr_gen; 12515d4f98a2SYan Zheng u64 new_ptr_gen; 12525d4f98a2SYan Zheng u64 last_snapshot; 12535d4f98a2SYan Zheng u32 blocksize; 12545d4f98a2SYan Zheng int level; 12555d4f98a2SYan Zheng int ret; 12565d4f98a2SYan Zheng int slot; 12575d4f98a2SYan Zheng 12585d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 12595d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 12605d4f98a2SYan Zheng BUG_ON(lowest_level > 1 && leaf); 12615d4f98a2SYan Zheng 12625d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 12635d4f98a2SYan Zheng 12645d4f98a2SYan Zheng slot = path->slots[lowest_level]; 12655d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 12665d4f98a2SYan Zheng 12675d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 12685d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 12695d4f98a2SYan Zheng level = btrfs_header_level(eb); 12705d4f98a2SYan Zheng 12715d4f98a2SYan Zheng if (level < lowest_level) { 12725d4f98a2SYan Zheng btrfs_tree_unlock(eb); 12735d4f98a2SYan Zheng free_extent_buffer(eb); 12745d4f98a2SYan Zheng return 0; 12755d4f98a2SYan Zheng } 12765d4f98a2SYan Zheng 12775d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 12785d4f98a2SYan Zheng BUG_ON(ret); 12795d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 12805d4f98a2SYan Zheng 12815d4f98a2SYan Zheng if (next_key) { 12825d4f98a2SYan Zheng next_key->objectid = (u64)-1; 12835d4f98a2SYan Zheng next_key->type = (u8)-1; 12845d4f98a2SYan Zheng next_key->offset = (u64)-1; 12855d4f98a2SYan Zheng } 12865d4f98a2SYan Zheng 12875d4f98a2SYan Zheng parent = eb; 12885d4f98a2SYan Zheng while (1) { 12895d4f98a2SYan Zheng level = btrfs_header_level(parent); 12905d4f98a2SYan Zheng BUG_ON(level < lowest_level); 12915d4f98a2SYan Zheng 12925d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 12935d4f98a2SYan Zheng if (ret && slot > 0) 12945d4f98a2SYan Zheng slot--; 12955d4f98a2SYan Zheng 12965d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 12975d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 12985d4f98a2SYan Zheng 12995d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 13005d4f98a2SYan Zheng blocksize = btrfs_level_size(dest, level - 1); 13015d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 13025d4f98a2SYan Zheng 13035d4f98a2SYan Zheng if (level <= max_level) { 13045d4f98a2SYan Zheng eb = path->nodes[level]; 13055d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 13065d4f98a2SYan Zheng path->slots[level]); 13075d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 13085d4f98a2SYan Zheng path->slots[level]); 13095d4f98a2SYan Zheng } else { 13105d4f98a2SYan Zheng new_bytenr = 0; 13115d4f98a2SYan Zheng new_ptr_gen = 0; 13125d4f98a2SYan Zheng } 13135d4f98a2SYan Zheng 13145d4f98a2SYan Zheng if (new_bytenr > 0 && new_bytenr == old_bytenr) { 13155d4f98a2SYan Zheng WARN_ON(1); 13165d4f98a2SYan Zheng ret = level; 13175d4f98a2SYan Zheng break; 13185d4f98a2SYan Zheng } 13195d4f98a2SYan Zheng 13205d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 13215d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 13225d4f98a2SYan Zheng if (level <= lowest_level && !leaf) { 13235d4f98a2SYan Zheng ret = 0; 13245d4f98a2SYan Zheng break; 13255d4f98a2SYan Zheng } 13265d4f98a2SYan Zheng 13275d4f98a2SYan Zheng eb = read_tree_block(dest, old_bytenr, blocksize, 13285d4f98a2SYan Zheng old_ptr_gen); 13295d4f98a2SYan Zheng btrfs_tree_lock(eb); 13305d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 13315d4f98a2SYan Zheng slot, &eb); 13325d4f98a2SYan Zheng BUG_ON(ret); 13335d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 13345d4f98a2SYan Zheng 13355d4f98a2SYan Zheng if (level <= lowest_level) { 13365d4f98a2SYan Zheng *leaf = eb; 13375d4f98a2SYan Zheng ret = 0; 13385d4f98a2SYan Zheng break; 13395d4f98a2SYan Zheng } 13405d4f98a2SYan Zheng 13415d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13425d4f98a2SYan Zheng free_extent_buffer(parent); 13435d4f98a2SYan Zheng 13445d4f98a2SYan Zheng parent = eb; 13455d4f98a2SYan Zheng continue; 13465d4f98a2SYan Zheng } 13475d4f98a2SYan Zheng 13485d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 13495d4f98a2SYan Zheng path->slots[level]); 13505d4f98a2SYan Zheng btrfs_release_path(src, path); 13515d4f98a2SYan Zheng 13525d4f98a2SYan Zheng path->lowest_level = level; 13535d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 13545d4f98a2SYan Zheng path->lowest_level = 0; 13555d4f98a2SYan Zheng BUG_ON(ret); 13565d4f98a2SYan Zheng 13575d4f98a2SYan Zheng /* 13585d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 13595d4f98a2SYan Zheng */ 13605d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 13615d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 13625d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 13635d4f98a2SYan Zheng 13645d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 13655d4f98a2SYan Zheng path->slots[level], old_bytenr); 13665d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 13675d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 13685d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 13695d4f98a2SYan Zheng 13705d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize, 13715d4f98a2SYan Zheng path->nodes[level]->start, 13725d4f98a2SYan Zheng src->root_key.objectid, level - 1, 0); 13735d4f98a2SYan Zheng BUG_ON(ret); 13745d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize, 13755d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 13765d4f98a2SYan Zheng 0); 13775d4f98a2SYan Zheng BUG_ON(ret); 13785d4f98a2SYan Zheng 13795d4f98a2SYan Zheng ret = btrfs_free_extent(trans, src, new_bytenr, blocksize, 13805d4f98a2SYan Zheng path->nodes[level]->start, 13815d4f98a2SYan Zheng src->root_key.objectid, level - 1, 0); 13825d4f98a2SYan Zheng BUG_ON(ret); 13835d4f98a2SYan Zheng 13845d4f98a2SYan Zheng ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize, 13855d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 13865d4f98a2SYan Zheng 0); 13875d4f98a2SYan Zheng BUG_ON(ret); 13885d4f98a2SYan Zheng 13895d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 13905d4f98a2SYan Zheng 13915d4f98a2SYan Zheng ret = level; 13925d4f98a2SYan Zheng break; 13935d4f98a2SYan Zheng } 13945d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13955d4f98a2SYan Zheng free_extent_buffer(parent); 13965d4f98a2SYan Zheng return ret; 13975d4f98a2SYan Zheng } 13985d4f98a2SYan Zheng 13995d4f98a2SYan Zheng /* 14005d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 14015d4f98a2SYan Zheng */ 14025d4f98a2SYan Zheng static noinline_for_stack 14035d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14045d4f98a2SYan Zheng int *level) 14055d4f98a2SYan Zheng { 14065d4f98a2SYan Zheng struct extent_buffer *eb; 14075d4f98a2SYan Zheng int i; 14085d4f98a2SYan Zheng u64 last_snapshot; 14095d4f98a2SYan Zheng u32 nritems; 14105d4f98a2SYan Zheng 14115d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14125d4f98a2SYan Zheng 14135d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 14145d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14155d4f98a2SYan Zheng path->nodes[i] = NULL; 14165d4f98a2SYan Zheng } 14175d4f98a2SYan Zheng 14185d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 14195d4f98a2SYan Zheng eb = path->nodes[i]; 14205d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14215d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 14225d4f98a2SYan Zheng path->slots[i]++; 14235d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 14245d4f98a2SYan Zheng last_snapshot) 14255d4f98a2SYan Zheng continue; 14265d4f98a2SYan Zheng 14275d4f98a2SYan Zheng *level = i; 14285d4f98a2SYan Zheng return 0; 14295d4f98a2SYan Zheng } 14305d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14315d4f98a2SYan Zheng path->nodes[i] = NULL; 14325d4f98a2SYan Zheng } 14335d4f98a2SYan Zheng return 1; 14345d4f98a2SYan Zheng } 14355d4f98a2SYan Zheng 14365d4f98a2SYan Zheng /* 14375d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 14385d4f98a2SYan Zheng */ 14395d4f98a2SYan Zheng static noinline_for_stack 14405d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14415d4f98a2SYan Zheng int *level) 14425d4f98a2SYan Zheng { 14435d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 14445d4f98a2SYan Zheng int i; 14455d4f98a2SYan Zheng u64 bytenr; 14465d4f98a2SYan Zheng u64 ptr_gen = 0; 14475d4f98a2SYan Zheng u64 last_snapshot; 14485d4f98a2SYan Zheng u32 blocksize; 14495d4f98a2SYan Zheng u32 nritems; 14505d4f98a2SYan Zheng 14515d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14525d4f98a2SYan Zheng 14535d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 14545d4f98a2SYan Zheng eb = path->nodes[i]; 14555d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14565d4f98a2SYan Zheng while (path->slots[i] < nritems) { 14575d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 14585d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 14595d4f98a2SYan Zheng break; 14605d4f98a2SYan Zheng path->slots[i]++; 14615d4f98a2SYan Zheng } 14625d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 14635d4f98a2SYan Zheng if (i == *level) 14645d4f98a2SYan Zheng break; 14655d4f98a2SYan Zheng *level = i + 1; 14665d4f98a2SYan Zheng return 0; 14675d4f98a2SYan Zheng } 14685d4f98a2SYan Zheng if (i == 1) { 14695d4f98a2SYan Zheng *level = i; 14705d4f98a2SYan Zheng return 0; 14715d4f98a2SYan Zheng } 14725d4f98a2SYan Zheng 14735d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 14745d4f98a2SYan Zheng blocksize = btrfs_level_size(root, i - 1); 14755d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, ptr_gen); 14765d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 14775d4f98a2SYan Zheng path->nodes[i - 1] = eb; 14785d4f98a2SYan Zheng path->slots[i - 1] = 0; 14795d4f98a2SYan Zheng } 14805d4f98a2SYan Zheng return 1; 14815d4f98a2SYan Zheng } 14825d4f98a2SYan Zheng 14835d4f98a2SYan Zheng /* 14845d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 14855d4f98a2SYan Zheng * [min_key, max_key) 14865d4f98a2SYan Zheng */ 14875d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 14885d4f98a2SYan Zheng struct btrfs_key *min_key, 14895d4f98a2SYan Zheng struct btrfs_key *max_key) 14905d4f98a2SYan Zheng { 14915d4f98a2SYan Zheng struct inode *inode = NULL; 14925d4f98a2SYan Zheng u64 objectid; 14935d4f98a2SYan Zheng u64 start, end; 14945d4f98a2SYan Zheng 14955d4f98a2SYan Zheng objectid = min_key->objectid; 14965d4f98a2SYan Zheng while (1) { 14975d4f98a2SYan Zheng cond_resched(); 14985d4f98a2SYan Zheng iput(inode); 14995d4f98a2SYan Zheng 15005d4f98a2SYan Zheng if (objectid > max_key->objectid) 15015d4f98a2SYan Zheng break; 15025d4f98a2SYan Zheng 15035d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 15045d4f98a2SYan Zheng if (!inode) 15055d4f98a2SYan Zheng break; 15065d4f98a2SYan Zheng 15075d4f98a2SYan Zheng if (inode->i_ino > max_key->objectid) { 15085d4f98a2SYan Zheng iput(inode); 15095d4f98a2SYan Zheng break; 15105d4f98a2SYan Zheng } 15115d4f98a2SYan Zheng 15125d4f98a2SYan Zheng objectid = inode->i_ino + 1; 15135d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 15145d4f98a2SYan Zheng continue; 15155d4f98a2SYan Zheng 15165d4f98a2SYan Zheng if (unlikely(min_key->objectid == inode->i_ino)) { 15175d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 15185d4f98a2SYan Zheng continue; 15195d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 15205d4f98a2SYan Zheng start = 0; 15215d4f98a2SYan Zheng else { 15225d4f98a2SYan Zheng start = min_key->offset; 15235d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(start, root->sectorsize)); 15245d4f98a2SYan Zheng } 15255d4f98a2SYan Zheng } else { 15265d4f98a2SYan Zheng start = 0; 15275d4f98a2SYan Zheng } 15285d4f98a2SYan Zheng 15295d4f98a2SYan Zheng if (unlikely(max_key->objectid == inode->i_ino)) { 15305d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 15315d4f98a2SYan Zheng continue; 15325d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 15335d4f98a2SYan Zheng end = (u64)-1; 15345d4f98a2SYan Zheng } else { 15355d4f98a2SYan Zheng if (max_key->offset == 0) 15365d4f98a2SYan Zheng continue; 15375d4f98a2SYan Zheng end = max_key->offset; 15385d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 15395d4f98a2SYan Zheng end--; 15405d4f98a2SYan Zheng } 15415d4f98a2SYan Zheng } else { 15425d4f98a2SYan Zheng end = (u64)-1; 15435d4f98a2SYan Zheng } 15445d4f98a2SYan Zheng 15455d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 15465d4f98a2SYan Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 15475d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 1); 15485d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 15495d4f98a2SYan Zheng } 15505d4f98a2SYan Zheng return 0; 15515d4f98a2SYan Zheng } 15525d4f98a2SYan Zheng 15535d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 15545d4f98a2SYan Zheng struct btrfs_key *key) 15555d4f98a2SYan Zheng 15565d4f98a2SYan Zheng { 15575d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 15585d4f98a2SYan Zheng if (!path->nodes[level]) 15595d4f98a2SYan Zheng break; 15605d4f98a2SYan Zheng if (path->slots[level] + 1 < 15615d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 15625d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 15635d4f98a2SYan Zheng path->slots[level] + 1); 15645d4f98a2SYan Zheng return 0; 15655d4f98a2SYan Zheng } 15665d4f98a2SYan Zheng level++; 15675d4f98a2SYan Zheng } 15685d4f98a2SYan Zheng return 1; 15695d4f98a2SYan Zheng } 15705d4f98a2SYan Zheng 15715d4f98a2SYan Zheng /* 15725d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 15735d4f98a2SYan Zheng * fs tree. 15745d4f98a2SYan Zheng */ 15755d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 15765d4f98a2SYan Zheng struct btrfs_root *root) 15775d4f98a2SYan Zheng { 15785d4f98a2SYan Zheng LIST_HEAD(inode_list); 15795d4f98a2SYan Zheng struct btrfs_key key; 15805d4f98a2SYan Zheng struct btrfs_key next_key; 15815d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 15825d4f98a2SYan Zheng struct btrfs_root *reloc_root; 15835d4f98a2SYan Zheng struct btrfs_root_item *root_item; 15845d4f98a2SYan Zheng struct btrfs_path *path; 15855d4f98a2SYan Zheng struct extent_buffer *leaf = NULL; 15865d4f98a2SYan Zheng unsigned long nr; 15875d4f98a2SYan Zheng int level; 15885d4f98a2SYan Zheng int max_level; 15895d4f98a2SYan Zheng int replaced = 0; 15905d4f98a2SYan Zheng int ret; 15915d4f98a2SYan Zheng int err = 0; 15925d4f98a2SYan Zheng 15935d4f98a2SYan Zheng path = btrfs_alloc_path(); 15945d4f98a2SYan Zheng if (!path) 15955d4f98a2SYan Zheng return -ENOMEM; 15965d4f98a2SYan Zheng 15975d4f98a2SYan Zheng reloc_root = root->reloc_root; 15985d4f98a2SYan Zheng root_item = &reloc_root->root_item; 15995d4f98a2SYan Zheng 16005d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 16015d4f98a2SYan Zheng level = btrfs_root_level(root_item); 16025d4f98a2SYan Zheng extent_buffer_get(reloc_root->node); 16035d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 16045d4f98a2SYan Zheng path->slots[level] = 0; 16055d4f98a2SYan Zheng } else { 16065d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 16075d4f98a2SYan Zheng 16085d4f98a2SYan Zheng level = root_item->drop_level; 16095d4f98a2SYan Zheng BUG_ON(level == 0); 16105d4f98a2SYan Zheng path->lowest_level = level; 16115d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 16125d4f98a2SYan Zheng if (ret < 0) { 16135d4f98a2SYan Zheng btrfs_free_path(path); 16145d4f98a2SYan Zheng return ret; 16155d4f98a2SYan Zheng } 16165d4f98a2SYan Zheng 16175d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 16185d4f98a2SYan Zheng path->slots[level]); 16195d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 16205d4f98a2SYan Zheng 16215d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 16225d4f98a2SYan Zheng } 16235d4f98a2SYan Zheng 16245d4f98a2SYan Zheng if (level == 0 && rc->stage == UPDATE_DATA_PTRS) { 16255d4f98a2SYan Zheng trans = btrfs_start_transaction(root, 1); 16265d4f98a2SYan Zheng 16275d4f98a2SYan Zheng leaf = path->nodes[0]; 16285d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, 0); 16295d4f98a2SYan Zheng btrfs_release_path(reloc_root, path); 16305d4f98a2SYan Zheng 16315d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, &key, path, 0, 1); 16325d4f98a2SYan Zheng if (ret < 0) { 16335d4f98a2SYan Zheng err = ret; 16345d4f98a2SYan Zheng goto out; 16355d4f98a2SYan Zheng } 16365d4f98a2SYan Zheng 16375d4f98a2SYan Zheng leaf = path->nodes[0]; 16385d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 1); 16395d4f98a2SYan Zheng ret = replace_file_extents(trans, rc, root, leaf, 16405d4f98a2SYan Zheng &inode_list); 16415d4f98a2SYan Zheng if (ret < 0) 16425d4f98a2SYan Zheng err = ret; 16435d4f98a2SYan Zheng goto out; 16445d4f98a2SYan Zheng } 16455d4f98a2SYan Zheng 16465d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 16475d4f98a2SYan Zheng 16485d4f98a2SYan Zheng while (1) { 16495d4f98a2SYan Zheng leaf = NULL; 16505d4f98a2SYan Zheng replaced = 0; 16515d4f98a2SYan Zheng trans = btrfs_start_transaction(root, 1); 16525d4f98a2SYan Zheng max_level = level; 16535d4f98a2SYan Zheng 16545d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 16555d4f98a2SYan Zheng if (ret < 0) { 16565d4f98a2SYan Zheng err = ret; 16575d4f98a2SYan Zheng goto out; 16585d4f98a2SYan Zheng } 16595d4f98a2SYan Zheng if (ret > 0) 16605d4f98a2SYan Zheng break; 16615d4f98a2SYan Zheng 16625d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 16635d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 16645d4f98a2SYan Zheng ret = 0; 16655d4f98a2SYan Zheng } else if (level == 1 && rc->stage == UPDATE_DATA_PTRS) { 16665d4f98a2SYan Zheng ret = replace_path(trans, root, reloc_root, 16675d4f98a2SYan Zheng path, &next_key, &leaf, 16685d4f98a2SYan Zheng level, max_level); 16695d4f98a2SYan Zheng } else { 16705d4f98a2SYan Zheng ret = replace_path(trans, root, reloc_root, 16715d4f98a2SYan Zheng path, &next_key, NULL, 16725d4f98a2SYan Zheng level, max_level); 16735d4f98a2SYan Zheng } 16745d4f98a2SYan Zheng if (ret < 0) { 16755d4f98a2SYan Zheng err = ret; 16765d4f98a2SYan Zheng goto out; 16775d4f98a2SYan Zheng } 16785d4f98a2SYan Zheng 16795d4f98a2SYan Zheng if (ret > 0) { 16805d4f98a2SYan Zheng level = ret; 16815d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 16825d4f98a2SYan Zheng path->slots[level]); 16835d4f98a2SYan Zheng replaced = 1; 16845d4f98a2SYan Zheng } else if (leaf) { 16855d4f98a2SYan Zheng /* 16865d4f98a2SYan Zheng * no block got replaced, try replacing file extents 16875d4f98a2SYan Zheng */ 16885d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, 0); 16895d4f98a2SYan Zheng ret = replace_file_extents(trans, rc, root, leaf, 16905d4f98a2SYan Zheng &inode_list); 16915d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 16925d4f98a2SYan Zheng free_extent_buffer(leaf); 16935d4f98a2SYan Zheng BUG_ON(ret < 0); 16945d4f98a2SYan Zheng } 16955d4f98a2SYan Zheng 16965d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 16975d4f98a2SYan Zheng if (ret > 0) 16985d4f98a2SYan Zheng break; 16995d4f98a2SYan Zheng 17005d4f98a2SYan Zheng BUG_ON(level == 0); 17015d4f98a2SYan Zheng /* 17025d4f98a2SYan Zheng * save the merging progress in the drop_progress. 17035d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 17045d4f98a2SYan Zheng */ 17055d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 17065d4f98a2SYan Zheng path->slots[level]); 17075d4f98a2SYan Zheng root_item->drop_level = level; 17085d4f98a2SYan Zheng 17095d4f98a2SYan Zheng nr = trans->blocks_used; 17105d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 17115d4f98a2SYan Zheng 17125d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 17135d4f98a2SYan Zheng 17145d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17155d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17165d4f98a2SYan Zheng } 17175d4f98a2SYan Zheng 17185d4f98a2SYan Zheng /* 17195d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 17205d4f98a2SYan Zheng * relocated and the block is tree root. 17215d4f98a2SYan Zheng */ 17225d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 17235d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 17245d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 17255d4f98a2SYan Zheng free_extent_buffer(leaf); 17265d4f98a2SYan Zheng if (ret < 0) 17275d4f98a2SYan Zheng err = ret; 17285d4f98a2SYan Zheng out: 17295d4f98a2SYan Zheng btrfs_free_path(path); 17305d4f98a2SYan Zheng 17315d4f98a2SYan Zheng if (err == 0) { 17325d4f98a2SYan Zheng memset(&root_item->drop_progress, 0, 17335d4f98a2SYan Zheng sizeof(root_item->drop_progress)); 17345d4f98a2SYan Zheng root_item->drop_level = 0; 17355d4f98a2SYan Zheng btrfs_set_root_refs(root_item, 0); 17365d4f98a2SYan Zheng } 17375d4f98a2SYan Zheng 17385d4f98a2SYan Zheng nr = trans->blocks_used; 17395d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 17405d4f98a2SYan Zheng 17415d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 17425d4f98a2SYan Zheng 17435d4f98a2SYan Zheng /* 17445d4f98a2SYan Zheng * put inodes while we aren't holding the tree locks 17455d4f98a2SYan Zheng */ 17465d4f98a2SYan Zheng while (!list_empty(&inode_list)) { 17475d4f98a2SYan Zheng struct inodevec *ivec; 17485d4f98a2SYan Zheng ivec = list_entry(inode_list.next, struct inodevec, list); 17495d4f98a2SYan Zheng list_del(&ivec->list); 17505d4f98a2SYan Zheng while (ivec->nr > 0) { 17515d4f98a2SYan Zheng ivec->nr--; 17525d4f98a2SYan Zheng iput(ivec->inode[ivec->nr]); 17535d4f98a2SYan Zheng } 17545d4f98a2SYan Zheng kfree(ivec); 17555d4f98a2SYan Zheng } 17565d4f98a2SYan Zheng 17575d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17585d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17595d4f98a2SYan Zheng 17605d4f98a2SYan Zheng return err; 17615d4f98a2SYan Zheng } 17625d4f98a2SYan Zheng 17635d4f98a2SYan Zheng /* 17645d4f98a2SYan Zheng * callback for the work threads. 17655d4f98a2SYan Zheng * this function merges reloc tree with corresponding fs tree, 17665d4f98a2SYan Zheng * and then drops the reloc tree. 17675d4f98a2SYan Zheng */ 17685d4f98a2SYan Zheng static void merge_func(struct btrfs_work *work) 17695d4f98a2SYan Zheng { 17705d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 17715d4f98a2SYan Zheng struct btrfs_root *root; 17725d4f98a2SYan Zheng struct btrfs_root *reloc_root; 17735d4f98a2SYan Zheng struct async_merge *async; 17745d4f98a2SYan Zheng 17755d4f98a2SYan Zheng async = container_of(work, struct async_merge, work); 17765d4f98a2SYan Zheng reloc_root = async->root; 17775d4f98a2SYan Zheng 17785d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 17795d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, 17805d4f98a2SYan Zheng reloc_root->root_key.offset); 17815d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 17825d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 17835d4f98a2SYan Zheng 17845d4f98a2SYan Zheng merge_reloc_root(async->rc, root); 17855d4f98a2SYan Zheng 17865d4f98a2SYan Zheng trans = btrfs_start_transaction(root, 1); 17875d4f98a2SYan Zheng btrfs_update_reloc_root(trans, root); 17885d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 17895d4f98a2SYan Zheng } 17905d4f98a2SYan Zheng 17912c47e605SYan Zheng btrfs_drop_snapshot(reloc_root, 0); 17925d4f98a2SYan Zheng 17935d4f98a2SYan Zheng if (atomic_dec_and_test(async->num_pending)) 17945d4f98a2SYan Zheng complete(async->done); 17955d4f98a2SYan Zheng 17965d4f98a2SYan Zheng kfree(async); 17975d4f98a2SYan Zheng } 17985d4f98a2SYan Zheng 17995d4f98a2SYan Zheng static int merge_reloc_roots(struct reloc_control *rc) 18005d4f98a2SYan Zheng { 18015d4f98a2SYan Zheng struct async_merge *async; 18025d4f98a2SYan Zheng struct btrfs_root *root; 18035d4f98a2SYan Zheng struct completion done; 18045d4f98a2SYan Zheng atomic_t num_pending; 18055d4f98a2SYan Zheng 18065d4f98a2SYan Zheng init_completion(&done); 18075d4f98a2SYan Zheng atomic_set(&num_pending, 1); 18085d4f98a2SYan Zheng 18095d4f98a2SYan Zheng while (!list_empty(&rc->reloc_roots)) { 18105d4f98a2SYan Zheng root = list_entry(rc->reloc_roots.next, 18115d4f98a2SYan Zheng struct btrfs_root, root_list); 18125d4f98a2SYan Zheng list_del_init(&root->root_list); 18135d4f98a2SYan Zheng 18145d4f98a2SYan Zheng async = kmalloc(sizeof(*async), GFP_NOFS); 18155d4f98a2SYan Zheng BUG_ON(!async); 18165d4f98a2SYan Zheng async->work.func = merge_func; 18175d4f98a2SYan Zheng async->work.flags = 0; 18185d4f98a2SYan Zheng async->rc = rc; 18195d4f98a2SYan Zheng async->root = root; 18205d4f98a2SYan Zheng async->done = &done; 18215d4f98a2SYan Zheng async->num_pending = &num_pending; 18225d4f98a2SYan Zheng atomic_inc(&num_pending); 18235d4f98a2SYan Zheng btrfs_queue_worker(&rc->workers, &async->work); 18245d4f98a2SYan Zheng } 18255d4f98a2SYan Zheng 18265d4f98a2SYan Zheng if (!atomic_dec_and_test(&num_pending)) 18275d4f98a2SYan Zheng wait_for_completion(&done); 18285d4f98a2SYan Zheng 18295d4f98a2SYan Zheng BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 18305d4f98a2SYan Zheng return 0; 18315d4f98a2SYan Zheng } 18325d4f98a2SYan Zheng 18335d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 18345d4f98a2SYan Zheng { 18355d4f98a2SYan Zheng struct tree_block *block; 18365d4f98a2SYan Zheng struct rb_node *rb_node; 18375d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 18385d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 18395d4f98a2SYan Zheng rb_erase(rb_node, blocks); 18405d4f98a2SYan Zheng kfree(block); 18415d4f98a2SYan Zheng } 18425d4f98a2SYan Zheng } 18435d4f98a2SYan Zheng 18445d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 18455d4f98a2SYan Zheng struct btrfs_root *reloc_root) 18465d4f98a2SYan Zheng { 18475d4f98a2SYan Zheng struct btrfs_root *root; 18485d4f98a2SYan Zheng 18495d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 18505d4f98a2SYan Zheng return 0; 18515d4f98a2SYan Zheng 18525d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset); 18535d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 18545d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 18555d4f98a2SYan Zheng 18565d4f98a2SYan Zheng return btrfs_record_root_in_trans(trans, root); 18575d4f98a2SYan Zheng } 18585d4f98a2SYan Zheng 18595d4f98a2SYan Zheng /* 18605d4f98a2SYan Zheng * select one tree from trees that references the block. 18615d4f98a2SYan Zheng * for blocks in refernce counted trees, we preper reloc tree. 18625d4f98a2SYan Zheng * if no reloc tree found and reloc_only is true, NULL is returned. 18635d4f98a2SYan Zheng */ 18645d4f98a2SYan Zheng static struct btrfs_root *__select_one_root(struct btrfs_trans_handle *trans, 18655d4f98a2SYan Zheng struct backref_node *node, 18665d4f98a2SYan Zheng struct backref_edge *edges[], 18675d4f98a2SYan Zheng int *nr, int reloc_only) 18685d4f98a2SYan Zheng { 18695d4f98a2SYan Zheng struct backref_node *next; 18705d4f98a2SYan Zheng struct btrfs_root *root; 18715d4f98a2SYan Zheng int index; 18725d4f98a2SYan Zheng int loop = 0; 18735d4f98a2SYan Zheng again: 18745d4f98a2SYan Zheng index = 0; 18755d4f98a2SYan Zheng next = node; 18765d4f98a2SYan Zheng while (1) { 18775d4f98a2SYan Zheng cond_resched(); 18785d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 18795d4f98a2SYan Zheng root = next->root; 18805d4f98a2SYan Zheng if (!root) { 18815d4f98a2SYan Zheng BUG_ON(!node->old_root); 18825d4f98a2SYan Zheng goto skip; 18835d4f98a2SYan Zheng } 18845d4f98a2SYan Zheng 18855d4f98a2SYan Zheng /* no other choice for non-refernce counted tree */ 18865d4f98a2SYan Zheng if (!root->ref_cows) { 18875d4f98a2SYan Zheng BUG_ON(reloc_only); 18885d4f98a2SYan Zheng break; 18895d4f98a2SYan Zheng } 18905d4f98a2SYan Zheng 18915d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 18925d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 18935d4f98a2SYan Zheng break; 18945d4f98a2SYan Zheng } 18955d4f98a2SYan Zheng 18965d4f98a2SYan Zheng if (loop) { 18975d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 18985d4f98a2SYan Zheng break; 18995d4f98a2SYan Zheng } 19005d4f98a2SYan Zheng 19015d4f98a2SYan Zheng if (reloc_only || next != node) { 19025d4f98a2SYan Zheng if (!root->reloc_root) 19035d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 19045d4f98a2SYan Zheng root = root->reloc_root; 19055d4f98a2SYan Zheng /* 19065d4f98a2SYan Zheng * if the reloc tree was created in current 19075d4f98a2SYan Zheng * transation, there is no node in backref tree 19085d4f98a2SYan Zheng * corresponds to the root of the reloc tree. 19095d4f98a2SYan Zheng */ 19105d4f98a2SYan Zheng if (btrfs_root_last_snapshot(&root->root_item) == 19115d4f98a2SYan Zheng trans->transid - 1) 19125d4f98a2SYan Zheng break; 19135d4f98a2SYan Zheng } 19145d4f98a2SYan Zheng skip: 19155d4f98a2SYan Zheng root = NULL; 19165d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 19175d4f98a2SYan Zheng if (!next || next->level <= node->level) 19185d4f98a2SYan Zheng break; 19195d4f98a2SYan Zheng } 19205d4f98a2SYan Zheng 19215d4f98a2SYan Zheng if (!root && !loop && !reloc_only) { 19225d4f98a2SYan Zheng loop = 1; 19235d4f98a2SYan Zheng goto again; 19245d4f98a2SYan Zheng } 19255d4f98a2SYan Zheng 19265d4f98a2SYan Zheng if (root) 19275d4f98a2SYan Zheng *nr = index; 19285d4f98a2SYan Zheng else 19295d4f98a2SYan Zheng *nr = 0; 19305d4f98a2SYan Zheng 19315d4f98a2SYan Zheng return root; 19325d4f98a2SYan Zheng } 19335d4f98a2SYan Zheng 19345d4f98a2SYan Zheng static noinline_for_stack 19355d4f98a2SYan Zheng struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans, 19365d4f98a2SYan Zheng struct backref_node *node) 19375d4f98a2SYan Zheng { 19385d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 19395d4f98a2SYan Zheng int nr; 19405d4f98a2SYan Zheng return __select_one_root(trans, node, edges, &nr, 0); 19415d4f98a2SYan Zheng } 19425d4f98a2SYan Zheng 19435d4f98a2SYan Zheng static noinline_for_stack 19445d4f98a2SYan Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 19455d4f98a2SYan Zheng struct backref_node *node, 19465d4f98a2SYan Zheng struct backref_edge *edges[], int *nr) 19475d4f98a2SYan Zheng { 19485d4f98a2SYan Zheng return __select_one_root(trans, node, edges, nr, 1); 19495d4f98a2SYan Zheng } 19505d4f98a2SYan Zheng 19515d4f98a2SYan Zheng static void grab_path_buffers(struct btrfs_path *path, 19525d4f98a2SYan Zheng struct backref_node *node, 19535d4f98a2SYan Zheng struct backref_edge *edges[], int nr) 19545d4f98a2SYan Zheng { 19555d4f98a2SYan Zheng int i = 0; 19565d4f98a2SYan Zheng while (1) { 19575d4f98a2SYan Zheng drop_node_buffer(node); 19585d4f98a2SYan Zheng node->eb = path->nodes[node->level]; 19595d4f98a2SYan Zheng BUG_ON(!node->eb); 19605d4f98a2SYan Zheng if (path->locks[node->level]) 19615d4f98a2SYan Zheng node->locked = 1; 19625d4f98a2SYan Zheng path->nodes[node->level] = NULL; 19635d4f98a2SYan Zheng path->locks[node->level] = 0; 19645d4f98a2SYan Zheng 19655d4f98a2SYan Zheng if (i >= nr) 19665d4f98a2SYan Zheng break; 19675d4f98a2SYan Zheng 19685d4f98a2SYan Zheng edges[i]->blockptr = node->eb->start; 19695d4f98a2SYan Zheng node = edges[i]->node[UPPER]; 19705d4f98a2SYan Zheng i++; 19715d4f98a2SYan Zheng } 19725d4f98a2SYan Zheng } 19735d4f98a2SYan Zheng 19745d4f98a2SYan Zheng /* 19755d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 19765d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 19775d4f98a2SYan Zheng * 19785d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 19795d4f98a2SYan Zheng * in that case this function just updates pointers. 19805d4f98a2SYan Zheng */ 19815d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 19825d4f98a2SYan Zheng struct backref_node *node, 19835d4f98a2SYan Zheng struct btrfs_key *key, 19845d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 19855d4f98a2SYan Zheng { 19865d4f98a2SYan Zheng struct backref_node *upper; 19875d4f98a2SYan Zheng struct backref_edge *edge; 19885d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 19895d4f98a2SYan Zheng struct btrfs_root *root; 19905d4f98a2SYan Zheng struct extent_buffer *eb; 19915d4f98a2SYan Zheng u32 blocksize; 19925d4f98a2SYan Zheng u64 bytenr; 19935d4f98a2SYan Zheng u64 generation; 19945d4f98a2SYan Zheng int nr; 19955d4f98a2SYan Zheng int slot; 19965d4f98a2SYan Zheng int ret; 19975d4f98a2SYan Zheng int err = 0; 19985d4f98a2SYan Zheng 19995d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 20005d4f98a2SYan Zheng 20015d4f98a2SYan Zheng path->lowest_level = node->level + 1; 20025d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 20035d4f98a2SYan Zheng cond_resched(); 20045d4f98a2SYan Zheng if (node->eb && node->eb->start == edge->blockptr) 20055d4f98a2SYan Zheng continue; 20065d4f98a2SYan Zheng 20075d4f98a2SYan Zheng upper = edge->node[UPPER]; 20085d4f98a2SYan Zheng root = select_reloc_root(trans, upper, edges, &nr); 20095d4f98a2SYan Zheng if (!root) 20105d4f98a2SYan Zheng continue; 20115d4f98a2SYan Zheng 20125d4f98a2SYan Zheng if (upper->eb && !upper->locked) 20135d4f98a2SYan Zheng drop_node_buffer(upper); 20145d4f98a2SYan Zheng 20155d4f98a2SYan Zheng if (!upper->eb) { 20165d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 20175d4f98a2SYan Zheng if (ret < 0) { 20185d4f98a2SYan Zheng err = ret; 20195d4f98a2SYan Zheng break; 20205d4f98a2SYan Zheng } 20215d4f98a2SYan Zheng BUG_ON(ret > 0); 20225d4f98a2SYan Zheng 20235d4f98a2SYan Zheng slot = path->slots[upper->level]; 20245d4f98a2SYan Zheng 20255d4f98a2SYan Zheng btrfs_unlock_up_safe(path, upper->level + 1); 20265d4f98a2SYan Zheng grab_path_buffers(path, upper, edges, nr); 20275d4f98a2SYan Zheng 20285d4f98a2SYan Zheng btrfs_release_path(NULL, path); 20295d4f98a2SYan Zheng } else { 20305d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 20315d4f98a2SYan Zheng &slot); 20325d4f98a2SYan Zheng BUG_ON(ret); 20335d4f98a2SYan Zheng } 20345d4f98a2SYan Zheng 20355d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 20365d4f98a2SYan Zheng if (!lowest) { 20375d4f98a2SYan Zheng if (node->eb->start == bytenr) { 20385d4f98a2SYan Zheng btrfs_tree_unlock(upper->eb); 20395d4f98a2SYan Zheng upper->locked = 0; 20405d4f98a2SYan Zheng continue; 20415d4f98a2SYan Zheng } 20425d4f98a2SYan Zheng } else { 20435d4f98a2SYan Zheng BUG_ON(node->bytenr != bytenr); 20445d4f98a2SYan Zheng } 20455d4f98a2SYan Zheng 20465d4f98a2SYan Zheng blocksize = btrfs_level_size(root, node->level); 20475d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 20485d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, generation); 20495d4f98a2SYan Zheng btrfs_tree_lock(eb); 20505d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 20515d4f98a2SYan Zheng 20525d4f98a2SYan Zheng if (!node->eb) { 20535d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 20545d4f98a2SYan Zheng slot, &eb); 20555d4f98a2SYan Zheng if (ret < 0) { 20565d4f98a2SYan Zheng err = ret; 20575d4f98a2SYan Zheng break; 20585d4f98a2SYan Zheng } 20595d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 20605d4f98a2SYan Zheng node->eb = eb; 20615d4f98a2SYan Zheng node->locked = 1; 20625d4f98a2SYan Zheng } else { 20635d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 20645d4f98a2SYan Zheng node->eb->start); 20655d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 20665d4f98a2SYan Zheng trans->transid); 20675d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 20685d4f98a2SYan Zheng 20695d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, 20705d4f98a2SYan Zheng node->eb->start, blocksize, 20715d4f98a2SYan Zheng upper->eb->start, 20725d4f98a2SYan Zheng btrfs_header_owner(upper->eb), 20735d4f98a2SYan Zheng node->level, 0); 20745d4f98a2SYan Zheng BUG_ON(ret); 20755d4f98a2SYan Zheng 20765d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 20775d4f98a2SYan Zheng BUG_ON(ret); 20785d4f98a2SYan Zheng } 20795d4f98a2SYan Zheng if (!lowest) { 20805d4f98a2SYan Zheng btrfs_tree_unlock(upper->eb); 20815d4f98a2SYan Zheng upper->locked = 0; 20825d4f98a2SYan Zheng } 20835d4f98a2SYan Zheng } 20845d4f98a2SYan Zheng path->lowest_level = 0; 20855d4f98a2SYan Zheng return err; 20865d4f98a2SYan Zheng } 20875d4f98a2SYan Zheng 20885d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 20895d4f98a2SYan Zheng struct backref_node *node, 20905d4f98a2SYan Zheng struct btrfs_path *path) 20915d4f98a2SYan Zheng { 20925d4f98a2SYan Zheng struct btrfs_key key; 20935d4f98a2SYan Zheng if (!node->eb || list_empty(&node->upper)) 20945d4f98a2SYan Zheng return 0; 20955d4f98a2SYan Zheng 20965d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 20975d4f98a2SYan Zheng return do_relocation(trans, node, &key, path, 0); 20985d4f98a2SYan Zheng } 20995d4f98a2SYan Zheng 21005d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 21015d4f98a2SYan Zheng struct backref_cache *cache, 21025d4f98a2SYan Zheng struct btrfs_path *path) 21035d4f98a2SYan Zheng { 21045d4f98a2SYan Zheng struct backref_node *node; 21055d4f98a2SYan Zheng int level; 21065d4f98a2SYan Zheng int ret; 21075d4f98a2SYan Zheng int err = 0; 21085d4f98a2SYan Zheng 21095d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 21105d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 21115d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 21125d4f98a2SYan Zheng struct backref_node, lower); 21135d4f98a2SYan Zheng BUG_ON(node->level != level); 21145d4f98a2SYan Zheng 21155d4f98a2SYan Zheng ret = link_to_upper(trans, node, path); 21165d4f98a2SYan Zheng if (ret < 0) 21175d4f98a2SYan Zheng err = ret; 21185d4f98a2SYan Zheng /* 21195d4f98a2SYan Zheng * this remove the node from the pending list and 21205d4f98a2SYan Zheng * may add some other nodes to the level + 1 21215d4f98a2SYan Zheng * pending list 21225d4f98a2SYan Zheng */ 21235d4f98a2SYan Zheng remove_backref_node(cache, node); 21245d4f98a2SYan Zheng } 21255d4f98a2SYan Zheng } 21265d4f98a2SYan Zheng BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root)); 21275d4f98a2SYan Zheng return err; 21285d4f98a2SYan Zheng } 21295d4f98a2SYan Zheng 21305d4f98a2SYan Zheng static void mark_block_processed(struct reloc_control *rc, 21315d4f98a2SYan Zheng struct backref_node *node) 21325d4f98a2SYan Zheng { 21335d4f98a2SYan Zheng u32 blocksize; 21345d4f98a2SYan Zheng if (node->level == 0 || 21355d4f98a2SYan Zheng in_block_group(node->bytenr, rc->block_group)) { 21365d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, node->level); 21375d4f98a2SYan Zheng set_extent_bits(&rc->processed_blocks, node->bytenr, 21385d4f98a2SYan Zheng node->bytenr + blocksize - 1, EXTENT_DIRTY, 21395d4f98a2SYan Zheng GFP_NOFS); 21405d4f98a2SYan Zheng } 21415d4f98a2SYan Zheng node->processed = 1; 21425d4f98a2SYan Zheng } 21435d4f98a2SYan Zheng 21445d4f98a2SYan Zheng /* 21455d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 21465d4f98a2SYan Zheng * as processed. 21475d4f98a2SYan Zheng */ 21485d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 21495d4f98a2SYan Zheng struct backref_node *node) 21505d4f98a2SYan Zheng { 21515d4f98a2SYan Zheng struct backref_node *next = node; 21525d4f98a2SYan Zheng struct backref_edge *edge; 21535d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 21545d4f98a2SYan Zheng int index = 0; 21555d4f98a2SYan Zheng 21565d4f98a2SYan Zheng while (next) { 21575d4f98a2SYan Zheng cond_resched(); 21585d4f98a2SYan Zheng while (1) { 21595d4f98a2SYan Zheng if (next->processed) 21605d4f98a2SYan Zheng break; 21615d4f98a2SYan Zheng 21625d4f98a2SYan Zheng mark_block_processed(rc, next); 21635d4f98a2SYan Zheng 21645d4f98a2SYan Zheng if (list_empty(&next->upper)) 21655d4f98a2SYan Zheng break; 21665d4f98a2SYan Zheng 21675d4f98a2SYan Zheng edge = list_entry(next->upper.next, 21685d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 21695d4f98a2SYan Zheng edges[index++] = edge; 21705d4f98a2SYan Zheng next = edge->node[UPPER]; 21715d4f98a2SYan Zheng } 21725d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 21735d4f98a2SYan Zheng } 21745d4f98a2SYan Zheng } 21755d4f98a2SYan Zheng 21765d4f98a2SYan Zheng static int tree_block_processed(u64 bytenr, u32 blocksize, 21775d4f98a2SYan Zheng struct reloc_control *rc) 21785d4f98a2SYan Zheng { 21795d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 21805d4f98a2SYan Zheng bytenr + blocksize - 1, EXTENT_DIRTY, 1)) 21815d4f98a2SYan Zheng return 1; 21825d4f98a2SYan Zheng return 0; 21835d4f98a2SYan Zheng } 21845d4f98a2SYan Zheng 21855d4f98a2SYan Zheng /* 21865d4f98a2SYan Zheng * check if there are any file extent pointers in the leaf point to 21875d4f98a2SYan Zheng * data require processing 21885d4f98a2SYan Zheng */ 21895d4f98a2SYan Zheng static int check_file_extents(struct reloc_control *rc, 21905d4f98a2SYan Zheng u64 bytenr, u32 blocksize, u64 ptr_gen) 21915d4f98a2SYan Zheng { 21925d4f98a2SYan Zheng struct btrfs_key found_key; 21935d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 21945d4f98a2SYan Zheng struct extent_buffer *leaf; 21955d4f98a2SYan Zheng u32 nritems; 21965d4f98a2SYan Zheng int i; 21975d4f98a2SYan Zheng int ret = 0; 21985d4f98a2SYan Zheng 21995d4f98a2SYan Zheng leaf = read_tree_block(rc->extent_root, bytenr, blocksize, ptr_gen); 22005d4f98a2SYan Zheng 22015d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 22025d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 22035d4f98a2SYan Zheng cond_resched(); 22045d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &found_key, i); 22055d4f98a2SYan Zheng if (found_key.type != BTRFS_EXTENT_DATA_KEY) 22065d4f98a2SYan Zheng continue; 22075d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 22085d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 22095d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 22105d4f98a2SYan Zheng continue; 22115d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 22125d4f98a2SYan Zheng if (bytenr == 0) 22135d4f98a2SYan Zheng continue; 22145d4f98a2SYan Zheng if (in_block_group(bytenr, rc->block_group)) { 22155d4f98a2SYan Zheng ret = 1; 22165d4f98a2SYan Zheng break; 22175d4f98a2SYan Zheng } 22185d4f98a2SYan Zheng } 22195d4f98a2SYan Zheng free_extent_buffer(leaf); 22205d4f98a2SYan Zheng return ret; 22215d4f98a2SYan Zheng } 22225d4f98a2SYan Zheng 22235d4f98a2SYan Zheng /* 22245d4f98a2SYan Zheng * scan child blocks of a given block to find blocks require processing 22255d4f98a2SYan Zheng */ 22265d4f98a2SYan Zheng static int add_child_blocks(struct btrfs_trans_handle *trans, 22275d4f98a2SYan Zheng struct reloc_control *rc, 22285d4f98a2SYan Zheng struct backref_node *node, 22295d4f98a2SYan Zheng struct rb_root *blocks) 22305d4f98a2SYan Zheng { 22315d4f98a2SYan Zheng struct tree_block *block; 22325d4f98a2SYan Zheng struct rb_node *rb_node; 22335d4f98a2SYan Zheng u64 bytenr; 22345d4f98a2SYan Zheng u64 ptr_gen; 22355d4f98a2SYan Zheng u32 blocksize; 22365d4f98a2SYan Zheng u32 nritems; 22375d4f98a2SYan Zheng int i; 22385d4f98a2SYan Zheng int err = 0; 22395d4f98a2SYan Zheng 22405d4f98a2SYan Zheng nritems = btrfs_header_nritems(node->eb); 22415d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, node->level - 1); 22425d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 22435d4f98a2SYan Zheng cond_resched(); 22445d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(node->eb, i); 22455d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(node->eb, i); 22465d4f98a2SYan Zheng if (ptr_gen == trans->transid) 22475d4f98a2SYan Zheng continue; 22485d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group) && 22495d4f98a2SYan Zheng (node->level > 1 || rc->stage == MOVE_DATA_EXTENTS)) 22505d4f98a2SYan Zheng continue; 22515d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 22525d4f98a2SYan Zheng continue; 22535d4f98a2SYan Zheng 22545d4f98a2SYan Zheng readahead_tree_block(rc->extent_root, 22555d4f98a2SYan Zheng bytenr, blocksize, ptr_gen); 22565d4f98a2SYan Zheng } 22575d4f98a2SYan Zheng 22585d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 22595d4f98a2SYan Zheng cond_resched(); 22605d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(node->eb, i); 22615d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(node->eb, i); 22625d4f98a2SYan Zheng if (ptr_gen == trans->transid) 22635d4f98a2SYan Zheng continue; 22645d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group) && 22655d4f98a2SYan Zheng (node->level > 1 || rc->stage == MOVE_DATA_EXTENTS)) 22665d4f98a2SYan Zheng continue; 22675d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 22685d4f98a2SYan Zheng continue; 22695d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group) && 22705d4f98a2SYan Zheng !check_file_extents(rc, bytenr, blocksize, ptr_gen)) 22715d4f98a2SYan Zheng continue; 22725d4f98a2SYan Zheng 22735d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 22745d4f98a2SYan Zheng if (!block) { 22755d4f98a2SYan Zheng err = -ENOMEM; 22765d4f98a2SYan Zheng break; 22775d4f98a2SYan Zheng } 22785d4f98a2SYan Zheng block->bytenr = bytenr; 22795d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &block->key, i); 22805d4f98a2SYan Zheng block->level = node->level - 1; 22815d4f98a2SYan Zheng block->key_ready = 1; 22825d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 22835d4f98a2SYan Zheng BUG_ON(rb_node); 22845d4f98a2SYan Zheng } 22855d4f98a2SYan Zheng if (err) 22865d4f98a2SYan Zheng free_block_list(blocks); 22875d4f98a2SYan Zheng return err; 22885d4f98a2SYan Zheng } 22895d4f98a2SYan Zheng 22905d4f98a2SYan Zheng /* 22915d4f98a2SYan Zheng * find adjacent blocks require processing 22925d4f98a2SYan Zheng */ 22935d4f98a2SYan Zheng static noinline_for_stack 22945d4f98a2SYan Zheng int add_adjacent_blocks(struct btrfs_trans_handle *trans, 22955d4f98a2SYan Zheng struct reloc_control *rc, 22965d4f98a2SYan Zheng struct backref_cache *cache, 22975d4f98a2SYan Zheng struct rb_root *blocks, int level, 22985d4f98a2SYan Zheng struct backref_node **upper) 22995d4f98a2SYan Zheng { 23005d4f98a2SYan Zheng struct backref_node *node; 23015d4f98a2SYan Zheng int ret = 0; 23025d4f98a2SYan Zheng 23035d4f98a2SYan Zheng WARN_ON(!list_empty(&cache->pending[level])); 23045d4f98a2SYan Zheng 23055d4f98a2SYan Zheng if (list_empty(&cache->pending[level + 1])) 23065d4f98a2SYan Zheng return 1; 23075d4f98a2SYan Zheng 23085d4f98a2SYan Zheng node = list_entry(cache->pending[level + 1].next, 23095d4f98a2SYan Zheng struct backref_node, lower); 23105d4f98a2SYan Zheng if (node->eb) 23115d4f98a2SYan Zheng ret = add_child_blocks(trans, rc, node, blocks); 23125d4f98a2SYan Zheng 23135d4f98a2SYan Zheng *upper = node; 23145d4f98a2SYan Zheng return ret; 23155d4f98a2SYan Zheng } 23165d4f98a2SYan Zheng 23175d4f98a2SYan Zheng static int get_tree_block_key(struct reloc_control *rc, 23185d4f98a2SYan Zheng struct tree_block *block) 23195d4f98a2SYan Zheng { 23205d4f98a2SYan Zheng struct extent_buffer *eb; 23215d4f98a2SYan Zheng 23225d4f98a2SYan Zheng BUG_ON(block->key_ready); 23235d4f98a2SYan Zheng eb = read_tree_block(rc->extent_root, block->bytenr, 23245d4f98a2SYan Zheng block->key.objectid, block->key.offset); 23255d4f98a2SYan Zheng WARN_ON(btrfs_header_level(eb) != block->level); 23265d4f98a2SYan Zheng if (block->level == 0) 23275d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 23285d4f98a2SYan Zheng else 23295d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 23305d4f98a2SYan Zheng free_extent_buffer(eb); 23315d4f98a2SYan Zheng block->key_ready = 1; 23325d4f98a2SYan Zheng return 0; 23335d4f98a2SYan Zheng } 23345d4f98a2SYan Zheng 23355d4f98a2SYan Zheng static int reada_tree_block(struct reloc_control *rc, 23365d4f98a2SYan Zheng struct tree_block *block) 23375d4f98a2SYan Zheng { 23385d4f98a2SYan Zheng BUG_ON(block->key_ready); 23395d4f98a2SYan Zheng readahead_tree_block(rc->extent_root, block->bytenr, 23405d4f98a2SYan Zheng block->key.objectid, block->key.offset); 23415d4f98a2SYan Zheng return 0; 23425d4f98a2SYan Zheng } 23435d4f98a2SYan Zheng 23445d4f98a2SYan Zheng /* 23455d4f98a2SYan Zheng * helper function to relocate a tree block 23465d4f98a2SYan Zheng */ 23475d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 23485d4f98a2SYan Zheng struct reloc_control *rc, 23495d4f98a2SYan Zheng struct backref_node *node, 23505d4f98a2SYan Zheng struct btrfs_key *key, 23515d4f98a2SYan Zheng struct btrfs_path *path) 23525d4f98a2SYan Zheng { 23535d4f98a2SYan Zheng struct btrfs_root *root; 23545d4f98a2SYan Zheng int ret; 23555d4f98a2SYan Zheng 23565d4f98a2SYan Zheng root = select_one_root(trans, node); 23575d4f98a2SYan Zheng if (unlikely(!root)) { 23585d4f98a2SYan Zheng rc->found_old_snapshot = 1; 23595d4f98a2SYan Zheng update_processed_blocks(rc, node); 23605d4f98a2SYan Zheng return 0; 23615d4f98a2SYan Zheng } 23625d4f98a2SYan Zheng 23635d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 23645d4f98a2SYan Zheng ret = do_relocation(trans, node, key, path, 1); 23655d4f98a2SYan Zheng if (ret < 0) 23665d4f98a2SYan Zheng goto out; 23675d4f98a2SYan Zheng if (node->level == 0 && rc->stage == UPDATE_DATA_PTRS) { 23685d4f98a2SYan Zheng ret = replace_file_extents(trans, rc, root, 23695d4f98a2SYan Zheng node->eb, NULL); 23705d4f98a2SYan Zheng if (ret < 0) 23715d4f98a2SYan Zheng goto out; 23725d4f98a2SYan Zheng } 23735d4f98a2SYan Zheng drop_node_buffer(node); 23745d4f98a2SYan Zheng } else if (!root->ref_cows) { 23755d4f98a2SYan Zheng path->lowest_level = node->level; 23765d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 23775d4f98a2SYan Zheng btrfs_release_path(root, path); 23785d4f98a2SYan Zheng if (ret < 0) 23795d4f98a2SYan Zheng goto out; 23805d4f98a2SYan Zheng } else if (root != node->root) { 23815d4f98a2SYan Zheng WARN_ON(node->level > 0 || rc->stage != UPDATE_DATA_PTRS); 23825d4f98a2SYan Zheng } 23835d4f98a2SYan Zheng 23845d4f98a2SYan Zheng update_processed_blocks(rc, node); 23855d4f98a2SYan Zheng ret = 0; 23865d4f98a2SYan Zheng out: 23875d4f98a2SYan Zheng drop_node_buffer(node); 23885d4f98a2SYan Zheng return ret; 23895d4f98a2SYan Zheng } 23905d4f98a2SYan Zheng 23915d4f98a2SYan Zheng /* 23925d4f98a2SYan Zheng * relocate a list of blocks 23935d4f98a2SYan Zheng */ 23945d4f98a2SYan Zheng static noinline_for_stack 23955d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 23965d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 23975d4f98a2SYan Zheng { 23985d4f98a2SYan Zheng struct backref_cache *cache; 23995d4f98a2SYan Zheng struct backref_node *node; 24005d4f98a2SYan Zheng struct btrfs_path *path; 24015d4f98a2SYan Zheng struct tree_block *block; 24025d4f98a2SYan Zheng struct rb_node *rb_node; 24035d4f98a2SYan Zheng int level = -1; 24045d4f98a2SYan Zheng int ret; 24055d4f98a2SYan Zheng int err = 0; 24065d4f98a2SYan Zheng 24075d4f98a2SYan Zheng path = btrfs_alloc_path(); 24085d4f98a2SYan Zheng if (!path) 24095d4f98a2SYan Zheng return -ENOMEM; 24105d4f98a2SYan Zheng 24115d4f98a2SYan Zheng cache = kmalloc(sizeof(*cache), GFP_NOFS); 24125d4f98a2SYan Zheng if (!cache) { 24135d4f98a2SYan Zheng btrfs_free_path(path); 24145d4f98a2SYan Zheng return -ENOMEM; 24155d4f98a2SYan Zheng } 24165d4f98a2SYan Zheng 24175d4f98a2SYan Zheng backref_cache_init(cache); 24185d4f98a2SYan Zheng 24195d4f98a2SYan Zheng rb_node = rb_first(blocks); 24205d4f98a2SYan Zheng while (rb_node) { 24215d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 24225d4f98a2SYan Zheng if (level == -1) 24235d4f98a2SYan Zheng level = block->level; 24245d4f98a2SYan Zheng else 24255d4f98a2SYan Zheng BUG_ON(level != block->level); 24265d4f98a2SYan Zheng if (!block->key_ready) 24275d4f98a2SYan Zheng reada_tree_block(rc, block); 24285d4f98a2SYan Zheng rb_node = rb_next(rb_node); 24295d4f98a2SYan Zheng } 24305d4f98a2SYan Zheng 24315d4f98a2SYan Zheng rb_node = rb_first(blocks); 24325d4f98a2SYan Zheng while (rb_node) { 24335d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 24345d4f98a2SYan Zheng if (!block->key_ready) 24355d4f98a2SYan Zheng get_tree_block_key(rc, block); 24365d4f98a2SYan Zheng rb_node = rb_next(rb_node); 24375d4f98a2SYan Zheng } 24385d4f98a2SYan Zheng 24395d4f98a2SYan Zheng rb_node = rb_first(blocks); 24405d4f98a2SYan Zheng while (rb_node) { 24415d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 24425d4f98a2SYan Zheng 24435d4f98a2SYan Zheng node = build_backref_tree(rc, cache, &block->key, 24445d4f98a2SYan Zheng block->level, block->bytenr); 24455d4f98a2SYan Zheng if (IS_ERR(node)) { 24465d4f98a2SYan Zheng err = PTR_ERR(node); 24475d4f98a2SYan Zheng goto out; 24485d4f98a2SYan Zheng } 24495d4f98a2SYan Zheng 24505d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 24515d4f98a2SYan Zheng path); 24525d4f98a2SYan Zheng if (ret < 0) { 24535d4f98a2SYan Zheng err = ret; 24545d4f98a2SYan Zheng goto out; 24555d4f98a2SYan Zheng } 24565d4f98a2SYan Zheng remove_backref_node(cache, node); 24575d4f98a2SYan Zheng rb_node = rb_next(rb_node); 24585d4f98a2SYan Zheng } 24595d4f98a2SYan Zheng 24605d4f98a2SYan Zheng if (level > 0) 24615d4f98a2SYan Zheng goto out; 24625d4f98a2SYan Zheng 24635d4f98a2SYan Zheng free_block_list(blocks); 24645d4f98a2SYan Zheng 24655d4f98a2SYan Zheng /* 24665d4f98a2SYan Zheng * now backrefs of some upper level tree blocks have been cached, 24675d4f98a2SYan Zheng * try relocating blocks referenced by these upper level blocks. 24685d4f98a2SYan Zheng */ 24695d4f98a2SYan Zheng while (1) { 24705d4f98a2SYan Zheng struct backref_node *upper = NULL; 24715d4f98a2SYan Zheng if (trans->transaction->in_commit || 24725d4f98a2SYan Zheng trans->transaction->delayed_refs.flushing) 24735d4f98a2SYan Zheng break; 24745d4f98a2SYan Zheng 24755d4f98a2SYan Zheng ret = add_adjacent_blocks(trans, rc, cache, blocks, level, 24765d4f98a2SYan Zheng &upper); 24775d4f98a2SYan Zheng if (ret < 0) 24785d4f98a2SYan Zheng err = ret; 24795d4f98a2SYan Zheng if (ret != 0) 24805d4f98a2SYan Zheng break; 24815d4f98a2SYan Zheng 24825d4f98a2SYan Zheng rb_node = rb_first(blocks); 24835d4f98a2SYan Zheng while (rb_node) { 24845d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 24855d4f98a2SYan Zheng if (trans->transaction->in_commit || 24865d4f98a2SYan Zheng trans->transaction->delayed_refs.flushing) 24875d4f98a2SYan Zheng goto out; 24885d4f98a2SYan Zheng BUG_ON(!block->key_ready); 24895d4f98a2SYan Zheng node = build_backref_tree(rc, cache, &block->key, 24905d4f98a2SYan Zheng level, block->bytenr); 24915d4f98a2SYan Zheng if (IS_ERR(node)) { 24925d4f98a2SYan Zheng err = PTR_ERR(node); 24935d4f98a2SYan Zheng goto out; 24945d4f98a2SYan Zheng } 24955d4f98a2SYan Zheng 24965d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, 24975d4f98a2SYan Zheng &block->key, path); 24985d4f98a2SYan Zheng if (ret < 0) { 24995d4f98a2SYan Zheng err = ret; 25005d4f98a2SYan Zheng goto out; 25015d4f98a2SYan Zheng } 25025d4f98a2SYan Zheng remove_backref_node(cache, node); 25035d4f98a2SYan Zheng rb_node = rb_next(rb_node); 25045d4f98a2SYan Zheng } 25055d4f98a2SYan Zheng free_block_list(blocks); 25065d4f98a2SYan Zheng 25075d4f98a2SYan Zheng if (upper) { 25085d4f98a2SYan Zheng ret = link_to_upper(trans, upper, path); 25095d4f98a2SYan Zheng if (ret < 0) { 25105d4f98a2SYan Zheng err = ret; 25115d4f98a2SYan Zheng break; 25125d4f98a2SYan Zheng } 25135d4f98a2SYan Zheng remove_backref_node(cache, upper); 25145d4f98a2SYan Zheng } 25155d4f98a2SYan Zheng } 25165d4f98a2SYan Zheng out: 25175d4f98a2SYan Zheng free_block_list(blocks); 25185d4f98a2SYan Zheng 25195d4f98a2SYan Zheng ret = finish_pending_nodes(trans, cache, path); 25205d4f98a2SYan Zheng if (ret < 0) 25215d4f98a2SYan Zheng err = ret; 25225d4f98a2SYan Zheng 25235d4f98a2SYan Zheng kfree(cache); 25245d4f98a2SYan Zheng btrfs_free_path(path); 25255d4f98a2SYan Zheng return err; 25265d4f98a2SYan Zheng } 25275d4f98a2SYan Zheng 25285d4f98a2SYan Zheng static noinline_for_stack 25295d4f98a2SYan Zheng int relocate_inode_pages(struct inode *inode, u64 start, u64 len) 25305d4f98a2SYan Zheng { 25315d4f98a2SYan Zheng u64 page_start; 25325d4f98a2SYan Zheng u64 page_end; 25335d4f98a2SYan Zheng unsigned long i; 25345d4f98a2SYan Zheng unsigned long first_index; 25355d4f98a2SYan Zheng unsigned long last_index; 25365d4f98a2SYan Zheng unsigned int total_read = 0; 25375d4f98a2SYan Zheng unsigned int total_dirty = 0; 25385d4f98a2SYan Zheng struct page *page; 25395d4f98a2SYan Zheng struct file_ra_state *ra; 25405d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 25415d4f98a2SYan Zheng struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree; 25425d4f98a2SYan Zheng int ret = 0; 25435d4f98a2SYan Zheng 25445d4f98a2SYan Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 25455d4f98a2SYan Zheng if (!ra) 25465d4f98a2SYan Zheng return -ENOMEM; 25475d4f98a2SYan Zheng 25485d4f98a2SYan Zheng mutex_lock(&inode->i_mutex); 25495d4f98a2SYan Zheng first_index = start >> PAGE_CACHE_SHIFT; 25505d4f98a2SYan Zheng last_index = (start + len - 1) >> PAGE_CACHE_SHIFT; 25515d4f98a2SYan Zheng 25525d4f98a2SYan Zheng /* make sure the dirty trick played by the caller work */ 25535d4f98a2SYan Zheng ret = invalidate_inode_pages2_range(inode->i_mapping, 25545d4f98a2SYan Zheng first_index, last_index); 25555d4f98a2SYan Zheng if (ret) 25565d4f98a2SYan Zheng goto out_unlock; 25575d4f98a2SYan Zheng 25585d4f98a2SYan Zheng file_ra_state_init(ra, inode->i_mapping); 25595d4f98a2SYan Zheng 25605d4f98a2SYan Zheng for (i = first_index ; i <= last_index; i++) { 25615d4f98a2SYan Zheng if (total_read % ra->ra_pages == 0) { 25625d4f98a2SYan Zheng btrfs_force_ra(inode->i_mapping, ra, NULL, i, 25635d4f98a2SYan Zheng min(last_index, ra->ra_pages + i - 1)); 25645d4f98a2SYan Zheng } 25655d4f98a2SYan Zheng total_read++; 25665d4f98a2SYan Zheng again: 25675d4f98a2SYan Zheng if (((u64)i << PAGE_CACHE_SHIFT) > i_size_read(inode)) 25685d4f98a2SYan Zheng BUG_ON(1); 25695d4f98a2SYan Zheng page = grab_cache_page(inode->i_mapping, i); 25705d4f98a2SYan Zheng if (!page) { 25715d4f98a2SYan Zheng ret = -ENOMEM; 25725d4f98a2SYan Zheng goto out_unlock; 25735d4f98a2SYan Zheng } 25745d4f98a2SYan Zheng if (!PageUptodate(page)) { 25755d4f98a2SYan Zheng btrfs_readpage(NULL, page); 25765d4f98a2SYan Zheng lock_page(page); 25775d4f98a2SYan Zheng if (!PageUptodate(page)) { 25785d4f98a2SYan Zheng unlock_page(page); 25795d4f98a2SYan Zheng page_cache_release(page); 25805d4f98a2SYan Zheng ret = -EIO; 25815d4f98a2SYan Zheng goto out_unlock; 25825d4f98a2SYan Zheng } 25835d4f98a2SYan Zheng } 25845d4f98a2SYan Zheng wait_on_page_writeback(page); 25855d4f98a2SYan Zheng 25865d4f98a2SYan Zheng page_start = (u64)page->index << PAGE_CACHE_SHIFT; 25875d4f98a2SYan Zheng page_end = page_start + PAGE_CACHE_SIZE - 1; 25885d4f98a2SYan Zheng lock_extent(io_tree, page_start, page_end, GFP_NOFS); 25895d4f98a2SYan Zheng 25905d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, page_start); 25915d4f98a2SYan Zheng if (ordered) { 25925d4f98a2SYan Zheng unlock_extent(io_tree, page_start, page_end, GFP_NOFS); 25935d4f98a2SYan Zheng unlock_page(page); 25945d4f98a2SYan Zheng page_cache_release(page); 25955d4f98a2SYan Zheng btrfs_start_ordered_extent(inode, ordered, 1); 25965d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 25975d4f98a2SYan Zheng goto again; 25985d4f98a2SYan Zheng } 25995d4f98a2SYan Zheng set_page_extent_mapped(page); 26005d4f98a2SYan Zheng 26015d4f98a2SYan Zheng if (i == first_index) 26025d4f98a2SYan Zheng set_extent_bits(io_tree, page_start, page_end, 26035d4f98a2SYan Zheng EXTENT_BOUNDARY, GFP_NOFS); 26045d4f98a2SYan Zheng btrfs_set_extent_delalloc(inode, page_start, page_end); 26055d4f98a2SYan Zheng 26065d4f98a2SYan Zheng set_page_dirty(page); 26075d4f98a2SYan Zheng total_dirty++; 26085d4f98a2SYan Zheng 26095d4f98a2SYan Zheng unlock_extent(io_tree, page_start, page_end, GFP_NOFS); 26105d4f98a2SYan Zheng unlock_page(page); 26115d4f98a2SYan Zheng page_cache_release(page); 26125d4f98a2SYan Zheng } 26135d4f98a2SYan Zheng out_unlock: 26145d4f98a2SYan Zheng mutex_unlock(&inode->i_mutex); 26155d4f98a2SYan Zheng kfree(ra); 26165d4f98a2SYan Zheng balance_dirty_pages_ratelimited_nr(inode->i_mapping, total_dirty); 26175d4f98a2SYan Zheng return ret; 26185d4f98a2SYan Zheng } 26195d4f98a2SYan Zheng 26205d4f98a2SYan Zheng static noinline_for_stack 26215d4f98a2SYan Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key) 26225d4f98a2SYan Zheng { 26235d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 26245d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 26255d4f98a2SYan Zheng struct extent_map *em; 26265d4f98a2SYan Zheng u64 start = extent_key->objectid - BTRFS_I(inode)->index_cnt; 26275d4f98a2SYan Zheng u64 end = start + extent_key->offset - 1; 26285d4f98a2SYan Zheng 26295d4f98a2SYan Zheng em = alloc_extent_map(GFP_NOFS); 26305d4f98a2SYan Zheng em->start = start; 26315d4f98a2SYan Zheng em->len = extent_key->offset; 26325d4f98a2SYan Zheng em->block_len = extent_key->offset; 26335d4f98a2SYan Zheng em->block_start = extent_key->objectid; 26345d4f98a2SYan Zheng em->bdev = root->fs_info->fs_devices->latest_bdev; 26355d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 26365d4f98a2SYan Zheng 26375d4f98a2SYan Zheng /* setup extent map to cheat btrfs_readpage */ 26385d4f98a2SYan Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 26395d4f98a2SYan Zheng while (1) { 26405d4f98a2SYan Zheng int ret; 26415d4f98a2SYan Zheng spin_lock(&em_tree->lock); 26425d4f98a2SYan Zheng ret = add_extent_mapping(em_tree, em); 26435d4f98a2SYan Zheng spin_unlock(&em_tree->lock); 26445d4f98a2SYan Zheng if (ret != -EEXIST) { 26455d4f98a2SYan Zheng free_extent_map(em); 26465d4f98a2SYan Zheng break; 26475d4f98a2SYan Zheng } 26485d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 0); 26495d4f98a2SYan Zheng } 26505d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 26515d4f98a2SYan Zheng 26525d4f98a2SYan Zheng return relocate_inode_pages(inode, start, extent_key->offset); 26535d4f98a2SYan Zheng } 26545d4f98a2SYan Zheng 26555d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 26565d4f98a2SYan Zheng static int get_ref_objectid_v0(struct reloc_control *rc, 26575d4f98a2SYan Zheng struct btrfs_path *path, 26585d4f98a2SYan Zheng struct btrfs_key *extent_key, 26595d4f98a2SYan Zheng u64 *ref_objectid, int *path_change) 26605d4f98a2SYan Zheng { 26615d4f98a2SYan Zheng struct btrfs_key key; 26625d4f98a2SYan Zheng struct extent_buffer *leaf; 26635d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 26645d4f98a2SYan Zheng int ret; 26655d4f98a2SYan Zheng int slot; 26665d4f98a2SYan Zheng 26675d4f98a2SYan Zheng leaf = path->nodes[0]; 26685d4f98a2SYan Zheng slot = path->slots[0]; 26695d4f98a2SYan Zheng while (1) { 26705d4f98a2SYan Zheng if (slot >= btrfs_header_nritems(leaf)) { 26715d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 26725d4f98a2SYan Zheng if (ret < 0) 26735d4f98a2SYan Zheng return ret; 26745d4f98a2SYan Zheng BUG_ON(ret > 0); 26755d4f98a2SYan Zheng leaf = path->nodes[0]; 26765d4f98a2SYan Zheng slot = path->slots[0]; 26775d4f98a2SYan Zheng if (path_change) 26785d4f98a2SYan Zheng *path_change = 1; 26795d4f98a2SYan Zheng } 26805d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, slot); 26815d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 26825d4f98a2SYan Zheng return -ENOENT; 26835d4f98a2SYan Zheng 26845d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_REF_V0_KEY) { 26855d4f98a2SYan Zheng slot++; 26865d4f98a2SYan Zheng continue; 26875d4f98a2SYan Zheng } 26885d4f98a2SYan Zheng ref0 = btrfs_item_ptr(leaf, slot, 26895d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 26905d4f98a2SYan Zheng *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0); 26915d4f98a2SYan Zheng break; 26925d4f98a2SYan Zheng } 26935d4f98a2SYan Zheng return 0; 26945d4f98a2SYan Zheng } 26955d4f98a2SYan Zheng #endif 26965d4f98a2SYan Zheng 26975d4f98a2SYan Zheng /* 26985d4f98a2SYan Zheng * helper to add a tree block to the list. 26995d4f98a2SYan Zheng * the major work is getting the generation and level of the block 27005d4f98a2SYan Zheng */ 27015d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 27025d4f98a2SYan Zheng struct btrfs_key *extent_key, 27035d4f98a2SYan Zheng struct btrfs_path *path, 27045d4f98a2SYan Zheng struct rb_root *blocks) 27055d4f98a2SYan Zheng { 27065d4f98a2SYan Zheng struct extent_buffer *eb; 27075d4f98a2SYan Zheng struct btrfs_extent_item *ei; 27085d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 27095d4f98a2SYan Zheng struct tree_block *block; 27105d4f98a2SYan Zheng struct rb_node *rb_node; 27115d4f98a2SYan Zheng u32 item_size; 27125d4f98a2SYan Zheng int level = -1; 27135d4f98a2SYan Zheng int generation; 27145d4f98a2SYan Zheng 27155d4f98a2SYan Zheng eb = path->nodes[0]; 27165d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 27175d4f98a2SYan Zheng 27185d4f98a2SYan Zheng if (item_size >= sizeof(*ei) + sizeof(*bi)) { 27195d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 27205d4f98a2SYan Zheng struct btrfs_extent_item); 27215d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 27225d4f98a2SYan Zheng generation = btrfs_extent_generation(eb, ei); 27235d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 27245d4f98a2SYan Zheng } else { 27255d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 27265d4f98a2SYan Zheng u64 ref_owner; 27275d4f98a2SYan Zheng int ret; 27285d4f98a2SYan Zheng 27295d4f98a2SYan Zheng BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 27305d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, extent_key, 27315d4f98a2SYan Zheng &ref_owner, NULL); 27325d4f98a2SYan Zheng BUG_ON(ref_owner >= BTRFS_MAX_LEVEL); 27335d4f98a2SYan Zheng level = (int)ref_owner; 27345d4f98a2SYan Zheng /* FIXME: get real generation */ 27355d4f98a2SYan Zheng generation = 0; 27365d4f98a2SYan Zheng #else 27375d4f98a2SYan Zheng BUG(); 27385d4f98a2SYan Zheng #endif 27395d4f98a2SYan Zheng } 27405d4f98a2SYan Zheng 27415d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 27425d4f98a2SYan Zheng 27435d4f98a2SYan Zheng BUG_ON(level == -1); 27445d4f98a2SYan Zheng 27455d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 27465d4f98a2SYan Zheng if (!block) 27475d4f98a2SYan Zheng return -ENOMEM; 27485d4f98a2SYan Zheng 27495d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 27505d4f98a2SYan Zheng block->key.objectid = extent_key->offset; 27515d4f98a2SYan Zheng block->key.offset = generation; 27525d4f98a2SYan Zheng block->level = level; 27535d4f98a2SYan Zheng block->key_ready = 0; 27545d4f98a2SYan Zheng 27555d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 27565d4f98a2SYan Zheng BUG_ON(rb_node); 27575d4f98a2SYan Zheng 27585d4f98a2SYan Zheng return 0; 27595d4f98a2SYan Zheng } 27605d4f98a2SYan Zheng 27615d4f98a2SYan Zheng /* 27625d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 27635d4f98a2SYan Zheng */ 27645d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 27655d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 27665d4f98a2SYan Zheng struct rb_root *blocks) 27675d4f98a2SYan Zheng { 27685d4f98a2SYan Zheng struct btrfs_path *path; 27695d4f98a2SYan Zheng struct btrfs_key key; 27705d4f98a2SYan Zheng int ret; 27715d4f98a2SYan Zheng 27725d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 27735d4f98a2SYan Zheng return 0; 27745d4f98a2SYan Zheng 27755d4f98a2SYan Zheng if (tree_search(blocks, bytenr)) 27765d4f98a2SYan Zheng return 0; 27775d4f98a2SYan Zheng 27785d4f98a2SYan Zheng path = btrfs_alloc_path(); 27795d4f98a2SYan Zheng if (!path) 27805d4f98a2SYan Zheng return -ENOMEM; 27815d4f98a2SYan Zheng 27825d4f98a2SYan Zheng key.objectid = bytenr; 27835d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 27845d4f98a2SYan Zheng key.offset = blocksize; 27855d4f98a2SYan Zheng 27865d4f98a2SYan Zheng path->search_commit_root = 1; 27875d4f98a2SYan Zheng path->skip_locking = 1; 27885d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 27895d4f98a2SYan Zheng if (ret < 0) 27905d4f98a2SYan Zheng goto out; 27915d4f98a2SYan Zheng BUG_ON(ret); 27925d4f98a2SYan Zheng 27935d4f98a2SYan Zheng btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 27945d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 27955d4f98a2SYan Zheng out: 27965d4f98a2SYan Zheng btrfs_free_path(path); 27975d4f98a2SYan Zheng return ret; 27985d4f98a2SYan Zheng } 27995d4f98a2SYan Zheng 28005d4f98a2SYan Zheng /* 28015d4f98a2SYan Zheng * helper to check if the block use full backrefs for pointers in it 28025d4f98a2SYan Zheng */ 28035d4f98a2SYan Zheng static int block_use_full_backref(struct reloc_control *rc, 28045d4f98a2SYan Zheng struct extent_buffer *eb) 28055d4f98a2SYan Zheng { 28065d4f98a2SYan Zheng struct btrfs_path *path; 28075d4f98a2SYan Zheng struct btrfs_extent_item *ei; 28085d4f98a2SYan Zheng struct btrfs_key key; 28095d4f98a2SYan Zheng u64 flags; 28105d4f98a2SYan Zheng int ret; 28115d4f98a2SYan Zheng 28125d4f98a2SYan Zheng if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) || 28135d4f98a2SYan Zheng btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV) 28145d4f98a2SYan Zheng return 1; 28155d4f98a2SYan Zheng 28165d4f98a2SYan Zheng path = btrfs_alloc_path(); 28175d4f98a2SYan Zheng BUG_ON(!path); 28185d4f98a2SYan Zheng 28195d4f98a2SYan Zheng key.objectid = eb->start; 28205d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 28215d4f98a2SYan Zheng key.offset = eb->len; 28225d4f98a2SYan Zheng 28235d4f98a2SYan Zheng path->search_commit_root = 1; 28245d4f98a2SYan Zheng path->skip_locking = 1; 28255d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, 28265d4f98a2SYan Zheng &key, path, 0, 0); 28275d4f98a2SYan Zheng BUG_ON(ret); 28285d4f98a2SYan Zheng 28295d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 28305d4f98a2SYan Zheng struct btrfs_extent_item); 28315d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 28325d4f98a2SYan Zheng BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)); 28335d4f98a2SYan Zheng if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) 28345d4f98a2SYan Zheng ret = 1; 28355d4f98a2SYan Zheng else 28365d4f98a2SYan Zheng ret = 0; 28375d4f98a2SYan Zheng btrfs_free_path(path); 28385d4f98a2SYan Zheng return ret; 28395d4f98a2SYan Zheng } 28405d4f98a2SYan Zheng 28415d4f98a2SYan Zheng /* 28425d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY 28435d4f98a2SYan Zheng * this function scans fs tree to find blocks reference the data extent 28445d4f98a2SYan Zheng */ 28455d4f98a2SYan Zheng static int find_data_references(struct reloc_control *rc, 28465d4f98a2SYan Zheng struct btrfs_key *extent_key, 28475d4f98a2SYan Zheng struct extent_buffer *leaf, 28485d4f98a2SYan Zheng struct btrfs_extent_data_ref *ref, 28495d4f98a2SYan Zheng struct rb_root *blocks) 28505d4f98a2SYan Zheng { 28515d4f98a2SYan Zheng struct btrfs_path *path; 28525d4f98a2SYan Zheng struct tree_block *block; 28535d4f98a2SYan Zheng struct btrfs_root *root; 28545d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 28555d4f98a2SYan Zheng struct rb_node *rb_node; 28565d4f98a2SYan Zheng struct btrfs_key key; 28575d4f98a2SYan Zheng u64 ref_root; 28585d4f98a2SYan Zheng u64 ref_objectid; 28595d4f98a2SYan Zheng u64 ref_offset; 28605d4f98a2SYan Zheng u32 ref_count; 28615d4f98a2SYan Zheng u32 nritems; 28625d4f98a2SYan Zheng int err = 0; 28635d4f98a2SYan Zheng int added = 0; 28645d4f98a2SYan Zheng int counted; 28655d4f98a2SYan Zheng int ret; 28665d4f98a2SYan Zheng 28675d4f98a2SYan Zheng path = btrfs_alloc_path(); 28685d4f98a2SYan Zheng if (!path) 28695d4f98a2SYan Zheng return -ENOMEM; 28705d4f98a2SYan Zheng 28715d4f98a2SYan Zheng ref_root = btrfs_extent_data_ref_root(leaf, ref); 28725d4f98a2SYan Zheng ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref); 28735d4f98a2SYan Zheng ref_offset = btrfs_extent_data_ref_offset(leaf, ref); 28745d4f98a2SYan Zheng ref_count = btrfs_extent_data_ref_count(leaf, ref); 28755d4f98a2SYan Zheng 28765d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, ref_root); 28775d4f98a2SYan Zheng if (IS_ERR(root)) { 28785d4f98a2SYan Zheng err = PTR_ERR(root); 28795d4f98a2SYan Zheng goto out; 28805d4f98a2SYan Zheng } 28815d4f98a2SYan Zheng 28825d4f98a2SYan Zheng key.objectid = ref_objectid; 28835d4f98a2SYan Zheng key.offset = ref_offset; 28845d4f98a2SYan Zheng key.type = BTRFS_EXTENT_DATA_KEY; 28855d4f98a2SYan Zheng 28865d4f98a2SYan Zheng path->search_commit_root = 1; 28875d4f98a2SYan Zheng path->skip_locking = 1; 28885d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 28895d4f98a2SYan Zheng if (ret < 0) { 28905d4f98a2SYan Zheng err = ret; 28915d4f98a2SYan Zheng goto out; 28925d4f98a2SYan Zheng } 28935d4f98a2SYan Zheng 28945d4f98a2SYan Zheng leaf = path->nodes[0]; 28955d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 28965d4f98a2SYan Zheng /* 28975d4f98a2SYan Zheng * the references in tree blocks that use full backrefs 28985d4f98a2SYan Zheng * are not counted in 28995d4f98a2SYan Zheng */ 29005d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 29015d4f98a2SYan Zheng counted = 0; 29025d4f98a2SYan Zheng else 29035d4f98a2SYan Zheng counted = 1; 29045d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 29055d4f98a2SYan Zheng if (rb_node) { 29065d4f98a2SYan Zheng if (counted) 29075d4f98a2SYan Zheng added = 1; 29085d4f98a2SYan Zheng else 29095d4f98a2SYan Zheng path->slots[0] = nritems; 29105d4f98a2SYan Zheng } 29115d4f98a2SYan Zheng 29125d4f98a2SYan Zheng while (ref_count > 0) { 29135d4f98a2SYan Zheng while (path->slots[0] >= nritems) { 29145d4f98a2SYan Zheng ret = btrfs_next_leaf(root, path); 29155d4f98a2SYan Zheng if (ret < 0) { 29165d4f98a2SYan Zheng err = ret; 29175d4f98a2SYan Zheng goto out; 29185d4f98a2SYan Zheng } 29195d4f98a2SYan Zheng if (ret > 0) { 29205d4f98a2SYan Zheng WARN_ON(1); 29215d4f98a2SYan Zheng goto out; 29225d4f98a2SYan Zheng } 29235d4f98a2SYan Zheng 29245d4f98a2SYan Zheng leaf = path->nodes[0]; 29255d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 29265d4f98a2SYan Zheng added = 0; 29275d4f98a2SYan Zheng 29285d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 29295d4f98a2SYan Zheng counted = 0; 29305d4f98a2SYan Zheng else 29315d4f98a2SYan Zheng counted = 1; 29325d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 29335d4f98a2SYan Zheng if (rb_node) { 29345d4f98a2SYan Zheng if (counted) 29355d4f98a2SYan Zheng added = 1; 29365d4f98a2SYan Zheng else 29375d4f98a2SYan Zheng path->slots[0] = nritems; 29385d4f98a2SYan Zheng } 29395d4f98a2SYan Zheng } 29405d4f98a2SYan Zheng 29415d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 29425d4f98a2SYan Zheng if (key.objectid != ref_objectid || 29435d4f98a2SYan Zheng key.type != BTRFS_EXTENT_DATA_KEY) { 29445d4f98a2SYan Zheng WARN_ON(1); 29455d4f98a2SYan Zheng break; 29465d4f98a2SYan Zheng } 29475d4f98a2SYan Zheng 29485d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 29495d4f98a2SYan Zheng struct btrfs_file_extent_item); 29505d4f98a2SYan Zheng 29515d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 29525d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 29535d4f98a2SYan Zheng goto next; 29545d4f98a2SYan Zheng 29555d4f98a2SYan Zheng if (btrfs_file_extent_disk_bytenr(leaf, fi) != 29565d4f98a2SYan Zheng extent_key->objectid) 29575d4f98a2SYan Zheng goto next; 29585d4f98a2SYan Zheng 29595d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 29605d4f98a2SYan Zheng if (key.offset != ref_offset) 29615d4f98a2SYan Zheng goto next; 29625d4f98a2SYan Zheng 29635d4f98a2SYan Zheng if (counted) 29645d4f98a2SYan Zheng ref_count--; 29655d4f98a2SYan Zheng if (added) 29665d4f98a2SYan Zheng goto next; 29675d4f98a2SYan Zheng 29685d4f98a2SYan Zheng if (!tree_block_processed(leaf->start, leaf->len, rc)) { 29695d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 29705d4f98a2SYan Zheng if (!block) { 29715d4f98a2SYan Zheng err = -ENOMEM; 29725d4f98a2SYan Zheng break; 29735d4f98a2SYan Zheng } 29745d4f98a2SYan Zheng block->bytenr = leaf->start; 29755d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &block->key, 0); 29765d4f98a2SYan Zheng block->level = 0; 29775d4f98a2SYan Zheng block->key_ready = 1; 29785d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, 29795d4f98a2SYan Zheng &block->rb_node); 29805d4f98a2SYan Zheng BUG_ON(rb_node); 29815d4f98a2SYan Zheng } 29825d4f98a2SYan Zheng if (counted) 29835d4f98a2SYan Zheng added = 1; 29845d4f98a2SYan Zheng else 29855d4f98a2SYan Zheng path->slots[0] = nritems; 29865d4f98a2SYan Zheng next: 29875d4f98a2SYan Zheng path->slots[0]++; 29885d4f98a2SYan Zheng 29895d4f98a2SYan Zheng } 29905d4f98a2SYan Zheng out: 29915d4f98a2SYan Zheng btrfs_free_path(path); 29925d4f98a2SYan Zheng return err; 29935d4f98a2SYan Zheng } 29945d4f98a2SYan Zheng 29955d4f98a2SYan Zheng /* 29965d4f98a2SYan Zheng * hepler to find all tree blocks that reference a given data extent 29975d4f98a2SYan Zheng */ 29985d4f98a2SYan Zheng static noinline_for_stack 29995d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 30005d4f98a2SYan Zheng struct btrfs_key *extent_key, 30015d4f98a2SYan Zheng struct btrfs_path *path, 30025d4f98a2SYan Zheng struct rb_root *blocks) 30035d4f98a2SYan Zheng { 30045d4f98a2SYan Zheng struct btrfs_key key; 30055d4f98a2SYan Zheng struct extent_buffer *eb; 30065d4f98a2SYan Zheng struct btrfs_extent_data_ref *dref; 30075d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 30085d4f98a2SYan Zheng unsigned long ptr; 30095d4f98a2SYan Zheng unsigned long end; 30105d4f98a2SYan Zheng u32 blocksize; 30115d4f98a2SYan Zheng int ret; 30125d4f98a2SYan Zheng int err = 0; 30135d4f98a2SYan Zheng 30145d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, NULL, extent_key->objectid, 30155d4f98a2SYan Zheng extent_key->offset); 30165d4f98a2SYan Zheng BUG_ON(ret < 0); 30175d4f98a2SYan Zheng if (ret > 0) { 30185d4f98a2SYan Zheng /* the relocated data is fragmented */ 30195d4f98a2SYan Zheng rc->extents_skipped++; 30205d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 30215d4f98a2SYan Zheng return 0; 30225d4f98a2SYan Zheng } 30235d4f98a2SYan Zheng 30245d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, 0); 30255d4f98a2SYan Zheng 30265d4f98a2SYan Zheng eb = path->nodes[0]; 30275d4f98a2SYan Zheng ptr = btrfs_item_ptr_offset(eb, path->slots[0]); 30285d4f98a2SYan Zheng end = ptr + btrfs_item_size_nr(eb, path->slots[0]); 30295d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 30305d4f98a2SYan Zheng if (ptr + sizeof(struct btrfs_extent_item_v0) == end) 30315d4f98a2SYan Zheng ptr = end; 30325d4f98a2SYan Zheng else 30335d4f98a2SYan Zheng #endif 30345d4f98a2SYan Zheng ptr += sizeof(struct btrfs_extent_item); 30355d4f98a2SYan Zheng 30365d4f98a2SYan Zheng while (ptr < end) { 30375d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 30385d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 30395d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 30405d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 30415d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 30425d4f98a2SYan Zheng blocks); 30435d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 30445d4f98a2SYan Zheng dref = (struct btrfs_extent_data_ref *)(&iref->offset); 30455d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 30465d4f98a2SYan Zheng eb, dref, blocks); 30475d4f98a2SYan Zheng } else { 30485d4f98a2SYan Zheng BUG(); 30495d4f98a2SYan Zheng } 30505d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 30515d4f98a2SYan Zheng } 30525d4f98a2SYan Zheng WARN_ON(ptr > end); 30535d4f98a2SYan Zheng 30545d4f98a2SYan Zheng while (1) { 30555d4f98a2SYan Zheng cond_resched(); 30565d4f98a2SYan Zheng eb = path->nodes[0]; 30575d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(eb)) { 30585d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 30595d4f98a2SYan Zheng if (ret < 0) { 30605d4f98a2SYan Zheng err = ret; 30615d4f98a2SYan Zheng break; 30625d4f98a2SYan Zheng } 30635d4f98a2SYan Zheng if (ret > 0) 30645d4f98a2SYan Zheng break; 30655d4f98a2SYan Zheng eb = path->nodes[0]; 30665d4f98a2SYan Zheng } 30675d4f98a2SYan Zheng 30685d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path->slots[0]); 30695d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 30705d4f98a2SYan Zheng break; 30715d4f98a2SYan Zheng 30725d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 30735d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY || 30745d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 30755d4f98a2SYan Zheng #else 30765d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 30775d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 30785d4f98a2SYan Zheng #endif 30795d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 30805d4f98a2SYan Zheng blocks); 30815d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 30825d4f98a2SYan Zheng dref = btrfs_item_ptr(eb, path->slots[0], 30835d4f98a2SYan Zheng struct btrfs_extent_data_ref); 30845d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 30855d4f98a2SYan Zheng eb, dref, blocks); 30865d4f98a2SYan Zheng } else { 30875d4f98a2SYan Zheng ret = 0; 30885d4f98a2SYan Zheng } 30895d4f98a2SYan Zheng if (ret) { 30905d4f98a2SYan Zheng err = ret; 30915d4f98a2SYan Zheng break; 30925d4f98a2SYan Zheng } 30935d4f98a2SYan Zheng path->slots[0]++; 30945d4f98a2SYan Zheng } 30955d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 30965d4f98a2SYan Zheng if (err) 30975d4f98a2SYan Zheng free_block_list(blocks); 30985d4f98a2SYan Zheng return err; 30995d4f98a2SYan Zheng } 31005d4f98a2SYan Zheng 31015d4f98a2SYan Zheng /* 31025d4f98a2SYan Zheng * hepler to find next unprocessed extent 31035d4f98a2SYan Zheng */ 31045d4f98a2SYan Zheng static noinline_for_stack 31055d4f98a2SYan Zheng int find_next_extent(struct btrfs_trans_handle *trans, 31065d4f98a2SYan Zheng struct reloc_control *rc, struct btrfs_path *path) 31075d4f98a2SYan Zheng { 31085d4f98a2SYan Zheng struct btrfs_key key; 31095d4f98a2SYan Zheng struct extent_buffer *leaf; 31105d4f98a2SYan Zheng u64 start, end, last; 31115d4f98a2SYan Zheng int ret; 31125d4f98a2SYan Zheng 31135d4f98a2SYan Zheng last = rc->block_group->key.objectid + rc->block_group->key.offset; 31145d4f98a2SYan Zheng while (1) { 31155d4f98a2SYan Zheng cond_resched(); 31165d4f98a2SYan Zheng if (rc->search_start >= last) { 31175d4f98a2SYan Zheng ret = 1; 31185d4f98a2SYan Zheng break; 31195d4f98a2SYan Zheng } 31205d4f98a2SYan Zheng 31215d4f98a2SYan Zheng key.objectid = rc->search_start; 31225d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31235d4f98a2SYan Zheng key.offset = 0; 31245d4f98a2SYan Zheng 31255d4f98a2SYan Zheng path->search_commit_root = 1; 31265d4f98a2SYan Zheng path->skip_locking = 1; 31275d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 31285d4f98a2SYan Zheng 0, 0); 31295d4f98a2SYan Zheng if (ret < 0) 31305d4f98a2SYan Zheng break; 31315d4f98a2SYan Zheng next: 31325d4f98a2SYan Zheng leaf = path->nodes[0]; 31335d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 31345d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 31355d4f98a2SYan Zheng if (ret != 0) 31365d4f98a2SYan Zheng break; 31375d4f98a2SYan Zheng leaf = path->nodes[0]; 31385d4f98a2SYan Zheng } 31395d4f98a2SYan Zheng 31405d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 31415d4f98a2SYan Zheng if (key.objectid >= last) { 31425d4f98a2SYan Zheng ret = 1; 31435d4f98a2SYan Zheng break; 31445d4f98a2SYan Zheng } 31455d4f98a2SYan Zheng 31465d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_ITEM_KEY || 31475d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 31485d4f98a2SYan Zheng path->slots[0]++; 31495d4f98a2SYan Zheng goto next; 31505d4f98a2SYan Zheng } 31515d4f98a2SYan Zheng 31525d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 31535d4f98a2SYan Zheng key.objectid, &start, &end, 31545d4f98a2SYan Zheng EXTENT_DIRTY); 31555d4f98a2SYan Zheng 31565d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 31575d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 31585d4f98a2SYan Zheng rc->search_start = end + 1; 31595d4f98a2SYan Zheng } else { 31605d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 31615d4f98a2SYan Zheng return 0; 31625d4f98a2SYan Zheng } 31635d4f98a2SYan Zheng } 31645d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 31655d4f98a2SYan Zheng return ret; 31665d4f98a2SYan Zheng } 31675d4f98a2SYan Zheng 31685d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 31695d4f98a2SYan Zheng { 31705d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31715d4f98a2SYan Zheng mutex_lock(&fs_info->trans_mutex); 31725d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 31735d4f98a2SYan Zheng mutex_unlock(&fs_info->trans_mutex); 31745d4f98a2SYan Zheng } 31755d4f98a2SYan Zheng 31765d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 31775d4f98a2SYan Zheng { 31785d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 31795d4f98a2SYan Zheng mutex_lock(&fs_info->trans_mutex); 31805d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 31815d4f98a2SYan Zheng mutex_unlock(&fs_info->trans_mutex); 31825d4f98a2SYan Zheng } 31835d4f98a2SYan Zheng 31845d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 31855d4f98a2SYan Zheng { 31865d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 31875d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 31885d4f98a2SYan Zheng return 1; 31895d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 31905d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 31915d4f98a2SYan Zheng return 1; 31925d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 31935d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 31945d4f98a2SYan Zheng return 1; 31955d4f98a2SYan Zheng return 0; 31965d4f98a2SYan Zheng } 31975d4f98a2SYan Zheng 31985d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 31995d4f98a2SYan Zheng { 32005d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 32015d4f98a2SYan Zheng struct btrfs_key key; 32025d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 32035d4f98a2SYan Zheng struct btrfs_path *path; 32045d4f98a2SYan Zheng struct btrfs_extent_item *ei; 32055d4f98a2SYan Zheng unsigned long nr; 32065d4f98a2SYan Zheng u64 flags; 32075d4f98a2SYan Zheng u32 item_size; 32085d4f98a2SYan Zheng int ret; 32095d4f98a2SYan Zheng int err = 0; 32105d4f98a2SYan Zheng 32115d4f98a2SYan Zheng path = btrfs_alloc_path(); 32125d4f98a2SYan Zheng if (!path) 32135d4f98a2SYan Zheng return -ENOMEM; 32145d4f98a2SYan Zheng 32155d4f98a2SYan Zheng rc->search_start = rc->block_group->key.objectid; 32165d4f98a2SYan Zheng clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY, 32175d4f98a2SYan Zheng GFP_NOFS); 32185d4f98a2SYan Zheng 32195d4f98a2SYan Zheng rc->create_reloc_root = 1; 32205d4f98a2SYan Zheng set_reloc_control(rc); 32215d4f98a2SYan Zheng 32225d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 32235d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 32245d4f98a2SYan Zheng 32255d4f98a2SYan Zheng while (1) { 32265d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 32275d4f98a2SYan Zheng 32285d4f98a2SYan Zheng ret = find_next_extent(trans, rc, path); 32295d4f98a2SYan Zheng if (ret < 0) 32305d4f98a2SYan Zheng err = ret; 32315d4f98a2SYan Zheng if (ret != 0) 32325d4f98a2SYan Zheng break; 32335d4f98a2SYan Zheng 32345d4f98a2SYan Zheng rc->extents_found++; 32355d4f98a2SYan Zheng 32365d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 32375d4f98a2SYan Zheng struct btrfs_extent_item); 32385d4f98a2SYan Zheng btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 32395d4f98a2SYan Zheng item_size = btrfs_item_size_nr(path->nodes[0], 32405d4f98a2SYan Zheng path->slots[0]); 32415d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 32425d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 32435d4f98a2SYan Zheng ret = check_extent_flags(flags); 32445d4f98a2SYan Zheng BUG_ON(ret); 32455d4f98a2SYan Zheng 32465d4f98a2SYan Zheng } else { 32475d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 32485d4f98a2SYan Zheng u64 ref_owner; 32495d4f98a2SYan Zheng int path_change = 0; 32505d4f98a2SYan Zheng 32515d4f98a2SYan Zheng BUG_ON(item_size != 32525d4f98a2SYan Zheng sizeof(struct btrfs_extent_item_v0)); 32535d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, &key, &ref_owner, 32545d4f98a2SYan Zheng &path_change); 32555d4f98a2SYan Zheng if (ref_owner < BTRFS_FIRST_FREE_OBJECTID) 32565d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_TREE_BLOCK; 32575d4f98a2SYan Zheng else 32585d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_DATA; 32595d4f98a2SYan Zheng 32605d4f98a2SYan Zheng if (path_change) { 32615d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 32625d4f98a2SYan Zheng 32635d4f98a2SYan Zheng path->search_commit_root = 1; 32645d4f98a2SYan Zheng path->skip_locking = 1; 32655d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, 32665d4f98a2SYan Zheng &key, path, 0, 0); 32675d4f98a2SYan Zheng if (ret < 0) { 32685d4f98a2SYan Zheng err = ret; 32695d4f98a2SYan Zheng break; 32705d4f98a2SYan Zheng } 32715d4f98a2SYan Zheng BUG_ON(ret > 0); 32725d4f98a2SYan Zheng } 32735d4f98a2SYan Zheng #else 32745d4f98a2SYan Zheng BUG(); 32755d4f98a2SYan Zheng #endif 32765d4f98a2SYan Zheng } 32775d4f98a2SYan Zheng 32785d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 32795d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 32805d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 32815d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 32825d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 32835d4f98a2SYan Zheng } else { 32845d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 32855d4f98a2SYan Zheng ret = 0; 32865d4f98a2SYan Zheng } 32875d4f98a2SYan Zheng if (ret < 0) { 32885d4f98a2SYan Zheng err = 0; 32895d4f98a2SYan Zheng break; 32905d4f98a2SYan Zheng } 32915d4f98a2SYan Zheng 32925d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 32935d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 32945d4f98a2SYan Zheng if (ret < 0) { 32955d4f98a2SYan Zheng err = ret; 32965d4f98a2SYan Zheng break; 32975d4f98a2SYan Zheng } 32985d4f98a2SYan Zheng } 32995d4f98a2SYan Zheng 33005d4f98a2SYan Zheng nr = trans->blocks_used; 33015d4f98a2SYan Zheng btrfs_end_transaction_throttle(trans, rc->extent_root); 33025d4f98a2SYan Zheng trans = NULL; 33035d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 33045d4f98a2SYan Zheng 33055d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 33065d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 33075d4f98a2SYan Zheng rc->found_file_extent = 1; 33085d4f98a2SYan Zheng ret = relocate_data_extent(rc->data_inode, &key); 33095d4f98a2SYan Zheng if (ret < 0) { 33105d4f98a2SYan Zheng err = ret; 33115d4f98a2SYan Zheng break; 33125d4f98a2SYan Zheng } 33135d4f98a2SYan Zheng } 33145d4f98a2SYan Zheng } 33155d4f98a2SYan Zheng btrfs_free_path(path); 33165d4f98a2SYan Zheng 33175d4f98a2SYan Zheng if (trans) { 33185d4f98a2SYan Zheng nr = trans->blocks_used; 33195d4f98a2SYan Zheng btrfs_end_transaction(trans, rc->extent_root); 33205d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 33215d4f98a2SYan Zheng } 33225d4f98a2SYan Zheng 33235d4f98a2SYan Zheng rc->create_reloc_root = 0; 33245d4f98a2SYan Zheng smp_mb(); 33255d4f98a2SYan Zheng 33265d4f98a2SYan Zheng if (rc->extents_found > 0) { 33275d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 33285d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 33295d4f98a2SYan Zheng } 33305d4f98a2SYan Zheng 33315d4f98a2SYan Zheng merge_reloc_roots(rc); 33325d4f98a2SYan Zheng 33335d4f98a2SYan Zheng unset_reloc_control(rc); 33345d4f98a2SYan Zheng 33355d4f98a2SYan Zheng /* get rid of pinned extents */ 33365d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 33375d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 33385d4f98a2SYan Zheng 33395d4f98a2SYan Zheng return err; 33405d4f98a2SYan Zheng } 33415d4f98a2SYan Zheng 33425d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 33435d4f98a2SYan Zheng struct btrfs_root *root, 33445d4f98a2SYan Zheng u64 objectid, u64 size) 33455d4f98a2SYan Zheng { 33465d4f98a2SYan Zheng struct btrfs_path *path; 33475d4f98a2SYan Zheng struct btrfs_inode_item *item; 33485d4f98a2SYan Zheng struct extent_buffer *leaf; 33495d4f98a2SYan Zheng int ret; 33505d4f98a2SYan Zheng 33515d4f98a2SYan Zheng path = btrfs_alloc_path(); 33525d4f98a2SYan Zheng if (!path) 33535d4f98a2SYan Zheng return -ENOMEM; 33545d4f98a2SYan Zheng 33555d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 33565d4f98a2SYan Zheng if (ret) 33575d4f98a2SYan Zheng goto out; 33585d4f98a2SYan Zheng 33595d4f98a2SYan Zheng leaf = path->nodes[0]; 33605d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 33615d4f98a2SYan Zheng memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item)); 33625d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 33635d4f98a2SYan Zheng btrfs_set_inode_size(leaf, item, size); 33645d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 33655d4f98a2SYan Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS); 33665d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 33675d4f98a2SYan Zheng btrfs_release_path(root, path); 33685d4f98a2SYan Zheng out: 33695d4f98a2SYan Zheng btrfs_free_path(path); 33705d4f98a2SYan Zheng return ret; 33715d4f98a2SYan Zheng } 33725d4f98a2SYan Zheng 33735d4f98a2SYan Zheng /* 33745d4f98a2SYan Zheng * helper to create inode for data relocation. 33755d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 33765d4f98a2SYan Zheng */ 33775d4f98a2SYan Zheng static struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 33785d4f98a2SYan Zheng struct btrfs_block_group_cache *group) 33795d4f98a2SYan Zheng { 33805d4f98a2SYan Zheng struct inode *inode = NULL; 33815d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 33825d4f98a2SYan Zheng struct btrfs_root *root; 33835d4f98a2SYan Zheng struct btrfs_key key; 33845d4f98a2SYan Zheng unsigned long nr; 33855d4f98a2SYan Zheng u64 objectid = BTRFS_FIRST_FREE_OBJECTID; 33865d4f98a2SYan Zheng int err = 0; 33875d4f98a2SYan Zheng 33885d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 33895d4f98a2SYan Zheng if (IS_ERR(root)) 33905d4f98a2SYan Zheng return ERR_CAST(root); 33915d4f98a2SYan Zheng 33925d4f98a2SYan Zheng trans = btrfs_start_transaction(root, 1); 33935d4f98a2SYan Zheng BUG_ON(!trans); 33945d4f98a2SYan Zheng 33955d4f98a2SYan Zheng err = btrfs_find_free_objectid(trans, root, objectid, &objectid); 33965d4f98a2SYan Zheng if (err) 33975d4f98a2SYan Zheng goto out; 33985d4f98a2SYan Zheng 33995d4f98a2SYan Zheng err = __insert_orphan_inode(trans, root, objectid, group->key.offset); 34005d4f98a2SYan Zheng BUG_ON(err); 34015d4f98a2SYan Zheng 34025d4f98a2SYan Zheng err = btrfs_insert_file_extent(trans, root, objectid, 0, 0, 0, 34035d4f98a2SYan Zheng group->key.offset, 0, group->key.offset, 34045d4f98a2SYan Zheng 0, 0, 0); 34055d4f98a2SYan Zheng BUG_ON(err); 34065d4f98a2SYan Zheng 34075d4f98a2SYan Zheng key.objectid = objectid; 34085d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 34095d4f98a2SYan Zheng key.offset = 0; 34105d4f98a2SYan Zheng inode = btrfs_iget(root->fs_info->sb, &key, root); 34115d4f98a2SYan Zheng BUG_ON(IS_ERR(inode) || is_bad_inode(inode)); 34125d4f98a2SYan Zheng BTRFS_I(inode)->index_cnt = group->key.objectid; 34135d4f98a2SYan Zheng 34145d4f98a2SYan Zheng err = btrfs_orphan_add(trans, inode); 34155d4f98a2SYan Zheng out: 34165d4f98a2SYan Zheng nr = trans->blocks_used; 34175d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 34185d4f98a2SYan Zheng 34195d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 34205d4f98a2SYan Zheng if (err) { 34215d4f98a2SYan Zheng if (inode) 34225d4f98a2SYan Zheng iput(inode); 34235d4f98a2SYan Zheng inode = ERR_PTR(err); 34245d4f98a2SYan Zheng } 34255d4f98a2SYan Zheng return inode; 34265d4f98a2SYan Zheng } 34275d4f98a2SYan Zheng 34285d4f98a2SYan Zheng /* 34295d4f98a2SYan Zheng * function to relocate all extents in a block group. 34305d4f98a2SYan Zheng */ 34315d4f98a2SYan Zheng int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start) 34325d4f98a2SYan Zheng { 34335d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = extent_root->fs_info; 34345d4f98a2SYan Zheng struct reloc_control *rc; 34355d4f98a2SYan Zheng int ret; 34365d4f98a2SYan Zheng int err = 0; 34375d4f98a2SYan Zheng 34385d4f98a2SYan Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 34395d4f98a2SYan Zheng if (!rc) 34405d4f98a2SYan Zheng return -ENOMEM; 34415d4f98a2SYan Zheng 34425d4f98a2SYan Zheng mapping_tree_init(&rc->reloc_root_tree); 34435d4f98a2SYan Zheng extent_io_tree_init(&rc->processed_blocks, NULL, GFP_NOFS); 34445d4f98a2SYan Zheng INIT_LIST_HEAD(&rc->reloc_roots); 34455d4f98a2SYan Zheng 34465d4f98a2SYan Zheng rc->block_group = btrfs_lookup_block_group(fs_info, group_start); 34475d4f98a2SYan Zheng BUG_ON(!rc->block_group); 34485d4f98a2SYan Zheng 34495d4f98a2SYan Zheng btrfs_init_workers(&rc->workers, "relocate", 34505d4f98a2SYan Zheng fs_info->thread_pool_size); 34515d4f98a2SYan Zheng 34525d4f98a2SYan Zheng rc->extent_root = extent_root; 34535d4f98a2SYan Zheng btrfs_prepare_block_group_relocation(extent_root, rc->block_group); 34545d4f98a2SYan Zheng 34555d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 34565d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 34575d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 34585d4f98a2SYan Zheng rc->data_inode = NULL; 34595d4f98a2SYan Zheng goto out; 34605d4f98a2SYan Zheng } 34615d4f98a2SYan Zheng 34625d4f98a2SYan Zheng printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n", 34635d4f98a2SYan Zheng (unsigned long long)rc->block_group->key.objectid, 34645d4f98a2SYan Zheng (unsigned long long)rc->block_group->flags); 34655d4f98a2SYan Zheng 34665d4f98a2SYan Zheng btrfs_start_delalloc_inodes(fs_info->tree_root); 34675d4f98a2SYan Zheng btrfs_wait_ordered_extents(fs_info->tree_root, 0); 34685d4f98a2SYan Zheng 34695d4f98a2SYan Zheng while (1) { 34705d4f98a2SYan Zheng mutex_lock(&fs_info->cleaner_mutex); 34715d4f98a2SYan Zheng btrfs_clean_old_snapshots(fs_info->tree_root); 34725d4f98a2SYan Zheng mutex_unlock(&fs_info->cleaner_mutex); 34735d4f98a2SYan Zheng 34745d4f98a2SYan Zheng rc->extents_found = 0; 34755d4f98a2SYan Zheng rc->extents_skipped = 0; 34765d4f98a2SYan Zheng 34775d4f98a2SYan Zheng ret = relocate_block_group(rc); 34785d4f98a2SYan Zheng if (ret < 0) { 34795d4f98a2SYan Zheng err = ret; 34805d4f98a2SYan Zheng break; 34815d4f98a2SYan Zheng } 34825d4f98a2SYan Zheng 34835d4f98a2SYan Zheng if (rc->extents_found == 0) 34845d4f98a2SYan Zheng break; 34855d4f98a2SYan Zheng 34865d4f98a2SYan Zheng printk(KERN_INFO "btrfs: found %llu extents\n", 34875d4f98a2SYan Zheng (unsigned long long)rc->extents_found); 34885d4f98a2SYan Zheng 34895d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 34905d4f98a2SYan Zheng btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1); 34915d4f98a2SYan Zheng invalidate_mapping_pages(rc->data_inode->i_mapping, 34925d4f98a2SYan Zheng 0, -1); 34935d4f98a2SYan Zheng rc->stage = UPDATE_DATA_PTRS; 34945d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 34955d4f98a2SYan Zheng rc->extents_skipped >= rc->extents_found) { 34965d4f98a2SYan Zheng iput(rc->data_inode); 34975d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, 34985d4f98a2SYan Zheng rc->block_group); 34995d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 35005d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 35015d4f98a2SYan Zheng rc->data_inode = NULL; 35025d4f98a2SYan Zheng break; 35035d4f98a2SYan Zheng } 35045d4f98a2SYan Zheng rc->stage = MOVE_DATA_EXTENTS; 35055d4f98a2SYan Zheng rc->found_file_extent = 0; 35065d4f98a2SYan Zheng } 35075d4f98a2SYan Zheng } 35085d4f98a2SYan Zheng 35095d4f98a2SYan Zheng filemap_fdatawrite_range(fs_info->btree_inode->i_mapping, 35105d4f98a2SYan Zheng rc->block_group->key.objectid, 35115d4f98a2SYan Zheng rc->block_group->key.objectid + 35125d4f98a2SYan Zheng rc->block_group->key.offset - 1); 35135d4f98a2SYan Zheng 35145d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 35155d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 35165d4f98a2SYan Zheng WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0); 35175d4f98a2SYan Zheng out: 35185d4f98a2SYan Zheng iput(rc->data_inode); 35195d4f98a2SYan Zheng btrfs_stop_workers(&rc->workers); 35205d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 35215d4f98a2SYan Zheng kfree(rc); 35225d4f98a2SYan Zheng return err; 35235d4f98a2SYan Zheng } 35245d4f98a2SYan Zheng 35255d4f98a2SYan Zheng /* 35265d4f98a2SYan Zheng * recover relocation interrupted by system crash. 35275d4f98a2SYan Zheng * 35285d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 35295d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 35305d4f98a2SYan Zheng */ 35315d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 35325d4f98a2SYan Zheng { 35335d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 35345d4f98a2SYan Zheng struct btrfs_key key; 35355d4f98a2SYan Zheng struct btrfs_root *fs_root; 35365d4f98a2SYan Zheng struct btrfs_root *reloc_root; 35375d4f98a2SYan Zheng struct btrfs_path *path; 35385d4f98a2SYan Zheng struct extent_buffer *leaf; 35395d4f98a2SYan Zheng struct reloc_control *rc = NULL; 35405d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 35415d4f98a2SYan Zheng int ret; 35425d4f98a2SYan Zheng int err = 0; 35435d4f98a2SYan Zheng 35445d4f98a2SYan Zheng path = btrfs_alloc_path(); 35455d4f98a2SYan Zheng if (!path) 35465d4f98a2SYan Zheng return -ENOMEM; 35475d4f98a2SYan Zheng 35485d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 35495d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 35505d4f98a2SYan Zheng key.offset = (u64)-1; 35515d4f98a2SYan Zheng 35525d4f98a2SYan Zheng while (1) { 35535d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, 35545d4f98a2SYan Zheng path, 0, 0); 35555d4f98a2SYan Zheng if (ret < 0) { 35565d4f98a2SYan Zheng err = ret; 35575d4f98a2SYan Zheng goto out; 35585d4f98a2SYan Zheng } 35595d4f98a2SYan Zheng if (ret > 0) { 35605d4f98a2SYan Zheng if (path->slots[0] == 0) 35615d4f98a2SYan Zheng break; 35625d4f98a2SYan Zheng path->slots[0]--; 35635d4f98a2SYan Zheng } 35645d4f98a2SYan Zheng leaf = path->nodes[0]; 35655d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 35665d4f98a2SYan Zheng btrfs_release_path(root->fs_info->tree_root, path); 35675d4f98a2SYan Zheng 35685d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 35695d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 35705d4f98a2SYan Zheng break; 35715d4f98a2SYan Zheng 35725d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root, &key); 35735d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 35745d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 35755d4f98a2SYan Zheng goto out; 35765d4f98a2SYan Zheng } 35775d4f98a2SYan Zheng 35785d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 35795d4f98a2SYan Zheng 35805d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 35815d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 35825d4f98a2SYan Zheng reloc_root->root_key.offset); 35835d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 35845d4f98a2SYan Zheng err = PTR_ERR(fs_root); 35855d4f98a2SYan Zheng goto out; 35865d4f98a2SYan Zheng } 35875d4f98a2SYan Zheng } 35885d4f98a2SYan Zheng 35895d4f98a2SYan Zheng if (key.offset == 0) 35905d4f98a2SYan Zheng break; 35915d4f98a2SYan Zheng 35925d4f98a2SYan Zheng key.offset--; 35935d4f98a2SYan Zheng } 35945d4f98a2SYan Zheng btrfs_release_path(root->fs_info->tree_root, path); 35955d4f98a2SYan Zheng 35965d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 35975d4f98a2SYan Zheng goto out; 35985d4f98a2SYan Zheng 35995d4f98a2SYan Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 36005d4f98a2SYan Zheng if (!rc) { 36015d4f98a2SYan Zheng err = -ENOMEM; 36025d4f98a2SYan Zheng goto out; 36035d4f98a2SYan Zheng } 36045d4f98a2SYan Zheng 36055d4f98a2SYan Zheng mapping_tree_init(&rc->reloc_root_tree); 36065d4f98a2SYan Zheng INIT_LIST_HEAD(&rc->reloc_roots); 36075d4f98a2SYan Zheng btrfs_init_workers(&rc->workers, "relocate", 36085d4f98a2SYan Zheng root->fs_info->thread_pool_size); 36095d4f98a2SYan Zheng rc->extent_root = root->fs_info->extent_root; 36105d4f98a2SYan Zheng 36115d4f98a2SYan Zheng set_reloc_control(rc); 36125d4f98a2SYan Zheng 36135d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 36145d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 36155d4f98a2SYan Zheng struct btrfs_root, root_list); 36165d4f98a2SYan Zheng list_del(&reloc_root->root_list); 36175d4f98a2SYan Zheng 36185d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 36195d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 36205d4f98a2SYan Zheng &rc->reloc_roots); 36215d4f98a2SYan Zheng continue; 36225d4f98a2SYan Zheng } 36235d4f98a2SYan Zheng 36245d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 36255d4f98a2SYan Zheng reloc_root->root_key.offset); 36265d4f98a2SYan Zheng BUG_ON(IS_ERR(fs_root)); 36275d4f98a2SYan Zheng 36285d4f98a2SYan Zheng __add_reloc_root(reloc_root); 36295d4f98a2SYan Zheng fs_root->reloc_root = reloc_root; 36305d4f98a2SYan Zheng } 36315d4f98a2SYan Zheng 36325d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 36335d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 36345d4f98a2SYan Zheng 36355d4f98a2SYan Zheng merge_reloc_roots(rc); 36365d4f98a2SYan Zheng 36375d4f98a2SYan Zheng unset_reloc_control(rc); 36385d4f98a2SYan Zheng 36395d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 36405d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 36415d4f98a2SYan Zheng out: 36425d4f98a2SYan Zheng if (rc) { 36435d4f98a2SYan Zheng btrfs_stop_workers(&rc->workers); 36445d4f98a2SYan Zheng kfree(rc); 36455d4f98a2SYan Zheng } 36465d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 36475d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 36485d4f98a2SYan Zheng struct btrfs_root, root_list); 36495d4f98a2SYan Zheng list_del(&reloc_root->root_list); 36505d4f98a2SYan Zheng free_extent_buffer(reloc_root->node); 36515d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 36525d4f98a2SYan Zheng kfree(reloc_root); 36535d4f98a2SYan Zheng } 36545d4f98a2SYan Zheng btrfs_free_path(path); 36555d4f98a2SYan Zheng 36565d4f98a2SYan Zheng if (err == 0) { 36575d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 36585d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 36595d4f98a2SYan Zheng BTRFS_DATA_RELOC_TREE_OBJECTID); 36605d4f98a2SYan Zheng if (IS_ERR(fs_root)) 36615d4f98a2SYan Zheng err = PTR_ERR(fs_root); 36625d4f98a2SYan Zheng } 36635d4f98a2SYan Zheng return err; 36645d4f98a2SYan Zheng } 36655d4f98a2SYan Zheng 36665d4f98a2SYan Zheng /* 36675d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 36685d4f98a2SYan Zheng * 36695d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 36705d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 36715d4f98a2SYan Zheng */ 36725d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 36735d4f98a2SYan Zheng { 36745d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 36755d4f98a2SYan Zheng struct btrfs_sector_sum *sector_sum; 36765d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 36775d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 36785d4f98a2SYan Zheng size_t offset; 36795d4f98a2SYan Zheng int ret; 36805d4f98a2SYan Zheng u64 disk_bytenr; 36815d4f98a2SYan Zheng LIST_HEAD(list); 36825d4f98a2SYan Zheng 36835d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 36845d4f98a2SYan Zheng BUG_ON(ordered->file_offset != file_pos || ordered->len != len); 36855d4f98a2SYan Zheng 36865d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 36875d4f98a2SYan Zheng ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr, 36885d4f98a2SYan Zheng disk_bytenr + len - 1, &list); 36895d4f98a2SYan Zheng 36905d4f98a2SYan Zheng while (!list_empty(&list)) { 36915d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 36925d4f98a2SYan Zheng list_del_init(&sums->list); 36935d4f98a2SYan Zheng 36945d4f98a2SYan Zheng sector_sum = sums->sums; 36955d4f98a2SYan Zheng sums->bytenr = ordered->start; 36965d4f98a2SYan Zheng 36975d4f98a2SYan Zheng offset = 0; 36985d4f98a2SYan Zheng while (offset < sums->len) { 36995d4f98a2SYan Zheng sector_sum->bytenr += ordered->start - disk_bytenr; 37005d4f98a2SYan Zheng sector_sum++; 37015d4f98a2SYan Zheng offset += root->sectorsize; 37025d4f98a2SYan Zheng } 37035d4f98a2SYan Zheng 37045d4f98a2SYan Zheng btrfs_add_ordered_sum(inode, ordered, sums); 37055d4f98a2SYan Zheng } 37065d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 37075d4f98a2SYan Zheng return 0; 37085d4f98a2SYan Zheng } 3709